Change up the logic a bit to handle the Red Storm case a bit better
(still need the real yod environment variable name), and add in some lengthy comments explaining the two different methods of debugger attach that we're using. MPI handle debugging doesn't seem to be working at the moment; still checking into it... This commit was SVN r18672.
Этот коммит содержится в:
Jeff Squyres
родитель
188c8bce5d
Коммит
da6aa57efb
@ -110,9 +110,7 @@ OMPI_DECLSPEC ompi_datatype_t* ompi_datatype_t_type_inclusion = NULL;
|
||||
|
||||
OMPI_DECLSPEC volatile int MPIR_debug_gate=0;
|
||||
|
||||
/* we don't believe we need MPIR_being_debugged here */
|
||||
|
||||
/* Check for a file in few dirrect ways for portability */
|
||||
/* Check for a file in few direct ways for portability */
|
||||
static void check(char *dir, char *file, char **locations)
|
||||
{
|
||||
char *str;
|
||||
@ -144,8 +142,34 @@ static void check(char *dir, char *file, char **locations)
|
||||
}
|
||||
|
||||
|
||||
/**
|
||||
* Wait for a debugger if asked.
|
||||
/*
|
||||
* Wait for a debugger if asked. We support two ways of waiting for
|
||||
* attaching debuggers:
|
||||
*
|
||||
|
||||
* 1. If using orterun: MPI processes will have the
|
||||
* ompi_mpi_being_debugged MCA param set to true. The HNP will call
|
||||
* MPIR_Breakpoint() and then RML send a message to VPID 0 (MCW rank
|
||||
* 0) when it returns (MPIR_Breakpoint() doesn't return until the
|
||||
* debugger has attached to all relevant processes). Meanwhile, VPID
|
||||
* 0 blocks waiting for the RML message. All other VPIDs immediately
|
||||
* call the grpcomm barrier (and therefore block until the debugger
|
||||
* attaches). Once VPID 0 receives the RML message, we know that the
|
||||
* debugger has attached to all processes that it cares about, and
|
||||
* VPID 0 then joins the grpcomm barrier, allowing the job to
|
||||
* continue. This scheme has the side effect of nicely supporting
|
||||
* partial attaches by parallel debuggers (i.e., attaching to only
|
||||
* some of the MPI processes; not necessarily all of them).
|
||||
*
|
||||
* 2. If not using orterun: in this case, ORTE_DISABLE_FULL_SUPPORT
|
||||
* will be true, and we know that there will not be an RML message
|
||||
* sent to VPID 0. So we have to look for a magic environment
|
||||
* variable from the launcher to know if the jobs will be attached by
|
||||
* a debugger (e.g., set by yod, srun, ...etc.), and if so, spin on
|
||||
* MPIR_debug_gate.
|
||||
*
|
||||
* Note that neither of these schemes use MPIR_being_debugged; it
|
||||
* doesn't seem useful to us.
|
||||
*/
|
||||
void ompi_wait_for_debugger(void)
|
||||
{
|
||||
@ -160,6 +184,12 @@ void ompi_wait_for_debugger(void)
|
||||
"is being debugged (default: false)",
|
||||
false, false, (int) false,
|
||||
&debugger);
|
||||
|
||||
/* Add in environment variables for other launchers, such as yod,
|
||||
srun, ...etc. */
|
||||
if (NULL != getenv("yod_you_are_being_debugged")) {
|
||||
debugger = 1;
|
||||
}
|
||||
|
||||
if (!debugger) {
|
||||
/* if not, just return */
|
||||
@ -167,7 +197,7 @@ void ompi_wait_for_debugger(void)
|
||||
}
|
||||
|
||||
/* if we are being debugged, then we need to find
|
||||
* the correct plug-in
|
||||
* the correct plug-ins
|
||||
*/
|
||||
a = strdup(opal_install_dirs.pkglibdir);
|
||||
mca_base_param_reg_string_name("ompi",
|
||||
@ -184,53 +214,42 @@ void ompi_wait_for_debugger(void)
|
||||
check(dirs[i], OMPI_MSGQ_DLL_PREFIX, mpimsgq_dll_locations);
|
||||
}
|
||||
}
|
||||
|
||||
/* only the rank=0 proc waits for the debugger - everyone else will just
|
||||
* spin in the barrier in mpi_init until rank=0 joins them
|
||||
*/
|
||||
if (0 != ORTE_PROC_MY_NAME->vpid) {
|
||||
return;
|
||||
}
|
||||
|
||||
/* we have to support at least two ways of completing the
|
||||
* debug attachment - either we will get a message from
|
||||
* the HNP telling us it is okay to release, or the debugger
|
||||
* itself will reach into us and set a gate.
