The willingness of the kernel devs to refactor stuff is both a huge strength and weakness for the kernel. The strength is in the extraordinary continual optimization and improvement in the codebase, not just locally to an area of code but for cross-kernel concepts, like the recent workqueue changes.
But this has a pretty harsh cost for people writing or maintaining code that is outside the kernel tree and which therefore does not get the reworking applied to it as part of the core kernel. Whatever code they put out is invalidated and broken again and again sometimes in just the space of a few weeks.
The freedom to refactor despite breaking external code is a huge luxury for the devs seldom seen elsewhere in the coding world. Some projects take some care to allow compilation of their drivers for all recent kernels, using conditional compilation based on the kernel tree it is being compiled against, but other projects have an attitude that it will only compile against the current Linus tree.
The foaming churn of change makes for pretty hard work trying to make any kernel code that is not in the main tree work for any length of time. Greg KH at least is on record that his concept of the solution is to bring everything inside the kernel tree, but I don't know how that will ever scale, and it loads the devs with having to understand a work with an ever growing amount of device-specific code. Aside from that, it makes the kernel devs gatekeepers for what will be accepted, and since not everything that can exist will be deemed acceptable, there will always be a class of device driver that is living out of the comforts of the main tree.
Anyway the end result is that for many projects that people need drivers for, the shelf-life of any instance of the driver sources is extremely narrowly defined. A Wifi driver for example touches many subsections of the kernel that have a history of changes in the recent past, yet requires a pretty recent kernel to compile at all with the stuff that it actually needs. So each driver tree has a quite narrow slot of kernel versions that it will work with, annoyingly current CVS from many drivers will not compile against current kernel source, not -git either but -rc versions. It means that out of tree drivers are a lottery for any recent kernel any kind of driver is a high commitment project, that needs constant revisiting to keep it alive.
There doesn't seem to be an answer except that over time more and more critical subsystems in the kernel will surely mature to the point that they get fiddled with less and less, and things should therefore die down on the breakage front. But in truth the adolescent codebase of Linux shows no signs at the moment of slowing down its crazed foaming froth of reinvention and massive damage and breakage to the code around it.
