Re: [Qemu-devel] [PATCH v2 15/15] net: invoke qemu_can_send_packet only before net queue sending function

[Zhi Yong Wu]

On Thu, May 24, 2012 at 6:07 PM, Paolo Bonzini <address@hidden> wrote:
> Il 24/05/2012 06:05, Zhi Yong Wu ha scritto:
>> On Thu, May 24, 2012 at 12:00 AM, Paolo Bonzini <address@hidden> wrote:
>>> Il 23/05/2012 17:14, address@hidden ha scritto:
>>>> From: Zhi Yong Wu <address@hidden>
>>>>
>>>> Signed-off-by: Zhi Yong Wu <address@hidden>
>>>> ---
>>>>  net/queue.c      |    4 ++--
>>>>  net/slirp.c      |    7 -------
>>>>  net/tap.c        |    2 +-
>>>>  slirp/if.c       |    5 -----
>>>>  slirp/libslirp.h |    1 -
>>>>  5 files changed, 3 insertions(+), 16 deletions(-)
>>>>
>>>> diff --git a/net/queue.c b/net/queue.c
>>>> index 0afd783..d2e57de 100644
>>>> --- a/net/queue.c
>>>> +++ b/net/queue.c
>>>> @@ -176,7 +176,7 @@ ssize_t qemu_net_queue_send(NetQueue *queue,
>>>>  {
>>>>      ssize_t ret;
>>>>
>>>> -    if (queue->delivering) {
>>>> +    if (queue->delivering || !qemu_can_send_packet(sender)) {
>>>>          return qemu_net_queue_append(queue, sender, flags, data, size, 
>>>> NULL);
>>>>      }
>>>>
>>>> @@ -200,7 +200,7 @@ ssize_t qemu_net_queue_send_iov(NetQueue *queue,
>>>>  {
>>>>      ssize_t ret;
>>>>
>>>> -    if (queue->delivering) {
>>>> +    if (queue->delivering || !qemu_can_send_packet(sender)) {
>>>>          return qemu_net_queue_append_iov(queue, sender, flags, iov, 
>>>> iovcnt, NULL);
>>>>      }
>>>>
>>>> diff --git a/net/slirp.c b/net/slirp.c
>>>> index a6ede2b..248f7ff 100644
>>>> --- a/net/slirp.c
>>>> +++ b/net/slirp.c
>>>> @@ -96,13 +96,6 @@ static void slirp_smb_cleanup(SlirpState *s);
>>>>  static inline void slirp_smb_cleanup(SlirpState *s) { }
>>>>  #endif
>>>>
>>>> -int slirp_can_output(void *opaque)
>>>> -{
>>>> -    SlirpState *s = opaque;
>>>> -
>>>> -    return qemu_can_send_packet(&s->nc);
>>>> -}
>>>> -
>>>>  void slirp_output(void *opaque, const uint8_t *pkt, int pkt_len)
>>>>  {
>>>>      SlirpState *s = opaque;
>>>> diff --git a/net/tap.c b/net/tap.c
>>>> index 65f45b8..7b1992b 100644
>>>> --- a/net/tap.c
>>>> +++ b/net/tap.c
>>>> @@ -210,7 +210,7 @@ static void tap_send(void *opaque)
>>>>          if (size == 0) {
>>>>              tap_read_poll(s, 0);
>>>>          }
>>>> -    } while (size > 0 && qemu_can_send_packet(&s->nc));
>>>> +    } while (size > 0);
>>>
>>> Can you explain this?  Also, have you benchmarked the change to see what
>> Its code execution flow is like below:
>> tap_send --> qemu_send_packet_async
>> ->qemu_send_packet_async_with_flags ->qemu_net_queue_send
>>
>> So it will finally invoke qemu_can_send_packet to determine if it can
>> send packets. this code change delay this determination.
>
> But you will copy packets uselessly.  The code before the patch simply
> left them on the tap file descriptor.  This is better because it
> involves the kernel in flow control.  You are introducing "bufferbloat".
You are correct, but can_send_packet will be invoked twice for one
packet delivery.
>
>>> Also, can you explain why you didn't implement this?
>> Hub can now do its own flow control if it provides its can_recieve.
>
> But you didn't add can_receive.
>
>> Why need we add some counts to track in-flight packets?
>
> To implement can_receive.
Let me try.
>
> Paolo



-- 
Regards,

Zhi Yong Wu