RE: [PATCH V1 2/3] perf/x86/intel/bm.c: Add Intel Branch Monitoring support

[Thomas Gleixner]

On Mon, 13 Nov 2017, Dey, Megha wrote:
> >-----Original Message-----
> >From: Peter Zijlstra [mailto:peterz@xxxxxxxxxxxxx]
> >Sent: Monday, November 13, 2017 1:00 AM
> >To: Megha Dey <megha.dey@xxxxxxxxxxxxxxx>
> >Cc: x86@xxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; linux-

Please fix your mail client so it does not add this complete useless
information to the reply.

> >On Sat, Nov 11, 2017 at 01:20:05PM -0800, Megha Dey wrote:
> >> +/*
> >> + * Unmask the NMI bit of the local APIC the first time task is
> >> +scheduled
> >> + * on a particular CPU.
> >> + */
> >> +static void intel_bm_unmask_nmi(void) {
> >> +	this_cpu_write(bm_unmask_apic, 0);
> >> +
> >> +	if (!(this_cpu_read(bm_unmask_apic))) {
> >> +		apic_write(APIC_LVTPC, APIC_DM_NMI);
> >> +		this_cpu_inc(bm_unmask_apic);
> >> +	}
> >> +}
> >
> >What? Why?
> 

> Normally, other drivers using perf create an event on every CPU (thereby
> calling perf_init on every CPU), where this bit(APIC_DM_NMI)is explicitly
> unmasked.  In our driver, we do not do this (since we are worried only
> about a particular task) and hence this bit is only disabled on the local
> APIC where the perf event is initialized.
>
> As such, if the task is scheduled out to some other CPU, this bit is set
> and hence would stop the interrupt from reaching the processing core.

Still that code makes no sense at all and certainly does not do what you
claim it does:

> >> +	this_cpu_write(bm_unmask_apic, 0);
> >> +
> >> +	if (!(this_cpu_read(bm_unmask_apic))) {

So first you write the per cpu variable to 0 and then you check whether it
is zero, which is pointless obviously.

> >
> >> +static int intel_bm_event_add(struct perf_event *event, int mode) {

Please move the opening bracket of the function into the next line. See the
kernel coding style documentation.

Thanks,

	tglx