This is the mail archive of the
mailing list for the Cygwin project.
Re: cygwin source-patch fixing deadlock while writing to serial port
- From: "H. Henning Schmidt" <henning at hhschmidt dot de>
- To: cygwin at cygwin dot com
- Date: Tue, 20 Jan 2004 00:16:50 +0100
- Subject: Re: cygwin source-patch fixing deadlock while writing to serial port
> On Mon, Jan 19, 2004 at 10:08:39PM +0100, H. Henning Schmidt wrote:
> >I found a potential deadlock while writing to a serial port (e.g.
> >/dev/com1) that has been opened as O_RDWR. The deadlock occurs from time
> >to time (not sure about exact conditions) when I write to that port,
> >while there is data coming in (e.g. from an external device) and I do
> >not read away that data fast enough from the port.
> >I did provide a test case a while ago in
> >http://sources.redhat.com/ml/cygwin/2003-03/msg01529.html. I digged into
> >the issue some more now and found that the executing thread got
> >sometimes deadlocked in fhandler_serial::raw_write(). It basically ends
> >up in a for(;;) loop and just never hits the break;
Exactly. When the input buffer overflows, all serial communications cease
and calls exit with ERROR_OPERATION_ABORTED. If you only call write, then
the ClearCommError() necessary to start things up again is never called,
and you stick in that infinite loop.
Ok, I will be happy to learn what the proper fix looks like in your opinion.
However, I have a hard time accepting the fact that I cannot write OUT the
serial port because my INput buffer overflows. At least when I have switched
off all kinds of flow control (which I have), then I consider these to be two
independent streams that happen to share one filedesc. Or should I perhaps
open two independent filedescriptors in this case, one write-only, one read-only?
Are you saying that the underlying Win-API enforces what you state above, or does
this really make sense in some way and I just didn't get it yet? Your statement
sounds to me like I am forced to keep a second thread around, and if only to get
me out of that trap, once I get hung there ... does not sound too elegant to me ...
> >The applied patch adds a safety exit to that for(;;) loop.
> >This fixes the testcase referenced above.
Yuck! No, this is not the proper fix.
> >This might not be the last problem lingering in the serial access code
> >(there are some FIXME tokens still around ...), but it is definitely an
> >improvement for me. I thought I'd share that with you.
> Can you convince me that this isn't just a band-aid? I don't understand
> why cygwin *shouldn't* hang in a situation like this. There are
> certainly similar situations where this happens on linux.
> Perhaps we need a low_priority_sleep (10) in the loop in that situation
> or something.
No. I have a partial patch for the above, but I am in the process of
getting a new Windows box and shuffling all my data. I'll try to submit
it when things settle if no one beats me to it.
who am I to beat you ...
but be assured that I'll be waiting eagerly for this patch.
This issue has been kicking me for a long time (letting my app hang up every once in
a while), and I really want get this fixed. So now that I know that there is a better
way to do it ... that's what I'm longing for then :-)
Unsubscribe info: http://cygwin.com/ml/#unsubscribe-simple
Problem reports: http://cygwin.com/problems.html