[ovs-dev] [PATCH v4 3/3] netdev-dpdk: Add TCP Segmentation Offload support

Ilya Maximets i.maximets at ovn.org
Fri Jan 17 17:58:56 UTC 2020


On 16.01.2020 18:00, Flavio Leitner wrote:
> Abbreviated as TSO, TCP Segmentation Offload is a feature which enables
> the network stack to delegate the TCP segmentation to the NIC reducing
> the per packet CPU overhead.
> 
> A guest using vhostuser interface with TSO enabled can send TCP packets
> much bigger than the MTU, which saves CPU cycles normally used to break
> the packets down to MTU size and to calculate checksums.
> 
> It also saves CPU cycles used to parse multiple packets/headers during
> the packet processing inside virtual switch.
> 
> If the destination of the packet is another guest in the same host, then
> the same big packet can be sent through a vhostuser interface skipping
> the segmentation completely. However, if the destination is not local,
> the NIC hardware is instructed to do the TCP segmentation and checksum
> calculation.
> 
> It is recommended to check if NIC hardware supports TSO before enabling
> the feature, which is off by default. For additional information please
> check the tso.rst document.
> 
> Signed-off-by: Flavio Leitner <fbl at sysclose.org>


I still didn't check computation of offsets and device configuration.
Few comments inline.

In general, I'd like if Ben will take a look to this patch, especially
at extensive netdev-linux code changes.

Also, suggesting to rename the patch to:
'userspace: Add TCP Segmentation Offload support.'

Best regards, Ilya maximets. 


