Re: perf test "object code reading" segfaulting via usercopy check

[Jiri Olsa]

On Fri, Sep 09, 2016 at 12:47:20PM -0300, Arnaldo Carvalho de Melo wrote:
> Em Fri, Sep 09, 2016 at 05:41:25PM +0200, Jiri Olsa escreveu:
> > On Fri, Sep 09, 2016 at 12:36:26PM -0300, Arnaldo Carvalho de Melo wrote:
> > > Hi Adrian,
> > > 
> > > 	I noticed that 'perf test "object code reading"' is segfaulting
> > > here:
> > > 
> > > [root@jouet linux]# perf test -F "object code reading"
> > > 21: Test object code reading :Segmentation fault
> > > [root@jouet linux]# 
> > > 
> > > dmesg output below, trying to figure this out...
> > 
> > heya,
> > it's the new hardened user copy check.. I sent patches for that:
> > 
> >   http://marc.info/?l=linux-kernel&m=147332143929289&w=2
> >   http://marc.info/?l=linux-kernel&m=147332145229291&w=2
> 
> Cool, but that is for the kernel, without thinking too much about it, is
> there a way to change that 'perf test' entry to avoid doing what
> triggers the segfault?
> 
> My first thought was that it was reading 4K all the way to the end,
> where it should instead read just what is remaining, but I haven't
> checked this theory at all.

it's actually reading within the bounds of kernel text that triggers
that, it's the new CONFIG_HARDENED_USERCOPY feature:
  f5509cc18daa mm: Hardened usercopy

check the change log, there's list of conditions and
one of them is:
  - object must not overlap with kernel text

jirka