If freezing of a cgroup races with waking of a task from the frozen state (like waiting in vfork() or in do_signal_stop()), a spurious transition of the cgroup state can happen. The task enters cgroup_leave_frozen(true), the cgroup->nr_frozen_tasks counter decrements, and the cgroup is switched to the unfrozen state. To prevent it, let's reserve cgroup_leave_frozen(true) for terminating processes and use cgroup_leave_frozen(false) otherwise. To avoid busy-looping in the signal handling loop waiting for JOBCTL_TRAP_FREEZE set from the cgroup freezing path, let's do it explicitly in cgroup_leave_frozen(), if the task is going to stay frozen. Suggested-by: Oleg Nesterov <oleg@xxxxxxxxxx> Signed-off-by: Roman Gushchin <guro@xxxxxx> --- kernel/cgroup/freezer.c | 16 +++++----------- kernel/signal.c | 2 +- 2 files changed, 6 insertions(+), 12 deletions(-) diff --git a/kernel/cgroup/freezer.c b/kernel/cgroup/freezer.c index 3bfbb3c8baf3..c321e768f8d3 100644 --- a/kernel/cgroup/freezer.c +++ b/kernel/cgroup/freezer.c @@ -139,19 +139,13 @@ void cgroup_leave_frozen(bool always_leave) cgroup_update_frozen(cgrp); WARN_ON_ONCE(!current->frozen); current->frozen = false; + } else if (!(current->jobctl & JOBCTL_TRAP_FREEZE)) { + spin_lock(¤t->sighand->siglock); + current->jobctl |= JOBCTL_TRAP_FREEZE; + set_thread_flag(TIF_SIGPENDING); + spin_unlock(¤t->sighand->siglock); } spin_unlock_irq(&css_set_lock); - - if (unlikely(current->frozen)) { - /* - * If the task remained in the frozen state, - * make sure it won't reach userspace without - * entering the signal handling loop. - */ - spin_lock_irq(¤t->sighand->siglock); - recalc_sigpending(); - spin_unlock_irq(¤t->sighand->siglock); - } } /* diff --git a/kernel/signal.c b/kernel/signal.c index 095e0fc57b25..16b72f4f14df 100644 --- a/kernel/signal.c +++ b/kernel/signal.c @@ -2514,7 +2514,7 @@ bool get_signal(struct ksignal *ksig) */ if (unlikely(cgroup_task_frozen(current))) { spin_unlock_irq(&sighand->siglock); - cgroup_leave_frozen(true); + cgroup_leave_frozen(false); goto relock; } -- 2.20.1