Message-ID: <b689b8d4-5fe1-6fae-c873-6e62a273cd07@intel.com>
Date:   Fri, 10 Aug 2018 09:25:02 -0700
From:   Reinette Chatre <reinette.chatre@...el.com>
To:     Peter Zijlstra <peterz@...radead.org>
Cc:     Dave Hansen <dave.hansen@...el.com>, tglx@...utronix.de,
        mingo@...hat.com, fenghua.yu@...el.com, tony.luck@...el.com,
        vikas.shivappa@...ux.intel.com, gavin.hindman@...el.com,
        jithu.joseph@...el.com, hpa@...or.com, x86@...nel.org,
        linux-kernel@...r.kernel.org
Subject: Re: [PATCH 0/2] x86/intel_rdt and perf/x86: Fix lack of coordination
 with perf

Hi Peter,

On 8/8/2018 10:33 AM, Reinette Chatre wrote:
> On 8/8/2018 12:51 AM, Peter Zijlstra wrote:
>> On Tue, Aug 07, 2018 at 03:47:15PM -0700, Reinette Chatre wrote:
>>>>  - I don't much fancy people accessing the guts of events like that;
>>>>    would not an inline function like:
>>>>
>>>>    static inline u64 x86_perf_rdpmc(struct perf_event *event)
>>>>    {
>>>> 	u64 val;
>>>>
>>>> 	lockdep_assert_irqs_disabled();
>>>>
>>>> 	rdpmcl(event->hw.event_base_rdpmc, val);
>>>> 	return val;
>>>>    }
>>>>
>>>>    Work for you?
>>>
>>> No. This does not provide accurate results. Implementing the above produces:
>>> pseudo_lock_mea-366   [002] ....    34.950740: pseudo_lock_l2: hits=4096
>>> miss=4
>>
>> But it being an inline function should allow the compiler to optimize
>> and lift the event->hw.event_base_rdpmc load like you now do manually.
>> Also, like Tony already suggested, you can prime that load just fine by
>> doing an extra invocation.
>>
>> (and note that the above function is _much_ simpler than
>> perf_event_read_local())
> 
> Unfortunately I do not find this to be the case. When I implement
> x86_perf_rdpmc() _exactly_ as you suggest above and do the measurement like:
> 
> l2_hits_before = x86_perf_rdpmc(l2_hit_event);
> l2_miss_before = x86_perf_rdpmc(l2_miss_event);
> l2_hits_before = x86_perf_rdpmc(l2_hit_event);
> l2_miss_before = x86_perf_rdpmc(l2_miss_event);
> /* read memory */
> l2_hits_after = x86_perf_rdpmc(l2_hit_event);
> l2_miss_after = x86_perf_rdpmc(l2_miss_event);
> 
> 
> Then the results are not accurate, neither are the consistently
> inaccurate to consider a constant adjustment:
> 
> pseudo_lock_mea-409   [002] ....   194.322611: pseudo_lock_l2: hits=4100
> miss=0
> pseudo_lock_mea-412   [002] ....   195.520203: pseudo_lock_l2: hits=4096
> miss=3
> pseudo_lock_mea-415   [002] ....   196.571114: pseudo_lock_l2: hits=4097
> miss=3
> pseudo_lock_mea-422   [002] ....   197.629118: pseudo_lock_l2: hits=4097
> miss=3
> pseudo_lock_mea-425   [002] ....   198.687160: pseudo_lock_l2: hits=4096
> miss=3
> pseudo_lock_mea-428   [002] ....   199.744156: pseudo_lock_l2: hits=4096
> miss=2
> pseudo_lock_mea-431   [002] ....   200.801131: pseudo_lock_l2: hits=4097
> miss=2
> pseudo_lock_mea-434   [002] ....   201.858141: pseudo_lock_l2: hits=4097
> miss=2
> pseudo_lock_mea-437   [002] ....   202.917168: pseudo_lock_l2: hits=4096
> miss=2
> 
> I was able to test Tony's theory and replacing the reading of the
> "after" counts with a direct rdpmcl() improve the results. What I mean
> is this:
> 
> l2_hit_pmcnum = x86_perf_rdpmc_ctr_get(l2_hit_event);
> l2_miss_pmcnum = x86_perf_rdpmc_ctr_get(l2_miss_event);
> l2_hits_before = x86_perf_rdpmc(l2_hit_event);
> l2_miss_before = x86_perf_rdpmc(l2_miss_event);
> l2_hits_before = x86_perf_rdpmc(l2_hit_event);
> l2_miss_before = x86_perf_rdpmc(l2_miss_event);
> /* read memory */
> rdpmcl(l2_hit_pmcnum, l2_hits_after);
> rdpmcl(l2_miss_pmcnum, l2_miss_after);
> 
> I did not run my full tests with the above but a simple read of 256KB
> pseudo-locked memory gives:
> pseudo_lock_mea-492   [002] ....   372.001385: pseudo_lock_l2: hits=4096
> miss=0
> pseudo_lock_mea-495   [002] ....   373.059748: pseudo_lock_l2: hits=4096
> miss=0
> pseudo_lock_mea-498   [002] ....   374.117027: pseudo_lock_l2: hits=4096
> miss=0
> pseudo_lock_mea-501   [002] ....   375.182864: pseudo_lock_l2: hits=4096
> miss=0
> pseudo_lock_mea-504   [002] ....   376.243958: pseudo_lock_l2: hits=4096
> miss=0
> 
> We thus seem to be encountering the issue Tony predicted where the
> memory being tested is evicting the earlier measurement code and data.

