[PATCH 0/2] sched/fair: Reset incompatible burst on quota change

[PATCH 0/2] sched/fair: Reset incompatible burst on quota change

[Zhe Liu]

A burst configured while a cgroup has unlimited CPU bandwidth can prevent
a later finite quota from being installed.  On cgroup v2, for instance:

  # echo 100000000 > cpu.max.burst
  # echo "50000 100000" > cpu.max
  sh: write error: Invalid argument

The quota stays unlimited because tg_set_bandwidth() validates the existing
burst against the new quota, and the only recovery is for userspace to know
that it must clear the burst before retrying the quota update.  cgroup v1 has
the same problem through cpu.cfs_quota_us.

Patch 1 resets the existing burst to zero when a valid finite quota is written
that the burst is incompatible with, while preserving it when it remains
compatible or when the new quota is unlimited.  This lets a quota update take
effect regardless of the order in which userspace writes the two files, and
still allows a burst to be staged before bandwidth control is enabled.  A
cgroup v2 selftest is added for the behavior.

Patch 2 documents the reset behavior in the cgroup v1 and v2 CPU bandwidth
documentation.

Zhe Liu (2):
  sched/fair: Reset incompatible burst on quota change
  Documentation: describe burst reset on quota changes

 Documentation/admin-guide/cgroup-v2.rst   |  5 +-
 Documentation/scheduler/sched-bwc.rst     | 14 ++---
 kernel/sched/core.c                       | 17 ++++++-
 tools/testing/selftests/cgroup/test_cpu.c | 62 +++++++++++++++++++++++
 4 files changed, 90 insertions(+), 8 deletions(-)

-- 
2.25.1

[PATCH 1/2] sched/fair: Reset incompatible burst on quota change

[Zhe Liu]

A burst configured while a cgroup has unlimited CPU bandwidth can prevent
a later finite quota from being installed.  For example, on cgroup v2:

  # echo 100000000 > cpu.max.burst
  # echo "50000 100000" > cpu.max
  sh: write error: Invalid argument

The quota remains unlimited because tg_set_bandwidth() validates the
existing burst against the new quota.  Recovering requires userspace to
know that it must clear the burst before retrying the quota update.  The
same problem affects cpu.cfs_quota_us on cgroup v1.

When changing the quota, reset the existing burst to zero if it is
incompatible with a valid finite quota.  Preserve it when it remains
compatible or when the new quota is unlimited.  This lets a quota update
take effect without depending on the order in which userspace writes the
two files.

Rejecting the quota would retain this ordering dependency.  Clamping the
burst would instead silently choose a different nonzero policy on behalf
of userspace.  Resetting it to zero provides the existing no-burst default
while leaving compatible bursts untouched.  Keeping the burst while the
quota is unlimited also allows userspace to stage a burst before enabling
bandwidth control.

Add a cgroup v2 regression test for the quota update behavior.

Fixes: f4183717b370 ("sched/fair: Introduce the burstable CFS controller")
Signed-off-by: Zhe Liu <liuzhe1@kylinos.cn>
---
 kernel/sched/core.c                       | 17 ++++++-
 tools/testing/selftests/cgroup/test_cpu.c | 62 +++++++++++++++++++++++
 2 files changed, 78 insertions(+), 1 deletion(-)

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 2e7cde033a31..324a4d22f8d1 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -10081,6 +10081,18 @@ static u64 cpu_period_read_u64(struct cgroup_subsys_state *css,
 	return period_us;
 }
 
+static u64 tg_burst_on_quota_change(u64 quota_us, u64 burst_us)
+{
+	if (quota_us == RUNTIME_INF || quota_us > max_bw_runtime_us)
+		return burst_us;
+
+	if (burst_us > quota_us ||
+	    burst_us > max_bw_runtime_us - quota_us)
+		return 0;
+
+	return burst_us;
+}
+
 static int tg_set_bandwidth(struct task_group *tg,
 			    u64 period_us, u64 quota_us, u64 burst_us)
 {
@@ -10169,6 +10181,7 @@ static int cpu_quota_write_s64(struct cgroup_subsys_state *css,
 		quota_us = RUNTIME_INF;
 
 	tg_bandwidth(tg, &period_us, NULL, &burst_us);
+	burst_us = tg_burst_on_quota_change(quota_us, burst_us);
 	return tg_set_bandwidth(tg, period_us, quota_us, burst_us);
 }
 