> ---
>  Documentation/automake.mk              |   1 +
>  Documentation/topics/index.rst         |   1 +
>  Documentation/topics/userspace-tso.rst |  98 +++++++
>  NEWS                                   |   1 +
>  lib/automake.mk                        |   2 +
>  lib/conntrack.c                        |  29 +-
>  lib/dp-packet.h                        | 186 +++++++++++-
>  lib/ipf.c                              |  32 +-
>  lib/netdev-dpdk.c                      | 349 +++++++++++++++++++---
>  lib/netdev-linux-private.h             |   5 +
>  lib/netdev-linux.c                     | 386 ++++++++++++++++++++++---
>  lib/netdev-provider.h                  |   9 +
>  lib/netdev.c                           |  78 ++++-
>  lib/userspace-tso.c                    |  48 +++
>  lib/userspace-tso.h                    |  23 ++
>  vswitchd/bridge.c                      |   2 +
>  vswitchd/vswitch.xml                   |  17 ++
>  17 files changed, 1143 insertions(+), 124 deletions(-)
>  create mode 100644 Documentation/topics/userspace-tso.rst
>  create mode 100644 lib/userspace-tso.c
>  create mode 100644 lib/userspace-tso.h
> 
> Changelog:
> - v4
>  * rebased on top of master (recvmmsg)
>  * fixed URL in doc to point to 19.11
>  * renamed tso to userspace-tso
>  * renamed the option to userspace-tso-enable
>  * removed prototype that left over from v2
>  * fixed function style declaration
>  * renamed dp_packet_hwol_tx_ip_checksum to dp_packet_hwol_tx_ipv4_checksum
>  * dp_packet_hwol_tx_ipv4_checksum now checks for PKT_TX_IPV4.
>  * account for drops while preping the batch for TX.
>  * don't prep the batch for TX if TSO is disabled.
>  * simplified setsockopt error checking
>  * fixed af_packet_sock error checking to not call setsockopt on
>       closed sockets.
>  * fixed ol_flags comment.
>  * used VLOG_ERR_BUF() to pass error messages.
>  * fixed packet leak at netdev_send_prepare_batch()
>  * added a coverage counter to account drops while preparing a batch
>    at netdev.c
>  * fixed netdev_send() to not call ->send() if the batch is empty.
>  * fixed packet leak at netdev_push_header and account for the drops.
>  * removed DPDK requirement to enable userspace TSO support.
>  * fixed parameter documentation in vswitch.xml.
>  * renamed tso.rst to userspace-tso.rst and moved to topics/
>  * added comments documeting the functions in dp-packet.h
>  * fixed dp_packet_hwol_is_tso to check only PKT_TX_TCP_SEG
> 
> - v3
>  * Improved the documentation.
>  * Updated copyright year to 2020.
>  * TSO offloaded msg now includes the netdev's name.
>  * Added period at the end of all code comments.
>  * Warn and drop encapsulation of TSO packets.
>  * Fixed travis issue with restricted virtio types.
>  * Fixed double headroom allocation in dpdk_copy_dp_packet_to_mbuf()
>    which caused packet corruption.
>  * Fixed netdev_dpdk_prep_hwol_packet() to unconditionally set
>    PKT_TX_IP_CKSUM only for IPv4 packets.
> 
> 
> diff --git a/Documentation/automake.mk b/Documentation/automake.mk
> index f2ca17bad..22976a3cd 100644
> --- a/Documentation/automake.mk
> +++ b/Documentation/automake.mk
> @@ -57,6 +57,7 @@ DOC_SOURCE = \
>  	Documentation/topics/ovsdb-replication.rst \
>  	Documentation/topics/porting.rst \
>  	Documentation/topics/tracing.rst \
> +	Documentation/topics/userspace-tso.rst \
>  	Documentation/topics/windows.rst \
>  	Documentation/howto/index.rst \
>  	Documentation/howto/dpdk.rst \
> diff --git a/Documentation/topics/index.rst b/Documentation/topics/index.rst
> index 34c4b10e0..08af3a24d 100644
> --- a/Documentation/topics/index.rst
> +++ b/Documentation/topics/index.rst
> @@ -50,5 +50,6 @@ OVS
>     language-bindings
>     testing
>     tracing
> +   userspace-tso
>     idl-compound-indexes
>     ovs-extensions
> diff --git a/Documentation/topics/userspace-tso.rst b/Documentation/topics/userspace-tso.rst
> new file mode 100644
> index 000000000..893c64839
> --- /dev/null
> +++ b/Documentation/topics/userspace-tso.rst
> @@ -0,0 +1,98 @@
> +..
> +      Copyright 2020, Red Hat, Inc.
> +
> +      Licensed under the Apache License, Version 2.0 (the "License"); you may
> +      not use this file except in compliance with the License. You may obtain
> +      a copy of the License at
> +
> +          http://www.apache.org/licenses/LICENSE-2.0
> +
> +      Unless required by applicable law or agreed to in writing, software
