4 ms·

Using Pseudo-Terminals to Control Interactive Programs (2010)

49 points by robertelder 9y ago

liveoneggs 9y ago
a less featured expect for some reason
jevinskie 9y ago
The problems that are solved by PTYs finally "clicked" for me when I read their example of the problematic fseek/flushing. I hadn't seen such a cogent expiation until now.
- drudru11 9y ago
  True, but be warned, that fseek is not a real example. You cannot fseek/lseek on character devices. You would have to use a different technique.
Animats 9y ago
Controlling a human-interactive program with another program is pounding a screw, but it does work for simple programs.
- daveguy 9y ago
  ...or the interaction challenges that will give rise to significant advances in AI.
megapatch 9y ago
Expect?
- eltoozero 9y ago
  Yes.[0] [0]: https://en.m.wikipedia.org/wiki/Expect https://en.m.wikipedia.org/wiki/Expect
bch 9y ago
For those seeing "Expect" and not knowing what's going on: https://en.wikipedia.org/wiki/Expect https://en.wikipedia.org/wiki/Expect One of Tcl's first "killer apps". Don Libes' (creator of Expect) book "Exploring Expect" is also one of those technical masterpieces that are very enjoyable, entertaining technical reads. Recommended if only for the discussions of terminals and interactive i/o. [edit: typo/punctuation]
801699 9y ago
in case not want to install tcl https://github.com/drudru/pty4 https://github.com/drudru/pty4 http://empty.sourceforge.net http://empty.sourceforge.net
cyphar 9y ago
This is a bit of a tangent, but one issue with the POSIX (and underlying BSD/Linux) APIs is that they don't act nicely with mount namespacing (and thus containers). Since every /dev/pts mountpoint has a different set of ptys associated with it, you need to join a mount namespace in order to operate on /dev/pts/... (and the only way to get a slavefd is to open /dev/pts/$num). This is of particular concern to container runtimes, so I recently wrote a patch that added a new tty_ioctl(4)[1] to make things _slightly_ better (though I think the pty namespacing concept really needs to be reworked -- not to mention the countless lovely problems with Linux's namespace primitives). Effectively it allows you to safely open a slavefd given a masterfd (without needing to worry about mount namespace issues). Short story, if you're an author of a container runtime I would take a look at using that feature. [1]: https://marc.info/?l=linux-kernel&m=149649933504954&w=2 https://marc.info/?l=linux-kernel&m=149649933504954&w=2

49 points by robertelder 9y ago

a less featured expect for some reason

The problems that are solved by PTYs finally "clicked" for me when I read their example of the problematic fseek/flushing. I hadn't seen such a cogent expiation until now.

drudru11 9y ago
True, but be warned, that fseek is not a real example. You cannot fseek/lseek on character devices. You would have to use a different technique.

Animats 9y ago

Controlling a human-interactive program with another program is pounding a screw, but it does work for simple programs.

daveguy 9y ago
...or the interaction challenges that will give rise to significant advances in AI.

megapatch 9y ago

Expect?

eltoozero 9y ago
Yes.[0] [0]: https://en.m.wikipedia.org/wiki/Expect https://en.m.wikipedia.org/wiki/Expect

bch 9y ago

For those seeing "Expect" and not knowing what's going on: https://en.wikipedia.org/wiki/Expect https://en.wikipedia.org/wiki/Expect One of Tcl's first "killer apps". Don Libes' (creator of Expect) book "Exploring Expect" is also one of those technical masterpieces that are very enjoyable, entertaining technical reads. Recommended if only for the discussions of terminals and interactive i/o. [edit: typo/punctuation]

801699 9y ago

in case not want to install tcl https://github.com/drudru/pty4 https://github.com/drudru/pty4 http://empty.sourceforge.net http://empty.sourceforge.net

cyphar 9y ago

This is a bit of a tangent, but one issue with the POSIX (and underlying BSD/Linux) APIs is that they don't act nicely with mount namespacing (and thus containers). Since every /dev/pts mountpoint has a different set of ptys associated with it, you need to join a mount namespace in order to operate on /dev/pts/... (and the only way to get a slavefd is to open /dev/pts/$num). This is of particular concern to container runtimes, so I recently wrote a patch that added a new tty_ioctl(4)[1] to make things _slightly_ better (though I think the pty namespacing concept really needs to be reworked -- not to mention the countless lovely problems with Linux's namespace primitives). Effectively it allows you to safely open a slavefd given a masterfd (without needing to worry about mount namespace issues). Short story, if you're an author of a container runtime I would take a look at using that feature. [1]: https://marc.info/?l=linux-kernel&m=149649933504954&w=2 https://marc.info/?l=linux-kernel&m=149649933504954&w=2