regression caused by 1d2024f61ec14bdb0c57a97a3fe73685abc2d198?

regression caused by 1d2024f61ec14bdb0c57a97a3fe73685abc2d198?

[Michael S. Tsirkin]

It seems that starting with kernel 3.3 ixgbe sets gso_size for
incoming frames. It seems that this might result in gso_size
being set even when gso_type is 0.
This in turn leads to a crash at macvtap_skb_to_vnet_hdr
drivers/net/macvtap.c:628
which has this code:

       if (skb_is_gso(skb)) {
                struct skb_shared_info *sinfo = skb_shinfo(skb);

                /* This is a hint as to how much should be linear. */
                vnet_hdr->hdr_len = skb_headlen(skb);
                vnet_hdr->gso_size = sinfo->gso_size;
                if (sinfo->gso_type & SKB_GSO_TCPV4)
                        vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
                else if (sinfo->gso_type & SKB_GSO_TCPV6)
                        vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
                else if (sinfo->gso_type & SKB_GSO_UDP)
                        vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
                else
                        BUG();
                if (sinfo->gso_type & SKB_GSO_TCP_ECN)
                        vnet_hdr->gso_type |= VIRTIO_NET_HDR_GSO_ECN;
        } else
                vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_NONE;


Since skb_is_gso tests gso_size.

What's the right way to handle this? Should skb_is_gso be
changed to test gso_type != 0?

-- 
MST

Re: regression caused by 1d2024f61ec14bdb0c57a97a3fe73685abc2d198?

[Eric Dumazet]

On Wed, 2013-02-06 at 13:43 +0200, Michael S. Tsirkin wrote:
> It seems that starting with kernel 3.3 ixgbe sets gso_size for
> incoming frames. It seems that this might result in gso_size
> being set even when gso_type is 0.
> This in turn leads to a crash at macvtap_skb_to_vnet_hdr
> drivers/net/macvtap.c:628
> which has this code:
> 
>        if (skb_is_gso(skb)) {
>                 struct skb_shared_info *sinfo = skb_shinfo(skb);
> 
>                 /* This is a hint as to how much should be linear. */
>                 vnet_hdr->hdr_len = skb_headlen(skb);
>                 vnet_hdr->gso_size = sinfo->gso_size;
>                 if (sinfo->gso_type & SKB_GSO_TCPV4)
>                         vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
>                 else if (sinfo->gso_type & SKB_GSO_TCPV6)
>                         vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
>                 else if (sinfo->gso_type & SKB_GSO_UDP)
>                         vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
>                 else
>                         BUG();
>                 if (sinfo->gso_type & SKB_GSO_TCP_ECN)
>                         vnet_hdr->gso_type |= VIRTIO_NET_HDR_GSO_ECN;
>         } else
>                 vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_NONE;
> 
> 
> Since skb_is_gso tests gso_size.
> 
> What's the right way to handle this? Should skb_is_gso be
> changed to test gso_type != 0?
> 

Or fix ixgbe to set gso_type in ixgbe_get_headlen(), as it does all the
dissection.

Re: regression caused by 1d2024f61ec14bdb0c57a97a3fe73685abc2d198?

[Michael S. Tsirkin]

On Wed, Feb 06, 2013 at 05:07:39AM -0800, Eric Dumazet wrote:
> On Wed, 2013-02-06 at 13:43 +0200, Michael S. Tsirkin wrote:
> > It seems that starting with kernel 3.3 ixgbe sets gso_size for
> > incoming frames. It seems that this might result in gso_size
> > being set even when gso_type is 0.
> > This in turn leads to a crash at macvtap_skb_to_vnet_hdr
> > drivers/net/macvtap.c:628
> > which has this code:
> > 
> >        if (skb_is_gso(skb)) {
> >                 struct skb_shared_info *sinfo = skb_shinfo(skb);
> > 
> >                 /* This is a hint as to how much should be linear. */
> >                 vnet_hdr->hdr_len = skb_headlen(skb);
> >                 vnet_hdr->gso_size = sinfo->gso_size;
> >                 if (sinfo->gso_type & SKB_GSO_TCPV4)
> >                         vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
> >                 else if (sinfo->gso_type & SKB_GSO_TCPV6)
> >                         vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
> >                 else if (sinfo->gso_type & SKB_GSO_UDP)
> >                         vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
> >                 else
> >                         BUG();
> >                 if (sinfo->gso_type & SKB_GSO_TCP_ECN)
> >                         vnet_hdr->gso_type |= VIRTIO_NET_HDR_GSO_ECN;
> >         } else
> >                 vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_NONE;
> > 
> > 
> > Since skb_is_gso tests gso_size.
> > 
> > What's the right way to handle this? Should skb_is_gso be
> > changed to test gso_type != 0?
> > 
> 
> Or fix ixgbe to set gso_type in ixgbe_get_headlen(), as it does all the
> dissection.


Hmm, ixgbe_get_headlen isn't run on linear skbs though.

