The submit queue polling threads are userland threads that just never exit to the userland. In case the creating task is part of a cgroup with the cpuset controller enabled, the poller should also stay within that cpuset. This also holds, as the poller belongs to the same cgroup as the task that started it.

With the current implementation, a process can "break out" of the defined cpuset by creating sq pollers consuming CPU time on other CPUs, which is especially problematic for realtime applications.

Part of this problem was fixed in a5fc1441 by dropping the PF_NO_SETAFFINITY flag, but this only becomes effective after the first modification of the cpuset (i.e. the pollers cpuset is correct after the first update of the enclosing cgroups cpuset).

By inheriting the cpuset of the creating tasks, we ensure that the poller is created with a cpumask that is a subset of the cgroups mask. Inheriting the creators cpumask is reasonable, as other userland tasks also inherit the mask.

Fixes: 37d1e2e3642e ("io_uring: move SQPOLL thread io-wq forked worker") Cc: stable@vger.kernel.org # 6.1+ Signed-off-by: Felix Moessbauer felix.moessbauer@siemens.com --- Changes since v1:

- do not set poller thread cpuset in non-pinning case, as the default is already correct (the mask is inherited from the parent). - Remove incorrect term "kernel thread" from the commit message

I tested this without pinning, explicit pinning of the parent task and non-all cgroup cpusets (and all combinations).

Best regards, Felix Moessbauer Siemens AG

io_uring/sqpoll.c | 2 -- 1 file changed, 2 deletions(-)

diff --git a/io_uring/sqpoll.c b/io_uring/sqpoll.c index 6ea21b503113..5a002fa1d953 100644 --- a/io_uring/sqpoll.c +++ b/io_uring/sqpoll.c @@ -231,8 +231,6 @@ static int io_sq_thread(void *data)

if (sqd->sq_cpu != -1) set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu)); - else - set_cpus_allowed_ptr(current, cpu_online_mask);

/* * Force audit context to get setup, in case we do prep side async