Re: [PATCH] percpu: Fix hint invariant breakage

[Dennis Zhou <dennis@kernel.org>]

Hello,

Sorry for the delay, I've been a bit sick.

On Mon, Mar 23, 2026 at 02:05:14PM +0000, Joonwon Kang wrote:
> > Hello,
> > 
> > On Fri, Mar 20, 2026 at 11:52:14AM +0000, Joonwon Kang wrote:
> > > The invariant "scan_hint_start > contig_hint_start if and only if
> > > scan_hint == contig_hint" should be kept for hint management. However,
> > > it could be broken in some cases:
> > > 
> > 
> > First I'd just like to apologize. I spent an hour yesterday trying to
> > remember why the invariant exists and the reality is this code is more
> > clever than it needs to be.
> 
> Thanks for taking time for this and sharing more context. While you are at
> it, I have a fundamental question on the invariant. I had deliberation and
> discussion on what benefits the invariant gets to the percpu allocator by
> its existence. My understanding is that if we put contig_hint before
> scan_hint when they are the same, it is more likely that contig_hint is
> broken by a future allocation, which leads to a linear scan after the
> scan_hint for hints update, although we could save scanning upto scan_hint
> when contig_hint is not broken. On the other hand, if we put scan_hint
> before contig_hint instead, it is more likely that scan_hint is broken
> while keeping contig_hint, which does not lead to the linear scan for
> hints update, although we could not save the scanning that could be saved
> in the other case.
> 
> In other words, if contig_hint breaking allocations occur a lot in general
> with the current invariant, the performance may more suffer than without
> the invariant. I also think that there would be no strict reason of having
> the invariant.
> 

I think the original premise is that percpu memory is quite expensive, 1
allocation costs nr_cpus * sizeof(allocation). So we do our best to bin
pack at the cost of faster allocations. We could always just break the
contig_hint but then over time we could cause more fragmentation.

The case that triggered this was netdev needing 8 byte objects with 16
byte alignment [1].

> So, could you clarify the necessity of the invariant? If there is no must
> reason, then I could post another spin-off patch to remove the invariant
> at all so that we could simplify the code and experiment the result. How
> do you think?
> 

I can't really recall the exact reasoning for the invariant, but it was
probably along the lines of wanting to not lose information if possible.

Say an earlier area becomes free that is the same size as the
contig_hint but with better alignment, we ant to use that as the
contig_hint but then we either have to lose the scan_hint or keep it
with the invariant. Given the premise above, I believe we want to
continue bin packing, I think the general idea of scanning next time
around isn't the worst thing.

Sadly because it's already there, and has worked for quite some time,
it's kind of on us today to provide data / reasoning to delete it. I'd
wager that some upcoming work is going to change how percpu gives out
objects either through some sort of slab caching that we can revisit
this more in that context.

Thanks,
Dennis