Also, I'm not sure I understand when should drivers set gso size
for incoming messages and what is a reasonable value.
Commit log talks about improved performance for lossy connections,
in this case, isn't this something net core should set?

I see 3 in-tree drivers that do this:

drivers/net/ethernet/broadcom/bnx2x/bnx2x_cmn.c:                skb_shinfo(skb)->gso_size = bnx2x
drivers/net/ethernet/intel/ixgbe/ixgbe_main.c:  skb_shinfo(skb)->gso_size = DIV_ROUND_UP((skb->le
drivers/net/ethernet/qlogic/qlcnic/qlcnic_io.c:         skb_shinfo(skb)->gso_size = qlcnic_get_lr

It seems likely the same issue applies there?


-- 
MST

Re: regression caused by 1d2024f61ec14bdb0c57a97a3fe73685abc2d198?

[Eric Dumazet]

On Wed, 2013-02-06 at 17:50 +0200, Michael S. Tsirkin wrote:

> Also, I'm not sure I understand when should drivers set gso size
> for incoming messages and what is a reasonable value.
> Commit log talks about improved performance for lossy connections,
> in this case, isn't this something net core should set?
> 
> I see 3 in-tree drivers that do this:
> 
> drivers/net/ethernet/broadcom/bnx2x/bnx2x_cmn.c:                skb_shinfo(skb)->gso_size = bnx2x

bnx2x is fine, take a look at lines 464

> drivers/net/ethernet/intel/ixgbe/ixgbe_main.c:  skb_shinfo(skb)->gso_size = DIV_ROUND_UP((skb->le
> drivers/net/ethernet/qlogic/qlcnic/qlcnic_io.c:         skb_shinfo(skb)->gso_size = qlcnic_get_lr
> 
> It seems likely the same issue applies there?

Yes

Re: regression caused by 1d2024f61ec14bdb0c57a97a3fe73685abc2d198?

[Eric Dumazet]

On Wed, 2013-02-06 at 17:50 +0200, Michael S. Tsirkin wrote:

> Commit log talks about improved performance for lossy connections,
> in this case, isn't this something net core should set?
> 

I fail to see where net core could do that.

These drivers kind of bypass GRO, so they must provide skbs that are
100% correct.

Re: regression caused by 1d2024f61ec14bdb0c57a97a3fe73685abc2d198?

[Michael S. Tsirkin]

On Wed, Feb 06, 2013 at 08:15:32AM -0800, Eric Dumazet wrote:
> On Wed, 2013-02-06 at 17:50 +0200, Michael S. Tsirkin wrote:
> 
> > Also, I'm not sure I understand when should drivers set gso size
> > for incoming messages and what is a reasonable value.
> > Commit log talks about improved performance for lossy connections,
> > in this case, isn't this something net core should set?
> > 
> > I see 3 in-tree drivers that do this:
> > 
> > drivers/net/ethernet/broadcom/bnx2x/bnx2x_cmn.c:                skb_shinfo(skb)->gso_size = bnx2x
> 
> bnx2x is fine, take a look at lines 464

Is this what you mean?

        /* This is needed in order to enable forwarding support */
        if (frag_size) {
                skb_shinfo(skb)->gso_size = bnx2x_set_lro_mss(bp,
                                        tpa_info->parsing_flags, len_on_bd);

                /* set for GRO */
                if (fp->mode == TPA_MODE_GRO)
                        skb_shinfo(skb)->gso_type =
                            (GET_FLAG(tpa_info->parsing_flags,
                                      PARSING_FLAGS_OVER_ETHERNET_PROTOCOL) ==
                                                PRS_FLAG_OVERETH_IPV6) ?
                                SKB_GSO_TCPV6 : SKB_GSO_TCPV4;
        }


I see it sets gso_type but apparently only if mode is GRO?
Will this still break if mode is set to LRO?




> > drivers/net/ethernet/intel/ixgbe/ixgbe_main.c:  skb_shinfo(skb)->gso_size = DIV_ROUND_UP((skb->le
> > drivers/net/ethernet/qlogic/qlcnic/qlcnic_io.c:         skb_shinfo(skb)->gso_size = qlcnic_get_lr
> > 
> > It seems likely the same issue applies there?
> 
> Yes

Re: regression caused by 1d2024f61ec14bdb0c57a97a3fe73685abc2d198?

