DISCLAIMER:I have no idea if what I have here is a good idea beyond the fact that my video card is working properly under linux-rt.  My computer froze a couple times with the following setup but I’m not sure if it’s the fix or the fact that linux-rt apparently doesn’t play well with multiple core machines. Or the computer freezy gnomes. Whatever. Use at your own risk.

I’m running Kubuntu 9.04 on AMD64.

While setting up my linux audio recording setup on Kubuntu 9.04 I installed the linux-rt meta-package (which right now points to linux-2.6.28-3-rt) for real time process scheduling.  I was dismayed and not at all surprised that the proprietary nVidia driver for my GeForce 7950  GT did not work immediately when I booted up out of GRUB with the new kernel.  So, I downloaded the latest nVidia driver from http://www.nvidia.com/object/linux_display_amd64_185.18.14.html and tried to install as I have to do every time there is a new linux kernel update by running this with the downloaded script:

$> sudo sh NVIDIA-Linux-x86_64-185.18.14-pkg2.run

The script informs you that it needs to compile the kernel module as usual, but dies saying compilation was unsuccessful. Following the advice of some forum posts I found, I extracted the script into its source tree with:

$> sudo sh NVIDIA-Linux-x86_64-185.18.14-pkg2.run -x

which creates a directory with a script nvidia-installer.sh and the source tree. I ran the installer manually with:

$> cd NVIDIA-Linux-x86_64-185.18.14-pkg2
$> sudo ./nvidia-installer


Compilation fails again at the same step. Inspecting /var/log/nvidia-installer.log tells us:

/home/labadorf/video_drivers/NVIDIA-Linux-x86_64-185.18.14-pkg2/usr/src/nv/os-interface.c:130: error: incompatible types in assignment

After poking around some forums there were some suggestions about replacing some SEMAPHORE code in certain driver source files, but they were for the wrong driver version (177) and didn’t work. Alright, well I have the source for both the kernel and driver, I thought, so let’s see what this incompatible type assignment is. At the end of the day, I traced the error to a couple lines of code earlier in usr/src/nv/os-interface.c. Apparently, the driver is trying to pass some raw_spinlock_t kernel struct to a function that can only accept spinlock_t structs when it detects you’re using a real time kernel. So, I thought a little commenting might do the trick…

Original os-interface.c source:


86 typedef struct os_sema_s
87 {
88 nv_stack_t *sp;
89 struct completion completion;
90 #if defined(CONFIG_PREEMPT_RT)
91 raw_spinlock_t lock;
92 #else
93 spinlock_t lock;
94 #endif
95 S032 count;
96 } os_sema_t;
97


Modified code:

86 typedef struct os_sema_s
87 {
88 nv_stack_t *sp;
89 struct completion completion;
90 //#if defined(CONFIG_PREEMPT_RT)
91 // raw_spinlock_t lock;
92 //#else
93 spinlock_t lock;
94 //#endif
95 S032 count;
96 } os_sema_t;
97


Lo and behold, on recompiling the driver everything works fine. Now, I suppose there may have been a good reason to use the raw version of the struct, but I got all of my audio stuff running with much lower latency AND using both monitors on high resolution (which is really handy since some of these audio apps take up a lot of screen real estate). There are a couple bugs open on this issue so I didn’t feel the need to post anything (though none of the comments there helped with my problem), hopefully someone will figure out what the problem is and fix it soon.

Note: By default the installer replaces whatever driver you have with the new one. You have to unpack the driver source and use the installer manually if you want to use different kernels and not recompile the driver every time. Look at this post and specifically Switcher’s comment 12 for instructions on how to do that.

Like I said, the commenting out I’ve done may be a terrible, terrible idea, but as I know nothing about graphics drivers and even less about kernel building I’m playing the ignorant fool. Good luck.