On Jul 21, 2015, at 12:22 AM, Gilles Gouaillardet <gil...@rist.or.jp
<mailto:gil...@rist.or.jp>> wrote:
Ralph,
here is some more detailed information.
from orte_ess_base_app_finalize()
first orte_rml_base_close() is invoked(via
mca_base_framework_close(&orte_rml_base_framework);
and it does
while (NULL != (item =
opal_list_remove_first(&orte_rml_base.posted_recvs))) {
OBJ_RELEASE(item);
}
then, opal_stop_progress_thread() is invoked
that means that when orte_rml_base_close is invoked, the progress
thread is up and running,
and can potentially invoke orte_rml_base_post_recv that does
if (req->cancel) {
OPAL_LIST_FOREACH(recv, &orte_rml_base.posted_recvs,
orte_rml_posted_recv_t) {
if (OPAL_EQUAL == orte_util_compare_name_fields(mask,
&post->peer, &recv->peer) &&
post->tag == recv->tag) {
opal_output_verbose(5,
orte_rml_base_framework.framework_output,
"%s canceling recv %d for peer %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
post->tag,
ORTE_NAME_PRINT(&recv->peer));
/* got a match - remove it */
opal_list_remove_item(&orte_rml_base.posted_recvs, &recv->super);
OBJ_RELEASE(recv);
break;
}
}
OBJ_RELEASE(req);
return;
}
/* this is where the assert fails */
since there is no lock, there is room for a race condition.
before all that happen, and in orte_ess_base_app_finalize(),
mca_base_framework_close(&orte_grpcomm_base_framework) invokes
finalize from grpcomm_rcd.c
that does
orte_rml.recv_cancel(ORTE_NAME_WILDCARD, ORTE_RML_TAG_ALLGATHER_RCD)
that is
orte_rml_oob_recv_cancel that ends up doing opal_event_set(...,
orte_rml_base_post_recv)
my first and naive attempt was to stop the opal_async progress thread
before closing the rml_base framework:
diff --git a/orte/mca/ess/base/ess_base_std_app.c
b/orte/mca/ess/base/ess_base_std_app.c
index d3bc6e6..4e09b47 100644
--- a/orte/mca/ess/base/ess_base_std_app.c
+++ b/orte/mca/ess/base/ess_base_std_app.c
@@ -353,18 +353,18 @@ int orte_ess_base_app_finalize(void)
(void) mca_base_framework_close(&orte_dfs_base_framework);
(void) mca_base_framework_close(&orte_routed_base_framework);
- (void) mca_base_framework_close(&orte_rml_base_framework);
- (void) mca_base_framework_close(&orte_oob_base_framework);
- (void) mca_base_framework_close(&orte_state_base_framework);
-
- orte_session_dir_finalize(ORTE_PROC_MY_NAME);
-
/* release the event base */
if (progress_thread_running) {
opal_stop_progress_thread("opal_async", true);
progress_thread_running = false;
}
+ (void) mca_base_framework_close(&orte_rml_base_framework);
+ (void) mca_base_framework_close(&orte_oob_base_framework);
+ (void) mca_base_framework_close(&orte_state_base_framework);
+
+ orte_session_dir_finalize(ORTE_PROC_MY_NAME);
+
return ORTE_SUCCESS;
}
that did not work : the opal_async progress thread is also used by
pmix, so at this stage,
invoking opal_stop_progress_thread only decrements the refcount (e.g.
no pthread_join() )
my second and dumb attempt was to finalize pmix before ess_base_app,
and that did not work
(crash)
i ran out of idea on how to fix this issue, but i found a simple
workaround :
adding a short sleep (10 ms) in orte_rml_base_close() seems enough to
avoid the race condition.
diff --git a/orte/mca/rml/base/rml_base_frame.c
b/orte/mca/rml/base/rml_base_frame.c
index 54dc454..050154c 100644
--- a/orte/mca/rml/base/rml_base_frame.c
+++ b/orte/mca/rml/base/rml_base_frame.c
@@ -17,6 +17,7 @@
#include "orte_config.h"
+#include <sys/poll.h>
#include <string.h>
#include "opal/dss/dss.h"
@@ -78,6 +79,7 @@ static int orte_rml_base_close(void)
{
opal_list_item_t *item;
+ poll(NULL,0,10);
while (NULL != (item =
opal_list_remove_first(&orte_rml_base.posted_recvs))) {
OBJ_RELEASE(item);
}
incidentally, i found two OPAL_LIST_FOREACH "loops" in which
opal_list_remove_item is invoked.
per a comment in opal_list.h, this is unsafe and
OPAL_LIST_FOREACH_SAFE should be used :
diff --git a/orte/mca/rml/base/rml_base_msg_handlers.c
b/orte/mca/rml/base/rml_base_msg_handlers.c
index 758bf91..22c7601 100644
--- a/orte/mca/rml/base/rml_base_msg_handlers.c
+++ b/orte/mca/rml/base/rml_base_msg_handlers.c
@@ -12,7 +12,9 @@
* All rights reserved.
