Date:	Fri, 29 Jan 2016 20:07:44 -0800
From:	Alexander Duyck <alexander.duyck@...il.com>
To:	Marcelo Ricardo Leitner <marcelo.leitner@...il.com>
Cc:	Netdev <netdev@...r.kernel.org>,
	Neil Horman <nhorman@...driver.com>,
	Vlad Yasevich <vyasevich@...il.com>,
	David Miller <davem@...emloft.net>,
	Jesper Dangaard Brouer <brouer@...hat.com>,
	Alexei Starovoitov <alexei.starovoitov@...il.com>,
	Daniel Borkmann <borkmann@...earbox.net>, marek@...udflare.com,
	Hannes Frederic Sowa <hannes@...essinduktion.org>,
	Florian Westphal <fw@...len.de>, pabeni@...hat.com,
	John Fastabend <john.r.fastabend@...el.com>,
	linux-sctp@...r.kernel.org, Tom Herbert <tom@...bertland.com>
Subject: Re: [RFC PATCH net-next 3/3] sctp: Add GSO support

On Fri, Jan 29, 2016 at 11:42 AM, Marcelo Ricardo Leitner
<marcelo.leitner@...il.com> wrote:
> On Fri, Jan 29, 2016 at 11:15:54AM -0800, Alexander Duyck wrote:
>> On Wed, Jan 27, 2016 at 9:06 AM, Marcelo Ricardo Leitner
>> <marcelo.leitner@...il.com> wrote:
>> > This patch enables SCTP to do GSO.
>> >
>> > SCTP has this pecualiarty that its packets cannot be just segmented to
>> > (P)MTU. Its chunks must be contained in IP segments, padding respected.
>> > So we can't just generate a big skb, set gso_size to the fragmentation
>> > point and deliver it to IP layer.
>> >
>> > Instead, this patch proposes that SCTP build a skb as it would be if it
>> > was received using GRO. That is, there will be a cover skb with the
>> > headers (incluing SCTP one) and children ones containing the actual SCTP
>> > chunks, already segmented in a way that respects SCTP RFCs and MTU.
>> >
>> > This way SCTP can benefit from GSO and instead of passing several
>> > packets through the stack, it can pass a single large packet if there
>> > are enough data queued and cwnd allows.
>> >
>> > Main points that need help:
>> > - Usage of skb_gro_receive()
>> >   It fits nicely in there and properly handles offsets/lens, though the
>> >   name means another thing. If you agree with this usage, we can rename
>> >   it to something like skb_coalesce
>> >
>> > - Checksum handling
>> >   Why only packets with checksum offloaded can be GSOed? Most of the
>> >   NICs doesn't support SCTP CRC offloading and this will nearly defeat
>> >   this feature. If checksum is being computed in sw, it doesn't really
>> >   matter if it's earlier or later, right?
>> >   This patch hacks skb_needs_check() to allow using GSO with sw-computed
>> >   checksums.
>> >   Also the meaning of UNNECESSARY and NONE are quite foggy to me yet and
>> >   its usage may be wrong.
>> >
>> > - gso_size = 1
>> >   There is skb_is_gso() all over the stack and it basically checks for
>> >   non-zero skb_shinfo(skb)->gso_size. Setting it to 1 is the hacky way I
>> >   found to keep skb_is_gso() working while being able to signal to
>> >   skb_segment() that it shouldn't use gso_size but instead the fragment
>> >   sizes themselves. skb_segment() will mainly just unpack the skb then.
>>
>> Instead of 1 why not use 0xFFFF?  It is a value that can never be used
>> for a legitimate segment size since IP total length is a 16 bit value
>> and includes the IP header in the size.
>
> Just felt that 1 was unpractical. But perhaps with no hard restriction
> like the one for 0xFFFF you said. I can replace it, 0xFFFF is better.
>
>> > - socket / gso max values
>> >   usage of sk_setup_caps() still needs a review
>> >
>> > Signed-off-by: Marcelo Ricardo Leitner <marcelo.leitner@...il.com>
>> > ---
>> >  include/linux/netdev_features.h |   7 +-
>> >  include/linux/netdevice.h       |   1 +
>> >  net/core/dev.c                  |   6 +-
>> >  net/core/skbuff.c               |  12 +-
>> >  net/ipv4/af_inet.c              |   1 +
>> >  net/sctp/offload.c              |  53 +++++++
>> >  net/sctp/output.c               | 338 +++++++++++++++++++++++++---------------
>> >  net/sctp/socket.c               |   2 +
