[v2,2/2] zswap: shrinks zswap pool based on memory pressure

  Currently, we only shrink the zswap pool when the user-defined limit is
hit. This means that if we set the limit too high, cold data that are
unlikely to be used again will reside in the pool, wasting precious
memory. It is hard to predict how much zswap space will be needed ahead
of time, as this depends on the workload (specifically, on factors such
as memory access patterns and compressibility of the memory pages).

This patch implements a memcg- and NUMA-aware shrinker for zswap, that
is initiated when there is memory pressure. The shrinker does not
have any parameter that must be tuned by the user, and can be opted in
or out on a per-memcg basis.

Furthermore, to make it more robust for many workloads and prevent
overshrinking (i.e evicting warm pages that might be refaulted into
memory), we build in the following heuristics:

* Estimate the number of warm pages residing in zswap, and attempt to
  protect this region of the zswap LRU.
* Scale the number of freeable objects by an estimate of the memory
  saving factor. The better zswap compresses the data, the fewer pages
  we will evict to swap (as we will otherwise incur IO for relatively
  small memory saving).
* During reclaim, if the shrinker encounters a page that is also being
  brought into memory, the shrinker will cautiously terminate its
  shrinking action, as this is a sign that it is touching the warmer
  region of the zswap LRU.

On a benchmark that we have run:

(without the shrinker)
real -- mean: 153.27s, median: 153.199s
sys -- mean: 541.652s, median: 541.903s
user -- mean: 4384.9673999999995s, median: 4385.471s

(with the shrinker)
real -- mean: 151.4956s, median: 151.456s
sys -- mean: 461.14639999999997s, median: 465.656s
user -- mean: 4384.7118s, median: 4384.675s

We observed a 14-15% reduction in kernel CPU time, which translated to
over 1% reduction in real time.

On another benchmark, where there was a lot more cold memory residing in
zswap, we observed even more pronounced gains:

(without the shrinker)
real -- mean: 157.52519999999998s, median: 157.281s
sys -- mean: 769.3082s, median: 780.545s
user -- mean: 4378.1622s, median: 4378.286s

(with the shrinker)
real -- mean: 152.9608s, median: 152.845s
sys -- mean: 517.4446s, median: 506.749s
user -- mean: 4387.694s, median: 4387.935s

Here, we saw around 32-35% reduction in kernel CPU time, which
translated to 2.8% reduction in real time. These results confirm our
hypothesis that the shrinker is more helpful the more cold memory we
have.

Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Nhat Pham <nphamcs@gmail.com>
---
 Documentation/admin-guide/mm/zswap.rst |  12 ++
 include/linux/memcontrol.h             |   1 +
 include/linux/mmzone.h                 |  14 ++
 mm/memcontrol.c                        |  33 +++++
 mm/swap_state.c                        |  31 ++++-
 mm/zswap.c                             | 180 ++++++++++++++++++++++++-
 6 files changed, 263 insertions(+), 8 deletions(-)
  

On Tue, Sep 19, 2023 at 10:14 AM Nhat Pham <nphamcs@gmail.com> wrote:
>
> Currently, we only shrink the zswap pool when the user-defined limit is
> hit. This means that if we set the limit too high, cold data that are
> unlikely to be used again will reside in the pool, wasting precious
> memory. It is hard to predict how much zswap space will be needed ahead
> of time, as this depends on the workload (specifically, on factors such
> as memory access patterns and compressibility of the memory pages).
>
> This patch implements a memcg- and NUMA-aware shrinker for zswap, that
> is initiated when there is memory pressure. The shrinker does not
> have any parameter that must be tuned by the user, and can be opted in
> or out on a per-memcg basis.

What's the use case for having per-memcg opt-in/out?

If there is memory pressure, reclaiming swap-backed pages will push
pages out of zswap anyway, regardless of this patch. With this patch,
any sort of reclaim can push pages out of zswap. Wouldn't that be
preferable to reclaiming memory that is currently resident in memory
(so arguably hotter than the pages in zswap)? Why would this decision
be different per-memcg?

>
> Furthermore, to make it more robust for many workloads and prevent
> overshrinking (i.e evicting warm pages that might be refaulted into
> memory), we build in the following heuristics:
>
> * Estimate the number of warm pages residing in zswap, and attempt to
>   protect this region of the zswap LRU.
> * Scale the number of freeable objects by an estimate of the memory
>   saving factor. The better zswap compresses the data, the fewer pages
>   we will evict to swap (as we will otherwise incur IO for relatively
>   small memory saving).
> * During reclaim, if the shrinker encounters a page that is also being
>   brought into memory, the shrinker will cautiously terminate its
>   shrinking action, as this is a sign that it is touching the warmer
>   region of the zswap LRU.

I don't have an opinion about the reclaim heuristics here, I will let
reclaim experts chip in.

>
> On a benchmark that we have run:

Please add more details (as much as possible) about the benchmarks used here.

>
> (without the shrinker)
> real -- mean: 153.27s, median: 153.199s
> sys -- mean: 541.652s, median: 541.903s
> user -- mean: 4384.9673999999995s, median: 4385.471s
>
> (with the shrinker)
> real -- mean: 151.4956s, median: 151.456s
> sys -- mean: 461.14639999999997s, median: 465.656s
> user -- mean: 4384.7118s, median: 4384.675s
>
> We observed a 14-15% reduction in kernel CPU time, which translated to
> over 1% reduction in real time.
>
> On another benchmark, where there was a lot more cold memory residing in
> zswap, we observed even more pronounced gains:
>
> (without the shrinker)
> real -- mean: 157.52519999999998s, median: 157.281s
> sys -- mean: 769.3082s, median: 780.545s
> user -- mean: 4378.1622s, median: 4378.286s
>
> (with the shrinker)
> real -- mean: 152.9608s, median: 152.845s
> sys -- mean: 517.4446s, median: 506.749s
> user -- mean: 4387.694s, median: 4387.935s
>
> Here, we saw around 32-35% reduction in kernel CPU time, which
> translated to 2.8% reduction in real time. These results confirm our
> hypothesis that the shrinker is more helpful the more cold memory we
> have.
>
> Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
> Signed-off-by: Nhat Pham <nphamcs@gmail.com>
> ---
>  Documentation/admin-guide/mm/zswap.rst |  12 ++
>  include/linux/memcontrol.h             |   1 +
>  include/linux/mmzone.h                 |  14 ++
>  mm/memcontrol.c                        |  33 +++++
>  mm/swap_state.c                        |  31 ++++-
>  mm/zswap.c                             | 180 ++++++++++++++++++++++++-
>  6 files changed, 263 insertions(+), 8 deletions(-)
>
> diff --git a/Documentation/admin-guide/mm/zswap.rst b/Documentation/admin-guide/mm/zswap.rst
> index 45b98390e938..ae8597a67804 100644
> --- a/Documentation/admin-guide/mm/zswap.rst
> +++ b/Documentation/admin-guide/mm/zswap.rst
> @@ -153,6 +153,18 @@ attribute, e. g.::
>
>  Setting this parameter to 100 will disable the hysteresis.
>
> +When there is a sizable amount of cold memory residing in the zswap pool, it
> +can be advantageous to proactively write these cold pages to swap and reclaim
> +the memory for other use cases. By default, the zswap shrinker is disabled.
> +User can enable it by first switching on the global knob:
> +
> +  echo Y > /sys/module/zswap/par meters/shrinker_enabled
> +
> +When the kernel is compiled with CONFIG_MEMCG_KMEM, user needs to further turn
> +it on for each cgroup that the shrinker should target:
> +
> +  echo 1 > /sys/fs/cgroup/<cgroup-name>/memory.zswap.shrinker.enabled
> +
>  A debugfs interface is provided for various statistic about pool size, number
>  of pages stored, same-value filled pages and various counters for the reasons
>  pages are rejected.
> diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
> index 05d34b328d9d..f005ea667863 100644
> --- a/include/linux/memcontrol.h
> +++ b/include/linux/memcontrol.h
> @@ -219,6 +219,7 @@ struct mem_cgroup {
>
>  #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
>         unsigned long zswap_max;
> +       atomic_t zswap_shrinker_enabled;
>  #endif
>
>         unsigned long soft_limit;
> diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
> index 4106fbc5b4b3..81f4c5ea3e16 100644
> --- a/include/linux/mmzone.h
> +++ b/include/linux/mmzone.h
> @@ -637,6 +637,20 @@ struct lruvec {
>  #ifdef CONFIG_MEMCG
>         struct pglist_data *pgdat;
>  #endif
> +#ifdef CONFIG_ZSWAP
> +       /*
> +        * Number of pages in zswap that should be protected from the shrinker.
> +        * This number is an estimate of the following counts:
> +        *
> +        * a) Recent page faults.
> +        * b) Recent insertion to the zswap LRU. This includes new zswap stores,
> +        *    as well as recent zswap LRU rotations.
> +        *
> +        * These pages are likely to be warm, and might incur IO if the are written
> +        * to swap.
> +        */
> +       unsigned long nr_zswap_protected;
> +#endif

Would this be better abstracted in a zswap lruvec struct?

>  };
>
>  /* Isolate unmapped pages */
> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index 9f84b3f7b469..1a2c97cf396f 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -5352,6 +5352,8 @@ mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
>         WRITE_ONCE(memcg->soft_limit, PAGE_COUNTER_MAX);
>  #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
>         memcg->zswap_max = PAGE_COUNTER_MAX;
> +       /* Disable the shrinker by default */
> +       atomic_set(&memcg->zswap_shrinker_enabled, 0);
>  #endif
>         page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
>         if (parent) {
> @@ -7877,6 +7879,31 @@ static ssize_t zswap_max_write(struct kernfs_open_file *of,
>         return nbytes;
>  }
>
> +static int zswap_shrinker_enabled_show(struct seq_file *m, void *v)
> +{
> +       struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
> +
> +       seq_printf(m, "%d\n", atomic_read(&memcg->zswap_shrinker_enabled));
> +       return 0;
> +}
> +
> +static ssize_t zswap_shrinker_enabled_write(struct kernfs_open_file *of,
> +                              char *buf, size_t nbytes, loff_t off)
> +{
> +       struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
> +       int zswap_shrinker_enabled;
> +       ssize_t parse_ret = kstrtoint(strstrip(buf), 0, &zswap_shrinker_enabled);
> +
> +       if (parse_ret)
> +               return parse_ret;
> +
> +       if (zswap_shrinker_enabled < 0 || zswap_shrinker_enabled > 1)
> +               return -ERANGE;
> +
> +       atomic_set(&memcg->zswap_shrinker_enabled, zswap_shrinker_enabled);
> +       return nbytes;
> +}
> +
>  static struct cftype zswap_files[] = {
>         {
>                 .name = "zswap.current",
> @@ -7889,6 +7916,12 @@ static struct cftype zswap_files[] = {
>                 .seq_show = zswap_max_show,
>                 .write = zswap_max_write,
>         },
> +       {
> +               .name = "zswap.shrinker.enabled",
> +               .flags = CFTYPE_NOT_ON_ROOT,
> +               .seq_show = zswap_shrinker_enabled_show,
> +               .write = zswap_shrinker_enabled_write,
> +       },
>         { }     /* terminate */
>  };
>  #endif /* CONFIG_MEMCG_KMEM && CONFIG_ZSWAP */
> diff --git a/mm/swap_state.c b/mm/swap_state.c
> index 1c826737aacb..788e36a06c34 100644
> --- a/mm/swap_state.c
> +++ b/mm/swap_state.c
> @@ -618,6 +618,22 @@ static unsigned long swapin_nr_pages(unsigned long offset)
>         return pages;
>  }
>
> +#ifdef CONFIG_ZSWAP
> +/*
> + * Refault is an indication that warmer pages are not resident in memory.
> + * Increase the size of zswap's protected area.
> + */
> +static void inc_nr_protected(struct page *page)
> +{
> +       struct lruvec *lruvec = folio_lruvec(page_folio(page));
> +       unsigned long flags;
> +
> +       spin_lock_irqsave(&lruvec->lru_lock, flags);
> +       lruvec->nr_zswap_protected++;
> +       spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> +}
> +#endif
> +

A few questions:
- Why is this function named in such a generic way?
- Why is this function here instead of in mm/zswap.c?
- Why is this protected by the heavily contested lruvec lock instead
of being an atomic?

>  /**
>   * swap_cluster_readahead - swap in pages in hope we need them soon
>   * @entry: swap entry of this memory
> @@ -686,7 +702,12 @@ struct page *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask,
>         lru_add_drain();        /* Push any new pages onto the LRU now */
>  skip:
>         /* The page was likely read above, so no need for plugging here */
> -       return read_swap_cache_async(entry, gfp_mask, vma, addr, NULL);
> +       page = read_swap_cache_async(entry, gfp_mask, vma, addr, NULL);
> +#ifdef CONFIG_ZSWAP
> +       if (page)
> +               inc_nr_protected(page);
> +#endif
> +       return page;
>  }
>
>  int init_swap_address_space(unsigned int type, unsigned long nr_pages)
> @@ -853,8 +874,12 @@ static struct page *swap_vma_readahead(swp_entry_t fentry, gfp_t gfp_mask,
>         lru_add_drain();
>  skip:
>         /* The page was likely read above, so no need for plugging here */
> -       return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address,
> -                                    NULL);
> +       page = read_swap_cache_async(fentry, gfp_mask, vma, vmf->address, NULL);
> +#ifdef CONFIG_ZSWAP
> +       if (page)
> +               inc_nr_protected(page);
> +#endif
> +       return page;
>  }
>
>  /**
> diff --git a/mm/zswap.c b/mm/zswap.c
> index 1a469e5d5197..79cb18eeb8bf 100644
> --- a/mm/zswap.c
> +++ b/mm/zswap.c
> @@ -145,6 +145,26 @@ module_param_named(exclusive_loads, zswap_exclusive_loads_enabled, bool, 0644);
>  /* Number of zpools in zswap_pool (empirically determined for scalability) */
>  #define ZSWAP_NR_ZPOOLS 32
>
> +/*
> + * Global flag to enable/disable memory pressure-based shrinker for all memcgs.
> + * If CONFIG_MEMCG_KMEM is on, we can further selectively disable
> + * the shrinker for each memcg.
> + */
> +static bool zswap_shrinker_enabled;
> +module_param_named(shrinker_enabled, zswap_shrinker_enabled, bool, 0644);
> +#ifdef CONFIG_MEMCG_KMEM
> +static bool is_shrinker_enabled(struct mem_cgroup *memcg)
> +{
> +       return zswap_shrinker_enabled &&
> +               atomic_read(&memcg->zswap_shrinker_enabled);
> +}
> +#else
> +static bool is_shrinker_enabled(struct mem_cgroup *memcg)
> +{
> +       return zswap_shrinker_enabled;
> +}
> +#endif
> +
>  /*********************************
>  * data structures
>  **********************************/
> @@ -174,6 +194,8 @@ struct zswap_pool {
>         char tfm_name[CRYPTO_MAX_ALG_NAME];
>         struct list_lru list_lru;
>         struct mem_cgroup *next_shrink;
> +       struct shrinker *shrinker;
> +       atomic_t nr_stored;
>  };
>
>  /*
> @@ -273,17 +295,26 @@ static bool zswap_can_accept(void)
>                         DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
>  }
>
> +static u64 get_zswap_pool_size(struct zswap_pool *pool)
> +{
> +       u64 pool_size = 0;
> +       int i;
> +
> +       for (i = 0; i < ZSWAP_NR_ZPOOLS; i++)
> +               pool_size += zpool_get_total_size(pool->zpools[i]);
> +
> +       return pool_size;
> +}
> +
>  static void zswap_update_total_size(void)
>  {
>         struct zswap_pool *pool;
>         u64 total = 0;
> -       int i;
>
>         rcu_read_lock();
>
>         list_for_each_entry_rcu(pool, &zswap_pools, list)
> -               for (i = 0; i < ZSWAP_NR_ZPOOLS; i++)
> -                       total += zpool_get_total_size(pool->zpools[i]);
> +               total += get_zswap_pool_size(pool);
>
>         rcu_read_unlock();
>
> @@ -318,8 +349,23 @@ static bool zswap_lru_add(struct list_lru *list_lru, struct zswap_entry *entry)
>  {
>         struct mem_cgroup *memcg = entry->objcg ?
>                 get_mem_cgroup_from_objcg(entry->objcg) : NULL;
> +       struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(entry->nid));
>         bool added = __list_lru_add(list_lru, &entry->lru, entry->nid, memcg);
> +       unsigned long flags, lru_size;
> +
> +       if (added) {
> +               lru_size = list_lru_count_one(list_lru, entry->nid, memcg);
> +               spin_lock_irqsave(&lruvec->lru_lock, flags);
> +               lruvec->nr_zswap_protected++;
>
> +               /*
> +                * Decay to avoid overflow and adapt to changing workloads.
> +                * This is based on LRU reclaim cost decaying heuristics.
> +                */
> +               if (lruvec->nr_zswap_protected > lru_size / 4)
> +                       lruvec->nr_zswap_protected /= 2;
> +               spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> +       }
>         mem_cgroup_put(memcg);
>         return added;
>  }
> @@ -420,6 +466,7 @@ static void zswap_free_entry(struct zswap_entry *entry)
>         else {
>                 zswap_lru_del(&entry->pool->list_lru, entry);
>                 zpool_free(zswap_find_zpool(entry), entry->handle);
> +               atomic_dec(&entry->pool->nr_stored);
>                 zswap_pool_put(entry->pool);
>         }
>         zswap_entry_cache_free(entry);
> @@ -461,6 +508,98 @@ static struct zswap_entry *zswap_entry_find_get(struct rb_root *root,
>         return entry;
>  }
>
> +/*********************************
> +* shrinker functions
> +**********************************/
> +static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_one *l,
> +                                      spinlock_t *lock, void *arg);
> +
> +static unsigned long zswap_shrinker_scan(struct shrinker *shrinker,
> +               struct shrink_control *sc)
> +{
> +       struct zswap_pool *pool = shrinker->private_data;
> +       unsigned long shrink_ret, nr_zswap_protected, flags,
> +               lru_size = list_lru_shrink_count(&pool->list_lru, sc);
> +       struct lruvec *lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid));
> +       bool encountered_page_in_swapcache = false;
> +
> +       spin_lock_irqsave(&lruvec->lru_lock, flags);
> +       nr_zswap_protected = lruvec->nr_zswap_protected;
> +       spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> +
> +       /*
> +        * Abort if the shrinker is disabled or if we are shrinking into the
> +        * protected region.
> +        */
> +       if (!is_shrinker_enabled(sc->memcg) ||
> +                       nr_zswap_protected >= lru_size - sc->nr_to_scan) {
> +               sc->nr_scanned = 0;
> +               return SHRINK_STOP;
> +       }
> +
> +       shrink_ret = list_lru_shrink_walk(&pool->list_lru, sc, &shrink_memcg_cb,
> +               &encountered_page_in_swapcache);
> +
> +       if (encountered_page_in_swapcache)
> +               return SHRINK_STOP;
> +
> +       return shrink_ret ? shrink_ret : SHRINK_STOP;
> +}
> +
> +static unsigned long zswap_shrinker_count(struct shrinker *shrinker,
> +               struct shrink_control *sc)
> +{
> +       struct zswap_pool *pool = shrinker->private_data;
> +       struct mem_cgroup *memcg = sc->memcg;
> +       struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(sc->nid));
> +       unsigned long nr_backing, nr_stored, nr_freeable, flags;
> +
> +#ifdef CONFIG_MEMCG_KMEM
> +       cgroup_rstat_flush(memcg->css.cgroup);
> +       nr_backing = memcg_page_state(memcg, MEMCG_ZSWAP_B) >> PAGE_SHIFT;
> +       nr_stored = memcg_page_state(memcg, MEMCG_ZSWAPPED);
> +#else
> +       /* use pool stats instead of memcg stats */
> +       nr_backing = get_zswap_pool_size(pool) >> PAGE_SHIFT;
> +       nr_stored = atomic_read(&pool->nr_stored);
> +#endif
> +
> +       if (!is_shrinker_enabled(memcg) || !nr_stored)
> +               return 0;
> +
> +       nr_freeable = list_lru_shrink_count(&pool->list_lru, sc);
> +       /*
> +        * Subtract the lru size by an estimate of the number of pages
> +        * that should be protected.
> +        */
> +       spin_lock_irqsave(&lruvec->lru_lock, flags);
> +       nr_freeable = nr_freeable > lruvec->nr_zswap_protected ?
> +               nr_freeable - lruvec->nr_zswap_protected : 0;
> +       spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> +
> +       /*
> +        * Scale the number of freeable pages by the memory saving factor.
> +        * This ensures that the better zswap compresses memory, the fewer
> +        * pages we will evict to swap (as it will otherwise incur IO for
> +        * relatively small memory saving).
> +        */
> +       return mult_frac(nr_freeable, nr_backing, nr_stored);
> +}
> +
> +static void zswap_alloc_shrinker(struct zswap_pool *pool)
> +{
> +       pool->shrinker =
> +               shrinker_alloc(SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE, "mm-zswap");
> +       if (!pool->shrinker)
> +               return;
> +
> +       pool->shrinker->private_data = pool;
> +       pool->shrinker->scan_objects = zswap_shrinker_scan;
> +       pool->shrinker->count_objects = zswap_shrinker_count;
> +       pool->shrinker->batch = 0;
> +       pool->shrinker->seeks = DEFAULT_SEEKS;
> +}
> +
>  /*********************************
>  * per-cpu code
>  **********************************/
> @@ -656,11 +795,14 @@ static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_o
>                                        spinlock_t *lock, void *arg)
>  {
>         struct zswap_entry *entry = container_of(item, struct zswap_entry, lru);
> +       bool *encountered_page_in_swapcache = (bool *)arg;
>         struct mem_cgroup *memcg;
>         struct zswap_tree *tree;
> +       struct lruvec *lruvec;
>         pgoff_t swpoffset;
>         enum lru_status ret = LRU_REMOVED_RETRY;
>         int writeback_result;
> +       unsigned long flags;
>
>         /*
>          * Once the lru lock is dropped, the entry might get freed. The
> @@ -696,8 +838,24 @@ static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_o
>                 /* we cannot use zswap_lru_add here, because it increments node's lru count */
>                 list_lru_putback(&entry->pool->list_lru, item, entry->nid, memcg);
>                 spin_unlock(lock);
> -               mem_cgroup_put(memcg);
>                 ret = LRU_RETRY;
> +
> +               /*
> +                * Encountering a page already in swap cache is a sign that we are shrinking
> +                * into the warmer region. We should terminate shrinking (if we're in the dynamic
> +                * shrinker context).
> +                */
> +               if (writeback_result == -EEXIST && encountered_page_in_swapcache) {
> +                       ret = LRU_SKIP;
> +                       *encountered_page_in_swapcache = true;
> +               }
> +               lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(entry->nid));
> +               spin_lock_irqsave(&lruvec->lru_lock, flags);
> +               /* Increment the protection area to account for the LRU rotation. */
> +               lruvec->nr_zswap_protected++;
> +               spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> +
> +               mem_cgroup_put(memcg);
>                 goto put_unlock;
>         }
>
> @@ -828,6 +986,11 @@ static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
>                                        &pool->node);
>         if (ret)
>                 goto error;
> +
> +       zswap_alloc_shrinker(pool);
> +       if (!pool->shrinker)
> +               goto error;
> +
>         pr_debug("using %s compressor\n", pool->tfm_name);
>
>         /* being the current pool takes 1 ref; this func expects the
> @@ -836,12 +999,17 @@ static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
>         kref_init(&pool->kref);
>         INIT_LIST_HEAD(&pool->list);
>         INIT_WORK(&pool->shrink_work, shrink_worker);
> -       list_lru_init_memcg(&pool->list_lru, NULL);
> +       if (list_lru_init_memcg(&pool->list_lru, pool->shrinker))
> +               goto lru_fail;
> +       shrinker_register(pool->shrinker);
>
>         zswap_pool_debug("created", pool);
>
>         return pool;
>
> +lru_fail:
> +       list_lru_destroy(&pool->list_lru);
> +       shrinker_free(pool->shrinker);
>  error:
>         if (pool->acomp_ctx)
>                 free_percpu(pool->acomp_ctx);
> @@ -899,6 +1067,7 @@ static void zswap_pool_destroy(struct zswap_pool *pool)
>
>         zswap_pool_debug("destroying", pool);
>
> +       shrinker_free(pool->shrinker);
>         cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
>         free_percpu(pool->acomp_ctx);
>         list_lru_destroy(&pool->list_lru);
> @@ -1431,6 +1600,7 @@ bool zswap_store(struct folio *folio)
>         if (entry->length) {
>                 INIT_LIST_HEAD(&entry->lru);
>                 zswap_lru_add(&pool->list_lru, entry);
> +               atomic_inc(&pool->nr_stored);
>         }
>         spin_unlock(&tree->lock);
>
> --
> 2.34.1
  

On Mon, Sep 25, 2023 at 1:38 PM Yosry Ahmed <yosryahmed@google.com> wrote:
>
> On Tue, Sep 19, 2023 at 10:14 AM Nhat Pham <nphamcs@gmail.com> wrote:
> >
> > Currently, we only shrink the zswap pool when the user-defined limit is
> > hit. This means that if we set the limit too high, cold data that are
> > unlikely to be used again will reside in the pool, wasting precious
> > memory. It is hard to predict how much zswap space will be needed ahead
> > of time, as this depends on the workload (specifically, on factors such
> > as memory access patterns and compressibility of the memory pages).
> >
> > This patch implements a memcg- and NUMA-aware shrinker for zswap, that
> > is initiated when there is memory pressure. The shrinker does not
> > have any parameter that must be tuned by the user, and can be opted in
> > or out on a per-memcg basis.
>
> What's the use case for having per-memcg opt-in/out?
>
> If there is memory pressure, reclaiming swap-backed pages will push
> pages out of zswap anyway, regardless of this patch. With this patch,
> any sort of reclaim can push pages out of zswap. Wouldn't that be
> preferable to reclaiming memory that is currently resident in memory
> (so arguably hotter than the pages in zswap)? Why would this decision
> be different per-memcg?
I'm not quite following your argument here. The point of having this
be done on a per-memcg basis is that we have different workloads
with different memory access pattern (and as a result, different memory
coldness distribution).