@@ -10492,8 +10505,10 @@ static ssize_t cpu_max_write(struct kernfs_open_file *of,
 
 	tg_bandwidth(tg, &period_us, NULL, &burst_us);
 	ret = cpu_period_quota_parse(buf, &period_us, &quota_us);
-	if (!ret)
+	if (!ret) {
+		burst_us = tg_burst_on_quota_change(quota_us, burst_us);
 		ret = tg_set_bandwidth(tg, period_us, quota_us, burst_us);
+	}
 	return ret ?: nbytes;
 }
 #endif /* CONFIG_CFS_BANDWIDTH */
diff --git a/tools/testing/selftests/cgroup/test_cpu.c b/tools/testing/selftests/cgroup/test_cpu.c
index 7a40d76b9548..2686dd79941b 100644
--- a/tools/testing/selftests/cgroup/test_cpu.c
+++ b/tools/testing/selftests/cgroup/test_cpu.c
@@ -703,6 +703,67 @@ static int test_cpucg_max(const char *root)
 	return ret;
 }
 
+/*
+ * This test verifies that writing a finite cpu.max resets an incompatible
+ * cpu.max.burst, while preserving a compatible burst.
+ */
+static int test_cpucg_max_burst_reset(const char *root)
+{
+	char *cpucg = NULL;
+	int ret = KSFT_FAIL;
+
+	cpucg = cg_name(root, "cpucg_max_burst_reset_test");
+	if (!cpucg)
+		goto cleanup;
+
+	if (cg_create(cpucg))
+		goto cleanup;
+
+	/* An unconstrained group may retain a burst for later use. */
+	if (cg_write(cpucg, "cpu.max.burst", "100000000"))
+		goto cleanup;
+	if (cg_read_long(cpucg, "cpu.max.burst") != 100000000)
+		goto cleanup;
+
+	/* A finite quota must not be blocked by the incompatible burst. */
+	if (cg_write(cpucg, "cpu.max", "50000 100000"))
+		goto cleanup;
+	if (cg_read_long(cpucg, "cpu.max.burst") != 0)
+		goto cleanup;
+	if (cg_read_strcmp(cpucg, "cpu.max", "50000 100000\n"))
+		goto cleanup;
+
+	/* Keep a burst which remains valid across a quota update. */
+	if (cg_write(cpucg, "cpu.max", "100000 100000"))
+		goto cleanup;
+	if (cg_write(cpucg, "cpu.max.burst", "50000"))
+		goto cleanup;
+	if (cg_write(cpucg, "cpu.max", "75000 100000"))
+		goto cleanup;
+	if (cg_read_long(cpucg, "cpu.max.burst") != 50000)
+		goto cleanup;
+
+	/* An unlimited quota preserves burst until it becomes incompatible. */
+	if (cg_write(cpucg, "cpu.max", "max 100000"))
+		goto cleanup;
+	if (cg_read_long(cpucg, "cpu.max.burst") != 50000)
+		goto cleanup;
+
+	/* Reset the same burst when a later finite quota conflicts. */
+	if (cg_write(cpucg, "cpu.max", "25000 100000"))
+		goto cleanup;
+	if (cg_read_long(cpucg, "cpu.max.burst") != 0)
+		goto cleanup;
+
+	ret = KSFT_PASS;
+
+cleanup:
+	cg_destroy(cpucg);
+	free(cpucg);
+
+	return ret;
+}
+
 /*
  * This test verifies that a process inside of a nested cgroup whose parent
  * group has a cpu.max value set, is properly throttled.
@@ -789,6 +850,7 @@ struct cpucg_test {
 	T(test_cpucg_nested_weight_overprovisioned),
 	T(test_cpucg_nested_weight_underprovisioned),
 	T(test_cpucg_max),
+	T(test_cpucg_max_burst_reset),
 	T(test_cpucg_max_nested),
 };
 #undef T
-- 
2.25.1

[PATCH 2/2] Documentation: describe burst reset on quota changes

[Zhe Liu]

Document that an incompatible burst is reset when a finite quota is
written through the cgroup v1 or cgroup v2 CPU bandwidth interface, while
compatible bursts and unlimited quota updates preserve the existing value.

