Bug#915381: Shouldn't nvidia-persistenced be daemonized (using Debian standard way to do so) ?
On 03/12/2018 11:58, Andreas Beckmann wrote:
> Patches welcome! (Or a pull request against the repo on salsa.)
Well, maybe even simpler than a patch:
sed 's/__USER__/nvpd/g' \
init/systemd/nvidia-persistenced.service.template \
> debian/nvidia-persistenced.service
sed 's/__USER__/nvpd/g' \
init/sysv/nvidia-persistenced.template \
> debian/nvidia-persistenced.init
cat << EOF > debian/nvidia-persistenced.postinst
#!/bin/sh
set -e
case "$1" in
'configure')
if ! getent passwd nvpd >/dev/null; then
# Create ad-hoc system user/group
adduser --system --group \
--home /var/run/nvpd/ \
--gecos 'nVidia Persistence daemon' \
--no-create-home \
nvpd
fi
;;
esac
#DEBHELPER#
EOF
PS: you picked my curiosity; I'll have a look at what "salsa" might be ;-)
> We will probably not "fix" this for stretch (but stretch-backports),
> unless someone gets this approved by the release team (but let's have
> this in sid first.)
Totally understand.
It'd be nice to have it in Buster though.
> And it should work with both systemd and legacy sysv.
Applied to latest (Unstable) 390.25 source package, this passes compilation
using pbuilder and produces a package with the ad-hoc /etc/init.d script
and /lib/systemd/system service file (debhelper >= 10 takes care of the
SysV/SystemD configuration; no need for dh-systemd build dependency).
And it installs and works OK on Stretch (SystemD).
I can't tell about SysV, though the nVidia init script looks simple
enough. Maybe an additional debian/nvidia-persistenced.default with a
"START={yes,no}" stanza might be needed (vs Debian Policy ?)
> Andreas
Thanks for your addressing this issue on such short notice.
Cédric
PS: In my effort to backport the latest upstream (410.78) drivers to Stretch
(we need CUDA 10 and serie 20** support), I personally used the (pre-compiled)
nvidia-persistenced available in the nVidia "*.run" package and included it
in my custom nvidia-graphics-driver source package
Bug#915381: Shouldn't nvidia-persistenced be daemonized (using Debian standard way to do so) ?
Control: tag -1 help On 2018-12-03 11:27, Cédric Dufour - Idiap Research Institute wrote: > Is there a reason the Debian package does not enable this binary as a > (sysv+systemd) daemon, along creating the ad-hoc system user (e.g. "nvpd") > for it ? Yes. Nobody has done it so far. :-) Maybe nobody actually used it until you came around. (I don't use it.) > If not, could this be looked into (I can provide input based on my own > re-packaging of the daemon) ? Patches welcome! (Or a pull request against the repo on salsa.) We will probably not "fix" this for stretch (but stretch-backports), unless someone gets this approved by the release team (but let's have this in sid first.) And it should work with both systemd and legacy sysv. Andreas
Bug#915381: Shouldn't nvidia-persistenced be daemonized (using Debian standard way to do so) ?
Package: nvidia-persistenced Version: 390.25-1 Hello, nVidia made it clear to us that nvidia-persistenced *really* should be running on (headless/no X server) CUDA nodes, because it does *more* than just keep the driver "up" (in particular, it fixes erroneous outputs from nvidia-smi; e.g. wrong GPU perf-state/load/memory accross multi-GPU setups). I see the current (Stretch) and latest (Unstable) Debian packaging of nvidia-persistenced does not enable the binary as a daemon (but merely provides example sysv/systemd/upstart scripts, in /usr/share/doc). Is there a reason the Debian package does not enable this binary as a (sysv+systemd) daemon, along creating the ad-hoc system user (e.g. "nvpd") for it ? If not, could this be looked into (I can provide input based on my own re-packaging of the daemon) ? Thanks for your work and best regards, Cédric -- Cédric Dufour @ Idiap Research Institute