>> >  8 files changed, 292 insertions(+), 128 deletions(-)
>> >
>> > diff --git a/include/linux/netdev_features.h b/include/linux/netdev_features.h
>> > index d9654f0eecb3519383441afa6b131ff9a5898485..f678998841f1800e0f2fe416a79935197d4ed305 100644
>> > --- a/include/linux/netdev_features.h
>> > +++ b/include/linux/netdev_features.h
>> > @@ -48,8 +48,9 @@ enum {
>> >         NETIF_F_GSO_UDP_TUNNEL_BIT,     /* ... UDP TUNNEL with TSO */
>> >         NETIF_F_GSO_UDP_TUNNEL_CSUM_BIT,/* ... UDP TUNNEL with TSO & CSUM */
>> >         NETIF_F_GSO_TUNNEL_REMCSUM_BIT, /* ... TUNNEL with TSO & REMCSUM */
>> > +       NETIF_F_GSO_SCTP_BIT,           /* ... SCTP fragmentation */
>> >         /**/NETIF_F_GSO_LAST =          /* last bit, see GSO_MASK */
>> > -               NETIF_F_GSO_TUNNEL_REMCSUM_BIT,
>> > +               NETIF_F_GSO_SCTP_BIT,
>> >
>> >         NETIF_F_FCOE_CRC_BIT,           /* FCoE CRC32 */
>> >         NETIF_F_SCTP_CRC_BIT,           /* SCTP checksum offload */
>> > @@ -119,6 +120,7 @@ enum {
>> >  #define NETIF_F_GSO_UDP_TUNNEL __NETIF_F(GSO_UDP_TUNNEL)
>> >  #define NETIF_F_GSO_UDP_TUNNEL_CSUM __NETIF_F(GSO_UDP_TUNNEL_CSUM)
>> >  #define NETIF_F_GSO_TUNNEL_REMCSUM __NETIF_F(GSO_TUNNEL_REMCSUM)
>> > +#define NETIF_F_GSO_SCTP       __NETIF_F(GSO_SCTP)
>> >  #define NETIF_F_HW_VLAN_STAG_FILTER __NETIF_F(HW_VLAN_STAG_FILTER)
>> >  #define NETIF_F_HW_VLAN_STAG_RX        __NETIF_F(HW_VLAN_STAG_RX)
>> >  #define NETIF_F_HW_VLAN_STAG_TX        __NETIF_F(HW_VLAN_STAG_TX)
>> > @@ -144,7 +146,8 @@ enum {
>> >
>> >  /* List of features with software fallbacks. */
>> >  #define NETIF_F_GSO_SOFTWARE   (NETIF_F_TSO | NETIF_F_TSO_ECN | \
>> > -                                NETIF_F_TSO6 | NETIF_F_UFO)
>> > +                                NETIF_F_TSO6 | NETIF_F_UFO | \
>> > +                                NETIF_F_GSO_SCTP)
>> >
>> >  /* List of IP checksum features. Note that NETIF_F_ HW_CSUM should not be
>> >   * set in features when NETIF_F_IP_CSUM or NETIF_F_IPV6_CSUM are set--
>> > diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
>> > index 289c2314d76668b8357728382bb33d6828617458..ce14fab858bf96dd0f85aca237350c8d8317756e 100644
>> > --- a/include/linux/netdevice.h
>> > +++ b/include/linux/netdevice.h
>> > @@ -3928,6 +3928,7 @@ static inline bool net_gso_ok(netdev_features_t features, int gso_type)
>> >         BUILD_BUG_ON(SKB_GSO_UDP_TUNNEL != (NETIF_F_GSO_UDP_TUNNEL >> NETIF_F_GSO_SHIFT));
>> >         BUILD_BUG_ON(SKB_GSO_UDP_TUNNEL_CSUM != (NETIF_F_GSO_UDP_TUNNEL_CSUM >> NETIF_F_GSO_SHIFT));
>> >         BUILD_BUG_ON(SKB_GSO_TUNNEL_REMCSUM != (NETIF_F_GSO_TUNNEL_REMCSUM >> NETIF_F_GSO_SHIFT));
>> > +       BUILD_BUG_ON(SKB_GSO_SCTP    != (NETIF_F_GSO_SCTP >> NETIF_F_GSO_SHIFT));
>> >
>> >         return (features & feature) == feature;
>> >  }
>> > diff --git a/net/core/dev.c b/net/core/dev.c
>> > index 8cba3d852f251c503b193823b71b27aaef3fb3ae..9583284086967c0746de5f553535e25e125714a5 100644
>> > --- a/net/core/dev.c
>> > +++ b/net/core/dev.c
>> > @@ -2680,7 +2680,11 @@ EXPORT_SYMBOL(skb_mac_gso_segment);
>> >  static inline bool skb_needs_check(struct sk_buff *skb, bool tx_path)
>> >  {
>> >         if (tx_path)
>> > -               return skb->ip_summed != CHECKSUM_PARTIAL;
>> > +               /* FIXME: Why only packets with checksum offloading are
>> > +                * supported for GSO?
>> > +                */
>> > +               return skb->ip_summed != CHECKSUM_PARTIAL &&
>> > +                      skb->ip_summed != CHECKSUM_UNNECESSARY;
>> >         else
>> >                 return skb->ip_summed == CHECKSUM_NONE;
>> >  }
>>
>> Tom Herbert just got rid of the use of CHECKSUM_UNNECESSARY in the
>> transmit path a little while ago.  Please don't reintroduce it.
>
> Can you give me some pointers on that? I cannot find such change.
> skb_needs_check() seems to be like that since beginning.

