% tar -jx < *.bz

Écrit le dimanche 27 octobre 2024.

For $WORK I had to setup a Linux machine (a Raspberry Pi) so that right after boot, an interactive CLI tool would start, ready to receive input and show information.

Of course if you start a script at boot, its stdin is closed and stdout goes to the Journal. There exists openvt(1) to launch that script in a new VT (or to take over tty1). (But then you have to mind that openvt needs to run as root to open a new VT, so it should wrap sudo to drop privileges, and that get a bit fiddly.)

Then I happened upon the TTYPath setting, which does exactly what I need; and of course it combines perfectly fine with settings for User= or WorkingDirectory. I ended up with the following:

[Unit]
Description=Interactive script started on boot

[Install]
WantedBy=multi-user.target

[Service]
# The magic to have prompts displayed to the user
TTYPath=/dev/tty1
StandardInput=tty
StandardOutput=tty

User=someuser
WorkingDirectory=/opt/software/
ExecStart=/opt/software/bin/start_soft

# Always restart on exit, disable restart counter, delay a little bit
# the restart to avoid spam in case of repeating error
Restart=always
StartLimitIntervalSec=0
RestartSec=5s

It only occured to me to have a look at how getty.service is configured, and indeed it uses the same options:

StandardInput=tty
StandardOutput=tty
TTYPath=/dev/%I
TTYReset=yes          # reset $TTYPath before and after execution
TTYVHangup=yes        # disconnect all clients which have open $TTYPath
TTYVTDisallocate=yes  # deallocate before and after execution; this ensures screen and scrollback buffer is clear

A inconvenience is that if you start on tty1, then messages from systemd or the kernel get mixed up with your program output. If you start on tty2, I tried to add ExecStartPre=chvt 2, but the VT would switch back to tty1 depending on exact boot order.

Écrit le lundi 21 octobre 2024, modifié le dimanche 27 octobre 2024.

Par exercise (pour apprendre un peu de JavaScript), et un petit peu par utilité (devant résoudre des ambiguités autour de l'encodage de données au travail), j'ai créé une page pour explorer quelques jeux de charactères.

Écrit le vendredi 27 septembre 2024.

I want to import a file (maybe generated at run-time, maybe even the name only known at run-time) that does not exist in sys.path. In a couple of previous jobs that has happened with Python bindings generated from protocol descriptions. In the past I've done that by temporarily adding the parent to sys.path then running importlib.import_module, while being annoyed there is no function import_file(name: str, file_path: Path) -> ModuleType. It turns out that importlib.util has all the bits necessary, it's just a matter of putting them together:

import importlib.util
import os
import sys
from types import ModuleType

def import_file(module_name: str, file_path: os.PathLike) -> ModuleType:
    spec = importlib.util.spec_from_file_location(module_name, file_path)
    # spec: ModuleSpec(name='bar', loader=<_frozen_importlib_external.SourceFileLoader object at 0x7f30b1d10310>, origin='/tmp/tmp.WpcbUEkm32/foo.py')
    # if spec_from_file_location fails, it returns None rather than raise an exception
    if spec is None:
        raise ModuleNotFoundError(f"Can't import {module_name} from {file_path}")
    if spec.loader is None:
        raise ModuleNotFoundError(f"Can't import {module_name} from {file_path}")
    module = importlib.util.module_from_spec(spec)
    sys.modules[module_name] = module
    spec.loader.exec_module(module)
    return module

I suppose it is optional to add the module to sys.modules. If we do, that will allow future import module_name to find the same module (which might even be useful for the purpose of injecting different code into a 3rd party library…). If we don't, then we get a new copy of the module every time we import it again. Eg. if /tmp/tmp.WpcbUEkm32/foo.py defined a global:

MODULE_GOBAL = 1

def myname():
    return __name__

And I try to imort it twice under the same name, same path, the module globals are disconnected:

foo1 = import_file("foo", "/tmp/tmp.WpcbUEkm32/foo.py")
foo2 = import_file("foo", "/tmp/tmp.WpcbUEkm32/foo.py")

foo1.myname()
# 'foo'
foo2.myname()
# 'foo'

foo1.MODULE_GOBAL += 1
print(foo1.MODULE_GOBAL)
# 2, as expected
print(foo2.MODULE_GOBAL)
# 1, unmodified

Écrit le vendredi 9 août 2024, modifié le lundi 21 octobre 2024.

