Re: [PATCH] perf/x86/intel/uncore: Fix oops when counting IMC uncore events on some TGL

[Liang, Kan]

On 5/27/2020 11:17 AM, David Laight wrote:
> From: Liang, Kan
> > Sent: 27 May 2020 16:01
> > On 5/27/2020 10:51 AM, David Laight wrote:
> > > From: Liang, Kan
> > > > Sent: 27 May 2020 15:47
> > > > On 5/27/2020 8:59 AM, David Laight wrote:
> > > > > From: kan.liang@xxxxxxxxxxxxxxx
> > > > > > Sent: 27 May 2020 13:31
> > > > > >
> > > > > > From: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
> > > > > >
> > > > > > When counting IMC uncore events on some TGL machines, an oops will be
> > > > > > triggered.
> > > > > >      [ 393.101262] BUG: unable to handle page fault for address:
> > > > > >      ffffb45200e15858
> > > > > >      [ 393.101269] #PF: supervisor read access in kernel mode
> > > > > >      [ 393.101271] #PF: error_code(0x0000) - not-present page
> > > > > >
> > > > > > Current perf uncore driver still use the IMC MAP SIZE inherited from
> > > > > > SNB, which is 0x6000.
> > > > > > However, the offset of IMC uncore counters for some TGL machines is
> > > > > > larger than 0x6000, e.g. 0xd8a0.
> > > > > >
> > > > > > Enlarge the IMC MAP SIZE for TGL to 0xe000.
> > > > >
> > > > > Replacing one 'random' constant with a different one
> > > > > doesn't seem like a proper fix.
> > > > >
> > > > > Surely the actual bounds of the 'memory' area are properly
> > > > > defined somewhere.
> > > > > Or at least should come from a table.
> > > > >
> > > > > You also need to verify that the offsets are within the mapped area.
> > > > > An unexpected offset shouldn't try to access an invalid address.
> > > >
> > > > Thanks for the review.
> > > >
> > > > I agree that we should add a check before mapping the area to prevent
> > > > the issue happens again.
> > > >
> > > > I think the check should be a generic check for all platforms which try
> > > > to map an area, not just for TGL. I will submit a separate patch for the
> > > > check.
> > >
> > > You need a check that the actual access is withing the mapped area.
> > > So instead of getting an OOPS you get a error.
> > >
> > > This is after you've mapped it.
> >
> > Sure. Will add a WARN_ONCE() before the actual access.
>
> No that will still panic some systems.
> pr_warn() is all you need.

If we print a warning for each access, there will be too many warnings.
I think I will use pr_warn_once instead.

Thanks,
Kan