Re: [RFC] perf/x86/amd: add support for Large Increment per Cycle Events

[Peter Zijlstra]

On Mon, Aug 26, 2019 at 02:59:15PM -0500, Kim Phillips wrote:
> The core AMD PMU has a 4-bit wide per-cycle increment for each
> performance monitor counter.  That works for most counters, but
> now with AMD Family 17h and above processors, for some, more than 15
> events can occur in a cycle.  Those events are called "Large
> Increment per Cycle" events, and one example is the number of
> SSE/AVX FLOPs retired (event code 0x003).  In order to count these
> events, two adjacent h/w PMCs get their count signals merged
> to form 8 bits per cycle total. 

*groan*

> In addition, the PERF_CTR count
> registers are merged to be able to count up to 64 bits.

That is daft; why can't you extend the existing MSR to 64bit?

> Normally, events like instructions retired, get programmed on a single
> counter like so:
> 
> PERF_CTL0 (MSR 0xc0010200) 0x000000000053ff0c # event 0x0c, umask 0xff
> PERF_CTR0 (MSR 0xc0010201) 0x0000800000000001 # r/w 48-bit count
> 
> The next counter at MSRs 0xc0010202-3 remains unused, or can be used
> independently to count something else.
> 
> When counting Large Increment per Cycle events, such as FLOPs,
> however, we now have to reserve the next counter and program the
> PERF_CTL (config) register with the Merge event (0xFFF), like so:
> 
> PERF_CTL0 (msr 0xc0010200) 0x000000000053ff03 # FLOPs event, umask 0xff
> PERF_CTR0 (msr 0xc0010201) 0x0000800000000001 # read 64-bit count, wr low 48b
> PERF_CTL1 (msr 0xc0010202) 0x0000000f004000ff # Merge event, enable bit
> PERF_CTR1 (msr 0xc0010203) 0x0000000000000000 # write higher 16-bits of count
> 
> The count is widened from the normal 48-bits to 64 bits by having the
> second counter carry the higher 16 bits of the count in its lower 16
> bits of its counter register.  Support for mixed 48- and 64-bit counting
> is not supported in this version.

This is diguisting.. please talk to your hardware people. I sort of
understand the pairing, but that upper 16 bit split for writes is just
woeful crap.

> For more details, search a Family 17h PPR for the "Large Increment per
> Cycle Events" section, e.g., section 2.1.15.3 on p. 173 in this version:
> 
> https://www.amd.com/system/files/TechDocs/56176_ppr_Family_17h_Model_71h_B0_pub_Rev_3.06.zip

My mama told me not to open random zip files of the interweb :-)

Also; afaict the only additional information there is that it works in
odd/even pairs and you have to program the odd one before the even one.
Surely you could've included that here.

> In order to support reserving the extra counter for a single Large
> Increment per Cycle event in the perf core, we:
> 
> 1. Add a f17h get_event_constraints() that returns only an even counter
> bitmask, since Large Increment events can only be placed on counters 0,
> 2, and 4 out of the currently available 0-5.

So hereby you promise that all LI events are unconstrained, right?
Also, what marks the paired counter in the used mask? Aaah, you modify
__perf_sched_find_counter(). Comments below.

> 2. We add a commit_scheduler hook that adds the Merge event (0xFFF) to
> any Large Increment event being scheduled.  If the event being scheduled
> is not a Large Increment event, we check for, and remove any
> pre-existing Large Increment events on the next counter.

That is weird at best; the scheduling hooks shouldn't be the one doing
the programming; that should be done in x86_pmu_enable(). Can't you do
this by changing amd_pmu::{en,dis}able() ?

(also; we really should rename some of those x86_pmu::ops :/)

> 3. In the main x86 scheduler, we reduce the number of available
> counters by the number of Large Increment per Cycle events being added.
> This improves the counter scheduler success rate.
> 
> 4. In perf_assign_events(), if a counter is assigned to a Large
> Increment event, we increment the current counter variable, so the
> counter used for the Merge event is skipped.
> 
> 5. In find_counter(), if a counter has been found for the
> Large Increment event, we set the next counter as used, to
> prevent other events from using it.
> 
> A side-effect of assigning a new get_constraints function for f17h
> disables calling the old (prior to f15h) amd_get_event_constraints
> implementation left enabled by commit e40ed1542dd7 ("perf/x86: Add perf
> support for AMD family-17h processors"), which is no longer
> necessary since those North Bridge events are obsolete.

