GitLab

This fixes several memory leaks, improves the test suite, and add support for the ARMCI profiling interface.

r9650 fixed this in the corresponding Makefile.mk, so this file is no
longer needed for reference.

No reviewer.

It's causing hangs and segfaults in a few tests that passed for me on my
laptop.  This isn't worth holding up the 1.5b1 release for, so I'm
reverting this for now until I can cook up a version that passes the
nightlies.

No reviewer.

Reviewed by buntinas@.

This is a copy of the non-SMP-aware code in MPIR_Allreduce_intra that
has been modified to perform rank translation immediately prior to any
communication calls.  For cases where incremental group rank translation
is abnormally slow compared to batch translation, this is a non-optimal
method.  A better approach would be to compute the list of ranks that we
communicate with and then translate them all at once.  Such an approach
makes the code much uglier though, so I'm punting on that until we
experience such a case.

No reviewer.

were two problems.

1. We were waiting till the proxy connected back to us, which will
never happen.

2. We were waiting for all the proxy launching threads to naturally
exit.  This will never happen if, for example, the ssh hangs.

We switched shared library support on Darwin back to using
"-Wl,-flat_namespace" by default, so this necessitated a
non-backwards-compatible ABI bump from 5:0:0 to 6:0:0.

No reviewer.

[svn-r9745] Removed bonus comma from end of enum definition in mpi.h.in to supress gcc -pedantic warnings.

Previously we were just getting lucky about the ordering.

No reviewer.

[svn-r9740] Revert the attribute fix, including an explanation of the problem so that a full fix can be provided before the next release

Also add a FIXME for the ordering of the man & www dir creation rules.

No reviewer.

[svn-r9732] Fixes bug in r9707 where a line for NEEDS_GETPAGESIZE_DECL was not generated in config.h.in because of the way hwloc is configured.

deadlock prevention fix added in #9590.  This splits the context ID space into
eager and safe segments and replaces the barrier from #9590 with a collective
attempt to allocate from the eager segment.  If threads block on eager
allocation, un-blocked threads will still be able to proceed and allocate
context IDs from the safe segment, avoiding the hold-and-wait deadlock scenario
possible in the pre-#9590 algorithm.

Reviewer: goodell

[svn-r9720] Fixes for Fortran INTEGER types, along with proper handling of output arguments for some of the new MPI-3 routines with Fortran interfaces

This is fallout from tt#1591/r9647.  We now were failing handle leak
checks when running these tests.

No reviewer.

No reviewer.

Without this change we are vulnerable to a nasty bit of compiler
behavior.  MPIDI_CH3_PktGeneric_t sometimes included padding between the
"kind" and "pktptrs" fields.  Then whole-structure assignments using the
generic packet (e.g., ch3_istartmsg.c:96) of buffers that are actually
MPIDI_CH3_Pkt_t unions under the hood will can fail to copy key parts of
the MPIDI_CH3_Pkt_t.

This specifically manifested itself for #1599 by transmitting garbage
"source_win_handle" values in a LOCK_GRANTED packet, which in turn
caused segfaults.

Removing this second type and directly using the actual packet union
eliminates this source of bugs.  The minor downside is that the
mpidpkt.h header becomes more tightly integrated into the rest of the
device.

Reviewed by buntinas@.

No reviewer.

[svn-r9690] disable yielding for linux by default because usleep(0) doesn't work well as a substitute for sched_yield

The problem is that the (previously on by default) collective error
propagation functionality is implemented in a way that is fundamentally
incompatible with the way that nonblocking collectives are implemented.

Thanks to Jed Brown for the bug report and reproducer.

No reviewer.

Based on discussion with Neil Fortner on the opa-core@ list and careful
(although by no means bulletproof) discussion, Darius and I came up with
this new set of barriers.  The new barriers shouldn't be any slower or
faster, but they avoid an insidious bug that Neil found.

Reviewed by buntinas@.

 + Fixed types in allocation routines to match new armci_size_t interface
 + Fixed a few memory leaks