[Eric Dumazet]

On Wed, 2013-02-06 at 19:42 +0200, Michael S. Tsirkin wrote:
> On Wed, Feb 06, 2013 at 08:15:32AM -0800, Eric Dumazet wrote:
> > On Wed, 2013-02-06 at 17:50 +0200, Michael S. Tsirkin wrote:
> > 
> > > Also, I'm not sure I understand when should drivers set gso size
> > > for incoming messages and what is a reasonable value.
> > > Commit log talks about improved performance for lossy connections,
> > > in this case, isn't this something net core should set?
> > > 
> > > I see 3 in-tree drivers that do this:
> > > 
> > > drivers/net/ethernet/broadcom/bnx2x/bnx2x_cmn.c:                skb_shinfo(skb)->gso_size = bnx2x
> > 
> > bnx2x is fine, take a look at lines 464
> 
> Is this what you mean?
> 
>         /* This is needed in order to enable forwarding support */
>         if (frag_size) {
>                 skb_shinfo(skb)->gso_size = bnx2x_set_lro_mss(bp,
>                                         tpa_info->parsing_flags, len_on_bd);
> 
>                 /* set for GRO */
>                 if (fp->mode == TPA_MODE_GRO)
>                         skb_shinfo(skb)->gso_type =
>                             (GET_FLAG(tpa_info->parsing_flags,
>                                       PARSING_FLAGS_OVER_ETHERNET_PROTOCOL) ==
>                                                 PRS_FLAG_OVERETH_IPV6) ?
>                                 SKB_GSO_TCPV6 : SKB_GSO_TCPV4;
>         }
> 
> 
> I see it sets gso_type but apparently only if mode is GRO?
> Will this still break if mode is set to LRO?
> 
> 


In net-next tree, line 464 looks like :

static void bnx2x_set_gro_params(struct sk_buff *skb, u16 parsing_flags,
                                 u16 len_on_bd, unsigned int pkt_len)
{
        /* TPA aggregation won't have either IP options or TCP options
         * other than timestamp or IPv6 extension headers.
         */
        u16 hdrs_len = ETH_HLEN + sizeof(struct tcphdr);

        if (GET_FLAG(parsing_flags, PARSING_FLAGS_OVER_ETHERNET_PROTOCOL) ==
            PRS_FLAG_OVERETH_IPV6) {
                hdrs_len += sizeof(struct ipv6hdr);
<HERE>          skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
        } else {
                hdrs_len += sizeof(struct iphdr);
<HERE>          skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
        }

        /* Check if there was a TCP timestamp, if there is it's will
         * always be 12 bytes length: nop nop kind length echo val.
         *
         * Otherwise FW would close the aggregation.
         */
        if (parsing_flags & PARSING_FLAGS_TIME_STAMP_EXIST_FLAG)
                hdrs_len += TPA_TSTAMP_OPT_LEN;

        skb_shinfo(skb)->gso_size = len_on_bd - hdrs_len;
...

Re: regression caused by 1d2024f61ec14bdb0c57a97a3fe73685abc2d198?

[Michael S. Tsirkin]

On Wed, Feb 06, 2013 at 09:42:45AM -0800, Eric Dumazet wrote:
> On Wed, 2013-02-06 at 19:42 +0200, Michael S. Tsirkin wrote:
> > On Wed, Feb 06, 2013 at 08:15:32AM -0800, Eric Dumazet wrote:
> > > On Wed, 2013-02-06 at 17:50 +0200, Michael S. Tsirkin wrote:
> > > 
> > > > Also, I'm not sure I understand when should drivers set gso size
> > > > for incoming messages and what is a reasonable value.
> > > > Commit log talks about improved performance for lossy connections,
> > > > in this case, isn't this something net core should set?
> > > > 
> > > > I see 3 in-tree drivers that do this:
> > > > 
> > > > drivers/net/ethernet/broadcom/bnx2x/bnx2x_cmn.c:                skb_shinfo(skb)->gso_size = bnx2x
> > > 
> > > bnx2x is fine, take a look at lines 464
> > 
> > Is this what you mean?
> > 
> >         /* This is needed in order to enable forwarding support */
> >         if (frag_size) {
> >                 skb_shinfo(skb)->gso_size = bnx2x_set_lro_mss(bp,
> >                                         tpa_info->parsing_flags, len_on_bd);
> > 
> >                 /* set for GRO */
> >                 if (fp->mode == TPA_MODE_GRO)
> >                         skb_shinfo(skb)->gso_type =
> >                             (GET_FLAG(tpa_info->parsing_flags,
> >                                       PARSING_FLAGS_OVER_ETHERNET_PROTOCOL) ==
> >                                                 PRS_FLAG_OVERETH_IPV6) ?
> >                                 SKB_GSO_TCPV6 : SKB_GSO_TCPV4;
> >         }
> > 
> > 
> > I see it sets gso_type but apparently only if mode is GRO?
> > Will this still break if mode is set to LRO?
> > 
> > 
> 
> 
> In net-next tree, line 464 looks like :
> 
> static void bnx2x_set_gro_params(struct sk_buff *skb, u16 parsing_flags,
>                                  u16 len_on_bd, unsigned int pkt_len)
> {
>         /* TPA aggregation won't have either IP options or TCP options
>          * other than timestamp or IPv6 extension headers.
>          */
>         u16 hdrs_len = ETH_HLEN + sizeof(struct tcphdr);
> 
>         if (GET_FLAG(parsing_flags, PARSING_FLAGS_OVER_ETHERNET_PROTOCOL) ==
>             PRS_FLAG_OVERETH_IPV6) {
>                 hdrs_len += sizeof(struct ipv6hdr);
> <HERE>          skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
>         } else {
>                 hdrs_len += sizeof(struct iphdr);
> <HERE>          skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
>         }
> 
>         /* Check if there was a TCP timestamp, if there is it's will
>          * always be 12 bytes length: nop nop kind length echo val.
>          *
>          * Otherwise FW would close the aggregation.
>          */
>         if (parsing_flags & PARSING_FLAGS_TIME_STAMP_EXIST_FLAG)
>                 hdrs_len += TPA_TSTAMP_OPT_LEN;
> 
>         skb_shinfo(skb)->gso_size = len_on_bd - hdrs_len;
> ...
> 

