Le jeudi 09 décembre 2010 à 16:44 -0800, Andrew Morton a écrit :

The name "kthread_create_on_cpu" is pretty misleading.

One would expect such a function to create a kthread which is bound to
that CPU.  But what it in fact does is to create a kthread which is
bound to all CPUs and whose stack, task_struct and thread_info were
allocated from the node which contains `cpu'.

Also, a saner interface would be one which takes the numa_node_id, not
the cpu number.

quoted

...

 /**
- * kthread_create - create a kthread.
+ * kthread_create_on_cpu - create a kthread.
  * @threadfn: the function to run until signal_pending(current).
  * @data: data ptr for @threadfn.
+ * @cpu: cpu number.
  * @namefmt: printf-style name for the thread.
  *
  * Description: This helper function creates and names a kernel
  * thread.  The thread will be stopped: use wake_up_process() to start
  * it.  See also kthread_run().
  *
+ * If thread is going to be bound on a particular cpu, give its number
+ * in @cpu, to get NUMA affinity for kthread stack, or else give -1.

This is a bit presumptuous.  The caller might wish to later bind this
thread to some or all of the CPUs on the node, rather than to a single
CPU (eg, kswapd()).


So what to do?  Maybe add a new kthread_create_node() which prepares a
kthread whose memory is bound to that node, then add a
kthread_create_cpu() convenience wrapper around that?

We probably can add the "bind to cpu" as a fifth patch, to avoid one 
kthread_bind(p, cpu);  done by these callers.

My reasoning not including this kthread_bind(p, cpu) in initial patch
series that I was focusing on NUMA properties first, not on scheduling
(this part already runs correctly as far as I know)

Thanks for taking the patch series, I was about to resubmit it today :)