In a workload where there is a lot of cold data, we can really benefit
from reclaiming all of those pages and repurpose the memory reclaimed
(for e.g for filecache).

On the other hand, in a workload where there aren't a lot of cold data,
reclaiming its zswapped pages will at best do nothing (wasting CPU
cycles on compression/decompression), and at worst hurt performance
(due to increased IO when we need those writtenback pages again).

Such different workloads could co-exist in the same system, and having
a per-memcg knob allows us to crank on the shrinker only on workloads
where it makes sense.
>
> >
> > Furthermore, to make it more robust for many workloads and prevent
> > overshrinking (i.e evicting warm pages that might be refaulted into
> > memory), we build in the following heuristics:
> >
> > * Estimate the number of warm pages residing in zswap, and attempt to
> >   protect this region of the zswap LRU.
> > * Scale the number of freeable objects by an estimate of the memory
> >   saving factor. The better zswap compresses the data, the fewer pages
> >   we will evict to swap (as we will otherwise incur IO for relatively
> >   small memory saving).
> > * During reclaim, if the shrinker encounters a page that is also being
> >   brought into memory, the shrinker will cautiously terminate its
> >   shrinking action, as this is a sign that it is touching the warmer
> >   region of the zswap LRU.
>
> I don't have an opinion about the reclaim heuristics here, I will let
> reclaim experts chip in.
>
> >
> > On a benchmark that we have run:
>
> Please add more details (as much as possible) about the benchmarks used here.
Sure! I built the kernel in a memory-limited cgroup a couple times,
then measured the build time.

To simulate conditions where there are cold, unused data, I
also generated a bunch of data in tmpfs (and never touch them
again).
>
> >
> > (without the shrinker)
> > real -- mean: 153.27s, median: 153.199s
> > sys -- mean: 541.652s, median: 541.903s
> > user -- mean: 4384.9673999999995s, median: 4385.471s
> >
> > (with the shrinker)
> > real -- mean: 151.4956s, median: 151.456s
> > sys -- mean: 461.14639999999997s, median: 465.656s
> > user -- mean: 4384.7118s, median: 4384.675s
> >
> > We observed a 14-15% reduction in kernel CPU time, which translated to
> > over 1% reduction in real time.
> >
> > On another benchmark, where there was a lot more cold memory residing in
> > zswap, we observed even more pronounced gains:
> >
> > (without the shrinker)
> > real -- mean: 157.52519999999998s, median: 157.281s
> > sys -- mean: 769.3082s, median: 780.545s
> > user -- mean: 4378.1622s, median: 4378.286s
> >
> > (with the shrinker)
> > real -- mean: 152.9608s, median: 152.845s
> > sys -- mean: 517.4446s, median: 506.749s
> > user -- mean: 4387.694s, median: 4387.935s
> >
> > Here, we saw around 32-35% reduction in kernel CPU time, which
> > translated to 2.8% reduction in real time. These results confirm our
> > hypothesis that the shrinker is more helpful the more cold memory we
> > have.
> >
> > Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
> > Signed-off-by: Nhat Pham <nphamcs@gmail.com>
> > ---
> >  Documentation/admin-guide/mm/zswap.rst |  12 ++
> >  include/linux/memcontrol.h             |   1 +
> >  include/linux/mmzone.h                 |  14 ++
> >  mm/memcontrol.c                        |  33 +++++
> >  mm/swap_state.c                        |  31 ++++-
> >  mm/zswap.c                             | 180 ++++++++++++++++++++++++-
> >  6 files changed, 263 insertions(+), 8 deletions(-)
> >
> > diff --git a/Documentation/admin-guide/mm/zswap.rst b/Documentation/admin-guide/mm/zswap.rst
> > index 45b98390e938..ae8597a67804 100644
> > --- a/Documentation/admin-guide/mm/zswap.rst
> > +++ b/Documentation/admin-guide/mm/zswap.rst
> > @@ -153,6 +153,18 @@ attribute, e. g.::
> >
> >  Setting this parameter to 100 will disable the hysteresis.
> >
> > +When there is a sizable amount of cold memory residing in the zswap pool, it
> > +can be advantageous to proactively write these cold pages to swap and reclaim
> > +the memory for other use cases. By default, the zswap shrinker is disabled.
> > +User can enable it by first switching on the global knob:
> > +
> > +  echo Y > /sys/module/zswap/par meters/shrinker_enabled
> > +
> > +When the kernel is compiled with CONFIG_MEMCG_KMEM, user needs to further turn
> > +it on for each cgroup that the shrinker should target:
> > +
> > +  echo 1 > /sys/fs/cgroup/<cgroup-name>/memory.zswap.shrinker.enabled
> > +
> >  A debugfs interface is provided for various statistic about pool size, number
> >  of pages stored, same-value filled pages and various counters for the reasons
> >  pages are rejected.
> > diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
> > index 05d34b328d9d..f005ea667863 100644
> > --- a/include/linux/memcontrol.h
> > +++ b/include/linux/memcontrol.h
> > @@ -219,6 +219,7 @@ struct mem_cgroup {
> >
> >  #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
> >         unsigned long zswap_max;
> > +       atomic_t zswap_shrinker_enabled;
> >  #endif
> >
> >         unsigned long soft_limit;
> > diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
> > index 4106fbc5b4b3..81f4c5ea3e16 100644
> > --- a/include/linux/mmzone.h
> > +++ b/include/linux/mmzone.h
> > @@ -637,6 +637,20 @@ struct lruvec {
> >  #ifdef CONFIG_MEMCG
> >         struct pglist_data *pgdat;
> >  #endif
> > +#ifdef CONFIG_ZSWAP
> > +       /*
> > +        * Number of pages in zswap that should be protected from the shrinker.
> > +        * This number is an estimate of the following counts:
> > +        *
> > +        * a) Recent page faults.
> > +        * b) Recent insertion to the zswap LRU. This includes new zswap stores,
> > +        *    as well as recent zswap LRU rotations.
> > +        *
> > +        * These pages are likely to be warm, and might incur IO if the are written
> > +        * to swap.
> > +        */
> > +       unsigned long nr_zswap_protected;
> > +#endif
>
> Would this be better abstracted in a zswap lruvec struct?
There is just one field, so that sounds like overkill to me.
But if we need to store more data (for smarter heuristics),
that'll be a good idea. I'll keep this in mind. Thanks for the
suggestion, Yosry!
>
> >  };
> >
> >  /* Isolate unmapped pages */
> > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > index 9f84b3f7b469..1a2c97cf396f 100644
> > --- a/mm/memcontrol.c
> > +++ b/mm/memcontrol.c
> > @@ -5352,6 +5352,8 @@ mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
> >         WRITE_ONCE(memcg->soft_limit, PAGE_COUNTER_MAX);
> >  #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
> >         memcg->zswap_max = PAGE_COUNTER_MAX;
> > +       /* Disable the shrinker by default */
> > +       atomic_set(&memcg->zswap_shrinker_enabled, 0);
> >  #endif
> >         page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
> >         if (parent) {
> > @@ -7877,6 +7879,31 @@ static ssize_t zswap_max_write(struct kernfs_open_file *of,
> >         return nbytes;
> >  }
> >
> > +static int zswap_shrinker_enabled_show(struct seq_file *m, void *v)
> > +{
> > +       struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
> > +
> > +       seq_printf(m, "%d\n", atomic_read(&memcg->zswap_shrinker_enabled));
> > +       return 0;
> > +}
> > +
> > +static ssize_t zswap_shrinker_enabled_write(struct kernfs_open_file *of,
> > +                              char *buf, size_t nbytes, loff_t off)
> > +{
> > +       struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
> > +       int zswap_shrinker_enabled;
> > +       ssize_t parse_ret = kstrtoint(strstrip(buf), 0, &zswap_shrinker_enabled);
> > +
> > +       if (parse_ret)
> > +               return parse_ret;
> > +
> > +       if (zswap_shrinker_enabled < 0 || zswap_shrinker_enabled > 1)
> > +               return -ERANGE;
> > +
> > +       atomic_set(&memcg->zswap_shrinker_enabled, zswap_shrinker_enabled);
> > +       return nbytes;
> > +}
> > +
> >  static struct cftype zswap_files[] = {
> >         {
> >                 .name = "zswap.current",
> > @@ -7889,6 +7916,12 @@ static struct cftype zswap_files[] = {
> >                 .seq_show = zswap_max_show,
> >                 .write = zswap_max_write,
> >         },
> > +       {
> > +               .name = "zswap.shrinker.enabled",
> > +               .flags = CFTYPE_NOT_ON_ROOT,
> > +               .seq_show = zswap_shrinker_enabled_show,
> > +               .write = zswap_shrinker_enabled_write,
> > +       },
> >         { }     /* terminate */
> >  };
> >  #endif /* CONFIG_MEMCG_KMEM && CONFIG_ZSWAP */
> > diff --git a/mm/swap_state.c b/mm/swap_state.c
> > index 1c826737aacb..788e36a06c34 100644
> > --- a/mm/swap_state.c
> > +++ b/mm/swap_state.c
> > @@ -618,6 +618,22 @@ static unsigned long swapin_nr_pages(unsigned long offset)
> >         return pages;
> >  }
> >
> > +#ifdef CONFIG_ZSWAP
> > +/*
> > + * Refault is an indication that warmer pages are not resident in memory.
> > + * Increase the size of zswap's protected area.
> > + */
> > +static void inc_nr_protected(struct page *page)
> > +{
> > +       struct lruvec *lruvec = folio_lruvec(page_folio(page));
> > +       unsigned long flags;
> > +
> > +       spin_lock_irqsave(&lruvec->lru_lock, flags);
> > +       lruvec->nr_zswap_protected++;
> > +       spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > +}
> > +#endif
> > +
>
> A few questions:
> - Why is this function named in such a generic way?
Perhaps inc_nr_zswap_protected would be better? :)
> - Why is this function here instead of in mm/zswap.c?
No particular reason :) It's not being used anywhere else,
so I just put it as a static function here.
> - Why is this protected by the heavily contested lruvec lock instead
> of being an atomic?
nr_zswap_protected can be decayed (see zswap_lru_add), which
I don't think it can be implemented with atomics :( It'd be much
cleaner indeed.

I'm wary of adding new locks, so I just re-use this existing lock.
But if lruvec lock is heavily congested (I'm not aware/familar with
this issue), then perhaps a new, dedicated lock would help?
>
> >  /**
> >   * swap_cluster_readahead - swap in pages in hope we need them soon
> >   * @entry: swap entry of this memory
> > @@ -686,7 +702,12 @@ struct page *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask,
> >         lru_add_drain();        /* Push any new pages onto the LRU now */
> >  skip:
> >         /* The page was likely read above, so no need for plugging here */
> > -       return read_swap_cache_async(entry, gfp_mask, vma, addr, NULL);
> > +       page = read_swap_cache_async(entry, gfp_mask, vma, addr, NULL);
> > +#ifdef CONFIG_ZSWAP
> > +       if (page)
> > +               inc_nr_protected(page);
> > +#endif
> > +       return page;
> >  }
> >
> >  int init_swap_address_space(unsigned int type, unsigned long nr_pages)
> > @@ -853,8 +874,12 @@ static struct page *swap_vma_readahead(swp_entry_t fentry, gfp_t gfp_mask,
> >         lru_add_drain();
> >  skip:
> >         /* The page was likely read above, so no need for plugging here */
> > -       return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address,
> > -                                    NULL);
> > +       page = read_swap_cache_async(fentry, gfp_mask, vma, vmf->address, NULL);
> > +#ifdef CONFIG_ZSWAP
> > +       if (page)
> > +               inc_nr_protected(page);
> > +#endif
> > +       return page;
> >  }
> >
> >  /**
> > diff --git a/mm/zswap.c b/mm/zswap.c
> > index 1a469e5d5197..79cb18eeb8bf 100644
> > --- a/mm/zswap.c
> > +++ b/mm/zswap.c
> > @@ -145,6 +145,26 @@ module_param_named(exclusive_loads, zswap_exclusive_loads_enabled, bool, 0644);
> >  /* Number of zpools in zswap_pool (empirically determined for scalability) */
> >  #define ZSWAP_NR_ZPOOLS 32
> >
> > +/*
> > + * Global flag to enable/disable memory pressure-based shrinker for all memcgs.
> > + * If CONFIG_MEMCG_KMEM is on, we can further selectively disable
> > + * the shrinker for each memcg.
> > + */
> > +static bool zswap_shrinker_enabled;
> > +module_param_named(shrinker_enabled, zswap_shrinker_enabled, bool, 0644);
> > +#ifdef CONFIG_MEMCG_KMEM
> > +static bool is_shrinker_enabled(struct mem_cgroup *memcg)
> > +{
> > +       return zswap_shrinker_enabled &&
> > +               atomic_read(&memcg->zswap_shrinker_enabled);
> > +}
> > +#else
> > +static bool is_shrinker_enabled(struct mem_cgroup *memcg)
> > +{
> > +       return zswap_shrinker_enabled;
> > +}
> > +#endif
> > +
> >  /*********************************
> >  * data structures
> >  **********************************/
> > @@ -174,6 +194,8 @@ struct zswap_pool {
> >         char tfm_name[CRYPTO_MAX_ALG_NAME];
> >         struct list_lru list_lru;
> >         struct mem_cgroup *next_shrink;
> > +       struct shrinker *shrinker;
> > +       atomic_t nr_stored;
> >  };
> >
> >  /*
> > @@ -273,17 +295,26 @@ static bool zswap_can_accept(void)
> >                         DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
> >  }
> >
> > +static u64 get_zswap_pool_size(struct zswap_pool *pool)
> > +{
> > +       u64 pool_size = 0;
> > +       int i;
> > +
> > +       for (i = 0; i < ZSWAP_NR_ZPOOLS; i++)
> > +               pool_size += zpool_get_total_size(pool->zpools[i]);
> > +
> > +       return pool_size;
> > +}
> > +
> >  static void zswap_update_total_size(void)
> >  {
> >         struct zswap_pool *pool;
> >         u64 total = 0;
> > -       int i;
> >
> >         rcu_read_lock();
> >
> >         list_for_each_entry_rcu(pool, &zswap_pools, list)
> > -               for (i = 0; i < ZSWAP_NR_ZPOOLS; i++)
> > -                       total += zpool_get_total_size(pool->zpools[i]);
> > +               total += get_zswap_pool_size(pool);
> >
> >         rcu_read_unlock();
> >
> > @@ -318,8 +349,23 @@ static bool zswap_lru_add(struct list_lru *list_lru, struct zswap_entry *entry)
> >  {
> >         struct mem_cgroup *memcg = entry->objcg ?
> >                 get_mem_cgroup_from_objcg(entry->objcg) : NULL;
> > +       struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(entry->nid));
> >         bool added = __list_lru_add(list_lru, &entry->lru, entry->nid, memcg);
> > +       unsigned long flags, lru_size;
> > +
> > +       if (added) {
> > +               lru_size = list_lru_count_one(list_lru, entry->nid, memcg);
> > +               spin_lock_irqsave(&lruvec->lru_lock, flags);
> > +               lruvec->nr_zswap_protected++;
> >
> > +               /*
> > +                * Decay to avoid overflow and adapt to changing workloads.
> > +                * This is based on LRU reclaim cost decaying heuristics.
> > +                */
> > +               if (lruvec->nr_zswap_protected > lru_size / 4)
> > +                       lruvec->nr_zswap_protected /= 2;
> > +               spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > +       }
> >         mem_cgroup_put(memcg);
> >         return added;
> >  }
> > @@ -420,6 +466,7 @@ static void zswap_free_entry(struct zswap_entry *entry)
> >         else {
> >                 zswap_lru_del(&entry->pool->list_lru, entry);
> >                 zpool_free(zswap_find_zpool(entry), entry->handle);
> > +               atomic_dec(&entry->pool->nr_stored);
> >                 zswap_pool_put(entry->pool);
> >         }
> >         zswap_entry_cache_free(entry);
> > @@ -461,6 +508,98 @@ static struct zswap_entry *zswap_entry_find_get(struct rb_root *root,
> >         return entry;
> >  }
> >
> > +/*********************************
> > +* shrinker functions
> > +**********************************/
> > +static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_one *l,
> > +                                      spinlock_t *lock, void *arg);
> > +
> > +static unsigned long zswap_shrinker_scan(struct shrinker *shrinker,
> > +               struct shrink_control *sc)
> > +{
> > +       struct zswap_pool *pool = shrinker->private_data;
> > +       unsigned long shrink_ret, nr_zswap_protected, flags,
> > +               lru_size = list_lru_shrink_count(&pool->list_lru, sc);
> > +       struct lruvec *lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid));
> > +       bool encountered_page_in_swapcache = false;
> > +
> > +       spin_lock_irqsave(&lruvec->lru_lock, flags);
> > +       nr_zswap_protected = lruvec->nr_zswap_protected;
> > +       spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > +
> > +       /*
> > +        * Abort if the shrinker is disabled or if we are shrinking into the
> > +        * protected region.
> > +        */
> > +       if (!is_shrinker_enabled(sc->memcg) ||
> > +                       nr_zswap_protected >= lru_size - sc->nr_to_scan) {
> > +               sc->nr_scanned = 0;
> > +               return SHRINK_STOP;
> > +       }
> > +
> > +       shrink_ret = list_lru_shrink_walk(&pool->list_lru, sc, &shrink_memcg_cb,
> > +               &encountered_page_in_swapcache);
> > +
> > +       if (encountered_page_in_swapcache)
> > +               return SHRINK_STOP;
> > +
> > +       return shrink_ret ? shrink_ret : SHRINK_STOP;
> > +}
> > +
> > +static unsigned long zswap_shrinker_count(struct shrinker *shrinker,
> > +               struct shrink_control *sc)
> > +{
> > +       struct zswap_pool *pool = shrinker->private_data;
> > +       struct mem_cgroup *memcg = sc->memcg;
> > +       struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(sc->nid));
> > +       unsigned long nr_backing, nr_stored, nr_freeable, flags;
> > +
> > +#ifdef CONFIG_MEMCG_KMEM
> > +       cgroup_rstat_flush(memcg->css.cgroup);
> > +       nr_backing = memcg_page_state(memcg, MEMCG_ZSWAP_B) >> PAGE_SHIFT;
> > +       nr_stored = memcg_page_state(memcg, MEMCG_ZSWAPPED);
> > +#else
> > +       /* use pool stats instead of memcg stats */
> > +       nr_backing = get_zswap_pool_size(pool) >> PAGE_SHIFT;
> > +       nr_stored = atomic_read(&pool->nr_stored);
> > +#endif
> > +
> > +       if (!is_shrinker_enabled(memcg) || !nr_stored)
> > +               return 0;
> > +
> > +       nr_freeable = list_lru_shrink_count(&pool->list_lru, sc);
> > +       /*
> > +        * Subtract the lru size by an estimate of the number of pages
> > +        * that should be protected.
> > +        */
> > +       spin_lock_irqsave(&lruvec->lru_lock, flags);
> > +       nr_freeable = nr_freeable > lruvec->nr_zswap_protected ?
> > +               nr_freeable - lruvec->nr_zswap_protected : 0;
> > +       spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > +
> > +       /*
> > +        * Scale the number of freeable pages by the memory saving factor.
> > +        * This ensures that the better zswap compresses memory, the fewer
> > +        * pages we will evict to swap (as it will otherwise incur IO for
> > +        * relatively small memory saving).
> > +        */
> > +       return mult_frac(nr_freeable, nr_backing, nr_stored);
> > +}
> > +
> > +static void zswap_alloc_shrinker(struct zswap_pool *pool)
> > +{
> > +       pool->shrinker =
> > +               shrinker_alloc(SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE, "mm-zswap");
> > +       if (!pool->shrinker)
> > +               return;
> > +
> > +       pool->shrinker->private_data = pool;
> > +       pool->shrinker->scan_objects = zswap_shrinker_scan;
> > +       pool->shrinker->count_objects = zswap_shrinker_count;
> > +       pool->shrinker->batch = 0;
> > +       pool->shrinker->seeks = DEFAULT_SEEKS;
> > +}
> > +
> >  /*********************************
> >  * per-cpu code
> >  **********************************/
> > @@ -656,11 +795,14 @@ static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_o
> >                                        spinlock_t *lock, void *arg)
> >  {
> >         struct zswap_entry *entry = container_of(item, struct zswap_entry, lru);
> > +       bool *encountered_page_in_swapcache = (bool *)arg;
> >         struct mem_cgroup *memcg;
> >         struct zswap_tree *tree;
> > +       struct lruvec *lruvec;
> >         pgoff_t swpoffset;
> >         enum lru_status ret = LRU_REMOVED_RETRY;
> >         int writeback_result;
> > +       unsigned long flags;
> >
> >         /*
> >          * Once the lru lock is dropped, the entry might get freed. The
> > @@ -696,8 +838,24 @@ static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_o
> >                 /* we cannot use zswap_lru_add here, because it increments node's lru count */
> >                 list_lru_putback(&entry->pool->list_lru, item, entry->nid, memcg);
> >                 spin_unlock(lock);
> > -               mem_cgroup_put(memcg);
> >                 ret = LRU_RETRY;
> > +
> > +               /*
> > +                * Encountering a page already in swap cache is a sign that we are shrinking
> > +                * into the warmer region. We should terminate shrinking (if we're in the dynamic
> > +                * shrinker context).
> > +                */
> > +               if (writeback_result == -EEXIST && encountered_page_in_swapcache) {
> > +                       ret = LRU_SKIP;
> > +                       *encountered_page_in_swapcache = true;
> > +               }
> > +               lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(entry->nid));
> > +               spin_lock_irqsave(&lruvec->lru_lock, flags);
> > +               /* Increment the protection area to account for the LRU rotation. */
> > +               lruvec->nr_zswap_protected++;
> > +               spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > +
> > +               mem_cgroup_put(memcg);
> >                 goto put_unlock;
> >         }
> >
> > @@ -828,6 +986,11 @@ static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
> >                                        &pool->node);
> >         if (ret)
> >                 goto error;
> > +
> > +       zswap_alloc_shrinker(pool);
> > +       if (!pool->shrinker)
> > +               goto error;
> > +
> >         pr_debug("using %s compressor\n", pool->tfm_name);
> >
> >         /* being the current pool takes 1 ref; this func expects the
> > @@ -836,12 +999,17 @@ static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
> >         kref_init(&pool->kref);
> >         INIT_LIST_HEAD(&pool->list);
> >         INIT_WORK(&pool->shrink_work, shrink_worker);
> > -       list_lru_init_memcg(&pool->list_lru, NULL);
> > +       if (list_lru_init_memcg(&pool->list_lru, pool->shrinker))
> > +               goto lru_fail;
> > +       shrinker_register(pool->shrinker);
> >
> >         zswap_pool_debug("created", pool);
> >
> >         return pool;
> >
> > +lru_fail:
> > +       list_lru_destroy(&pool->list_lru);
> > +       shrinker_free(pool->shrinker);
> >  error:
> >         if (pool->acomp_ctx)
> >                 free_percpu(pool->acomp_ctx);
> > @@ -899,6 +1067,7 @@ static void zswap_pool_destroy(struct zswap_pool *pool)
> >
> >         zswap_pool_debug("destroying", pool);
> >
> > +       shrinker_free(pool->shrinker);
> >         cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
> >         free_percpu(pool->acomp_ctx);
> >         list_lru_destroy(&pool->list_lru);
> > @@ -1431,6 +1600,7 @@ bool zswap_store(struct folio *folio)
> >         if (entry->length) {
> >                 INIT_LIST_HEAD(&entry->lru);
> >                 zswap_lru_add(&pool->list_lru, entry);
> > +               atomic_inc(&pool->nr_stored);
> >         }
> >         spin_unlock(&tree->lock);
> >
> > --
> > 2.34.1
Thanks for the comments/suggestion, Yosry!
  

