Re: [PATCH 2/2] mm: zswap.c: remove RB tree

[Yosry Ahmed <yosryahmed@google.com>]

On Thu, Jan 18, 2024 at 9:43 PM Chris Li <chrisl@kernel.org> wrote:
>
> Hi Yosry,
>
> On Wed, Jan 17, 2024 at 10:35 PM Yosry Ahmed <yosryahmed@google.com> wrote:
> >
> > > @@ -493,45 +471,47 @@ static struct zswap_entry *zswap_search(struct zswap_tree *tree, pgoff_t offset)
> > >  static int zswap_insert(struct zswap_tree *tree, struct zswap_entry *entry,
> > >                         struct zswap_entry **dupentry)
> > >  {
> > > -       struct rb_root *root = &tree->rbroot;
> > > -       struct rb_node **link = &root->rb_node, *parent = NULL;
> > > -       struct zswap_entry *myentry, *old;
> > > -       pgoff_t myentry_offset, entry_offset = swp_offset(entry->swpentry);
> > > -
> > > -
> > > -       while (*link) {
> > > -               parent = *link;
> > > -               myentry = rb_entry(parent, struct zswap_entry, rbnode);
> > > -               myentry_offset = swp_offset(myentry->swpentry);
> > > -               if (myentry_offset > entry_offset)
> > > -                       link = &(*link)->rb_left;
> > > -               else if (myentry_offset < entry_offset)
> > > -                       link = &(*link)->rb_right;
> > > -               else {
> > > -                       old = xa_load(&tree->xarray, entry_offset);
> > > -                       BUG_ON(old != myentry);
> > > -                       *dupentry = myentry;
> > > +       struct zswap_entry *e;
> > > +       pgoff_t offset = swp_offset(entry->swpentry);
> > > +       XA_STATE(xas, &tree->xarray, offset);
> > > +
> > > +       do {
> > > +               xas_lock_irq(&xas);
> > > +               do {
> > > +                       e = xas_load(&xas);
> > > +                       if (xa_is_zero(e))
> > > +                               e = NULL;
> > > +               } while (xas_retry(&xas, e));
> > > +               if (xas_valid(&xas) && e) {
> > > +                       xas_unlock_irq(&xas);
> > > +                       *dupentry = e;
> > >                         return -EEXIST;
> > >                 }
> > > -       }
> > > -       rb_link_node(&entry->rbnode, parent, link);
> > > -       rb_insert_color(&entry->rbnode, root);
> > > -       old = xa_store(&tree->xarray, entry_offset, entry, GFP_KERNEL);
> > > -       return 0;
> > > +               xas_store(&xas, entry);
> > > +               xas_unlock_irq(&xas);
> > > +       } while (xas_nomem(&xas, GFP_KERNEL));
> > > +       return xas_error(&xas);
> >
> > I think using the xas_* APIs can be avoided here. The only reason we
> > need it is that we want to check if there's an existing entry first,
> > and return -EEXIST. However, in that case, the caller will replace it
> > anyway (and do some operations on the dupentry):
>
> We might be able to for the insert case if we don't mind changing the
> code behavior a bit. My original intent is to keep close to the
> original zswap code and not stir the pot too much for the xarray
> replacement. We can always make more adjustment once the RB tree is
> gone.

I don't see how this changes code behavior though. The current code in
zswap_store() will do the following:

- Hold the tree lock to make sure no one else modifies it.
- Try to insert, check if there is already a dupentry at the index and
return -EEXIST.
- Warn, increment zswap_duplicate_entry, and invalidate the dupentry.
- Try to insert again (this should always succeed since we are holding
the lock).

What I am proposing is:
- zswap_xa_insert() is a thin wrapper around xa_store() (or we can
remove it completely).
- zswap_store() does the following:
  - Use zswap_xa_insert() and check if there is a returned dupentry.
  - Warn, increment zswap_duplicate_entry, and invalidate the dupentry.

Either way, we always place the entry we have in the tree, and if
there is a dupentry we warn and invalidate it. If anything, the latter
is more straightforward.

Am I missing something?

> >
> > >  }
> > >
> > >  static bool zswap_erase(struct zswap_tree *tree, struct zswap_entry *entry)
> > >  {
> > > +       struct zswap_entry *e;
> > >         pgoff_t offset = swp_offset(entry->swpentry);
> > > -       if (!RB_EMPTY_NODE(&entry->rbnode)) {
> > > -               struct zswap_entry *old;
> > > -               old = xa_erase(&tree->xarray, offset);
> > > -               BUG_ON(old != entry);
> > > -               rb_erase(&entry->rbnode, &tree->rbroot);
> > > -               RB_CLEAR_NODE(&entry->rbnode);
> > > -               return true;
> > > -       }
> > > -       return false;
> > > +       XA_STATE(xas, &tree->xarray, offset);
> > > +
> > > +       do {
> > > +               xas_lock_irq(&xas);
> > > +               do {
> > > +                       e = xas_load(&xas);
> > > +               } while (xas_retry(&xas, e));
> > > +               if (xas_valid(&xas) && e != entry) {
> > > +                       xas_unlock_irq(&xas);
> > > +                       return false;
> > > +               }
> > > +               xas_store(&xas, NULL);
> > > +               xas_unlock_irq(&xas);
> > > +       } while (xas_nomem(&xas, GFP_KERNEL));
> > > +       return !xas_error(&xas);
> > >  }
> >
> > Same here, I think we just want:
> >
> > return !!xa_erase(..);
>
> For the erase case it is tricky.
> The current zswap code does not erase an entry if the tree entry at
> the same offset has been changed. It should be fine if the new entry
> is NULL. Basically some race to remove the entry already. However, if
> the entry is not NULL, then force resetting it to NULL will change
> behavior compared to the current.

I see, very good point. I think we can use xa_cmpxchg() and pass in NULL?