OK, it looks like intel/qlogic can just look at skb protocol
and set gso_type to TCPV4/V6.
I'll try it and see if this works.

-- 
MST

Re: regression caused by 1d2024f61ec14bdb0c57a97a3fe73685abc2d198?

[Michael S. Tsirkin]

On Wed, Feb 06, 2013 at 09:42:45AM -0800, Eric Dumazet wrote:
> On Wed, 2013-02-06 at 19:42 +0200, Michael S. Tsirkin wrote:
> > On Wed, Feb 06, 2013 at 08:15:32AM -0800, Eric Dumazet wrote:
> > > On Wed, 2013-02-06 at 17:50 +0200, Michael S. Tsirkin wrote:
> > > 
> > > > Also, I'm not sure I understand when should drivers set gso size
> > > > for incoming messages and what is a reasonable value.
> > > > Commit log talks about improved performance for lossy connections,
> > > > in this case, isn't this something net core should set?
> > > > 
> > > > I see 3 in-tree drivers that do this:
> > > > 
> > > > drivers/net/ethernet/broadcom/bnx2x/bnx2x_cmn.c:                skb_shinfo(skb)->gso_size = bnx2x
> > > 
> > > bnx2x is fine, take a look at lines 464
> > 
> > Is this what you mean?
> > 
> >         /* This is needed in order to enable forwarding support */
> >         if (frag_size) {
> >                 skb_shinfo(skb)->gso_size = bnx2x_set_lro_mss(bp,
> >                                         tpa_info->parsing_flags, len_on_bd);
> > 
> >                 /* set for GRO */
> >                 if (fp->mode == TPA_MODE_GRO)
> >                         skb_shinfo(skb)->gso_type =
> >                             (GET_FLAG(tpa_info->parsing_flags,
> >                                       PARSING_FLAGS_OVER_ETHERNET_PROTOCOL) ==
> >                                                 PRS_FLAG_OVERETH_IPV6) ?
> >                                 SKB_GSO_TCPV6 : SKB_GSO_TCPV4;
> >         }
> > 
> > 
> > I see it sets gso_type but apparently only if mode is GRO?
> > Will this still break if mode is set to LRO?
> > 
> > 
> 
> 
> In net-next tree, line 464 looks like :
> 
> static void bnx2x_set_gro_params(struct sk_buff *skb, u16 parsing_flags,
>                                  u16 len_on_bd, unsigned int pkt_len)
> {
>         /* TPA aggregation won't have either IP options or TCP options
>          * other than timestamp or IPv6 extension headers.
>          */
>         u16 hdrs_len = ETH_HLEN + sizeof(struct tcphdr);
> 
>         if (GET_FLAG(parsing_flags, PARSING_FLAGS_OVER_ETHERNET_PROTOCOL) ==
>             PRS_FLAG_OVERETH_IPV6) {
>                 hdrs_len += sizeof(struct ipv6hdr);
> <HERE>          skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
>         } else {
>                 hdrs_len += sizeof(struct iphdr);
> <HERE>          skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
>         }
> 
>         /* Check if there was a TCP timestamp, if there is it's will
>          * always be 12 bytes length: nop nop kind length echo val.
>          *
>          * Otherwise FW would close the aggregation.
>          */
>         if (parsing_flags & PARSING_FLAGS_TIME_STAMP_EXIST_FLAG)
>                 hdrs_len += TPA_TSTAMP_OPT_LEN;
> 
>         skb_shinfo(skb)->gso_size = len_on_bd - hdrs_len;
> ...

OK this means cbf1de72324a8105ddcc3d9ce9acbc613faea17e might be needed
in 3.8 and maybe -stable, otherwise macvtap crashes when LRO is set.

-- 
MST

Re: regression caused by 1d2024f61ec14bdb0c57a97a3fe73685abc2d198?

[Eric Dumazet]

On Wed, 2013-02-06 at 20:05 +0200, Michael S. Tsirkin wrote:

> 
> OK this means cbf1de72324a8105ddcc3d9ce9acbc613faea17e might be needed
> in 3.8 and maybe -stable, otherwise macvtap crashes when LRO is set.
> 

You may be right : My issue was an accounting error in a rarely used
path (qdisc on ingress). Yours is definitely more serious ;)

Re: regression caused by 1d2024f61ec14bdb0c57a97a3fe73685abc2d198?