On Mon, Sep 25, 2023 at 4:29 PM Nhat Pham <nphamcs@gmail.com> wrote:
>
> On Mon, Sep 25, 2023 at 1:38 PM Yosry Ahmed <yosryahmed@google.com> wrote:
> >
> > On Tue, Sep 19, 2023 at 10:14 AM Nhat Pham <nphamcs@gmail.com> wrote:
> > >
> > > Currently, we only shrink the zswap pool when the user-defined limit is
> > > hit. This means that if we set the limit too high, cold data that are
> > > unlikely to be used again will reside in the pool, wasting precious
> > > memory. It is hard to predict how much zswap space will be needed ahead
> > > of time, as this depends on the workload (specifically, on factors such
> > > as memory access patterns and compressibility of the memory pages).
> > >
> > > This patch implements a memcg- and NUMA-aware shrinker for zswap, that
> > > is initiated when there is memory pressure. The shrinker does not
> > > have any parameter that must be tuned by the user, and can be opted in
> > > or out on a per-memcg basis.
> >
> > What's the use case for having per-memcg opt-in/out?
> >
> > If there is memory pressure, reclaiming swap-backed pages will push
> > pages out of zswap anyway, regardless of this patch. With this patch,
> > any sort of reclaim can push pages out of zswap. Wouldn't that be
> > preferable to reclaiming memory that is currently resident in memory
> > (so arguably hotter than the pages in zswap)? Why would this decision
> > be different per-memcg?
> I'm not quite following your argument here. The point of having this
> be done on a per-memcg basis is that we have different workloads
> with different memory access pattern (and as a result, different memory
> coldness distribution).
>
> In a workload where there is a lot of cold data, we can really benefit
> from reclaiming all of those pages and repurpose the memory reclaimed
> (for e.g for filecache).
>
> On the other hand, in a workload where there aren't a lot of cold data,
> reclaiming its zswapped pages will at best do nothing (wasting CPU
> cycles on compression/decompression), and at worst hurt performance
> (due to increased IO when we need those writtenback pages again).
>
> Such different workloads could co-exist in the same system, and having
> a per-memcg knob allows us to crank on the shrinker only on workloads
> where it makes sense.

I am not sure we are on the same page here.

What you're describing sounds more like proactive reclaim, which we
wouldn't invoke unless the workload has cold data anyway.

IIUC, outside of that, this shrinker will run when there is memory
pressure. This means that we need to free memory anyway, regardless of
its absolute coldness. We want to evict the colder pages in the memcg.
It seems to be that in ~all cases, evicting pages in zswap will be
better than evicting pages in memory, as the pages in memory are
arguably hotter (since they weren't reclaimed first). This seems to be
something that would be true for all workloads.

What am I missing?

> >
> > >
> > > Furthermore, to make it more robust for many workloads and prevent
> > > overshrinking (i.e evicting warm pages that might be refaulted into
> > > memory), we build in the following heuristics:
> > >
> > > * Estimate the number of warm pages residing in zswap, and attempt to
> > >   protect this region of the zswap LRU.
> > > * Scale the number of freeable objects by an estimate of the memory
> > >   saving factor. The better zswap compresses the data, the fewer pages
> > >   we will evict to swap (as we will otherwise incur IO for relatively
> > >   small memory saving).
> > > * During reclaim, if the shrinker encounters a page that is also being
> > >   brought into memory, the shrinker will cautiously terminate its
> > >   shrinking action, as this is a sign that it is touching the warmer
> > >   region of the zswap LRU.
> >
> > I don't have an opinion about the reclaim heuristics here, I will let
> > reclaim experts chip in.
> >
> > >
> > > On a benchmark that we have run:
> >
> > Please add more details (as much as possible) about the benchmarks used here.
> Sure! I built the kernel in a memory-limited cgroup a couple times,
> then measured the build time.
>
> To simulate conditions where there are cold, unused data, I
> also generated a bunch of data in tmpfs (and never touch them
> again).

Please include such details in the commit message, there is also
another reference below to "another" benchmark.


> >
> > >
> > > (without the shrinker)
> > > real -- mean: 153.27s, median: 153.199s
> > > sys -- mean: 541.652s, median: 541.903s
> > > user -- mean: 4384.9673999999995s, median: 4385.471s
> > >
> > > (with the shrinker)
> > > real -- mean: 151.4956s, median: 151.456s
> > > sys -- mean: 461.14639999999997s, median: 465.656s
> > > user -- mean: 4384.7118s, median: 4384.675s
> > >
> > > We observed a 14-15% reduction in kernel CPU time, which translated to
> > > over 1% reduction in real time.
> > >
> > > On another benchmark, where there was a lot more cold memory residing in
> > > zswap, we observed even more pronounced gains:
> > >
> > > (without the shrinker)
> > > real -- mean: 157.52519999999998s, median: 157.281s
> > > sys -- mean: 769.3082s, median: 780.545s
> > > user -- mean: 4378.1622s, median: 4378.286s
> > >
> > > (with the shrinker)
> > > real -- mean: 152.9608s, median: 152.845s
> > > sys -- mean: 517.4446s, median: 506.749s
> > > user -- mean: 4387.694s, median: 4387.935s
> > >
> > > Here, we saw around 32-35% reduction in kernel CPU time, which
> > > translated to 2.8% reduction in real time. These results confirm our
> > > hypothesis that the shrinker is more helpful the more cold memory we
> > > have.
> > >
> > > Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
> > > Signed-off-by: Nhat Pham <nphamcs@gmail.com>
> > > ---
> > >  Documentation/admin-guide/mm/zswap.rst |  12 ++
> > >  include/linux/memcontrol.h             |   1 +
> > >  include/linux/mmzone.h                 |  14 ++
> > >  mm/memcontrol.c                        |  33 +++++
> > >  mm/swap_state.c                        |  31 ++++-
> > >  mm/zswap.c                             | 180 ++++++++++++++++++++++++-
> > >  6 files changed, 263 insertions(+), 8 deletions(-)
> > >
> > > diff --git a/Documentation/admin-guide/mm/zswap.rst b/Documentation/admin-guide/mm/zswap.rst
> > > index 45b98390e938..ae8597a67804 100644
> > > --- a/Documentation/admin-guide/mm/zswap.rst
> > > +++ b/Documentation/admin-guide/mm/zswap.rst
> > > @@ -153,6 +153,18 @@ attribute, e. g.::
> > >
> > >  Setting this parameter to 100 will disable the hysteresis.
> > >
> > > +When there is a sizable amount of cold memory residing in the zswap pool, it
> > > +can be advantageous to proactively write these cold pages to swap and reclaim
> > > +the memory for other use cases. By default, the zswap shrinker is disabled.
> > > +User can enable it by first switching on the global knob:
> > > +
> > > +  echo Y > /sys/module/zswap/par meters/shrinker_enabled
> > > +
> > > +When the kernel is compiled with CONFIG_MEMCG_KMEM, user needs to further turn
> > > +it on for each cgroup that the shrinker should target:
> > > +
> > > +  echo 1 > /sys/fs/cgroup/<cgroup-name>/memory.zswap.shrinker.enabled
> > > +
> > >  A debugfs interface is provided for various statistic about pool size, number
> > >  of pages stored, same-value filled pages and various counters for the reasons
> > >  pages are rejected.
> > > diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
> > > index 05d34b328d9d..f005ea667863 100644
> > > --- a/include/linux/memcontrol.h
> > > +++ b/include/linux/memcontrol.h
> > > @@ -219,6 +219,7 @@ struct mem_cgroup {
> > >
> > >  #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
> > >         unsigned long zswap_max;
> > > +       atomic_t zswap_shrinker_enabled;
> > >  #endif
> > >
> > >         unsigned long soft_limit;
> > > diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
> > > index 4106fbc5b4b3..81f4c5ea3e16 100644
> > > --- a/include/linux/mmzone.h
> > > +++ b/include/linux/mmzone.h
> > > @@ -637,6 +637,20 @@ struct lruvec {
> > >  #ifdef CONFIG_MEMCG
> > >         struct pglist_data *pgdat;
> > >  #endif
> > > +#ifdef CONFIG_ZSWAP
> > > +       /*
> > > +        * Number of pages in zswap that should be protected from the shrinker.
> > > +        * This number is an estimate of the following counts:
> > > +        *
> > > +        * a) Recent page faults.
> > > +        * b) Recent insertion to the zswap LRU. This includes new zswap stores,
> > > +        *    as well as recent zswap LRU rotations.
> > > +        *
> > > +        * These pages are likely to be warm, and might incur IO if the are written
> > > +        * to swap.
> > > +        */
> > > +       unsigned long nr_zswap_protected;
> > > +#endif
> >
> > Would this be better abstracted in a zswap lruvec struct?
> There is just one field, so that sounds like overkill to me.
> But if we need to store more data (for smarter heuristics),
> that'll be a good idea. I'll keep this in mind. Thanks for the
> suggestion, Yosry!

(A space between the quoted text and the reply usually helps visually :)

It wasn't really about the number of fields, but rather place this
struct in zswap.h (with the long comment explaining what it's doing),
and adding an abstracted struct member here. The comment will live in
an appropriate file, further modifications don't need to touch
mmzone.h, and struct lruvec is less cluttered for readers that don't
care about zswap (and we can avoid the ifdef).

Anyway, this is all mostly aesthetic so I don't feel strongly.

> >
> > >  };
> > >
> > >  /* Isolate unmapped pages */
> > > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > > index 9f84b3f7b469..1a2c97cf396f 100644
> > > --- a/mm/memcontrol.c
> > > +++ b/mm/memcontrol.c
> > > @@ -5352,6 +5352,8 @@ mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
> > >         WRITE_ONCE(memcg->soft_limit, PAGE_COUNTER_MAX);
> > >  #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
> > >         memcg->zswap_max = PAGE_COUNTER_MAX;
> > > +       /* Disable the shrinker by default */
> > > +       atomic_set(&memcg->zswap_shrinker_enabled, 0);
> > >  #endif
> > >         page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
> > >         if (parent) {
> > > @@ -7877,6 +7879,31 @@ static ssize_t zswap_max_write(struct kernfs_open_file *of,
> > >         return nbytes;
> > >  }
> > >
> > > +static int zswap_shrinker_enabled_show(struct seq_file *m, void *v)
> > > +{
> > > +       struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
> > > +
> > > +       seq_printf(m, "%d\n", atomic_read(&memcg->zswap_shrinker_enabled));
> > > +       return 0;
> > > +}
> > > +
> > > +static ssize_t zswap_shrinker_enabled_write(struct kernfs_open_file *of,
> > > +                              char *buf, size_t nbytes, loff_t off)
> > > +{
> > > +       struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
> > > +       int zswap_shrinker_enabled;
> > > +       ssize_t parse_ret = kstrtoint(strstrip(buf), 0, &zswap_shrinker_enabled);
> > > +
> > > +       if (parse_ret)
> > > +               return parse_ret;
> > > +
> > > +       if (zswap_shrinker_enabled < 0 || zswap_shrinker_enabled > 1)
> > > +               return -ERANGE;
> > > +
> > > +       atomic_set(&memcg->zswap_shrinker_enabled, zswap_shrinker_enabled);
> > > +       return nbytes;
> > > +}
> > > +
> > >  static struct cftype zswap_files[] = {
> > >         {
> > >                 .name = "zswap.current",
> > > @@ -7889,6 +7916,12 @@ static struct cftype zswap_files[] = {
> > >                 .seq_show = zswap_max_show,
> > >                 .write = zswap_max_write,
> > >         },
> > > +       {
> > > +               .name = "zswap.shrinker.enabled",
> > > +               .flags = CFTYPE_NOT_ON_ROOT,
> > > +               .seq_show = zswap_shrinker_enabled_show,
> > > +               .write = zswap_shrinker_enabled_write,
> > > +       },
> > >         { }     /* terminate */
> > >  };
> > >  #endif /* CONFIG_MEMCG_KMEM && CONFIG_ZSWAP */
> > > diff --git a/mm/swap_state.c b/mm/swap_state.c
> > > index 1c826737aacb..788e36a06c34 100644
> > > --- a/mm/swap_state.c
> > > +++ b/mm/swap_state.c
> > > @@ -618,6 +618,22 @@ static unsigned long swapin_nr_pages(unsigned long offset)
> > >         return pages;
> > >  }
> > >
> > > +#ifdef CONFIG_ZSWAP
> > > +/*
> > > + * Refault is an indication that warmer pages are not resident in memory.
> > > + * Increase the size of zswap's protected area.
> > > + */
> > > +static void inc_nr_protected(struct page *page)
> > > +{
> > > +       struct lruvec *lruvec = folio_lruvec(page_folio(page));
> > > +       unsigned long flags;
> > > +
> > > +       spin_lock_irqsave(&lruvec->lru_lock, flags);
> > > +       lruvec->nr_zswap_protected++;
> > > +       spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > > +}
> > > +#endif
> > > +
> >
> > A few questions:
> > - Why is this function named in such a generic way?
> Perhaps inc_nr_zswap_protected would be better? :)

If we use an atomic, the function can go away anyway. See below.

> > - Why is this function here instead of in mm/zswap.c?
> No particular reason :) It's not being used anywhere else,
> so I just put it as a static function here.

It is inline in mm/zswap.c in one place. I personally would have
preferred nr_zswap_protected and the helper to be defined in
zswap.h/zswap.c as I mentioned below. Anyway, this function can go
away.

> > - Why is this protected by the heavily contested lruvec lock instead
> > of being an atomic?
> nr_zswap_protected can be decayed (see zswap_lru_add), which
> I don't think it can be implemented with atomics :( It'd be much
> cleaner indeed.

I think a cmpxchg (or a try_cmpxchg) loop can be used in this case to
implement it using an atomic?

See https://docs.kernel.org/core-api/wrappers/atomic_t.html.

> > > +               lruvec->nr_zswap_protected++;
> > >
> > > +               /*
> > > +                * Decay to avoid overflow and adapt to changing workloads.
> > > +                * This is based on LRU reclaim cost decaying heuristics.
> > > +                */
> > > +               if (lruvec->nr_zswap_protected > lru_size / 4)
> > > +                       lruvec->nr_zswap_protected /= 2;

