[AGENT++] Patch to fix Mib::cleanup() and prevent possible deadlock
Claus Klein
claus.klein at arcormail.de
Tue Dec 16 23:56:59 CET 2014
Hi Frank,
I have all this done:
void init_signals()
{
signal(SIGTERM, sig);
signal(SIGINT, sig);
signal(SIGSEGV, sig);
// catch SIGPIPE and ignore it this can occur when the subagent dies when
// we are trying to response a request after timeout
#ifndef WIN32
signal(SIGPIPE, SIG_IGN);
#endif
}
The main loop looks like this:
Request* req;
do {
retries = 0;
while ((continueWork) && (!mib->get_agentx()->quit())) {
req = reqList->receive(20000); // ms
if (req) {
mib->process_request(req);
} else {
mib->ping_master();
//NOTE possible deadlock at cleanup() -> lock_mib() after get request timeout with Mib::process_request() -> lock_mib()!
//FIXME reqList->timeout_set_requests(); // does only help for set request timeout!
mib->cleanup();
// here we add or remove rows at shared ifTable
test_index_allocation();
}
}
// Make sure that all pending set requests which may
// have locked any resources are terminated and resources
// are freed before connection to master is reestablished.
reqList->terminate_set_requests();
//NOTE this calls may end in a deadlock after row destroy timeout or set request timeout! ck
mib->cleanup();
mib->save_all();
while ((continueWork) && (retries++ < 10) && (!mib->init())) {
Thread::sleep(retries * 1000); // ms (up to 10 sec)
}
} while ((continueWork) && (retries < 10));
Its my modified subagent example code!
If this is wrong, what is the right way to handle this?
Best regards,
Claus
On 16.12.2014, at 23:29, Frank Fock <fock at agentpp.com> wrote:
> Hi Claus,
>
> The root question to answer is: Why is the thread that runs through the MibEntry::update(..) method
> is terminated without unlocking its locks?
> I would expect the SIGPIPE handler to ignore the signal and set a flag for the main-loop to reconnect
> with the master. The reconnection (and terminate_set_requests) must then not be started before the
> requestList of the sub-agent runs empty.
>
> Best regards,
> Frank
More information about the AGENTPP
mailing list