Tcl Source Code

I have some memory channel code that does:

puts $sendchannel some-data  ; # non-blocking
gets $readchannel myvar  ; # blocking, line buffered

This no longer works on macos.  the fileevent on the read handle
is not triggered, and the gets returns an empty string.  

If I put an 'update idletasks' inbetween the puts/gets, it works.

I have not been able to test on Linux as yet.

Edit 1: Changed subject, as this appears to be a problem on both macos and
windows.

An attempt to reproduce any sort of bug based on the provided description and
attached "memsock.test" was fruitless.  This ticket can be closed as invalid.

As was indicated by the original description of resolving the issue using
[update idletasks], what's needed is some event processing between the write to
one channel and the read from the other.  The attached file,
memsock_vwait.test, accomplishes that by entering an event loop at the top
level, scheduling a main coroutine to be run by that event loop, and then
yielding the coroutine at the key moment to allow some event processing.

Thank you!

No bug has yet been demonstrated, but I'm working now on a new test case. Perhaps it will expose something.

I used to port more things back to 8.6, but got pushback about making unnecessary changes to a patch release. I no longer want to invest my time in the 8.6 series, but instead focus on 8.7 and above. Other people who care more about 8.6 can port fixes if they consider it worth their time.

> There may indeed be an outstanding issue regarding writing to a nonblocking
> channel and then attempting to immediately read from it, but such an issue can
> be fixed far more gracefully than by making the revert in question.

Fully agreed. So, I suggest to fix this outstanding issue, and then put it in again.

Another remark: If this is a bug, it should have been applied to Tcl 8.6 as well. Why wasn't that done at the time? Then this outstanding issue would have been discovered much earlier.

For now (just before Tcl 8.7a5) I agree with reverting this. But - as soon as it is corrected - it should be re-applied, but then to both 8.6 and 8.7

Also, no test case was developed for this report, indicating that an expedient resolution was undertaken rather than a correct one.

The behaviour this report describes is not a bug. The provided test sets an event handler but then never enters the event loop:

fileevent $ch readable [list ::memsock::readhandler $ch $port $handler]

This behaviour is not supported. Ad event handler on a channel can not not be processed until the event loop is entered. The example may have worked in previous versions of Tcl, but only because of the bug described in [67a5eabbd3d19591] and [TCL_FILE_EVENTS cannot drive async I/O alone].

In the attached "memsock.test" example, "sendget" relies on this event handler noted above to ferry data between the sending channel and the receiving channel. This is an error. That handler should not be activated until the event loop is entered.

This "memsock.test" example attached to this report is in essence a restatment of [refchan, coroutine, and postevent from the "watch" proc], and that issue was already fixed. Unfortunately, in response to this new issue, that fix was reverted, effectively unfixing issue [67a5eabbd3d19591], and therefore unfixing the real bug illustreated by "memsock.test". The work that was reverted represents an important change in the direction of correcting some fundamental issues in Tcl's IO subsystem. That work should be reinstated immediately so that it makes it into the 8.7 release.

There may indeed be an outstanding issue regarding writing to a nonblocking channel and then attempting to immediately read from it, but such an issue can be fixed far more gracefully than by making the revert in question.

Fixed [a785b5bf5e11645d|here]

Proposed fix [3f466e55c8|here]

Thanks for the testcase! I did a bisect with the following result:

 12 BAD     2019-04-24 04:52:22 [9bcec7cd880540c3]
 13 GOOD    2019-04-23 06:50:02 [c4804bce46c2b672]

Conclusion: This [Bugfix] introduced the regression.

Assigning to pouryorick for further analysis.

Attached a non-minimal program that demonstrates the failure.

Confirmed that this fails on MacOS, Windows and Linux.