Re: Optimizing small reads

[Linus Torvalds <torvalds@linux-foundation.org>]

On Fri, 3 Oct 2025 at 02:55, Kiryl Shutsemau <kirill@shutemov.name> wrote:
>
> So, you intentionally bypassing refcount bump on the folio which is part
> of the usual page cache lookup protocol.

Yes. The big cost that Willy reported originally was the cacheline
bouncing of the refcount (due to having tons and tons of very small
reads on the same page in a very threaded load).

My initial reaction was that it was a completely unrealistic
microbenchmark (who would read the same page from  multiple threads at
the same time unless they were just doing some silly benchmark with no
real-world application?), but apparently there are real loads.

> But without the pin, what prevents the folio from being freed and
> reallocated in the same spot under you? Do we wait for a grace period
> somewhere in this reallocation cycle? I don't see it.

That was one of the theories and it might be the right one. There was
some discussion about having a sequence number. It's a long time ago,
I forget.

I had some reason I thought it wasn't an issue, but after all this
time I can't recall.

There's a reason that patch was marked 'Not-yet-signed-off-by' :)

         Linus