Review Request 129984: Prevent misdetection of EOF on Linux

[Peter Wu]

-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://git.reviewboard.kde.org/r/129984/
-----------------------------------------------------------

(Updated March 8, 2017, 11:52 a.m.)


Status
------

This change has been discarded.


Review request for Konsole and Kurt Hindenburg.


Bugs: 372991
    https://bugs.kde.org/show_bug.cgi?id=372991


Repository: kpty


Description
-------

On Linux, even if slave activity is detected on a master fd, FIONREAD
can still return zero available bytes in the buffer. Previously this
would prevent further signals from being emitted even if new IO arrived
(since the readNotifier is disabled) which in its turn prevented the
Konsole output from being updated.

A race condition in the kernel is suspected, the first known broken
kernel is v4.1.10-89-g5eb491ba5d06.

BUG: 372991


Diffs
-----

  src/kptydevice.cpp 22233a505d32d988795839cb592098c38f7fab33 

Diff: https://git.reviewboard.kde.org/r/129984/diff/


Testing
-------

Tested in conjunction with konsole v16.12.1-60-g151215a9 with this debug patch applied:
```
diff --git a/src/kptydevice.cpp b/src/kptydevice.cpp
index 92b443b..4dfcd02 100644
--- a/src/kptydevice.cpp
+++ b/src/kptydevice.cpp
@@ -273,6 +273,7 @@ struct KPtyDevicePrivate : public KPtyPrivate {
     KRingBuffer writeBuffer;
 };
 
+#include <cstdio>
 bool KPtyDevicePrivate::_k_canRead()
 {
     Q_Q(KPtyDevice);
@@ -310,6 +311,9 @@ bool KPtyDevicePrivate::_k_canRead()
         // select() will trigger again anyway if an EOF condition was found, and
         // only then we will accept it.
         if (!available && !maybeEof) {
+            FILE *fp = fopen("/tmp/EOF", "a");
+            fprintf(fp, "Maybe EOF!?\n");
+            fclose(fp);
             maybeEof = true;
             return true;
         } else if (available) {

```

Indeed, the /tmp/EOF file is being written and the Konsole output no longer "hangs".

Closing a tab (via EOF in bash or closing the tab) somehow does not trigger this debug print, perhaps the watcher is disabled elsewhere. (This is an observation and not a problem.)


Thanks,

Peter Wu

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.kde.org/pipermail/konsole-devel/attachments/20170308/a2b645b6/attachment-0001.html>