> RFC because I'd like input on the approach, including how to add support
> for mixed-width (48- and 64-bit) counting for a single PMU.

Ideally I'd tell you to wait for sane hardware :/


> diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
> index 325959d19d9a..4596c141f348 100644
> --- a/arch/x86/events/core.c
> +++ b/arch/x86/events/core.c
> @@ -787,6 +787,18 @@ static bool __perf_sched_find_counter(struct perf_sched *sched)
>  		if (!__test_and_set_bit(idx, sched->state.used)) {
>  			if (sched->state.nr_gp++ >= sched->max_gp)
>  				return false;
> +			if (c->flags & PERF_X86_EVENT_LARGE_INC) {

Can we please call that something like:

  PERF_X86_EVENT_PAIR

> +				/*
> +				 * merged events need the Merge event
> +				 * on the next counter
> +				 */
> +				if (__test_and_set_bit(idx + 1,
> +						       sched->state.used))
> +					/* next counter already used */
> +					return false;

Coding Style wants { } there. Also, remove that line-break.

> +
> +				set_bit(idx + 1, sched->state.used);

  __set_bit() surely

> +			}
>  
>  			goto done;
>  		}
> @@ -849,14 +861,20 @@ int perf_assign_events(struct event_constraint **constraints, int n,
>  			int wmin, int wmax, int gpmax, int *assign)
>  {
>  	struct perf_sched sched;
> +	struct event_constraint *c;
> +
>  
>  	perf_sched_init(&sched, constraints, n, wmin, wmax, gpmax);
>  
>  	do {
>  		if (!perf_sched_find_counter(&sched))
>  			break;	/* failed */
> -		if (assign)
> +		if (assign) {
>  			assign[sched.state.event] = sched.state.counter;
> +			c = constraints[sched.state.event];
> +			if (c->flags & PERF_X86_EVENT_LARGE_INC)
> +				sched.state.counter++;
> +		}

How about you make __perf_sched_find_count() set the right value? That
already knows it did this.

>  	} while (perf_sched_next_event(&sched));
>  
>  	return sched.state.unassigned;
> @@ -952,6 +970,18 @@ int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign)
>  		    READ_ONCE(cpuc->excl_cntrs->exclusive_present))
>  			gpmax /= 2;
>  
> +		/*
> +		 * reduce the amount of available counters
> +		 * to allow fitting the Merge event along
> +		 * with their large increment event
> +		 */
> +		if (x86_pmu.flags & PMU_FL_MERGE) {
> +			for (i = 0; i < n; i++) {
> +				hwc = &cpuc->event_list[i]->hw;
> +				if (is_lg_inc_event(hwc) && gpmax > 1)

It should not be possible to hit !gpmax; make that a WARN.

> +					gpmax--;
> +			}

Alternatively you could have collect_events() cound the number of
'lg_inc' (we really have to come up with a better name) events on the
cpuc. Then you can do a simple subtraction and avoid the loop.

> +		}
>  		unsched = perf_assign_events(cpuc->event_constraint, n, wmin,
>  					     wmax, gpmax, assign);
>  	}
> @@ -1210,6 +1240,13 @@ int x86_perf_event_set_period(struct perf_event *event)
>  
>  	wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
>  
> +	/*
> +	 * Clear the Merge event counter's upper 16 bits since
> +	 * we currently declare a 48-bit counter width
> +	 */
> +	if (is_lg_inc_event(hwc))
> +		wrmsrl(x86_pmu_event_addr(idx + 1), 0);
> +

*yuck*...

>  	/*
>  	 * Due to erratum on certan cpu we need
>  	 * a second write to be sure the register