[Ben Hutchings]

On Wed, 2013-02-06 at 17:50 +0200, Michael S. Tsirkin wrote:
> On Wed, Feb 06, 2013 at 05:07:39AM -0800, Eric Dumazet wrote:
> > On Wed, 2013-02-06 at 13:43 +0200, Michael S. Tsirkin wrote:
> > > It seems that starting with kernel 3.3 ixgbe sets gso_size for
> > > incoming frames. It seems that this might result in gso_size
> > > being set even when gso_type is 0.
> > > This in turn leads to a crash at macvtap_skb_to_vnet_hdr
> > > drivers/net/macvtap.c:628
> > > which has this code:
> > > 
> > >        if (skb_is_gso(skb)) {
> > >                 struct skb_shared_info *sinfo = skb_shinfo(skb);
> > > 
> > >                 /* This is a hint as to how much should be linear. */
> > >                 vnet_hdr->hdr_len = skb_headlen(skb);
> > >                 vnet_hdr->gso_size = sinfo->gso_size;
> > >                 if (sinfo->gso_type & SKB_GSO_TCPV4)
> > >                         vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
> > >                 else if (sinfo->gso_type & SKB_GSO_TCPV6)
> > >                         vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
> > >                 else if (sinfo->gso_type & SKB_GSO_UDP)
> > >                         vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
> > >                 else
> > >                         BUG();
> > >                 if (sinfo->gso_type & SKB_GSO_TCP_ECN)
> > >                         vnet_hdr->gso_type |= VIRTIO_NET_HDR_GSO_ECN;
> > >         } else
> > >                 vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_NONE;
> > > 
> > > 
> > > Since skb_is_gso tests gso_size.
> > > 
> > > What's the right way to handle this? Should skb_is_gso be
> > > changed to test gso_type != 0?
> > > 
> > 
> > Or fix ixgbe to set gso_type in ixgbe_get_headlen(), as it does all the
> > dissection.
> 
> 
> Hmm, ixgbe_get_headlen isn't run on linear skbs though.
> 
> Also, I'm not sure I understand when should drivers set gso size
> for incoming messages and what is a reasonable value.
> Commit log talks about improved performance for lossy connections,
> in this case, isn't this something net core should set?
[...]

It should be set to the segment size on the wire, so TCP gets a correct
picture of packet loss.  The networking core has no idea what hardware/
firmware LRO did.

I've previously raised this issue of macvlan vs LRO (which is the same
issue we previously had with IP forwarding and with bridging):
     http://thread.gmane.org/gmane.linux.network/221695

Ben.

-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

Re: regression caused by 1d2024f61ec14bdb0c57a97a3fe73685abc2d198?

[Michael S. Tsirkin]

On Wed, Feb 06, 2013 at 07:58:21PM +0000, Ben Hutchings wrote:
> On Wed, 2013-02-06 at 17:50 +0200, Michael S. Tsirkin wrote:
> > On Wed, Feb 06, 2013 at 05:07:39AM -0800, Eric Dumazet wrote:
> > > On Wed, 2013-02-06 at 13:43 +0200, Michael S. Tsirkin wrote:
> > > > It seems that starting with kernel 3.3 ixgbe sets gso_size for
> > > > incoming frames. It seems that this might result in gso_size
> > > > being set even when gso_type is 0.
> > > > This in turn leads to a crash at macvtap_skb_to_vnet_hdr
> > > > drivers/net/macvtap.c:628
> > > > which has this code:
> > > > 
> > > >        if (skb_is_gso(skb)) {
> > > >                 struct skb_shared_info *sinfo = skb_shinfo(skb);
> > > > 
> > > >                 /* This is a hint as to how much should be linear. */
> > > >                 vnet_hdr->hdr_len = skb_headlen(skb);
> > > >                 vnet_hdr->gso_size = sinfo->gso_size;
> > > >                 if (sinfo->gso_type & SKB_GSO_TCPV4)
> > > >                         vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
> > > >                 else if (sinfo->gso_type & SKB_GSO_TCPV6)
> > > >                         vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
> > > >                 else if (sinfo->gso_type & SKB_GSO_UDP)
> > > >                         vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
> > > >                 else
> > > >                         BUG();
> > > >                 if (sinfo->gso_type & SKB_GSO_TCP_ECN)
> > > >                         vnet_hdr->gso_type |= VIRTIO_NET_HDR_GSO_ECN;
> > > >         } else
> > > >                 vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_NONE;
> > > > 
> > > > 
> > > > Since skb_is_gso tests gso_size.
> > > > 
> > > > What's the right way to handle this? Should skb_is_gso be
> > > > changed to test gso_type != 0?
> > > > 
> > > 
> > > Or fix ixgbe to set gso_type in ixgbe_get_headlen(), as it does all the
> > > dissection.
> > 
> > 
> > Hmm, ixgbe_get_headlen isn't run on linear skbs though.
> > 
> > Also, I'm not sure I understand when should drivers set gso size
> > for incoming messages and what is a reasonable value.
> > Commit log talks about improved performance for lossy connections,
> > in this case, isn't this something net core should set?
> [...]
> 
> It should be set to the segment size on the wire, so TCP gets a correct
> picture of packet loss.  The networking core has no idea what hardware/
> firmware LRO did.
> 
> I've previously raised this issue of macvlan vs LRO (which is the same
> issue we previously had with IP forwarding and with bridging):
>      http://thread.gmane.org/gmane.linux.network/221695
> 
> Ben.

