linux-kernel - Re: [PATCH 4/9] perf/x86/intel: Support hardware TopDown metrics

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Tue, 28 May 2019 14:24:14 -0400
From:   "Liang, Kan" <kan.liang@...ux.intel.com>
To:     Peter Zijlstra <peterz@...radead.org>
Cc:     acme@...nel.org, mingo@...hat.com, linux-kernel@...r.kernel.org,
        tglx@...utronix.de, jolsa@...nel.org, eranian@...gle.com,
        alexander.shishkin@...ux.intel.com, ak@...ux.intel.com
Subject: Re: [PATCH 4/9] perf/x86/intel: Support hardware TopDown metrics



On 5/28/2019 9:30 AM, Peter Zijlstra wrote:
> On Tue, May 21, 2019 at 02:40:50PM -0700, kan.liang@...ux.intel.com wrote:
>> +static u64 icl_metric_update_event(struct perf_event *event, u64 val)
>> +{
>> +	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
>> +	struct hw_perf_event *hwc = &event->hw;
>> +	u64 newval, metric, slots_val = 0, new, last;
>> +	bool nmi = in_nmi();
>> +	int txn_flags = nmi ? 0 : cpuc->txn_flags;
>> +
>> +	/*
>> +	 * Use cached value for transaction.
>> +	 */
>> +	newval = 0;
>> +	if (txn_flags) {
>> +		newval = cpuc->txn_metric;
>> +		slots_val = cpuc->txn_slots;
>> +	} else if (nmi) {
>> +		newval = cpuc->nmi_metric;
>> +		slots_val = cpuc->nmi_slots;
>> +	}
>> +
>> +	if (!newval) {
>> +		slots_val = val;
>> +
>> +		rdpmcl((1<<29), newval);
>> +		if (txn_flags) {
>> +			cpuc->txn_metric = newval;
>> +			cpuc->txn_slots = slots_val;
>> +		} else if (nmi) {
>> +			cpuc->nmi_metric = newval;
>> +			cpuc->nmi_slots = slots_val;
>> +		}
>> +
>> +		if (!(txn_flags & PERF_PMU_TXN_REMOVE)) {
>> +			/* Reset the metric value when reading
>> +			 * The SLOTS register must be reset when PERF_METRICS reset,
>> +			 * otherwise PERF_METRICS may has wrong output.
>> +			 */
> 
> broken comment style.. (and grammer)

Missed a full stop.
Should be "Reset the metric value for each read."
> 
>> +			wrmsrl(MSR_PERF_METRICS, 0);
>> +			wrmsrl(MSR_CORE_PERF_FIXED_CTR3, 0);
> 
> I don't get this, overflow happens on when we flip sign, so why is
> programming 0 a sane thing to do?

Reset the counters (programming 0) don't trigger overflow.
We have to reset both registers for each read to avoid the known 
PERF_METRICS issue.


> 
>> +			hwc->saved_metric = 0;
>> +			hwc->saved_slots = 0;
>> +		} else {
>> +			/* saved metric and slots for context switch */
>> +			hwc->saved_metric = newval;
>> +			hwc->saved_slots = val;
>> +
>> +		}
>> +		/* cache the last metric and slots */
>> +		cpuc->last_metric = hwc->last_metric;
>> +		cpuc->last_slots = hwc->last_slots;
>> +		hwc->last_metric = 0;
>> +		hwc->last_slots = 0;
>> +	}
>> +
>> +	/* The METRICS and SLOTS have been reset when reading */
>> +	if (!(txn_flags & PERF_PMU_TXN_REMOVE))
>> +		local64_set(&hwc->prev_count, 0);
>> +
>> +	if (is_slots_event(event))
>> +		return (slots_val - cpuc->last_slots);
>> +
>> +	/*
>> +	 * The metric is reported as an 8bit integer percentage
>> +	 * suming up to 0xff. As the counter is less than 64bits
>> +	 * we can use the not used bits to get the needed precision.
>> +	 * Use 16bit fixed point arithmetic for
>> +	 * slots-in-metric = (MetricPct / 0xff) * val
>> +	 * This works fine for upto 48bit counters, but will
>> +	 * lose precision above that.
>> +	 */
>> +
>> +	metric = (cpuc->last_metric >> ((hwc->idx - INTEL_PMC_IDX_FIXED_METRIC_BASE)*8)) & 0xff;
>> +	last = (((metric * 0xffff) >> 8) * cpuc->last_slots) >> 16;
> 
> How is that cpuc->last_* crap not broken for NMIs ?

There should be no NMI for slots or metric events at the moment, because 
the MSR_PERF_METRICS and MSR_CORE_PERF_FIXED_CTR3 are reset in first read.
Other NMIs will not touch the codes here.

Thanks,
Kan

> 
>> +
>> +	metric = (newval >> ((hwc->idx - INTEL_PMC_IDX_FIXED_METRIC_BASE)*8)) & 0xff;
>> +	new = (((metric * 0xffff) >> 8) * slots_val) >> 16;
>> +
>> +	return (new - last);
>> +}
> 
> 
> This is diguisting.. and unreadable.
> 
> mul_u64_u32_shr() is actually really fast, use it.
>