>
> I'm wary of adding new locks, so I just re-use this existing lock.
> But if lruvec lock is heavily congested (I'm not aware/familar with
> this issue), then perhaps a new, dedicated lock would help?
> >
> > >  /**
> > >   * swap_cluster_readahead - swap in pages in hope we need them soon
> > >   * @entry: swap entry of this memory
> > > @@ -686,7 +702,12 @@ struct page *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask,
> > >         lru_add_drain();        /* Push any new pages onto the LRU now */
> > >  skip:
> > >         /* The page was likely read above, so no need for plugging here */
> > > -       return read_swap_cache_async(entry, gfp_mask, vma, addr, NULL);
> > > +       page = read_swap_cache_async(entry, gfp_mask, vma, addr, NULL);
> > > +#ifdef CONFIG_ZSWAP
> > > +       if (page)
> > > +               inc_nr_protected(page);
> > > +#endif
> > > +       return page;
> > >  }
> > >
> > >  int init_swap_address_space(unsigned int type, unsigned long nr_pages)
> > > @@ -853,8 +874,12 @@ static struct page *swap_vma_readahead(swp_entry_t fentry, gfp_t gfp_mask,
> > >         lru_add_drain();
> > >  skip:
> > >         /* The page was likely read above, so no need for plugging here */
> > > -       return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address,
> > > -                                    NULL);
> > > +       page = read_swap_cache_async(fentry, gfp_mask, vma, vmf->address, NULL);
> > > +#ifdef CONFIG_ZSWAP
> > > +       if (page)
> > > +               inc_nr_protected(page);
> > > +#endif
> > > +       return page;
> > >  }
> > >
> > >  /**
> > > diff --git a/mm/zswap.c b/mm/zswap.c
> > > index 1a469e5d5197..79cb18eeb8bf 100644
> > > --- a/mm/zswap.c
> > > +++ b/mm/zswap.c
> > > @@ -145,6 +145,26 @@ module_param_named(exclusive_loads, zswap_exclusive_loads_enabled, bool, 0644);
> > >  /* Number of zpools in zswap_pool (empirically determined for scalability) */
> > >  #define ZSWAP_NR_ZPOOLS 32
> > >
> > > +/*
> > > + * Global flag to enable/disable memory pressure-based shrinker for all memcgs.
> > > + * If CONFIG_MEMCG_KMEM is on, we can further selectively disable
> > > + * the shrinker for each memcg.
> > > + */
> > > +static bool zswap_shrinker_enabled;
> > > +module_param_named(shrinker_enabled, zswap_shrinker_enabled, bool, 0644);
> > > +#ifdef CONFIG_MEMCG_KMEM
> > > +static bool is_shrinker_enabled(struct mem_cgroup *memcg)
> > > +{
> > > +       return zswap_shrinker_enabled &&
> > > +               atomic_read(&memcg->zswap_shrinker_enabled);
> > > +}
> > > +#else
> > > +static bool is_shrinker_enabled(struct mem_cgroup *memcg)
> > > +{
> > > +       return zswap_shrinker_enabled;
> > > +}
> > > +#endif
> > > +
> > >  /*********************************
> > >  * data structures
> > >  **********************************/
> > > @@ -174,6 +194,8 @@ struct zswap_pool {
> > >         char tfm_name[CRYPTO_MAX_ALG_NAME];
> > >         struct list_lru list_lru;
> > >         struct mem_cgroup *next_shrink;
> > > +       struct shrinker *shrinker;
> > > +       atomic_t nr_stored;
> > >  };
> > >
> > >  /*
> > > @@ -273,17 +295,26 @@ static bool zswap_can_accept(void)
> > >                         DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
> > >  }
> > >
> > > +static u64 get_zswap_pool_size(struct zswap_pool *pool)
> > > +{
> > > +       u64 pool_size = 0;
> > > +       int i;
> > > +
> > > +       for (i = 0; i < ZSWAP_NR_ZPOOLS; i++)
> > > +               pool_size += zpool_get_total_size(pool->zpools[i]);
> > > +
> > > +       return pool_size;
> > > +}
> > > +
> > >  static void zswap_update_total_size(void)
> > >  {
> > >         struct zswap_pool *pool;
> > >         u64 total = 0;
> > > -       int i;
> > >
> > >         rcu_read_lock();
> > >
> > >         list_for_each_entry_rcu(pool, &zswap_pools, list)
> > > -               for (i = 0; i < ZSWAP_NR_ZPOOLS; i++)
> > > -                       total += zpool_get_total_size(pool->zpools[i]);
> > > +               total += get_zswap_pool_size(pool);
> > >
> > >         rcu_read_unlock();
> > >
> > > @@ -318,8 +349,23 @@ static bool zswap_lru_add(struct list_lru *list_lru, struct zswap_entry *entry)
> > >  {
> > >         struct mem_cgroup *memcg = entry->objcg ?
> > >                 get_mem_cgroup_from_objcg(entry->objcg) : NULL;
> > > +       struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(entry->nid));
> > >         bool added = __list_lru_add(list_lru, &entry->lru, entry->nid, memcg);
> > > +       unsigned long flags, lru_size;
> > > +
> > > +       if (added) {
> > > +               lru_size = list_lru_count_one(list_lru, entry->nid, memcg);
> > > +               spin_lock_irqsave(&lruvec->lru_lock, flags);
> > > +               lruvec->nr_zswap_protected++;
> > >
> > > +               /*
> > > +                * Decay to avoid overflow and adapt to changing workloads.
> > > +                * This is based on LRU reclaim cost decaying heuristics.
> > > +                */
> > > +               if (lruvec->nr_zswap_protected > lru_size / 4)
> > > +                       lruvec->nr_zswap_protected /= 2;
> > > +               spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > > +       }
> > >         mem_cgroup_put(memcg);
> > >         return added;
> > >  }
> > > @@ -420,6 +466,7 @@ static void zswap_free_entry(struct zswap_entry *entry)
> > >         else {
> > >                 zswap_lru_del(&entry->pool->list_lru, entry);
> > >                 zpool_free(zswap_find_zpool(entry), entry->handle);
> > > +               atomic_dec(&entry->pool->nr_stored);
> > >                 zswap_pool_put(entry->pool);
> > >         }
> > >         zswap_entry_cache_free(entry);
> > > @@ -461,6 +508,98 @@ static struct zswap_entry *zswap_entry_find_get(struct rb_root *root,
> > >         return entry;
> > >  }
> > >
> > > +/*********************************
> > > +* shrinker functions
> > > +**********************************/
> > > +static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_one *l,
> > > +                                      spinlock_t *lock, void *arg);
> > > +
> > > +static unsigned long zswap_shrinker_scan(struct shrinker *shrinker,
> > > +               struct shrink_control *sc)
> > > +{
> > > +       struct zswap_pool *pool = shrinker->private_data;
> > > +       unsigned long shrink_ret, nr_zswap_protected, flags,
> > > +               lru_size = list_lru_shrink_count(&pool->list_lru, sc);
> > > +       struct lruvec *lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid));
> > > +       bool encountered_page_in_swapcache = false;
> > > +
> > > +       spin_lock_irqsave(&lruvec->lru_lock, flags);
> > > +       nr_zswap_protected = lruvec->nr_zswap_protected;
> > > +       spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > > +
> > > +       /*
> > > +        * Abort if the shrinker is disabled or if we are shrinking into the
> > > +        * protected region.
> > > +        */
> > > +       if (!is_shrinker_enabled(sc->memcg) ||
> > > +                       nr_zswap_protected >= lru_size - sc->nr_to_scan) {
> > > +               sc->nr_scanned = 0;
> > > +               return SHRINK_STOP;
> > > +       }
> > > +
> > > +       shrink_ret = list_lru_shrink_walk(&pool->list_lru, sc, &shrink_memcg_cb,
> > > +               &encountered_page_in_swapcache);
> > > +
> > > +       if (encountered_page_in_swapcache)
> > > +               return SHRINK_STOP;
> > > +
> > > +       return shrink_ret ? shrink_ret : SHRINK_STOP;
> > > +}
> > > +
> > > +static unsigned long zswap_shrinker_count(struct shrinker *shrinker,
> > > +               struct shrink_control *sc)
> > > +{
> > > +       struct zswap_pool *pool = shrinker->private_data;
> > > +       struct mem_cgroup *memcg = sc->memcg;
> > > +       struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(sc->nid));
> > > +       unsigned long nr_backing, nr_stored, nr_freeable, flags;
> > > +
> > > +#ifdef CONFIG_MEMCG_KMEM
> > > +       cgroup_rstat_flush(memcg->css.cgroup);
> > > +       nr_backing = memcg_page_state(memcg, MEMCG_ZSWAP_B) >> PAGE_SHIFT;
> > > +       nr_stored = memcg_page_state(memcg, MEMCG_ZSWAPPED);
> > > +#else
> > > +       /* use pool stats instead of memcg stats */
> > > +       nr_backing = get_zswap_pool_size(pool) >> PAGE_SHIFT;
> > > +       nr_stored = atomic_read(&pool->nr_stored);
> > > +#endif
> > > +
> > > +       if (!is_shrinker_enabled(memcg) || !nr_stored)
> > > +               return 0;
> > > +
> > > +       nr_freeable = list_lru_shrink_count(&pool->list_lru, sc);
> > > +       /*
> > > +        * Subtract the lru size by an estimate of the number of pages
> > > +        * that should be protected.
> > > +        */
> > > +       spin_lock_irqsave(&lruvec->lru_lock, flags);
> > > +       nr_freeable = nr_freeable > lruvec->nr_zswap_protected ?
> > > +               nr_freeable - lruvec->nr_zswap_protected : 0;
> > > +       spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > > +
> > > +       /*
> > > +        * Scale the number of freeable pages by the memory saving factor.
> > > +        * This ensures that the better zswap compresses memory, the fewer
> > > +        * pages we will evict to swap (as it will otherwise incur IO for
> > > +        * relatively small memory saving).
> > > +        */
> > > +       return mult_frac(nr_freeable, nr_backing, nr_stored);
> > > +}
> > > +
> > > +static void zswap_alloc_shrinker(struct zswap_pool *pool)
> > > +{
> > > +       pool->shrinker =
> > > +               shrinker_alloc(SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE, "mm-zswap");
> > > +       if (!pool->shrinker)
> > > +               return;
> > > +
> > > +       pool->shrinker->private_data = pool;
> > > +       pool->shrinker->scan_objects = zswap_shrinker_scan;
> > > +       pool->shrinker->count_objects = zswap_shrinker_count;
> > > +       pool->shrinker->batch = 0;
> > > +       pool->shrinker->seeks = DEFAULT_SEEKS;
> > > +}
> > > +
> > >  /*********************************
> > >  * per-cpu code
> > >  **********************************/
> > > @@ -656,11 +795,14 @@ static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_o
> > >                                        spinlock_t *lock, void *arg)
> > >  {
> > >         struct zswap_entry *entry = container_of(item, struct zswap_entry, lru);
> > > +       bool *encountered_page_in_swapcache = (bool *)arg;
> > >         struct mem_cgroup *memcg;
> > >         struct zswap_tree *tree;
> > > +       struct lruvec *lruvec;
> > >         pgoff_t swpoffset;
> > >         enum lru_status ret = LRU_REMOVED_RETRY;
> > >         int writeback_result;
> > > +       unsigned long flags;
> > >
> > >         /*
> > >          * Once the lru lock is dropped, the entry might get freed. The
> > > @@ -696,8 +838,24 @@ static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_o
> > >                 /* we cannot use zswap_lru_add here, because it increments node's lru count */
> > >                 list_lru_putback(&entry->pool->list_lru, item, entry->nid, memcg);
> > >                 spin_unlock(lock);
> > > -               mem_cgroup_put(memcg);
> > >                 ret = LRU_RETRY;
> > > +
> > > +               /*
> > > +                * Encountering a page already in swap cache is a sign that we are shrinking
> > > +                * into the warmer region. We should terminate shrinking (if we're in the dynamic
> > > +                * shrinker context).
> > > +                */
> > > +               if (writeback_result == -EEXIST && encountered_page_in_swapcache) {
> > > +                       ret = LRU_SKIP;
> > > +                       *encountered_page_in_swapcache = true;
> > > +               }
> > > +               lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(entry->nid));
> > > +               spin_lock_irqsave(&lruvec->lru_lock, flags);
> > > +               /* Increment the protection area to account for the LRU rotation. */
> > > +               lruvec->nr_zswap_protected++;
> > > +               spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > > +
> > > +               mem_cgroup_put(memcg);
> > >                 goto put_unlock;
> > >         }
> > >
> > > @@ -828,6 +986,11 @@ static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
> > >                                        &pool->node);
> > >         if (ret)
> > >                 goto error;
> > > +
> > > +       zswap_alloc_shrinker(pool);
> > > +       if (!pool->shrinker)
> > > +               goto error;
> > > +
> > >         pr_debug("using %s compressor\n", pool->tfm_name);
> > >
> > >         /* being the current pool takes 1 ref; this func expects the
> > > @@ -836,12 +999,17 @@ static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
> > >         kref_init(&pool->kref);
> > >         INIT_LIST_HEAD(&pool->list);
> > >         INIT_WORK(&pool->shrink_work, shrink_worker);
> > > -       list_lru_init_memcg(&pool->list_lru, NULL);
> > > +       if (list_lru_init_memcg(&pool->list_lru, pool->shrinker))
> > > +               goto lru_fail;
> > > +       shrinker_register(pool->shrinker);
> > >
> > >         zswap_pool_debug("created", pool);
> > >
> > >         return pool;
> > >
> > > +lru_fail:
> > > +       list_lru_destroy(&pool->list_lru);
> > > +       shrinker_free(pool->shrinker);
> > >  error:
> > >         if (pool->acomp_ctx)
> > >                 free_percpu(pool->acomp_ctx);
> > > @@ -899,6 +1067,7 @@ static void zswap_pool_destroy(struct zswap_pool *pool)
> > >
> > >         zswap_pool_debug("destroying", pool);
> > >
> > > +       shrinker_free(pool->shrinker);
> > >         cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
> > >         free_percpu(pool->acomp_ctx);
> > >         list_lru_destroy(&pool->list_lru);
> > > @@ -1431,6 +1600,7 @@ bool zswap_store(struct folio *folio)
> > >         if (entry->length) {
> > >                 INIT_LIST_HEAD(&entry->lru);
> > >                 zswap_lru_add(&pool->list_lru, entry);
> > > +               atomic_inc(&pool->nr_stored);
> > >         }
> > >         spin_unlock(&tree->lock);
> > >
> > > --
> > > 2.34.1
> Thanks for the comments/suggestion, Yosry!
  

On Mon, Sep 25, 2023 at 5:00 PM Yosry Ahmed <yosryahmed@google.com> wrote:
>
> On Mon, Sep 25, 2023 at 4:29 PM Nhat Pham <nphamcs@gmail.com> wrote:
> >
> > On Mon, Sep 25, 2023 at 1:38 PM Yosry Ahmed <yosryahmed@google.com> wrote:
> > >
> > > On Tue, Sep 19, 2023 at 10:14 AM Nhat Pham <nphamcs@gmail.com> wrote:
> > > >
> > > > Currently, we only shrink the zswap pool when the user-defined limit is
> > > > hit. This means that if we set the limit too high, cold data that are
> > > > unlikely to be used again will reside in the pool, wasting precious
> > > > memory. It is hard to predict how much zswap space will be needed ahead
> > > > of time, as this depends on the workload (specifically, on factors such
> > > > as memory access patterns and compressibility of the memory pages).
> > > >
> > > > This patch implements a memcg- and NUMA-aware shrinker for zswap, that
> > > > is initiated when there is memory pressure. The shrinker does not
> > > > have any parameter that must be tuned by the user, and can be opted in
> > > > or out on a per-memcg basis.
> > >
> > > What's the use case for having per-memcg opt-in/out?
> > >
> > > If there is memory pressure, reclaiming swap-backed pages will push
> > > pages out of zswap anyway, regardless of this patch. With this patch,
> > > any sort of reclaim can push pages out of zswap. Wouldn't that be
> > > preferable to reclaiming memory that is currently resident in memory
> > > (so arguably hotter than the pages in zswap)? Why would this decision
> > > be different per-memcg?
> > I'm not quite following your argument here. The point of having this
> > be done on a per-memcg basis is that we have different workloads
> > with different memory access pattern (and as a result, different memory
> > coldness distribution).
> >
> > In a workload where there is a lot of cold data, we can really benefit
> > from reclaiming all of those pages and repurpose the memory reclaimed
> > (for e.g for filecache).
> >
> > On the other hand, in a workload where there aren't a lot of cold data,
> > reclaiming its zswapped pages will at best do nothing (wasting CPU
> > cycles on compression/decompression), and at worst hurt performance
> > (due to increased IO when we need those writtenback pages again).
> >
> > Such different workloads could co-exist in the same system, and having
> > a per-memcg knob allows us to crank on the shrinker only on workloads
> > where it makes sense.
>
> I am not sure we are on the same page here.
>
> What you're describing sounds more like proactive reclaim, which we
> wouldn't invoke unless the workload has cold data anyway.
>
> IIUC, outside of that, this shrinker will run when there is memory
> pressure. This means that we need to free memory anyway, regardless of
> its absolute coldness. We want to evict the colder pages in the memcg.
> It seems to be that in ~all cases, evicting pages in zswap will be
> better than evicting pages in memory, as the pages in memory are
> arguably hotter (since they weren't reclaimed first). This seems to be
> something that would be true for all workloads.
>
> What am I missing?

Yup, the shrinker is initiated under memory pressure.
And with it, we can reclaim memory from zswap when
it's (often) not at max capacity.

The kernel has no knowledge of absolute coldness, only relative
coldness thanks to LRU. We don't have a global LRU of all possible
memory pages/objects for a particular memcg either, so we cannot
compare the coldness of objects from different sources.

The "coldest" pages in zswap LRU could very well be warm enough
that swapping them out degrades performance, while there are even
colder memory from other sources (other shrinkers registered for this
memcg). Alternatively, we can also "evict" uncompressed anonymous
memory, which will go to the zswap pool. This also saves memory,
and could potentially be better than zswap reclaim (2 compressed
pages might be better performance-wise than 1 uncompressed,
1 swapped out)

All of this depends on the memory access pattern of the workloads,
which could differ cgroup-by-cgroup within the same system.
Having a per-memcg knob is a way for admins to influence this
decision from userspace, if the admins have knowledge about
workload memory access patterns.

For e.g, if we know that there is one particular cgroup that populates
a bunch of single-use tmpfs pages, then we can target that cgroup
specifically, while leaving the other cgroups in the system alone.

>
> > >
> > > >
> > > > Furthermore, to make it more robust for many workloads and prevent
> > > > overshrinking (i.e evicting warm pages that might be refaulted into
> > > > memory), we build in the following heuristics:
> > > >
> > > > * Estimate the number of warm pages residing in zswap, and attempt to
> > > >   protect this region of the zswap LRU.
> > > > * Scale the number of freeable objects by an estimate of the memory
> > > >   saving factor. The better zswap compresses the data, the fewer pages
> > > >   we will evict to swap (as we will otherwise incur IO for relatively
> > > >   small memory saving).
> > > > * During reclaim, if the shrinker encounters a page that is also being
> > > >   brought into memory, the shrinker will cautiously terminate its
> > > >   shrinking action, as this is a sign that it is touching the warmer
> > > >   region of the zswap LRU.
> > >
> > > I don't have an opinion about the reclaim heuristics here, I will let
> > > reclaim experts chip in.
> > >
> > > >
> > > > On a benchmark that we have run:
> > >
> > > Please add more details (as much as possible) about the benchmarks used here.
> > Sure! I built the kernel in a memory-limited cgroup a couple times,
> > then measured the build time.
> >
> > To simulate conditions where there are cold, unused data, I
> > also generated a bunch of data in tmpfs (and never touch them
> > again).
>
> Please include such details in the commit message, there is also
> another reference below to "another" benchmark.

Will do if/when I send v3.
The "another" benchmark is just generating even more tmpfs cold data :)

>
>
> > >
> > > >
> > > > (without the shrinker)
> > > > real -- mean: 153.27s, median: 153.199s
> > > > sys -- mean: 541.652s, median: 541.903s
> > > > user -- mean: 4384.9673999999995s, median: 4385.471s
> > > >
> > > > (with the shrinker)
> > > > real -- mean: 151.4956s, median: 151.456s
> > > > sys -- mean: 461.14639999999997s, median: 465.656s
> > > > user -- mean: 4384.7118s, median: 4384.675s
> > > >
> > > > We observed a 14-15% reduction in kernel CPU time, which translated to
> > > > over 1% reduction in real time.
> > > >
> > > > On another benchmark, where there was a lot more cold memory residing in
> > > > zswap, we observed even more pronounced gains:
> > > >
> > > > (without the shrinker)
> > > > real -- mean: 157.52519999999998s, median: 157.281s
> > > > sys -- mean: 769.3082s, median: 780.545s
> > > > user -- mean: 4378.1622s, median: 4378.286s
> > > >
> > > > (with the shrinker)
> > > > real -- mean: 152.9608s, median: 152.845s
> > > > sys -- mean: 517.4446s, median: 506.749s
> > > > user -- mean: 4387.694s, median: 4387.935s
> > > >
> > > > Here, we saw around 32-35% reduction in kernel CPU time, which
> > > > translated to 2.8% reduction in real time. These results confirm our
> > > > hypothesis that the shrinker is more helpful the more cold memory we
> > > > have.
> > > >
> > > > Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
> > > > Signed-off-by: Nhat Pham <nphamcs@gmail.com>
> > > > ---
> > > >  Documentation/admin-guide/mm/zswap.rst |  12 ++
> > > >  include/linux/memcontrol.h             |   1 +
> > > >  include/linux/mmzone.h                 |  14 ++
> > > >  mm/memcontrol.c                        |  33 +++++
> > > >  mm/swap_state.c                        |  31 ++++-
> > > >  mm/zswap.c                             | 180 ++++++++++++++++++++++++-
> > > >  6 files changed, 263 insertions(+), 8 deletions(-)
> > > >
> > > > diff --git a/Documentation/admin-guide/mm/zswap.rst b/Documentation/admin-guide/mm/zswap.rst
> > > > index 45b98390e938..ae8597a67804 100644
> > > > --- a/Documentation/admin-guide/mm/zswap.rst
> > > > +++ b/Documentation/admin-guide/mm/zswap.rst
> > > > @@ -153,6 +153,18 @@ attribute, e. g.::
> > > >
> > > >  Setting this parameter to 100 will disable the hysteresis.
> > > >
> > > > +When there is a sizable amount of cold memory residing in the zswap pool, it
> > > > +can be advantageous to proactively write these cold pages to swap and reclaim
> > > > +the memory for other use cases. By default, the zswap shrinker is disabled.
> > > > +User can enable it by first switching on the global knob:
> > > > +
> > > > +  echo Y > /sys/module/zswap/par meters/shrinker_enabled
> > > > +
> > > > +When the kernel is compiled with CONFIG_MEMCG_KMEM, user needs to further turn
> > > > +it on for each cgroup that the shrinker should target:
> > > > +
> > > > +  echo 1 > /sys/fs/cgroup/<cgroup-name>/memory.zswap.shrinker.enabled
> > > > +
> > > >  A debugfs interface is provided for various statistic about pool size, number
> > > >  of pages stored, same-value filled pages and various counters for the reasons
> > > >  pages are rejected.
> > > > diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
> > > > index 05d34b328d9d..f005ea667863 100644
> > > > --- a/include/linux/memcontrol.h
> > > > +++ b/include/linux/memcontrol.h
> > > > @@ -219,6 +219,7 @@ struct mem_cgroup {
> > > >
> > > >  #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
> > > >         unsigned long zswap_max;
> > > > +       atomic_t zswap_shrinker_enabled;
> > > >  #endif
> > > >
> > > >         unsigned long soft_limit;
> > > > diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
> > > > index 4106fbc5b4b3..81f4c5ea3e16 100644
> > > > --- a/include/linux/mmzone.h
> > > > +++ b/include/linux/mmzone.h
> > > > @@ -637,6 +637,20 @@ struct lruvec {
> > > >  #ifdef CONFIG_MEMCG
> > > >         struct pglist_data *pgdat;
> > > >  #endif
> > > > +#ifdef CONFIG_ZSWAP
> > > > +       /*
> > > > +        * Number of pages in zswap that should be protected from the shrinker.
> > > > +        * This number is an estimate of the following counts:
> > > > +        *
> > > > +        * a) Recent page faults.
> > > > +        * b) Recent insertion to the zswap LRU. This includes new zswap stores,
> > > > +        *    as well as recent zswap LRU rotations.
> > > > +        *
> > > > +        * These pages are likely to be warm, and might incur IO if the are written
> > > > +        * to swap.
> > > > +        */
> > > > +       unsigned long nr_zswap_protected;
> > > > +#endif
> > >
> > > Would this be better abstracted in a zswap lruvec struct?
> > There is just one field, so that sounds like overkill to me.
> > But if we need to store more data (for smarter heuristics),
> > that'll be a good idea. I'll keep this in mind. Thanks for the
> > suggestion, Yosry!
>
> (A space between the quoted text and the reply usually helps visually :)
>
> It wasn't really about the number of fields, but rather place this
> struct in zswap.h (with the long comment explaining what it's doing),
> and adding an abstracted struct member here. The comment will live in
> an appropriate file, further modifications don't need to touch
> mmzone.h, and struct lruvec is less cluttered for readers that don't
> care about zswap (and we can avoid the ifdef).
>
> Anyway, this is all mostly aesthetic so I don't feel strongly.
>
> > >
> > > >  };
> > > >
> > > >  /* Isolate unmapped pages */
> > > > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > > > index 9f84b3f7b469..1a2c97cf396f 100644
> > > > --- a/mm/memcontrol.c
> > > > +++ b/mm/memcontrol.c
> > > > @@ -5352,6 +5352,8 @@ mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
> > > >         WRITE_ONCE(memcg->soft_limit, PAGE_COUNTER_MAX);
> > > >  #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
> > > >         memcg->zswap_max = PAGE_COUNTER_MAX;
> > > > +       /* Disable the shrinker by default */
> > > > +       atomic_set(&memcg->zswap_shrinker_enabled, 0);
> > > >  #endif
> > > >         page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
> > > >         if (parent) {
> > > > @@ -7877,6 +7879,31 @@ static ssize_t zswap_max_write(struct kernfs_open_file *of,
> > > >         return nbytes;
> > > >  }
> > > >
> > > > +static int zswap_shrinker_enabled_show(struct seq_file *m, void *v)
> > > > +{
> > > > +       struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
> > > > +
> > > > +       seq_printf(m, "%d\n", atomic_read(&memcg->zswap_shrinker_enabled));
> > > > +       return 0;
> > > > +}
> > > > +
> > > > +static ssize_t zswap_shrinker_enabled_write(struct kernfs_open_file *of,
> > > > +                              char *buf, size_t nbytes, loff_t off)
> > > > +{
> > > > +       struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
> > > > +       int zswap_shrinker_enabled;
> > > > +       ssize_t parse_ret = kstrtoint(strstrip(buf), 0, &zswap_shrinker_enabled);
> > > > +
> > > > +       if (parse_ret)
> > > > +               return parse_ret;
> > > > +
> > > > +       if (zswap_shrinker_enabled < 0 || zswap_shrinker_enabled > 1)
> > > > +               return -ERANGE;
> > > > +
> > > > +       atomic_set(&memcg->zswap_shrinker_enabled, zswap_shrinker_enabled);
> > > > +       return nbytes;
> > > > +}
> > > > +
> > > >  static struct cftype zswap_files[] = {
> > > >         {
> > > >                 .name = "zswap.current",
> > > > @@ -7889,6 +7916,12 @@ static struct cftype zswap_files[] = {
> > > >                 .seq_show = zswap_max_show,
> > > >                 .write = zswap_max_write,
> > > >         },
> > > > +       {
> > > > +               .name = "zswap.shrinker.enabled",
> > > > +               .flags = CFTYPE_NOT_ON_ROOT,
> > > > +               .seq_show = zswap_shrinker_enabled_show,
> > > > +               .write = zswap_shrinker_enabled_write,
> > > > +       },
> > > >         { }     /* terminate */
> > > >  };
> > > >  #endif /* CONFIG_MEMCG_KMEM && CONFIG_ZSWAP */
> > > > diff --git a/mm/swap_state.c b/mm/swap_state.c
> > > > index 1c826737aacb..788e36a06c34 100644
> > > > --- a/mm/swap_state.c
> > > > +++ b/mm/swap_state.c
> > > > @@ -618,6 +618,22 @@ static unsigned long swapin_nr_pages(unsigned long offset)
> > > >         return pages;
> > > >  }
> > > >
> > > > +#ifdef CONFIG_ZSWAP
> > > > +/*
> > > > + * Refault is an indication that warmer pages are not resident in memory.
> > > > + * Increase the size of zswap's protected area.
> > > > + */
> > > > +static void inc_nr_protected(struct page *page)
> > > > +{
> > > > +       struct lruvec *lruvec = folio_lruvec(page_folio(page));
> > > > +       unsigned long flags;
> > > > +
> > > > +       spin_lock_irqsave(&lruvec->lru_lock, flags);
> > > > +       lruvec->nr_zswap_protected++;
> > > > +       spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > > > +}
> > > > +#endif
> > > > +
> > >
> > > A few questions:
> > > - Why is this function named in such a generic way?
> > Perhaps inc_nr_zswap_protected would be better? :)
>
> If we use an atomic, the function can go away anyway. See below.
>
> > > - Why is this function here instead of in mm/zswap.c?
> > No particular reason :) It's not being used anywhere else,
> > so I just put it as a static function here.
>
> It is inline in mm/zswap.c in one place. I personally would have
> preferred nr_zswap_protected and the helper to be defined in
> zswap.h/zswap.c as I mentioned below. Anyway, this function can go
> away.
>
> > > - Why is this protected by the heavily contested lruvec lock instead
> > > of being an atomic?
> > nr_zswap_protected can be decayed (see zswap_lru_add), which
> > I don't think it can be implemented with atomics :( It'd be much
> > cleaner indeed.
>
> I think a cmpxchg (or a try_cmpxchg) loop can be used in this case to
> implement it using an atomic?
>
> See https://docs.kernel.org/core-api/wrappers/atomic_t.html.