Signed-off-by: Zhe Liu <liuzhe1@kylinos.cn>
---
 Documentation/admin-guide/cgroup-v2.rst |  5 ++++-
 Documentation/scheduler/sched-bwc.rst   | 14 ++++++++------
 2 files changed, 12 insertions(+), 7 deletions(-)

diff --git a/Documentation/admin-guide/cgroup-v2.rst b/Documentation/admin-guide/cgroup-v2.rst
index aed195a71cbf..9f5e43d96476 100644
--- a/Documentation/admin-guide/cgroup-v2.rst
+++ b/Documentation/admin-guide/cgroup-v2.rst
@@ -1200,7 +1200,10 @@ will be referred to. All time durations are in microseconds.
 
 	which indicates that the group may consume up to $MAX in each
 	$PERIOD duration.  "max" for $MAX indicates no limit.  If only
-	one number is written, $MAX is updated.
+	one number is written, $MAX is updated.  When a finite $MAX is
+	written, an existing cpu.max.burst value which is incompatible
+	with the new limit is reset to zero.  Writing "max" leaves the
+	burst value unchanged.
 
 	This file affects only processes under the fair-class scheduler.
 
diff --git a/Documentation/scheduler/sched-bwc.rst b/Documentation/scheduler/sched-bwc.rst
index e881a945c188..79bd5f254e50 100644
--- a/Documentation/scheduler/sched-bwc.rst
+++ b/Documentation/scheduler/sched-bwc.rst
@@ -90,14 +90,16 @@ bandwidth restriction in place, such a group is described as an unconstrained
 bandwidth group. This represents the traditional work-conserving behavior for
 CFS.
 
-Writing any (valid) positive value(s) no smaller than cpu.cfs_burst_us will
-enact the specified bandwidth limit. The minimum quota allowed for the quota or
-period is 1ms. There is also an upper bound on the period length of 1s.
-Additional restrictions exist when bandwidth limits are used in a hierarchical
-fashion, these are explained in more detail below.
+Writing any valid positive value will enact the specified bandwidth limit. If
+the existing cpu.cfs_burst_us value is incompatible with the new quota, it is
+reset to zero. The minimum quota allowed for the quota or period is 1ms. There
+is also an upper bound on the period length of 1s. Additional restrictions
+exist when bandwidth limits are used in a hierarchical fashion, these are
+explained in more detail below.
 
 Writing any negative value to cpu.cfs_quota_us will remove the bandwidth limit
-and return the group to an unconstrained state once more.
+and return the group to an unconstrained state once more. The existing
+cpu.cfs_burst_us value remains unchanged.
 
 A value of 0 for cpu.cfs_burst_us indicates that the group can not accumulate
 any unused bandwidth. It makes the traditional bandwidth control behavior for
-- 
2.25.1

Re: [PATCH 1/2] sched/fair: Reset incompatible burst on quota change

[Michal Koutný]

[-- Attachment #1: Type: text/plain, Size: 1564 bytes --]

On Thu, Aug 20, 2026 at 11:32:17AM +0800, Zhe Liu <liuzhe1@kylinos.cn> wrote:
> A burst configured while a cgroup has unlimited CPU bandwidth can prevent
> a later finite quota from being installed.  For example, on cgroup v2:
> 
>   # echo 100000000 > cpu.max.burst
>   # echo "50000 100000" > cpu.max
>   sh: write error: Invalid argument
> 
> The quota remains unlimited because tg_set_bandwidth() validates the
> existing burst against the new quota.  Recovering requires userspace to
> know that it must clear the burst before retrying the quota update.  The
> same problem affects cpu.cfs_quota_us on cgroup v1.
> 
> When changing the quota, reset the existing burst to zero if it is
> incompatible with a valid finite quota.  Preserve it when it remains
> compatible or when the new quota is unlimited.  This lets a quota update
> take effect without depending on the order in which userspace writes the
> two files.
> 
> Rejecting the quota would retain this ordering dependency.  Clamping the
> burst would instead silently choose a different nonzero policy on behalf
> of userspace.

Why not clamp the burst_us to quota_us? That's quite natural to me.

> Resetting it to zero provides the existing no-burst default while
> leaving compatible bursts untouched.

Like Sashiko said, the user configured values should not get lost, the
resulting burst value (0 or quota or whatever makes sense) might be
applied effectively (to allow configuration order independence) but not
overwrite what was configured.

Thanks,
Michal

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 265 bytes --]