|
||||
*
|
||||
* First, attempt to get a message-based release
|
||||
*/
|
||||
OBJ_CONSTRUCT(&buf, opal_buffer_t);
|
||||
rc = orte_rml.recv_buffer(ORTE_NAME_WILDCARD, &buf, ORTE_RML_TAG_DEBUGGER_RELEASE, 0);
|
||||
OBJ_DESTRUCT(&buf); /* don't care about contents of message */
|
||||
|
||||
if (rc > 0) {
|
||||
/* message received - we can go! */
|
||||
return;
|
||||
} else if (ORTE_ERR_NOT_SUPPORTED == rc) {
|
||||
/* if the recv isn't supported, then we fall back
|
||||
* to the alternative method for waiting
|
||||
*/
|
||||
goto spin_wait;
|
||||
} else {
|
||||
/* if it failed for some other reason, then we are
|
||||
* in trouble - for now, just report the problem
|
||||
* and give up waiting
|
||||
*/
|
||||
opal_output(0, "Debugger_attach[rank=%ld]: could not wait for debugger - error %s!",
|
||||
(long)ORTE_PROC_MY_NAME->vpid, ORTE_ERROR_NAME(rc));
|
||||
return;
|
||||
}
|
||||
|
||||
spin_wait:
|
||||
/* spin until debugger attaches and releases us */
|
||||
while (MPIR_debug_gate == 0) {
|
||||
#if defined(__WINDOWS__)
|
||||
Sleep(100); /* milliseconds */
|
||||
#elif defined(HAVE_USLEEP)
|
||||
usleep(100000); /* microseconds */
|
||||
#else
|
||||
sleep(1); /* seconds */
|
||||
#endif
|
||||
}
|
||||
}
|
||||
|
||||
if (ORTE_DISABLE_FULL_SUPPORT) {
|
||||
/* spin until debugger attaches and releases us */
|
||||
while (MPIR_debug_gate == 0) {
|
||||
#if defined(__WINDOWS__)
|
||||
Sleep(100); /* milliseconds */
|
||||
#elif defined(HAVE_USLEEP)
|
||||
usleep(100000); /* microseconds */
|
||||
#else
|
||||
sleep(1); /* seconds */
|
||||
#endif
|
||||
}
|
||||
} else {
|
||||
|
||||
/* only the rank=0 proc waits for either a message from the
|
||||
* HNP or for the debugger to attach - everyone else will just
|
||||
* spin in * the grpcomm barrier in ompi_mpi_init until rank=0
|
||||
* joins them.
|
||||
*/
|
||||
if (0 != ORTE_PROC_MY_NAME->vpid) {
|
||||
return;
|
||||
}
|
||||
|
||||
/* VPID 0 waits for a message from the HNP */
|
||||
OBJ_CONSTRUCT(&buf, opal_buffer_t);
|
||||
rc = orte_rml.recv_buffer(ORTE_NAME_WILDCARD, &buf,
|
||||
ORTE_RML_TAG_DEBUGGER_RELEASE, 0);
|
||||
OBJ_DESTRUCT(&buf); /* don't care about contents of message */
|
||||
if (rc < 0) {
|
||||
/* if it failed for some reason, then we are in trouble -
|
||||
* for now, just report the problem and give up waiting
|
||||
*/
|
||||
opal_output(0, "Debugger_attach[rank=%ld]: could not wait for debugger - error %s!",
|
||||
(long)ORTE_PROC_MY_NAME->vpid, ORTE_ERROR_NAME(rc));
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
|
||||
Загрузка…
x
Ссылка в новой задаче
Block a user