Ah I did think about this, but that seems overkill at the time.
But if lruvec lock is indeed hotly contested, this should help.

>
> > > > +               lruvec->nr_zswap_protected++;
> > > >
> > > > +               /*
> > > > +                * Decay to avoid overflow and adapt to changing workloads.
> > > > +                * This is based on LRU reclaim cost decaying heuristics.
> > > > +                */
> > > > +               if (lruvec->nr_zswap_protected > lru_size / 4)
> > > > +                       lruvec->nr_zswap_protected /= 2;
>
> >
> > I'm wary of adding new locks, so I just re-use this existing lock.
> > But if lruvec lock is heavily congested (I'm not aware/familar with
> > this issue), then perhaps a new, dedicated lock would help?
> > >
> > > >  /**
> > > >   * swap_cluster_readahead - swap in pages in hope we need them soon
> > > >   * @entry: swap entry of this memory
> > > > @@ -686,7 +702,12 @@ struct page *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask,
> > > >         lru_add_drain();        /* Push any new pages onto the LRU now */
> > > >  skip:
> > > >         /* The page was likely read above, so no need for plugging here */
> > > > -       return read_swap_cache_async(entry, gfp_mask, vma, addr, NULL);
> > > > +       page = read_swap_cache_async(entry, gfp_mask, vma, addr, NULL);
> > > > +#ifdef CONFIG_ZSWAP
> > > > +       if (page)
> > > > +               inc_nr_protected(page);
> > > > +#endif
> > > > +       return page;
> > > >  }
> > > >
> > > >  int init_swap_address_space(unsigned int type, unsigned long nr_pages)
> > > > @@ -853,8 +874,12 @@ static struct page *swap_vma_readahead(swp_entry_t fentry, gfp_t gfp_mask,
> > > >         lru_add_drain();
> > > >  skip:
> > > >         /* The page was likely read above, so no need for plugging here */
> > > > -       return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address,
> > > > -                                    NULL);
> > > > +       page = read_swap_cache_async(fentry, gfp_mask, vma, vmf->address, NULL);
> > > > +#ifdef CONFIG_ZSWAP
> > > > +       if (page)
> > > > +               inc_nr_protected(page);
> > > > +#endif
> > > > +       return page;
> > > >  }
> > > >
> > > >  /**
> > > > diff --git a/mm/zswap.c b/mm/zswap.c
> > > > index 1a469e5d5197..79cb18eeb8bf 100644
> > > > --- a/mm/zswap.c
> > > > +++ b/mm/zswap.c
> > > > @@ -145,6 +145,26 @@ module_param_named(exclusive_loads, zswap_exclusive_loads_enabled, bool, 0644);
> > > >  /* Number of zpools in zswap_pool (empirically determined for scalability) */
> > > >  #define ZSWAP_NR_ZPOOLS 32
> > > >
> > > > +/*
> > > > + * Global flag to enable/disable memory pressure-based shrinker for all memcgs.
> > > > + * If CONFIG_MEMCG_KMEM is on, we can further selectively disable
> > > > + * the shrinker for each memcg.
> > > > + */
> > > > +static bool zswap_shrinker_enabled;
> > > > +module_param_named(shrinker_enabled, zswap_shrinker_enabled, bool, 0644);
> > > > +#ifdef CONFIG_MEMCG_KMEM
> > > > +static bool is_shrinker_enabled(struct mem_cgroup *memcg)
> > > > +{
> > > > +       return zswap_shrinker_enabled &&
> > > > +               atomic_read(&memcg->zswap_shrinker_enabled);
> > > > +}
> > > > +#else
> > > > +static bool is_shrinker_enabled(struct mem_cgroup *memcg)
> > > > +{
> > > > +       return zswap_shrinker_enabled;
> > > > +}
> > > > +#endif
> > > > +
> > > >  /*********************************
> > > >  * data structures
> > > >  **********************************/
> > > > @@ -174,6 +194,8 @@ struct zswap_pool {
> > > >         char tfm_name[CRYPTO_MAX_ALG_NAME];
> > > >         struct list_lru list_lru;
> > > >         struct mem_cgroup *next_shrink;
> > > > +       struct shrinker *shrinker;
> > > > +       atomic_t nr_stored;
> > > >  };
> > > >
> > > >  /*
> > > > @@ -273,17 +295,26 @@ static bool zswap_can_accept(void)
> > > >                         DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
> > > >  }
> > > >
> > > > +static u64 get_zswap_pool_size(struct zswap_pool *pool)
> > > > +{
> > > > +       u64 pool_size = 0;
> > > > +       int i;
> > > > +
> > > > +       for (i = 0; i < ZSWAP_NR_ZPOOLS; i++)
> > > > +               pool_size += zpool_get_total_size(pool->zpools[i]);
> > > > +
> > > > +       return pool_size;
> > > > +}
> > > > +
> > > >  static void zswap_update_total_size(void)
> > > >  {
> > > >         struct zswap_pool *pool;
> > > >         u64 total = 0;
> > > > -       int i;
> > > >
> > > >         rcu_read_lock();
> > > >
> > > >         list_for_each_entry_rcu(pool, &zswap_pools, list)
> > > > -               for (i = 0; i < ZSWAP_NR_ZPOOLS; i++)
> > > > -                       total += zpool_get_total_size(pool->zpools[i]);
> > > > +               total += get_zswap_pool_size(pool);
> > > >
> > > >         rcu_read_unlock();
> > > >
> > > > @@ -318,8 +349,23 @@ static bool zswap_lru_add(struct list_lru *list_lru, struct zswap_entry *entry)
> > > >  {
> > > >         struct mem_cgroup *memcg = entry->objcg ?
> > > >                 get_mem_cgroup_from_objcg(entry->objcg) : NULL;
> > > > +       struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(entry->nid));
> > > >         bool added = __list_lru_add(list_lru, &entry->lru, entry->nid, memcg);
> > > > +       unsigned long flags, lru_size;
> > > > +
> > > > +       if (added) {
> > > > +               lru_size = list_lru_count_one(list_lru, entry->nid, memcg);
> > > > +               spin_lock_irqsave(&lruvec->lru_lock, flags);
> > > > +               lruvec->nr_zswap_protected++;
> > > >
> > > > +               /*
> > > > +                * Decay to avoid overflow and adapt to changing workloads.
> > > > +                * This is based on LRU reclaim cost decaying heuristics.
> > > > +                */
> > > > +               if (lruvec->nr_zswap_protected > lru_size / 4)
> > > > +                       lruvec->nr_zswap_protected /= 2;
> > > > +               spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > > > +       }
> > > >         mem_cgroup_put(memcg);
> > > >         return added;
> > > >  }
> > > > @@ -420,6 +466,7 @@ static void zswap_free_entry(struct zswap_entry *entry)
> > > >         else {
> > > >                 zswap_lru_del(&entry->pool->list_lru, entry);
> > > >                 zpool_free(zswap_find_zpool(entry), entry->handle);
> > > > +               atomic_dec(&entry->pool->nr_stored);
> > > >                 zswap_pool_put(entry->pool);
> > > >         }
> > > >         zswap_entry_cache_free(entry);
> > > > @@ -461,6 +508,98 @@ static struct zswap_entry *zswap_entry_find_get(struct rb_root *root,
> > > >         return entry;
> > > >  }
> > > >
> > > > +/*********************************
> > > > +* shrinker functions
> > > > +**********************************/
> > > > +static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_one *l,
> > > > +                                      spinlock_t *lock, void *arg);
> > > > +
> > > > +static unsigned long zswap_shrinker_scan(struct shrinker *shrinker,
> > > > +               struct shrink_control *sc)
> > > > +{
> > > > +       struct zswap_pool *pool = shrinker->private_data;
> > > > +       unsigned long shrink_ret, nr_zswap_protected, flags,
> > > > +               lru_size = list_lru_shrink_count(&pool->list_lru, sc);
> > > > +       struct lruvec *lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid));
> > > > +       bool encountered_page_in_swapcache = false;
> > > > +
> > > > +       spin_lock_irqsave(&lruvec->lru_lock, flags);
> > > > +       nr_zswap_protected = lruvec->nr_zswap_protected;
> > > > +       spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > > > +
> > > > +       /*
> > > > +        * Abort if the shrinker is disabled or if we are shrinking into the
> > > > +        * protected region.
> > > > +        */
> > > > +       if (!is_shrinker_enabled(sc->memcg) ||
> > > > +                       nr_zswap_protected >= lru_size - sc->nr_to_scan) {
> > > > +               sc->nr_scanned = 0;
> > > > +               return SHRINK_STOP;
> > > > +       }
> > > > +
> > > > +       shrink_ret = list_lru_shrink_walk(&pool->list_lru, sc, &shrink_memcg_cb,
> > > > +               &encountered_page_in_swapcache);
> > > > +
> > > > +       if (encountered_page_in_swapcache)
> > > > +               return SHRINK_STOP;
> > > > +
> > > > +       return shrink_ret ? shrink_ret : SHRINK_STOP;
> > > > +}
> > > > +
> > > > +static unsigned long zswap_shrinker_count(struct shrinker *shrinker,
> > > > +               struct shrink_control *sc)
> > > > +{
> > > > +       struct zswap_pool *pool = shrinker->private_data;
> > > > +       struct mem_cgroup *memcg = sc->memcg;
> > > > +       struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(sc->nid));
> > > > +       unsigned long nr_backing, nr_stored, nr_freeable, flags;
> > > > +
> > > > +#ifdef CONFIG_MEMCG_KMEM
> > > > +       cgroup_rstat_flush(memcg->css.cgroup);
> > > > +       nr_backing = memcg_page_state(memcg, MEMCG_ZSWAP_B) >> PAGE_SHIFT;
> > > > +       nr_stored = memcg_page_state(memcg, MEMCG_ZSWAPPED);
> > > > +#else
> > > > +       /* use pool stats instead of memcg stats */
> > > > +       nr_backing = get_zswap_pool_size(pool) >> PAGE_SHIFT;
> > > > +       nr_stored = atomic_read(&pool->nr_stored);
> > > > +#endif
> > > > +
> > > > +       if (!is_shrinker_enabled(memcg) || !nr_stored)
> > > > +               return 0;
> > > > +
> > > > +       nr_freeable = list_lru_shrink_count(&pool->list_lru, sc);
> > > > +       /*
> > > > +        * Subtract the lru size by an estimate of the number of pages
> > > > +        * that should be protected.
> > > > +        */
> > > > +       spin_lock_irqsave(&lruvec->lru_lock, flags);
> > > > +       nr_freeable = nr_freeable > lruvec->nr_zswap_protected ?
> > > > +               nr_freeable - lruvec->nr_zswap_protected : 0;
> > > > +       spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > > > +
> > > > +       /*
> > > > +        * Scale the number of freeable pages by the memory saving factor.
> > > > +        * This ensures that the better zswap compresses memory, the fewer
> > > > +        * pages we will evict to swap (as it will otherwise incur IO for
> > > > +        * relatively small memory saving).
> > > > +        */
> > > > +       return mult_frac(nr_freeable, nr_backing, nr_stored);
> > > > +}
> > > > +
> > > > +static void zswap_alloc_shrinker(struct zswap_pool *pool)
> > > > +{
> > > > +       pool->shrinker =
> > > > +               shrinker_alloc(SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE, "mm-zswap");
> > > > +       if (!pool->shrinker)
> > > > +               return;
> > > > +
> > > > +       pool->shrinker->private_data = pool;
> > > > +       pool->shrinker->scan_objects = zswap_shrinker_scan;
> > > > +       pool->shrinker->count_objects = zswap_shrinker_count;
> > > > +       pool->shrinker->batch = 0;
> > > > +       pool->shrinker->seeks = DEFAULT_SEEKS;
> > > > +}
> > > > +
> > > >  /*********************************
> > > >  * per-cpu code
> > > >  **********************************/
> > > > @@ -656,11 +795,14 @@ static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_o
> > > >                                        spinlock_t *lock, void *arg)
> > > >  {
> > > >         struct zswap_entry *entry = container_of(item, struct zswap_entry, lru);
> > > > +       bool *encountered_page_in_swapcache = (bool *)arg;
> > > >         struct mem_cgroup *memcg;
> > > >         struct zswap_tree *tree;
> > > > +       struct lruvec *lruvec;
> > > >         pgoff_t swpoffset;
> > > >         enum lru_status ret = LRU_REMOVED_RETRY;
> > > >         int writeback_result;
> > > > +       unsigned long flags;
> > > >
> > > >         /*
> > > >          * Once the lru lock is dropped, the entry might get freed. The
> > > > @@ -696,8 +838,24 @@ static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_o
> > > >                 /* we cannot use zswap_lru_add here, because it increments node's lru count */
> > > >                 list_lru_putback(&entry->pool->list_lru, item, entry->nid, memcg);
> > > >                 spin_unlock(lock);
> > > > -               mem_cgroup_put(memcg);
> > > >                 ret = LRU_RETRY;
> > > > +
> > > > +               /*
> > > > +                * Encountering a page already in swap cache is a sign that we are shrinking
> > > > +                * into the warmer region. We should terminate shrinking (if we're in the dynamic
> > > > +                * shrinker context).
> > > > +                */
> > > > +               if (writeback_result == -EEXIST && encountered_page_in_swapcache) {
> > > > +                       ret = LRU_SKIP;
> > > > +                       *encountered_page_in_swapcache = true;
> > > > +               }
> > > > +               lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(entry->nid));
> > > > +               spin_lock_irqsave(&lruvec->lru_lock, flags);
> > > > +               /* Increment the protection area to account for the LRU rotation. */
> > > > +               lruvec->nr_zswap_protected++;
> > > > +               spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > > > +
> > > > +               mem_cgroup_put(memcg);
> > > >                 goto put_unlock;
> > > >         }
> > > >
> > > > @@ -828,6 +986,11 @@ static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
> > > >                                        &pool->node);
> > > >         if (ret)
> > > >                 goto error;
> > > > +
> > > > +       zswap_alloc_shrinker(pool);
> > > > +       if (!pool->shrinker)
> > > > +               goto error;
> > > > +
> > > >         pr_debug("using %s compressor\n", pool->tfm_name);
> > > >
> > > >         /* being the current pool takes 1 ref; this func expects the
> > > > @@ -836,12 +999,17 @@ static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
> > > >         kref_init(&pool->kref);
> > > >         INIT_LIST_HEAD(&pool->list);
> > > >         INIT_WORK(&pool->shrink_work, shrink_worker);
> > > > -       list_lru_init_memcg(&pool->list_lru, NULL);
> > > > +       if (list_lru_init_memcg(&pool->list_lru, pool->shrinker))
> > > > +               goto lru_fail;
> > > > +       shrinker_register(pool->shrinker);
> > > >
> > > >         zswap_pool_debug("created", pool);
> > > >
> > > >         return pool;
> > > >
> > > > +lru_fail:
> > > > +       list_lru_destroy(&pool->list_lru);
> > > > +       shrinker_free(pool->shrinker);
> > > >  error:
> > > >         if (pool->acomp_ctx)
> > > >                 free_percpu(pool->acomp_ctx);
> > > > @@ -899,6 +1067,7 @@ static void zswap_pool_destroy(struct zswap_pool *pool)
> > > >
> > > >         zswap_pool_debug("destroying", pool);
> > > >
> > > > +       shrinker_free(pool->shrinker);
> > > >         cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
> > > >         free_percpu(pool->acomp_ctx);
> > > >         list_lru_destroy(&pool->list_lru);
> > > > @@ -1431,6 +1600,7 @@ bool zswap_store(struct folio *folio)
> > > >         if (entry->length) {
> > > >                 INIT_LIST_HEAD(&entry->lru);
> > > >                 zswap_lru_add(&pool->list_lru, entry);
> > > > +               atomic_inc(&pool->nr_stored);
> > > >         }
> > > >         spin_unlock(&tree->lock);
> > > >
> > > > --
> > > > 2.34.1
> > Thanks for the comments/suggestion, Yosry!
  

On Mon, Sep 25, 2023 at 5:43 PM Nhat Pham <nphamcs@gmail.com> wrote:
>
> On Mon, Sep 25, 2023 at 5:00 PM Yosry Ahmed <yosryahmed@google.com> wrote:
> >
> > On Mon, Sep 25, 2023 at 4:29 PM Nhat Pham <nphamcs@gmail.com> wrote:
> > >
> > > On Mon, Sep 25, 2023 at 1:38 PM Yosry Ahmed <yosryahmed@google.com> wrote:
> > > >
> > > > On Tue, Sep 19, 2023 at 10:14 AM Nhat Pham <nphamcs@gmail.com> wrote:
> > > > >
> > > > > Currently, we only shrink the zswap pool when the user-defined limit is
> > > > > hit. This means that if we set the limit too high, cold data that are
> > > > > unlikely to be used again will reside in the pool, wasting precious
> > > > > memory. It is hard to predict how much zswap space will be needed ahead
> > > > > of time, as this depends on the workload (specifically, on factors such
> > > > > as memory access patterns and compressibility of the memory pages).
> > > > >
> > > > > This patch implements a memcg- and NUMA-aware shrinker for zswap, that
> > > > > is initiated when there is memory pressure. The shrinker does not
> > > > > have any parameter that must be tuned by the user, and can be opted in
> > > > > or out on a per-memcg basis.
> > > >
> > > > What's the use case for having per-memcg opt-in/out?
> > > >
> > > > If there is memory pressure, reclaiming swap-backed pages will push
> > > > pages out of zswap anyway, regardless of this patch. With this patch,
> > > > any sort of reclaim can push pages out of zswap. Wouldn't that be
> > > > preferable to reclaiming memory that is currently resident in memory
> > > > (so arguably hotter than the pages in zswap)? Why would this decision
> > > > be different per-memcg?
> > > I'm not quite following your argument here. The point of having this
> > > be done on a per-memcg basis is that we have different workloads
> > > with different memory access pattern (and as a result, different memory
> > > coldness distribution).
> > >
> > > In a workload where there is a lot of cold data, we can really benefit
> > > from reclaiming all of those pages and repurpose the memory reclaimed
> > > (for e.g for filecache).
> > >
> > > On the other hand, in a workload where there aren't a lot of cold data,
> > > reclaiming its zswapped pages will at best do nothing (wasting CPU
> > > cycles on compression/decompression), and at worst hurt performance
> > > (due to increased IO when we need those writtenback pages again).
> > >
> > > Such different workloads could co-exist in the same system, and having
> > > a per-memcg knob allows us to crank on the shrinker only on workloads
> > > where it makes sense.
> >
> > I am not sure we are on the same page here.
> >
> > What you're describing sounds more like proactive reclaim, which we
> > wouldn't invoke unless the workload has cold data anyway.
> >
> > IIUC, outside of that, this shrinker will run when there is memory
> > pressure. This means that we need to free memory anyway, regardless of
> > its absolute coldness. We want to evict the colder pages in the memcg.
> > It seems to be that in ~all cases, evicting pages in zswap will be
> > better than evicting pages in memory, as the pages in memory are
> > arguably hotter (since they weren't reclaimed first). This seems to be
> > something that would be true for all workloads.
> >
> > What am I missing?
>
> Yup, the shrinker is initiated under memory pressure.
> And with it, we can reclaim memory from zswap when
> it's (often) not at max capacity.
>
> The kernel has no knowledge of absolute coldness, only relative
> coldness thanks to LRU. We don't have a global LRU of all possible
> memory pages/objects for a particular memcg either, so we cannot
> compare the coldness of objects from different sources.
>
> The "coldest" pages in zswap LRU could very well be warm enough
> that swapping them out degrades performance, while there are even
> colder memory from other sources (other shrinkers registered for this
> memcg). Alternatively, we can also "evict" uncompressed anonymous
> memory, which will go to the zswap pool. This also saves memory,
> and could potentially be better than zswap reclaim (2 compressed
> pages might be better performance-wise than 1 uncompressed,
> 1 swapped out)
>
> All of this depends on the memory access pattern of the workloads,
> which could differ cgroup-by-cgroup within the same system.
> Having a per-memcg knob is a way for admins to influence this
> decision from userspace, if the admins have knowledge about
> workload memory access patterns.
>
> For e.g, if we know that there is one particular cgroup that populates
> a bunch of single-use tmpfs pages, then we can target that cgroup
> specifically, while leaving the other cgroups in the system alone.

