diff --git a/_multibuild b/_multibuild index 606536d..a8e6772 100644 --- a/_multibuild +++ b/_multibuild @@ -1,6 +1,6 @@ - standard - testsuite + ucx + ucx-testsuite ofi ofi-testsuite diff --git a/ch4-shm-fix-data-type-for-recv_bytes-in-MPIDI_POSIX_mpi_release_gather_release.patch b/ch4-shm-fix-data-type-for-recv_bytes-in-MPIDI_POSIX_mpi_release_gather_release.patch new file mode 100644 index 0000000..282f867 --- /dev/null +++ b/ch4-shm-fix-data-type-for-recv_bytes-in-MPIDI_POSIX_mpi_release_gather_release.patch @@ -0,0 +1,47 @@ +commit 564c0affae6a79b3c99fce5717c402182d74daa1 +Author: Nicolas Morey +Date: Sat Nov 8 23:34:58 2025 +0100 + + ch4: shm: fix data type for recv_bytes in MPIDI_POSIX_mpi_release_gather_release + + The number of received bytes in release_gather_release is badly cast between + int and MPI_Aint. On most arch this is not an issue, but for Big-Endian 64b arch (s390x) + it ends up losing the actual value as we only copy the first 4 MSB. + Fix the issue by writing the whole MPI_AInt in the shm_buf instead of just an int. + + Signed-off-by: Nicolas Morey + +diff --git src/mpid/ch4/shm/posix/release_gather/release_gather.h src/mpid/ch4/shm/posix/release_gather/release_gather.h +index ac966cb9772e..ff1308830d00 100644 +--- src/mpid/ch4/shm/posix/release_gather/release_gather.h ++++ src/mpid/ch4/shm/posix/release_gather/release_gather.h +@@ -121,7 +121,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_POSIX_mpi_release_gather_release(void *local_ + datatype, root, MPIR_BCAST_TAG, comm_ptr, &status); + MPIR_ERR_CHECK(mpi_errno); + MPIR_Get_count_impl(&status, MPI_BYTE, &recv_bytes); +- MPIR_Typerep_copy(bcast_data_addr, &recv_bytes, sizeof(int), ++ MPIR_Typerep_copy(bcast_data_addr, &recv_bytes, sizeof(MPI_Aint), + MPIR_TYPEREP_FLAG_NONE); + /* It is necessary to copy the errflag as well to handle the case when non-root + * becomes temporary root as part of compositions (or smp aware colls). These temp +@@ -146,7 +146,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_POSIX_mpi_release_gather_release(void *local_ + /* When error checking is enabled, place the datasize in shm_buf first, followed by the + * errflag, followed by the actual data with an offset of (2*cacheline_size) bytes from + * the starting address */ +- MPIR_Typerep_copy(bcast_data_addr, &count, sizeof(int), MPIR_TYPEREP_FLAG_NONE); ++ MPIR_Typerep_copy(bcast_data_addr, &count, sizeof(MPI_Aint), MPIR_TYPEREP_FLAG_NONE); + /* It is necessary to copy the errflag as well to handle the case when non-root + * becomes root as part of compositions (or smp aware colls). These roots might + * expect same data as other ranks but different from the actual root. So only +@@ -218,8 +218,9 @@ MPL_STATIC_INLINE_PREFIX int MPIDI_POSIX_mpi_release_gather_release(void *local_ + * datasize is copied out from shm_buffer and compared against the count a rank was + * expecting. Also, the errflag is copied out. In case of mismatch mpi_errno is set. + * Actual data starts after (2*cacheline_size) bytes */ +- int recv_bytes, recv_errflag; +- MPIR_Typerep_copy(&recv_bytes, bcast_data_addr, sizeof(int), MPIR_TYPEREP_FLAG_NONE); ++ MPI_Aint recv_bytes; ++ int recv_errflag; ++ MPIR_Typerep_copy(&recv_bytes, bcast_data_addr, sizeof(MPI_Aint), MPIR_TYPEREP_FLAG_NONE); + MPIR_Typerep_copy(&recv_errflag, (char *) bcast_data_addr + MPIDU_SHM_CACHE_LINE_LEN, + sizeof(int), MPIR_TYPEREP_FLAG_NONE); + MPIR_ERR_CHKANDJUMP2(recv_bytes != count, mpi_errno, MPI_ERR_OTHER, diff --git a/mpich.changes b/mpich.changes index 2aed6ed..d15cfe4 100644 --- a/mpich.changes +++ b/mpich.changes @@ -1,3 +1,14 @@ +------------------------------------------------------------------- +Tue Nov 4 17:54:18 UTC 2025 - Nicolas Morey + +- Use external hwloc +- Rename standard multibuild flavor to ucx for more clarity +- Add ch4-shm-fix-data-type-for-recv_bytes-in-MPIDI_POSIX_mpi_release_gather_release.patch + to fix a datatype issue on s390x +- Speed up testsuite by not rebuilding mpich completely +- Fix random configure failure on aarch64 +- Fix FFLAGS + ------------------------------------------------------------------- Tue Oct 28 17:54:04 UTC 2025 - Nicolas Morey diff --git a/mpich.spec b/mpich.spec index 9513b1a..6752049 100644 --- a/mpich.spec +++ b/mpich.spec @@ -29,10 +29,10 @@ ExclusiveArch: do_not_build %endif -%if "%{flavor}" == "standard" +%if "%{flavor}" == "ucx" %define build_flavor ucx %endif -%if "%{flavor}" == "testsuite" +%if "%{flavor}" == "ucx-testsuite" %define build_flavor ucx %define testsuite 1 %endif @@ -84,6 +84,7 @@ Source101: README.md Patch1: autogen-only-deal-with-json-yaksa-if-enabled.patch Patch2: autoconf-pull-dynamic-and-not-static-libs-from-pkg-config.patch Patch3: romio-test-fix-bad-snprintf-arguments.patch +Patch4: ch4-shm-fix-data-type-for-recv_bytes-in-MPIDI_POSIX_mpi_release_gather_release.patch BuildRoot: %{_tmppath}/%{name}-%{version}-build BuildRequires: fdupes @@ -101,6 +102,7 @@ BuildRequires: libtool BuildRequires: libtool BuildRequires: mpi-selector BuildRequires: python3-devel +BuildRequires: hwloc-devel >= 2.0 %if "%{build_flavor}" == "ofi" BuildRequires: libfabric-devel @@ -114,7 +116,6 @@ BuildRequires: libuct-devel >= 1.7.0 # UCX is only available for 64b archs ExcludeArch: %ix86 %arm %endif - Provides: mpi BuildRequires: Modules BuildRequires: gcc-c++ @@ -187,11 +188,15 @@ rm -R modules/{ucx,libfabric,json-c} # GCC10 needs an extra flag to allow badly passed parameters %if 0%{?suse_version} > 1500 -export FFLAGS="-fallow-argument-mismatch $FFLAGS" -export FCFLAGS="-fallow-argument-mismatch $FCFLAGS" +export FFLAGS="-fallow-argument-mismatch %{optflags}" +export FCFLAGS="-fallow-argument-mismatch %{optflags}" %endif - -./autogen.sh --without-ucx --without-ofi --without-json +%ifarch aarch64 +# For some reason, configure has random issue with defining this +# on aarch64 only. Set it to avoid random failure +export CROSS_F77_SIZEOF_INTEGER=4 +%endif +autoreconf -fi %configure \ --prefix=%{p_prefix} \ --exec-prefix=%{p_prefix} \ @@ -205,6 +210,7 @@ export FCFLAGS="-fallow-argument-mismatch $FCFLAGS" --sysconfdir=%{p_sysconfdir} \ --disable-rpath \ --disable-wrapper-rpath \ + --with-hwloc \ %if "%{build_flavor}" == "ofi" --with-ofi \ --with-device=ch4:ofi \ @@ -218,6 +224,21 @@ export FCFLAGS="-fallow-argument-mismatch $FCFLAGS" MPICHLIB_CFLAGS="%{optflags}" \ MPICHLIB_CXXFLAGS="%{optflags}" +%if 0%{?testsuite} +%install +rm -rf %{buildroot}/* + +%check +# Disable CMA. Modern kernels require specific ptrace capabilities +# that are not available in OBS +export MPIR_CVAR_CH4_CMA_ENABLE=0 +for dir in src/mpl src/mpi/romio/test; do + ( + cd $dir && make check + ) +done + +%else make %{?_smp_mflags} VERBOSE=1 %install @@ -251,16 +272,6 @@ find %{buildroot} -name "*.a" -delete %fdupes %{buildroot}%{p_datadir} %fdupes %{buildroot}%{p_libdir}/pkgconfig -%if 0%{?testsuite} -# Remove everything from testsuite package -# It is all contained by mpich packages -rm -rf %{buildroot}/* - -%check -make check - -%else - # make and install mpivars files install -m 0755 -d %{buildroot}%{_bindir} sed -e 's,prefix,%p_prefix,g' -e 's,libdir,%{p_libdir},g' %{S:1} > %{buildroot}%{p_bindir}/mpivars.sh