Messages in this thread | | | Date | Sun, 1 Feb 2015 14:26:11 +0200 | From | "Michael S. Tsirkin" <> | Subject | Re: [PATCH] tun: orphan an skb on tx |
| |
On Sun, Feb 01, 2015 at 11:20:33AM +0000, David Woodhouse wrote: > On Wed, 2010-04-14 at 08:58 +0800, Herbert Xu wrote: > > On Tue, Apr 13, 2010 at 08:31:03PM +0200, Eric Dumazet wrote: > > > > > > Herbert Acked your patch, so I guess its OK, but I think it can be > > > dangerous. > > > > The tun socket accounting was never designed to stop it from > > flooding another tun interface. It's there to stop it from > > transmitting above a destination interface TX bandwidth and > > cause unnecessary packet drops. It also limits the total amount > > of kernel memory that can be pinned down by a single tun interface. > > > > In this case, all we're doing is shifting the accounting from the > > "hardware" queue to the qdisc queue. > > > > So your ability to flood a tun interface is essentially unchanged. > > I've just been looking at VPN performance, using netperf to flood an > openconnect/ocserv connection over GigE and profiling my VPN client. > > If I run netperf over the *unencrypted* link, it only sends 1Gb/s of > packets — because the packets are correctly accounted to netperf's UDP > socket until the moment they're actually transmitted on the wire, and > the backpressure works correctly. > > When I run over the VPN, netperf thinks it sent 2½ times the amount of > TX traffic.
At some level, it's expected: netperf's manual actually says: A UDP_STREAM test has no end-to-end flow control - UDP provides none and neither does netperf. However, if you wish, you can configure netperf with --enable-intervals=yes to enable the global command-line -b and -w options to pace bursts of traffic onto the network.
> Packets are being dropped by the tun device before even > feeding them up to the VPN client to be sent — presumably because of > this skb_orphan() call. (The client itself should do the right thing, > and only suck packets out of the tun at the rate it can shove them out > *its* UDP socket.)
A simple work-around is to limit the rate using a non work conservig qdisc.
> Did we ever look at the alternative solution of taking ownership only > after a timeout, or on demand when we need to shut down the device?
I've been thinking about this on and off, but didn't find a good safe solution yet.
For timeout, the difficulty is to find a good timer value, low enough to avoid DOS attacks but high enough to avoid spurious packet drops (and expensive timer interrupts).
-- MST
| |