[PATCH v4 1/2] slub: introduce count_partial_free_approx()

[PATCH v4 1/2] slub: introduce count_partial_free_approx()

[Jianfeng Wang]

When reading "/proc/slabinfo", the kernel needs to report the number
of free objects for each kmem_cache. The current implementation uses
count_partial() to get it by scanning each kmem_cache_node's partial
slab list and summing free objects from every partial slab. This
process must hold per-kmem_cache_node spinlock and disable IRQ, and
may take a long time. Consequently, it can block slab allocations on
other CPUs and cause timeouts for network devices, when the partial
list is long. In production, even NMI watchdog can be triggered due
to this matter: e.g., for "buffer_head", the number of partial slabs
was observed to be ~1M in one kmem_cache_node. This problem was also
confirmed by others [1-3].

Iterating a partial list to get the exact count of objects can cause
soft lockups for a long list with or without the lock (e.g., if
preemption is disabled), and may not be very useful: the object count
can change after the lock is released. The approach of maintaining
free-object counters requires atomic operations on the fast path [3].

So, the fix is to introduce count_partial_free_approx(). This function
can be used for getting the free object count in a kmem_cache_node's
partial list. It limits the number of slabs to scan and avoids scanning
the whole list by giving an approximation for a long list. Suppose the
limit is N. If the list's length is not greater than N, output the exact
count by traversing the list; if its length is greater than N, output an
approximated count by traversing a subset of the list. The proposed
method is to scan N/2 slabs from the list's head and N/2 slabs from
the tail. For a partial list with ~280K slabs, benchmarks show that
it performs better than just counting from the list's head, after slabs
get sorted by kmem_cache_shrink(). Default the limit to 10000, as it
produces an approximation within 1% of the exact count for both
scenarios. Then, use count_partial_free_approx() in get_slabinfo().

Benchmarks: Diff = (exact - approximated) / exact
* Normal case (w/o kmem_cache_shrink()):
| MAX_TO_SCAN | Diff (count from head)| Diff (count head+tail)|
| 1000        |  0.43  %              |  1.09  %              |
| 5000        |  0.06  %              |  0.37  %              |
| 10000       |  0.02  %              |  0.16  %              |
| 20000       |  0.009 %              | -0.003 %              |

* Skewed case (w/ kmem_cache_shrink()):
| MAX_TO_SCAN | Diff (count from head)| Diff (count head+tail)|
| 1000        |  12.46 %              |  6.75  %              |
| 5000        |  5.38  %              |  1.27  %              |
| 10000       |  4.99  %              |  0.22  %              |
| 20000       |  4.86  %              | -0.06  %              |

[1] https://lore.kernel.org/linux-mm/alpine.DEB.2.21.2003031602460.1537@www.lameter.com/T/
[2] https://lore.kernel.org/lkml/alpine.DEB.2.22.394.2008071258020.55871@www.lameter.com/T/
[3] https://lore.kernel.org/lkml/1e01092b-140d-2bab-aeba-321a74a194ee@linux.com/T/

Signed-off-by: Jianfeng Wang <jianfeng.w.wang@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
---
 mm/slub.c | 39 ++++++++++++++++++++++++++++++++++++++-
 1 file changed, 38 insertions(+), 1 deletion(-)

diff --git a/mm/slub.c b/mm/slub.c
index 1bb2a93cf7b6..6d8ecad07daf 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -3229,6 +3229,43 @@ static unsigned long count_partial(struct kmem_cache_node *n,
 #endif /* CONFIG_SLUB_DEBUG || SLAB_SUPPORTS_SYSFS */
 
