Hmm. First time I hear this being an issue. Seems indeed it's not being executed in %postun/%preun. May I ask how you ran dracut exactly?

(In reply to Stefan Dirsch from comment #1)
> May I ask how you ran dracut exactly?

dracut -f --regenerate-all

Related - on installation of NVIDIA KMP new initrd is generated which includes NVDIA modules. On removal of KMP initrd is not regenerated so we are left with initrd including NVIDIA drivers. Which can be quite surprising for users and it actually did cause issues. So probably package should simply call 

/usr/lib/module-init-tools/kernel-scriptlets/kmp-postun

which will also take care of depmod. I bet there are some magic RPM macros to take care of it.

(In reply to Andrei Borzenkov from comment #3)
> Related - on installation of NVIDIA KMP new initrd is generated which
> includes NVDIA modules. On removal of KMP initrd is not regenerated so we
> are left with initrd including NVIDIA drivers. Which can be quite surprising
> for users and it actually did cause issues. So probably package should
> simply call 
> 
> /usr/lib/module-init-tools/kernel-scriptlets/kmp-postun
> 
> which will also take care of depmod. I bet there are some magic RPM macros
> to take care of it.

Actually this is being done.

postuninstall scriptlet (using /bin/sh):
[...]
run_if_exists /usr/lib/module-init-tools/kernel-scriptlets/kmp-postun --name "nvidia-driver-G06-kmp-default" \
  --version "535.86.05_k6.4.3_1" --release "10.1" --kernelrelease "6.4.3-1" \
  --flavor "default" --usrmerged "01" "$@"

(In reply to Stefan Dirsch from comment #4)
>   --version "535.86.05_k6.4.3_1" --release "10.1" --kernelrelease "6.4.3-1" 

Kernel release is hardcoded in RPM and it is skipped for any other release. There is some code in %triggerpostun, but I honestly fail to understand what it does. Oh, and for G06 it is using the wrong package name - it says

triggerpostun scriptlet (using /bin/bash) -- nvidia-gfxG06-kmp-default

while package is named nvidia-driver-G06-kmp-default. May be this trigger is explicitly for migrating from old package naming convention, do not know.

Thinking more about it. On Tumbleweed NVIDIA driver is compiled and installed into kernel tree every time *kernel* is updated. What should happen when KMP is removed? Should all compiled binaries installed for every kernel version be removed? Should only the binary from the latest release be removed? I rather expect all compiled binaries should be removed. This is what e.g. Ubuntu nvidia-dkms package does.

(In reply to Andrei Borzenkov from comment #6)
> (In reply to Stefan Dirsch from comment #4)
> >   --version "535.86.05_k6.4.3_1" --release "10.1" --kernelrelease "6.4.3-1" 
> 
> Kernel release is hardcoded in RPM and it is skipped for any other release.
> There is some code in %triggerpostun, but I honestly fail to understand what
> it does. Oh, and for G06 it is using the wrong package name - it says
> 
> triggerpostun scriptlet (using /bin/bash) -- nvidia-gfxG06-kmp-default
> 
> while package is named nvidia-driver-G06-kmp-default. May be this trigger is
> explicitly for migrating from old package naming convention, do not know.

This was needed only for the migration from the old to the new package name.

(In reply to Andrei Borzenkov from comment #7)
> Thinking more about it. On Tumbleweed NVIDIA driver is compiled and
> installed into kernel tree every time *kernel* is updated. What should
> happen when KMP is removed? Should all compiled binaries installed for every
> kernel version be removed? Should only the binary from the latest release be
> removed? I rather expect all compiled binaries should be removed. This is
> what e.g. Ubuntu nvidia-dkms package does.

I tried to remove nvidia modules for no longer existing kernel trees, but this fails and I needed to revert it.

# rpm --triggers -qp nvidia-driver-G06-kmp-default-535.86.05_k6.4.3_1-10.1.x86_64.rpm
[...]
triggerpostun scriptlet (using /bin/sh) -- kernel-default
#
# Unfortunately doesn't work since kernel updates are not considered "atomar"
# when using YaST/zypper (only safe when using rpm) [boo#1182666]
#
#for dir in $(find /lib/modules  -mindepth 1 -maxdepth 1 -type d); do
#       if [ ! -d $dir/kernel ]; then
#               test -d $dir/updates && rm -f  $dir/updates/nvidia*.ko
#       fi
#done

Ok. Indeed I can reproduce this issue with current Tumbleweed easily.

(In reply to Stefan Dirsch from comment #9)
> (In reply to Andrei Borzenkov from comment #7)
> > Thinking more about it. On Tumbleweed NVIDIA driver is compiled and
> > installed into kernel tree every time *kernel* is updated. What should
> > happen when KMP is removed? Should all compiled binaries installed for every
> > kernel version be removed? Should only the binary from the latest release be
> > removed? I rather expect all compiled binaries should be removed. This is
> > what e.g. Ubuntu nvidia-dkms package does.
> 
> I tried to remove nvidia modules for no longer existing kernel trees, but
> this fails and I needed to revert it.
> 
> # rpm --triggers -qp
> nvidia-driver-G06-kmp-default-535.86.05_k6.4.3_1-10.1.x86_64.rpm
> [...]
> triggerpostun scriptlet (using /bin/sh) -- kernel-default
> #
> # Unfortunately doesn't work since kernel updates are not considered "atomar"
> # when using YaST/zypper (only safe when using rpm) [boo#1182666]
> #
> #for dir in $(find /lib/modules  -mindepth 1 -maxdepth 1 -type d); do
> #       if [ ! -d $dir/kernel ]; then
> #               test -d $dir/updates && rm -f  $dir/updates/nvidia*.ko
> #       fi
> #done

Actually I'm doing this when removing the package completely (%postun). For that case I added now a run of depmod for the affected kernel module trees.

This will get fixed with the following RPM changelog entry

-------------------------------------------------------------------
Wed Aug  2 12:23:27 UTC 2023 - Stefan Dirsch <sndirsch@suse.com>

- %postun: regenerate modules.dep, etc. to avoid dracut failures
  later (boo#1213765)

Closing ...