Re: [PATCH net-next v2 9/9] xen-netback: Aggregate TX unmap operations

From: Zoltan Kiss
Date: Mon Dec 16 2013 - 11:30:16 EST


On 13/12/13 15:44, Wei Liu wrote:
On Thu, Dec 12, 2013 at 11:48:17PM +0000, Zoltan Kiss wrote:
Unmapping causes TLB flushing, therefore we should make it in the largest
possible batches. However we shouldn't starve the guest for too long. So if
the guest has space for at least two big packets and we don't have at least a
quarter ring to unmap, delay it for at most 1 milisec.


Is this solution temporary or permanent? If it is permanent would it
make sense to make these parameter tunable?

Well, I'm not entirely sure yet this is the best way to do this, so in this sense it's temporary. But generally we should do some sort of batching, as TLB flush cannot be avoided every time. If we settle on something we should make the tunable parameters tunable.
The problem is that it is a thin red line we should find here. My first approach was that I left the tx_dealloc_work_todo as it was, and after the thread woke up but before anything were done I made it sleep for 50 ns and measured how fast the guest is running out of free slots:

if (kthread_should_stop())
break;

+i=0;
+do {
+ ++i;
+ prev_free_slots = nr_free_slots(&vif->tx);
+ __set_current_state(TASK_UNINTERRUPTIBLE);
+ rc = schedule_hrtimeout_range(&tx_dealloc_delay_ktime, 10, HRTIMER_MODE_REL);
+ if (rc) trace_printk("%s sleep were interrupted! %d\n",vif->dev->name, rc);
+ curr_free_slots = nr_free_slots(&vif->tx);
+} while ( (curr_free_slots < 4 * (prev_free_slots - curr_free_slots)) && i < 11);
+
xenvif_tx_dealloc_action(vif);

And worst case after 500 ns I let the thread to do the unmap anyway. But I was a bit worried about this approach, so I choose a bit more conservative one for this patch.

There are also ideas to use some other instrument for unmapping instead of the current separate thread approach. Putting it into the NAPI instance was the original idea, which caused problems. Placing it into the another thread where RX work happens also doesn't sound too good, these things can and should happen in parallel.
Other ideas were work queues and tasklets, I'll spend some more time to check if they are feasible.

Regards,

Zoli
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/