[PATCH RFC] mm, slab: add an optimistic __slab_try_return_freelist()

["Vlastimil Babka (SUSE)" <vbabka@kernel.org>]

When we end up returning extraneous objects during refill to a slab
where we just did a get_freelist_nofreeze(), it is likely no other CPU
has freed objects to it meanwhile. We can then reattach the remainder of
the freelist without having to walk the (potentially cache cold)
freelist to to find its tail to connect slab->freelist to it.

Add a __slab_try_return_freelist() function that does that. It's a bit
like __slab_free() but many situations from there cannot occur in this
specific scenario so it's simpler.

Signed-off-by: Vlastimil Babka (SUSE) <vbabka@kernel.org>
---
I've got this idea during the discussions on refill spilling, but not
sure if I've mentioned it on-list. Anyway it seems like there should be
no downsides (in theory...) so please test if it indeed improves things
anywhere and then it could be a better baseline before trying anything
that comes with tradeoffs?

I've included SLUB_STAT items that show to me the optimistic path is
indeed successful, but maybe they don't need to make it in final version
due to being rather low-level.

Git version here:

https://git.kernel.org/pub/scm/linux/kernel/git/vbabka/linux.git/log/?h=b4/refill-optimistic-return

It's Linus' upstream few days ago with "mm/slub: defer freelist
construction until after bulk allocation from a new slab" which is
heading to 7.2 so also should be considered part of baseline.
---
 mm/slub.c | 76 +++++++++++++++++++++++++++++++++++++++++++++++++++++++--------
 1 file changed, 67 insertions(+), 9 deletions(-)

diff --git a/mm/slub.c b/mm/slub.c
index 35b6cd0efc3b..176bc4936d03 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -373,6 +373,8 @@ enum stat_item {
 	SHEAF_PREFILL_OVERSIZE,	/* Allocation of oversize sheaf for prefill */
 	SHEAF_RETURN_FAST,	/* Sheaf return reattached spare sheaf */
 	SHEAF_RETURN_SLOW,	/* Sheaf return could not reattach spare */
+	REFILL_RETURN_FAST,
+	REFILL_RETURN_SLOW,
 	NR_SLUB_STAT_ITEMS
 };
 
@@ -4323,7 +4325,8 @@ static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags)
  * Assumes this is performed only for caches without debugging so we
  * don't need to worry about adding the slab to the full list.
  */
-static inline void *get_freelist_nofreeze(struct kmem_cache *s, struct slab *slab)
+static inline void *get_freelist_nofreeze(struct kmem_cache *s, struct slab *slab,
+					  unsigned int *count)
 {
 	struct freelist_counters old, new;
 
@@ -4339,6 +4342,7 @@ static inline void *get_freelist_nofreeze(struct kmem_cache *s, struct slab *sla
 
 	} while (!slab_update_freelist(s, slab, &old, &new, "get_freelist_nofreeze"));
 
+	*count = old.objects - old.inuse;
 	return old.freelist;
 }
 
@@ -5502,6 +5506,50 @@ static noinline void free_to_partial_list(
 	}
 }
 
+/*
+ * Try returning a (part of) freelist that we just detached from the slab.
+ * Optimistically assume the slab is still full so we don't need to know
+ * the tail of the freelist. Return to the partial list unconditionally even
+ * if it became empty.
+ *
+ * Fail if the slab isn't full anymore due to a cocurrent free.
+ *
+ * Can be only used for non-debug caches
+ */
+static bool __slab_try_return_freelist(struct kmem_cache *s, struct slab *slab,
+				       void *head, int cnt)
+{
+	struct freelist_counters old, new;
+	struct kmem_cache_node *n = NULL;
+	unsigned long flags;
+
+	old.freelist = slab->freelist;
+	old.counters = slab->counters;
+
+	if (old.freelist)
+		return false;
+
+	new.freelist = head;
+	new.counters = old.counters;
+	new.inuse -= cnt;
+
+	n = get_node(s, slab_nid(slab));
+
+	spin_lock_irqsave(&n->list_lock, flags);
+
+	if (!slab_update_freelist(s, slab, &old, &new, "__slab_try_return_freelist")) {
+		spin_unlock_irqrestore(&n->list_lock, flags);
+		return false;
+	}
+
+	add_partial(n, slab, ADD_TO_HEAD);
+	stat(s, FREE_ADD_PARTIAL);
+
+	spin_unlock_irqrestore(&n->list_lock, flags);
+	stat(s, REFILL_RETURN_FAST);
+	return true;
+}
+
 /*
  * Slow path handling. This may still be called frequently since objects
  * have a longer lifetime than the cpu slabs in most processing loads.
@@ -7113,34 +7161,40 @@ __refill_objects_node(struct kmem_cache *s, void **p, gfp_t gfp, unsigned int mi
 
 	list_for_each_entry_safe(slab, slab2, &pc.slabs, slab_list) {
 
+		unsigned int count;
+
 		list_del(&slab->slab_list);
 
-		object = get_freelist_nofreeze(s, slab);
+		object = get_freelist_nofreeze(s, slab, &count);
 
-		while (object && refilled < max) {
+		while (count && refilled < max) {
 			p[refilled] = object;
 			object = get_freepointer(s, object);
 			maybe_wipe_obj_freeptr(s, p[refilled]);
 
 			refilled++;
+			count--;
 		}
 
 		/*
 		 * Freelist had more objects than we can accommodate, we need to
-		 * free them back. We can treat it like a detached freelist, just
-		 * need to find the tail object.
+		 * free them back. First we try to be optimistic and assume the
+		 * slab is stil full since we just detached its freelist.
+		 * Otherwise we must need to find the tail object.
 		 */
-		if (unlikely(object)) {
+		if (unlikely(count)) {
 			void *head = object;
 			void *tail;
-			int cnt = 0;
+
+			if (__slab_try_return_freelist(s, slab, head, count))
+				break;
 
 			do {
 				tail = object;
-				cnt++;
 				object = get_freepointer(s, object);
 			} while (object);
-			__slab_free(s, slab, head, tail, cnt, _RET_IP_);
+			__slab_free(s, slab, head, tail, count, _RET_IP_);
+			stat(s, REFILL_RETURN_SLOW);
 		}
 
 		if (refilled >= max)
@@ -9366,6 +9420,8 @@ STAT_ATTR(SHEAF_PREFILL_SLOW, sheaf_prefill_slow);
 STAT_ATTR(SHEAF_PREFILL_OVERSIZE, sheaf_prefill_oversize);
 STAT_ATTR(SHEAF_RETURN_FAST, sheaf_return_fast);
 STAT_ATTR(SHEAF_RETURN_SLOW, sheaf_return_slow);
+STAT_ATTR(REFILL_RETURN_FAST, refill_return_fast);
+STAT_ATTR(REFILL_RETURN_SLOW, refill_return_slow);
 #endif	/* CONFIG_SLUB_STATS */
 
 #ifdef CONFIG_KFENCE
@@ -9454,6 +9510,8 @@ static const struct attribute *const slab_attrs[] = {
 	&sheaf_prefill_oversize_attr.attr,
 	&sheaf_return_fast_attr.attr,
 	&sheaf_return_slow_attr.attr,
+	&refill_return_fast_attr.attr,
+	&refill_return_slow_attr.attr,
 #endif
 #ifdef CONFIG_FAILSLAB
 	&failslab_attr.attr,

---
base-commit: da993c58f9bde299a5e1a4e7900125b32dccd2a6
change-id: 20260421-b4-refill-optimistic-return-f44d3b74cc49

Best regards,
--  
Vlastimil Babka (SUSE) <vbabka@kernel.org>