Why does connector use a work queue???

Why does connector use a work queue???

[Eric W. Biederman]

These days netlink callbacks for messages happen in the context of the
process who sent the message.  Things like permission checks are much
more complicated if we don't use that process context.

I was looking at removing NETLINK_CB(skb).eff_cap but I discovered
that connected takes the netlink messages, them from their
perfectly good process context, and puts the into a workqueue
for reasons that are not apparent to me.

Unless I am misreading something we should just be able to remove the
work queues and greatly simplify the connector code.

Something like:

diff --git a/drivers/connector/connector.c b/drivers/connector/connector.c
index 537c29a..4c9af0a 100644
--- a/drivers/connector/connector.c
+++ b/drivers/connector/connector.c
@@ -120,52 +120,28 @@ EXPORT_SYMBOL_GPL(cn_netlink_send);
  */
 static int cn_call_callback(struct sk_buff *skb)
 {
-	struct cn_callback_entry *__cbq, *__new_cbq;
+	struct cn_callback_entry *__cbq;
 	struct cn_dev *dev = &cdev;
 	struct cn_msg *msg = NLMSG_DATA(nlmsg_hdr(skb));
 	int err = -ENODEV;
+	void (*callback)(struct cn_msg *req, struct netlink_skb_parms *nsp);
 
+	callback = NULL;
 	spin_lock_bh(&dev->cbdev->queue_lock);
 	list_for_each_entry(__cbq, &dev->cbdev->queue_list, callback_entry) {
 		if (cn_cb_equal(&__cbq->id.id, &msg->id)) {
-			if (likely(!work_pending(&__cbq->work) &&
-					__cbq->data.skb == NULL)) {
-				__cbq->data.skb = skb;
-
-				if (queue_cn_work(__cbq, &__cbq->work))
-					err = 0;
-				else
-					err = -EINVAL;
-			} else {
-				struct cn_callback_data *d;
-
-				err = -ENOMEM;
-				__new_cbq = kzalloc(sizeof(struct cn_callback_entry), GFP_ATOMIC);
-				if (__new_cbq) {
-					d = &__new_cbq->data;
-					d->skb = skb;
-					d->callback = __cbq->data.callback;
-					d->free = __new_cbq;
-
-					__new_cbq->pdev = __cbq->pdev;
-
-					INIT_WORK(&__new_cbq->work,
-							&cn_queue_wrapper);
-
-					if (queue_cn_work(__new_cbq,
-						    &__new_cbq->work))
-						err = 0;
-					else {
-						kfree(__new_cbq);
-						err = -EINVAL;
-					}
-				}
-			}
+			callback = __cbq->data.callback;
+			err = 0;
 			break;
 		}
 	}
 	spin_unlock_bh(&dev->cbdev->queue_lock);
 
+	if (!err) {
+		callback(msg, &NETLINK_CB(skb));
+		kfree_skb(skb);
+	}
+
 	return err;
 }
 

Re: Why does connector use a work queue???

[Evgeniy Polyakov]

On Sat, Feb 27, 2010 at 06:57:02PM -0800, Eric W. Biederman (ebiederm@xmission.com) wrote:
> These days netlink callbacks for messages happen in the context of the
> process who sent the message.  Things like permission checks are much
> more complicated if we don't use that process context.
> 
> I was looking at removing NETLINK_CB(skb).eff_cap but I discovered
> that connected takes the netlink messages, them from their
> perfectly good process context, and puts the into a workqueue
> for reasons that are not apparent to me.

Netlink was made synchronous rather recently and connector was not
changed to take advantage of that. Previously it used work queue to
postpone work processing into always-process context.

> Unless I am misreading something we should just be able to remove the
> work queues and greatly simplify the connector code.
> 
> Something like:

Idea looks very good, thank you, but it misses structure changes
to eliminate now unneded members.

-- 
	Evgeniy Polyakov

Re: Why does connector use a work queue???

[Eric W. Biederman]

Evgeniy Polyakov <zbr@ioremap.net> writes:

> On Sat, Feb 27, 2010 at 06:57:02PM -0800, Eric W. Biederman (ebiederm@xmission.com) wrote:
>> These days netlink callbacks for messages happen in the context of the
>> process who sent the message.  Things like permission checks are much
>> more complicated if we don't use that process context.
>> 
>> I was looking at removing NETLINK_CB(skb).eff_cap but I discovered
>> that connected takes the netlink messages, them from their
>> perfectly good process context, and puts the into a workqueue
>> for reasons that are not apparent to me.
>
> Netlink was made synchronous rather recently and connector was not
> changed to take advantage of that. Previously it used work queue to
> postpone work processing into always-process context.

I suspected it was something like that but I want to be certain.

>> Unless I am misreading something we should just be able to remove the
>> work queues and greatly simplify the connector code.
>> 
>> Something like:
>
> Idea looks very good, thank you, but it misses structure changes
> to eliminate now unneded members.

Right. I didn't trust myself to write a patch that would rip out half
of the connector code until I understood what the constraints were.

Eric