Permalink
Commits on Jan 18, 2017
  1. Remove inject API from the shmmod API.

    We had already removed this API from the netmod interface, but forgot
    to remove it from the shmmod interface.
    
    Signed-off-by: Ken Raffenetti <raffenet@mcs.anl.gov>
    pavanbalaji committed Jan 16, 2017
  2. Fix array initialization in structure.

    Structure initialization needs an additional level of braces.
    
    Signed-off-by: Ken Raffenetti <raffenet@mcs.anl.gov>
    pavanbalaji committed Jan 18, 2017
  3. warning squash.

    Remove unnecessary stack variable that needs to be explicitly
    initialized.
    
    Signed-off-by: Ken Raffenetti <raffenet@mcs.anl.gov>
    pavanbalaji committed Jan 16, 2017
  4. Update Charles' name to be consistent with other names.

    Signed-off-by: Ken Raffenetti <raffenet@mcs.anl.gov>
    pavanbalaji committed Jan 16, 2017
  5. MPID_Abort fixes.

    The previous fixes to MPID_Abort were causing it to fail to build on a
    mac.  This patch allows MPID_Abort to be defined in a C file, with a
    prototype in the overall header to allow all files to access it
    correctly.
    
    This fix is not yet complete.  The build still fails inexplicably if
    the MPID_Abort is moved to a different file.  But at least it gets the
    build going for now.
    
    Signed-off-by: Ken Raffenetti <raffenet@mcs.anl.gov>
    pavanbalaji committed Jan 18, 2017
  6. CH4/OFI: Fix typo in debug output

    Signed-off-by: Ken Raffenetti <raffenet@mcs.anl.gov>
    wesbland committed with raffenet Jan 13, 2017
  7. CH4/OFI: Remove fprintf in FI_EP_RDM check

    This fprintf could generate a lot of noise depending on the providers
    available in libfabric. Remove it and rely on the debug logging info
    instead.
    
    Signed-off-by: Wesley Bland <wesley.bland@intel.com>
    raffenet committed with raffenet Jan 18, 2017
  8. CH4/OFI: Fix the runtime check for FI_EP_RDM

    Somehow this check hasn't caused problems up to now. We should be
    checking that the provider supports FI_EP_RDM, not the opposite.
    
    Signed-off-by: Ken Raffenetti <raffenet@mcs.anl.gov>
    wesbland committed with raffenet Jan 13, 2017
Commits on Jan 17, 2017
  1. ROMIO: Abort the comm associated with a file on error

    Pass the comm associated with the MPI_File to MPIR_Abort in the event of
    fatal error. Previous versions of this code passed either NULL or
    MPI_COMM_WORLD, but the comm used to create the file is a more accurate
    scope for the error.
    
    Signed-off-by: Pavan Balaji <balaji@anl.gov>
    raffenet committed with raffenet Jan 17, 2017
  2. glue/romio: Expose abort functionality

    Mostly a correction of [c2e3d7e]. Expose a wrapper for MPID_Abort
    that ROMIO can safely call.
    
    Signed-off-by: Pavan Balaji <balaji@anl.gov>
    raffenet committed with raffenet Jan 16, 2017
  3. CH4: Remove extra definition of MPID_Abort

    Signed-off-by: Pavan Balaji <balaji@anl.gov>
    raffenet committed with raffenet Jan 16, 2017
Commits on Jan 13, 2017
  1. CH4: Fix unused label warning in mpidig_init.c

    Squash warning about unused fn_fail label.
    
    Signed-off-by: Ken Raffenetti <raffenet@mcs.anl.gov>
    wesbland committed with raffenet Jan 13, 2017
  2. CH4/OFI: Fix fn_fail warning in imrecv

    Remove an unused label to squash a warning.
    
    Signed-off-by: Ken Raffenetti <raffenet@mcs.anl.gov>
    wesbland committed with raffenet Jan 13, 2017
Commits on Jan 12, 2017
  1. mpl: Add Utility Thread Offloading Interface

    You can add the capability of Utility Thread Offloading to pthread by
    specifying '--with-thread-package=uti'. The locations of the header
    file and library are specified by either '--with-uti=yes', which means
    the standard directories, or '--with-uti=PATH', which means
    PATH/{lib,include}.
    
    Signed-off-by: Ken Raffenetti <raffenet@mcs.anl.gov>
    masamichi-takagi committed with raffenet Nov 1, 2016
Commits on Jan 11, 2017
  1. CH4: Remove namespace shift for ADI functions

    Signed-off-by: Wesley Bland <wesley.bland@intel.com>
    raffenet committed with raffenet Jan 4, 2017
  2. ROMIO: Remove call to MPID_Abort

    Remove the only instance of ROMIO calling ADI functionality. Even in
    MPICH-specific glue code, this can cause build problems. E.g. in CH4
    where ADI functions are defined as static inlines, the compiler will
    complain about the non-static definition in mpir_ext.h. Instead, we
    print an error message and call MPI_Abort.
    
    Signed-off-by: Wesley Bland <wesley.bland@intel.com>
    raffenet committed with raffenet Jan 5, 2017
  3. handlemem: Move finalize handle checks out of header

    Avoid some header dependency issues caused by the call to MPID_Abort at
    the end of MPIR_check_handles_on_finalize. MPID_Abort is defined in
    mpiimpl.h, but after mpir_handlemem.h is included. Since this is debug
    related code, any performance loss from losing the static inline should
    be okay.
    