> +      distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
> +      WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
> +      License for the specific language governing permissions and limitations
> +      under the License.
> +
> +      Convention for heading levels in Open vSwitch documentation:
> +
> +      =======  Heading 0 (reserved for the title in a document)
> +      -------  Heading 1
> +      ~~~~~~~  Heading 2
> +      +++++++  Heading 3
> +      '''''''  Heading 4
> +
> +      Avoid deeper levels because they do not render well.
> +
> +========================
> +Userspace Datapath - TSO
> +========================
> +
> +**Note:** This feature is considered experimental.
> +
> +TCP Segmentation Offload (TSO) enables a network stack to delegate segmentation
> +of an oversized TCP segment to the underlying physical NIC. Offload of frame
> +segmentation achieves computational savings in the core, freeing up CPU cycles
> +for more useful work.
> +
> +A common use case for TSO is when using virtualization, where traffic that's
> +coming in from a VM can offload the TCP segmentation, thus avoiding the
> +fragmentation in software. Additionally, if the traffic is headed to a VM
> +within the same host further optimization can be expected. As the traffic never
> +leaves the machine, no MTU needs to be accounted for, and thus no segmentation
> +and checksum calculations are required, which saves yet more cycles. Only when
> +the traffic actually leaves the host the segmentation needs to happen, in which
> +case it will be performed by the egress NIC. Consult your controller's
> +datasheet for compatibility. Secondly, the NIC must have an associated DPDK
> +Poll Mode Driver (PMD) which supports `TSO`. For a list of features per PMD,
> +refer to the `DPDK documentation`__.
> +
> +__ https://doc.dpdk.org/guides-19.11/nics/overview.html
> +
> +Enabling TSO
> +~~~~~~~~~~~~
> +
> +The TSO support may be enabled via a global config value
> +``userspace-tso-enable``.  Setting this to ``true`` enables TSO support for
> +all ports.
> +
> +    $ ovs-vsctl set Open_vSwitch . other_config:userspace-tso-enable=true
> +
> +The default value is ``false``.
> +
> +Changing ``userspace-tso-enable`` requires restarting the daemon.
> +
> +When using :doc:`vHost User ports <dpdk/vhost-user>`, TSO may be enabled
> +as follows.
> +
> +`TSO` is enabled in OvS by the DPDK vHost User backend; when a new guest
> +connection is established, `TSO` is thus advertised to the guest as an
> +available feature:
> +
> +QEMU Command Line Parameter::
> +
> +    $ sudo $QEMU_DIR/x86_64-softmmu/qemu-system-x86_64 \
> +    ...
> +    -device virtio-net-pci,mac=00:00:00:00:00:01,netdev=mynet1,\
> +    csum=on,guest_csum=on,guest_tso4=on,guest_tso6=on\
> +    ...
> +
> +2. Ethtool. Assuming that the guest's OS also supports `TSO`, ethtool can be
> +used to enable same::
> +
> +    $ ethtool -K eth0 sg on     # scatter-gather is a prerequisite for TSO
> +    $ ethtool -K eth0 tso on
> +    $ ethtool -k eth0
> +
> +~~~~~~~~~~~
> +Limitations
> +~~~~~~~~~~~
> +
> +The current OvS userspace `TSO` implementation supports flat and VLAN networks
> +only (i.e. no support for `TSO` over tunneled connection [VxLAN, GRE, IPinIP,
> +etc.]).
> +
> +There is no software implementation of TSO, so all ports attached to the
> +datapath must support TSO or packets using that feature will be dropped
> +on ports without TSO support.  That also means guests using vhost-user
> +in client mode will receive TSO packet regardless of TSO being enabled
> +or disabled within the guest.
> diff --git a/NEWS b/NEWS
> index e8d662a0c..586d81173 100644
> --- a/NEWS
> +++ b/NEWS
> @@ -26,6 +26,7 @@ Post-v2.12.0
>       * DPDK ring ports (dpdkr) are deprecated and will be removed in next
>         releases.
>       * Add support for DPDK 19.11.
> +     * Add experimental support for TSO.
>     - RSTP:
>       * The rstp_statistics column in Port table will only be updated every
>         stats-update-interval configured in Open_vSwtich table.
> diff --git a/lib/automake.mk b/lib/automake.mk
> index ebf714501..b80de9fc4 100644
> --- a/lib/automake.mk
> +++ b/lib/automake.mk
> @@ -304,6 +304,8 @@ lib_libopenvswitch_la_SOURCES = \
>  	lib/tnl-neigh-cache.h \
>  	lib/tnl-ports.c \
>  	lib/tnl-ports.h \
> +	lib/userspace-tso.c \
> +	lib/userspace-tso.h \