Maybe you need to update your kernel.  All this stuff was changed in
December and has been this way for a little while now.

Commits:
7a6ae71b24905 "net: Elaborate on checksum offload interface description"
253aab0597d9e "fcoe: Use CHECKSUM_PARTIAL to indicate CRC offload"
53692b1de419c "sctp: Rename NETIF_F_SCTP_CSUM to NETIF_F_SCTP_CRC"

The main reason I even noticed it is because of some of the work I did
on the Intel NIC offloads.

>> > diff --git a/net/core/skbuff.c b/net/core/skbuff.c
>> > index 704b69682085dec77f3d0f990aaf0024afd705b9..96f223f8d769d2765fd64348830c76cb222906c8 100644
>> > --- a/net/core/skbuff.c
>> > +++ b/net/core/skbuff.c
>> > @@ -3017,8 +3017,16 @@ struct sk_buff *skb_segment(struct sk_buff *head_skb,
>> >                 int size;
>> >
>> >                 len = head_skb->len - offset;
>> > -               if (len > mss)
>> > -                       len = mss;
>> > +               if (len > mss) {
>> > +                       /* FIXME: A define is surely welcomed, but maybe
>> > +                        * shinfo->txflags is better for this flag, but
>> > +                        * we need to expand it then
>> > +                        */
>> > +                       if (mss == 1)
>> > +                               len = list_skb->len;
>> > +                       else
>> > +                               len = mss;
>> > +               }
>> >
>>
>> Using 0xFFFF here as a flag with the MSS value would likely be much
>> more readable.
>
> Either way it will be replaced by a define/name instead.

Yeah, that would probably be good.