I thoroughly reviewed this email thread to ensure that all your feedback
is being addressed. At this time I believe the current solution does so
since it addresses all requirements I was able to capture:
- Use in-kernel interface to perf.
- Do not write directly to PMU registers.
- Do not introduce another PMU owner. perf maintains role as performing
  resource arbitration for PMU.
- User space is able to use perf and resctrl at the same time.
- event_base_rdpmc is accessed and used only within an interrupts
  disabled section.
- Internals of events are never accessed directly, inline function used.
- Due to "pinned" usage the scheduling of event may have failed.  Error
  state is checked in recommended way and have a credible error
  handling.
- use X86_CONFIG

The pseudocode of the current solution is presented below. With this
solution I am able to address our customer requirement to be able to
measure a pseudo-locked region accurately while also addressing your
requirements to use perf correctly.

Is this solution acceptable to you?

#include "../../events/perf_event.h" /* For X86_CONFIG() */

/*
 * The X86_CONFIG() macro cannot be used in
 * a designated initializer as below - the initialization of
 * the .config attribute is thus deferred to later in order
 * to use X86_CONFIG
 */

static struct perf_event_attr l2_miss_attr = {
        .type           = PERF_TYPE_RAW,
        .size           = sizeof(struct perf_event_attr),
        .pinned         = 1,
        .disabled       = 0,
        .exclude_user   = 1
};

static struct perf_event_attr l2_hit_attr = {
        .type           = PERF_TYPE_RAW,
        .size           = sizeof(struct perf_event_attr),
        .pinned         = 1,
        .disabled       = 0,
        .exclude_user   = 1
};

static inline int x86_perf_rdpmc_ctr_get(struct perf_event *event)
{
        lockdep_assert_irqs_disabled();

        return IS_ERR(event) ? 0 : event->hw.event_base_rdpmc;
}

static inline int x86_perf_event_error_state(struct perf_event *event)
{
        int ret = 0;
        u64 tmp;

        ret = perf_event_read_local(event, &tmp, NULL, NULL);
        if (ret < 0)
                return ret;

        if (event->attr.pinned && event->oncpu != smp_processor_id())
                return -EBUSY;

        return ret;
}

/*
 * Below is run by kernel thread on correct CPU as triggered
 * by user via debugfs
 */
static int measure_cycles_perf_fn(...)
{
        u64 l2_hits_before, l2_hits_after, l2_miss_before, l2_miss_after;
        struct perf_event *l2_miss_event, *l2_hit_event;
        int l2_hit_pmcnum, l2_miss_pmcnum;
        /* Other vars */

        l2_miss_attr.config = X86_CONFIG(.event=0xd1, .umask=0x10);
        l2_hit_attr.config = X86_CONFIG(.event=0xd1, .umask=0x2);
        l2_miss_event = perf_event_create_kernel_counter(&l2_miss_attr,
                                                         cpu,
                                                         NULL, NULL, NULL);
        if (IS_ERR(l2_miss_event))
                goto out;

        l2_hit_event = perf_event_create_kernel_counter(&l2_hit_attr,
                                                        cpu,
                                                        NULL, NULL, NULL);
        if (IS_ERR(l2_hit_event))
                goto out_l2_miss;

        local_irq_disable();
        if (x86_perf_event_error_state(l2_miss_event)) {
                local_irq_enable();
                goto out_l2_hit;
        }
        if (x86_perf_event_error_state(l2_hit_event)) {
                local_irq_enable();
                goto out_l2_hit;
        }
        /* Disable hardware prefetchers */
        /* Initialize local variables */
        l2_hit_pmcnum = x86_perf_rdpmc_ctr_get(l2_hit_event);
        l2_miss_pmcnum = x86_perf_rdpmc_ctr_get(l2_miss_event);
        rdpmcl(l2_hit_pmcnum, l2_hits_before);
        rdpmcl(l2_miss_pmcnum, l2_miss_before);
        /*
         * From SDM: Performing back-to-back fast reads are not guaranteed
         * to be monotonic. To guarantee monotonicity on back-toback reads,
         * a serializing instruction must be placed between the two
         * RDPMC instructions
         */
        rmb();
        rdpmcl(l2_hit_pmcnum, l2_hits_before);
        rdpmcl(l2_miss_pmcnum, l2_miss_before);
        rmb();
        /* Loop through pseudo-locked memory */
        rdpmcl(l2_hit_pmcnum, l2_hits_after);
        rdpmcl(l2_miss_pmcnum, l2_miss_after);
        rmb();
        /* Re-enable hardware prefetchers */
        local_irq_enable();
        /* Write results to kernel tracepoints */
out_l2_hit:
        perf_event_release_kernel(l2_hit_event);
out_l2_miss:
        perf_event_release_kernel(l2_miss_event);
out:
        /* Cleanup */
}

Your feedback has been valuable and greatly appreciated.

Reinette

Powered by blists - more mailing lists