| 1 |
* System.Console.Regions handling of outputConcurrent and errorConcurrent |
| 2 |
when the output lacks a newline is surprising. The output may never be |
| 3 |
displayed. A better approach: |
| 4 |
|
| 5 |
When output arrives that lacks a newline, allocate a region for it |
| 6 |
and buffer it in that region, so it displays immediately and more |
| 7 |
output will go to the same region. When a newline arrives, close the |
| 8 |
region and scroll up the completed line the same way any completed region |
| 9 |
is displayed. |
| 10 |
|
| 11 |
* Parts of System.Console.Concurrent may not be async exception safe. |
| 12 |
|
| 13 |
If a thread is running an action from that module and an async exception |
| 14 |
is sent to it, it may result in a deadlock or other problem. |
| 15 |
|
| 16 |
(System.Console.Regions has been made safe, |
| 17 |
also outputConcurrent and errorConcurrent.) |
| 18 |
|
| 19 |
Particularly problematic is takeOutputLock', which takes the lock |
| 20 |
and then outputs buffers to the console. If the emitOutputBuffer |
| 21 |
calls are interrupted by async exception, it will be left locked |
| 22 |
and the buffered output is also lost. But masking them is not good |
| 23 |
because emitOutputBuffer could run for a long time in some situations. |
| 24 |
|
| 25 |
fgProcess and bgProcess also do stuff with registerOutputThread/unregisterOutputThread |
| 26 |
that may not be async exception safe. |
| 27 |
|
| 28 |
And createProcessForeground uses takeOutputLock but then calls fgProcess, |
| 29 |
which could be interrupted (during its call to registerOutputThread) |
| 30 |
before it starts the async thread that drops the lock. |
| 31 |
|
| 32 |
One approach to all this might be to fork off a worker thread, |
| 33 |
which will thus be immune to any async exception directed at the calling |
| 34 |
thread. |
| 35 |
|
| 36 |
* Calling setConsoleRegion with something that throws an error |
| 37 |
will cause no further display updates to happen. |
| 38 |
|
| 39 |
The exception in the thunk will crash the displayThread, |
| 40 |
and that is not waited on until the action passed to displayConsoleRegions |
| 41 |
finishes, so the exception is deferred to that point. Of course if the |
| 42 |
action never finishes, that hides the exception. |
| 43 |
|
| 44 |
It does seem this should be improved, by catching the crashing |
| 45 |
displayThread and ideally propigating the exception immediately. |
| 46 |
But displayConsoleRegions would need to use a different class than |
| 47 |
the current MonadIO to be able to do that. Eg, MonadBaseControl |
| 48 |
so lifted-async can be used, or a class that lets the action be |
| 49 |
canceled. |
| 50 |
|
| 51 |
Or it could be dealt with by having setConsoleRegion force the thunk |
| 52 |
before passing it over to the displayThread. But the instance |
| 53 |
ToRegionContent (STM Text) seems to prevent doing that. |