I think it's useful to break down the discussion here for cgroup
reclaim and global reclaim.

For cgroup reclaim, the kernel knows that the pages in the LRUs are
relatively hotter than the pages in zswap. So I don't see why
userspace would opt out specific cgroups from zswap shrinking. In my
experience, most memory usage comes from LRU pages, so let's ignore
other shrinkers for a second. Yes, in some cases compressing another
page might be better than moving a compressed page to swap, but how
would userspace have the intuition to decide this? It varies not only
based on workload, but also the point in time, the compressibility of
pages, etc.

In other words, how would a system admin choose to opt a cgroup in or out?

For global reclaim, IIUC you are saying that we want to protect some
cgroups under global memory pressure because we know that their "cold"
memory in zswap is hotter than memory elsewhere in the hierarchy,
right?

Isn't this the case for LRU reclaim as well? I would assume memory
protections would be used to tune this, not opting a cgroup completely
from zswap shrinking. Global reclaim can end up reclaiming LRU pages
from that cgroup if protection is not set up correctly anyway. What do
we gain by protecting pages in zswap if hotter pages in the LRUs are
not protected?

>
> >
> > > >
> > > > >
> > > > > Furthermore, to make it more robust for many workloads and prevent
> > > > > overshrinking (i.e evicting warm pages that might be refaulted into
> > > > > memory), we build in the following heuristics:
> > > > >
> > > > > * Estimate the number of warm pages residing in zswap, and attempt to
> > > > >   protect this region of the zswap LRU.
> > > > > * Scale the number of freeable objects by an estimate of the memory
> > > > >   saving factor. The better zswap compresses the data, the fewer pages
> > > > >   we will evict to swap (as we will otherwise incur IO for relatively
> > > > >   small memory saving).
> > > > > * During reclaim, if the shrinker encounters a page that is also being
> > > > >   brought into memory, the shrinker will cautiously terminate its
> > > > >   shrinking action, as this is a sign that it is touching the warmer
> > > > >   region of the zswap LRU.
> > > >
> > > > I don't have an opinion about the reclaim heuristics here, I will let
> > > > reclaim experts chip in.
> > > >
> > > > >
> > > > > On a benchmark that we have run:
> > > >
> > > > Please add more details (as much as possible) about the benchmarks used here.
> > > Sure! I built the kernel in a memory-limited cgroup a couple times,
> > > then measured the build time.
> > >
> > > To simulate conditions where there are cold, unused data, I
> > > also generated a bunch of data in tmpfs (and never touch them
> > > again).
> >
> > Please include such details in the commit message, there is also
> > another reference below to "another" benchmark.
>
> Will do if/when I send v3.
> The "another" benchmark is just generating even more tmpfs cold data :)

Those benchmarks are heavily synthetic, which is not a showstopper,
but describing them in the commit message helps people reason about
the change.

>
> >
> >
> > > >
> > > > >
> > > > > (without the shrinker)
> > > > > real -- mean: 153.27s, median: 153.199s
> > > > > sys -- mean: 541.652s, median: 541.903s
> > > > > user -- mean: 4384.9673999999995s, median: 4385.471s
> > > > >
> > > > > (with the shrinker)
> > > > > real -- mean: 151.4956s, median: 151.456s
> > > > > sys -- mean: 461.14639999999997s, median: 465.656s
> > > > > user -- mean: 4384.7118s, median: 4384.675s
> > > > >
> > > > > We observed a 14-15% reduction in kernel CPU time, which translated to
> > > > > over 1% reduction in real time.
> > > > >
> > > > > On another benchmark, where there was a lot more cold memory residing in
> > > > > zswap, we observed even more pronounced gains:
> > > > >
> > > > > (without the shrinker)
> > > > > real -- mean: 157.52519999999998s, median: 157.281s
> > > > > sys -- mean: 769.3082s, median: 780.545s
> > > > > user -- mean: 4378.1622s, median: 4378.286s
> > > > >
> > > > > (with the shrinker)
> > > > > real -- mean: 152.9608s, median: 152.845s
> > > > > sys -- mean: 517.4446s, median: 506.749s
> > > > > user -- mean: 4387.694s, median: 4387.935s
> > > > >
> > > > > Here, we saw around 32-35% reduction in kernel CPU time, which
> > > > > translated to 2.8% reduction in real time. These results confirm our
> > > > > hypothesis that the shrinker is more helpful the more cold memory we
> > > > > have.
> > > > >
> > > > > Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
> > > > > Signed-off-by: Nhat Pham <nphamcs@gmail.com>
> > > > > ---
> > > > >  Documentation/admin-guide/mm/zswap.rst |  12 ++
> > > > >  include/linux/memcontrol.h             |   1 +
> > > > >  include/linux/mmzone.h                 |  14 ++
> > > > >  mm/memcontrol.c                        |  33 +++++
> > > > >  mm/swap_state.c                        |  31 ++++-
> > > > >  mm/zswap.c                             | 180 ++++++++++++++++++++++++-
> > > > >  6 files changed, 263 insertions(+), 8 deletions(-)
> > > > >
> > > > > diff --git a/Documentation/admin-guide/mm/zswap.rst b/Documentation/admin-guide/mm/zswap.rst
> > > > > index 45b98390e938..ae8597a67804 100644
> > > > > --- a/Documentation/admin-guide/mm/zswap.rst
> > > > > +++ b/Documentation/admin-guide/mm/zswap.rst
> > > > > @@ -153,6 +153,18 @@ attribute, e. g.::
> > > > >
> > > > >  Setting this parameter to 100 will disable the hysteresis.
> > > > >
> > > > > +When there is a sizable amount of cold memory residing in the zswap pool, it
> > > > > +can be advantageous to proactively write these cold pages to swap and reclaim
> > > > > +the memory for other use cases. By default, the zswap shrinker is disabled.
> > > > > +User can enable it by first switching on the global knob:
> > > > > +
> > > > > +  echo Y > /sys/module/zswap/par meters/shrinker_enabled
> > > > > +
> > > > > +When the kernel is compiled with CONFIG_MEMCG_KMEM, user needs to further turn
> > > > > +it on for each cgroup that the shrinker should target:
> > > > > +
> > > > > +  echo 1 > /sys/fs/cgroup/<cgroup-name>/memory.zswap.shrinker.enabled
> > > > > +
> > > > >  A debugfs interface is provided for various statistic about pool size, number
> > > > >  of pages stored, same-value filled pages and various counters for the reasons
> > > > >  pages are rejected.
> > > > > diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
> > > > > index 05d34b328d9d..f005ea667863 100644
> > > > > --- a/include/linux/memcontrol.h
> > > > > +++ b/include/linux/memcontrol.h
> > > > > @@ -219,6 +219,7 @@ struct mem_cgroup {
> > > > >
> > > > >  #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
> > > > >         unsigned long zswap_max;
> > > > > +       atomic_t zswap_shrinker_enabled;
> > > > >  #endif
> > > > >
> > > > >         unsigned long soft_limit;
> > > > > diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
> > > > > index 4106fbc5b4b3..81f4c5ea3e16 100644
> > > > > --- a/include/linux/mmzone.h
> > > > > +++ b/include/linux/mmzone.h
> > > > > @@ -637,6 +637,20 @@ struct lruvec {
> > > > >  #ifdef CONFIG_MEMCG
> > > > >         struct pglist_data *pgdat;
> > > > >  #endif
> > > > > +#ifdef CONFIG_ZSWAP
> > > > > +       /*
> > > > > +        * Number of pages in zswap that should be protected from the shrinker.
> > > > > +        * This number is an estimate of the following counts:
> > > > > +        *
> > > > > +        * a) Recent page faults.
> > > > > +        * b) Recent insertion to the zswap LRU. This includes new zswap stores,
> > > > > +        *    as well as recent zswap LRU rotations.
> > > > > +        *
> > > > > +        * These pages are likely to be warm, and might incur IO if the are written
> > > > > +        * to swap.
> > > > > +        */
> > > > > +       unsigned long nr_zswap_protected;
> > > > > +#endif
> > > >
> > > > Would this be better abstracted in a zswap lruvec struct?
> > > There is just one field, so that sounds like overkill to me.
> > > But if we need to store more data (for smarter heuristics),
> > > that'll be a good idea. I'll keep this in mind. Thanks for the
> > > suggestion, Yosry!
> >
> > (A space between the quoted text and the reply usually helps visually :)
> >
> > It wasn't really about the number of fields, but rather place this
> > struct in zswap.h (with the long comment explaining what it's doing),
> > and adding an abstracted struct member here. The comment will live in
> > an appropriate file, further modifications don't need to touch
> > mmzone.h, and struct lruvec is less cluttered for readers that don't
> > care about zswap (and we can avoid the ifdef).
> >
> > Anyway, this is all mostly aesthetic so I don't feel strongly.
> >
> > > >
> > > > >  };
> > > > >
> > > > >  /* Isolate unmapped pages */
> > > > > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > > > > index 9f84b3f7b469..1a2c97cf396f 100644
> > > > > --- a/mm/memcontrol.c
> > > > > +++ b/mm/memcontrol.c
> > > > > @@ -5352,6 +5352,8 @@ mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
> > > > >         WRITE_ONCE(memcg->soft_limit, PAGE_COUNTER_MAX);
> > > > >  #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
> > > > >         memcg->zswap_max = PAGE_COUNTER_MAX;
> > > > > +       /* Disable the shrinker by default */
> > > > > +       atomic_set(&memcg->zswap_shrinker_enabled, 0);
> > > > >  #endif
> > > > >         page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
> > > > >         if (parent) {
> > > > > @@ -7877,6 +7879,31 @@ static ssize_t zswap_max_write(struct kernfs_open_file *of,
> > > > >         return nbytes;
> > > > >  }
> > > > >
> > > > > +static int zswap_shrinker_enabled_show(struct seq_file *m, void *v)
> > > > > +{
> > > > > +       struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
> > > > > +
> > > > > +       seq_printf(m, "%d\n", atomic_read(&memcg->zswap_shrinker_enabled));
> > > > > +       return 0;
> > > > > +}
> > > > > +
> > > > > +static ssize_t zswap_shrinker_enabled_write(struct kernfs_open_file *of,
> > > > > +                              char *buf, size_t nbytes, loff_t off)
> > > > > +{
> > > > > +       struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
> > > > > +       int zswap_shrinker_enabled;
> > > > > +       ssize_t parse_ret = kstrtoint(strstrip(buf), 0, &zswap_shrinker_enabled);
> > > > > +
> > > > > +       if (parse_ret)
> > > > > +               return parse_ret;
> > > > > +
> > > > > +       if (zswap_shrinker_enabled < 0 || zswap_shrinker_enabled > 1)
> > > > > +               return -ERANGE;
> > > > > +
> > > > > +       atomic_set(&memcg->zswap_shrinker_enabled, zswap_shrinker_enabled);
> > > > > +       return nbytes;
> > > > > +}
> > > > > +
> > > > >  static struct cftype zswap_files[] = {
> > > > >         {
> > > > >                 .name = "zswap.current",
> > > > > @@ -7889,6 +7916,12 @@ static struct cftype zswap_files[] = {
> > > > >                 .seq_show = zswap_max_show,
> > > > >                 .write = zswap_max_write,
> > > > >         },
> > > > > +       {
> > > > > +               .name = "zswap.shrinker.enabled",
> > > > > +               .flags = CFTYPE_NOT_ON_ROOT,
> > > > > +               .seq_show = zswap_shrinker_enabled_show,
> > > > > +               .write = zswap_shrinker_enabled_write,
> > > > > +       },
> > > > >         { }     /* terminate */
> > > > >  };
> > > > >  #endif /* CONFIG_MEMCG_KMEM && CONFIG_ZSWAP */
> > > > > diff --git a/mm/swap_state.c b/mm/swap_state.c
> > > > > index 1c826737aacb..788e36a06c34 100644
> > > > > --- a/mm/swap_state.c
> > > > > +++ b/mm/swap_state.c
> > > > > @@ -618,6 +618,22 @@ static unsigned long swapin_nr_pages(unsigned long offset)
> > > > >         return pages;
> > > > >  }
> > > > >
> > > > > +#ifdef CONFIG_ZSWAP
> > > > > +/*
> > > > > + * Refault is an indication that warmer pages are not resident in memory.
> > > > > + * Increase the size of zswap's protected area.
> > > > > + */
> > > > > +static void inc_nr_protected(struct page *page)
> > > > > +{
> > > > > +       struct lruvec *lruvec = folio_lruvec(page_folio(page));
> > > > > +       unsigned long flags;
> > > > > +
> > > > > +       spin_lock_irqsave(&lruvec->lru_lock, flags);
> > > > > +       lruvec->nr_zswap_protected++;
> > > > > +       spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > > > > +}
> > > > > +#endif
> > > > > +
> > > >
> > > > A few questions:
> > > > - Why is this function named in such a generic way?
> > > Perhaps inc_nr_zswap_protected would be better? :)
> >
> > If we use an atomic, the function can go away anyway. See below.
> >
> > > > - Why is this function here instead of in mm/zswap.c?
> > > No particular reason :) It's not being used anywhere else,
> > > so I just put it as a static function here.
> >
> > It is inline in mm/zswap.c in one place. I personally would have
> > preferred nr_zswap_protected and the helper to be defined in
> > zswap.h/zswap.c as I mentioned below. Anyway, this function can go
> > away.
> >
> > > > - Why is this protected by the heavily contested lruvec lock instead
> > > > of being an atomic?
> > > nr_zswap_protected can be decayed (see zswap_lru_add), which
> > > I don't think it can be implemented with atomics :( It'd be much
> > > cleaner indeed.
> >
> > I think a cmpxchg (or a try_cmpxchg) loop can be used in this case to
> > implement it using an atomic?
> >
> > See https://docs.kernel.org/core-api/wrappers/atomic_t.html.
>
> Ah I did think about this, but that seems overkill at the time.
> But if lruvec lock is indeed hotly contested, this should help.

I wouldn't say so, we can drop numerous calls to grab/drop the lock,
and drop the helper. A try_cmpxchg loop here would only be a couple of
lines, I suspect it would be more concise than the code now:

old = atomic_inc_return(&lruvec->nr_zswap_protected);
do {
       if (old > lru_size / 4)
               new = old / 2;
} while (atomic_try_cmpxchg(&lruvec->nr_zswap_protected, &old, new));

>
> >
> > > > > +               lruvec->nr_zswap_protected++;
> > > > >
> > > > > +               /*
> > > > > +                * Decay to avoid overflow and adapt to changing workloads.
> > > > > +                * This is based on LRU reclaim cost decaying heuristics.
> > > > > +                */
> > > > > +               if (lruvec->nr_zswap_protected > lru_size / 4)
> > > > > +                       lruvec->nr_zswap_protected /= 2;
> >
> > >
> > > I'm wary of adding new locks, so I just re-use this existing lock.
> > > But if lruvec lock is heavily congested (I'm not aware/familar with
> > > this issue), then perhaps a new, dedicated lock would help?
> > > >
> > > > >  /**
> > > > >   * swap_cluster_readahead - swap in pages in hope we need them soon
> > > > >   * @entry: swap entry of this memory
> > > > > @@ -686,7 +702,12 @@ struct page *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask,
> > > > >         lru_add_drain();        /* Push any new pages onto the LRU now */
> > > > >  skip:
> > > > >         /* The page was likely read above, so no need for plugging here */
> > > > > -       return read_swap_cache_async(entry, gfp_mask, vma, addr, NULL);
> > > > > +       page = read_swap_cache_async(entry, gfp_mask, vma, addr, NULL);
> > > > > +#ifdef CONFIG_ZSWAP
> > > > > +       if (page)
> > > > > +               inc_nr_protected(page);
> > > > > +#endif
> > > > > +       return page;
> > > > >  }
> > > > >
> > > > >  int init_swap_address_space(unsigned int type, unsigned long nr_pages)
> > > > > @@ -853,8 +874,12 @@ static struct page *swap_vma_readahead(swp_entry_t fentry, gfp_t gfp_mask,
> > > > >         lru_add_drain();
> > > > >  skip:
> > > > >         /* The page was likely read above, so no need for plugging here */
> > > > > -       return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address,
> > > > > -                                    NULL);
> > > > > +       page = read_swap_cache_async(fentry, gfp_mask, vma, vmf->address, NULL);
> > > > > +#ifdef CONFIG_ZSWAP
> > > > > +       if (page)
> > > > > +               inc_nr_protected(page);
> > > > > +#endif
> > > > > +       return page;
> > > > >  }
> > > > >
> > > > >  /**
> > > > > diff --git a/mm/zswap.c b/mm/zswap.c
> > > > > index 1a469e5d5197..79cb18eeb8bf 100644
> > > > > --- a/mm/zswap.c
> > > > > +++ b/mm/zswap.c
> > > > > @@ -145,6 +145,26 @@ module_param_named(exclusive_loads, zswap_exclusive_loads_enabled, bool, 0644);
> > > > >  /* Number of zpools in zswap_pool (empirically determined for scalability) */
> > > > >  #define ZSWAP_NR_ZPOOLS 32
> > > > >
> > > > > +/*
> > > > > + * Global flag to enable/disable memory pressure-based shrinker for all memcgs.
> > > > > + * If CONFIG_MEMCG_KMEM is on, we can further selectively disable
> > > > > + * the shrinker for each memcg.
> > > > > + */
> > > > > +static bool zswap_shrinker_enabled;
> > > > > +module_param_named(shrinker_enabled, zswap_shrinker_enabled, bool, 0644);
> > > > > +#ifdef CONFIG_MEMCG_KMEM
> > > > > +static bool is_shrinker_enabled(struct mem_cgroup *memcg)
> > > > > +{
> > > > > +       return zswap_shrinker_enabled &&
> > > > > +               atomic_read(&memcg->zswap_shrinker_enabled);
> > > > > +}
> > > > > +#else
> > > > > +static bool is_shrinker_enabled(struct mem_cgroup *memcg)
> > > > > +{
> > > > > +       return zswap_shrinker_enabled;
> > > > > +}
> > > > > +#endif
> > > > > +
> > > > >  /*********************************
> > > > >  * data structures
> > > > >  **********************************/
> > > > > @@ -174,6 +194,8 @@ struct zswap_pool {
> > > > >         char tfm_name[CRYPTO_MAX_ALG_NAME];
> > > > >         struct list_lru list_lru;
> > > > >         struct mem_cgroup *next_shrink;
> > > > > +       struct shrinker *shrinker;
> > > > > +       atomic_t nr_stored;
> > > > >  };
> > > > >
> > > > >  /*
> > > > > @@ -273,17 +295,26 @@ static bool zswap_can_accept(void)
> > > > >                         DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
> > > > >  }
> > > > >
> > > > > +static u64 get_zswap_pool_size(struct zswap_pool *pool)
> > > > > +{
> > > > > +       u64 pool_size = 0;
> > > > > +       int i;
> > > > > +
> > > > > +       for (i = 0; i < ZSWAP_NR_ZPOOLS; i++)
> > > > > +               pool_size += zpool_get_total_size(pool->zpools[i]);
> > > > > +
> > > > > +       return pool_size;
> > > > > +}
> > > > > +
> > > > >  static void zswap_update_total_size(void)
> > > > >  {
> > > > >         struct zswap_pool *pool;
> > > > >         u64 total = 0;
> > > > > -       int i;
> > > > >
> > > > >         rcu_read_lock();
> > > > >
> > > > >         list_for_each_entry_rcu(pool, &zswap_pools, list)
> > > > > -               for (i = 0; i < ZSWAP_NR_ZPOOLS; i++)
> > > > > -                       total += zpool_get_total_size(pool->zpools[i]);
> > > > > +               total += get_zswap_pool_size(pool);
> > > > >
> > > > >         rcu_read_unlock();
> > > > >
> > > > > @@ -318,8 +349,23 @@ static bool zswap_lru_add(struct list_lru *list_lru, struct zswap_entry *entry)
> > > > >  {
> > > > >         struct mem_cgroup *memcg = entry->objcg ?
> > > > >                 get_mem_cgroup_from_objcg(entry->objcg) : NULL;
> > > > > +       struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(entry->nid));
> > > > >         bool added = __list_lru_add(list_lru, &entry->lru, entry->nid, memcg);
> > > > > +       unsigned long flags, lru_size;
> > > > > +
> > > > > +       if (added) {
> > > > > +               lru_size = list_lru_count_one(list_lru, entry->nid, memcg);
> > > > > +               spin_lock_irqsave(&lruvec->lru_lock, flags);
> > > > > +               lruvec->nr_zswap_protected++;
> > > > >
> > > > > +               /*
> > > > > +                * Decay to avoid overflow and adapt to changing workloads.
> > > > > +                * This is based on LRU reclaim cost decaying heuristics.
> > > > > +                */
> > > > > +               if (lruvec->nr_zswap_protected > lru_size / 4)
> > > > > +                       lruvec->nr_zswap_protected /= 2;
> > > > > +               spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > > > > +       }
> > > > >         mem_cgroup_put(memcg);
> > > > >         return added;
> > > > >  }
> > > > > @@ -420,6 +466,7 @@ static void zswap_free_entry(struct zswap_entry *entry)
> > > > >         else {
> > > > >                 zswap_lru_del(&entry->pool->list_lru, entry);
> > > > >                 zpool_free(zswap_find_zpool(entry), entry->handle);
> > > > > +               atomic_dec(&entry->pool->nr_stored);
> > > > >                 zswap_pool_put(entry->pool);
> > > > >         }
> > > > >         zswap_entry_cache_free(entry);
> > > > > @@ -461,6 +508,98 @@ static struct zswap_entry *zswap_entry_find_get(struct rb_root *root,
> > > > >         return entry;
> > > > >  }
> > > > >
> > > > > +/*********************************
> > > > > +* shrinker functions
> > > > > +**********************************/
> > > > > +static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_one *l,
> > > > > +                                      spinlock_t *lock, void *arg);
> > > > > +
> > > > > +static unsigned long zswap_shrinker_scan(struct shrinker *shrinker,
> > > > > +               struct shrink_control *sc)
> > > > > +{
> > > > > +       struct zswap_pool *pool = shrinker->private_data;
> > > > > +       unsigned long shrink_ret, nr_zswap_protected, flags,
> > > > > +               lru_size = list_lru_shrink_count(&pool->list_lru, sc);
> > > > > +       struct lruvec *lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid));
> > > > > +       bool encountered_page_in_swapcache = false;
> > > > > +
> > > > > +       spin_lock_irqsave(&lruvec->lru_lock, flags);
> > > > > +       nr_zswap_protected = lruvec->nr_zswap_protected;
> > > > > +       spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > > > > +
> > > > > +       /*
> > > > > +        * Abort if the shrinker is disabled or if we are shrinking into the
> > > > > +        * protected region.
> > > > > +        */
> > > > > +       if (!is_shrinker_enabled(sc->memcg) ||
> > > > > +                       nr_zswap_protected >= lru_size - sc->nr_to_scan) {
> > > > > +               sc->nr_scanned = 0;
> > > > > +               return SHRINK_STOP;
> > > > > +       }
> > > > > +
> > > > > +       shrink_ret = list_lru_shrink_walk(&pool->list_lru, sc, &shrink_memcg_cb,
> > > > > +               &encountered_page_in_swapcache);
> > > > > +
> > > > > +       if (encountered_page_in_swapcache)
> > > > > +               return SHRINK_STOP;
> > > > > +
> > > > > +       return shrink_ret ? shrink_ret : SHRINK_STOP;
> > > > > +}
> > > > > +
> > > > > +static unsigned long zswap_shrinker_count(struct shrinker *shrinker,
> > > > > +               struct shrink_control *sc)
> > > > > +{
> > > > > +       struct zswap_pool *pool = shrinker->private_data;
> > > > > +       struct mem_cgroup *memcg = sc->memcg;
> > > > > +       struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(sc->nid));
> > > > > +       unsigned long nr_backing, nr_stored, nr_freeable, flags;
> > > > > +
> > > > > +#ifdef CONFIG_MEMCG_KMEM
> > > > > +       cgroup_rstat_flush(memcg->css.cgroup);
> > > > > +       nr_backing = memcg_page_state(memcg, MEMCG_ZSWAP_B) >> PAGE_SHIFT;
> > > > > +       nr_stored = memcg_page_state(memcg, MEMCG_ZSWAPPED);
> > > > > +#else
> > > > > +       /* use pool stats instead of memcg stats */
> > > > > +       nr_backing = get_zswap_pool_size(pool) >> PAGE_SHIFT;
> > > > > +       nr_stored = atomic_read(&pool->nr_stored);
> > > > > +#endif
> > > > > +
> > > > > +       if (!is_shrinker_enabled(memcg) || !nr_stored)
> > > > > +               return 0;
> > > > > +
> > > > > +       nr_freeable = list_lru_shrink_count(&pool->list_lru, sc);
> > > > > +       /*
> > > > > +        * Subtract the lru size by an estimate of the number of pages
> > > > > +        * that should be protected.
> > > > > +        */
> > > > > +       spin_lock_irqsave(&lruvec->lru_lock, flags);
> > > > > +       nr_freeable = nr_freeable > lruvec->nr_zswap_protected ?
> > > > > +               nr_freeable - lruvec->nr_zswap_protected : 0;
> > > > > +       spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > > > > +
> > > > > +       /*
> > > > > +        * Scale the number of freeable pages by the memory saving factor.
> > > > > +        * This ensures that the better zswap compresses memory, the fewer
> > > > > +        * pages we will evict to swap (as it will otherwise incur IO for
> > > > > +        * relatively small memory saving).
> > > > > +        */
> > > > > +       return mult_frac(nr_freeable, nr_backing, nr_stored);
> > > > > +}
> > > > > +
> > > > > +static void zswap_alloc_shrinker(struct zswap_pool *pool)
> > > > > +{
> > > > > +       pool->shrinker =
> > > > > +               shrinker_alloc(SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE, "mm-zswap");
> > > > > +       if (!pool->shrinker)
> > > > > +               return;
> > > > > +
> > > > > +       pool->shrinker->private_data = pool;
> > > > > +       pool->shrinker->scan_objects = zswap_shrinker_scan;
> > > > > +       pool->shrinker->count_objects = zswap_shrinker_count;
> > > > > +       pool->shrinker->batch = 0;
> > > > > +       pool->shrinker->seeks = DEFAULT_SEEKS;
> > > > > +}
> > > > > +
> > > > >  /*********************************
> > > > >  * per-cpu code
> > > > >  **********************************/
> > > > > @@ -656,11 +795,14 @@ static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_o
> > > > >                                        spinlock_t *lock, void *arg)
> > > > >  {
> > > > >         struct zswap_entry *entry = container_of(item, struct zswap_entry, lru);
> > > > > +       bool *encountered_page_in_swapcache = (bool *)arg;
> > > > >         struct mem_cgroup *memcg;
> > > > >         struct zswap_tree *tree;
> > > > > +       struct lruvec *lruvec;
> > > > >         pgoff_t swpoffset;
> > > > >         enum lru_status ret = LRU_REMOVED_RETRY;
> > > > >         int writeback_result;
> > > > > +       unsigned long flags;
> > > > >
> > > > >         /*
> > > > >          * Once the lru lock is dropped, the entry might get freed. The
> > > > > @@ -696,8 +838,24 @@ static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_o
> > > > >                 /* we cannot use zswap_lru_add here, because it increments node's lru count */
> > > > >                 list_lru_putback(&entry->pool->list_lru, item, entry->nid, memcg);
> > > > >                 spin_unlock(lock);
> > > > > -               mem_cgroup_put(memcg);
> > > > >                 ret = LRU_RETRY;
> > > > > +
> > > > > +               /*
> > > > > +                * Encountering a page already in swap cache is a sign that we are shrinking
> > > > > +                * into the warmer region. We should terminate shrinking (if we're in the dynamic
> > > > > +                * shrinker context).
> > > > > +                */
> > > > > +               if (writeback_result == -EEXIST && encountered_page_in_swapcache) {
> > > > > +                       ret = LRU_SKIP;
> > > > > +                       *encountered_page_in_swapcache = true;
> > > > > +               }
> > > > > +               lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(entry->nid));
> > > > > +               spin_lock_irqsave(&lruvec->lru_lock, flags);
> > > > > +               /* Increment the protection area to account for the LRU rotation. */
> > > > > +               lruvec->nr_zswap_protected++;
> > > > > +               spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > > > > +
> > > > > +               mem_cgroup_put(memcg);
> > > > >                 goto put_unlock;
> > > > >         }
> > > > >
> > > > > @@ -828,6 +986,11 @@ static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
> > > > >                                        &pool->node);
> > > > >         if (ret)
> > > > >                 goto error;
> > > > > +
> > > > > +       zswap_alloc_shrinker(pool);
> > > > > +       if (!pool->shrinker)
> > > > > +               goto error;
> > > > > +
> > > > >         pr_debug("using %s compressor\n", pool->tfm_name);
> > > > >
> > > > >         /* being the current pool takes 1 ref; this func expects the
> > > > > @@ -836,12 +999,17 @@ static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
> > > > >         kref_init(&pool->kref);
> > > > >         INIT_LIST_HEAD(&pool->list);
> > > > >         INIT_WORK(&pool->shrink_work, shrink_worker);
> > > > > -       list_lru_init_memcg(&pool->list_lru, NULL);
> > > > > +       if (list_lru_init_memcg(&pool->list_lru, pool->shrinker))
> > > > > +               goto lru_fail;
> > > > > +       shrinker_register(pool->shrinker);
> > > > >
> > > > >         zswap_pool_debug("created", pool);
> > > > >
> > > > >         return pool;
> > > > >
> > > > > +lru_fail:
> > > > > +       list_lru_destroy(&pool->list_lru);
> > > > > +       shrinker_free(pool->shrinker);
> > > > >  error:
> > > > >         if (pool->acomp_ctx)
> > > > >                 free_percpu(pool->acomp_ctx);
> > > > > @@ -899,6 +1067,7 @@ static void zswap_pool_destroy(struct zswap_pool *pool)
> > > > >
> > > > >         zswap_pool_debug("destroying", pool);
> > > > >
> > > > > +       shrinker_free(pool->shrinker);
> > > > >         cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
> > > > >         free_percpu(pool->acomp_ctx);
> > > > >         list_lru_destroy(&pool->list_lru);
> > > > > @@ -1431,6 +1600,7 @@ bool zswap_store(struct folio *folio)
> > > > >         if (entry->length) {
> > > > >                 INIT_LIST_HEAD(&entry->lru);
> > > > >                 zswap_lru_add(&pool->list_lru, entry);
> > > > > +               atomic_inc(&pool->nr_stored);
> > > > >         }
> > > > >         spin_unlock(&tree->lock);
> > > > >
> > > > > --
> > > > > 2.34.1
> > > Thanks for the comments/suggestion, Yosry!
  

