Per discussion on https://github.com/open-mpi/ompi/pull/1767 (and some
subsequent phone calls and off-issue email discussions), the PSM
library is hijacking signal handlers by default. Specifically: unless
the environment variables `IPATH_NO_BACKTRACE=1` (for PSM / Intel
TrueScale) is set, the library constructor for this library will
hijack various signal handlers for the purpose of invoking its own
error reporting mechanisms.
This may be a bit *surprising*, but is not a *problem*, per se. The
real problem is that older versions of at least the PSM library do not
unregister these signal handlers upon being unloaded from memory.
Hence, a segv can actually result in a double segv (i.e., the original
segv and then another segv when the now-non-existent signal handler is
invoked).
This PSM signal hijacking subverts Open MPI's own signal reporting
mechanism, which may be a bit surprising for some users (particularly
those who do not have Intel TrueScale). As such, we disable it by
default so that Open MPI's own error-reporting mechanisms are used.
Additionally, there is a typo in the library destructor for the PSM2
library that may cause problems in the unloading of its signal
handlers. This problem can be avoided by setting `HFI_NO_BACKTRACE=1`
(for PSM2 / Intel OmniPath).
This is further compounded by the fact that the PSM / PSM2 libraries
can be loaded by the OFI MTL and the usNIC BTL (because they are
loaded by libfabric), even when there is no Intel networking hardware
present. Having the PSM/PSM2 libraries behave this way when no Intel
hardware is present is clearly undesirable (and is likely to be fixed
in future releases of the PSM/PSM2 libraries).
This commit sets the following two environment variables to disable
this behavior from the PSM/PSM2 libraries (if they are not already
set):
* IPATH_NO_BACKTRACE=1
* HFI_NO_BACKTRACE=1
If the user has set these variables before invoking Open MPI, we will
not override their values (i.e., their preferences will be honored).
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
This commit adds the opal_atomic_swap_32 and opal_atomic_swap_64
functions. This should improve the performance of btl/vader.
Signed-off-by: Nathan Hjelm <hjelmn@me.com>
Before dynamic add_procs the openib_btl_size_queues was called exactly
once for non-dynamic jobs. Now the function is called on each new
connection so the calculation was wrong. Re-wrote the function to
correctly calculate the CQ size and only attempt to adjust the CQ if
the requested size has changed. This fixes a bug when using the openib
btl on psm2 hardware that is caused by the time needed to resize a
CQ. The overhead was causing udcm to timeout and fail.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit improves the CMA detection when the installed glibc doesn't
have support for CMA. In this case we need to verify that the syscall
numbers in opal/include/opal/sys/cma.h are valid for the architecture.
This verification is done by attempting to use CMA while including the
internal header.
Signed-off-by: Nathan Hjelm <hjelmn@me.com>
Compiler implementations are free to include support for atomics that
use locks. Unfortunately lock-free and lock atomics do not mix. Older
versions of llvm on OS X use locks to provide
__atomic_compare_exchange on 128-bit values but are lock-free on
64-bit values. This screws up our lifo implementation which mixes
64-bit and 128-bit atomics on the same values to improve
performance. This commit adds a configure-time check if 128-bit
atomics are lock free. If they are not then the 128-bit __atomic CAS
is disabled and we check for the __sync version as a fallback.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
clang 7.0 with the picky option on is extremely verbose, and complains
about almost everything. Trying to make him happy, at least regarding
the datatype engine.
This commit fixes a bug in opal progress registration that can cause
crashes when a progress function is registered while another thread is
in opal_progress(). Before this commit realloc is used to allocate
more space for progress functions but it is possible for a thread in
opal_progress() to try to read from the array that is freed by realloc
before the array is re-assigned when realloc returns. To prevent this
race use malloc + memcpy to fill the new array and atomically swap out
the old and new array pointers.
Per suggestion we now allocate a default of 8 slots for callbacks and
double the current number when we run out of space.
This commit also fixes leaking the callbacks_lp array.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
* atomic: add support for __atomic builtins
This commit adds support for the gcc __atomic builtins. The __sync
builtins are deprecated and have been replaced by these atomics. In
addition, the new atomics support atomic exchange which was not
supported by __sync.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
* atomic: add support for transactional memory
This commit adds support for using transactional memory when using
opal atomic locks. This feature is enabled if the __HLE__ feature is
available and the gcc builtin atomics are in use.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This fixes https://github.com/open-mpi/ompi/issues/1732: i.e., the
case where the outer project has its own check for
<valgrind/valgrind.h>, but also supplements CPPFLAGS (to find
Valgrind's header files) before doing that check.
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
Ideally, we would tell OMPI to disable autoconf's caching of our
valgrind check result so that its check gets the right result after
adding CPPFLAGS. Not sure if we can do that.
For now, just disable our Valgrind code in embedded mode.
This will keep the x86 backend enabled under Valgrind but
it will auto-disable itself when finding identical APIC ids anyway
(because CPUID returns same outputs for all PUs).
Signed-off-by: Brice Goglin <Brice.Goglin@inria.fr>
Fixesopen-mpi/ompi#1732
(cherry picked from commit open-mpi/hwloc@8b44fb1c81)
This commit fixes a programming error when using an aries nic. The
documentation of ugni shows that only the local alignment restriction
for get was lifted on aries. There is still a remote address alignment
restriction.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit fixes a programming error when using an aries nic. The
documentation of ugni shows that only the local alignment restriction
for get was lifted on aries. There is still a remote address alignment
restriction.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit adds support for Cray Aries atomic operations. This
includes 32-bit and floating point support.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit add support for more atomic operations and type. The
operations added are logical and, logical or, logical xor, swap, min,
and max. New types are 32-bit int by using the
MCA_BTL_ATOMIC_FLAG_32BIT flag, 64-bit float by using the
MCA_BTL_ATOMIC_FLAG_FLOAT flag, and 32-bit float by using both
flags. Floating point numbers are supported by packing the number in
as an int64_t or int32_t. We will update the btl interface in the
future to make this less confusing.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
Before dynamic add_procs support was committed to master we called
add_procs with every proc in the job. The XRC code in the openib btl
was taking advantage of this and setting the number of work queue
entries (WQE) based on all the procs on a remote node. Since that is
no longer the case we can not simply increment the sd_wqe field on the
queue pair. To fix the issue a new field has been added to the xrc
queue pair structure to keep track of how many wqes there are total on
the queue pair. If a new endpoint is added that increases the number
of wqes and the xrc queue pair is already connected the code will
attempt to modify the number of wqes on the queue pair. A failure is
ignored because all that will happen is the number of active send work
requests on an XRC queue pair will be more limited.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