Python3 has bytes, which are sequences of 8-bit integers, and str, sequences of Unicode code-points. To go from one to the other, you need to encode or decode by giving an explicit encoding. There are many protocols where the markup is ASCII, even if the data is some other encoding that you don't know. If you know that other encoding is ASCII-compatible, it is useful to be able to parse, split etc. the markup, and you just need to pass-through the payload.

An initial search on the Internet brought up an article by Eric S. Raymond that touches on that, and suggests to decode the data as ISO-8859-1, handle it as str, then the payload can be losslessly recovered by reencoding it. The first 256 codepoints of Unicode are exactly the ISO-8859-1 codepoints (a bit more on that further down). As a result, the following is idempotent:

>>> by = bytes(range(0xff))
>>> by2 = bytes(str(by, encoding="iso-8859-1"), encoding="iso-8859-1")
>>> by == by2
True

A few days later I came across that article by Alyssa Coghlan, which mentions the same idea, and also the existence of the errors="surrogateencoding" (PEP 383) error handler (also: codecs in the Python documentation), which is designed to allow exactly what I needed:

>>> by = bytes(range(0xff))
>>> by2 = bytes(str(by, encoding="ascii", errors="surrogateescape"), encoding="ascii", errors="surrogateescape")
>>> by == by2
True

Alyssa Coghlan has some discussion about the merits of each approach, I can't really say that functionally they have any meaningful difference. As she points out, if you do anything but re-encode the non-ASCII compatible parts using the same codec, you risk getting Mobijake (or if you're lucky, a UnicodeDecodeError rather than silently producing garbage).

Performance-wise, let's see:

>>> import timeit
>>> timeit.timeit("""bytes(str(by, encoding="ISO-8859-1"), encoding="ISO-8859-1")""", setup="import random; by = random.randbytes(10_000)")
0.8885893229962676
>>> timeit.timeit("""bytes(str(by, encoding="ascii", errors="surrogateescape"), encoding="ascii", errors="surrogateescape")""", setup="import random; by = random.randbytes(10_000)")
125.00223343299876

That's… a very large difference. ESR's article points out that ISO-8859-1 has some properties that make it some efficient (it maps bytes 0x80—0xff to Unicode code-points of the same numeric value, so there is no translation cost, and the in-memory representation is more efficient). Trying increasing sizes:

>>> for size in 10, 100, 1_000, 2_000, 5_000, 10_000, 20_000:
...     by = random.randbytes(size)
...     duration = timeit.timeit("""bytes(str(by, encoding="ascii", errors="surrogateescape"), encoding="ascii", errors="surrogateescape")""", globals={"by": by}, number=100_000)
...     print(f"{size}\t{duration}")
... 
10	0.0650910490003298
100	0.1047916579991579
1000	0.5472217770002317
2000	1.5103355319952243
5000	5.779067411000142
10000	12.497241530996689
20000	25.78209423399676

That seems to grow faster than O(n); the duration seems to be ×3 when the size goes ×2, is it growing like O(n^1.5)?

In contrast, using the ISO-8859-1 method seems to having a complexity O(n):

>>> for size in 10, 100, 1_000, 2_000, 5_000, 10_000, 20_000, 50_000, 100_000:
...     by = random.randbytes(size)
...     duration = timeit.timeit("""bytes(str(by, encoding="iso-8859-1"), encoding="iso-8859-1")""", globals={"by": by}, number=100_000)
...     print(f"{size}\t{duration}")
... 
10	0.05453772499458864
100	0.037617702000716235
1000	0.05454556500626495
2000	0.05654650100041181
5000	0.06352802200126462
10000	0.0898260960020707
20000	0.23981017799815163
50000	0.4997737009980483
100000	0.9646763860000647

By design of Unicode, the first 256 codepoints are the same as those of ISO 8859-1, and the conversion between Unicode and ISO 8859-1 is implemented in C in CPython. That's why the technique describes above works losslessly, and is so much faster than any mapping to surrogate escapes (even if I still don't understand how that latter approach is so slow).

