Re: [PATCH 16/17] x86-64: Remove the PDA

[Ingo Molnar <mingo@elte.hu>]

* Tejun Heo <tj@kernel.org> wrote:

> Hello, Brian.
> 
> Brian Gerst wrote:
> >> How about something like the following?
> >>
> >> #define CANARY_OFFSET   40
> >> #define CANARY_SIZE     8
> >>
> >> DECLARE_PER_CPU(unsigned long, stack_canary);
> >>
> >> and in linker script,
> >>
> >> PERCPU_VADDR_PREALLOC(0, :percpu, CANARY_OFFSET + CANARY_SIZE)
> >> per_cpu__stack_canary = __per_cpu_start + CANARY_OFFSET;
> >>
> > 
> > The thing I don't like about the prealloc method is that it puts the
> > page-aligned variables at the end.  This leaves a gap which is
> > unavailable for dynamic allocations.  Stealing 48 bytes from the
> > bottom of the irqstack (which is 16k) keeps the page-aligned section
> > at the start.  It's really no different than how the thread_info
> > structure sits at the bottom of the process stack.
> > 
> > How about something like:
> > union irq_stack_union {
> > 	char irq_stack[IRQSTACKSIZE];
> > 	struct {
> > 		char pad[40];
> > 		unsigned long stack_canary;
> > 	}
> > };
> > 
> > That documents the overlay better, and avoids having to touch the
> > linker script.
> 
> I have no objection as long as it's sufficiently documented.

There is another advantage from Brian's trick of reusing the IRQ stack 
bottom: if we ever overflow the IRQ stack the kernel will likely stomp on 
the canary and overwrite it, and then (if the user runs on a 
stackprotector kernel) we will get an instantaneous assert and backtrace, 
exactly where the overflow happened.

Small overflows are otherwise rather hard to catch right on the spot so 
this is a bonus.

	Ingo