Re: [PATCH] perf/x86/amd: Warn only on new bits set

[Peter Zijlstra]

On Tue, Jun 25, 2024 at 07:47:06AM -0700, Paul E. McKenney wrote:
> On Tue, Jun 25, 2024 at 01:57:34PM +0200, Peter Zijlstra wrote:
> > On Fri, May 24, 2024 at 07:10:20AM -0700, Breno Leitao wrote:
> > > Warning at every leaking bits can cause a flood of message, triggering
> > > vairous stall-warning mechanisms to fire, including CSD locks, which
> > > makes the machine to be unusable.
> > > 
> > > Track the bits that are being leaked, and only warn when a new bit is
> > > set.
> > > 
> > > Suggested-by: Paul E. McKenney <paulmck@xxxxxxxxxx>
> > > Signed-off-by: Breno Leitao <leitao@xxxxxxxxxx>
> > > ---
> > >  arch/x86/events/amd/core.c | 9 +++++++--
> > >  1 file changed, 7 insertions(+), 2 deletions(-)
> > > 
> > > diff --git a/arch/x86/events/amd/core.c b/arch/x86/events/amd/core.c
> > > index 1fc4ce44e743..df0ba2382d13 100644
> > > --- a/arch/x86/events/amd/core.c
> > > +++ b/arch/x86/events/amd/core.c
> > > @@ -941,11 +941,12 @@ static int amd_pmu_v2_snapshot_branch_stack(struct perf_branch_entry *entries, u
> > >  static int amd_pmu_v2_handle_irq(struct pt_regs *regs)
> > >  {
> > >  	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
> > > +	static atomic64_t status_warned = ATOMIC64_INIT(0);
> > > +	u64 reserved, status, mask, new_bits;
> > >  	struct perf_sample_data data;
> > >  	struct hw_perf_event *hwc;
> > >  	struct perf_event *event;
> > >  	int handled = 0, idx;
> > > -	u64 reserved, status, mask;
> > >  	bool pmu_enabled;
> > >  
> > >  	/*
> > > @@ -1010,7 +1011,11 @@ static int amd_pmu_v2_handle_irq(struct pt_regs *regs)
> > >  	 * the corresponding PMCs are expected to be inactive according to the
> > >  	 * active_mask
> > >  	 */
> > > -	WARN_ON(status > 0);
> > > +	if (status > 0) {
> > > +		new_bits = atomic64_fetch_or(status, &status_warned) ^ atomic64_read(&status_warned);
> > > +		// A new bit was set for the very first time.
> > > +		WARN(new_bits, "New overflows for inactive PMCs: %llx\n", new_bits);
> > > +	}
> > 
> > Why not just a WARN_ON_ONCE() instead? This really shouldn't be
> > happening in the first place.
> 
> We did consider that, but seeing the full set of bits that shouldn't
> have been happening in the first place helps with debuggging.
> 
> But is there a better way to accumulate and print the full set of
> unexpected bits?

Dunno, I was just wondering if the whole thing wasn't massive overkill.
The changelog wasn't really explaining much here.