Re: [PATCH 0/2] mm/zswap: optimize the scalability of zswap rb-tree

[Yosry Ahmed]

On Wed, Jan 17, 2024 at 4:18 PM Chris Li <chrisl@xxxxxxxxxx> wrote:
>
> Hi Yosry,
>
> On Wed, Jan 17, 2024 at 3:48 PM Yosry Ahmed <yosryahmed@xxxxxxxxxx> wrote:
> >
> > > Currently the xarray patch should have everything it takes to use RCU
> > > read lock. However taking out the tree spinlock is more work than
> > > previously. If we are going to remove the tree spinlock, I think we
> > > should revert back to doing a zswap tree lookup and return the zswap
> > > entry with reference increased. The tree mapping can still decouple
> > > from the zswap entry reference count drop to zero.  Anyway, my V1 of
> > > the xarray patch will not include removing the tree spinlock.
> >
> > Interesting. What do you mean by removing the tree spinlock? My
> > assumption was that the xarray reduces lock contention because we do
> > not need a lock to do lookups, but we still need the lock otherwise.
> > Did you have something in mind to completely remove the tree lock?
>
> In my current xarray series, it adds the xarray alongside the rb tree.
> Xarray has its own internal lock as well. Effectively zswap now has
> two locks instead of just one previously.  The xarray lock will not
> have any contention due to the xarray lock taken inside the zswap rb
> tree lock. The eventual goal is reducing the two locks back to
> one(xarray lock), which is not in my V1 patch. Your understanding is
> correct, the xarray still needs to have one lock for protecting the
> write path.

Hmm I don't understand. What's the point of keeping the rbtree if we
have the xarray? Doesn't it end up being more expensive and bug-prone
to maintain both trees?

When you say "eventual goal", do you mean what the patch would morph
into in later versions (as in v1 is just a proof of concept without
removing the rbtree), or follow up patches?