This commit is related to an RFC from June 2014. Disscussion can be
found at:
http://www.open-mpi.org/community/lists/devel/2014/07/15140.php
The finalize function is set using either the linker option -fini or
__attribute__((destructor)) depending on compiler support. I have
confirmed that this hybrid approach works with all the major
compilers. The attribute is supported by gcc, clang, llvm, xlc, and
icc. The fini function will support pgi. If a compiler/linker
combination does not support either the destructor or fini function a
message will be printed on re-init indicating it is not supported (an
improvement over the current behavior-- SEGV).
I moved the following to the destructor function:
- Class system finalize. This solves a bug when MPI_T_finalize is
called before MPI_Init. The only downside to this change is we will
leave the footprint of the opal class system after
MPI_Finalize. This footprint should be relatively small.
This is an alternative to #517 but the two PRs are not
mutually-exclusive (with some modifications). This commit should also
be safe for 1.8.x as it does not change internal or external ABI (#517
changes internal ABI).
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit is a rework of the component repository. The changes
included in this commit are:
- Remove the component dependency code based off .ompi_info
files. This code is legacy code dating back 10 years that and is no
longer used.
- Move the plugin scanning code to the component repository. New
calls have been added to add new scanning paths, query available
components, and dlopen/load components.
- Pass the framework down to mca_base_component_find/filter. Eventually
the framework structure will be used to further validate components
before they are used.
- Add support to the MCA framework system to disable scanning for
dlopened components on open (support already existed in
register). This is really only relevant to installdirs as it has no
register function and no DSO components.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit fixes a typo in mca_btl_vader_progress_endpoints where
OPAL_THREAD_LOCK was used when OPAL_THREAD_UNLOCK was intended.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit fixes several vagrind errors. Included:
- installdirs did not correctly reinitialize all pointers to NULL
at close. This causes valgrind errors on a subsequent call to
opal_init_tool.
- several opal strings were leaked by opal_deregister_params which
was setting them to NULL instead of letting them be freed by the
MCA variable system.
- move opal_net_init to AFTER the variable system is initialized and
opal's MCA variables have been registered. opal_net_init uses a
variable registered by opal_register_params!
- do not leak ompi_mpi_main_thread when it is allocated by
MPI_T_init_thread.
- do not overwrite ompi_mpi_main_thread if it is already set (by
MPI_T_init_thread).
- mca_base_var: read_files was overwritting mca_base_var_file_list
even if it was non-NULL.
- mca_base_var: set all file global variables to initial states on
finalize.
- btl/vader: decrement enumerator reference count to ensure that it
is freed.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit fixes a bug when opal_error_init is called with the same
values multiple times. If opal_error_init is called too many times it
will start failing with OPAL_ERR_OUT_OF_RESOURCE. To fix the problem
check if an existing convertor matching the requested one and return
that one instead.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit fixes the following bugs:
- opal_output_finalize did not properly set internal state. This
caused problems when calling the sequence opal_output_init (),
opal_output_finalize (), opal_output_init ().
- opal_info support called mca_base_open () but never called the
matching mca_base_close (). mca_base_open () and mca_base_close ()
have been updated to use a open count instead of an open flag to
allow mca_base_open to be called through multiple paths (as may be
the case when MPI_T is in use).
- orte_info support did not register opal variables. This can cause
orte-info to not return opal variables.
- opal_info, orte_info, and ompi_info support have been updated to
use a register count.
- When opening the dl framework the reference count was added to
ensure the framework stuck around. The framework being closed
prematurely was a bug in the MCA base that has since been
corrected. The increment (and associated decrement) have been
removed.
- dl/dlopen did not set the value of
mca_dl_dlopen_component.filename_suffixes_mca_storage on each call
to register. Instead the value was set in the component
structure. This caused the value to be lost when re-loading the
component. Fixed by setting the default value in register.
- Reset shmem framework state on close to avoid returning a stale
component after reloading opal/shmem.
- MCA base parameters were not properly deregistered when the MCA
base was closed.
This commit may fix#374.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
linux: only use the device-tree on Power machines
It's available on ARM but the assumption that cpus' "reg" start at 0
is invalid.
We could make that work but the device-tree doesn't currently
bring anything better than sysfs on ARM, so don't bother for now.