Alphabetical order is broken.
I guess, could be fixed while applying?

>  	lib/netdev-native-tnl.c \
>  	lib/netdev-native-tnl.h \
>  	lib/token-bucket.c \
> diff --git a/lib/conntrack.c b/lib/conntrack.c
> index b80080e72..742d2ad4f 100644
> --- a/lib/conntrack.c
> +++ b/lib/conntrack.c
> @@ -2022,7 +2022,8 @@ conn_key_extract(struct conntrack *ct, struct dp_packet *pkt, ovs_be16 dl_type,
>          if (hwol_bad_l3_csum) {
>              ok = false;
>          } else {
> -            bool hwol_good_l3_csum = dp_packet_ip_checksum_valid(pkt);
> +            bool hwol_good_l3_csum = dp_packet_ip_checksum_valid(pkt)
> +                                     || dp_packet_hwol_tx_ipv4_checksum(pkt);
>              /* Validate the checksum only when hwol is not supported. */
>              ok = extract_l3_ipv4(&ctx->key, l3, dp_packet_l3_size(pkt), NULL,
>                                   !hwol_good_l3_csum);
> @@ -2036,7 +2037,8 @@ conn_key_extract(struct conntrack *ct, struct dp_packet *pkt, ovs_be16 dl_type,
>      if (ok) {
>          bool hwol_bad_l4_csum = dp_packet_l4_checksum_bad(pkt);
>          if (!hwol_bad_l4_csum) {
> -            bool  hwol_good_l4_csum = dp_packet_l4_checksum_valid(pkt);
> +            bool  hwol_good_l4_csum = dp_packet_l4_checksum_valid(pkt)
> +                                      || dp_packet_hwol_tx_l4_checksum(pkt);
>              /* Validate the checksum only when hwol is not supported. */
>              if (extract_l4(&ctx->key, l4, dp_packet_l4_size(pkt),
>                             &ctx->icmp_related, l3, !hwol_good_l4_csum,
> @@ -3237,8 +3239,11 @@ handle_ftp_ctl(struct conntrack *ct, const struct conn_lookup_ctx *ctx,
>                  }
>                  if (seq_skew) {
>                      ip_len = ntohs(l3_hdr->ip_tot_len) + seq_skew;
> -                    l3_hdr->ip_csum = recalc_csum16(l3_hdr->ip_csum,
> -                                          l3_hdr->ip_tot_len, htons(ip_len));
> +                    if (!dp_packet_hwol_tx_ipv4_checksum(pkt)) {
> +                        l3_hdr->ip_csum = recalc_csum16(l3_hdr->ip_csum,
> +                                                        l3_hdr->ip_tot_len,
> +                                                        htons(ip_len));
> +                    }
>                      l3_hdr->ip_tot_len = htons(ip_len);
>                  }
>              }
> @@ -3256,13 +3261,15 @@ handle_ftp_ctl(struct conntrack *ct, const struct conn_lookup_ctx *ctx,
>      }
>  
>      th->tcp_csum = 0;
> -    if (ctx->key.dl_type == htons(ETH_TYPE_IPV6)) {
> -        th->tcp_csum = packet_csum_upperlayer6(nh6, th, ctx->key.nw_proto,
> -                           dp_packet_l4_size(pkt));
> -    } else {
> -        uint32_t tcp_csum = packet_csum_pseudoheader(l3_hdr);
> -        th->tcp_csum = csum_finish(
> -             csum_continue(tcp_csum, th, dp_packet_l4_size(pkt)));
> +    if (!dp_packet_hwol_tx_l4_checksum(pkt)) {
> +        if (ctx->key.dl_type == htons(ETH_TYPE_IPV6)) {
> +            th->tcp_csum = packet_csum_upperlayer6(nh6, th, ctx->key.nw_proto,
> +                               dp_packet_l4_size(pkt));
> +        } else {
> +            uint32_t tcp_csum = packet_csum_pseudoheader(l3_hdr);
> +            th->tcp_csum = csum_finish(
> +                 csum_continue(tcp_csum, th, dp_packet_l4_size(pkt)));
> +        }
>      }
>  
>      if (seq_skew) {
> diff --git a/lib/dp-packet.h b/lib/dp-packet.h
> index 133942155..3e995f505 100644
> --- a/lib/dp-packet.h
> +++ b/lib/dp-packet.h
> @@ -456,7 +456,7 @@ dp_packet_init_specific(struct dp_packet *p)
>  {
>      /* This initialization is needed for packets that do not come from DPDK
>       * interfaces, when vswitchd is built with --with-dpdk. */
> -    p->mbuf.tx_offload = p->mbuf.packet_type = 0;
> +    p->mbuf.ol_flags = p->mbuf.tx_offload = p->mbuf.packet_type = 0;

I'm not very comfortable with this change since dp_packet_init_specific()
always called after dp_packet_reset_offloads() that should clear ol_flags.
However, since we're not clearing non-offloading flags now, this is needed.

