Process trees and termination

[roger riggs]

Hi Peter,

Thanks for the comments...

On 6/26/2014 4:55 AM, Peter Levart wrote:
> On 06/26/2014 12:05 AM, David M. Lloyd wrote:
>> On 06/25/2014 04:52 PM, roger riggs wrote:
>>> Hi,
>>>
>>> The next step for JEP 102, Process API Updates is to provide an API to
>>> enumerate the hierarchy of processes and to be able to terminate a
>>> process tree.
>>>
>>> This draft javadoc <http://bussund0416/%7Erriggs/pdoc/>update [1] 
>>> includes:
>>>   - ProcessHandle class to provide handles to processes, their 
>>> children,
>>>      parent, process id and liveness
>>>   - Process extends ProcessHandle and inherits the new behaviors
>>>   - Method destroyProcessTree() terminates a process tree
>>>
>>> Additional functions related to termination and information about 
>>> processes
>>> will be added in a later update.
>>>
>>> Comments appreciated, Roger
>>
>> I still have a not-great feeling about the lack of atomicity inherent 
>> in these operations.  At least when you're dealing with direct child 
>> processes (on UNIX-likes), there is some measure of control, in that 
>> a process is zombified until reaped by wait() and friends.
>>
>> Also I have a much higher level of concern about 
>> destroyProcessTree(), which I don't think has any place in this API.  
>> Control of any process belongs solely with their parent (or PID 1), 
>> in my opinion.
>>
>> Those concerns aside, what about making the allChildren() and 
>> allProcesses() methods lazy in some way?
>>
> Hi Roger, David,
>
> - Will there be a guarantee that ProcessHandle objects returned from 
> factory methods: |current()| 
> <http://cr.openjdk.java.net/%7Erriggs/ph-apidraft/java/lang/ProcessHandle.html#current-->, 
> |of(long)| 
> <http://cr.openjdk.java.net/%7Erriggs/ph-apidraft/java/lang/ProcessHandle.html#of-long->, 
> |children()| 
> <http://cr.openjdk.java.net/%7Erriggs/ph-apidraft/java/lang/ProcessHandle.html#children-->, 
> |allChildren()| 
> <http://cr.openjdk.java.net/%7Erriggs/ph-apidraft/java/lang/ProcessHandle.html#allChildren-->, 
> |getParent()| 
> <http://cr.openjdk.java.net/%7Erriggs/ph-apidraft/java/lang/ProcessHandle.html#getParent--> 
> and |allProcesses()| 
> <http://cr.openjdk.java.net/%7Erriggs/ph-apidraft/java/lang/ProcessHandle.html#allProcesses--> 
> representing those processes that were started by ProcessBuilder API 
> are actually the same Process objects that were returned from the 
> ProcessBuilder API? The Process object can be viewed as a proxy of an 
> OS process so as long as we have a reference to it, we can obtain it's 
> exit status even though it has already been waited for (reaped) and we 
> can reliably terminate it, because Process object is tracking the 
> state of OS process and only forwards the termination request if it 
> knows the OS process is still alive or zombified and so minimizes the 
> probability to kill wrong process that just reused the same pid.
The initial design does not have them returning Process instances.
Process instances are capabilities and their reference accessibility is 
currently controlled
by the code that created the Process.  If ProcessHandle gave out those 
Process instances
it would break existing encapsulation.

>
> The following expression: ProcessHandle.current().children() would 
> therefore return just canonicalized Process objects (unless some 
> native code embedded in the JVM process started sub-processes too).
>
> If returning canonicalized Process objects from factory methods is not 
> desirable (some part of the program could mess with Input/Output 
> streams of Process-es it did not have access to before), then perhaps 
> ProcessHandle representing a process created by ProcessBuilder API 
> would just have an internal private reference to the Process object to 
> which it would forward method invocations...
yes, that would break encapsulation; not desirable and potential 
security issue.

And yes, the implementation can maintain the association and delegate 
the behavior.

It does not/should not make any difference in practice.
Regardless of whether the TERM or KILL signal is delivered by an external
process or via the ProcessHandle API, the Process mechanics should be 
the same.
An application using the Process should see the same behavior, cleanup
and termination, in either case.
>
>
> - There is an inconsistency in the javadoc for 
> ProcessHandle.destroyForcibly():
>
> public abstractProcessHandle  <http://cr.openjdk.java.net/%7Erriggs/ph-apidraft/java/lang/ProcessHandle.html>  destroyForcibly()
> Kills the process. The process represented by this |ProcessHandle| 
> object is forcibly terminated. If the process is not alive no action 
> is taken.
>
> Note: The process may not terminate immediately. i.e. |isAlive()| may 
> return true for a brief period after |destroyForcibly()| is called. 
> This method may be chained to |waitFor()| if needed.
>
>
> ...the method has a co-variant return type and there's no waitFor() 
> method on the ProcessHandle. Why not? If ProcessHandle is a (or 
> delegates to) Process then it could use it's waitFor(), if it is just 
> a "foreign" ProcessHandle it could be implemented as a poll-ing wait 
> for the isAlive() to return false...
>
I plan to address waitFor of ProcessHandle a bit later due to the 
complexities
previously discussed for non-Process processes and when I can spend more 
time
on the implementations and alternatives.
Also requested are either Future style returns or lambda enabled onExit 
behavior.

>
> - Will there not be any destroyGracefully() method (which would be a 
> no-op on platforms not supporting graceful termination)?
An alternative is to add a method to report whether destroy is graceful 
or not.
Either way an application needs to be aware of whether it worked or not so
it can take some alternate action.

>
>
> - Depending on the OS facilities used to obtain process information 
> (/proc/<pid> ?) there might be some other attributes already 
> available. The name of a process would be nice to have.
yep, also on the list and it builds naturally on top of ProcessHandle
when the relationship between ProcessHandle and Process is settled.

Thanks, Roger

>
>
> Regards, Peter
>