> On 4 Nov 2014, at 09:50, Alexander Bluhm wrote:
>
> On Thu, Oct 30, 2014 at 09:55:35PM -0400, Ted Unangst wrote:
>> On Thu, Oct 30, 2014 at 22:10, Alexander Bluhm wrote:
>>> +
>>> + /* Avoid user land starvation. */
>>> + yield();
>>
>> I think this is the responsibility of the taskq thre
On Thu, Oct 30, 2014 at 09:55:35PM -0400, Ted Unangst wrote:
> On Thu, Oct 30, 2014 at 22:10, Alexander Bluhm wrote:
> > +
> > + /* Avoid user land starvation. */
> > + yield();
>
> I think this is the responsibility of the taskq thread, not the
> individual task.
I am not sure about this. W
> On 31 Oct 2014, at 22:37, Alexander Bluhm wrote:
>
> On Fri, Oct 31, 2014 at 02:50:00PM +1000, David Gwynne wrote:
>> so without splicing, the payloads from multiple tcp packets (at least all of
>> the ones in a single softnet run?) get bundled up into a buffer that
>> userland reads and the
On Fri, Oct 31, 2014 at 02:50:00PM +1000, David Gwynne wrote:
> so without splicing, the payloads from multiple tcp packets (at least all of
> the ones in a single softnet run?) get bundled up into a buffer that userland
> reads and then writes out again in a single go. right?
>
> you're suggest
> On 31 Oct 2014, at 07:10, Alexander Bluhm wrote:
>
> Hi,
>
> Some performance measurements showed that socket splicing for TCP
> can be made faster. The main slowdown was that tcp_output() got
> called for every incomming packet. When copying through user-land
> this cannot happen as the sc
On Thu, Oct 30, 2014 at 22:10, Alexander Bluhm wrote:
> +
> + /* Avoid user land starvation. */
> + yield();
I think this is the responsibility of the taskq thread, not the
individual task.
Regarding the second pool diff, that looks very reasonable. Quite a
savings for a feature many mach
Hi,
Some performance measurements showed that socket splicing for TCP
can be made faster. The main slowdown was that tcp_output() got
called for every incomming packet. When copying through user-land
this cannot happen as the scheduler gets involved.
So my idea is to do the socket splicing for
