Re: ARM64: Regression with commit e3067861ba66 ("arm64: add basic VMAP_STACK support")

From: Leo Yan
Date: Mon Oct 16 2017 - 20:31:13 EST


On Mon, Oct 16, 2017 at 03:35:46PM +0100, Robin Murphy wrote:
> On 16/10/17 15:26, Mark Rutland wrote:
> > On Mon, Oct 16, 2017 at 03:12:45PM +0100, Robin Murphy wrote:
> >> On 16/10/17 14:48, Mark Rutland wrote:
> >>> Hi Leo,
> >>>
> >>> On Mon, Oct 16, 2017 at 09:17:23AM +0800, Leo Yan wrote:
> >>>> On Tue, Oct 10, 2017 at 05:03:44PM +0100, Robin Murphy wrote:
> >>>>> On 10/10/17 16:45, Mark Rutland wrote:
> >>>>>> On Tue, Oct 10, 2017 at 10:27:25PM +0800, Leo Yan wrote:
> >>>>>>> I work mainline kernel on Hikey620 board, I find it's easily to
> >>>>>>> introduce the panic and report the log as below. So I bisect the kernel
> >>>>>>> and finally narrow down the commit e3067861ba66 ("arm64: add basic
> >>>>>>> VMAP_STACK support") which introduce this issue.
> >>>>>>>
> >>>>>>> I tried to remove 'select HAVE_ARCH_VMAP_STACK' from
> >>>>>>> arch/arm64/Kconfig, then I can see the panic issue will dismiss. So
> >>>>>>> could you check this and have insight for this issue?
> >>>>>>
> >>>>>> Given the stuff in the backtrace, my suspicion is something is trying to
> >>>>>> perform DMA to/from the stack, getting junk addresses form the attempted
> >>>>>> virt<->phys conversions.
> >>>>>>
> >>>>>> Could you try enabling both VMAP_STACK and CONFIG_DEBUG_VIRTUAL?
> >>>>>
> >>>>> CONFIG_DMA_API_DEBUG should scream about drivers trying to use stack
> >>>>> addresses either way, too.
> >>>>
> >>>> Thanks for suggestions, Mark & Robin.
> >>>>
> >>>> I enabled these debugging configs but cannot get clue from it; but
> >>>> occasionally found this issue is quite likely related with CA53 errata,
> >>>> especialy ERRATA_A53_855873 is the relative one. So I changed to use
> >>>> ARM-TF mainline code with ERRATA fixing, this issue can be dismissed.
> >>>
> >>> Thanks for the update.
> >>>
> >>> Just to confirm, with the updated firmware you no longer see the issue?
> >>>
> >>> I can't immediately see how that would be related.
> >>
> >> Cores up to r0p2 have the other errata to which
> >> ARM64_WORKAROUND_CLEAN_CACHE also applies anyway; r3p0+ have an ACTLR
> >> bit to do thee CVAC->CIVAC upgrade in hardware, and our policy is that
> >> we expect firmware to enable such hardware workarounds where possible. I
> >> assume that's why we don't explicitly document 855873 anywhere in Linux.
> >
> > Sure, I also looked it up. ;)
> >
> > I meant that I couldn't immediately see why VMAP'd stacks were likely to
> > tickle issues with that more reliably.
>
> Ah, right - in context, "that" appeared to refer to "updated firmware",
> not "VMAP_STACK". Sorry.
>
> I guess the vmap addresses might tickle the "same L2 set" condition
> differently to when both stack and DMA buffer are linear map addresses.

A bit more info for this.

I can reproduce this memory abort panic, and the panic places are not
consistent; usually it's related with kmalloc address. Do you think
"VMAP_STACK" introduces much more operations for cache clean? If
so if might be in the same *set* with any other memory access (like
kmalloc operations), then trigger data abort.

Hikey has CA53 CPUs is r3 version so it's luck can directly apply the
ERRATA 855873 in ARM-TF.

BTW, in case I may mislead you guys, we should note there have another
two ERRATAs applied in ARM-TFv1.4 for Hikey:

ERRATA_A53_836870 := 1
ERRATA_A53_843419 := 1

Thanks,
Leo Yan