I see, you proposed disabling LRO the moment a macvlan is attached.
If I understand correctly, the difference as compared to bridge is that
bridge normally consumes all incoming packets, macvlan is often used in
parallel with the underlying interface.

BTW are there other issues with forwarding/bridging and LRO? If everyone
sets gso_type in packets it seems we can leave LRO set even with
bridging?

> -- 
> Ben Hutchings, Staff Engineer, Solarflare
> Not speaking for my employer; that's the marketing department's job.
> They asked us to note that Solarflare product names are trademarked.

Re: regression caused by 1d2024f61ec14bdb0c57a97a3fe73685abc2d198?

[Ben Hutchings]

On Wed, 2013-02-06 at 23:45 +0200, Michael S. Tsirkin wrote:
> On Wed, Feb 06, 2013 at 07:58:21PM +0000, Ben Hutchings wrote:
> > On Wed, 2013-02-06 at 17:50 +0200, Michael S. Tsirkin wrote:
> > > On Wed, Feb 06, 2013 at 05:07:39AM -0800, Eric Dumazet wrote:
> > > > On Wed, 2013-02-06 at 13:43 +0200, Michael S. Tsirkin wrote:
> > > > > It seems that starting with kernel 3.3 ixgbe sets gso_size for
> > > > > incoming frames. It seems that this might result in gso_size
> > > > > being set even when gso_type is 0.
> > > > > This in turn leads to a crash at macvtap_skb_to_vnet_hdr
> > > > > drivers/net/macvtap.c:628
> > > > > which has this code:
> > > > > 
> > > > >        if (skb_is_gso(skb)) {
> > > > >                 struct skb_shared_info *sinfo = skb_shinfo(skb);
> > > > > 
> > > > >                 /* This is a hint as to how much should be linear. */
> > > > >                 vnet_hdr->hdr_len = skb_headlen(skb);
> > > > >                 vnet_hdr->gso_size = sinfo->gso_size;
> > > > >                 if (sinfo->gso_type & SKB_GSO_TCPV4)
> > > > >                         vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
> > > > >                 else if (sinfo->gso_type & SKB_GSO_TCPV6)
> > > > >                         vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
> > > > >                 else if (sinfo->gso_type & SKB_GSO_UDP)
> > > > >                         vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
> > > > >                 else
> > > > >                         BUG();
> > > > >                 if (sinfo->gso_type & SKB_GSO_TCP_ECN)
> > > > >                         vnet_hdr->gso_type |= VIRTIO_NET_HDR_GSO_ECN;
> > > > >         } else
> > > > >                 vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_NONE;
> > > > > 
> > > > > 
> > > > > Since skb_is_gso tests gso_size.
> > > > > 
> > > > > What's the right way to handle this? Should skb_is_gso be
> > > > > changed to test gso_type != 0?
> > > > > 
> > > > 
> > > > Or fix ixgbe to set gso_type in ixgbe_get_headlen(), as it does all the
> > > > dissection.
> > > 
> > > 
> > > Hmm, ixgbe_get_headlen isn't run on linear skbs though.
> > > 
> > > Also, I'm not sure I understand when should drivers set gso size
> > > for incoming messages and what is a reasonable value.
> > > Commit log talks about improved performance for lossy connections,
> > > in this case, isn't this something net core should set?
> > [...]
> > 
> > It should be set to the segment size on the wire, so TCP gets a correct
> > picture of packet loss.  The networking core has no idea what hardware/
> > firmware LRO did.
> > 
> > I've previously raised this issue of macvlan vs LRO (which is the same
> > issue we previously had with IP forwarding and with bridging):
> >      http://thread.gmane.org/gmane.linux.network/221695
> > 
> > Ben.
> 
> I see, you proposed disabling LRO the moment a macvlan is attached.
> If I understand correctly, the difference as compared to bridge is that
> bridge normally consumes all incoming packets, macvlan is often used in
> parallel with the underlying interface.

Not so different from IP forwarding, though, in that a single interface
can both forward packets and deliver them locally.

> BTW are there other issues with forwarding/bridging and LRO? If everyone
> sets gso_type in packets it seems we can leave LRO set even with
> bridging?

Some implementations of LRO violate the end-to-end principle by merging
segments with varying lengths and TCP timestamps.  GRO is very careful
to ensure that the original packets can be recovered from the skbs it
produces.

