Re: [tip:x86/mm] [x86/mm/tlb] 209954cbc7: will-it-scale.per_thread_ops 13.2% regression

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

On 2024-11-28 21:52, Rik van Riel wrote:
> On Thu, 2024-11-28 at 14:46 -0500, Mathieu Desnoyers wrote:
>>
>> I suspect you could use a similar per-cpu data structure per-mm
>> to keep track of the pending TLB flush mask, and update it simply
>> with
>> load/store to per-CPU data rather than have to cacheline-bounce all
>> over
>> the place due to frequent mm_cpumask atomic updates.
>>
>> Then you get all the benefits without introducing a window where
>> useless
>> TLB flush IPIs get triggered.
>>
>> Of course it's slightly less compact in terms of memory footprint
>> than a
>> cpumask, but you gain a lot by removing cache line bouncing on this
>> frequent context switch code path.
>>
>> Thoughts ?
> 
> The first thought that comes to mind is that we already
> have a per-CPU variable indicating which is the currently
> loaded mm on that CPU.

Only on x86 though.

> 
> We could probably just skip sending IPIs to CPUs that do
> not have the mm_struct currently loaded.
> 
> This can race against switch_mm_irqs_off() on a CPU
> switching to that mm simultaneously with the TLB flush,
> which should be fine because that CPU cannot load TLB
> entries from previously cleared page tables.
> 
> However, it does mean we cannot safely clear bits
> out of the mm_cpumask, because a race between clearing
> the bit on one CPU, and setting it on another would not
> be something we could easily catch at all, unless we
> can figure out some clever memory ordering thing there.
> 

Or we just build a per-cpu mm_cpumask from per-CPU state
every time we want to use the mm_cpumask. But AFAIU this
is going to be a tradeoff between:

- Overhead of context switch at scale

(e.g. will-it-scale:)
for a in $(seq 1 2); do (./context_switch1_threads -t 192 -s 20 &); done

For reference, my POC reaches 50% performance improvement with this.

   vs

- Overhead of TLB flush

(e.g. will-it-scale:)
./tlb_flush2_threads -t 192 -s 20

For reference, my POC has about 33% regression on that test case due
to extra work when using mm_cpumask.

So I guess what we end up doing really depends which scenario we consider
most frequent.

Thanks,

Mathieu

-- 
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com