    Signed-off-by: Wesley Bland <wesley.bland@intel.com>
    raffenet committed with raffenet Jan 5, 2017
Commits on Jan 10, 2017
  1. Fix -Wmisleading-indentation in a few places

    Signed-off-by: Ken Raffenetti <raffenet@mcs.anl.gov>
    roblatham00 committed with raffenet Jan 10, 2017
Commits on Jan 9, 2017
  1. CH4/OFI: Add capability for fetch_atomic iovec limit

    For some providers, the limit for fetch_atomic operations might be
    different than for other RMA operations. Add a capability to allow the
    user/provider developer to set this value at runtime/compile-time,
    respectively.
    
    Fixes csr/mpich-ofi#442
    
    Signed-off-by: Ken Raffenetti <raffenet@mcs.anl.gov>
    wesbland committed with raffenet Dec 20, 2016
Commits on Jan 6, 2017
  1. CH4/UCX: Set source and tag status values in truncation case

    The MPI standard encourages setting the source and tag information in
    status for truncated messages. UCX itself does not guarantee the
    content of its completion info in the case of truncation errors, but
    passing it along can do no harm. In the event they are accurate, they
    can be of use. See tests in test/mpi/errors/pt2pt.
    
    No reviewer.
    raffenet committed with raffenet Jan 6, 2017
Commits on Jan 5, 2017
  1. ROMIO: Fix -Wmisleading-indentation in ad_read_coll.c

    (also see open-mpi/ompi@33d2988)
    
    Signed-off-by: Ken Raffenetti <raffenet@mcs.anl.gov>
    OMGtechy committed with raffenet Jan 5, 2017
  2. CH4/OFI: Add capability for fetch_atomic iovec limit

    For some providers, the limit for fetch_atomic operations might be
    different than for other RMA operations. Add a capability to allow the
    user/provider developer to set this value at runtime/compile-time,
    respectively.
    
    Fixes csr/mpich-ofi#442
    
    Signed-off-by: Ken Raffenetti <raffenet@mcs.anl.gov>
    wesbland committed with raffenet Dec 20, 2016
  3. CH4: Add debug output for capability sets

    Add the environment variable MPIR_CVAR_CH4_OFI_CAPABILITY_SETS_DEBUG to
    turn on debug output for capability sets configuration.
    
    Signed-off-by: Ken Raffenetti <raffenet@mcs.anl.gov>
    wesbland committed with raffenet Dec 19, 2016
  4. CH4: Add Capability Sets README documentation

    Signed-off-by: Ken Raffenetti <raffenet@mcs.anl.gov>
    wesbland committed with raffenet Dec 19, 2016
  5. CH4: Remove unnecessary temporary variable

    Signed-off-by: Wesley Bland <wesley.bland@intel.com>
    raffenet committed with raffenet Jan 4, 2017
  6. CH4: Fixup mrecv request completion

    1. MPIR_Request_complete already extracts the status for us, so we can
       remove the extra call.
    2. Capture and return the error code from MPIR_Request_complete to the
       upper layer.
    
    Signed-off-by: Wesley Bland <wesley.bland@intel.com>
    raffenet committed with raffenet Jan 4, 2017
  7. CH4: Remove extra container_of definition

    Signed-off-by: Wesley Bland <wesley.bland@intel.com>
    raffenet committed with raffenet Jan 4, 2017
Commits on Jan 3, 2017
  1. config: Move local debug options to top level

    Debug options used to control which processes communicate over shared
    memory are no longer specific to nemesis. Move the options to the top
    level configure while maintaining aliases for the old names.
    
    Signed-off-by: Yanfei Guo <yguo@anl.gov>
    raffenet committed with yfguo Dec 13, 2016
Commits on Dec 30, 2016
  1. CH4/OFI: Remove unneeded check in imrecv

    A NULL message pointer is handled by the CH4 layer.
    
    Signed-off-by: Lena Oden <loden@anl.gov>
    raffenet committed with raffenet Dec 19, 2016
  2. CH4: Remove MPI_PROC_NULL check in mrecv/imrecv

    This case cannot happen. We already check for MPI_PROC_NULL during
    mprobe/improbe and handle it accordingly.
    
    Signed-off-by: Lena Oden <loden@anl.gov>
    raffenet committed with raffenet Dec 19, 2016
  3. CH4/UCX: Fix improbe/imrecv

    Fix remaining issues with the improbe/imrecv functionality in the UCX
    netmod after refactoring.
    
     1. Eliminate extra request. Convert already allocated MPROBE request
        to RECV request and give back to upper layer instead.
     2. Implement new mrecv completion callback. Because we already have
        our request allocated, we need different logic.
    
    Signed-off-by: Lena Oden <loden@anl.gov>
    raffenet committed with raffenet Dec 16, 2016
  4. CH4: Make progress in nonblocking probe

    The MPI standard requires nonblocking probe functions make
    progress. If no message is found, call progress test before returning
    to the user, similar to CH3.
    
    Signed-off-by: Lena Oden <loden@anl.gov>
    raffenet committed with raffenet Dec 16, 2016
Commits on Dec 21, 2016
  1. CH4/UCX: Check for truncated messages

    1. Check that all non-contig data is unpacked in generic datatype
       callbacks, and report truncation if not.
    2. In recv completion callback, check if UCX reports truncation and set
       error in status if so.
    
    Signed-off-by: Ken Raffenetti <raffenet@mcs.anl.gov>
    LenaO committed with raffenet Dec 15, 2016
  2. mpid/common/datatype: Call device hooks for pairtypes

    These non-contig types are never committed by the user, so we need to
    call the device hooks manually.
    
    Signed-off-by: Ken Raffenetti <raffenet@mcs.anl.gov>
    LenaO committed with raffenet Dec 13, 2016