Yes, that is what I am suggesting. I believe it will make the code in PR 155 redundant, as well.
On Feb 8, 2017 00:50, "Jeremy Mitchell" <[email protected]> wrote: So you're suggesting a different solution that this? https://github.com/apache/incubator-trafficcontrol/commit/ abb936a85f6bb09169329b8daf1d2b2c094e07f1 you're suggesting that the code in that commit be reverted and a double-fork be pursued instead? jeremy On Tue, Feb 7, 2017 at 3:33 PM, Oren Shemesh <[email protected]> wrote: > I believe that the merge of PR 155 solved a problem, in the same way I > tried to solve the mail problem: Make sure that SIGCHLD is set to DEFAULT > and not to IGNORE. > > But after researching a bit about daemonizing in Python, I see no > reason for the $SIG{CHLD} = 'IGNORE' statement to be a part of basic > daemonizing. > Hence, I believe that this statement might be the root cause of the > problem. > Without setting SIGCHLD to IGNORE, both the problem PR 155 solves, and the > mail problem I stumbled upon, will be gone. > > This statement was added in this commit: > > commit abb936a85f6bb09169329b8daf1d2b2c094e07f1 > > Fixes an issue where forked processes become defunct rather than > exiting cleanly. > > I suggest to solve the problem of zombie processes with a double-fork > (Double-forking is a common technique in daemonizing, used to make sure the > daemon is completely detached from it's parent): > > 1. Parent forks first child > 2. Parent does waitpid() > 3. First child forks the daemon > 4. First child exits with a return code indicating the success of the > second fork. > 5. Parent gets this return code and knows that the second fork went OK. > 6. Because the first child is dead, the parent of the daemon is set to 1 > (init), so when the daemon exits, it does not become a zombie (init > reclaims it). > > This way the SIGCHLD handler is not affected, in both parent and daemon. > > Comments ? > > Thanks, Oren. > > > > On Tue, Feb 7, 2017 at 10:40 PM, Dave Neuman <[email protected]> wrote: > > > Thanks Jeremy. We did merge that PR not too long ago to fix one issue, > but > > it looks like it created another. I agree that it would be better to > use a > > different library. > > > > On Tue, Feb 7, 2017 at 1:29 PM, Jeremy Mitchell <[email protected]> > > wrote: > > > > > I'll have to defer to someone else as most of that was way over my head > > but > > > I do know we have this ticket outstanding: > > > https://issues.apache.org/jira/browse/TC-74 > > > > > > and also, this PR was merged a while back regarding $SIG{CHLD} - > > > https://github.com/apache/incubator-trafficcontrol/pull/155 > > > > > > I'm not really sure if it's a good or bad idea to wrap every call to > > > $self->mail(...); > > > as you suggested. maybe others have thoughts on that. > > > > > > preferably, it would be nice to get TC-74 done and remove our > dependency > > on > > > an outdated library. > > > > > > jeremy > > > > > > > > > On Tue, Feb 7, 2017 at 4:27 AM, Oren Shemesh <[email protected]> wrote: > > > > > > > Hi, > > > > > > > > I'm looking for someone to form an opinion about a problem I believe > I > > > have > > > > found, and a fix. > > > > > > > > The problem manifests itself as an error when TO attempts to send an > > > > E-mail. > > > > > > > > The details: > > > > When traffic ops runs, it daemonizes itself, and as part of that, it > > does > > > > $SIG{CHLD} = 'IGNORE'; : > > > > > > > > https://github.com/apache/incubator-trafficcontrol/blob/ > > > > master/traffic_ops/app/lib/MojoPlugins/Daemonize.pm#L28 > > > > > > > > But, when sending a mail, TO uses Mojolicous::Plugin::Mail, which in > > turn > > > > is a wrapper for the standard (And Obsolete) perl module MIME::Lite. > > > > > > > > MIME::Lite sends a mail by running the following piece of code > (Snippet > > > > from app/local/lib/perl5/MIME/Lite.pm): > > > > > > > > my $pid = open SENDMAIL, "|-"; > > > > defined($pid) or die "open of pipe failed: $!