Re: [RFC PATCH v4 0/2] mm/zsmalloc: defer object free to per-cpu batches
From: Sergey Senozhatsky
Date: Fri Nov 20 2026 - 02:48:43 EST
On (26/11/20 10:44), Sergey Senozhatsky wrote:
> On (26/11/19 14:35), Sergey Senozhatsky wrote:
> > On (26/11/09 18:59), Hongfeng Chen wrote:
> > > On Fri, Oct 23, 2026 at 11:19 AM Nhat Pham <nphamcs@xxxxxxxxx> wrote:
> > > > Is there any more juice we can squeeze from zsmalloc in that case?
> > > >
> > > > It would be nice if you can perf zsmalloc freeing. And yeah, maybe we
> > > > can't optimize any further, but let's eliminate that first before we
> > > > rely on offloading.
> > >
> > > Here is the perf breakdown of the free path:
> > >
> > > zram:
> > >
> > > __swap_cluster_free_entries 90.28%
> > > zram_slot_free_notify 86.01%
> > > slot_free 74.15%
> > > zs_free 56.01%
> > > __free_frozen_pages 19.85%
> > > kmem_cache_free 10.15%
> > > _raw_spin_lock 5.53%
> > > fix_fullness_group 3.31%
> > > dec_zone_page_state 2.43%
> > > obj_free 2.17%
> > > trylock_zspage 2.12%
> >
> > We removed one of kmem_cache_free() from zs_free() [1].
> >
> > [1] https://lore.kernel.org/linux-kernel/20261110054836.1515030-1-senozhatsky@xxxxxxxxxxxx
>
> I suppose we can batch some operations in zs_free(), something like
> this:
A slightly reworked patch below - handle incompressible and pow-of-two
zspages separately (w/o batching), because those are 1-page zspages,
nothing to batch.
---
mm/zsmalloc.c | 52 +++++++++++++++++++++++++++++++++++++++++++++------
1 file changed, 46 insertions(+), 6 deletions(-)
diff --git a/mm/zsmalloc.c b/mm/zsmalloc.c
index f571b3c9e9ca..9dcd1549c2fb 100644
--- a/mm/zsmalloc.c
+++ b/mm/zsmalloc.c
@@ -37,6 +37,7 @@
#include <linux/debugfs.h>
#include <linux/zsmalloc.h>
#include <linux/fs.h>
+#include <linux/folio_batch.h>
#include <linux/workqueue.h>
#include "zpdesc.h"
@@ -866,6 +867,44 @@ static int trylock_zspage(struct zspage *zspage)
return 0;
}
+/*
+ * Free the pages of a multi-page zspage in one go, see folios_put().
+ * Not inlined, so that the folio_batch is only set up (and, with
+ * CONFIG_INIT_STACK_ALL_ZERO, zero-filled) when there is something
+ * to batch.
+ */
+static noinline void free_page_chain(struct zpdesc *zpdesc)
+{
+ struct zone *zone = zpdesc_zone(zpdesc);
+ struct folio_batch fbatch;
+ int nr = 0;
+
+ static_assert(ZS_MAX_PAGES_PER_ZSPAGE <= FOLIO_BATCH_SIZE);
+
+ folio_batch_init(&fbatch);
+ do {
+ struct zpdesc *next = get_next_zpdesc(zpdesc);
+
+ VM_BUG_ON_PAGE(!zpdesc_is_locked(zpdesc), zpdesc_page(zpdesc));
+ reset_zpdesc(zpdesc);
+ zpdesc_unlock(zpdesc);
+
+ /* NR_ZSPAGES is per zone: one update per same-zone run */
+ if (unlikely(zpdesc_zone(zpdesc) != zone)) {
+ mod_zone_page_state(zone, NR_ZSPAGES, -nr);
+ zone = zpdesc_zone(zpdesc);
+ nr = 0;
+ }
+ nr++;
+
+ folio_batch_add(&fbatch, zpdesc_folio(zpdesc));
+ zpdesc = next;
+ } while (zpdesc);
+
+ mod_zone_page_state(zone, NR_ZSPAGES, -nr);
+ folios_put(&fbatch);
+}
+
/*
* Three free helpers, kept apart here:
*
@@ -884,21 +923,22 @@ static int trylock_zspage(struct zspage *zspage)
*/
static inline void __free_zspage_lockless(struct zspage *zspage)
{
- struct zpdesc *zpdesc, *next;
+ struct zpdesc *zpdesc;
VM_BUG_ON(get_zspage_inuse(zspage));
VM_BUG_ON(zspage->fullness != ZS_INUSE_RATIO_0);
- next = zpdesc = get_first_zpdesc(zspage);
- do {
+ zpdesc = get_first_zpdesc(zspage);
+ if (get_next_zpdesc(zpdesc)) {
+ free_page_chain(zpdesc);
+ } else {
+ /* Single-page zspage, e.g. a huge object: nothing to batch */
VM_BUG_ON_PAGE(!zpdesc_is_locked(zpdesc), zpdesc_page(zpdesc));
- next = get_next_zpdesc(zpdesc);
reset_zpdesc(zpdesc);
zpdesc_unlock(zpdesc);
zpdesc_dec_zone_page_state(zpdesc);
zpdesc_put(zpdesc);
- zpdesc = next;
- } while (zpdesc != NULL);
+ }
cache_free_zspage(zspage);
}
--
2.56.1.750.gf70dc682cf-goog