David Airlie had sent in a pull request today for the Linux 3.2 kernel that addresses an Intel VT-d issue with Ironlake graphics, a kexec fix where the GPU write-back engines need to be switched off, and then a trivial Radeon KMS fix. The kexec fix sent Linus over the edge.
In the past Linus Torvalds has been vocal about select Direct Rendering Manager (DRM) pull requests from David Airlie with calling the GEM memory management patches at first being untested crap, the lack of Nouveau DRM in the Linux kernel, questioning what's wrong with the whole DRM crowd, and showing why DRM has been problematic (late pull request), among other remarks.
Below is the response from Linus Torvalds for Airlie's pull request.
Quite frankly, I think it's too late for something like a kexec bugfix. Nobody cares. So kexec doesn't work - that's not something new. This doesn't smell like a regression to me. And the kcalloc things you mention *sound* like some kind of cleanup crap.
The DRM layer has been fairly good for a few releases, but I'm getting the feeling that I need to start pushing back, because I'm getting stuff that I don't think matters, and shouldn't be sent to me after -rc4.
So I'm not pulling this.
What the heck is up?
By now, I want fixes that either fix real regressions that people *care* about, or that help new unreleased hardware that people *will* care about and that cannot possibly mess up old users.
kexec? Who the f*ck cares? Really?
As it's not exactly a popular end-user Linux feature, for those that don't known, kexec allows a new kernel to be booted and switched to from an existing running Linux kernel without a full system reboot. Using kexec skips over the BIOS/UEFI hardware initialization -- this allows for a faster switch to a new kernel and other niche cases (e.g. there's a way to re-clock with the NVIDIA binary blob and then switching over to Nouveau, etc).
After Airlie replied to Linus, there was a soft response from Torvalds where he clarified his position.
So having looked at the patch itself, I don't dislike the notion of making sure certain fields are nicely initialized. So I don't hate the patch itself, but quite frankly, to me it doesn't smell even *remotely* like "regression fix". I don't think this is something that has ever worked.
And I do realize that some enterprise distros want to use kexec, but at the same time I say "that's *their* problem". We know kexec hasn't been horribly reliable, anybody who uses it should have be taking that into account.
I hope kexec gets more reliable, but I *also* really hope that our RC series will calm down, and on the whole, weighing the two concerns, when we're talking about something that has never worked before either, I think the thing is pretty clear.
That said, if there is some other real use-case ("this fixes problems with the BIOS from Xyz initializing things to random crap"), I'd have no real objections to the patch itself.
So my complaint is really a "I want to be more anal about the later -rc patches, I feel we're slipping", not a "I hate the patch per se".
Linus is basically wanting to reserve the post merge window now for "real" regression fixes and new hardware enablement, at least for the DRM subsystem.
The thread can be found on LKML.org for those wanting to read in full or check for updates.