2016-09-14 18:01 GMT+02:00 Richard Shaw <hobbes1069(a)gmail.com&gt;: > > I've had a few failures but not really sure as to the cause, but > I've been > toying around with doing the final RPM install with a systemd > service > instead of calling dnf directly with the kernel posttrans script.

> Thoughts? Pitfalls? I really don't understand the method here.

Can you reproduce any issue ? or do you have received informal informations that the posttrans script wasn't working for a reason ? Because for now It's working fine with me all the time, so there is probably an unknown corner case here.

For some known ones, the major reason why the kernel posttrans script might fails is because it's missing the matching kernel-devel for the new installed kernel. Please see: https://bugzilla.rpmfusion.org/show_bug.cgi?id=3386 (please note that while its reported against the akmods maintainer, anyone can take this and implement, not only Richard or me). In others word, you need to force to install the matching kernel- devel along with the kernel update (dnf update kernel will break).

Looking at the current akmods.service, can you please explain why: - It's using before display-manager.service (that's not the point with nvidia here, if the nvidia driver isn't here, we should better implement a fallback to anything else) - RemainAfterExit=yes, Can you point me why this was added ? akmods shouldn't remains once it has finished to build. - It has "ExecStart=-/usr/sbin/modprobe nvidia", this is not the way the nvidia driver should be loaded, it's loaded by the DDX driver (well not that true with modern nvidia driver,but at least it shoudn't be loaded that way). I think having more into systemd service might help (you can trigger a start akmods.service from posttrans), I don't think having yet another akmods script will help. Thx

On Wed, Sep 14, 2016 at 12:12 PM, Nicolas Chauvet <kwizart(a)gmail.com&gt; wrote: > 2016-09-14 18:01 GMT+02:00 Richard Shaw <hobbes1069(a)gmail.com&gt;: > > I've had a few failures but not really sure as to the cause, but > I've been > > toying around with doing the final RPM install with a systemd > service > > instead of calling dnf directly with the kernel posttrans script. > > > > Thoughts? Pitfalls? > I really don't understand the method here. > Can you reproduce any issue ? or do you have received informal > informations that the posttrans script wasn't working for a reason > ? > Because for now It's working fine with me all the time, so there is > probably an unknown corner case here. Reproduce, no, but I've had a couple of occasions where I was not able to get to graphical boot. I just read about the need to run depmod on a Fedora thread. I didn't think about that so I manually erased the RPMs and let akmod rebuild them which fixed the problem. 

  > > For some known ones, the major reason why the kernel posttrans script > > might fails is because it's missing the matching kernel-devel for the > > new installed kernel. Please see: >

> > while its reported against the akmods maintainer, anyone can take this > > and implement, not only Richard or me). > > In others word, you need to force to install the matching kernel-devel > > along with the kernel update (dnf update kernel will break).> > I'm certainly not missing kernel-devel but I have had intermittent issues.  > > > Looking at the current akmods.service, can you please explain why: > > - It's using before display-manager.service (that's not the point with > > nvidia here, if the nvidia driver isn't here, we should better > > implement a fallback to anything else)> > Are you saying it should be dependent on an earlier target? I chose display-manager since nvidia is my primary use. >   > > - RemainAfterExit=yes, Can you point me why this was added ? akmods > > shouldn't remains once it has finished to build.> > Since it runs and exits (not a daemon) this option keeps systemd from thinking it failed. >   > > - It has "ExecStart=-/usr/sbin/modprobe nvidia", this is not the way > > the nvidia driver should be loaded, it's loaded by the DDX driver > > (well not that true with modern nvidia driver,but at least it shoudn't > > be loaded that way).> > I can't remember why I added that but if must have fixed a problem I was having at the time.  > > > I think having more into systemd service might help (you can trigger a > > start akmods.service from posttrans), I don't think having yet another > > akmods script will help.> > No new script, just alter the akmodsposttrans to call "systemctl start akmods-install" or whatever it will be called instead of calling akmods directly.  > Thanks, Richard

2016-09-14 20:43 GMT+02:00 Sérgio Basto <sergio(a)serjux.com&gt;: > > On Qua, 2016-09-14 at 13:16 -0500, Richard Shaw wrote: > > On Wed, Sep 14, 2016 at 12:12 PM, Nicolas Chauvet <kwizart(a)gmail.co > m> wrote: > > 2016-09-14 18:01 GMT+02:00 Richard Shaw <hobbes1069(a)gmail.com&gt;: > > > > I've had a few failures but not really sure as to the cause, but > > I've been > > toying around with doing the final RPM install with a systemd > > service > > instead of calling dnf directly with the kernel posttrans script. > > > > Thoughts? Pitfalls? > I really don't understand the method here. > Can you reproduce any issue ? or do you have received informal > informations that the posttrans script wasn't working for a reason > ? > Because for now It's working fine with me all the time, so there is > probably an unknown corner case here. > > > Reproduce, no, but I've had a couple of occasions where I was not > able to > get to graphical boot. I just read about the need to run depmod on > a Fedora > thread. I didn't think about that so I manually erased the RPMs and > let > akmod rebuild them which fixed the problem. > > > yes , after install kmod you should load the kmods modules , it may > work > without it , because X load it but we should ensure modprobe > nvidia.ko (not > depmod .. ) The documentation has always said that you need to reboot once the driver is installed/updated (and immediately after). 1/ Because that's easier than to restart any process that have the wrong (mesa) libGL. 2/ Because nvidia driver use some kind of contiguous memory that might not be available after the boot. 3/ Because it relies on certains kernel boot options that was set by the package, so you need to reboot anyway. > > for example VirtualBox instructions: > >     dnf install VirtualBox kernel-devel-$(uname -r) akmod- > VirtualBox >     akmods >     systemctl restart systemd-modules-load.service It could be possible to use a trigger on systemd-modules-load.service after kmod-foo will install. At least it would be a generic way to solve the problem when a given kmod drops a /usr/lib/modules-load.d file This should be handled by kmodtool at least. But I would prefer that a reboot is documented (the less is done on user's back, the better).

On Wed, Sep 14, 2016 at 12:12 PM, Nicolas Chauvet <kwizart(a)gmail.com&gt; wrote: > > 2016-09-14 18:01 GMT+02:00 Richard Shaw <hobbes1069(a)gmail.com&gt;: > > I've had a few failures but not really sure as to the cause, but I've > > been > > toying around with doing the final RPM install with a systemd service > > instead of calling dnf directly with the kernel posttrans script. > > > > Thoughts? Pitfalls? > I really don't understand the method here. > Can you reproduce any issue ? or do you have received informal > informations that the posttrans script wasn't working for a reason ? > Because for now It's working fine with me all the time, so there is > probably an unknown corner case here. Reproduce, no, but I've had a couple of occasions where I was not able to get to graphical boot. I just read about the need to run depmod on a Fedora thread. I didn't think about that so I manually erased the RPMs and let akmod rebuild them which fixed the problem.

On Qua, 2016-09-14 at 13:16 -0500, Richard Shaw wrote: > No new script, just alter the akmodsposttrans to call "systemctl > start akmods-install" or whatever it will be called instead of > calling akmods directly. I don't deslike the idea , what you propose ?