Re: [patch V9 10/39] x86/entry: Provide helpers for execute on irqstack

From: Qian Cai
Date: Sat Jun 13 2020 - 09:56:02 EST


On Wed, Jun 10, 2020 at 09:38:56PM +0200, Thomas Gleixner wrote:
> Qian Cai <cai@xxxxxx> writes:
> > On Tue, Jun 09, 2020 at 10:50:50PM +0200, Thomas Gleixner wrote:
> >> Thomas Gleixner <tglx@xxxxxxxxxxxxx> writes:
> >> >> BAD (after oom02)
> >> >> # cat /sys/kernel/debug/stackdepot/info
> >> >> Unique stacks: 140476
> >> >
> >> > That's indeed odd. I try to reproduce and figure out what really breaks
> >> > here.
> >>
> >> I checked your config file and I think I know where this comes from. Can
> >> you plase disable KASAN just for testing purposes and compare before
> >> after again?
> >
> > It turns out I'll need a few days to be able to get ahold of those
> > affected systems again. I'll be reporting back as soon as possible.
>
> I figured it out. It has nothing to do with asm_call_on...(). It's also
> unrelated to page_owner. It's purely a KASAN issue.
>
> KASAN trims stack traces with the help of filter_irq_stacks() which
> relies on __irqentry_text_start/end
>
> The entry rework moved the interrupt entry points out of the irqentry
> section, which breaks that filter function.
>
> I made it work again. Here are the numbers which make that obvious:
>
> Mainline:
> Unique stacks: 23403
> Depot index: 244
> Depot offset: 4208
>
> x86/entry:
> Unique stacks: 38835
> Depot index: 464
> Depot offset: 7616
>
> x86/entry + fix:
> Unique stacks: 23607
> Depot index: 247
> Depot offset: 14224
>
> So with the non-working trimming this generates more unique stacks and
> because they are not trimmed they become larger and eat more storage
> space. The resulting average per stack shows that:
>
> Mainline: 171 bytes per stack
> x86/entry: 195 bytes per stack
> x86/entry + fix: 172 bytes per stack
>
> I'll point you to a test branch shortly.

Thomas, I get ahold of one of the affected systems now if you want some
testing there.