On 32-bit architectures loads/stores of fast box headers may take
multiple instructions. This can lead to a data race between the
sender/receiver when reading/writing the sequence number. This can
lead to a situation where the receiver could process incomplete
data. To fix the issue this commit re-orders the fast box header to
put the sequence number and the tag in the same 32-bits to ensure they
are always loaded/stored together.
Fixes#473
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
fi_av_insert() is invoked with a context containing each endpoint
USNIC_NUM_CHANNELS times. If the address on that endpoint fails to
resolve / is unreachable / has some error, we'll therefore get
USNIC_NUM_CHANNELS error completions with that same endpoint. We
therefore only want to warn about the unreachability of (and
OBJ_RELEASE) that endpoint the *first* time.
Fixes CSCut46822.
This commit adds support for project_framework_component_* parameter
matching. This is the first step in allowing the same framework name
in multiple projects. This change also bumps the MCA component version
to 2.1.0.
All master frameworks have been updated to use the new component
versioning macro. An mca.h has been added to each project to add a
project specific versioning macro of the form
PROJECT_MCA_VERSION_2_1_0.
Signed-off-by: Nathan Hjelm <hjelmn@me.com>
It was setup in the PCI backend before filtering,
and partially updated after filtering in the core.
Only setup once correctly after filtering in the core.
(cherry picked from commit open-mpi/hwloc@9659653d24)
Conflicts:
tests/hwloc/linux/40intel64-2g2n4c+pci.output
tests/hwloc/xml/192em64t-12gr2n8c2t-distancegroups.xml
tests/hwloc/xml/192em64t-24n8c2t-distancegroups.xml
tests/hwloc/xml/192em64t-24n8c2t-nodistancegroups.xml
tests/hwloc/xml/24em64t-2n6c2t-pci.xml
tests/hwloc/xml/32em64t-2n8c2t-pci-normalio.xml
tests/hwloc/xml/96em64t-4n4d3ca2co-pci.xml
utils/hwloc/test-hwloc-compress-dir.input.tar.gz
utils/hwloc/test-hwloc-compress-dir.output.tar.gz
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
If super_set contains more allocated ulongs than sub_set,
we did not check the last ulongs.
We would return true instead of false when sub_set is
infinite while the last ulongs in super_set are not full.
This fixes tests/hwloc_bitmap_compare_inclusion on some platforms.
(cherry picked from commit open-mpi/hwloc@299e6e846f)
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
Otherwise we get spurious bits for crazy topologies such as 8em64t-2s2ca2c-buggynuma.output
Will make debug asserts easier.
(cherry picked from commit open-mpi/hwloc@546cd9330a)
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
Make sure we define complete cpuset/nodeset when we define groups' main cpuset/nodeset
during later insert of groups (for PCI hostbridges or distances).
Otherwise they may end up clearing child/parent complete sets which
suddenly become incoherent while they were fixed earlier.
Needed to fix allowed_nodeset meaning.
(cherry picked from commit open-mpi/hwloc@7c88d17add)
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
When looking for PUs inside R_MAXSDL rads, some AIX 6.1 releases
return one first rad without any PU.
AIX 6.1 00F63F144C00 does (on quad-power7).
AIX 6.1 00CBAAC24C00 doesn't (on 16x power6).
So we can't assume rad #x contains PU #x. But we already have the right
code to fill the cpuset from the rad, so use that to obtain the PU os_index
as well.
Cannot be used to obtain NUMA node os_index since there's no way to directly
retrieve NUMA nodes from rads (mempools seem unrelated). Just keep using #rad
for NUMA nodes os_index and document that convention when converting back in
set_membind().
Thanks to Hendryk Bockelmann and Erik Schnetter for helping debugging.
(cherry picked from commit open-mpi/hwloc@60006c7b88)
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
Otherwise we'll have some NULL objects above, would be annoying.
No need to dig further, the distance matrix is likely buggy.
We still keep the inserted groups at this level (incomplete level)
because removing them is hard.
(cherry picked from commit open-mpi/hwloc@312a971ec9)
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
Commit 626129d2818693e62b83c1cfa2ba6e058e5bed66 fixed the hwloc
device/vendor numbers obtained from libpciaccess.