\n"; > > > > if ( !$pid ) { ### child > > > > exec(@cmd) or die "can't exec $p{Sendmail}: $!\n"; > > > > ### NOTREACHED > > > > } else { ### parent > > > > $self->print( \*SENDMAIL ); > > > > close SENDMAIL || die "error closing $p{Sendmail}: $! > (exit > > > > $?)\n"; > > > > $return = 1; > > > > } > > > > > > > > The problem with this code is that assumes that the SIGCHLD handler > is > > > left > > > > at it's DEFAULT handling, which means that when a child process > > > terminates, > > > > the parent process can reap it during the 'close' statement. > > > > Traffic Ops runs with the SIGCHLD handler set to IGNORE, which means > > that > > > > when a child process dies, the parent process forgets about it > > > immediately. > > > > So, if sendmail runs fast enough and finishes before the 'close' > > command > > > is > > > > executed, the close will fail to find the child process, and the > error > > > > propagates all the way to the calling code, looking something like > > this: > > > > > > > > Traffic Ops fatal error occurred while processing your request. > > > > Error at line 2732 ( close SENDMAIL || die "error closing > $p{Sendmail}: > > > $! > > > > (exit $?)\n";) > > > > error closing /usr/lib/sendmail: No child processes (exit -1) > > > > > > > > It should be noted that we are running TO on single-core machines for > > > > development, which means that the code is much more likely to fail in > > > this > > > > manner (In machines with more cores, the parent process usually does > > not > > > > get suspended when sendmail is spawned, hence it reaches the 'close' > > > > command immediately and waits for sendmail to finish properly). > > > > > > > > I am not sure why TO needs to completely ignore SIGCHLD, and I am not > > > sure > > > > if it is reasonable to ask from a module such as MIME::Lite to work > > when > > > > SIGCHLD is set to IGNORE. > > > > However, MIME::Lite is not being maintained any more (The owner > > > recommends > > > > using better modules, see this > > > > <http://search.cpan.org/~rjbs/MIME-Lite-3.030/lib/MIME/Lite.pm#WAIT > > !>), > > > > and > > > > the latest Mojolicous::Plugin::Mail version still uses MIME::Lite. > > > > (I guess Mojolicous is not very actively maintained these days). > > > > Which means that until we switch to the experimental, Angular-based, > > > > Traffic Ops, we are stuck with the current MIME::Lite code. > > > > > > > > Maybe a good fix for this issue is to not set SIGCHLD to IGNORE while > > > > daemonizing. However, this means that every time we create a new > > process, > > > > we must waitpid() for it, otherwise we will accumulate zombies. > > > > I guess it is not very practical to ensure this across the entire > code > > > base > > > > of TO (And, as we have seen, all Perl modules that it uses). > > > > > > > > So the low-footprint fix I want to make is to wrap every call to > > > > $self->mail(...) done in our Email.pm > > > > <https://github.com/apache/incubator-trafficcontrol/blob/ > > > > master/traffic_ops/app/lib/MojoPlugins/Email.pm> > > > > code, like this: > > > > > > > > $SIG{CHLD} = 'DEFAULT'; > > > > $self->mail(...); > > > > $SIG{CHLD} = 'IGNORE'; > > > > > > > > I tested this, and it fixes the problem. > > > > > > > > Thoughts ? > > > > > > > > I plan to create a JIRA and a PR, but I wanted to get someone's more > > > expert > > > > view on this before I go on. > > > > Being a bit of a novice, maybe I am missing some key thing here. > > > > > > > > Thanks ,Oren. > > > > -- > > > > > > > > *Oren Shemesh* > > > > Qwilt | Work: +972-72-2221637| Mobile: +972-50-2281168 | > > [email protected] > > > > <[email protected]> > > > > > > > > > > > > > -- > > *Oren Shemesh* > Qwilt | Work: +972-72-2221637| Mobile: +972-50-2281168 | [email protected] > <[email protected]> >