>> >                 hsize = skb_headlen(head_skb) - offset;
>> >                 if (hsize < 0)
>> > diff --git a/net/ipv4/af_inet.c b/net/ipv4/af_inet.c
>> > index 5c5db6636704daa0c49fc13e84b2c5b282a44ed3..ec1c779bb664d1399d74f2bd7016e30b648ce47d 100644
>> > --- a/net/ipv4/af_inet.c
>> > +++ b/net/ipv4/af_inet.c
>> > @@ -1220,6 +1220,7 @@ static struct sk_buff *inet_gso_segment(struct sk_buff *skb,
>> >                        SKB_GSO_UDP_TUNNEL |
>> >                        SKB_GSO_UDP_TUNNEL_CSUM |
>> >                        SKB_GSO_TUNNEL_REMCSUM |
>> > +                      SKB_GSO_SCTP |
>> >                        0)))
>> >                 goto out;
>> >
>> > diff --git a/net/sctp/offload.c b/net/sctp/offload.c
>> > index 7080a6318da7110c1688dd0c5bb240356dbd0cd3..3b96035fa180a4e7195f7b6e7a8be7b97c8f8b26 100644
>> > --- a/net/sctp/offload.c
>> > +++ b/net/sctp/offload.c
>> > @@ -36,8 +36,61 @@
>> >  #include <net/sctp/checksum.h>
>> >  #include <net/protocol.h>
>> >
>> > +static __le32 sctp_gso_make_checksum(struct sk_buff *skb)
>> > +{
>> > +       skb->ip_summed = CHECKSUM_NONE;
>> > +       return sctp_compute_cksum(skb, skb_transport_offset(skb));
>> > +}
>> > +
>>
>> I really despise the naming of this bit here.  SCTP does not use a
>> checksum.  It uses a CRC.  Please don't call this a checksum as it
>> will just make the code really confusing.   I think the name should be
>> something like gso_make_crc32c.
>
> Agreed. SCTP code still references it as 'cksum'. I'll change that in
> another patch.
>
>> I think we need to address the CRC issues before we can really get
>> into segmentation.  Specifically we need to be able to offload SCTP
>> and FCoE in software since they both use the CHECKSUM_PARTIAL value
>> and then we can start cleaning up more of this mess and move onto
>> segmentation.
>
> Hm? The mess on CRC issues here is caused by this patch alone. It's good
> as it is today. And a good part of this mess is caused by trying to GSO
> without offloading CRC too.
>
> Or you mean that SCTP and FCoE should stop using CHECKSUM_* at all?

Well after Tom's change both SCTP and FCoE use CHECKSUM_PARTIAL.
CHECKSUM_PARTIAL is what is used to indicate to the hardware that a
checksum offload has been requested so that is what is looked for at
the driver level.

My concern with all this is that we should probably be looking at
coming up with a means of offloading this in software when
skb_checksum_help is called.  Right now validate_xmit_skb doesn't have
any understanding of what to do with SCTP or FCoE and will try to just
compute a checksum for them.

>> > +static struct sk_buff *sctp_gso_segment(struct sk_buff *skb,
>> > +                                       netdev_features_t features)
>> > +{
>> > +       struct sk_buff *segs = ERR_PTR(-EINVAL);
>> > +       struct sctphdr *sh;
>> > +
>> > +       sh = sctp_hdr(skb);
>> > +       if (!pskb_may_pull(skb, sizeof(*sh)))
>> > +               goto out;
>> > +
>> > +       __skb_pull(skb, sizeof(*sh));
>> > +
>> > +       if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
>> > +               /* Packet is from an untrusted source, reset gso_segs. */
>> > +               int type = skb_shinfo(skb)->gso_type;
>> > +
>> > +               if (unlikely(type &
>> > +                            ~(SKB_GSO_SCTP | SKB_GSO_DODGY |
>> > +                              0) ||
>> > +                            !(type & (SKB_GSO_SCTP))))
>> > +                       goto out;
>> > +
>> > +               /* This should not happen as no NIC has SCTP GSO
>> > +                * offloading, it's always via software and thus we
>> > +                * won't send a large packet down the stack.
>> > +                */
>> > +               WARN_ONCE(1, "SCTP segmentation offloading to NICs is not supported.");
>> > +               goto out;
>> > +       }
>> > +
>>
>> So what you are going to end up needing here is some way to tell the
>> hardware that you are doing the checksum no matter what.  There is no
>> value in you computing a 1's compliment checksum for the payload if
>> you aren't going to use it.  What you can probably do is just clear
>> the standard checksum flags and then OR in NETIF_F_HW_CSUM if
>> NETIF_F_SCTP_CRC is set and that should get skb_segment to skip
>> offloading the checksum.
>
> Interesting, ok
>
>> One other bit that will make this more complicated is if we ever get
>> around to supporting SCTP in tunnels.  Then we will need to sort out
>> how things like remote checksum offload should impact SCTP, and how to
>> deal with needing to compute both a CRC and 1's compliment checksum.
>> What we would probably need to do is check for encap_hdr_csum and if
>> it is set and we are doing SCTP then we would need to clear the
>> NETIF_F_HW_CSUM, NETIF_F_IP_CSUM, and NETIF_F_IPV6_CSUM flags.
>
> Yup. And that includes on storing pointers to where to store each of it.