But the corresponding names are still retrieved from pciaccess numbers,
so fix these numbers inside pciaccess structures before retrieving the names.
(cherry picked from commit open-mpi/hwloc@85ea6e4acc)
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
Misc objects were used between system and machine in the past
but quickly got replaced with groups.
(cherry picked from commit open-mpi/hwloc@6c2aa6d1ea)
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
Misc are reserved for annotating the topology, the core
doesn't like merging them. Group is more appropriate.
(cherry picked from commit open-mpi/hwloc@3c47649591)
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
hwloc_get_first_largest_obj_inside_cpuset() returns the largest/highest object,
but it could still have a child with the same cpuset.
So check children as well in case there's a matching NUMA node there.
(cherry picked from commit open-mpi/hwloc@57a1c4fbe4)
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
When ignore_keep_structure is enabled, intermediate level can disappear
between parent and child, making the new child complete_cpuset smaller,
causing the child list to require a reorder just like in remove_ignored().
(cherry picked from commit open-mpi/hwloc@88afbe6b62)
Embed this related commit:
core: abstract out reorder_children(), needed when merging modifies the list of children
(cherry picked from commit open-mpi/hwloc@14db82d391)
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
If object A contains B + I/O as children, we can "ignore" I/Os and still
try to merge A and B. We now do the same for Misc objects without cpusets
instead of I/Os.
This fixes a corner case when export/reimport to XML creates a slightly
different topology (making hwloc_insert_misc fail inside a Linux cgroup).
Thanks to Dave Love for reporting the problem.
Fixes#118
(cherry picked from commit open-mpi/hwloc@650371e115)
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
When I/O are attached under a PU, removing the children's cpusets from
the parent cpuset doesn't give 0, it gives the PU cpuset.
The assertion fails on single-pu machines with I/O when --merge is given,
only one PU remains with I/O under it.
But if we insert Misc by cpuset under PU, it gives 0 as expected.
Fix the assertion accordingly.
Thanks to Thomas Van Doren for reporting the issue.
(cherry picked from commit open-mpi/hwloc@45c94c336d)
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
hwloc_x86_discover() calls hwloc_look_x86() twice, which calls hwloc_have_x86_cpuid().
If everything gets inlined, the asm label inside hwloc_have_x86_cpuid()
is duplicated.
Use a local label with f annotation in jumps to avoid the problem.
Thanks to Thomas Van Doren for reporting the issue (found with gcc -m32).
(cherry picked from commit open-mpi/hwloc@50e447f5bc)
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
x86: Not critical since BSDs that use this backend have no membind support,
but better fix it for uniformization.
(cherry picked from commit open-mpi/hwloc@a431361c7d)
OSF: Looks like nobody ever tried to play with memory binding on OSF/Tru64.
(cherry picked from commit open-mpi/hwloc@2d6c73356d)
Conflicts:
NEWS
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
At least some solaris enforce the need to #include X11/Xlib.h first.
Thanks to Siegmar Gross for reporting the issue.
(cherry picked from commit open-mpi/hwloc@005a7e89b6)
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
tolower needs <ctype.h>
Thanks to Ralph Castain for reporting the failure.
(cherry picked from commit open-mpi/hwloc@038c372a58)
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
strncasecmp() needs <strings.h>
Thanks to Pavan Balaji for reporting the failure.
(cherry picked from commit open-mpi/hwloc@37439c4801)
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
operands
Also added support for the xchg instruction. The instruction is
supported by ia32 and may benefit vader.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
Don't filter the topology by cpuset if you are mpirun until you know that no other compute nodes are involved. This deals with the corner case where mpirun is executing on a node of different topology from the compute nodes.
Simplify - don't mandate that all cpus in the given cpuset be present on every node. We can then run everything thru the filter as before, which ensures that any procs run on mpirun are also contained within the specified cpuset.
Correctly count the number of available PUs under each object when given a cpuset
Fix the default binding settings, and correctly count PUs when no cpuset is given
Ensure the binding policy gets set in all cases
When we use LIBADD for static libraries, the dependent libraries get
propagated properly. For example, the dl/dlopen component will almost
certainly require the -ldl library; when using LIBS, that doesn't get
propagated elsewhere in the tree, but when using LIBADD, it does
(e.g., when linking opal_wrapper_compiler).
