[PATCH] drivers/perf: Fix kernel panic when rmmod PMU modules during perf sampling

[Will Deacon]

On Thu, Jul 16, 2020 at 11:36:10AM +0100, John Garry wrote:
> On 16/07/2020 11:30, Will Deacon wrote:
> > On Thu, Jul 16, 2020 at 11:26:25AM +0100, Robin Murphy wrote:
> > > On 2020-07-16 10:41, Will Deacon wrote:
> > > > On Thu, Jul 16, 2020 at 05:19:25PM +0800, Qi Liu wrote:
> > > > > Kernel panic will also happen when users try to unbind PMU drivers with
> > > > > device. This unbind issue could be solved by another patch latter.
> > > > > 
> > > > >    drivers/perf/arm_smmuv3_pmu.c                 | 1 +
> > > > >    drivers/perf/fsl_imx8_ddr_perf.c              | 1 +
> > > > >    drivers/perf/hisilicon/hisi_uncore_ddrc_pmu.c | 1 +
> > > > >    drivers/perf/hisilicon/hisi_uncore_hha_pmu.c  | 1 +
> > > > >    drivers/perf/hisilicon/hisi_uncore_l3c_pmu.c  | 1 +
> > > > >    5 files changed, 5 insertions(+)
> > > > > 
> > > > > diff --git a/drivers/perf/arm_smmuv3_pmu.c b/drivers/perf/arm_smmuv3_pmu.c
> > > > > index 48e28ef..90caba56 100644
> > > > > --- a/drivers/perf/arm_smmuv3_pmu.c
> > > > > +++ b/drivers/perf/arm_smmuv3_pmu.c
> > > > > @@ -742,6 +742,7 @@ static int smmu_pmu_probe(struct platform_device *pdev)
> > > > >    	platform_set_drvdata(pdev, smmu_pmu);
> > > > > 
> > > > >    	smmu_pmu->pmu = (struct pmu) {
> > > > > +		.module		= THIS_MODULE,
> > > > I thought platform_driver_register() did this automatically?
> > > For the platform device itself, yes, but this is for the PMU device - perf
> > > needs to take a reference to the module, otherwise the platform device can
> > > still be pulled out from under its feet.
> > Urgh, gross.
> > 
> > > I can't remember if we ever discussed making perf_pmu_register() do the same
> > > trick as platform_device_register() and friends, but obviously it's a
> > > possibility.
> > Yeah, but I suppose this patch is the right thing to do for now. I'll queue
> > it as a fix.
> > 
> 
> Please also note what Qi Liu wrote about being able to unbind the driver and
> cause the same issue. I don't know if you can about that issue also.

I guess we have to throw in some '.suppress_bind_attrs = true,' lines as
well, then? I'll queue this as-is, but happy to take a follow-up if somebody
can test it.

Will