Ben.

-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

Re: regression caused by 1d2024f61ec14bdb0c57a97a3fe73685abc2d198?

[Michael S. Tsirkin]

On Wed, Feb 06, 2013 at 09:55:47PM +0000, Ben Hutchings wrote:
> On Wed, 2013-02-06 at 23:45 +0200, Michael S. Tsirkin wrote:
> > On Wed, Feb 06, 2013 at 07:58:21PM +0000, Ben Hutchings wrote:
> > > On Wed, 2013-02-06 at 17:50 +0200, Michael S. Tsirkin wrote:
> > > > On Wed, Feb 06, 2013 at 05:07:39AM -0800, Eric Dumazet wrote:
> > > > > On Wed, 2013-02-06 at 13:43 +0200, Michael S. Tsirkin wrote:
> > > > > > It seems that starting with kernel 3.3 ixgbe sets gso_size for
> > > > > > incoming frames. It seems that this might result in gso_size
> > > > > > being set even when gso_type is 0.
> > > > > > This in turn leads to a crash at macvtap_skb_to_vnet_hdr
> > > > > > drivers/net/macvtap.c:628
> > > > > > which has this code:
> > > > > > 
> > > > > >        if (skb_is_gso(skb)) {
> > > > > >                 struct skb_shared_info *sinfo = skb_shinfo(skb);
> > > > > > 
> > > > > >                 /* This is a hint as to how much should be linear. */
> > > > > >                 vnet_hdr->hdr_len = skb_headlen(skb);
> > > > > >                 vnet_hdr->gso_size = sinfo->gso_size;
> > > > > >                 if (sinfo->gso_type & SKB_GSO_TCPV4)
> > > > > >                         vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
> > > > > >                 else if (sinfo->gso_type & SKB_GSO_TCPV6)
> > > > > >                         vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
> > > > > >                 else if (sinfo->gso_type & SKB_GSO_UDP)
> > > > > >                         vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
> > > > > >                 else
> > > > > >                         BUG();
> > > > > >                 if (sinfo->gso_type & SKB_GSO_TCP_ECN)
> > > > > >                         vnet_hdr->gso_type |= VIRTIO_NET_HDR_GSO_ECN;
> > > > > >         } else
> > > > > >                 vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_NONE;
> > > > > > 
> > > > > > 
> > > > > > Since skb_is_gso tests gso_size.
> > > > > > 
> > > > > > What's the right way to handle this? Should skb_is_gso be
> > > > > > changed to test gso_type != 0?
> > > > > > 
> > > > > 
> > > > > Or fix ixgbe to set gso_type in ixgbe_get_headlen(), as it does all the
> > > > > dissection.
> > > > 
> > > > 
> > > > Hmm, ixgbe_get_headlen isn't run on linear skbs though.
> > > > 
> > > > Also, I'm not sure I understand when should drivers set gso size
> > > > for incoming messages and what is a reasonable value.
> > > > Commit log talks about improved performance for lossy connections,
> > > > in this case, isn't this something net core should set?
> > > [...]
> > > 
> > > It should be set to the segment size on the wire, so TCP gets a correct
> > > picture of packet loss.  The networking core has no idea what hardware/
> > > firmware LRO did.
> > > 
> > > I've previously raised this issue of macvlan vs LRO (which is the same
> > > issue we previously had with IP forwarding and with bridging):
> > >      http://thread.gmane.org/gmane.linux.network/221695
> > > 
> > > Ben.
> > 
> > I see, you proposed disabling LRO the moment a macvlan is attached.
> > If I understand correctly, the difference as compared to bridge is that
> > bridge normally consumes all incoming packets, macvlan is often used in
> > parallel with the underlying interface.
> 
> Not so different from IP forwarding, though, in that a single interface
> can both forward packets and deliver them locally.

Hmm, for ip forwarding we don't try do disable LRO on the device, do we?
What's the solution there?

> > BTW are there other issues with forwarding/bridging and LRO? If everyone
> > sets gso_type in packets it seems we can leave LRO set even with
> > bridging?
> 
> Some implementations of LRO violate the end-to-end principle by merging
> segments with varying lengths and TCP timestamps.  GRO is very careful
> to ensure that the original packets can be recovered from the skbs it
> produces.
> 
> Ben.
> 
> -- 
> Ben Hutchings, Staff Engineer, Solarflare
> Not speaking for my employer; that's the marketing department's job.
> They asked us to note that Solarflare product names are trademarked.

Re: regression caused by 1d2024f61ec14bdb0c57a97a3fe73685abc2d198?

