]> git.ipfire.org Git - thirdparty/linux.git/commit
sched_ext: Use READ_ONCE() for the read side of dsq->nr update
authorzhidao su <suzhidao@xiaomi.com>
Mon, 2 Mar 2026 09:14:39 +0000 (17:14 +0800)
committerTejun Heo <tj@kernel.org>
Mon, 2 Mar 2026 17:23:00 +0000 (07:23 -1000)
commit9adfcef334bf9c6ef68eaecfca5f45d18614efe0
treea794e00eaef9d0f53c472505bd6d77d116bae33d
parent032e084f0d43fda78c33abfc704ac13a0891a6e7
sched_ext: Use READ_ONCE() for the read side of dsq->nr update

scx_bpf_dsq_nr_queued() reads dsq->nr via READ_ONCE() without holding
any lock, making dsq->nr a lock-free concurrently accessed variable.
However, dsq_mod_nr(), the sole writer of dsq->nr, only uses
WRITE_ONCE() on the write side without the matching READ_ONCE() on the
read side:

    WRITE_ONCE(dsq->nr, dsq->nr + delta);
                        ^^^^^^^
                        plain read -- KCSAN data race

The KCSAN documentation requires that if one accessor uses READ_ONCE()
or WRITE_ONCE() on a variable to annotate lock-free access, all other
accesses must also use the appropriate accessor. A plain read on the
right-hand side of WRITE_ONCE() leaves the pair incomplete and will
trigger KCSAN warnings.

Fix by using READ_ONCE() for the read side of the update:

    WRITE_ONCE(dsq->nr, READ_ONCE(dsq->nr) + delta);

This is consistent with scx_bpf_dsq_nr_queued() and makes the
concurrent access annotation complete and KCSAN-clean.

Signed-off-by: zhidao su <suzhidao@xiaomi.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
kernel/sched/ext.c