On Mon, 22 Feb 2010, Miao Xie wrote:
> >>> guarantee_online_cpus() truly does require callback_mutex, the
> >>> cgroup_scan_tasks() iterator locking can protect changes in the cgroup
> >>> hierarchy but it doesn't protect a store to cs->cpus_allowed or for
> >>> hotplug.
> >>
> >> Right, but the callback_mutex was being removed by this patch.
> >>
> >
> > I was making the case for it to be readded :)
>
> But cgroup_mutex is held when someone changes cs->cpus_allowed or doing hotplug,
> so I think callback_mutex is not necessary in this case.
>
Then why is it taken in update_cpumask()?
> I think this patch can't fix this bug, because mems_allowed of tasks in the
> top group is set to node_possible_map by default, not when the task is
> attached.
>
Ok, I thought that all tasks get their ->attach() function called whenever
their cgroup is mounted.
> I made a new patch at the end of this email to fix it, but I have no machine
> to test it now. who can test it for me.
>
> ---
> diff --git a/init/main.c b/init/main.c
> index 4cb47a1..512ba15 100644
> --- a/init/main.c
> +++ b/init/main.c
> @@ -846,7 +846,7 @@ static int __init kernel_init(void * unused)
> /*
> * init can allocate pages on any node
> */
> - set_mems_allowed(node_possible_map);
> + set_mems_allowed(node_states[N_HIGH_MEMORY]);
> /*
> * init can run on any cpu.
> */
> diff --git a/kernel/cpuset.c b/kernel/cpuset.c
> index ba401fa..e29b440 100644
> --- a/kernel/cpuset.c
> +++ b/kernel/cpuset.c
> @@ -935,10 +935,12 @@ static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
> struct task_struct *tsk = current;
>
> tsk->mems_allowed = *to;
> + wmb();
>
> do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
>
> guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
> + wmb();
> }
>
> /*
> @@ -1391,11 +1393,10 @@ static void cpuset_attach(struct cgroup_subsys *ss, struct cgroup *cont,
>
> if (cs == &top_cpuset) {
> cpumask_copy(cpus_attach, cpu_possible_mask);
> - to = node_possible_map;
> } else {
> guarantee_online_cpus(cs, cpus_attach);
> - guarantee_online_mems(cs, &to);
> }
> + guarantee_online_mems(cs, &to);
>
> /* do per-task migration stuff possibly for each in the threadgroup */
> cpuset_attach_task(tsk, &to, cs);
Do we need to set cpus_attach to cpu_possible_mask? Why won't
cpu_active_mask suffice?
> @@ -2090,15 +2091,19 @@ static int cpuset_track_online_cpus(struct notifier_block *unused_nb,
> static int cpuset_track_online_nodes(struct notifier_block *self,
> unsigned long action, void *arg)
> {
> + nodemask_t oldmems;
Is it possible to use NODEMASK_ALLOC() instead?
> +
> cgroup_lock();
> switch (action) {
> case MEM_ONLINE:
> - case MEM_OFFLINE:
> + oldmems = top_cpuset.mems_allowed;
> mutex_lock(&callback_mutex);
> top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
> mutex_unlock(&callback_mutex);
> - if (action == MEM_OFFLINE)
> - scan_for_empty_cpusets(&top_cpuset);
> + update_tasks_nodemask(&top_cpuset, &oldmems, NULL);
> + break;
> + case MEM_OFFLINE:
> + scan_for_empty_cpusets(&top_cpuset);
> break;
> default:
> break;
Looks good.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>