Re: [PATCH v3 2/4] sched_clock: Optimize cache line usage

[Daniel Thompson <daniel.thompson@linaro.org>]

On 05/02/15 01:14, Stephen Boyd wrote:
> On 01/30, Daniel Thompson wrote:
>> diff --git a/kernel/time/sched_clock.c b/kernel/time/sched_clock.c
>> index 3d21a8719444..cb69a47dfee4 100644
>> --- a/kernel/time/sched_clock.c
>> +++ b/kernel/time/sched_clock.c
>> @@ -18,28 +18,44 @@
>>  #include <linux/seqlock.h>
>>  #include <linux/bitops.h>
>>  
>> -struct clock_data {
>> -	ktime_t wrap_kt;
>> +/**
>> + * struct clock_read_data - data required to read from sched_clock
>> + *
> 
> Nitpick: Won't kernel-doc complain that members aren't
> documented?

It does indeed. I'll add descriptions here...


>> + * Care must be taken when updating this structure; it is read by
>> + * some very hot code paths. It occupies <=48 bytes and, when combined
>> + * with the seqcount used to synchronize access, comfortably fits into
>> + * a 64 byte cache line.
>> + */
>> +struct clock_read_data {
>>  	u64 epoch_ns;
>>  	u64 epoch_cyc;
>> -	seqcount_t seq;
>> -	unsigned long rate;
>> +	u64 sched_clock_mask;
>> +	u64 (*read_sched_clock)(void);
>>  	u32 mult;
>>  	u32 shift;
>>  	bool suspended;
>>  };
>>  
>> +/**
>> + * struct clock_data - all data needed for sched_clock (including
>> + *                     registration of a new clock source)
>> + *
> 
> Same comment.

... and here.


>> + * The ordering of this structure has been chosen to optimize cache
>> + * performance. In particular seq and read_data (combined) should fit
>> + * into a single 64 byte cache line.
>> + */
>> +struct clock_data {
>> +	seqcount_t seq;
>> +	struct clock_read_data read_data;
>> +	ktime_t wrap_kt;
>> +	unsigned long rate;
>> +};
>> @@ -60,15 +79,16 @@ unsigned long long notrace sched_clock(void)
>>  {
>>  	u64 cyc, res;
>>  	unsigned long seq;
>> +	struct clock_read_data *rd = &cd.read_data;
>>  
>>  	do {
>>  		seq = raw_read_seqcount_begin(&cd.seq);
>>  
>> -		res = cd.epoch_ns;
>> -		if (!cd.suspended) {
>> -			cyc = read_sched_clock();
>> -			cyc = (cyc - cd.epoch_cyc) & sched_clock_mask;
>> -			res += cyc_to_ns(cyc, cd.mult, cd.shift);
>> +		res = rd->epoch_ns;
>> +		if (!rd->suspended) {
> 
> Should this have likely() treatment? It would be really nice if
> we could use static branches here to avoid any branch penalty at
> all. I guess that would need some sort of special cased
> stop_machine() though. Or I wonder if we could replace
> rd->read_sched_clock() with a dumb function that returns
> cd.epoch_cyc so that the math turns out to be 0?

Great idea.

Making this code branchless with a special function sounds very much
better than using likely().