ProcessReaper: single thread reaper

[roger riggs]

Hi,

Another approach was suggested by a member of the Solaris team.

If you open /proc/pid O_RDONLY for any process you wish to monitor
and use poll(2), you can wait for a hangup event which indicates that
the process has exited.  You can then reap that process's status w/
waitpid.  You'll also want to wait on a pipe; as you fork additional
processes you can write their pid to the pipe and the monitoring thread
can wake up, read the pipe and add that fd to the pollfds.  This will
work for any version of Solaris you support, and it uses a fd per
process; using the pipe mechanism no locking is required.

I haven't had a chance to followup but it is interesting that monitoring 
/proc
can give an indication of Process termination and some aspects should
work on Linux, not just Solaris.

Roger


On 4/14/2014 5:02 AM, Peter Levart wrote:
> Hi Martin, Roger,
>
> Just a thought. Would it be feasible to have two (ore more) built-in 
> strategies, selectable by system property? A backwards compatible 
> tread per child, using waitpid(pid, ...), a single reaper thread using 
> waitpid(-1, ...), maybe also single threaded strategy accessible only 
> on Linux/Solaris using waitid(-1, ..., WNOWAIT)... All packed nicely 
> in a package-private interface (ProcessReaper) with multiple 
> implementations?
>
> Regards, Peter
>
> On 04/12/2014 01:37 AM, Martin Buchholz wrote:
>> Let's step back again and try to check our goals...
>>
>> We could try to optimize the one-reaper-thread-per-subprocess thing. 
>>  But that is risky, and the cost of what we're doing today is not 
>> that high.
>>
>> We could try to implement the feature of killing off an entire 
>> subprocess tree.  But historically, any kind of behavior change like 
>> that has been vetoed.  I have tried and failed to make less 
>> incompatible changes.  We would have to add a new API.
>>
>> The reality is that Java does not give you real access to the 
>> underlying OS, and unless there's a seriously heterodox attempt to 
>> provide OS-specific extensions, people will have to continue to 
>> either write native code or delegate to an OS-savvy subprocess like a 
>> perl script.
>>
>>
>> On Fri, Apr 11, 2014 at 7:52 AM, Peter Levart <peter.levart at gmail.com 
>> <mailto:peter.levart at gmail.com>> wrote:
>>
>>     On 04/09/2014 07:02 PM, Martin Buchholz wrote:
>>>
>>>
>>>
>>>     On Tue, Apr 8, 2014 at 11:08 PM, Peter Levart
>>>     <peter.levart at gmail.com <mailto:peter.levart at gmail.com>> wrote:
>>>
>>>         Hi Martin,
>>>
>>>         As you might have seen in my later reply to Roger, there's
>>>         still hope on that front: setpgid() + wait(-pgid, ...) might
>>>         be the answer. I'm exploring in that direction. Shells are
>>>         doing it, so why can't JDK?
>>>
>>>         It's a little trickier for Process API, since I imagine that
>>>         shells form a group of processes from a pipeline which is
>>>         known in-advance while Process API will have to add
>>>         processes to the live group dynamically. So some races will
>>>         have to be resolved, but I think it's doable.
>>>
>>>
>>>     This is a clever idea, and it's arguably better to design
>>>     subprocesses so they live in separate process groups (emacs does
>>>     that), but:
>>>     Every time you create a process group, you change the effect of
>>>     a user signal like Ctrl-C, since it's sent to only one group.
>>>     Maybe propagate signals to the subprocess group?  It's starting
>>>     to get complicated...
>>>
>>
>>     Hi Martin,
>>
>>     Yes, shells send Ctrl-C (SIGINT) and other signals initiated by
>>     terminal to a (foreground) process group. A process group is
>>     formed from a pipeline of interconnected processes. Each pipeline
>>     is considered to be a separate "job", hence shells call this
>>     feature "job-control". Child processes by default inherit process
>>     group from it's parent, so children born with Process API (and
>>     their children) inherit the process group from the JVM process.
>>     Considering the intentions of shell job-controll, is propagating
>>     SIGTERM/SIGINT/SIGTSTP/SIGCONT signals to children spawned by
>>     Process API desirable? If so, then yes, handling those signals in
>>     JVM and propagating them to current process group that contains
>>     all children spawned by Process API and their descendants would
>>     have to be performed by JVM. That problem would certainly have to
>>     be addressed. But let's first see what I found out about
>>     sigaction(SIGCHLD, ...), setpgid(pid, pgid), waitpid(-pgid, ...),
>>     etc...
>>
>>     waitpid(-pgid, ...) alone seems to not be enough for our task.
>>     Mainly because a process can re-assign it's group and join some
>>     other group. I don't know if this is a situation that occurs in
>>     real world, but imagine if we have one live child process in a
>>     process group pgid1 and no unwaited exited children. If we issue:
>>
>>         waitpid(-pgid1, &status, 0);
>>
>>     Then this call blocks, because at the time it was given, there
>>     were >0 child processes in the pgid1 group and none of them has
>>     exited yet. Now if this one child process changes it's process
>>     group with:
>>
>>         setpgid(0, pgid2);
>>
>>     Then the waitpid call in the parent does not return (maybe this
>>     is a bug in Linux?) although there are no more live child
>>     processes in the pgid1 group any more. Even when this child
>>     exits, the call to waitpid does not return, since this child is
>>     not in the group we are waiting for when it exits. If all our
>>     children "escape" the group in such way, the tread doing waiting
>>     will never unblock. To solve this, we can employ signal handlers.
>>     In a signal handler for SIGCHLD signal we can invoke:
>>
>>         waitpid(-pgid1, &status, WNOHANG); // non-blocking call
>>
>>     ...in loop until it either returns (0) which means that there're
>>     no more unwaited exited children in the group at the momen or
>>     (-1) with errno == ECHILD, which means that there're no more
>>     children in the queried group any more - the group does not exist
>>     any more. Since signal handler is invoked whith SIGCHLD being
>>     masked and there is one bit of pending signal state in the
>>     kernel, no child exit can be "skipped" this way. Unless the child
>>     "escapes" by changing it's group. I don't know of a plausible
>>     reason for a program to change it's process group. If a program
>>     executing as JVM child wants to become a background daemon it
>>     usually behaves as follows:
>>
>>     - fork()s a grand-child and then exit()s (so we get notified via
>>     signal and waitpid(-pgid, ...) successfully for it's exitstatus)
>>     - the grand-child then changes it's session and group (becomes
>>     session and group leader), closes file descriptors, etc. The
>>     responsibility for waiting on the grand-child daemon is
>>     transferred to the init process (pid=1) since the grand-child
>>     becomes an orphan (has no parent).
>>
>>     Ignoring this still unsolved problem of possible ill-behaved
>>     child program that changes it's process group, I started
>>     constructing a proof-of-concept prototype. What I will do in the
>>     prototype is start throwing IllegalStateException from the
>>     methods of the Process API that pertain to such children. I think
>>     this is reasonable.
>>
>>     Stay tuned,
>>
>>     Peter
>>
>>
>>
>