I'd concider this as a bug in current code that we're not clearing memory
layout related flags in mbuf.ol_flags for packets that comes from non-DPDK
ports. This needs to be fixed, maybe by a separate patch for previous
OVS branches. 

>      p->mbuf.nb_segs = 1;
>      p->mbuf.next = NULL;
>  }
> @@ -519,6 +519,96 @@ dp_packet_set_allocated(struct dp_packet *b, uint16_t s)
>      b->mbuf.buf_len = s;
>  }
>  
> +/* Return true if packet 'b' offloads TCP segmentation. */

/* Returns 'true' if packet 'b' marked for TCP segmentation offloading. */ ?

> +static inline bool
> +dp_packet_hwol_is_tso(const struct dp_packet *b)
> +{
> +    return !!(b->mbuf.ol_flags & PKT_TX_TCP_SEG);
> +}
> +
> +/* Return true if packet 'b' is IPv4. The flag is required when
> + * offload is requested. */

/* Returns 'true' if packet 'b' is marked for IPv4 checksum offloading. */ ?

> +static inline bool
> +dp_packet_hwol_is_ipv4(const struct dp_packet *b)
> +{
> +    return !!(b->mbuf.ol_flags & PKT_TX_IPV4);
> +}
> +
> +/* Return the L4 cksum offload bitmask. */
> +static inline uint64_t
> +dp_packet_hwol_l4_mask(const struct dp_packet *b)
> +{
> +    return b->mbuf.ol_flags & PKT_TX_L4_MASK;
> +}
> +
> +/* Return true if the packet 'b' offloads TCP checksum calculation. */

/* Returns 'true' if packet 'b' is marked for TCP checksum offloading. */ ?

> +static inline bool
> +dp_packet_hwol_l4_is_tcp(const struct dp_packet *b)
> +{
> +    return (b->mbuf.ol_flags & PKT_TX_L4_MASK) == PKT_TX_TCP_CKSUM;
> +}
> +
> +/* Return true if the packet 'b' offloads UDP checksum calculation. */

/* Returns 'true' if packet 'b' is marked for UDP checksum offloading. */ ?

> +static inline bool
> +dp_packet_hwol_l4_is_udp(struct dp_packet *b)
> +{
> +    return (b->mbuf.ol_flags & PKT_TX_L4_MASK) == PKT_TX_UDP_CKSUM;
> +}
> +
> +/* Return true if the packet 'b' offloads SCTP checksum calculation. */

/* Returns 'true' if packet 'b' is marked for SCTP checksum offloading. */ ?

> +static inline bool
> +dp_packet_hwol_l4_is_sctp(struct dp_packet *b)
> +{
> +    return (b->mbuf.ol_flags & PKT_TX_L4_MASK) == PKT_TX_SCTP_CKSUM;
> +}
> +
> +/* Flag the packet 'b' as IPv4 necessary when offload is used. */

/* Mark packet 'b' for IPv4 checksum offloading. */ ?

> +static inline void
> +dp_packet_hwol_set_tx_ipv4(struct dp_packet *b)
> +{
> +    b->mbuf.ol_flags |= PKT_TX_IPV4;
> +}
> +
> +/* Flag the packet 'b' as IPv6 necessary when offload is used. */

/* Mark packet 'b' for IPv6 checksum offloading. */ ?

> +static inline void
> +dp_packet_hwol_set_tx_ipv6(struct dp_packet *b)
> +{
> +    b->mbuf.ol_flags |= PKT_TX_IPV6;
> +}
> +
> +/* Request TCP checksum offload for packet 'b'. It implies that
> + * either the packet 'b' is flagged as IPv4 or IPv6. */

/* Mark packet 'b' for TCP checksum offloading.  It implies that either
 * the packet 'b' is marked for IPv4 or IPv6 checksum offloading. */ ?

