On Thu, Nov 30, 2017 at 10:46:17AM +0800, Jason Wang wrote:
>
>
> On 2017年11月29日 23:31, Michael S. Tsirkin wrote:
> > On Wed, Nov 29, 2017 at 09:23:24AM -0500,[email protected] wrote:
> > > From: Wei Xu<[email protected]>
> > >
> > > Matthew found a roughly 40% tcp throughput regression with commit
> > > c67df11f(vhost_net: try batch dequing from skb array) as discussed
> > > in the following thread:
> > > https://www.mail-archive.com/[email protected]/msg187936.html
> > >
> > > Eventually we figured out that it was a skb leak in handle_rx()
> > > when sending packets to the VM. This usually happens when a guest
> > > can not drain out vq as fast as vhost fills in, afterwards it sets
> > > off the traffic jam and leaks skb(s) which occurs as no headcount
> > > to send on the vq from vhost side.
> > >
> > > This can be avoided by making sure we have got enough headcount
> > > before actually consuming a skb from the batched rx array while
> > > transmitting, which is simply done by moving checking the zero
> > > headcount a bit ahead.
> > >
> > > Also strengthen the small possibility of leak in case of recvmsg()
> > > fails by freeing the skb.
> > >
> > > Signed-off-by: Wei Xu<[email protected]>
> > > Reported-by: Matthew Rosato<[email protected]>
> > > ---
> > > drivers/vhost/net.c | 23 +++++++++++++----------
> > > 1 file changed, 13 insertions(+), 10 deletions(-)
> > >
> > > v2:
> > > - add Matthew as the reporter, thanks matthew.
> > > - moving zero headcount check ahead instead of defer consuming skb
> > > due to jason and mst's comment.
> > > - add freeing skb in favor of recvmsg() fails.
> > >
> > > diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
> > > index 8d626d7..e302e08 100644
> > > --- a/drivers/vhost/net.c
> > > +++ b/drivers/vhost/net.c
> > > @@ -778,16 +778,6 @@ static void handle_rx(struct vhost_net *net)
> > > /* On error, stop handling until the next kick. */
> > > if (unlikely(headcount < 0))
> > > goto out;
> > > - if (nvq->rx_array)
> > > - msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
> > > - /* On overrun, truncate and discard */
> > > - if (unlikely(headcount > UIO_MAXIOV)) {
> > > - iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
> > > - err = sock->ops->recvmsg(sock, &msg,
> > > - 1, MSG_DONTWAIT | MSG_TRUNC);
> > > - pr_debug("Discarded rx packet: len %zd\n", sock_len);
> > > - continue;
> > > - }
> > > /* OK, now we need to know about added descriptors. */
> > > if (!headcount) {
> > > if (unlikely(vhost_enable_notify(&net->dev, vq))) {
> > > @@ -800,6 +790,18 @@ static void handle_rx(struct vhost_net *net)
> > > * they refilled. */
> > > goto out;
> > > }
> > > + if (nvq->rx_array)
> > > + msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
> > > + /* On overrun, truncate and discard */
> > > + if (unlikely(headcount > UIO_MAXIOV)) {
> > > + iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
> > > + err = sock->ops->recvmsg(sock, &msg,
> > > + 1, MSG_DONTWAIT | MSG_TRUNC);
> > > + if (unlikely(err != 1))
> > Why 1? How is receiving 1 byte special or even possible?
> > Also, I wouldn't put an unlikely here. It's all error handling code anyway.
Vhost is dropping the skb by invoking a 1 byte recvmsg() here, while it
is kind of weird to free skb since it would have been freed in recvmsg()
for most cases, and the return value doesn't make sense too much.
> >
> > > + kfree_skb((struct sk_buff *)msg.msg_control);
> > You do not need a cast here.
Yes, exactly, I missed it.
> > Also, is it really safe to refer to msg_control here?
> > I'd rather keep a copy of the skb pointer and use it than assume
> > caller did not change it. But also see below.
It should be safe since msg is a local variable here, the callee has no
chance to modify it, except rx_array is not used by vhost and then it
becomes uncertain, but I don't know what the case is. Isn't vhost using
rx_array for all kinds of devices? Any clue rings the bell?
> >
> > > + pr_debug("Discarded rx packet: len %zd\n", sock_len);
> > > + continue;
> > > + }
> > > /* We don't need to be notified again. */
> > > iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
> > > fixup = msg.msg_iter;
> > > @@ -818,6 +820,7 @@ static void handle_rx(struct vhost_net *net)
> > > pr_debug("Discarded rx packet: "
> > > " len %d, expected %zd\n", err, sock_len);
> > > vhost_discard_vq_desc(vq, headcount);
> > > + kfree_skb((struct sk_buff *)msg.msg_control);
> > You do not need a cast here.
> >
> > Also, we have
> >
> > ret = tun_put_user(tun, tfile, skb, to);
> > if (unlikely(ret < 0))
> > kfree_skb(skb);
> > else
> > consume_skb(skb);
> >
> > return ret;
> >
> > So it looks like recvmsg actually always consumes the skb.
> > So I was wrong when I said you need to kfree it after
> > recv msg, and your original patch was good.
OK, I will repost it.
BTW, Per Jason's comments below, vhost has passed in proper flag,
iov, etc to both tun and tap device, so it would not be an issue.
The only case probably be missed would be removing tun dynamically
with traffic from vhost, while I am not sure how this can be reproduced,
I remember someone reported he got a similar issue by repeatedly
creating and destroying 1000+ VMs simultaneously. Has these kinds of
issues been fixed already? Or that might come about still?
if (!tun)
return -EBADFD;
Anyway, what about do it in recvmsg() and in another patch later on?
Wei
> >
> > Jason, what do you think?
> >
>
> tun_recvmsg() has the following check:
>
> static int tun_recvmsg(struct socket *sock, struct msghdr *m, size_t
> total_len,
> int flags)
> {
> struct tun_file *tfile = container_of(sock, struct tun_file, socket);
> struct tun_struct *tun = __tun_get(tfile);
> int ret;
>
> if (!tun)
> return -EBADFD;
>
> if (flags & ~(MSG_DONTWAIT|MSG_TRUNC|MSG_ERRQUEUE)) {
> ret = -EINVAL;
> goto out;
> }
>
> And tun_do_read() has:
>
> if (!iov_iter_count(to))
> return 0;
>
> So I think we need free skb in those cases.
>
> Thanks
From 1585463615569715912@xxx Thu Nov 30 04:25:29 +0000 2017
X-GM-THRID: 1585331132080725538
X-Gmail-Labels: Inbox,Category Forums,HistoricalUnread