[PATCH 00/16] sched/deadline: first bandwidth accounting fixes and selftests

From: Mathieu Desnoyers

Date: Wed Nov 18 2026 - 15:07:59 EST


Hi Ingo, Juri,

This series fixes 13 bugs of SCHED_DEADLINE. Then it shows the deadline
bandwidth of each runqueue in sched_debug, and adds two selftests. The
fixes come first, so that stable can take patches 1 to 13.

Two of the bugs take the kernel down: a division by zero (patch 4), and
an owner's CPU that locks up (patch 8). One leaves cpuset_mutex locked
(patch 6). One leaves a task that never runs again (patch 7). Most of
the others leave total_bw of a root domain, or this_bw of a runqueue,
wrong. The root domain then admits deadline tasks beyond its capacity,
or refuses every one.

These are the first 14 patches of "sched/deadline: bandwidth accounting
fixes and selftests" v7 [1], as Juri proposed [2]. Their v7 numbers are
1-8, 10, 13, 14, 16, 20 and 37. The rest of v7 will follow as v8 of that
series, on top of this one.

Patch 11 is Juri's, patch 13 is Andrea Parri's. I carry them unchanged.
Sebastian reviewed patch 13. Juri, three of your four fixes that Ingo
dropped from sched/urgent are here:
- 2/4 is patch 10.
- 3/4 is patch 1.
- 4/4 is patch 11.
1/4 stays in the second series.

The fixes
=========

1 Refuse SCHED_FLAG_KEEP_PARAMS requests that would change the policy
or the class.
2 Skip admission for parameters that SCHED_FLAG_KEEP_PARAMS keeps.
3 Let a task boosted by a deadline waiter change its policy.
4 Do not divide by zero active CPUs of a root domain.
5 Refuse a dl_server debugfs write while the runqueue is offline.
6 Take cpuset_mutex for a deadline policy change only with pi, so
that sysrq-n does not leave it locked.
7 Replenish a throttled task switched out and back, so that it runs
again.
8 Do not refill a boost from cleared parameters, so that the owner's
CPU no longer locks up.
9 Reset the deadline and the runtime with the parameters.
10 Release the bandwidth of a task that leaves the class asleep.
11 Reject SCHED_DEADLINE on exited tasks (Juri Lelli).
12 Keep this_bw right when a task changes during a remote wakeup.
13 Check the affinity of a deadline task under the runqueue lock, so
that sched_setaffinity() and sched_setattr() cannot both succeed
(Andrea Parri).

Patch 14 shows the deadline bandwidth of each runqueue in sched_debug.
Patch 15 adds deadline_switch_test. Patch 16 adds deadline_bw_test, with
the 47 of its 112 cases that pass with these fixes. The other cases
come with their fixes, in the second series.

Known limit
===========

Patch 10 clears the parameters of a task that leaves the class asleep.
On v7.4-rc2, such a task kept them. That hid a bug of the rtmutex chain
walk: an owner can keep as its donor a deadline task that left the
class. Another deadline task then gets too little CPU time. With patch
10, the case pi_chain_top_moves shows this in every run. It passes on
v7.4-rc2. The case is new in the second series, with the two rtmutex
fixes that make it pass with patch 10.

Testing
=======

The 52 cases pass wherever they run, with lockdep and no kernel
warning, in 25 QEMU guests:
- 9 guests of 8 vCPUs, one with proxy execution off, and 1 of 4 vCPUs.
- 5 guests with PREEMPT_RT, 4 of 8 vCPUs and 1 of 4.
- 3 guests without CONFIG_CPUSETS, of 4, 4 and 5 vCPUs.
- 4 guests of 5 vCPUs booted with isolcpus=domain,1,2 or
isolcpus=domain,0,1, 2 of them with PREEMPT_RT.
- 1 guest of 8 vCPUs booted with isolcpus=domain,6,7.
- 2 guests with CONFIG_RT_GROUP_SCHED, of 8 and 4 vCPUs.
move_cancelled runs only in the last group. Without CONFIG_CPUSETS, the
cases that need a cpuset skip. With 3 CPUs that are not isolated,
params_waking and rebuild_record_taken skip: they need 4.

On v7.4-rc2, six cases of deadline_bw_test fail in each of 8 guests of 8
vCPUs. Patch 16 lists them. overrun_caught fails in one guest. In
another one, hotplug_server crashed the kernel with the division by zero
that patch 4 fixes. throttled_round_trip and asleep_round_trip of
deadline_switch_test fail too. The sysrq cases did not run there:
sysrq-n leaves cpuset_mutex locked, which patch 6 fixes. In earlier runs
with them, sysrq_normalize failed.