> +static inline void
> +dp_packet_hwol_set_csum_tcp(struct dp_packet *b)
> +{
> +    b->mbuf.ol_flags |= PKT_TX_TCP_CKSUM;
> +}
> +
> +/* Request UDP checksum offload for packet 'b'. It implies that
> + * either the packet 'b' is flagged as IPv4 or IPv6. */

/* Mark packet 'b' for UDP checksum offloading.  It implies that either
 * the packet 'b' is marked for IPv4 or IPv6 checksum offloading. */ ?

> +static inline void
> +dp_packet_hwol_set_csum_udp(struct dp_packet *b)
> +{
> +    b->mbuf.ol_flags |= PKT_TX_UDP_CKSUM;
> +}
> +
> +/* Request SCTP checksum offload for packet 'b'. It implies that
> + * either the packet 'b' is flagged as IPv4 or IPv6. */

/* Mark packet 'b' for SCTP checksum offloading.  It implies that either
 * the packet 'b' is marked for IPv4 or IPv6 checksum offloading. */ ?

> +static inline void
> +dp_packet_hwol_set_csum_sctp(struct dp_packet *b)
> +{
> +    b->mbuf.ol_flags |= PKT_TX_SCTP_CKSUM;
> +}
> +
> +/* Request TCP segmentation offload for packet 'b'. It implies that
> + * either the packet 'b' is flagged as IPv4 or IPv6 and also implies
> + * that TCP checksum offload is flagged. */

/* Mark packet 'b' for TCP segmentation offloading.  It implies that
 * either the packet 'b' is marked for IPv4 or IPv6 checksum offloading
 * and also for TCP checksum offloading. */ ?

> +static inline void
> +dp_packet_hwol_set_tcp_seg(struct dp_packet *b)
> +{
> +    b->mbuf.ol_flags |= PKT_TX_TCP_SEG;
> +}
> +
>  /* Returns the RSS hash of the packet 'p'.  Note that the returned value is
>   * correct only if 'dp_packet_rss_valid(p)' returns true */
>  static inline uint32_t
> @@ -648,6 +738,84 @@ dp_packet_set_allocated(struct dp_packet *b, uint16_t s)
>      b->allocated_ = s;
>  }
>  
> +/* There are no implementation when not DPDK enabled datapath. */
> +static inline bool
> +dp_packet_hwol_is_tso(const struct dp_packet *b OVS_UNUSED)
> +{
> +    return false;
> +}
> +
> +/* There are no implementation when not DPDK enabled datapath. */
> +static inline bool
> +dp_packet_hwol_is_ipv4(const struct dp_packet *b OVS_UNUSED)
> +{
> +    return false;
> +}
> +
> +/* There are no implementation when not DPDK enabled datapath. */
> +static inline uint64_t
> +dp_packet_hwol_l4_mask(const struct dp_packet *b OVS_UNUSED)
> +{
> +    return 0;
> +}
> +
> +/* There are no implementation when not DPDK enabled datapath. */
> +static inline bool
> +dp_packet_hwol_l4_is_tcp(const struct dp_packet *b OVS_UNUSED)
> +{
> +    return false;
> +}
> +
> +/* There are no implementation when not DPDK enabled datapath. */
> +static inline bool
> +dp_packet_hwol_l4_is_udp(const struct dp_packet *b OVS_UNUSED)
> +{
> +    return false;
> +}
> +
> +/* There are no implementation when not DPDK enabled datapath. */
> +static inline bool
> +dp_packet_hwol_l4_is_sctp(const struct dp_packet *b OVS_UNUSED)
> +{
> +    return false;
> +}
> +
> +/* There are no implementation when not DPDK enabled datapath. */
> +static inline void
> +dp_packet_hwol_set_tx_ipv4(struct dp_packet *b OVS_UNUSED)
> +{
> +}
> +
> +/* There are no implementation when not DPDK enabled datapath. */
> +static inline void
> +dp_packet_hwol_set_tx_ipv6(struct dp_packet *b OVS_UNUSED)
> +{
> +}
> +
> +/* There are no implementation when not DPDK enabled datapath. */
> +static inline void
> +dp_packet_hwol_set_csum_tcp(struct dp_packet *b OVS_UNUSED)
> +{
> +}
> +
> +/* There are no implementation when not DPDK enabled datapath. */
> +static inline void
> +dp_packet_hwol_set_csum_udp(struct dp_packet *b OVS_UNUSED)
> +{
> +}
> +
> +/* There are no implementation when not DPDK enabled datapath. */
> +static inline void
> +dp_packet_hwol_set_csum_sctp(struct dp_packet *b OVS_UNUSED)
> +{
> +}
> +
> +/* There are no implementation when not DPDK enabled datapath. */
> +static inline void
> +dp_packet_hwol_set_tcp_seg(struct dp_packet *b OVS_UNUSED)
> +{
> +}
> +
>  /* Returns the RSS hash of the packet 'p'.  Note that the returned value is
>   * correct only if 'dp_packet_rss_valid(p)' returns true */
>  static inline uint32_t
> @@ -939,6 +1107,22 @@ dp_packet_batch_reset_cutlen(struct dp_packet_batch *batch)
>      }
>  }
>  
> +/* Return true if the packet 'b' requested IPv4 checksum offload. */
> +static inline bool
> +dp_packet_hwol_tx_ipv4_checksum(const struct dp_packet *b)

