NVIDIA Developer Talks Openly About Linux Support
Q: If the NVIDIA Linux driver developers had a wishlist for changes to Linux and accompanying libraries/software, what would the top few items be?
I'm not sure. I suppose some of the responses to the next question could be applied here.
Q: Which part of Linux / X.Org is most troublesome?
The hardest thing about distributing a proprietary driver for Linux is to build a binary that will run across as many Linux distributions as possible. The challenges with this are:
1) The lack of a stable API in the Linux kernel. This is not a large obstacle for us, though: the kernel interface layer of the NVIDIA kernel module is distributed as source code, and compiled at install time for the version and configuration of the kernel in use. This requires occasional maintenance to update for new kernel interface changes, but generally is not too much work.
That said, the kernel API churn sometimes seems unfortunate: in some cases, working interfaces are broken or replaced with broken ones for no seemingly good reason. In some other cases, APIs that were previously available to us are rendered unusable.
2) The (recently, more quickly) changing ABI in the X.Org DDX. As in #1, this isn't a large obstacle for us: in recent driver branches, we can fairly easily build in support for multiple X server ABIs.
3) Being very careful about library and symbol dependencies in any of the binaries we distribute. The classic newbie mistakes here are things like:
a) Compiling/linking something on a new distro against a fairly recent version of glibc, and then trying to run that binary on a different distro, with a slightly older version of glibc. Classic errors are things like:
undefined reference to `[email protected]@GLIBC_2.3.4'
undefined reference to `__ctype_b'
b) Linking against the C++ runtime library (libstdc++) but then a different distro having a different version of the libstdc++.
libstdc++.so.6: cannot open shared object file: No such file or directory
We avoid these problems by a) explicitly linking against a very old glibc, and b) avoiding use of the C++ runtime. However, it requires careful attention.
#1 and #2 are our own fault for trying to produce a binary-only Linux driver (I consider this the price we pay for leveraging so much of the core kernel module from NVIDIA's cross-platform code base).
However, I've seen #3 bite many others just trying to provide an application on Linux. Many people eventually give up and just certify their commercial application for a small controlled list of Linux distributions that all have the same glibc and libstdc++ versions, or rebuild their application for different targeted distributions.
Providing a more consistent runtime environment for applications I think is an area for improvement in the Linux user space. To be fair, my experiences with the problems in #3 are a bit stale, so there might be less churn in DSO interfaces these days. And I should note that the actual Linux kernel ABI to user space is quite robust -- the point is mostly just about the DSOs that applications link against. I should also note that Ulrich Drepper's DSO How To  is an excellent resource for anyone trying to portably use Linux DSOs (in addition to those of us trying to produce portable Linux DSOs).