EFA incorrectly implements FI_DELIVERY_COMPLETE in earlier libfabric
versions. While FI_DELIVERY_COMPLETE would be advertised by the
provider, completions would return too early by not accounting for
bounce buffers on the receive side. This would cause the BTL
to receive early completions that lead to correctness issues.
This is not an issue in the mtl/ofi as it does not require
FI_DELIVERY_COMPLETE.
Signed-off-by: William Zhang <wilzhang@amazon.com>
There are 2 reasons for this:
- pending CUDA events are not progressed by this BTL, so anything that becomes
asychronous will never be completed.
- we use the packed data on the shared memory backing file, and this will be
returned to the peer process upon return (thus if we copy asynchronously we
might not copy the right data).
Signed-off-by: George Bosilca <bosilca@icl.utk.edu>
We do not want to be patching upstream components anymore.
The proper method is to get this merged upstream, then
pull it in the next upstream release.
This reverts commit c39fb5758a772c062e20db9b42f2b06805884802.
Signed-off-by: Austen Lauria <awlauria@us.ibm.com>
It has been broken for months because of the lack of initialization of the
HWLOC library. The smcuda process creating the backing file (local rank 0)
uses opal_cache_line_size to align the objects in the backing file, and the
opal_cache_line_size is initialized by default to 128. Later on, when the rest
of the processes attach the same backing file, HWLOC has been called and the
cache size has now been updated to the correct value. If this value is
different than the default one (and they are as most cache sizes are 64 bytes
right now) the objects in the backing file will be misaligned.
Signed-off-by: George Bosilca <bosilca@icl.utk.edu>
This commit updates the btl interface to change the parameters
passed to receive callbacks. The interface used to pass the tag,
a btl base descriptor, and the callback context. Most of the
values in the btl base descriptor were unused and only helped
simplify the callbacks from the self btl. All of the arguments
have now been replaced with a single receive callback descriptor.
This descriptor contains the incoming endpoint, data segment(s),
tag, and callback context. All btls have been updated to use
the new callback and the btl interface version has been bumped
to v3.2.0.
As part of this change the descriptor argument (and the segments
contained within it) have been marked as const. The were treated
as const before but this change could allow the compiler to make
better optimization decisions and will enforce that the callback
does not attempt to change the data in the descriptor.
Signed-off-by: Nathan Hjelm <hjelmn@google.com>
NULL pointer arithmetic is undefined behaviour in c.
The payload_ptr can be NULL in the moment when mpool is not initialized.
References from the c11 standard:
- 6.5.6 Additive operators
- 6.3.2.3 Pointers
Signed-off-by: Christoph Niethammer <niethammer@hlrs.de>
bugfix: provider selection would not differentiate between ipv4
and ipv6 addresses which would cause some nodes to be unable
to communicate between each other. Adding a check for address
format to provider selection to ensure that all nodes use the
same address format.
Signed-off-by: Nikola Dancejic <dancejic@amazon.com>
- there is new API to detect missing memmory events.
Enabled using of new UCX API to detect missing events
Signed-off-by: Sergey Oblomov <sergeyo@mellanox.com>
This bug was first seen in a different product that's using the same
interception code as OMPI. But I think it's potentially in OMPI too.
In my vanilla build of OMPI master on RH8 if I "gdb libopen-pal.so" and
"disassemble intercept_brk", I'm seeing a suspicious extra instruction
in front of PATCHER_BEGIN:
0x00000000000d6778 <+40>: std r2,24(r1) // something gcc put in front
0x00000000000d677c <+44>: std r2,96(r1) // PATCHER_BEGIN's toc_save
0x00000000000d6780 <+48>: nop // NOPs from PATCHER_BEGIN
0x00000000000d6784 <+52>: nop // that get replaced
0x00000000000d6788 <+56>: nop // by instructions that
0x00000000000d678c <+60>: nop // change r2
0x00000000000d6790 <+64>: nop //
Later there are loads from that location like
0x000000000019e0e4 <+132>: ld r2,24(r1)
that make me nervous since that's the pre-updated value.
I believe this is the same thing Nathan is describing way back in a9bc692d
and his solution was to put a second call around each interception, where
the outer call is just
intercept_brk():
PATCHER_BEGIN
_intercept_brk()
PATCHER_END
and the inner call _intercept_brk() is where the bulk of the code goes.
What I'm seeing is that _intercept_brk() is being inlined and probably
negating Nathan's fix. So I want to add __opal_attribute_noinline__ to
restore the fix.
With this commit in place, the disassembly of intercept_brk becomes tiny
because it's no longer inlining _intercept_brk() and the susipicious
early save of r2 is gone. I made the same fix to all the intercept_*
functions, although intercept_brk was the only one that had a suspicious
save of r2.
As far as empirical failures though, we only have those from the non-OMPI
product that's using the same patcher code. I'm not actually getting OMPI
to fail from the above suspicious data being saved in r1+24.
Signed-off-by: Mark Allen <markalle@us.ibm.com>