Why we need this function that just calls another publically available
function without any additional processing?

> +{
> +
> +    return !!dp_packet_hwol_is_ipv4(b);

'dp_packet_hwol_is_ipv4()' returns boolean, no need to !!.

> +}
> +
> +/* Return true if the packet 'b' requested L4 checksum offload. */
> +static inline bool
> +dp_packet_hwol_tx_l4_checksum(const struct dp_packet *b)
> +{
> +

Redundant empty line.

> +    return !!dp_packet_hwol_l4_mask(b);
> +}
> +
>  #ifdef  __cplusplus
>  }
>  #endif
> diff --git a/lib/ipf.c b/lib/ipf.c
> index 45c489122..14df04374 100644
> --- a/lib/ipf.c
> +++ b/lib/ipf.c
> @@ -433,9 +433,11 @@ ipf_reassemble_v4_frags(struct ipf_list *ipf_list)
>      len += rest_len;
>      l3 = dp_packet_l3(pkt);
>      ovs_be16 new_ip_frag_off = l3->ip_frag_off & ~htons(IP_MORE_FRAGMENTS);
> -    l3->ip_csum = recalc_csum16(l3->ip_csum, l3->ip_frag_off,
> -                                new_ip_frag_off);
> -    l3->ip_csum = recalc_csum16(l3->ip_csum, l3->ip_tot_len, htons(len));
> +    if (!dp_packet_hwol_tx_ipv4_checksum(pkt)) {
> +        l3->ip_csum = recalc_csum16(l3->ip_csum, l3->ip_frag_off,
> +                                    new_ip_frag_off);
> +        l3->ip_csum = recalc_csum16(l3->ip_csum, l3->ip_tot_len, htons(len));
> +    }
>      l3->ip_tot_len = htons(len);
>      l3->ip_frag_off = new_ip_frag_off;
>      dp_packet_set_l2_pad_size(pkt, 0);
> @@ -606,6 +608,7 @@ ipf_is_valid_v4_frag(struct ipf *ipf, struct dp_packet *pkt)
>      }
>  
>      if (OVS_UNLIKELY(!dp_packet_ip_checksum_valid(pkt)
> +                     && !dp_packet_hwol_tx_ipv4_checksum(pkt)
>                       && csum(l3, ip_hdr_len) != 0)) {
>          goto invalid_pkt;
>      }
> @@ -1181,16 +1184,21 @@ ipf_post_execute_reass_pkts(struct ipf *ipf,
>                  } else {
>                      struct ip_header *l3_frag = dp_packet_l3(frag_0->pkt);
>                      struct ip_header *l3_reass = dp_packet_l3(pkt);
> -                    ovs_be32 reass_ip = get_16aligned_be32(&l3_reass->ip_src);
> -                    ovs_be32 frag_ip = get_16aligned_be32(&l3_frag->ip_src);
> -                    l3_frag->ip_csum = recalc_csum32(l3_frag->ip_csum,
> -                                                     frag_ip, reass_ip);
> -                    l3_frag->ip_src = l3_reass->ip_src;
> +                    if (!dp_packet_hwol_tx_ipv4_checksum(frag_0->pkt)) {
