Re: [PATCH] slub: avoid scanning all partial slabs in get_slabinfo()

[Jianfeng Wang <jianfeng.w.wang@oracle.com>]

On 2/22/24 7:02 PM, Christoph Lameter (Ampere) wrote:
> On Thu, 22 Feb 2024, Chengming Zhou wrote:
> 
>> Anyway, I put the code below for discussion...
> 
> Can we guestimate the free objects based on the number of partial slabs. That number is available.
> 

Yes.
I've thought about calculating the average number of free objects in a
partial slab (through sampling) and then estimating the total number of
free objects as (avg * n->nr_partial).

See the following.

---
 mm/slub.c | 20 ++++++++++++++++++--
 1 file changed, 18 insertions(+), 2 deletions(-)

diff --git a/mm/slub.c b/mm/slub.c
index 63d281dfacdb..13385761049c 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -2963,6 +2963,8 @@ static inline bool free_debug_processing(struct kmem_cache *s,
 #endif /* CONFIG_SLUB_DEBUG */
 
 #if defined(CONFIG_SLUB_DEBUG) || defined(SLAB_SUPPORTS_SYSFS)
+#define MAX_PARTIAL_TO_SCAN 10000
+
 static unsigned long count_partial(struct kmem_cache_node *n,
 					int (*get_count)(struct slab *))
 {
@@ -2971,8 +2973,22 @@ static unsigned long count_partial(struct kmem_cache_node *n,
 	struct slab *slab;
 
 	spin_lock_irqsave(&n->list_lock, flags);
-	list_for_each_entry(slab, &n->partial, slab_list)
-		x += get_count(slab);
+	if (n->nr_partial > MAX_PARTIAL_TO_SCAN) {
+		/* Estimate total count of objects via sampling */
+		unsigned long sample_rate = n->nr_partial / MAX_PARTIAL_TO_SCAN;
+		unsigned long scanned = 0;
+		unsigned long counted = 0;
+		list_for_each_entry(slab, &n->partial, slab_list) {
+			if (++scanned % sample_rate == 0) {
+				x += get_count(slab);
+				counted++;
+			}
+		}
+		x = mult_frac(x, n->nr_partial, counted);
+	} else {
+		list_for_each_entry(slab, &n->partial, slab_list)
+			x += get_count(slab);
+	}
 	spin_unlock_irqrestore(&n->list_lock, flags);
 	return x;
 }
-- 

> How accurate need the accounting be? We also have fuzzy accounting in the VM counters.
Based on my experience, for a |kmem_cache|, the total number of objects can tell
whether the |kmem_cache| has been heavily used by a workload. When the total
number is large: if the number of free objects is small, then either these objects
are really in-use or there is *memory leak* going on (which then must be further
diagnosed). However, if the number of free objects is large, we can only know
the slab memory fragmentation happens.

So, I think the object accounting needn't be accurate. We only have to tell
whether a large percentage of slab objects is free or not. The above code is a
sampling, which should do the job if we take enough samples.