Actually the pointers bit is easy.  The csum_start and csum_offset
values should be set up after you have segmented the skb and should be
updated after the skb has been segmented.  If nothing else you can
probably take a look at the TCP code tcp_gso_segment and
tcp4_gso_segment for inspiration.  Basically you need to make sure
that you set the ip_summed, csum_start, and csum_offset values for
your first frame before you start segmenting it into multiple frames.

>> > +       segs = skb_segment(skb, features);
>> > +       if (IS_ERR(segs))
>> > +               goto out;
>> > +
>> > +       /* All that is left is update SCTP CRC if necessary */
>> > +       for (skb = segs; skb; skb = skb->next) {
>> > +               if (skb->ip_summed != CHECKSUM_PARTIAL) {
>> > +                       sh = sctp_hdr(skb);
>> > +                       sh->checksum = sctp_gso_make_checksum(skb);
>> > +               }
>> > +       }
>> > +
>>
>> Okay, so it looks like you are doing the right thing here and leaving
>> this as CHECKSUM_PARTIAL.
>
> Actually no then. sctp_gso_make_checksum() replaces it:
> +static __le32 sctp_gso_make_checksum(struct sk_buff *skb)
> +{
> +       skb->ip_summed = CHECKSUM_NONE;
> +       return sctp_compute_cksum(skb, skb_transport_offset(skb));
>
> Why again would have to leave it as CHECKSUM_PARTIAL? IP header?

My earlier comment is actually incorrect.  This section is pretty much
broken since CHECKSUM_PARTIAL only reflects a 1's compliment checksum
in the case of skb_segment so whatever the value it is worthless.
CHECKSUM_PARTIAL is used to indicate if a given frame needs to be
offloaded.  It is meant to let the device know that it still needs to
compute a checksum or CRC beginning at csum_start and then storing the
new value at csum_offset.  However for skb_segment it is actually
referring to a 1's compliment checksum and if it returns CHECKSUM_NONE
it means it is stored in skb->csum which would really wreck things for
you since that was your skb->csum_start and skb->csum_offset values.
I have a patch to change this so that we update a checksum in the
SKB_GSO_CB, but I wasn't planning on submitting that until net-next
opens.

In the case of SCTP you probably don't even need to bother checking
the value since it is meaningless as skb_segment doesn't know how to
do an SCTP checksum anyway.  To that end for now what you could do is
just set NETIF_F_HW_CSUM.  This way skb_segment won't go and try to
compute a 1's compliment checksum on the payload since there is no
actual need for it.

One other bit you will need to do is to check the value of SCTP_CRC
outside of skb_segment.  You might look at how
__skb_udp_tunnel_segment does this to populate its own offload_csum
boolean value, though you would want to use features, not
skb->dev->features as that is a bit of a workaround since features is
stripped by hw_enc_features in some paths if I recall correctly.

Once the frames are segmented and if you don't support the offload you
could then call gso_make_crc32c() or whatever you want to name it to
perform the CRC calculation and populate the field.  One question by
the way.  Don't you need to initialize the checksum value to 0 before
you compute it?  I think you might have missed that step when you were
setting this up.

Powered by blists - more mailing lists