* af_packet: don't enable global timestamps
@ 2007-09-04 5:35 Stephen Hemminger
2007-09-04 7:59 ` Eric Dumazet
0 siblings, 1 reply; 3+ messages in thread
From: Stephen Hemminger @ 2007-09-04 5:35 UTC (permalink / raw)
To: David S. Miller, Andi Kleen; +Cc: netdev
Andi mentioned he did something like this already, but never
submitted it.
The dhcp client application uses AF_PACKET with a packet filter to
receive data. The application doesn't even use timestamps, but because
the AF_PACKET API has timestamps, they get turned on globally which
causes an expensive time of day lookup for every packet received
on any system that uses the standard DHCP client.
The fix is to not enable the timestamp (but use if if available).
This causes the time lookup to only occur on those packets
that are destined for the AF_PACKET socket.
The timestamping occurs after packet filtering
so all packets dropped by filtering to not cause a clock call.
The one downside of this a a few microseconds additional delay
added from the normal timestamping location (netif_rx) until the
receive callback in AF_PACKET. But since the offset is fairly consistent
it should not upset applications that do want really use timestamps,
like wireshark.
Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
--- a/net/packet/af_packet.c 2007-07-23 09:31:26.000000000 +0100
+++ b/net/packet/af_packet.c 2007-09-03 14:55:00.000000000 +0100
@@ -640,11 +640,10 @@ static int tpacket_rcv(struct sk_buff *s
h->tp_snaplen = snaplen;
h->tp_mac = macoff;
h->tp_net = netoff;
- if (skb->tstamp.tv64 == 0) {
- __net_timestamp(skb);
- sock_enable_timestamp(sk);
- }
- tv = ktime_to_timeval(skb->tstamp);
+ if (skb->tstamp.tv64)
+ tv = ktime_to_timeval(skb->tstamp);
+ else
+ do_gettimeofday(&tv);
h->tp_sec = tv.tv_sec;
h->tp_usec = tv.tv_usec;
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: af_packet: don't enable global timestamps
2007-09-04 5:35 af_packet: don't enable global timestamps Stephen Hemminger
@ 2007-09-04 7:59 ` Eric Dumazet
2007-09-06 12:56 ` David Miller
0 siblings, 1 reply; 3+ messages in thread
From: Eric Dumazet @ 2007-09-04 7:59 UTC (permalink / raw)
To: Stephen Hemminger; +Cc: David S. Miller, Andi Kleen, netdev
On Tue, 4 Sep 2007 06:35:25 +0100
Stephen Hemminger <shemminger@linux-foundation.org> wrote:
> Andi mentioned he did something like this already, but never
> submitted it.
>
> The dhcp client application uses AF_PACKET with a packet filter to
> receive data. The application doesn't even use timestamps, but because
> the AF_PACKET API has timestamps, they get turned on globally which
> causes an expensive time of day lookup for every packet received
> on any system that uses the standard DHCP client.
>
> The fix is to not enable the timestamp (but use if if available).
> This causes the time lookup to only occur on those packets
> that are destined for the AF_PACKET socket.
> The timestamping occurs after packet filtering
> so all packets dropped by filtering to not cause a clock call.
>
> The one downside of this a a few microseconds additional delay
> added from the normal timestamping location (netif_rx) until the
> receive callback in AF_PACKET. But since the offset is fairly consistent
> it should not upset applications that do want really use timestamps,
> like wireshark.
This patch seems the correct fix for this longstanding problem.
Please note that if wireshark/tcpdump processes really want precise timestamps,
they still can ask this by using setsockopt(SO_TIMESTAMP or SO_TIMESTAMPNS) on their private socket.
(We already know that running a sniffer has a cost anyway)
>
> Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org>
>
>
> --- a/net/packet/af_packet.c 2007-07-23 09:31:26.000000000 +0100
> +++ b/net/packet/af_packet.c 2007-09-03 14:55:00.000000000 +0100
> @@ -640,11 +640,10 @@ static int tpacket_rcv(struct sk_buff *s
> h->tp_snaplen = snaplen;
> h->tp_mac = macoff;
> h->tp_net = netoff;
> - if (skb->tstamp.tv64 == 0) {
> - __net_timestamp(skb);
> - sock_enable_timestamp(sk);
> - }
> - tv = ktime_to_timeval(skb->tstamp);
> + if (skb->tstamp.tv64)
> + tv = ktime_to_timeval(skb->tstamp);
> + else
> + do_gettimeofday(&tv);
> h->tp_sec = tv.tv_sec;
> h->tp_usec = tv.tv_usec;
>
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: af_packet: don't enable global timestamps
2007-09-04 7:59 ` Eric Dumazet
@ 2007-09-06 12:56 ` David Miller
0 siblings, 0 replies; 3+ messages in thread
From: David Miller @ 2007-09-06 12:56 UTC (permalink / raw)
To: dada1; +Cc: shemminger, ak, netdev
From: Eric Dumazet <dada1@cosmosbay.com>
Date: Tue, 4 Sep 2007 09:59:58 +0200
> This patch seems the correct fix for this longstanding problem.
>
> Please note that if wireshark/tcpdump processes really want precise timestamps,
> they still can ask this by using setsockopt(SO_TIMESTAMP or SO_TIMESTAMPNS) on their private socket.
>
> (We already know that running a sniffer has a cost anyway)
Agreed, let's give this a try.
I've applied Stephen's patch to net-2.6.24
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2007-09-06 12:57 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2007-09-04 5:35 af_packet: don't enable global timestamps Stephen Hemminger
2007-09-04 7:59 ` Eric Dumazet
2007-09-06 12:56 ` David Miller
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).