Re: [RFC PATCH 00/23] accel/tcg: Convert victim tlb to IntervalTree

[Pierrick Bouvier <pierrick.bouvier@linaro.org>]

On 10/9/24 10:10, Richard Henderson wrote:
> On 10/9/24 09:27, BALATON Zoltan wrote:
>> On Wed, 9 Oct 2024, Richard Henderson wrote:
>>> Based-on: 20241009000453.315652-1-richard.henderson@linaro.org
>>> ("[PATCH v3 00/20] accel/tcg: Introduce tlb_fill_align hook")
>>>
>>> The initial idea was: how much can we do with an intelligent data
>>> structure for the same cost as a linear search through an array?
>>>
>>> This is an initial installment along these lines.  This is about
>>> as far as I can go without first converting all targets to the
>>> new tlb_fill_align hook.  Indeed, the final two patches will not
>>> compile with all targets enabled, but hint at the direction of
>>> the next steps.
>>>
>>> I do not expect large perf changes with this patch set.  I will
>>> be happy if performance comes out even.
>>
>> Then what's the point?
> 
> Eventually fixing the page size > TARGET_PAGE_SIZE performance issues.
> 
> E.g. with a 16k or 64k aarch64 guest kernel, we still have TARGET_PAGE_SIZE at 4k, so all
> guest pages are "large", and so run into our current behaviour of flushing the entire tlb
> too often.
> 
> Even without that, I expect further cleanups to improve performance, we're just not there yet.
> 
> 
> r~
> 

Does merging pages over a given range be something we could benefit from 
too? In this case, entries in our tlbtree would have varying size, 
allowing us to cover more space with a single entry.

It would allow us to have a more shallow tlbtree (given that it's 
rebalanced when modified) and speed up walking operations.

I'm not sure if it can help performance wise though.