On Mon, Sep 25, 2023 at 6:12 PM Yosry Ahmed <yosryahmed@google.com> wrote:
>
> On Mon, Sep 25, 2023 at 5:43 PM Nhat Pham <nphamcs@gmail.com> wrote:
> >
> > On Mon, Sep 25, 2023 at 5:00 PM Yosry Ahmed <yosryahmed@google.com> wrote:
> > >
> > > On Mon, Sep 25, 2023 at 4:29 PM Nhat Pham <nphamcs@gmail.com> wrote:
> > > >
> > > > On Mon, Sep 25, 2023 at 1:38 PM Yosry Ahmed <yosryahmed@google.com> wrote:
> > > > >
> > > > > On Tue, Sep 19, 2023 at 10:14 AM Nhat Pham <nphamcs@gmail.com> wrote:
> > > > > >
> > > > > > Currently, we only shrink the zswap pool when the user-defined limit is
> > > > > > hit. This means that if we set the limit too high, cold data that are
> > > > > > unlikely to be used again will reside in the pool, wasting precious
> > > > > > memory. It is hard to predict how much zswap space will be needed ahead
> > > > > > of time, as this depends on the workload (specifically, on factors such
> > > > > > as memory access patterns and compressibility of the memory pages).
> > > > > >
> > > > > > This patch implements a memcg- and NUMA-aware shrinker for zswap, that
> > > > > > is initiated when there is memory pressure. The shrinker does not
> > > > > > have any parameter that must be tuned by the user, and can be opted in
> > > > > > or out on a per-memcg basis.
> > > > >
> > > > > What's the use case for having per-memcg opt-in/out?
> > > > >
> > > > > If there is memory pressure, reclaiming swap-backed pages will push
> > > > > pages out of zswap anyway, regardless of this patch. With this patch,
> > > > > any sort of reclaim can push pages out of zswap. Wouldn't that be
> > > > > preferable to reclaiming memory that is currently resident in memory
> > > > > (so arguably hotter than the pages in zswap)? Why would this decision
> > > > > be different per-memcg?
> > > > I'm not quite following your argument here. The point of having this
> > > > be done on a per-memcg basis is that we have different workloads
> > > > with different memory access pattern (and as a result, different memory
> > > > coldness distribution).
> > > >
> > > > In a workload where there is a lot of cold data, we can really benefit
> > > > from reclaiming all of those pages and repurpose the memory reclaimed
> > > > (for e.g for filecache).
> > > >
> > > > On the other hand, in a workload where there aren't a lot of cold data,
> > > > reclaiming its zswapped pages will at best do nothing (wasting CPU
> > > > cycles on compression/decompression), and at worst hurt performance
> > > > (due to increased IO when we need those writtenback pages again).
> > > >
> > > > Such different workloads could co-exist in the same system, and having
> > > > a per-memcg knob allows us to crank on the shrinker only on workloads
> > > > where it makes sense.
> > >
> > > I am not sure we are on the same page here.
> > >
> > > What you're describing sounds more like proactive reclaim, which we
> > > wouldn't invoke unless the workload has cold data anyway.
> > >
> > > IIUC, outside of that, this shrinker will run when there is memory
> > > pressure. This means that we need to free memory anyway, regardless of
> > > its absolute coldness. We want to evict the colder pages in the memcg.
> > > It seems to be that in ~all cases, evicting pages in zswap will be
> > > better than evicting pages in memory, as the pages in memory are
> > > arguably hotter (since they weren't reclaimed first). This seems to be
> > > something that would be true for all workloads.
> > >
> > > What am I missing?
> >
> > Yup, the shrinker is initiated under memory pressure.
> > And with it, we can reclaim memory from zswap when
> > it's (often) not at max capacity.
> >
> > The kernel has no knowledge of absolute coldness, only relative
> > coldness thanks to LRU. We don't have a global LRU of all possible
> > memory pages/objects for a particular memcg either, so we cannot
> > compare the coldness of objects from different sources.
> >
> > The "coldest" pages in zswap LRU could very well be warm enough
> > that swapping them out degrades performance, while there are even
> > colder memory from other sources (other shrinkers registered for this
> > memcg). Alternatively, we can also "evict" uncompressed anonymous
> > memory, which will go to the zswap pool. This also saves memory,
> > and could potentially be better than zswap reclaim (2 compressed
> > pages might be better performance-wise than 1 uncompressed,
> > 1 swapped out)
> >
> > All of this depends on the memory access pattern of the workloads,
> > which could differ cgroup-by-cgroup within the same system.
> > Having a per-memcg knob is a way for admins to influence this
> > decision from userspace, if the admins have knowledge about
> > workload memory access patterns.
> >
> > For e.g, if we know that there is one particular cgroup that populates
> > a bunch of single-use tmpfs pages, then we can target that cgroup
> > specifically, while leaving the other cgroups in the system alone.
>
> I think it's useful to break down the discussion here for cgroup
> reclaim and global reclaim.
>
> For cgroup reclaim, the kernel knows that the pages in the LRUs are
> relatively hotter than the pages in zswap. So I don't see why
> userspace would opt out specific cgroups from zswap shrinking. In my
> experience, most memory usage comes from LRU pages, so let's ignore
> other shrinkers for a second. Yes, in some cases compressing another
> page might be better than moving a compressed page to swap, but how
> would userspace have the intuition to decide this? It varies not only
> based on workload, but also the point in time, the compressibility of
> pages, etc.
>
> In other words, how would a system admin choose to opt a cgroup in or out?
>
> For global reclaim, IIUC you are saying that we want to protect some
> cgroups under global memory pressure because we know that their "cold"
> memory in zswap is hotter than memory elsewhere in the hierarchy,
> right?
>
> Isn't this the case for LRU reclaim as well? I would assume memory
> protections would be used to tune this, not opting a cgroup completely
> from zswap shrinking. Global reclaim can end up reclaiming LRU pages
> from that cgroup if protection is not set up correctly anyway. What do
> we gain by protecting pages in zswap if hotter pages in the LRUs are
> not protected?

Hmm you got a point. I guess our main motivation is just being
extra safe. It's a new feature, so we want to make sure that
we limit unintentional performance regression for everyone
(not just Meta) as much as possible.

However, as you have pointed out, per-cgroup knob might not
help any more than a simple, global knob. I'll remove it in v3
(and we can revisit this decision later on if it turns out to be
necessary after all).

>
> >
> > >
> > > > >
> > > > > >
> > > > > > Furthermore, to make it more robust for many workloads and prevent
> > > > > > overshrinking (i.e evicting warm pages that might be refaulted into
> > > > > > memory), we build in the following heuristics:
> > > > > >
> > > > > > * Estimate the number of warm pages residing in zswap, and attempt to
> > > > > >   protect this region of the zswap LRU.
> > > > > > * Scale the number of freeable objects by an estimate of the memory
> > > > > >   saving factor. The better zswap compresses the data, the fewer pages
> > > > > >   we will evict to swap (as we will otherwise incur IO for relatively
> > > > > >   small memory saving).
> > > > > > * During reclaim, if the shrinker encounters a page that is also being
> > > > > >   brought into memory, the shrinker will cautiously terminate its
> > > > > >   shrinking action, as this is a sign that it is touching the warmer
> > > > > >   region of the zswap LRU.
> > > > >
> > > > > I don't have an opinion about the reclaim heuristics here, I will let
> > > > > reclaim experts chip in.
> > > > >
> > > > > >
> > > > > > On a benchmark that we have run:
> > > > >
> > > > > Please add more details (as much as possible) about the benchmarks used here.
> > > > Sure! I built the kernel in a memory-limited cgroup a couple times,
> > > > then measured the build time.
> > > >
> > > > To simulate conditions where there are cold, unused data, I
> > > > also generated a bunch of data in tmpfs (and never touch them
> > > > again).
> > >
> > > Please include such details in the commit message, there is also
> > > another reference below to "another" benchmark.
> >
> > Will do if/when I send v3.
> > The "another" benchmark is just generating even more tmpfs cold data :)
>
> Those benchmarks are heavily synthetic, which is not a showstopper,
> but describing them in the commit message helps people reason about
> the change.
>
> >
> > >
> > >
> > > > >
> > > > > >
> > > > > > (without the shrinker)
> > > > > > real -- mean: 153.27s, median: 153.199s
> > > > > > sys -- mean: 541.652s, median: 541.903s
> > > > > > user -- mean: 4384.9673999999995s, median: 4385.471s
> > > > > >
> > > > > > (with the shrinker)
> > > > > > real -- mean: 151.4956s, median: 151.456s
> > > > > > sys -- mean: 461.14639999999997s, median: 465.656s
> > > > > > user -- mean: 4384.7118s, median: 4384.675s
> > > > > >
> > > > > > We observed a 14-15% reduction in kernel CPU time, which translated to
> > > > > > over 1% reduction in real time.
> > > > > >
> > > > > > On another benchmark, where there was a lot more cold memory residing in
> > > > > > zswap, we observed even more pronounced gains:
> > > > > >
> > > > > > (without the shrinker)
> > > > > > real -- mean: 157.52519999999998s, median: 157.281s
> > > > > > sys -- mean: 769.3082s, median: 780.545s
> > > > > > user -- mean: 4378.1622s, median: 4378.286s
> > > > > >
> > > > > > (with the shrinker)
> > > > > > real -- mean: 152.9608s, median: 152.845s
> > > > > > sys -- mean: 517.4446s, median: 506.749s
> > > > > > user -- mean: 4387.694s, median: 4387.935s
> > > > > >
> > > > > > Here, we saw around 32-35% reduction in kernel CPU time, which
> > > > > > translated to 2.8% reduction in real time. These results confirm our
> > > > > > hypothesis that the shrinker is more helpful the more cold memory we
> > > > > > have.
> > > > > >
> > > > > > Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
> > > > > > Signed-off-by: Nhat Pham <nphamcs@gmail.com>
> > > > > > ---
> > > > > >  Documentation/admin-guide/mm/zswap.rst |  12 ++
> > > > > >  include/linux/memcontrol.h             |   1 +
> > > > > >  include/linux/mmzone.h                 |  14 ++
> > > > > >  mm/memcontrol.c                        |  33 +++++
> > > > > >  mm/swap_state.c                        |  31 ++++-
> > > > > >  mm/zswap.c                             | 180 ++++++++++++++++++++++++-
> > > > > >  6 files changed, 263 insertions(+), 8 deletions(-)
> > > > > >
> > > > > > diff --git a/Documentation/admin-guide/mm/zswap.rst b/Documentation/admin-guide/mm/zswap.rst
> > > > > > index 45b98390e938..ae8597a67804 100644
> > > > > > --- a/Documentation/admin-guide/mm/zswap.rst
> > > > > > +++ b/Documentation/admin-guide/mm/zswap.rst
> > > > > > @@ -153,6 +153,18 @@ attribute, e. g.::
> > > > > >
> > > > > >  Setting this parameter to 100 will disable the hysteresis.
> > > > > >
> > > > > > +When there is a sizable amount of cold memory residing in the zswap pool, it
> > > > > > +can be advantageous to proactively write these cold pages to swap and reclaim
> > > > > > +the memory for other use cases. By default, the zswap shrinker is disabled.
> > > > > > +User can enable it by first switching on the global knob:
> > > > > > +
> > > > > > +  echo Y > /sys/module/zswap/par meters/shrinker_enabled
> > > > > > +
> > > > > > +When the kernel is compiled with CONFIG_MEMCG_KMEM, user needs to further turn
> > > > > > +it on for each cgroup that the shrinker should target:
> > > > > > +
> > > > > > +  echo 1 > /sys/fs/cgroup/<cgroup-name>/memory.zswap.shrinker.enabled
> > > > > > +
> > > > > >  A debugfs interface is provided for various statistic about pool size, number
> > > > > >  of pages stored, same-value filled pages and various counters for the reasons
> > > > > >  pages are rejected.
> > > > > > diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
> > > > > > index 05d34b328d9d..f005ea667863 100644
> > > > > > --- a/include/linux/memcontrol.h
> > > > > > +++ b/include/linux/memcontrol.h
> > > > > > @@ -219,6 +219,7 @@ struct mem_cgroup {
> > > > > >
> > > > > >  #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
> > > > > >         unsigned long zswap_max;
> > > > > > +       atomic_t zswap_shrinker_enabled;
> > > > > >  #endif
> > > > > >
> > > > > >         unsigned long soft_limit;
> > > > > > diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
> > > > > > index 4106fbc5b4b3..81f4c5ea3e16 100644
> > > > > > --- a/include/linux/mmzone.h
> > > > > > +++ b/include/linux/mmzone.h
> > > > > > @@ -637,6 +637,20 @@ struct lruvec {
> > > > > >  #ifdef CONFIG_MEMCG
> > > > > >         struct pglist_data *pgdat;
> > > > > >  #endif
> > > > > > +#ifdef CONFIG_ZSWAP
> > > > > > +       /*
> > > > > > +        * Number of pages in zswap that should be protected from the shrinker.
> > > > > > +        * This number is an estimate of the following counts:
> > > > > > +        *
> > > > > > +        * a) Recent page faults.
> > > > > > +        * b) Recent insertion to the zswap LRU. This includes new zswap stores,
> > > > > > +        *    as well as recent zswap LRU rotations.
> > > > > > +        *
> > > > > > +        * These pages are likely to be warm, and might incur IO if the are written
> > > > > > +        * to swap.
> > > > > > +        */
> > > > > > +       unsigned long nr_zswap_protected;
> > > > > > +#endif
> > > > >
> > > > > Would this be better abstracted in a zswap lruvec struct?
> > > > There is just one field, so that sounds like overkill to me.
> > > > But if we need to store more data (for smarter heuristics),
> > > > that'll be a good idea. I'll keep this in mind. Thanks for the
> > > > suggestion, Yosry!
> > >
> > > (A space between the quoted text and the reply usually helps visually :)
> > >
> > > It wasn't really about the number of fields, but rather place this
> > > struct in zswap.h (with the long comment explaining what it's doing),
> > > and adding an abstracted struct member here. The comment will live in
> > > an appropriate file, further modifications don't need to touch
> > > mmzone.h, and struct lruvec is less cluttered for readers that don't
> > > care about zswap (and we can avoid the ifdef).
> > >
> > > Anyway, this is all mostly aesthetic so I don't feel strongly.
> > >
> > > > >
> > > > > >  };
> > > > > >
> > > > > >  /* Isolate unmapped pages */
> > > > > > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > > > > > index 9f84b3f7b469..1a2c97cf396f 100644
> > > > > > --- a/mm/memcontrol.c
> > > > > > +++ b/mm/memcontrol.c
> > > > > > @@ -5352,6 +5352,8 @@ mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
> > > > > >         WRITE_ONCE(memcg->soft_limit, PAGE_COUNTER_MAX);
> > > > > >  #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_ZSWAP)
> > > > > >         memcg->zswap_max = PAGE_COUNTER_MAX;
> > > > > > +       /* Disable the shrinker by default */
> > > > > > +       atomic_set(&memcg->zswap_shrinker_enabled, 0);
> > > > > >  #endif
> > > > > >         page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
> > > > > >         if (parent) {
> > > > > > @@ -7877,6 +7879,31 @@ static ssize_t zswap_max_write(struct kernfs_open_file *of,
> > > > > >         return nbytes;
> > > > > >  }
> > > > > >
> > > > > > +static int zswap_shrinker_enabled_show(struct seq_file *m, void *v)
> > > > > > +{
> > > > > > +       struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
> > > > > > +
> > > > > > +       seq_printf(m, "%d\n", atomic_read(&memcg->zswap_shrinker_enabled));
> > > > > > +       return 0;
> > > > > > +}
> > > > > > +
> > > > > > +static ssize_t zswap_shrinker_enabled_write(struct kernfs_open_file *of,
> > > > > > +                              char *buf, size_t nbytes, loff_t off)
> > > > > > +{
> > > > > > +       struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
> > > > > > +       int zswap_shrinker_enabled;
> > > > > > +       ssize_t parse_ret = kstrtoint(strstrip(buf), 0, &zswap_shrinker_enabled);
> > > > > > +
> > > > > > +       if (parse_ret)
> > > > > > +               return parse_ret;
> > > > > > +
> > > > > > +       if (zswap_shrinker_enabled < 0 || zswap_shrinker_enabled > 1)
> > > > > > +               return -ERANGE;
> > > > > > +
> > > > > > +       atomic_set(&memcg->zswap_shrinker_enabled, zswap_shrinker_enabled);
> > > > > > +       return nbytes;
> > > > > > +}
> > > > > > +
> > > > > >  static struct cftype zswap_files[] = {
> > > > > >         {
> > > > > >                 .name = "zswap.current",
> > > > > > @@ -7889,6 +7916,12 @@ static struct cftype zswap_files[] = {
> > > > > >                 .seq_show = zswap_max_show,
> > > > > >                 .write = zswap_max_write,
> > > > > >         },
> > > > > > +       {
> > > > > > +               .name = "zswap.shrinker.enabled",
> > > > > > +               .flags = CFTYPE_NOT_ON_ROOT,
> > > > > > +               .seq_show = zswap_shrinker_enabled_show,
> > > > > > +               .write = zswap_shrinker_enabled_write,
> > > > > > +       },
> > > > > >         { }     /* terminate */
> > > > > >  };
> > > > > >  #endif /* CONFIG_MEMCG_KMEM && CONFIG_ZSWAP */
> > > > > > diff --git a/mm/swap_state.c b/mm/swap_state.c
> > > > > > index 1c826737aacb..788e36a06c34 100644
> > > > > > --- a/mm/swap_state.c
> > > > > > +++ b/mm/swap_state.c
> > > > > > @@ -618,6 +618,22 @@ static unsigned long swapin_nr_pages(unsigned long offset)
> > > > > >         return pages;
> > > > > >  }
> > > > > >
> > > > > > +#ifdef CONFIG_ZSWAP
> > > > > > +/*
> > > > > > + * Refault is an indication that warmer pages are not resident in memory.
> > > > > > + * Increase the size of zswap's protected area.
> > > > > > + */
> > > > > > +static void inc_nr_protected(struct page *page)
> > > > > > +{
> > > > > > +       struct lruvec *lruvec = folio_lruvec(page_folio(page));
> > > > > > +       unsigned long flags;
> > > > > > +
> > > > > > +       spin_lock_irqsave(&lruvec->lru_lock, flags);
> > > > > > +       lruvec->nr_zswap_protected++;
> > > > > > +       spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > > > > > +}
> > > > > > +#endif
> > > > > > +
> > > > >
> > > > > A few questions:
> > > > > - Why is this function named in such a generic way?
> > > > Perhaps inc_nr_zswap_protected would be better? :)
> > >
> > > If we use an atomic, the function can go away anyway. See below.
> > >
> > > > > - Why is this function here instead of in mm/zswap.c?
> > > > No particular reason :) It's not being used anywhere else,
> > > > so I just put it as a static function here.
> > >
> > > It is inline in mm/zswap.c in one place. I personally would have
> > > preferred nr_zswap_protected and the helper to be defined in
> > > zswap.h/zswap.c as I mentioned below. Anyway, this function can go
> > > away.
> > >
> > > > > - Why is this protected by the heavily contested lruvec lock instead
> > > > > of being an atomic?
> > > > nr_zswap_protected can be decayed (see zswap_lru_add), which
> > > > I don't think it can be implemented with atomics :( It'd be much
> > > > cleaner indeed.
> > >
> > > I think a cmpxchg (or a try_cmpxchg) loop can be used in this case to
> > > implement it using an atomic?
> > >
> > > See https://docs.kernel.org/core-api/wrappers/atomic_t.html.
> >
> > Ah I did think about this, but that seems overkill at the time.
> > But if lruvec lock is indeed hotly contested, this should help.
>
> I wouldn't say so, we can drop numerous calls to grab/drop the lock,
> and drop the helper. A try_cmpxchg loop here would only be a couple of
> lines, I suspect it would be more concise than the code now:
>
> old = atomic_inc_return(&lruvec->nr_zswap_protected);
> do {
>        if (old > lru_size / 4)
>                new = old / 2;
> } while (atomic_try_cmpxchg(&lruvec->nr_zswap_protected, &old, new));
>

