Rusty Russell wrote:
> On Thursday 26 June 2008 05:07:18 Anthony Liguori wrote:
>
>> Rusty Russell wrote:
>>
>>> @@ -1563,6 +1561,16 @@ static void setup_tun_net(char *arg)
>>> /* Tell Guest what MAC address to use. */
>>> add_feature(dev, VIRTIO_NET_F_MAC);
>>> add_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY);
>>> + /* Expect Guest to handle everything except UFO */
>>> + add_feature(dev, VIRTIO_NET_F_CSUM);
>>>
>> You're setting this feature twice.
>>
>
> Hmm, not in the version here?
>
Sorry, misread apparently.
>>> + add_feature(dev, VIRTIO_NET_F_GUEST_CSUM);
>>>
>> You set this feature, but I never see the virtio-net driver acknowledge
>> the feature.
I still don't see GUEST_CSUM ever get referenced in virtio_net.c.
What's the intention of this feature bit? Could I be missing a
virtio_net patch? I'm using the latest bits in Linus' tree.
>> Curiously, my implementation with KVM is struggling
>> because UDP packet checksums are not correct so the DHCP client is
>> ignoring them. If I disable CSUM offload, things it works fine (using
>> the virtio-net header). The problem is only host=>guest, guest=>host is
>> fine.
>>
>
> OK, found this: wrong args to skb_partial_csum_set. It was found by Mark
> McLoughlin before, I just lost the fix when I extracted this into a separate
> patch. I chose to move the call to skb_partial_csum_set(), rather than use
> his fix (which assumed a tap not tun device).
>
This still doesn't fix the problem. I can manually assign an IP address
and even do netperf runs but I cannot get a dhcp address (dhclient is
picky about the udp csum). Also, the RX performance is so low that I'm
sure a ton of packets are getting dropped. However, this patchset is
extremely promising, here are the results with KVM for TX:
w/o gso
[ 3] 0.0-10.0 sec 593 MBytes 498 Mbits/sec
w/gso
[ 5] 0.0-10.0 sec 1.86 GBytes 1.60 Gbits/sec
So that's a huge increase. Unfortunately, RX drops from 1.04 Gbits/sec
to only a few hundred Kbit/sec. I'm pretty sure this is the
checksumming issue.
Also, when I exit KVM, QEMU zombies and I notice:
Message from [EMAIL PROTECTED] at Jun 26 13:02:07 ...
kernel: unregister_netdevice: waiting for tap0 to become free. Usage
count = 3
Message from [EMAIL PROTECTED] at Jun 26 13:02:17 ...
kernel: unregister_netdevice: waiting for tap0 to become free. Usage
count = 0
Once the refcount drops to 0, the process exits. It looks fishy to me
though.
Regards,
Anthony Liguori
> Here's two fixes on top of previous patch:
>
> diff -u b/drivers/net/tun.c b/drivers/net/tun.c
> --- b/drivers/net/tun.c Thu Jun 26 00:21:59 2008 +1000
> +++ b/drivers/net/tun.c Thu Jun 26 14:35:03 2008 +1000
> @@ -298,11 +298,11 @@
> if ((len -= sizeof(gso)) > count)
> return -EINVAL;
>
> - if (gso.hdr_len > len)
> - return -EINVAL;
> -
> if (memcpy_fromiovec((void *)&gso, iv, sizeof(gso)))
> return -EFAULT;
> +
> + if (gso.hdr_len > len)
> + return -EINVAL;
> }
>
> if ((tun->flags & TUN_TYPE_MASK) == TUN_TAP_DEV) {
> @@ -324,6 +324,16 @@
> return -EFAULT;
> }
>
> + if (gso.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
> + if (!skb_partial_csum_set(skb, gso.csum_start,
> + gso.csum_offset)) {
> + tun->dev->stats.rx_dropped++;
> + kfree_skb(skb);
> + return -EINVAL;
> + }
> + } else if (tun->flags & TUN_NOCHECKSUM)
> + skb->ip_summed = CHECKSUM_UNNECESSARY;
> +
> switch (tun->flags & TUN_TYPE_MASK) {
> case TUN_TUN_DEV:
> skb_reset_mac_header(skb);
> @@ -335,16 +345,6 @@
> break;
> };
>
> - if (gso.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
> - if (!skb_partial_csum_set(skb, gso.csum_start,
> - gso.csum_offset)) {
> - tun->dev->stats.rx_dropped++;
> - kfree_skb(skb);
> - return -EINVAL;
> - }
> - } else if (tun->flags & TUN_NOCHECKSUM)
> - skb->ip_summed = CHECKSUM_UNNECESSARY;
> -
> if (gso.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
> pr_debug("GSO!\n");
> switch (gso.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
>
_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/virtualization