* Copyright (c) 2007-2013 Los Alamos National Security, LLC. All
rights
* reserved.
- * Copyright (c) 2015 Intel, Inc. All rights reserved.
+ * Copyright (c) 2015 Intel, Inc. All rights reserved.
+ * Copyright (c) 2015 Research Organization for Information Science
+ * and Technology (RIST). All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
@@ -55,7 +57,7 @@ static void msg_match_recv(orte_rml_posted_recv_t
*rcv, bool get_all);
void orte_rml_base_post_recv(int sd, short args, void *cbdata)
{
orte_rml_recv_request_t *req = (orte_rml_recv_request_t*)cbdata;
- orte_rml_posted_recv_t *post, *recv;
+ orte_rml_posted_recv_t *post, *recv, *next;
orte_ns_cmp_bitmask_t mask = ORTE_NS_CMP_ALL | ORTE_NS_CMP_WILD;
opal_output_verbose(5, orte_rml_base_framework.framework_output,
@@ -75,7 +77,7 @@ void orte_rml_base_post_recv(int sd, short args,
void *cbdata)
* and remove it from our list
*/
if (req->cancel) {
- OPAL_LIST_FOREACH(recv, &orte_rml_base.posted_recvs,
orte_rml_posted_recv_t) {
+ OPAL_LIST_FOREACH_SAFE(recv, next,
&orte_rml_base.posted_recvs, orte_rml_posted_recv_t) {
if (OPAL_EQUAL == orte_util_compare_name_fields(mask,
&post->peer, &recv->peer) &&
post->tag == recv->tag) {
opal_output_verbose(5,
orte_rml_base_framework.framework_output,
@@ -120,12 +122,12 @@ void orte_rml_base_post_recv(int sd, short
args, void *cbdata)
void orte_rml_base_complete_recv_msg (orte_rml_recv_t **recv_msg)
{
- orte_rml_posted_recv_t *post;
+ orte_rml_posted_recv_t *post, *next;
orte_ns_cmp_bitmask_t mask = ORTE_NS_CMP_ALL | ORTE_NS_CMP_WILD;
opal_buffer_t buf;
orte_rml_recv_t *msg = *recv_msg;
/* see if we have a waiting recv for this message */
- OPAL_LIST_FOREACH(post, &orte_rml_base.posted_recvs,
orte_rml_posted_recv_t) {
+ OPAL_LIST_FOREACH_SAFE(post, next, &orte_rml_base.posted_recvs,
orte_rml_posted_recv_t) {
/* since names could include wildcards, must use
* the more generalized comparison function
*/
i hope this helps,
Gilles
On 7/17/2015 11:04 PM, Ralph Castain wrote:
It’s probably a race condition caused by uniting the ORTE and OPAL
async threads, though I can’t confirm that yet.
On Jul 17, 2015, at 3:11 AM, Gilles Gouaillardet
<gilles.gouaillar...@gmail.com
<mailto:gilles.gouaillar...@gmail.com>> wrote:
Folks,
I noticed several errors such as
http://mtt.open-mpi.org/index.php?do_redir=2244
that did not make any sense to me (at first glance)
I was able to attach one process when the issue occurs.
the sigsegv occurs in thread 2, while thread 1 is invoking
ompi_rte_finalize.
All I can think is a scenario in which the progress thread (aka
thread 2) is still dealing with some memory that was just
freed/unmapped/corrupted by the main thread.
I empirically noticed the error is more likely to occur when there
are many tasks on one node
e.g. mpirun --oversubscribe -np 32 a.out
Cheers,
Gilles
_______________________________________________
devel mailing list
de...@open-mpi.org <mailto:de...@open-mpi.org>
Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel
Link to this post:
http://www.open-mpi.org/community/lists/devel/2015/07/17652.php
_______________________________________________
devel mailing list
de...@open-mpi.org
Subscription:http://www.open-mpi.org/mailman/listinfo.cgi/devel
Link to this
post:http://www.open-mpi.org/community/lists/devel/2015/07/17655.php
_______________________________________________
devel mailing list
de...@open-mpi.org <mailto:de...@open-mpi.org>
Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel
Link to this post:
http://www.open-mpi.org/community/lists/devel/2015/07/17668.php
