Re: Random panic in load_balance() with 3.16-rc

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Jul 24, 2014 at 11:47:17AM -0700, Linus Torvalds wrote:
> However, that constant spilling part just counts as "too stupid to
> live". The real bug is this:
> 
>         movq    $load_balance_mask, -136(%rbp)  #, %sfp
>         subq    $184, %rsp      #,
> 
> where gcc creates the stack frame *after* having already used it to
> save that constant *deep* below the stack frame.
> 
> The x86-64 ABI specifies a 128-byte red-zone under the stack pointer,
> and this is ok by that limit. It looks like it's illegal (136 > 128),
> but the fact is, we've had four "pushq"s to update %rsp since loading
> the frame pointer, so it's just *barely* legal with the red-zoning.
> 
> But we build the kernel with -mno-red-zone. We do *not* follow the
> x86-64 ABI wrt redzoning, because we *cannot*: interrupts while in
> kernel mode *will* use the stack without a redzone. So that
> "-mno-red-zone" is not some "optional guideline". It's a hard and
> harsh requirement for the kernel, and gcc-4.9 is a buggy piece of shit
> for ignoring it. And your bug happens becuase you happen to hit an
> interrupt _just_ in that single instruction window (or perhaps hit
> some other similar case and corrupted kernel data structures earlier).

Ooh, shiny, I so missed all that (also didn't know about red-zones
etc..).

Glad this got sorted.