 #ifdef CONFIG_SLUB_DEBUG
+#define MAX_PARTIAL_TO_SCAN 10000
+
+static unsigned long count_partial_free_approx(struct kmem_cache_node *n)
+{
+	unsigned long flags;
+	unsigned long x = 0;
+	struct slab *slab;
+
+	spin_lock_irqsave(&n->list_lock, flags);
+	if (n->nr_partial <= MAX_PARTIAL_TO_SCAN) {
+		list_for_each_entry(slab, &n->partial, slab_list)
+			x += slab->objects - slab->inuse;
+	} else {
+		/*
+		 * For a long list, approximate the total count of objects in
+		 * it to meet the limit on the number of slabs to scan.
+		 * Scan from both the list's head and tail for better accuracy.
+		 */
+		unsigned long scanned = 0;
+
+		list_for_each_entry(slab, &n->partial, slab_list) {
+			x += slab->objects - slab->inuse;
+			if (++scanned == MAX_PARTIAL_TO_SCAN / 2)
+				break;
+		}
+		list_for_each_entry_reverse(slab, &n->partial, slab_list) {
+			x += slab->objects - slab->inuse;
+			if (++scanned == MAX_PARTIAL_TO_SCAN)
+				break;
+		}
+		x = mult_frac(x, n->nr_partial, scanned);
+		x = min(x, node_nr_objs(n));
+	}
+	spin_unlock_irqrestore(&n->list_lock, flags);
+	return x;
+}
+
 static noinline void
 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
 {
@@ -7089,7 +7126,7 @@ void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
 	for_each_kmem_cache_node(s, node, n) {
 		nr_slabs += node_nr_slabs(n);
 		nr_objs += node_nr_objs(n);
-		nr_free += count_partial(n, count_free);
+		nr_free += count_partial_free_approx(n);
 	}
 
 	sinfo->active_objs = nr_objs - nr_free;
-- 
2.42.1

[PATCH v4 0/2] slub: introduce count_partial_free_approx()

[Jianfeng Wang]

This patch fixes a known issue in get_slabinfo() which relies on
count_partial() to get the exact count of free objects in a
kmem_cache_node's partial list. For some slubs, their partial lists
can be extremely long. Currently, count_partial() traverses a partial
list to get the exact count of objects. This process may take a long
time, during which slab allocations are blocked and IRQs are disabled.
In production, even NMI watchdog can be triggered due to this matter.

The fix is to limit the number of slabs to scan and output an
approximated count for a long partial list. The v1 patch counts N slabs
from the list's head and then uses it to estimate the total object
count in the list. As suggested by Vlastimil, the v2 patch uses an
alternative, i.e., counting N/2 from the list's head and tail, produces
a more accurate approximation after the partial list is sorted by
kmem_cache_shrink(). In the latest version, the implementation is moved
to a new function count_partial_free_approx() under CONFIG_SLUB_DEBUG.
count_partial() is still used in sysfs.

---
Changes since v3 [3]
 - Place count_partial_free_approx() under CONFIG_SLUB_DEBUG only

Changes since v2 [2]
 - Introduce count_partial_free_approx() and keep count_partial()
 - Use count_partial_free_approx() in get_slabinfo() and slab_out_of_memory()

Changes since v1 [1]
 - Update the approximation method by counting from the list's head and tail
 - Cap the approximation by the total object count
 - Update the commit message to add benchmark results and explain the choice

[1] https://lore.kernel.org/linux-mm/20240411164023.99368-1-jianfeng.w.wang@oracle.com/
[2] https://lore.kernel.org/linux-mm/20240417185938.5237-2-jianfeng.w.wang@oracle.com/
[3] https://lore.kernel.org/linux-mm/20240419175611.47413-1-jianfeng.w.wang@oracle.com/

Thanks,
--Jianfeng

Jianfeng Wang (2):
  slub: introduce count_partial_free_approx()
  slub: use count_partial_free_approx() in slab_out_of_memory()

 mm/slub.c | 41 +++++++++++++++++++++++++++++++++++++++--
 1 file changed, 39 insertions(+), 2 deletions(-)

-- 
2.42.1

[PATCH v4 1/2] slub: introduce count_partial_free_approx()

[Jianfeng Wang]

When reading "/proc/slabinfo", the kernel needs to report the number
of free objects for each kmem_cache. The current implementation uses
count_partial() to get it by scanning each kmem_cache_node's partial
slab list and summing free objects from every partial slab. This
process must hold per-kmem_cache_node spinlock and disable IRQ, and
may take a long time. Consequently, it can block slab allocations on
other CPUs and cause timeouts for network devices, when the partial
list is long. In production, even NMI watchdog can be triggered due
to this matter: e.g., for "buffer_head", the number of partial slabs
was observed to be ~1M in one kmem_cache_node. This problem was also
confirmed by others [1-3].

Iterating a partial list to get the exact count of objects can cause
soft lockups for a long list with or without the lock (e.g., if
preemption is disabled), and may not be very useful: the object count
can change after the lock is released. The approach of maintaining
free-object counters requires atomic operations on the fast path [3].

