On Tue, 25 May 2010, Nick Piggin wrote: > OK, you still never justified why that change is needed, or why it > is even a cleanup at all. You need actually a *good* reason to change > the user kernel API. A slight difference in opinion of what the sysctls > should be, or a slight change in implementation in the kernel, is not > a good reason in the slightest. > My personal opinion (and it's no longer what I'm advocating since I'm hoping that the much larger importance of this patchset can now be realized without this sideshow) is that if we have the opportunity to consolidate two virtually unused sysctls into one to cleanup procfs and make extending the oom killer easier in the future without the need to add additional sysctls for systems that cannot afford a tasklist scan that we should do it. I didn't think anyone was using them (nobody can be cited as using them) and I felt the two year deprecation period was enough to be able to cleanup the ever-growing list of VM sysctls as well as making it easier to extend in the future by adding a sysctl that had semantics specifically directed to its target audience. > The point about not many people using the parameters I don't think is > a good one. 2.6.32 is being used in the next enterprise kernels so they > are going to be in production for 5 or more years. How many people will > have written scripts by the time they upgrade? > If nothing else, this solidifies in my mind the notion that once a VM sysctl is added, it can never be removed. I regret adding oom_kill_allocating_task at the mere suggestion of SGI when I made cpuset ooms do a tasklist scan without developing a more extendable interface for those large systems that we could piggyback on later for the same reasons (we'd now need to disable oom_dump_tasks once it's the default for those systems). I felt the deprecation and removal would eventually be sufficient, but I'm no longer going to push it because it's sidetracking a much larger and seperate effort. > > I've made that change in my latest patch series which I'll be posting > > shortly. > > The other thing is that it makes perfect sense to put controversial > changes on hold even if you still think you can make a case for them. I personally like consistency even amongst architectures when it comes to the semantics of sysctls, so the pagefault oom handling changes were merely for compatibility until all architectures were converted to using the oom killer. You've done that work, so I no longer have a problem with panicking when the pagefault handler is called. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxxx For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>