Contents |
Debugging Applications and SAFplus Platform Components
The high availability features of the SAFplus Platform interfere with debugging applications because the act of debugging an application breaks the performance and availability "contract" between the SAFplus Platform and an application and therefore causes the SAFplus Platform to restart the application.
This HOWTO describes strategies that can be used to debug within an high availability environment. It is not a primer on gdb. In fact, it assumes an in-depth knowledge of debugging using gdb (or gdb derivatives such as Wind River's IDE) and therefore focuses exclusively on the interactions between the debugger and the SAFplus Platform.
It is useful to toggle the following debugging interventions on an as needed basis since many "defeat" the high availability features of the SAFplus Platform. Therefore the examples below use a global variable (that you could define in your component's clCompAppMain.c) called debuggingOn to control some features.
int debuggingOn = 1; /* Set to 0 to turn debugging off, 1 to turn it on */
Debugging Configuration Settings
To aid in debugging, options exist to "defeat" the normal SAFplus monitoring of processes and threads. The following settings are accessible either via global or environment variables. In the list below, the environment variable is in parenthesis. The global variable can be used to initiate debugging behavior some time during the execution of your process. The environment variable will cause debugging behavior across the entire SAFplus model. These variables are all "FALSE" (not in debug mode) by default, setting to TRUE will enable their effects.
- clDbgPauseOn (CL_DEBUG_PAUSE)
If a thread is going to abort or assert, SAFplus can cause the thread to block on a semaphore instead. This allows the engineer to attach to the process via gdb and look at the thread's state.
- clDbgPauseOnCodeError (CL_DEBUG_PAUSE_CODE_ERROR)
If clDbgPauseOn is TRUE, then this variable enables "pauses" on errors that are clearly application bugs -- such as passing a NULL pointer to a SAFplus API. Normally, SAFplus API calls return an error in these cases. But realistically, an application's error handling code can rarely solve bugs in itself -- other then via very granular solutions such as running with reduced functionality or by exiting gracefully. By pausing at the moment the bug is detected, the stack remains available for analysis through gdb.
- clDbgNoKillComponents (CL_DEBUG_COMP_NO_KILL)
This global variable exists in the AMF server. If the AMF was going to kill a component for any reason, it will not do so. This allows the developer to attach to it via gdb. However, note that the AMF will continue as if it had actually killed the component and so proper AMF operation after then event is not guaranteed. For this reason, other ways to avoid common cases where the AMF kills a component are discussed in the next sections.
- clDbgCompTimeoutOverride (CL_DEBUG_COMP_TIMEOUT_OVERRIDE)
This global variable exists in the AMF server. If the AMF was going to kill a component due to a timeout, it will not do so. This allows the developer to attach to it via gdb. However, note that the AMF will continue as if it had actually killed the component and so proper AMF operation after then event is not guaranteed. For this reason, other ways to avoid component timeouts are discussed in the next few sections.
Stopping Process Health Monitoring
It is extremely irritating to attach to your process via a debugger only to have the SAFplus's high availability features intervene and kill the process you are trying to debug! Certain techniques can be used to temporarily "defeat" this feature to allow debugging to occur. Of course, while the feature is "off" the high availability features do not work.
Turning off heartbeat
Use saAmfHealthcheckStop() to stop heartbeat, or do not call saAmfHealthcheckStart() to begin with. AMF heartbeating is discussed in detail in the SAF AMF Specification
"ACK"ing Messages Early
The SAFplus Platform also monitors the response of certain crucial messages to your component. If your component does not respond in time, the SAFplus Platform kills your component. This feature ensures that your application does not have a bug in the handling of these messages; however when debugging the message handling we must "defeat" this feature. The most common messages are the CSI (work) assignment messages, which trigger the "clCompAppAMFCSISet" (clCompAppMain.c) function call. When debugging your component, it is best to "ACK" the message right away so the SAFplus Platform will think that you have handled the message. However, it is important to NOT do so when not in debugging mode, to ensure that the HA features work.
In the routine below, note that the "clCpmResponse(cpmHandle, invocation, CL_OK);" function has been moved above the switch statement when debugging is on (and moved to below it, for succinctness, when debugging is off).
ClRcT clCompAppAMFCSISet(
ClInvocationT invocation,
const ClNameT *compName,
ClAmsHAStateT haState,
ClAmsCSIDescriptorT csiDescriptor)
{
/* If component hang detection is off, then reply with an OK first,
so that this call won't time out causing the controller to kill me.
"debuggingOn" is NOT part of SAFplus. It is a global variable that
you could define in your own application to control this behavior.
*/
if (debuggingOn)
{
if (haState != CL_AMS_HA_STATE_QUIESCING)
clCpmResponse(cpmHandle, invocation, CL_OK);
}
/* DO NOT SET BREAKPOINTS ABOVE THIS LINE (IN THIS FUNCTION)
(the controller will think that your process is dead)
To set a breakpoint in this routine:
1. First set "debuggingOn" to 1
2. Set your breakpoint below this comment
3. Continue
*/
/* Print information about the CSI Set */
clprintf ("Component [%s] : PID [%d]. CSI Set Received\n",
compName->value, mypid);
clCompAppAMFPrintCSI(csiDescriptor, haState);
/* Take appropriate action based on state */
switch ( haState )
{
case CL_AMS_HA_STATE_ACTIVE:
{
/*
AMF has requested application to take the active HA state
for the CSI.
