GitLab

Wesley Bland authored Jan 27, 2014 and Pavan Balaji committed Jan 27, 2014

Resets MPIDI_TAG_UB back to 0x7fffffff. This value was changed a while back,
but the change should have happened at the MPI layer instead of the CH3 layer.
This resets the value to allow CH3 to use the tag space.

Instead, the value is now set in the MPI layer during initthread. This means
that it will be safe regardless of the device being used. This prevents a
collision that was occurring on the pamid device where the values for
MPIR_TAG_ERROR_BIT and the MPIR_Process.attr.tagged_coll_mask values were the
same.

Fixes #2008
Signed-off-by: Pavan Balaji <balaji@mcs.anl.gov>

Dave Goodell authored Apr 04, 2013 and Rob Latham committed Oct 31, 2013

Also includes random fixes to `-Wshorten-64-to-32` warnings which might
need to be teased out.
Signed-off-by: Pavan Balaji <balaji@mcs.anl.gov>

Xin Zhao authored Jul 30, 2013 and Pavan Balaji committed Aug 01, 2013

Because CH3 layer needs to know if shared memory region is allocated in lower layer.
Signed-off-by: Pavan Balaji <balaji@mcs.anl.gov>

Xin Zhao authored Jul 22, 2013 and Pavan Balaji committed Jul 28, 2013

Add "alloc_shm" to window's info arguments and initialize it to FALSE.
In MPID_Win_allocate, if "alloc_shm" is set to true, call ALLOCATE_SHARED,
otherwise call ALLOCATE.

Free window memory only when SHM region is not allocated, therwise it is
already freed in MPIDI_CH3I_SHM_Win_free.
Signed-off-by: Pavan Balaji <balaji@mcs.anl.gov>

Wesley Bland authored Jun 06, 2013 and Pavan Balaji committed Jul 25, 2013

These macros are used to manipulate the error bit that has been added
to the tag space to propagate error notification on top of other
communication.
Signed-off-by: Pavan Balaji <balaji@mcs.anl.gov>

Wesley Bland authored May 31, 2013 and Pavan Balaji committed Jul 25, 2013

The highest bit of tag space will now be reserved for failure
notification. We are far above the required 32767, so this shouldn't be
a problem.
Signed-off-by: Pavan Balaji <balaji@mcs.anl.gov>

Myrank was caching win_ptr->comm_ptr->rank, so we now use that directly
rather than caching it in the MPID_Win object.

Reviewer: balaji

Several fields of MPID_Win were not used above the ADI.  These have been
moved to CH3 to reduce clutter.  Downstream devices should redefine
these fields in MPID_DEV_WIN_DECL if they are needed.

Reviewer: balaji

Message transfers now respect the communicator-specific threshold.  This
change has not been carefully checked for impact on our shared-memory
ping-pong latency.

Reviewed-by: goodell

The single_op_opt flag in the request object was previously used to
track whether an operation is a lock-op-unlock type, for the purposes of
completion.  Tracking this state has been merged into the packet header
flags, so the single_op_opt flag is no longer needed.

Reviewer: goodell

Added a flags field to MPID_Request that we can use to stash flags from
suspended RMA ops and retrieve them later when we complete the operation.

Reviewer: goodell

Prior to this patch, a lock entry was enqueued in the RMA ops list when
Win_lock was called.  This patch adds a new state tracking mechanism, which we
use to record the synchronization state with respect to each RMA target.  This
new mechanism absorbs tracking of lock operation and the lock state at the
target.  It significantly simplifies the RMA synchronization and ops list
processing.

Reviewer: goodell

Communication operations on shared memory windows now perform the op directly
on the shared buffer.  This requried the addition of a per-window interprocess
mutex to ensure that atomics and accumulates are performed atomically.

Reviewer: buntinas

[svn-r10770] removed channel_private field in VC and used MPIDI_CH3_VC_DECL macro which is overridden by channel. Reviewed by Dinan

This removes macro copy-paste clutter that was left around from when this state
moved from MPID_Win into a separate structure.

Reviewer: goodell

Added an additional error check and better documentation of how the state is
tracked and updated for Fence.

Reviewer: goodell

This refinement to the existing PSCW state tracking more cleanly captures the
PSCW state and further restricts erroneous usage.

Reviewer: goodell

This fixes the performance regression that was introduced by concatenation of
per-target lists.

Reviewer: goodell

Moved fence_issued and start_assert MPID_Win members into CH3.  These should
not be a required part of the ADI, since they are not needed above the ADI and
implementations should be free to choose different mechanisms for tracking
the state of synchronization operations.

Reviewer: buntinas

This patch adds code to track the RMA epoch state of the local process.
Currently, we are tracking the synchronization states that are allowed by
MPICH; in the future, we may want to restrict this to only states that are
allowed by the standard.  The addition of epoch tracking has several benefits:

 * It allows us to detect synchronization errors (implemented in this patch).
 * It allows us to implement lock_all more efficiently (implemented in this
   patch).
 * It will allow us to distinguish between active and passive target epochs and
   avoid O(p) op list concatenation (future patch).

Reviewer: balaji

Updated RMA code to remove trailing "_e" and "_s" on enum and struct type
names to match the MPICH style.

Reviewer: goodell

Reviewer: goodell

Updated RMA implementation to track the passive target status individually, for
each target.  Includes new implementation for lock/unlock_all.  Lock_all is
currently unoptimized, see #1734 for future plans.

Reviewer: buntinas

The use of a dense array is a temporary measure to support the reference
implementation.  This will be much improved by ticket #1735.

Reviewer: goodell

In this patch, I have refactored the RMA ops list again to use the MPL UTList
doubly-linked list and to treat the list as a proper object.  This should set
us up to work with multiple lists, as we will soon have one list per target.
Doubly-linking the list is a big help in terms of maintainability (no more
prevNext pointers) and flexibility (better implementation of request-based
ops and other optimizations).

Reviewer: goodell

Implementation of MPI-3 window info query routines.

Reviewer: balaji

No reviewer.

This commit implements MPI-3 RMA's flush and flush_all operations.

Reviewer: buntinas

When enabled, this mode of operation immediately requests the lock when
MPI_Win_lock is called.  Currently, this is enabled by setting the
MPICH_RMA_LOCK_IMMED environment variable.  In the future, we can also make
this mode of operation available though an info/assert.  This capability is
needed to implement MPI-3's flush operations.

Reviewer: buntinas

Reviewer: goodell

By setting "indent-tabs-mode:nil" we should hopefully begin to slowly
squeeze out hard tabs from the source without a disruptive (to
downstream projects) whitespace-fixing change.

No reviewer.

This adds the win_fns table to ch3, which allows the channel to override the
default implementation of window creation routines provided by ch3.  This also
pushes the implementation of shared memory windows down into Nemesis, includes
window functions for sock, and contains multiple improvements to the window
creation functions code.

Reviewer: buntinas

[svn-r10070] fixed up the collective function override mechanism and it's usage in ch3.  Fixes #1634.  Reviewed by goodell@

Nemesis.  In MPI-3, ch3 will require shared memory in order to support
MPI_Win_allocate_shared().

yet implemented below the ADI level.

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@.

Forum ticket #286) and updates MPI_Intercomm_create to use a collectives tag
space which does not conflict with point-to-point communication (MPI Forum
ticket #305).

In addition, this fixes a deadlock bug in context id allocation and adds
several regression tests (comm_dup_deadlock.c exercises this specific bug).

Reviewer: David Goodell