Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <20230606145611.704392-1-cerasuolodomenico@gmail.com>
 <20230606145611.704392-2-cerasuolodomenico@gmail.com> <20230608165250.GG352940@cmpxchg.org>
 <20230608170459.GH352940@cmpxchg.org> <20230608184516.GA356779@cmpxchg.org>
In-Reply-To: <20230608184516.GA356779@cmpxchg.org>
From:   Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Date:   Fri, 9 Jun 2023 10:39:55 +0200
Message-ID: <CA+CLi1gP6+tKrHSfUxBhLvYu=F7NMnuPp+gt-63cwonU4r25UA@mail.gmail.com>
Subject: Re: [RFC PATCH v2 1/7] mm: zswap: add pool shrinking mechanism
To:     Johannes Weiner <hannes@cmpxchg.org>
Cc:     vitaly.wool@konsulko.com, minchan@kernel.org,
        senozhatsky@chromium.org, yosryahmed@google.com,
        linux-mm@kvack.org, ddstreet@ieee.org, sjenning@redhat.com,
        nphamcs@gmail.com, akpm@linux-foundation.org,
        linux-kernel@vger.kernel.org, kernel-team@meta.com
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Thu, Jun 8, 2023 at 8:45=E2=80=AFPM Johannes Weiner <hannes@cmpxchg.org>=
 wrote:
>
> On Thu, Jun 08, 2023 at 01:05:00PM -0400, Johannes Weiner wrote:
> > On Thu, Jun 08, 2023 at 12:52:51PM -0400, Johannes Weiner wrote:
> > > On Tue, Jun 06, 2023 at 04:56:05PM +0200, Domenico Cerasuolo wrote:
> > > > @@ -584,14 +601,70 @@ static struct zswap_pool *zswap_pool_find_get=
(char *type, char *compressor)
> > > >   return NULL;
> > > >  }
> > > >
> > > > +static int zswap_shrink(struct zswap_pool *pool)
> > > > +{
> > > > + struct zswap_entry *lru_entry, *tree_entry =3D NULL;
> > > > + struct zswap_header *zhdr;
> > > > + struct zswap_tree *tree;
> > > > + int swpoffset;
> > > > + int ret;
> > > > +
> > > > + /* get a reclaimable entry from LRU */
> > > > + spin_lock(&pool->lru_lock);
> > > > + if (list_empty(&pool->lru)) {
> > > > +         spin_unlock(&pool->lru_lock);
> > > > +         return -EINVAL;
> > > > + }
> > > > + lru_entry =3D list_last_entry(&pool->lru, struct zswap_entry, lru=
);
> > > > + list_del_init(&lru_entry->lru);
> > > > + zhdr =3D zpool_map_handle(pool->zpool, lru_entry->handle, ZPOOL_M=
M_RO);
> > > > + tree =3D zswap_trees[swp_type(zhdr->swpentry)];
> > > > + zpool_unmap_handle(pool->zpool, lru_entry->handle);
> > > > + /*
> > > > +  * Once the pool lock is dropped, the lru_entry might get freed. =
The
> > > > +  * swpoffset is copied to the stack, and lru_entry isn't deref'd =
again
> > > > +  * until the entry is verified to still be alive in the tree.
> > > > +  */
> > > > + swpoffset =3D swp_offset(zhdr->swpentry);
> > > > + spin_unlock(&pool->lru_lock);
> > > > +
> > > > + /* hold a reference from tree so it won't be freed during writeba=
ck */
> > > > + spin_lock(&tree->lock);
> > > > + tree_entry =3D zswap_entry_find_get(&tree->rbroot, swpoffset);
> > > > + if (tree_entry !=3D lru_entry) {
> > > > +         if (tree_entry)
> > > > +                 zswap_entry_put(tree, tree_entry);
> > > > +         spin_unlock(&tree->lock);
> > > > +         return -EAGAIN;
> > > > + }
> > > > + spin_unlock(&tree->lock);
> > > > +
> > > > + ret =3D zswap_writeback_entry(pool->zpool, lru_entry->handle);
> > > > +
> > > > + spin_lock(&tree->lock);
> > > > + if (ret) {
> > > > +         spin_lock(&pool->lru_lock);
> > > > +         list_move(&lru_entry->lru, &pool->lru);
> > > > +         spin_unlock(&pool->lru_lock);
> > > > + }
> > > > + zswap_entry_put(tree, tree_entry);
> > >
> > > On re-reading this, I find the lru_entry vs tree_entry distinction
> > > unnecessarily complicated. Once it's known that the thing coming off
> > > the LRU is the same thing as in the tree, there is only "the entry".
> > >
> > > How about 'entry' and 'tree_entry', and after validation use 'entry'
> > > throughout the rest of the function?
> >
> > Even better, safe the tree_entry entirely by getting the reference
> > from the LRU already, and then just search the tree for a match:
> >
> >       /* Get an entry off the LRU */
> >       spin_lock(&pool->lru_lock);
> >       entry =3D list_last_entry();
> >       list_del(&entry->lru);
> >       zswap_entry_get(entry);
> >       spin_unlock(&pool->lru_lock);
> >
> >       /* Check for invalidate() race */
> >       spin_lock(&tree->lock);
> >       if (entry !=3D zswap_rb_search(&tree->rbroot, swpoffset)) {
> >               ret =3D -EAGAIN;
> >               goto put_unlock;
> >       }
> >       spin_unlock(&tree->lock);
>
> Eh, brainfart. It needs the tree lock to bump the ref, of course.
>
> But this should work, right?
>
>         /* Check for invalidate() race */
>         spin_lock(&tree->lock);
>         if (entry !=3D zswap_rb_search(&tree->rbroot, swpoffset)) {
>                 ret =3D -EAGAIN;
>                 goto unlock;
>         }
>         zswap_entry_get(entry);
>         spin_unlock(&tree->lock);

This should work indeed, it's much cleaner with just one local
zswap_entry, will update!