If we really get a catastrophic error from a libfabric call, don't
bother trying to continue (because data has been corrupted and there's
nothing sane left to do). Just call opal_btl_usnic_exit() (which
tries to call the PML error callback, but we're so early in the
module_init process that this likely hasn't been setup yet, so the job
will likely abort).
Nothing too substantial here, but two of the messages moved from
"libfabric API failed" to "internal error during init", just to be a
bit more descriptive.
When we get errors, the entry.data field tells us how many errors are
being reported. So decrement the loop count variable by that much.
This fixes CSCut30441.
Enabling the FT code breaks compilation (again). This series
tries to fix the compiler errors. This is again only fixing
the compiler errors without any warranty that the result
might actually support FT again.
This first patch moves orte_cr_continue_like_restart from ORTE
to opal_cr_continue_like_restart in OPAL. This only leaves three
calls from OPAL to ORTE in the FT code. As it is not yet 100%
clear how to handle these calls the code orte_sstore.set_attr()
has been #ifdef'd out for now.
The derived segment type (btl_openib_segment_t) was intended to store
the registration info needed for put and get. In BTL 3.0 this is no
longer required. I intended to remove this type as part of
open-mpi/ompi@74f1af4548 .
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit should resolve an issue seen with CUDA-aware support. The
problem came in with BTL 3.0. Before 3.0 the size of the copy was
stored in the incoming segment's des_remote_count field. This field
does not exist in BTL 3.0 so I stored the value in the
des_segment_count field. This caused problems with the cuda support
code. To fix the issue the endpoint pointer is now stored in the in
fragment's endpoint pointer which free's up the segment's des_cbdata
pointer for storing the transfer size.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
If we ended up with no modules (e.g., all usnic devices were
excluded), there was a race condition in that the connectivity agent
could tear down its local socket before one or more of the local
clients saw it. Therefore, the local clients would timeout waiting
for the socket to appear.
So move the connectivity checker init later in the bootstrapping
process (it *must* be setup before module_init()), and have it only
invoked if we actually ended up with one or more modules.
Re-structure the loop looking for duplicates a little so that we only
have a single free of the string that happens regardless of whether we
found a duplicate or not.
This was Coverity CID 1288090
Fix previously-unfinished error paths during startup/bootstrapping.
Instead of just blindly continuing on when an fi_* function call
fails, opal_show_help and skip that device.
Also, only check the usnic config minimums once. They're VIC-wide and
won't change on a per-device basis -- we only need to check them once.
Fixes CSCut19179.
There were a number of bugs in the framework/variable code that
affected deregistration:
- Frameworks could be erroneously closed if seperately registered and
opened then subsequently closed. This was a bug in the original
design which only reference counted opens but not
registrations. This would cause undefined behavior if
MPI_T_finalize actually calls ompi_info_close_components as
intended. Now both registrations and opens are reference counted
and frameworks/components are not torn down until the matching
number of close calls have been made.
- group_find_by_name did not pass the invalidok flags down
to mca_base_var_group_get_internal correctly.
- Group deregistration caused the group to be completely reset. This
does not match the behavior required by MPI_T as it could reduce
the number of variables/subgroups in a group.
This commit also updates MPI_T_finalize to call
ompi_info_close_components as originally intended.
Closes#374
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
Embedding libltdl without the use of Libtool bootstrapping has
proven... difficult. Instead, create a new simple "dl" framework. It
only provides 4 functions:
- open a DSO (very similar to lt_dlopenadvise())
- lookup a symbol in a previously-opened DSO (very similar to lt_dlsym())
- close a previously-opened DSO (very similar to lt_dlclose())
- iterate over all files in a directory (very similar to ld_dlforeachfile())
There will be follow-on commits with a simple dlopen-based component
(nowhere near as complete/functional as libltdl, but good enough for
Linux and OS X), and a libltdl-based component for all other
platforms.
The intent is that the dlopen-based component can be built by default
in almost all cases. But if libltdl is available, that component will
be built. End result: we still get DSO-based functionality by default
in (almost?) all cases. Without embedding libltdl. Which is what we
want.