Yeah this definitely seems quite clean. Lemme give this a try.

> >
> > >
> > > > > > +               lruvec->nr_zswap_protected++;
> > > > > >
> > > > > > +               /*
> > > > > > +                * Decay to avoid overflow and adapt to changing workloads.
> > > > > > +                * This is based on LRU reclaim cost decaying heuristics.
> > > > > > +                */
> > > > > > +               if (lruvec->nr_zswap_protected > lru_size / 4)
> > > > > > +                       lruvec->nr_zswap_protected /= 2;
> > >
> > > >
> > > > I'm wary of adding new locks, so I just re-use this existing lock.
> > > > But if lruvec lock is heavily congested (I'm not aware/familar with
> > > > this issue), then perhaps a new, dedicated lock would help?
> > > > >
> > > > > >  /**
> > > > > >   * swap_cluster_readahead - swap in pages in hope we need them soon
> > > > > >   * @entry: swap entry of this memory
> > > > > > @@ -686,7 +702,12 @@ struct page *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask,
> > > > > >         lru_add_drain();        /* Push any new pages onto the LRU now */
> > > > > >  skip:
> > > > > >         /* The page was likely read above, so no need for plugging here */
> > > > > > -       return read_swap_cache_async(entry, gfp_mask, vma, addr, NULL);
> > > > > > +       page = read_swap_cache_async(entry, gfp_mask, vma, addr, NULL);
> > > > > > +#ifdef CONFIG_ZSWAP
> > > > > > +       if (page)
> > > > > > +               inc_nr_protected(page);
> > > > > > +#endif
> > > > > > +       return page;
> > > > > >  }
> > > > > >
> > > > > >  int init_swap_address_space(unsigned int type, unsigned long nr_pages)
> > > > > > @@ -853,8 +874,12 @@ static struct page *swap_vma_readahead(swp_entry_t fentry, gfp_t gfp_mask,
> > > > > >         lru_add_drain();
> > > > > >  skip:
> > > > > >         /* The page was likely read above, so no need for plugging here */
> > > > > > -       return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address,
> > > > > > -                                    NULL);
> > > > > > +       page = read_swap_cache_async(fentry, gfp_mask, vma, vmf->address, NULL);
> > > > > > +#ifdef CONFIG_ZSWAP
> > > > > > +       if (page)
> > > > > > +               inc_nr_protected(page);
> > > > > > +#endif
> > > > > > +       return page;
> > > > > >  }
> > > > > >
> > > > > >  /**
> > > > > > diff --git a/mm/zswap.c b/mm/zswap.c
> > > > > > index 1a469e5d5197..79cb18eeb8bf 100644
> > > > > > --- a/mm/zswap.c
> > > > > > +++ b/mm/zswap.c
> > > > > > @@ -145,6 +145,26 @@ module_param_named(exclusive_loads, zswap_exclusive_loads_enabled, bool, 0644);
> > > > > >  /* Number of zpools in zswap_pool (empirically determined for scalability) */
> > > > > >  #define ZSWAP_NR_ZPOOLS 32
> > > > > >
> > > > > > +/*
> > > > > > + * Global flag to enable/disable memory pressure-based shrinker for all memcgs.
> > > > > > + * If CONFIG_MEMCG_KMEM is on, we can further selectively disable
> > > > > > + * the shrinker for each memcg.
> > > > > > + */
> > > > > > +static bool zswap_shrinker_enabled;
> > > > > > +module_param_named(shrinker_enabled, zswap_shrinker_enabled, bool, 0644);
> > > > > > +#ifdef CONFIG_MEMCG_KMEM
> > > > > > +static bool is_shrinker_enabled(struct mem_cgroup *memcg)
> > > > > > +{
> > > > > > +       return zswap_shrinker_enabled &&
> > > > > > +               atomic_read(&memcg->zswap_shrinker_enabled);
> > > > > > +}
> > > > > > +#else
> > > > > > +static bool is_shrinker_enabled(struct mem_cgroup *memcg)
> > > > > > +{
> > > > > > +       return zswap_shrinker_enabled;
> > > > > > +}
> > > > > > +#endif
> > > > > > +
> > > > > >  /*********************************
> > > > > >  * data structures
> > > > > >  **********************************/
> > > > > > @@ -174,6 +194,8 @@ struct zswap_pool {
> > > > > >         char tfm_name[CRYPTO_MAX_ALG_NAME];
> > > > > >         struct list_lru list_lru;
> > > > > >         struct mem_cgroup *next_shrink;
> > > > > > +       struct shrinker *shrinker;
> > > > > > +       atomic_t nr_stored;
> > > > > >  };
> > > > > >
> > > > > >  /*
> > > > > > @@ -273,17 +295,26 @@ static bool zswap_can_accept(void)
> > > > > >                         DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
> > > > > >  }
> > > > > >
> > > > > > +static u64 get_zswap_pool_size(struct zswap_pool *pool)
> > > > > > +{
> > > > > > +       u64 pool_size = 0;
> > > > > > +       int i;
> > > > > > +
> > > > > > +       for (i = 0; i < ZSWAP_NR_ZPOOLS; i++)
> > > > > > +               pool_size += zpool_get_total_size(pool->zpools[i]);
> > > > > > +
> > > > > > +       return pool_size;
> > > > > > +}
> > > > > > +
> > > > > >  static void zswap_update_total_size(void)
> > > > > >  {
> > > > > >         struct zswap_pool *pool;
> > > > > >         u64 total = 0;
> > > > > > -       int i;
> > > > > >
> > > > > >         rcu_read_lock();
> > > > > >
> > > > > >         list_for_each_entry_rcu(pool, &zswap_pools, list)
> > > > > > -               for (i = 0; i < ZSWAP_NR_ZPOOLS; i++)
> > > > > > -                       total += zpool_get_total_size(pool->zpools[i]);
> > > > > > +               total += get_zswap_pool_size(pool);
> > > > > >
> > > > > >         rcu_read_unlock();
> > > > > >
> > > > > > @@ -318,8 +349,23 @@ static bool zswap_lru_add(struct list_lru *list_lru, struct zswap_entry *entry)
> > > > > >  {
> > > > > >         struct mem_cgroup *memcg = entry->objcg ?
> > > > > >                 get_mem_cgroup_from_objcg(entry->objcg) : NULL;
> > > > > > +       struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(entry->nid));
> > > > > >         bool added = __list_lru_add(list_lru, &entry->lru, entry->nid, memcg);
> > > > > > +       unsigned long flags, lru_size;
> > > > > > +
> > > > > > +       if (added) {
> > > > > > +               lru_size = list_lru_count_one(list_lru, entry->nid, memcg);
> > > > > > +               spin_lock_irqsave(&lruvec->lru_lock, flags);
> > > > > > +               lruvec->nr_zswap_protected++;
> > > > > >
> > > > > > +               /*
> > > > > > +                * Decay to avoid overflow and adapt to changing workloads.
> > > > > > +                * This is based on LRU reclaim cost decaying heuristics.
> > > > > > +                */
> > > > > > +               if (lruvec->nr_zswap_protected > lru_size / 4)
> > > > > > +                       lruvec->nr_zswap_protected /= 2;
> > > > > > +               spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > > > > > +       }
> > > > > >         mem_cgroup_put(memcg);
> > > > > >         return added;
> > > > > >  }
> > > > > > @@ -420,6 +466,7 @@ static void zswap_free_entry(struct zswap_entry *entry)
> > > > > >         else {
> > > > > >                 zswap_lru_del(&entry->pool->list_lru, entry);
> > > > > >                 zpool_free(zswap_find_zpool(entry), entry->handle);
> > > > > > +               atomic_dec(&entry->pool->nr_stored);
> > > > > >                 zswap_pool_put(entry->pool);
> > > > > >         }
> > > > > >         zswap_entry_cache_free(entry);
> > > > > > @@ -461,6 +508,98 @@ static struct zswap_entry *zswap_entry_find_get(struct rb_root *root,
> > > > > >         return entry;
> > > > > >  }
> > > > > >
> > > > > > +/*********************************
> > > > > > +* shrinker functions
> > > > > > +**********************************/
> > > > > > +static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_one *l,
> > > > > > +                                      spinlock_t *lock, void *arg);
> > > > > > +
> > > > > > +static unsigned long zswap_shrinker_scan(struct shrinker *shrinker,
> > > > > > +               struct shrink_control *sc)
> > > > > > +{
> > > > > > +       struct zswap_pool *pool = shrinker->private_data;
> > > > > > +       unsigned long shrink_ret, nr_zswap_protected, flags,
> > > > > > +               lru_size = list_lru_shrink_count(&pool->list_lru, sc);
> > > > > > +       struct lruvec *lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid));
> > > > > > +       bool encountered_page_in_swapcache = false;
> > > > > > +
> > > > > > +       spin_lock_irqsave(&lruvec->lru_lock, flags);
> > > > > > +       nr_zswap_protected = lruvec->nr_zswap_protected;
> > > > > > +       spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > > > > > +
> > > > > > +       /*
> > > > > > +        * Abort if the shrinker is disabled or if we are shrinking into the
> > > > > > +        * protected region.
> > > > > > +        */
> > > > > > +       if (!is_shrinker_enabled(sc->memcg) ||
> > > > > > +                       nr_zswap_protected >= lru_size - sc->nr_to_scan) {
> > > > > > +               sc->nr_scanned = 0;
> > > > > > +               return SHRINK_STOP;
> > > > > > +       }
> > > > > > +
> > > > > > +       shrink_ret = list_lru_shrink_walk(&pool->list_lru, sc, &shrink_memcg_cb,
> > > > > > +               &encountered_page_in_swapcache);
> > > > > > +
> > > > > > +       if (encountered_page_in_swapcache)
> > > > > > +               return SHRINK_STOP;
> > > > > > +
> > > > > > +       return shrink_ret ? shrink_ret : SHRINK_STOP;
> > > > > > +}
> > > > > > +
> > > > > > +static unsigned long zswap_shrinker_count(struct shrinker *shrinker,
> > > > > > +               struct shrink_control *sc)
> > > > > > +{
> > > > > > +       struct zswap_pool *pool = shrinker->private_data;
> > > > > > +       struct mem_cgroup *memcg = sc->memcg;
> > > > > > +       struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(sc->nid));
> > > > > > +       unsigned long nr_backing, nr_stored, nr_freeable, flags;
> > > > > > +
> > > > > > +#ifdef CONFIG_MEMCG_KMEM
> > > > > > +       cgroup_rstat_flush(memcg->css.cgroup);
> > > > > > +       nr_backing = memcg_page_state(memcg, MEMCG_ZSWAP_B) >> PAGE_SHIFT;
> > > > > > +       nr_stored = memcg_page_state(memcg, MEMCG_ZSWAPPED);
> > > > > > +#else
> > > > > > +       /* use pool stats instead of memcg stats */
> > > > > > +       nr_backing = get_zswap_pool_size(pool) >> PAGE_SHIFT;
> > > > > > +       nr_stored = atomic_read(&pool->nr_stored);
> > > > > > +#endif
> > > > > > +
> > > > > > +       if (!is_shrinker_enabled(memcg) || !nr_stored)
> > > > > > +               return 0;
> > > > > > +
> > > > > > +       nr_freeable = list_lru_shrink_count(&pool->list_lru, sc);
> > > > > > +       /*
> > > > > > +        * Subtract the lru size by an estimate of the number of pages
> > > > > > +        * that should be protected.
> > > > > > +        */
> > > > > > +       spin_lock_irqsave(&lruvec->lru_lock, flags);
> > > > > > +       nr_freeable = nr_freeable > lruvec->nr_zswap_protected ?
> > > > > > +               nr_freeable - lruvec->nr_zswap_protected : 0;
> > > > > > +       spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > > > > > +
> > > > > > +       /*
> > > > > > +        * Scale the number of freeable pages by the memory saving factor.
> > > > > > +        * This ensures that the better zswap compresses memory, the fewer
> > > > > > +        * pages we will evict to swap (as it will otherwise incur IO for
> > > > > > +        * relatively small memory saving).
> > > > > > +        */
> > > > > > +       return mult_frac(nr_freeable, nr_backing, nr_stored);
> > > > > > +}
> > > > > > +
> > > > > > +static void zswap_alloc_shrinker(struct zswap_pool *pool)
> > > > > > +{
> > > > > > +       pool->shrinker =
> > > > > > +               shrinker_alloc(SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE, "mm-zswap");
> > > > > > +       if (!pool->shrinker)
> > > > > > +               return;
> > > > > > +
> > > > > > +       pool->shrinker->private_data = pool;
> > > > > > +       pool->shrinker->scan_objects = zswap_shrinker_scan;
> > > > > > +       pool->shrinker->count_objects = zswap_shrinker_count;
> > > > > > +       pool->shrinker->batch = 0;
> > > > > > +       pool->shrinker->seeks = DEFAULT_SEEKS;
> > > > > > +}
> > > > > > +
> > > > > >  /*********************************
> > > > > >  * per-cpu code
> > > > > >  **********************************/
> > > > > > @@ -656,11 +795,14 @@ static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_o
> > > > > >                                        spinlock_t *lock, void *arg)
> > > > > >  {
> > > > > >         struct zswap_entry *entry = container_of(item, struct zswap_entry, lru);
> > > > > > +       bool *encountered_page_in_swapcache = (bool *)arg;
> > > > > >         struct mem_cgroup *memcg;
> > > > > >         struct zswap_tree *tree;
> > > > > > +       struct lruvec *lruvec;
> > > > > >         pgoff_t swpoffset;
> > > > > >         enum lru_status ret = LRU_REMOVED_RETRY;
> > > > > >         int writeback_result;
> > > > > > +       unsigned long flags;
> > > > > >
> > > > > >         /*
> > > > > >          * Once the lru lock is dropped, the entry might get freed. The
> > > > > > @@ -696,8 +838,24 @@ static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_o
> > > > > >                 /* we cannot use zswap_lru_add here, because it increments node's lru count */
> > > > > >                 list_lru_putback(&entry->pool->list_lru, item, entry->nid, memcg);
> > > > > >                 spin_unlock(lock);
> > > > > > -               mem_cgroup_put(memcg);
> > > > > >                 ret = LRU_RETRY;
> > > > > > +
> > > > > > +               /*
> > > > > > +                * Encountering a page already in swap cache is a sign that we are shrinking
> > > > > > +                * into the warmer region. We should terminate shrinking (if we're in the dynamic
> > > > > > +                * shrinker context).
> > > > > > +                */
> > > > > > +               if (writeback_result == -EEXIST && encountered_page_in_swapcache) {
> > > > > > +                       ret = LRU_SKIP;
> > > > > > +                       *encountered_page_in_swapcache = true;
> > > > > > +               }
> > > > > > +               lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(entry->nid));
> > > > > > +               spin_lock_irqsave(&lruvec->lru_lock, flags);
> > > > > > +               /* Increment the protection area to account for the LRU rotation. */
> > > > > > +               lruvec->nr_zswap_protected++;
> > > > > > +               spin_unlock_irqrestore(&lruvec->lru_lock, flags);
> > > > > > +
> > > > > > +               mem_cgroup_put(memcg);
> > > > > >                 goto put_unlock;
> > > > > >         }
> > > > > >
> > > > > > @@ -828,6 +986,11 @@ static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
> > > > > >                                        &pool->node);
> > > > > >         if (ret)
> > > > > >                 goto error;
> > > > > > +
> > > > > > +       zswap_alloc_shrinker(pool);
> > > > > > +       if (!pool->shrinker)
> > > > > > +               goto error;
> > > > > > +
> > > > > >         pr_debug("using %s compressor\n", pool->tfm_name);
> > > > > >
> > > > > >         /* being the current pool takes 1 ref; this func expects the
> > > > > > @@ -836,12 +999,17 @@ static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
> > > > > >         kref_init(&pool->kref);
> > > > > >         INIT_LIST_HEAD(&pool->list);
> > > > > >         INIT_WORK(&pool->shrink_work, shrink_worker);
> > > > > > -       list_lru_init_memcg(&pool->list_lru, NULL);
> > > > > > +       if (list_lru_init_memcg(&pool->list_lru, pool->shrinker))
> > > > > > +               goto lru_fail;
> > > > > > +       shrinker_register(pool->shrinker);
> > > > > >
> > > > > >         zswap_pool_debug("created", pool);
> > > > > >
> > > > > >         return pool;
> > > > > >
> > > > > > +lru_fail:
> > > > > > +       list_lru_destroy(&pool->list_lru);
> > > > > > +       shrinker_free(pool->shrinker);
> > > > > >  error:
> > > > > >         if (pool->acomp_ctx)
> > > > > >                 free_percpu(pool->acomp_ctx);
> > > > > > @@ -899,6 +1067,7 @@ static void zswap_pool_destroy(struct zswap_pool *pool)
> > > > > >
> > > > > >         zswap_pool_debug("destroying", pool);
> > > > > >
> > > > > > +       shrinker_free(pool->shrinker);
> > > > > >         cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
> > > > > >         free_percpu(pool->acomp_ctx);
> > > > > >         list_lru_destroy(&pool->list_lru);
> > > > > > @@ -1431,6 +1600,7 @@ bool zswap_store(struct folio *folio)
> > > > > >         if (entry->length) {
> > > > > >                 INIT_LIST_HEAD(&entry->lru);
> > > > > >                 zswap_lru_add(&pool->list_lru, entry);
> > > > > > +               atomic_inc(&pool->nr_stored);
> > > > > >         }
> > > > > >         spin_unlock(&tree->lock);
> > > > > >
> > > > > > --
> > > > > > 2.34.1
> > > > Thanks for the comments/suggestion, Yosry!