[PATCHSET sched_ext/for-7.1-fixes] sched_ext: Fix sched_ext_dead() races with task initialization

[Tejun Heo]

Hello,

zhidao su reported a NULL deref and an ops.init_task() leak when
sched_ext_dead() races scx_root_enable_workfn() in CONFIG_EXT_SUB_SCHED
kernels [1]. The same race window also affects the analogous sub-sched paths
(scx_sub_enable_workfn()'s per-task init pass and scx_sub_disable()'s
migration loop), and the wrapper-disable paths trip on the NONE state that
scx_fail_parent() leaves behind. Closing all of these calls for a
state-machine extension rather than a localized fix.

The series introduces SCX_TASK_INIT_BEGIN as an explicit intermediate state
between NONE and INIT, and replaces the SCX_TASK_OFF_TASKS marker flag with
a real SCX_TASK_DEAD terminal state. With the state machine in place, every
init path uses the same handshake: write INIT_BEGIN under rq lock, init
outside the lock, recheck DEAD under rq lock, unwind via
scx_sub_init_cancel_task() on hit. The wrapper-disable and
switched_from_scx() paths get NONE early-returns to handle the
scx_fail_parent() residue.

It is more invasive than zhidao's patches but covers the related races
uniformly and avoids the implicit list_empty() check his approach relies
on. Credit to him for finding and reporting the bug.

 0001 sched_ext: Cleanups in preparation for the SCX_TASK_INIT_BEGIN/DEAD work
 0002 sched_ext: Inline scx_init_task() and move RESET_RUNNABLE_AT into scx_set_task_state()
 0003 sched_ext: Replace SCX_TASK_OFF_TASKS flag with SCX_TASK_DEAD state
 0004 sched_ext: Close root-enable vs sched_ext_dead() race with SCX_TASK_INIT_BEGIN
 0005 sched_ext: Close sub-sched init race with post-init DEAD recheck
 0006 sched_ext: Handle SCX_TASK_NONE in disable/switched_from paths

Based on sched_ext/for-7.1-fixes (ab28a0673daa).

Git tree: git://git.kernel.org/pub/scm/linux/kernel/git/tj/sched_ext.git for-7.1-fixes-dead-race

Verified with a debug patch that widens the unlocked init windows on the
root and sub-sched paths and counts post-init DEAD-recheck hits.
Reproducers exercise each of the original races plus the scx_fail_parent
NONE-state regression, followed by a multi-iteration stress under fork
churn. Counters show the windows are hit and no
BUG/WARNING/Oops/Invalid-task-state appears.

[1] https://lore.kernel.org/all/20260429133155.3825247-1-suzhidao@xxxxxxxxxx/

 include/linux/sched/ext.h |  17 ++--
 kernel/sched/ext.c        | 221 +++++++++++++++++++++++++++++++---------------
 2 files changed, 162 insertions(+), 76 deletions(-)

Thanks.

--
tejun