[Ben Hutchings]

On Thu, 2013-02-07 at 00:26 +0200, Michael S. Tsirkin wrote:
> On Wed, Feb 06, 2013 at 09:55:47PM +0000, Ben Hutchings wrote:
> > On Wed, 2013-02-06 at 23:45 +0200, Michael S. Tsirkin wrote:
> > > On Wed, Feb 06, 2013 at 07:58:21PM +0000, Ben Hutchings wrote:
> > > > On Wed, 2013-02-06 at 17:50 +0200, Michael S. Tsirkin wrote:
> > > > > On Wed, Feb 06, 2013 at 05:07:39AM -0800, Eric Dumazet wrote:
> > > > > > On Wed, 2013-02-06 at 13:43 +0200, Michael S. Tsirkin wrote:
> > > > > > > It seems that starting with kernel 3.3 ixgbe sets gso_size for
> > > > > > > incoming frames. It seems that this might result in gso_size
> > > > > > > being set even when gso_type is 0.
> > > > > > > This in turn leads to a crash at macvtap_skb_to_vnet_hdr
> > > > > > > drivers/net/macvtap.c:628
> > > > > > > which has this code:
> > > > > > > 
> > > > > > >        if (skb_is_gso(skb)) {
> > > > > > >                 struct skb_shared_info *sinfo = skb_shinfo(skb);
> > > > > > > 
> > > > > > >                 /* This is a hint as to how much should be linear. */
> > > > > > >                 vnet_hdr->hdr_len = skb_headlen(skb);
> > > > > > >                 vnet_hdr->gso_size = sinfo->gso_size;
> > > > > > >                 if (sinfo->gso_type & SKB_GSO_TCPV4)
> > > > > > >                         vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
> > > > > > >                 else if (sinfo->gso_type & SKB_GSO_TCPV6)
> > > > > > >                         vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
> > > > > > >                 else if (sinfo->gso_type & SKB_GSO_UDP)
> > > > > > >                         vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
> > > > > > >                 else
> > > > > > >                         BUG();
> > > > > > >                 if (sinfo->gso_type & SKB_GSO_TCP_ECN)
> > > > > > >                         vnet_hdr->gso_type |= VIRTIO_NET_HDR_GSO_ECN;
> > > > > > >         } else
> > > > > > >                 vnet_hdr->gso_type = VIRTIO_NET_HDR_GSO_NONE;
> > > > > > > 
> > > > > > > 
> > > > > > > Since skb_is_gso tests gso_size.
> > > > > > > 
> > > > > > > What's the right way to handle this? Should skb_is_gso be
> > > > > > > changed to test gso_type != 0?
> > > > > > > 
> > > > > > 
> > > > > > Or fix ixgbe to set gso_type in ixgbe_get_headlen(), as it does all the
> > > > > > dissection.
> > > > > 
> > > > > 
> > > > > Hmm, ixgbe_get_headlen isn't run on linear skbs though.
> > > > > 
> > > > > Also, I'm not sure I understand when should drivers set gso size
> > > > > for incoming messages and what is a reasonable value.
> > > > > Commit log talks about improved performance for lossy connections,
> > > > > in this case, isn't this something net core should set?
> > > > [...]
> > > > 
> > > > It should be set to the segment size on the wire, so TCP gets a correct
> > > > picture of packet loss.  The networking core has no idea what hardware/
> > > > firmware LRO did.
> > > > 
> > > > I've previously raised this issue of macvlan vs LRO (which is the same
> > > > issue we previously had with IP forwarding and with bridging):
> > > >      http://thread.gmane.org/gmane.linux.network/221695
> > > > 
> > > > Ben.
> > > 
> > > I see, you proposed disabling LRO the moment a macvlan is attached.
> > > If I understand correctly, the difference as compared to bridge is that
> > > bridge normally consumes all incoming packets, macvlan is often used in
> > > parallel with the underlying interface.
> > 
> > Not so different from IP forwarding, though, in that a single interface
> > can both forward packets and deliver them locally.
> 
> Hmm, for ip forwarding we don't try do disable LRO on the device, do we?
> What's the solution there?

Yes we do (or we did).

Ben.

> > > BTW are there other issues with forwarding/bridging and LRO? If everyone
> > > sets gso_type in packets it seems we can leave LRO set even with
> > > bridging?
> > 
> > Some implementations of LRO violate the end-to-end principle by merging
> > segments with varying lengths and TCP timestamps.  GRO is very careful
> > to ensure that the original packets can be recovered from the skbs it
> > produces.
> > 
> > Ben.
> > 
> > -- 
> > Ben Hutchings, Staff Engineer, Solarflare
> > Not speaking for my employer; that's the marketing department's job.
> > They asked us to note that Solarflare product names are trademarked.

-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

Re: regression caused by 1d2024f61ec14bdb0c57a97a3fe73685abc2d198?

[Cong Wang]

On Wed, 06 Feb 2013 at 17:42 GMT, Michael S. Tsirkin <mst@redhat.com> wrote:
>
> I see it sets gso_type but apparently only if mode is GRO?
> Will this still break if mode is set to LRO?

The comment inside skb_warn_if_lro() said:

        /* LRO sets gso_size but not gso_type, whereas if GSO is really
         * wanted then gso_type will be set. */