Use of this configuration option can cause crashing, hanging, and
(worse) incorrect results when btl/sm, btl/scif, or btl/vader are
in use. We discussed this at the January 2015 developers meeting
and it was decided to remove the option entirely. This commit does
just that. All usage of OPAL_WANT_SMP_LOCKS has been removed.
call the opal_common_verbs_mca_register function to make sure that
opal_common_verbs_want_fork_support mca parameter is created and therefore
can be used to control the fork support.
Both opal_shmem_base_select() and
opal_shmem_base_best_runnable_component_name() and were calling
opal_shmem_base_runtime_query(), which would do component selection
(and closing of losing components) twice.
Put protection in opal_shmem_base_runtime_query() to return the cached
results the second time. Additionally, make
opal_shmem_base_runtime_query() "own" the cached results
(vs. opal_shmem_base_select).
This allows the opal_shmem_base_module_t to be properly cast to an
mca_base_module_t.
(this commit is the rationale for the previous shmem C99 .member
initialization commit)
Also include two other minor changes:
1. More C99-style member initialization in the component struct
1. Fix the BTL module member initialization to not be redundant
Due to the nature of the cache architecture on power,
we don't export coherency_line_size for L2 in sysfs.
If we are unable to get the L2 cache line size, try L1.
See open-mpi/ompi#383 for more information.
In order to have an effect, ibv_fork_init should be called in the
beginning of the verbs initialization flow - before the calls to the
ibv_create_qp and ibv_create_cq verbs.
These functions are called from the oob/ud code and by the time the
other verbs components (btl openib, pml yalla, ...) call ibv_fork_init,
it's too late. This commit forces the call to ibv_fork_init (if it's
requested) right at the beginning of all the components that are using
verbs.
(ibv_fork_init() can be safely called multiple times)
This commit also removes the btl_openib_want_fork_support mca parameter
and adds a new mca parameter instead - opal_verbs_want_fork_support.
Through this new parameter, fork support may be requested for ALL
components.
The default value for this parameter is set to 1.
Before this commit the btl_openib_want_fork_support parameter didn't
provide fork support for the openib btl if its value was set to 1.
(because when openib called ibv_fork_init, it was already after the
calls to ibv_create_* in oob/ud and thereofre it failed).
Use of the old ompi_free_list_t and ompi_free_list_item_t is
deprecated. These classes will be removed in a future commit.
This commit updates the entire code base to use opal_free_list_t and
opal_free_list_item_t.
Notes:
OMPI_FREE_LIST_*_MT -> opal_free_list_* (uses opal_using_threads ())
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
Please verify your components have been updated correctly. Keep in
mind that in terms of threading:
OPAL_FREE_LIST_GET -> opal_free_list_get_st
OPAL_FREE_LIST_RETURN -> opal_free_list_return_st
I used the opal_using_threads() variant anytime it appeared multiple
threads could be operating on the free list. If this is not the case
update to _st. If multiple threads are always in use change to _mt.
Historically these two lists were different due to ompi_free_list_t
dependencies in ompi (mpool). Those dependencies have since been moved
to opal so it is safe to (finally) combine them. The combined free
list comes in three flavors:
- Single-threaded. Only to be used when it is guaranteed that no
concurrent access will be made to the free list. Single-threaded
functions are suffixed with _st.
- Mutli-threaded. To be used when the free list may be accessed by
multiple threads despite the setting of opal_using_threads.
Multi-threaded functins are suffixed with _mt.
- Conditionally multi-threaded. Common use case. These functions are
thread-safe if opal_using_threads is set to true.
Compatibility functions for the ompi_free_list_t and the old accessor
functions (OPAL_FREE_LIST_*) are available while the code base is
transitioned to the new class/functions.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit adds an owner file in each of the component directories
for each framework. This allows for a simple script to parse
the contents of the files and generate, among other things, tables
to be used on the project's wiki page. Currently there are two
"fields" in the file, an owner and a status. A tool to parse
the files and generate tables for the wiki page will be added
in a subsequent commit.
Some BTLs do not require local registration for some rdma
transactions. For example: inline put on openib, fma put on ugni. This
commit adds code to expose the local registration thresholds to BTL
users. Optimized code can take advantage of this information to
improve rdma performance.
