| Commit message (Collapse) | Author | Age |
... | |
| |
|
| |
|
| |
|
| |
|
| |
|
|
|
|
|
| |
Reviewed-by: Henry Estela <henry.r.estela@intel.com>
Signed-off-by: Michal Szarmach <michal.szarmach@intel.com>
|
|\
| |
| | |
A couple of GCC fixes
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
In file included from /«BUILDDIR»/psm/include/ipath_user.h:63:0,
from /«BUILDDIR»/psm/psm_user.h:45,
from am_reqrep_shmem.c:40:
am_reqrep_shmem.c: In function 'psmi_shm_attach':
/«BUILDDIR»/psm/include/ipath_udebug.h:79:28: error: format '%ld' expects argument of type 'long int', but argument 6 has type 'size_t {aka unsigned int}' [-Werror=format=]
fprintf(__ipath_dbgout, "%s%s: " fmt, __ipath_mylabel, __func__, \
^
/«BUILDDIR»/psm/include/ipath_udebug.h:95:31: note: in expansion of macro '__IPATH_DBG_WHICH'
#define _IPATH_PRDBG(fmt,...) __IPATH_DBG_WHICH(__IPATH_PROCDBG,fmt,##__VA_ARGS__)
^
am_reqrep_shmem.c:544:2: note: in expansion of macro '_IPATH_PRDBG'
_IPATH_PRDBG("Mapped and initialized shm object control page at %p,"
^
|
|/
|
|
|
|
|
|
|
|
| |
psm_diags.c: In function 'memcpy_check_size':
psm_diags.c:286:7: error: statement is indented as if it were guarded by... [-Werror=misleading-indentation]
if (dst) psmi_free(dst);
^~
psm_diags.c:284:5: note: ...this 'if' clause, but it is not
if (src == NULL || dst == NULL)
^~
|
|\
| |
| | |
Initialize variables
|
| | |
|
|/
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This patch fixes the following -Werror=maybe-uninitialized errors:
psm_diags.c:294:10: error: 'dst_p' may be used uninitialized in this function [-Werror=maybe-uninitialized]
if (dst_p) psmi_free(dst_p);
^
In file included from psm_user.h:54:0,
from psm_diags.c:35:
psm_utils.h:156:23: error: 'src_p' may be used uninitialized in this function [-Werror=maybe-uninitialized]
#define psmi_free(sz) psmi_free_internal(sz)
^
psm_diags.c:290:11: note: 'src_p' was declared here
void *src_p, *dst_p;
^
If posix_memalign fails to return a value different than zero,
the pointers are never initialized.
|
|
|
|
|
|
|
|
|
|
| |
During MPI_Finalize with Intel MPI, many processes would hang in
ips_proto_mq_isend.
This problem was only noticed at a larger amount of ranks and nodes.
This bug was introduced in 2b70a0f379c1a72b924bc4d5cf6ea19749c0bd57.
There were some bit shift and mask errors with the change.
The commidx base id was not properly used correctly in ips_epstate functions,
causing some messages to never get received properly.
|
|
|
|
|
|
|
|
| |
This reverts commit 554c2d6f297e3038339a5ee96384ef5f988e6a70.
That change introduced a hang when PSM_MULTIRAIL was set to 0:1,1:1,1:2.
The workaround for fatal temporary exhaustion of send tid dma
descriptors is to set PSM_MEMORY=large .
|
| |
|
|
|
|
|
|
|
| |
Since the default value is 1, we will just ignore the user input if the
user (or an MPI) tries to set PSM_SDMA to 1. This is a simple fix to
prevent PSM from using invalid internal values after getting a value of
1 from PSM_SDMA.
|
|
|
|
|
|
|
| |
Fixes issue reported in
https://www.open-mpi.org/community/lists/devel/2015/08/17838.php
The minor version is increased by 1 to allow for openmpi to check for
java support.
|
|
|
|
| |
Add compiled files and remove unused entries.
|
|
|
|
|
|
|
|
|
|
| |
PSM would allocate enough shared memory for 256 processes when compiled without
SCIF support. This patch instead uses the number of processors online or
the number of mpi processes running to determine how much memory to allocate.
However, if MPI_LOCALNRANKS or PSC_MPI_PPN are not set, then the
number of cores online will be used to set how much virtual memory
should be allocated. In that case the maximum number of jobs that can
run on a node will be limited to the number of cores online.
|
|\
| |
| | |
Prevent compiler warnings
|
| |
| |
| |
| | |
Happens when compiling with -Wall -Werror.
|
| | |
|
|\ \
| |/
|/| |
Fix executable stack reported by readelf -l.
|
|/
|
|
| |
Fixes 01org/psm#4
|
|
|
|
|
|
|
| |
The infinipath-psm rpm will always be built regardless of xeon phi support.
The xeon phi pacakge will provide intel-mic-psm instead of being a completely
different package. This allows for packages to require infinipath-psm and not
have to check for intel-mic-psm.
|
|
|
|
| |
Subject: [PATCH 3/3] Flush all the traffic before returning in PIO.
|
|
|
|
| |
Subject: [PATCH 2/3] Fix tid protocol sender side descriptor limitation.
|
|
|
|
|
|
|
|
|
| |
Subject: [PATCH 1/3] Add Support for more than 64K ranks,
commidx size increased from 16 to 20 bits to allow for 2^20 connection ids.
The base connection id was changed from a pseudo random number to 0.
This allows backwards compatibility with other PSM libraries, provided that
the job uses less than 64K ranks.
|
| |
|
|
|
|
|
|
|
| |
The rpm name and tarball name need to keep the
prior name.
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
|
| |
|
| |
|
| |
|
|
|
|
|
| |
The epoch was only need to transition to the new
version format.
|
| |
|
|
|
|
|
|
|
|
| |
These are compile issues associated with using the gcc 4.6.4 compiler:
Issues corrected:
- duplicate memory errors
- many used variable warnings
|
|
|
|
|
| |
Also adjust version/release for Makefile to use
git describe.
|
|\ |
|
| |
| |
| |
| | |
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
|
|/ |
|
| |
|
| |
|
| |
|
| |
|
|
|
|
| |
for bug 119761. The InfiniPath software minor version number is bumped to 12 to co-ordinate this change with the driver.
|
|
|
|
| |
message buffer. This is the case where the message fits into 2 PIO sends but is not a multiple of 4 in length. We split into two, roughly equal-sized, sends but ensure that the second one is a multiple of 4 in size so that we do not over-run in the dword copy routine. The first packet might not be a multiple of 4, but any over-run will just be into the second half of the message which is safe.
|
| |
|
|
|
|
| |
without LIDs) in calculation of number of contexts
|
|
|
|
| |
system call interface for bug 118844
|
|
|
|
| |
119298, but only for debug builds since this option can cause additional code to be generated
|