bug#59784: [version 1.4.0rc1] Retrying a failed install fails

[Ludovic Courtès]

Hi,

"pelzflorian (Florian Pelz)" <pelzflorian@pelzflorian.de> skribis:

> Desperately I tried also adding fsync, to no avail, both issues remain.
> Non-working patch attached.
>
> Maybe dynamic-wind is an inappropriate pattern here?
>
> If I interrupt installation using Ctrl-C (which I normally don’t,
> instead I unplug Ethernet), then I have to press Ctrl-C twice.  Maybe
> that could be related to why I need to resume twice?

One finding: when hitting C-c, the dynamic-wind exit handler (the one
that restores the database and umounts the cow store) is *not* executed.

This is because ‘call-with-mnt-container’ sets a SIGINT handler that
terminates that process with SIGKILL (I’m not entirely sure of the
rationale, but said process cannot handle signals in Scheme while it’s
in ‘waitpid’, called from ‘run-command’).

I did reproduce the issue in a VM by running “ifconfig ens3 down” in a
tty, or by killing the ‘guix substitute’ process, to cause failure of
‘guix system init’.  In that case the database is indeed restored, but I
occasionally get errors like “/gnu/store/….drv: No such file or
directory”.

Ludo’.