I recently learned (between $WORK making me investigate character encoding issues, and making a character table explorer as a JavaScript learning exercise) about the C1 controls characters. Wikipedia has a write-up, in short in 1973, during their work on 8-bit encodings for languages beyond English, ECMA and ISO came up with an extra series of control characters in the 0x80--0x9F range, which would allow things like terminal control (the 0x9B Control Sequence Introducer for ANSI escape sequences), switching to a different encoding (eg. SS2/SS2, used by EUC-JP), etc. However, it would still be necessary to represent those in a 7-bit enviroment; so each of those could also be written as 0x1B ␛ followed by an ASCII char (eg. Control Sequence Introducer could be spelled 0x1B [). Both of the following print HI in reverse video with xfce4-terminal and libvte-2.91 v0.70.6, but the second example doesn't work in XTerm version 379. I guess XTerm doesn't (or dropped) support for the C1 controls spelling of the ANSI escape sequences.

$ printf "\\u001b[7m HI \\u001b[m Normal\n"
HI  Normal
$ printf "\\u009B7m HI \\u009Bm Normal\n"
HI  Normal

Anyway, those C1 controls were almost never used, but are reserved in the ISO 8859-X encodings. Windows-1252, presumably noticing that lack of use, assigned printable glyphs at those positions. HTML5 aliases ISO 8859-1 to Windows-1252, I guess because it wouldn't make sense to use control charactes in an HTML document, so it those appear, that must be because the author actually meant Windows-1252 (and if they don't, then ISO 8859-1 and Windows-1252 are identical outside the range of C1 controls.

Écrit le lundi 1 avril 2024.

Trying a new Today I learned section… Today, I learned about set -u, which will fail a bash script if we try to use a variable that's not defined. No more deleting someone's files because a variable was not defined. Demo:

#!/bin/bash
set -u

GREETING=Hello

#oups I misspelt that
echo ${GREETINGS} world

If you run this:

$ ./script
./script: line 7: GREETINGS: unbound variable

That makes testing whether a variable is set slightly more complicated:

set -u
if [[ -z ${GREETINGS} ]]; then
    # oups we never get here, the previous line fails with "unbound variable"
    GREETINGS=$( ....compute greetings ...)
fi

Instead:

set -u
if [[ ! -v GREETINGS ]]; then
    GREETINGS=$( ... compute greetings ... )
fi

Where -v is a Bash Conditional Expression described thusly:

-v varname

True if the shell variable varname is set (has been assigned a value).

I suppose the actually equivalent test would be: \[\[ ! -v GREETINGS ]] || \[\[ -z ${GREETINGS} ]], to check whether the variable is not only set but also non-null.

I've always used set -e and set -o pipefail in shell scripts. The first will exit the entire script if any command (more specifically: any pipeline fails). However by default, quoting the manual: the "exit status of a pipeline is the exit status of the last command in the pipeline", meaning the following succeeds:

$ false | true
$ echo $?
0

But, if "pipefail is enabled, the pipeline's return status is the value of the last (rightmost) command to exit with a non-zero status, or zero if all commands exit successfully", so:

$ set -o pipefail
$ false | true
$ echo $?
1

TL;DR: Use set -euo pipefail at the top of your bash scripts. Julia Evans agrees.

Écrit le samedi 30 mars 2024.

RPi.GPIO is a Python module for controlling GPIO pins on Raspberry Pi. It is old and little maintained (they still use sourceforge.net!), but still useful today. It doesn't have type hints, so I added stubs for RPi.GPIO, now available on PyPI.

Incidentally I learned of distribution name normalisation. Package names can use ASCII letters, numbers, and ., - and _; but the normalised name is lower-cased with any run of punctuation characters replaced with a single dash, so rpi-gpio is the normal form for RPi.GPIO. pip the CLI tool and the PyPI web interface accept any form: pip install RPi.GPIO, or pip install rpi-gpio, or pip install rPi__GpIo will install the same package. pip-compile with use the normalized name when writing a constraints file.

Écrit le samedi 15 décembre 2012.

J'ai voulu avoir de l'IPv6 à la maison. Mon opérateur Internet ne fournit que de l'IPv4, il faut donc tunneller d'une façon ou d'une autre. J'ai une Kimsufi à disposition, avec un accès IPv6 natif, j'ai donc voulu l'utiliser comme point de sortie.

Bien sûr, une solution est d'utiliser un fournisseur de tunnel comme HE ou SixXS, mais où est l'intérêt de ne pas faire les choses soi-même ?

Écrit le samedi 2 juin 2012, modifié le lundi 11 mars 2024.

Avec FreeBSD 9.0 qui est sorti il y a presque six mois, il est grand temps de se mettre à jour. On ne peut pas à la fois prétendre au status de cyborg, et utiliser autre chose que la dernière version de son OS préféré !

