When cpu_populated_map was introduced, it was supposed that cwq->thread can
survive after CPU_DEAD, that is why we never shrink cpu_populated_map.
This is not very nice, we can safely remove the already dead CPU from the map.
The only required change is that destroy_workqueue() must hold the hotplug lock
until it destroys all cwq->thread's, to protect the cpu_populated_map. We could
make the local copy of cpu mask and drop the lock, but sizeof(cpumask_t) may be
very large.
Also, fix the comment near queue_work(). Unless _cpu_down() happens we do
guarantee the cpu-affinity of the work_struct, and we have users which rely on
this.
Signed-off-by: Oleg Nesterov <[email protected]>
--- 25/kernel/workqueue.c~2_WQ_CPU_MAP 2008-02-15 16:59:18.000000000 +0300
+++ 25/kernel/workqueue.c 2008-02-16 19:33:03.000000000 +0300
@@ -158,8 +158,8 @@ static void __queue_work(struct cpu_work
*
* Returns 0 if @work was already on a queue, non-zero otherwise.
*
- * We queue the work to the CPU it was submitted, but there is no
- * guarantee that it will be processed by that CPU.
+ * We queue the work to the CPU it was submitted, but if CPU dies
+ * it can be processed by another CPU.
*/
int queue_work(struct workqueue_struct *wq, struct work_struct *work)
{
@@ -813,12 +813,12 @@ void destroy_workqueue(struct workqueue_
spin_lock(&workqueue_lock);
list_del(&wq->list);
spin_unlock(&workqueue_lock);
- put_online_cpus();
for_each_cpu_mask(cpu, *cpu_map) {
cwq = per_cpu_ptr(wq->cpu_wq, cpu);
cleanup_workqueue_thread(cwq, cpu);
}
+ put_online_cpus();
free_percpu(wq->cpu_wq);
kfree(wq);
@@ -836,7 +836,6 @@ static int __devinit workqueue_cpu_callb
action &= ~CPU_TASKS_FROZEN;
switch (action) {
-
case CPU_UP_PREPARE:
cpu_set(cpu, cpu_populated_map);
}
@@ -864,6 +863,12 @@ static int __devinit workqueue_cpu_callb
}
}
+ switch (action) {
+ case CPU_UP_CANCELED:
+ case CPU_DEAD:
+ cpu_clear(cpu, cpu_populated_map);
+ }
+
return NOTIFY_OK;
}
Hi Oleg,
This patch looks OK to me. But while reading this I got some doubts
in nearby places, so BTW 2 small questions:
1) ... workqueue_cpu_callback(...)
{
...
list_for_each_entry(wq, &workqueues, list) {
cwq = per_cpu_ptr(wq->cpu_wq, cpu);
switch (action) {
case CPU_UP_PREPARE:
...
It looks like not all CPU_ cases are served here: shouldn't
list_for_each_entry() be omitted for them?
2) ... __create_workqueue_key(...)
{
...
if (singlethread) {
...
} else {
get_online_cpus();
spin_lock(&workqueue_lock);
list_add(&wq->list, &workqueues);
Shouldn't this list_add() be done after all these inits below?
spin_unlock(&workqueue_lock);
for_each_possible_cpu(cpu) {
cwq = init_cpu_workqueue(wq, cpu);
...
}
...
Thanks,
Jarek P.
On Sat, Feb 16, 2008 at 08:22:59PM +0300, Oleg Nesterov wrote:
> When cpu_populated_map was introduced, it was supposed that cwq->thread can
> survive after CPU_DEAD, that is why we never shrink cpu_populated_map.
>
> This is not very nice, we can safely remove the already dead CPU from the map.
> The only required change is that destroy_workqueue() must hold the hotplug lock
> until it destroys all cwq->thread's, to protect the cpu_populated_map. We could
> make the local copy of cpu mask and drop the lock, but sizeof(cpumask_t) may be
> very large.
>
> Also, fix the comment near queue_work(). Unless _cpu_down() happens we do
> guarantee the cpu-affinity of the work_struct, and we have users which rely on
> this.
>
> Signed-off-by: Oleg Nesterov <[email protected]>
On 02/17, Jarek Poplawski wrote:
>
> This patch looks OK to me.
Thanks for looking at this!
> But while reading this I got some doubts
> in nearby places, so BTW 2 small questions:
>
> 1) ... workqueue_cpu_callback(...)
> {
> ...
> list_for_each_entry(wq, &workqueues, list) {
> cwq = per_cpu_ptr(wq->cpu_wq, cpu);
>
> switch (action) {
> case CPU_UP_PREPARE:
> ...
>
> It looks like not all CPU_ cases are served here: shouldn't
> list_for_each_entry() be omitted for them?
Yes, but this is harmless. cpu-hotplug callbacks are not time-critical,
and cpu_down/cpu_up happens not often, and LIST_HEAD(workqueues) is not
very long, so ...
> 2) ... __create_workqueue_key(...)
> {
> ...
> if (singlethread) {
> ...
> } else {
> get_online_cpus();
> spin_lock(&workqueue_lock);
> list_add(&wq->list, &workqueues);
>
> Shouldn't this list_add() be done after all these inits below?
>
> spin_unlock(&workqueue_lock);
>
> for_each_possible_cpu(cpu) {
> cwq = init_cpu_workqueue(wq, cpu);
> ...
> }
> ...
This doesn't matter. Please note that get_online_cpus() blocks
cpu_up/cpu_down, they take cpu_hotplug_begin().
Oleg.
On Mon, Feb 18, 2008 at 02:45:56AM +0300, Oleg Nesterov wrote:
> On 02/17, Jarek Poplawski wrote:
...
> > 1) ... workqueue_cpu_callback(...)
...
> Yes, but this is harmless. cpu-hotplug callbacks are not time-critical,
> and cpu_down/cpu_up happens not often, and LIST_HEAD(workqueues) is not
> very long, so ...
>
> > 2) ... __create_workqueue_key(...)
...
> > Shouldn't this list_add() be done after all these inits below?
> This doesn't matter. Please note that get_online_cpus() blocks
> cpu_up/cpu_down, they take cpu_hotplug_begin().
You are completely right. It looks like this was only about ..."look".
(But adding only "default:" in 1) would make it look nicer to me...)
Regards,
Jarek P.