Each of the 14 kernel patches builds alone. The whole series also
builds without CONFIG_CPUSETS, with PREEMPT_RT and with
CONFIG_RT_GROUP_SCHED. The selftests build with no warning.

Not tested: cgroup v1 cpusets, the sched_ext server, and CPUs of
unequal capacity.

How this was developed
======================

An LLM coding assistant wrote these patches, their messages and the
tests, under my direction and review. It did not write patches 11 and
13. I know the concepts behind SCHED_DEADLINE, but not its
implementation well. So I ask those who know this code for a careful
reading of the fixes. A reproducer that stops failing shows that the
symptom goes away. It does not show that the fix is the one you would
have written.

The bugs came from:
- flaky cgroup selftests of a feature I am developing;
- the review of Juri's four fixes, and his review of v5;
- building deadline_bw_test, and reading the code;
- the Sashiko AI reviews of v1 to v7. Each finding was checked against
the code first.

Changes since v7
================

- Reviewed-by from Juri on 4, 5, 8 and 10. He reviewed the v7 code of
5, before the change listed below.
- 5: the refusal covers the whole CPU deactivation, from the Sashiko
review of v7.
- 8: the fix paragraph is Juri's wording.
- 10: its message names asleep_round_trip, which tests it without PI.
The two PI cases it named move to the second series.
- 15: the stack of the filler threads is at least PTHREAD_STACK_MIN.
The semaphores are destroyed.
- The messages credit the Sashiko AI review without the number of the
posting, as Juri suggested.

Based on tip sched/urgent, ec6f6731538e (v7.4-rc2).

[1] https://lore.kernel.org/all/cover.1794922146.git.mathieu.desnoyers@xxxxxxxxxxxx/
[2] https://lore.kernel.org/all/av3J3zMfkRplg-N4@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/

Andrea Parri (1):
sched/deadline: Check SCHED_DEADLINE affinity under the rq lock

Juri Lelli (1):
sched/deadline: Reject SCHED_DEADLINE on exited tasks

Mathieu Desnoyers (14):
sched: Refuse SCHED_FLAG_KEEP_PARAMS that would change the policy or
the class
sched/deadline: Do not admit bandwidth for parameters that are kept
sched: Let a task boosted by a deadline waiter change its policy
sched/deadline: Do not divide by zero active CPUs of a root domain
sched/deadline: Refuse dl_server debugfs writes while the rq is
offline
sched: Take cpuset_mutex for a deadline policy change only with pi
sched/deadline: Do not lose a throttled task switched out and back
sched/deadline: Do not replenish a boost from cleared parameters
sched/deadline: Reset the deadline and the runtime with the parameters
sched/deadline: Release the bandwidth of a task that leaves the class
asleep
sched/deadline: Fix this_bw when a task is changed during a remote
wakeup
sched/debug: Show the deadline bandwidth of each runqueue
selftests/sched: Add a test of SCHED_DEADLINE policy round trips
selftests/sched: Add a test of SCHED_DEADLINE bandwidth accounting

include/linux/sched.h | 5 +
kernel/sched/core.c | 11 +
kernel/sched/deadline.c | 125 +-
kernel/sched/debug.c | 13 +-
kernel/sched/sched.h | 1 +
kernel/sched/syscalls.c | 103 +-
tools/testing/selftests/Makefile | 1 +
.../selftests/sched/deadline/.gitignore | 2 +
.../testing/selftests/sched/deadline/Makefile | 8 +
tools/testing/selftests/sched/deadline/config | 5 +
.../sched/deadline/deadline_bw_test.c | 5468 +++++++++++++++++
.../sched/deadline/deadline_switch_test.c | 1115 ++++
.../testing/selftests/sched/deadline/settings | 1 +
13 files changed, 6829 insertions(+), 29 deletions(-)
create mode 100644 tools/testing/selftests/sched/deadline/.gitignore
create mode 100644 tools/testing/selftests/sched/deadline/Makefile
create mode 100644 tools/testing/selftests/sched/deadline/config
create mode 100644 tools/testing/selftests/sched/deadline/deadline_bw_test.c
create mode 100644 tools/testing/selftests/sched/deadline/deadline_switch_test.c
create mode 100644 tools/testing/selftests/sched/deadline/settings


base-commit: ec6f6731538e0d45872eb279fdeb477bed152d6f
--
2.47.3