and useAppend and all that

[Stefan Bodewig <bo...@apache.org>]

Hi,

we have some discussion going on in Bugzilla that we probably need to
pull together at a single place.

[BTW since minotaur is down I haven't received any mails to @apache.org
for several hours which includes most of my mailing lists subscriptions
so I may already have missed a separate thread.]

In order to fix Bug 5003 StreamPumper gained the useAvailable flag that
made it use a timed and interruptable read on input streams because
sometimes a process on Windows doesn't close its streams once it
finishes (the streams are handed down to child processes which keep them
open and useAvailable detaches Ant from those child processes).

There is an AntUnit test for this in trunk: exec-test
testDoesntWaitForChildren which silently fails on Windows if you set
useAvailable to false.  It fails if the test takes longer than 20
seconds, it should finish in about three.

[Next sidenote, we will probably need to add a timing assertion to
AntUnit to make the test fail not-so-silently].

A side effect of this is that now each execution on Windows takes at
least the amount of time we wait for joining threads.  Currently 200ms.
This makes tasks that perform many small executions like attrib a lot
slower than they used to be.  I don't see any workaround here.  This was
bug 48734.

In the meantime bug 48746 was fixed by setting useAvailable to true
outside of Windows as well, which brings bug 48734 to other OSes as well
- bug 49384.

It also seems to cause problems for NetBeans, see Jesse's comments in
bug 48746.

And then there is bug 46805 which claims that even useAvailable=true
doesn't fix a hanging process on Windows and suggests that we destroy
the process instance and not close the process' streams explicitly at
all.  I have no test cases for this bug and can't reproduce it myself.

process.destroy makes Ant lose output from the forked process, so it
doesn't seem a viable option.

Not closing the streams doesn't seem to do any harm on Windows but I
can't confirm that it helps in any way.  I did see some issues on the
only Linux box I can reach right now, but that may be due to other
problems (other exec stuff fails there as well).

If fiddling with useAvailable doesn't seem to be a decent approach maybe
there must be a completely different route to take.

Stefan

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@ant.apache.org
For additional commands, e-mail: dev-help@ant.apache.org

Re: and useAppend and all that

[Stefan Bodewig <bo...@apache.org>]

On 2010-06-14, Stefan Bodewig wrote:

> And then there is bug 46805 which claims that even useAvailable=true
> doesn't fix a hanging process on Windows and suggests that we destroy
> the process instance and not close the process' streams explicitly at
> all.  I have no test cases for this bug and can't reproduce it myself.

> process.destroy makes Ant lose output from the forked process, so it
> doesn't seem a viable option.

> Not closing the streams doesn't seem to do any harm on Windows but I
> can't confirm that it helps in any way.  I did see some issues on the
> only Linux box I can reach right now, but that may be due to other
> problems (other exec stuff fails there as well).

Tested on a "real" machine now and not closing streams doesn't seem to
cause any problems on Linux either.  "doesn't seem" because there always
is a possibility of a race-condition that I just haven't hit.

Stefan

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@ant.apache.org
For additional commands, e-mail: dev-help@ant.apache.org