*/
break;
}
case CL_AMS_HA_STATE_STANDBY:
{
/*
AMF has requested application to take the standby HA state
for this CSI.
*/
break;
}
case CL_AMS_HA_STATE_QUIESCED:
{
/*
AMF has requested application to quiesce the CSI currently
assigned the active or quiescing HA state. The application
must stop work associated with the CSI immediately.
*/
break;
}
case CL_AMS_HA_STATE_QUIESCING:
{
/*
AMF has requested application to quiesce the CSI currently
assigned the active HA state. The application must stop work
associated with the CSI gracefully and not accept any new
workloads while the work is being terminated.
*/
clCpmCSIQuiescingComplete(cpmHandle, invocation, CL_OK);
break;
}
default:
{
/*
Should never happen. Ignore.
*/
}
}
/* If I'm am killing components, then reply with an OK after
execution, to ensure that the handling did not hang */
if (!debuggingOn)
{
if (haState != CL_AMS_HA_STATE_QUIESCING)
clCpmResponse(cpmHandle, invocation, CL_OK);
}
return CL_OK;
}
Component State Transition Timeouts
The SAFplus Platform also ensures that your component transitions through various states within reasonable a time. To debug during these transitions (startup, shutdown, etc) it is necessary to increase these timeouts. These timeouts are available in the IDE, or directly in the XML configuration files for temporary changes.
Search for the following XML in your "etc/clAmfDefinitions.xml" file and change the numbers for every component you need to debug to large values.
<timeouts>
<instantiateTimeout>1000000</instantiateTimeout>
<terminateTimeout>1000000</terminateTimeout>
<cleanupTimeout>1000000</cleanupTimeout>
<amStartTimeout>1000000</amStartTimeout>
<amStopTimeout>1000000</amStopTimeout>
<quiescingCompleteTimeout>1000000</quiescingCompleteTimeout>
<csiSetTimeout>1000000</csiSetTimeout>
<csiRemoveTimeout>1000000</csiRemoveTimeout>
<proxiedCompInstantiateTimeout>1000000</proxiedCompInstantiateTimeout>
<proxiedCompCleanupTimeout>1000000</proxiedCompCleanupTimeout>
</timeouts>
Creating core files on failure
Core files are configured to be automatically created and will be found int the "var/run" directory. If core files are not being created, your environment configuration may be overriding our changes.
Make sure the core file size is "unlimited" by running "ulimit -c unlimited" before executing SAFplus.
You can test core creation by jamming a SEGFAULT into a program by doing "kill -SEGV <process id>".
Pausing and Resuming a Thread
Sometimes a problem occurs during startup -- before you have the opportunity to attach to a process, set a breakpoint, and resume the process. In a standard application this is easy to debug because "gdb" starts up with your program "stopped". But in a high availability framework, where the SAFplus Platform starts and stops programs, this is trickier.
To debug this, 2 functions are available from "clDebugApi.h", one which "pauses" the calling thread, the other "resumes" it. The programmer adds the "pause" function to the code and recompiles. When the function is called and the global variable clDbgPauseOn (or CL_DEBUG_PAUSE environment variable) is true, the thread will output a log and block on a semaphore. The programmer can then check the logs to see if a thread was paused. If so, he goes into gdb, attaches to the process, finds the thread that is paused -- "thread apply all backtrace" ("thr a a bt") and look for a call to "pause" -- switches to that thread ("thread x"), and then debugs the thread. The programmer can even "resume" the thread by calling the clDbgResume function from gdb's prompt:
gdb> call clDbgResume()
The "pause" function is actually a macro so that the file and line from which the function was called can be passed. This information is very useful if a log is all that is provided for debugging.
Intercepting the Automatic Start of a Component
It is possible to run your component within a wrapper application (such as gdb or purify). When the SAFplus Platform is deployed on a target, all applications are placed in the "bin" subdirectory of the SAFplus Platform tree (for example "/root/asp/bin"). To intercept the application start for debugging purposes:
- rename your application
- create a script and name it the same as the original application name
- edit the script to add your wrapper.
- Note that you will have to start a separate window for applications that need a console (like gdb)
For example the following simple script can be used to make your application start within gdb (note though that it must be used in conjunction with the techniques described in the "Stopping Process Health Monitoring" section above unless you are a very fast typist, since your application's initialization will be monitored by the OpenClovis HA infrastructure to ensure that it succeeds). This script assumes that your application is called "myapp" and that you moved it to "/root/asp/bin/myapp.orig". It also assumes that your Linux installation has X-windows and the gnome desktop environment installed (however, a similar technique can be used for other window managers):
#!/bin/bash
/usr/bin/gnome-terminal --window --command "gdb /root/asp/bin/myapp.orig $1 $2 $3 $4 $5 $6 $7 $8 $9"