[arm64] OOPS when using /proc/kcore to disassemble the kernel symbols in "perf top"

[Mark Rutland]

Hi,

On Fri, Apr 21, 2017 at 01:46:43PM +0800, Tan Xiaojun wrote:
> On 2017/4/20 9:38, Tan Xiaojun wrote:
> > On 2017/4/19 19:49, Mark Rutland wrote:
> >> Hi,
> >>
> >> Ard, this sseems to be a nomap issue. Please see below.
> >>
> >> Xiaojun, for some reason, the first message in this thread didn't seem
> >> to make it to LAKML (or to me). In future could you please Cc me for
> >> emails regarding perf on arm/arm64?
> >>
> > 
> > Sorry, this is my negligence.
> > 
> >> On Wed, Apr 19, 2017 at 09:44:56AM +0530, Pratyush Anand wrote:
> >>> On Saturday 15 April 2017 02:18 PM, Tan Xiaojun wrote:
> >>>> My test server is Hisilicon D03/D05 (arm64).
> >>>> Kernel source code is 4.11-rc6 (up to date) and config (as an attachment in the end) is generated by defconfig.
> >>>> (Old version does not seem to have this problem. Linux-4.1 is fine and other versions I have not tested yet.)
> >>>
> >>> I tested with mustang(ARM64) and 4.11-rc6 and could not reproduce it.
> >>>
> 
> Hi, 
> Pratyush,
> 
> Sorry, could you test it again? Because I tested it many times and found it is not triggered every time.
> And you can run "perf top -U" and try more kernel symbols to increase the probability of occurrence, or 
> maybe you can try Mark's way "cat /proc/kcore > /dev/null".
> 
> I would like to confirm whether this is hardware related, but I have no other arm64 boards except the
> boards of Hisilicon.

As I mentioned in my prior reply, this is a bug in the way we handle
nomap memory in the kernel.

This is not hardware related, and this is not specific to perf.

The kcore code expects that if a vmalloc mapping has a corresponding
struct page, that it can be accessed via the linear mapping. However,
this is not true for nomap memory.

> I found the patch which introduced the problem.
> The commit is:
> 
> commit f9040773b7bbbd9e98eb6184a263512a7cfc133f
> Author: Ard Biesheuvel <ard.biesheuvel at linaro.org>
> Date:   Tue Feb 16 13:52:40 2016 +0100
> 
>     arm64: move kernel image to base of vmalloc area
> 
>     This moves the module area to right before the vmalloc area, and moves
>     the kernel image to the base of the vmalloc area. This is an intermediate
>     step towards implementing KASLR, which allows the kernel image to be
>     located anywhere in the vmalloc area.
> 
>     Since other subsystems such as hibernate may still need to refer to the
>     kernel text or data segments via their linears addresses, both are mapped
>     in the linear region as well. The linear alias of the text region is
>     mapped read-only/non-executable to prevent inadvertent modification or
>     execution.
> 
>     Signed-off-by: Ard Biesheuvel <ard.biesheuvel at linaro.org>
>     Signed-off-by: Catalin Marinas <catalin.marinas at arm.com>
> 
> It can work well without this patch in Linux-4.5-rc4. And it can
> trigger an OOPS with this patch in Linux-4.5-rc4.
> 
> I try to revert it in v4.11-rc6, but it involves too much conflict.
> So I need to understand this patch fist. Then I can known where the problem is.

Reverting this patch is not the correct fix.

The fix will either be changing the way we set things up for nomap
memory, or with additions to the kcore or vread code to cater for nomap.

Thanks,
Mark.