[Rxtx] to flush or not to flush, that is a question

[Trent Jarvi]

On Fri, 29 Sep 2006, Dr. Douglas Lyon wrote:

> Hi All,
> Let me see if I can summarize the discussion on flushing so far:
>
> 1. Manual Flush: A close is just a close vs.
> 2. Automatic Flush: A close is a flush and then a close.
>
> On the one hand we can define a close for a serial port
> so that it closes the serial port and has no other role. If
> data has been left in the buffer, it will be lost unless a flush
> is done explicitly.
>
> While it is the case that this would seem to speed up
> the close method, some might argue that the loss of data is
> not good software engineering, thus impacting system reliability.
> People will use:
> sp.flush();
> sp.close();
> as a serial port idiom, if close does not automatically flush.
>
> If flush does not precede close,
> the possible loss of data should be taken into
> account at the programmers' level, or it is a semantic error.
>
> Others argue that a slow serial line will delay the delivery of
> data for so long that the close will appear stalled.
>
> Corner-point: The emperor has no close.
>
> Suppose the serial device gets all stuffed up and data
> cannot be delivered. In such a case, the close will never
> finish with the flush and the serial port will never be released.
>
> In such a case, other programs (and users) who need the resource
> will be unable to access it.
>
> Digression: once I went to a wedding on a boat where the head
> was all stuffed up. Users were unable to flush and the
> OutputStream never closed... :(
>
> A serial port is a system wide resource for which many applications
> may contend. Once it is assigned to a process, in theory, no other
> application may usurp it.
>
> Wont deadlock from unreleased resources impact system reliability?
>
> So, on the one hand we adversely impact system reliability with
> increased chance of deadlock, or adversely impact system reliability
> with increased chance of data loss.
>
> To address the corner-point, the flush might need a time-out. Perhaps
> this can be computed as a function of the serial port data rate and the
> amount of data in the buffer.
>
> Please let me know if I am missing something.
>

output buffer empty events are flushes on Solaris and some other Unix 
Systems for all practical purposes.  Every write has a flush on those 
systems as that is the only means of knowing when to send the event.

Any Java API should abstract these details away from the programmer and 
behave the same on all systems.

--
Trent Jarvi
tjarvi at qbang.org