Dans un épisode précédent on avait installé son serveur avec ZFS, justement pour pouvoir faire des mises à jour avec une solution de repli si les choses tournent mal. On a deux solutions : faire un snapshot puis mettre à jour, avec la possibilité de faire un rollback plus tard ; ou alors installer dans un nouveau volume et pointer le bootloader sur le nouvel environnement. J'ai exploré la deuxième option. Cette deuxième option permet de ne pas tout perdre s'il faut revenir en arrière. On verra dans la suite si cela devait se révéler utile…

Écrit le dimanche 30 octobre 2011.

EuroBSDcon 2011 s'est déroulé au début de ce mois d'octobre. Je viens de finir de mettre en forme mes notes sur les conférences que j'ai suivies.

On y retrouvera en particulier :

• sendmail, history and design, présenté par Eric Allman, dans lequel l'auteur présente l'histoire du MTA qui est encore aujourd'hui le plus utilisé sur Internet, les éléments de design qui ont permis à sendmail de survivre pendant 30 ans, et les élèments malgré lesquels sendmail a survécu pendant 30 ans ;
• Minix 3, ou comment construire un OS fiable, par Herbert Bos : en utilisant des principes systématiques de réduction des privilèges, de séparation des composants, et en acceptant que les composants peuvent crasher mais sont pour la plupart redémarrables ;
• status d'IPv6 dans FreeBSD de Bjoern Zeeb, comment FreeBSD a vécu le World IPv6 Day cette année, les leçons du noyau AF_INET6-only, et ce qui doit arriver dans le futur proche dans un FreeBSD près de chez vous ;
• portage d'OpenBSD vers UltraSPARC T1 et T2 de Mark Kettenis, une session très technique (au cours de laquelle je ne prétendrais pas avoir tout compris), qui décrit dans une première partie les atouts de cette plateforme, et comment OpenBSD est capable de tourner sur cette architecture ;
• NPF, un nouveau pf pour NetBSD 6.0 de Zoltan Arnold Nagy, qui explique pourquoi un nouveau pare-feu est nécessaire et comment NPF a été construit ;
• High Availability STorage in FreeBSD de Pawel Jakub Dawidek, du RAID1 over network, modes de synchronisation, démonstration, performance ;
• testing automatique de NetBSD de Martin Husemann, les composants logiciels mis en œuvre, les bindings dans différents languages, les bugs découverts grâce au testing systématique.

À noter également les slides de quelques développeurs OpenBSD :

• What's new in OpenSSH, de Damien Miller ;
• OpenBSD's New Suspend and Resume Framework (slides), de Paul Irofti ;
• OpenBSD SCSI Evolution de David Gwynne et Ken Westerback ;
• 10 years of pf, de Ryan McBride et Henning Brauer :
• OpenBSD Update, du même Henning Brauer.

Malheureusement les slides de Claudio Jeker sur MPLS ne sont pas en ligne.

Michael Dexter a parlé de virtualisation dans le monde BSD (plus de documents). Son exposé avait deux grands axes : une classification des techniques de virtualisation existante, et surtout une présentation de BHyVe, un hyperviseur de type 2 (à la Xen) pour FreeBSD.

Écrit le samedi 27 août 2011, modifié le dimanche 18 septembre 2011.

Comme ZFS c'est beau, ça rend le poil soyeux et ça permet à l'admin de diminuer sa consommation d'aspirine, j'ai profité de mon changement de Kimsufi pour l'installer en tout ZFS, depuis le boot. Bien sûr, une solution aurait été de laisser l'installation sur / et de ne mettre que les données sur un volume ZFS, mais c'est trop facile, ça. Et puis on garde de l'UFS, c'est so-XX^ième siècle.

Pour installer en tout ZFS, il doit y avoir au moins trois solutions pour faire cette installation :

• partir du système existant et installé, créer un volume, faire un cp de notre / UFS vers ce volume ZFS, pointer le chargeur de boot vers ce volume ;
• booter depuis le réseau (OVH permet de booter sur un FreeBSD 8.2, yay), et faire l'installation comme cela ;
• booter sur vKVM depuis une ISO.

On est d'accord que le premier, c'est tricher. Chez moi le vKVM perdait régulièrement la connexion, alors j'ai utilisé le boot sur le réseau (le mode rescue-pro).

Let's go.