Re: [RFC PATCH v2 1/7] mm: zswap: add pool shrinking mechanism

[Domenico Cerasuolo <cerasuolodomenico@gmail.com>]

On Wed, Jun 7, 2023 at 10:14 AM Yosry Ahmed <yosryahmed@google.com> wrote:
>
> On Tue, Jun 6, 2023 at 7:56 AM Domenico Cerasuolo
> <cerasuolodomenico@gmail.com> wrote:
> >
> > Each zpool driver (zbud, z3fold and zsmalloc) implements its own shrink
> > function, which is called from zpool_shrink. However, with this commit,
> > a unified shrink function is added to zswap. The ultimate goal is to
> > eliminate the need for zpool_shrink once all zpool implementations have
> > dropped their shrink code.
> >
> > To ensure the functionality of each commit, this change focuses solely
> > on adding the mechanism itself. No modifications are made to
> > the backends, meaning that functionally, there are no immediate changes.
> > The zswap mechanism will only come into effect once the backends have
> > removed their shrink code. The subsequent commits will address the
> > modifications needed in the backends.
> >
> > Signed-off-by: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
> > ---
> >  mm/zswap.c | 96 +++++++++++++++++++++++++++++++++++++++++++++++++++---
> >  1 file changed, 91 insertions(+), 5 deletions(-)
> >
> > diff --git a/mm/zswap.c b/mm/zswap.c
> > index bcb82e09eb64..c99bafcefecf 100644
> > --- a/mm/zswap.c
> > +++ b/mm/zswap.c
> > @@ -150,6 +150,12 @@ struct crypto_acomp_ctx {
> >         struct mutex *mutex;
> >  };
> >
> > +/*
> > + * The lock ordering is zswap_tree.lock -> zswap_pool.lru_lock.
> > + * The only case where lru_lock is not acquired while holding tree.lock is
> > + * when a zswap_entry is taken off the lru for writeback, in that case it
> > + * needs to be verified that it's still valid in the tree.
> > + */
> >  struct zswap_pool {
> >         struct zpool *zpool;
> >         struct crypto_acomp_ctx __percpu *acomp_ctx;
> > @@ -159,6 +165,8 @@ struct zswap_pool {
> >         struct work_struct shrink_work;
> >         struct hlist_node node;
> >         char tfm_name[CRYPTO_MAX_ALG_NAME];
> > +       struct list_head lru;
> > +       spinlock_t lru_lock;
> >  };
> >
> >  /*
> > @@ -176,10 +184,12 @@ struct zswap_pool {
> >   *            be held while changing the refcount.  Since the lock must
> >   *            be held, there is no reason to also make refcount atomic.
> >   * length - the length in bytes of the compressed page data.  Needed during
> > - *          decompression. For a same value filled page length is 0.
> > + *          decompression. For a same value filled page length is 0, and both
> > + *          pool and lru are invalid and must be ignored.
> >   * pool - the zswap_pool the entry's data is in
> >   * handle - zpool allocation handle that stores the compressed page data
> >   * value - value of the same-value filled pages which have same content
> > + * lru - handle to the pool's lru used to evict pages.
> >   */
> >  struct zswap_entry {
> >         struct rb_node rbnode;
> > @@ -192,6 +202,7 @@ struct zswap_entry {
> >                 unsigned long value;
> >         };
> >         struct obj_cgroup *objcg;
> > +       struct list_head lru;
> >  };
> >
> >  struct zswap_header {
> > @@ -364,6 +375,12 @@ static void zswap_free_entry(struct zswap_entry *entry)
> >         if (!entry->length)
> >                 atomic_dec(&zswap_same_filled_pages);
> >         else {
> > +       /* zpool_evictable will be removed once all 3 backends have migrated */
> > +               if (!zpool_evictable(entry->pool->zpool)) {
> > +                       spin_lock(&entry->pool->lru_lock);
> > +                       list_del(&entry->lru);
> > +                       spin_unlock(&entry->pool->lru_lock);
> > +               }
> >                 zpool_free(entry->pool->zpool, entry->handle);
> >                 zswap_pool_put(entry->pool);
> >         }
> > @@ -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)
>
> Nit: rename to zswap_shrink_one() so that it's clear we always
> writeback one entry per call?

I named it like that to mirror zpool_shrink but I think that you've got a point
in that it might not be very clear that it is shrinking by one page only.
What about zswap_reclaim_entry? I'm not a native speaker, but with
zswap_shrink_one I wouldn't obviously intend that the "one" refers to an
entry.

