[Installation Instruction] nvidia driver with kernel-rt-2.6.20-0119.rt8 on FC6

[JaXXoN]

Hi!

For those of you Fedora Core 6 users who like to try out realtime preemption,
but are not willing to patch, compile and install a kernel from scratch, Ingo Molnar
is now maintaining approptiate kernel RPM packages. However, if you like to
use the nvidia linux graphics driver, with the most recent precompiled -rt kernel,
then you need to get around a few obstacles. Here is a step-by-step installation
instruction:


0. Make sure your installation does 3D acceleration with a standard kernel
(i.e. 2.6.19-1.2911.fc6)


1. Install the RT kernel:

wget http://people.redhat.com/mingo/realt...9.rt8.i686.rpm
wget http://people.redhat.com/mingo/realt...9.rt8.i686.rpm
rpm -i kernel-rt-2.6.20-0119.rt8.i686.rpm
rpm -i kernel-rt-devel-2.6.20-0119.rt8.i686.rpm


2. Modifiy "modpost" in order to accept non-GPL modules:

2a. Goto to the modpost directory

cd /usr/src/kernels/2.6.20-0119.rt8-i686/scripts/mod

2b. Delete the following two lines (1197-1198) in file modpost.c:

if (!mod->gpl_compatible)
   check_for_gpl_usage(exp->export, basename, exp->name);

2c. Recompile modpost

gcc -o modpost modpost.c file2alias.c sumversion.c

3. Unpack, patch and compile the nvidia driver module

cd $PATH_TO_NVIDIA_DRIVER_PACKAGE
wget "http://www.nvnews.net/vbulletin/attachment.php?attachmentid=23993&d=1171124108"
sh NVIDIA-Linux-x86-1.0-9746-pkg1.run --extract-only
cd NVIDIA-Linux-x86-1.0-9746-pkg1/usr/src/nv
patch -p1 < ../../../../patch-nv-1.0-9746_realtime-preempt.txt
make SYSSRC=/usr/src/kernels/2.6.20-0119.rt8-i686 module

Where $PATH_TO_NVIDIA_DRIVER_PACKAGE is the directory where you have
saved the nvidia driver package.


4. Manually define the missing symbol entry for "paravirt_ops"

PARAVIRT_OPS=`grep "D paravirt_ops" /boot/System.map-2.6.20-0119.rt8 | colrm 9`
ld -m elf_i386 --defsym paravirt_ops=0x$PARAVIRT_OPS -r -o nvidia.ko nvidia.o nvidia.mod.o

5. Install the nvidia kernel module

cp nvidia.ko /lib/modules/2.6.20-0119.rt8/kernel/drivers/video/
depmod -a 2.6.20-0119.rt8

Limitations
This instruction has only beed tested with 32-bit FC6. It may work for 64-bit,
but make sure to use "colrm 17" instead of "colrm 9" in step 4.


Feedback appreciated

Bernhard

[r.c.]

Does 32bit fc6 not use gcc 4.1.1, or did Ingo start compiling with the newer compiler? The reason I ask is that I just went through the process of installing the latest driver on the rt8 kernel on an x86_64 system.

As for my experience on x86_64, I did not have to do most of what you wrote, but I did have to recompile the kernel itself, since the driver warned of being compiled with a different version of gcc than what the kernel was compiled with.

I also had to change a couple of flags in nv.c, as described elsewhere on these boards, but I don't see that in your instructions. Since I'm not doing any virtualization or SELinux, I didn't have to do anything else.

[JaXXoN]

Originally Posted by r.c.

As for my experience on x86_64, I did not have to do most of what you wrote, but I did have to recompile the kernel itself,

Right, when compiling your own kernel, then you can easily disable
paravirt_ops and won't run into troubles with GPL-only symbols.
However, this post was intended to support Ingo's precompiled RPMs ...

Originally Posted by r.c.

since the driver warned of being compiled with a different version of gcc than what the kernel was compiled with.

Right, but i haven't experienced any stability issues so far. The potential
compiler incompatibilities mostly apply to C++, AFAIK. Otherwise
you would be in serious trouble with the nv-kernel.o object file that
comes with the driver, or do you know which compiler nvidia used for it?

When doing a "cat nv-kernel.o | strings | grep -i gcc", you will see a lot
of "GCC: (GNU) 2.95.3 20010315 (release)" so the driver core might have
been compiled with that stinking old version of gcc :-)

Originally Posted by r.c.

I also had to change a couple of flags in nv.c, as described elsewhere on these boards, but I don't see that in your instructions.

Right, the patch you can download with the wget command as described
above already includes the change from SA_* to IRQF_*.

regards

Bernhard

[puterguy]

This seems like a real kluge & PITA.

Any idea when/if nvidia might release *fixed* 9631 binaries?

[JaXXoN]

Hi!

Just to let you know, it is not anymore necessary to patch the nvidia
kernel driver for version 1.0-9755 along with kernel 2.6.20-0119.rt8!

The situation is that nvidia replaced the problematic semaphore
synchronization method by "completions" some times ago, and the
high latency causing "wbinvd" instruction (when PAT support is
enabled) has been replaced in 2.6.19 by a fast and interruptible
sequence of "clflush" instructions.

However, you may still need to manually define missing symbols
for pre-compiled kernels (see earlier post for an example).

regards

Bernhard

[SiVaS]

hallo,

howto use this with Kernel 2.6.21 ?

modpost.c looks different:

if (!mod->gpl_compatible)
check_for_gpl_usage(exp->export, basename, exp->name);
check_for_unused(exp->export, basename, exp->name); '

what shall i remove - all 3 lines ?

i tested one time and i get this error during compile :

scripts/mod/modpost.c:1172: Warning: »check_for_gpl_usage« defined, but not used

scripts/mod/modpost.c:1197: Warning: »check_for_unused« defined, but not used

Thanks


CooSee ' Ya

[crespowu]

Thanks for the installation instruction.It saved me a lot of time.


DVD to iPod

[nathan3011]

Originally Posted by crespowu

Thanks for the installation instruction.It saved me a lot of time.


DVD to iPod

Yeah i must say i agree with this poster, thank you so much for taking the time out of your day to post this for everyone :-) and i really aprreciate you doing Landlord Insurence for us.

Many Thanks