Re: UBSAN: Undefined behaviour in arch/x86/events/intel/p6.c:116:29

[Jiri Olsa]

On Wed, Dec 04, 2019 at 04:06:56PM +0100, Peter Zijlstra wrote:
> On Wed, Dec 04, 2019 at 01:15:40PM +0100, Jiri Olsa wrote:
> > On Tue, Dec 03, 2019 at 03:39:49PM +0200, Meelis Roos wrote:
> > > > Does something like so fix it?
> > > 
> > > Unfortunately not (tested on top of todays git):
> > 
> > hi,
> > which p6 model are you seeing this on?
> > how do you trigger that?
> 
> Triggers on any p6 model. I hacked up perf and used "qemu-system-x86_64
> -cpu pentium2".
> 
> The below seems to cure things.
> 
> ---
> diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
> index 9a89d98c55bd..f17417644665 100644
> --- a/arch/x86/events/core.c
> +++ b/arch/x86/events/core.c
> @@ -1642,9 +1643,12 @@ static struct attribute_group x86_pmu_format_group __ro_after_init = {
>  
>  ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr, char *page)
>  {
> -	struct perf_pmu_events_attr *pmu_attr = \
> +	struct perf_pmu_events_attr *pmu_attr =

ugh, did this do something weird? ;-)

>  		container_of(attr, struct perf_pmu_events_attr, attr);
> -	u64 config = x86_pmu.event_map(pmu_attr->id);
> +	u64 config = 0;
> +
> +	if (pmu_attr->id < x86_pmu.max_events)
> +		x86_pmu.event_map(pmu_attr->id);

hum, should this be assigned to config?

		config = x86_pmu.event_map(pmu_attr->id);

jirka

>  
>  	/* string trumps id */
>  	if (pmu_attr->event_str)
> @@ -1713,6 +1717,9 @@ is_visible(struct kobject *kobj, struct attribute *attr, int idx)
>  {
>  	struct perf_pmu_events_attr *pmu_attr;
>  
> +	if (idx >= x86_pmu.max_events)
> +		return 0;
> +
>  	pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr.attr);
>  	/* str trumps id */
>  	return pmu_attr->event_str || x86_pmu.event_map(idx) ? attr->mode : 0;
>