>
> > +{
> > +       struct zswap_entry *lru_entry, *tree_entry = 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 = list_last_entry(&pool->lru, struct zswap_entry, lru);
> > +       list_del_init(&lru_entry->lru);
> > +       zhdr = zpool_map_handle(pool->zpool, lru_entry->handle, ZPOOL_MM_RO);
> > +       tree = 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
>
> Nit: lru lock*
>
> > +        * 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 = swp_offset(zhdr->swpentry);
> > +       spin_unlock(&pool->lru_lock);
> > +
> > +       /* hold a reference from tree so it won't be freed during writeback */
> > +       spin_lock(&tree->lock);
> > +       tree_entry = zswap_entry_find_get(&tree->rbroot, swpoffset);
> > +       if (tree_entry != lru_entry) {
> > +               if (tree_entry)
> > +                       zswap_entry_put(tree, tree_entry);
> > +               spin_unlock(&tree->lock);
> > +               return -EAGAIN;
> > +       }
> > +       spin_unlock(&tree->lock);
> > +
> > +       ret = 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);
> > +       spin_unlock(&tree->lock);
> > +
> > +       return ret ? -EAGAIN : 0;
> > +}
> > +
> >  static void shrink_worker(struct work_struct *w)
> >  {
> >         struct zswap_pool *pool = container_of(w, typeof(*pool),
> >                                                 shrink_work);
> >         int ret, failures = 0;
> >
> > +       /* zpool_evictable will be removed once all 3 backends have migrated */
> >         do {
> > -               ret = zpool_shrink(pool->zpool, 1, NULL);
> > +               if (zpool_evictable(pool->zpool))
> > +                       ret = zpool_shrink(pool->zpool, 1, NULL);
> > +               else
> > +                       ret = zswap_shrink(pool);
> >                 if (ret) {
> >                         zswap_reject_reclaim_fail++;
> >                         if (ret != -EAGAIN)
> > @@ -655,6 +728,8 @@ static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
> >          */
> >         kref_init(&pool->kref);
> >         INIT_LIST_HEAD(&pool->list);
> > +       INIT_LIST_HEAD(&pool->lru);
> > +       spin_lock_init(&pool->lru_lock);
> >         INIT_WORK(&pool->shrink_work, shrink_worker);
> >
> >         zswap_pool_debug("created", pool);
> > @@ -1270,7 +1345,7 @@ static int zswap_frontswap_store(unsigned type, pgoff_t offset,
> >         }
> >
> >         /* store */
> > -       hlen = zpool_evictable(entry->pool->zpool) ? sizeof(zhdr) : 0;
> > +       hlen = sizeof(zhdr);
> >         gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
> >         if (zpool_malloc_support_movable(entry->pool->zpool))
> >                 gfp |= __GFP_HIGHMEM | __GFP_MOVABLE;
> > @@ -1313,6 +1388,12 @@ static int zswap_frontswap_store(unsigned type, pgoff_t offset,
> >                         zswap_entry_put(tree, dupentry);
> >                 }
> >         } while (ret == -EEXIST);
> > +       /* zpool_evictable will be removed once all 3 backends have migrated */
> > +       if (entry->length && !zpool_evictable(entry->pool->zpool)) {
> > +               spin_lock(&entry->pool->lru_lock);
> > +               list_add(&entry->lru, &entry->pool->lru);
> > +               spin_unlock(&entry->pool->lru_lock);
> > +       }
> >         spin_unlock(&tree->lock);
> >
> >         /* update stats */
> > @@ -1384,8 +1465,7 @@ static int zswap_frontswap_load(unsigned type, pgoff_t offset,
> >         /* decompress */
> >         dlen = PAGE_SIZE;
> >         src = zpool_map_handle(entry->pool->zpool, entry->handle, ZPOOL_MM_RO);
> > -       if (zpool_evictable(entry->pool->zpool))
> > -               src += sizeof(struct zswap_header);
> > +       src += sizeof(struct zswap_header);
> >
> >         if (!zpool_can_sleep_mapped(entry->pool->zpool)) {
> >                 memcpy(tmp, src, entry->length);
> > @@ -1415,6 +1495,12 @@ static int zswap_frontswap_load(unsigned type, pgoff_t offset,
> >  freeentry:
> >         spin_lock(&tree->lock);
> >         zswap_entry_put(tree, entry);
> > +       /* zpool_evictable will be removed once all 3 backends have migrated */
> > +       if (entry->length && !zpool_evictable(entry->pool->zpool)) {
> > +               spin_lock(&entry->pool->lru_lock);
> > +               list_move(&entry->lru, &entry->pool->lru);
> > +               spin_unlock(&entry->pool->lru_lock);
> > +       }
>
> It's not really this patch's fault, but when merged with commit
> fe1d1f7d0fb5 ("mm: zswap: support exclusive loads") from mm-unstable
> [1], and with CONFIG_ZSWAP_EXCLUSIVE_LOADS=y, this causes a crash.
>
> This happens because fe1d1f7d0fb5 makes the loads exclusive, so
> zswap_entry_put(tree, entry) above the added code causes the entry to
> be freed, then we go ahead and deference multiple fields within it in
> the added chunk. Moving the chunk above zswap_entry_put() (and
> consequently also above zswap_invalidate_entry() from fe1d1f7d0fb5)
> makes this work correctly.
>
> Perhaps it would be useful to rebase on top of fe1d1f7d0fb5 for your
> next version(s), if any.

Will definitely rebase, I just now saw that you tested the suggested resolution
below, thanks, it does make sense.

>
> Maybe the outcome would be something like:
>
> zswap_entry_put(tree, entry);
> if (!ret && IS_ENABLED(CONFIG_ZSWAP_EXCLUSIVE_LOADS)) {
>         zswap_invalidate_entry(tree, entry);
> } else if (entry->length && !zpool_evictable(entry->pool->zpool)) {
>         spin_lock(&entry->pool->lru_lock);
>         list_move(&entry->lru, &entry->pool->lru);
>         spin_unlock(&entry->pool->lru_lock);
> }
>
> I am assuming if we are going to invalidate the entry anyway there is
> no need to move it to the front of the lru -- but I didn't really
> think it through.
>
> [1]https://lore.kernel.org/lkml/20230530210251.493194-1-yosryahmed@google.com/
>
> >         spin_unlock(&tree->lock);
> >
> >         return ret;
> > --
> > 2.34.1
> >