So, the fix is to introduce count_partial_free_approx(). This function
can be used for getting the free object count in a kmem_cache_node's
partial list. It limits the number of slabs to scan and avoids scanning
the whole list by giving an approximation for a long list. Suppose the
limit is N. If the list's length is not greater than N, output the exact
count by traversing the list; if its length is greater than N, output an
approximated count by traversing a subset of the list. The proposed
method is to scan N/2 slabs from the list's head and N/2 slabs from
the tail. For a partial list with ~280K slabs, benchmarks show that
it performs better than just counting from the list's head, after slabs
get sorted by kmem_cache_shrink(). Default the limit to 10000, as it
produces an approximation within 1% of the exact count for both
scenarios. Then, use count_partial_free_approx() in get_slabinfo().

Benchmarks: Diff = (exact - approximated) / exact
* Normal case (w/o kmem_cache_shrink()):
| MAX_TO_SCAN | Diff (count from head)| Diff (count head+tail)|
| 1000        |  0.43  %              |  1.09  %              |
| 5000        |  0.06  %              |  0.37  %              |
| 10000       |  0.02  %              |  0.16  %              |
| 20000       |  0.009 %              | -0.003 %              |

* Skewed case (w/ kmem_cache_shrink()):
| MAX_TO_SCAN | Diff (count from head)| Diff (count head+tail)|
| 1000        |  12.46 %              |  6.75  %              |
| 5000        |  5.38  %              |  1.27  %              |
| 10000       |  4.99  %              |  0.22  %              |
| 20000       |  4.86  %              | -0.06  %              |

[1] https://lore.kernel.org/linux-mm/alpine.DEB.2.21.2003031602460.1537@www.lameter.com/T/
[2] https://lore.kernel.org/lkml/alpine.DEB.2.22.394.2008071258020.55871@www.lameter.com/T/
[3] https://lore.kernel.org/lkml/1e01092b-140d-2bab-aeba-321a74a194ee@linux.com/T/

Signed-off-by: Jianfeng Wang <jianfeng.w.wang@oracle.com>
Acked-by: David Rientjes <rientjes@google.com>
---
 mm/slub.c | 39 ++++++++++++++++++++++++++++++++++++++-
 1 file changed, 38 insertions(+), 1 deletion(-)

diff --git a/mm/slub.c b/mm/slub.c
index 1bb2a93cf7b6..6d8ecad07daf 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -3229,6 +3229,43 @@ static unsigned long count_partial(struct kmem_cache_node *n,
 #endif /* CONFIG_SLUB_DEBUG || SLAB_SUPPORTS_SYSFS */
 
 #ifdef CONFIG_SLUB_DEBUG
+#define MAX_PARTIAL_TO_SCAN 10000
+
+static unsigned long count_partial_free_approx(struct kmem_cache_node *n)
+{
+	unsigned long flags;
+	unsigned long x = 0;
+	struct slab *slab;
+
+	spin_lock_irqsave(&n->list_lock, flags);
+	if (n->nr_partial <= MAX_PARTIAL_TO_SCAN) {
+		list_for_each_entry(slab, &n->partial, slab_list)
+			x += slab->objects - slab->inuse;
+	} else {
+		/*
+		 * For a long list, approximate the total count of objects in
+		 * it to meet the limit on the number of slabs to scan.
+		 * Scan from both the list's head and tail for better accuracy.
+		 */
+		unsigned long scanned = 0;
+
+		list_for_each_entry(slab, &n->partial, slab_list) {
+			x += slab->objects - slab->inuse;
+			if (++scanned == MAX_PARTIAL_TO_SCAN / 2)
+				break;
+		}
+		list_for_each_entry_reverse(slab, &n->partial, slab_list) {
+			x += slab->objects - slab->inuse;
+			if (++scanned == MAX_PARTIAL_TO_SCAN)
+				break;
+		}
+		x = mult_frac(x, n->nr_partial, scanned);
+		x = min(x, node_nr_objs(n));
+	}
+	spin_unlock_irqrestore(&n->list_lock, flags);
+	return x;
+}
+
 static noinline void
 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
 {
@@ -7089,7 +7126,7 @@ void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
 	for_each_kmem_cache_node(s, node, n) {
 		nr_slabs += node_nr_slabs(n);
 		nr_objs += node_nr_objs(n);
-		nr_free += count_partial(n, count_free);
+		nr_free += count_partial_free_approx(n);
 	}
 
 	sinfo->active_objs = nr_objs - nr_free;
-- 
2.42.1

Re: [PATCH v4 1/2] slub: introduce count_partial_free_approx()

[Christoph Lameter (Ampere)]

Acked-by: Christoph Lameter <cl@linux.com>