> +                        ovs_be32 reass_ip =
> +                            get_16aligned_be32(&l3_reass->ip_src);
> +                        ovs_be32 frag_ip =
> +                            get_16aligned_be32(&l3_frag->ip_src);
> +
> +                        l3_frag->ip_csum = recalc_csum32(l3_frag->ip_csum,
> +                                                         frag_ip, reass_ip);
> +                        reass_ip = get_16aligned_be32(&l3_reass->ip_dst);
> +                        frag_ip = get_16aligned_be32(&l3_frag->ip_dst);
> +                        l3_frag->ip_csum = recalc_csum32(l3_frag->ip_csum,
> +                                                         frag_ip, reass_ip);
> +                    }
>  
> -                    reass_ip = get_16aligned_be32(&l3_reass->ip_dst);
> -                    frag_ip = get_16aligned_be32(&l3_frag->ip_dst);
> -                    l3_frag->ip_csum = recalc_csum32(l3_frag->ip_csum,
> -                                                     frag_ip, reass_ip);
> +                    l3_frag->ip_src = l3_reass->ip_src;
>                      l3_frag->ip_dst = l3_reass->ip_dst;
>                  }
>  
> diff --git a/lib/netdev-dpdk.c b/lib/netdev-dpdk.c
> index d1469f6f2..48fd6c184 100644
> --- a/lib/netdev-dpdk.c
> +++ b/lib/netdev-dpdk.c
> @@ -70,6 +70,7 @@
>  #include "smap.h"
>  #include "sset.h"
>  #include "timeval.h"
> +#include "userspace-tso.h"

Slightly out of alphabetical order.

>  #include "unaligned.h"
>  #include "unixctl.h"
>  #include "util.h"
> @@ -201,6 +202,8 @@ struct netdev_dpdk_sw_stats {
>      uint64_t tx_qos_drops;
>      /* Packet drops in ingress policer processing. */
>      uint64_t rx_qos_drops;
> +    /* Packet drops in HWOL processing */

Period at the end of comment.

> +    uint64_t tx_invalid_hwol_drops;
>  };
>  
>  enum { DPDK_RING_SIZE = 256 };
> @@ -410,7 +413,8 @@ struct ingress_policer {
>  enum dpdk_hw_ol_features {
>      NETDEV_RX_CHECKSUM_OFFLOAD = 1 << 0,
>      NETDEV_RX_HW_CRC_STRIP = 1 << 1,
> -    NETDEV_RX_HW_SCATTER = 1 << 2
> +    NETDEV_RX_HW_SCATTER = 1 << 2,
> +    NETDEV_TX_TSO_OFFLOAD = 1 << 3,
>  };
>  
>  /*
> @@ -992,6 +996,12 @@ dpdk_eth_dev_port_config(struct netdev_dpdk *dev, int n_rxq, int n_txq)
>          conf.rxmode.offloads |= DEV_RX_OFFLOAD_KEEP_CRC;
>      }
>  
> +    if (dev->hw_ol_features & NETDEV_TX_TSO_OFFLOAD) {
> +        conf.txmode.offloads |= DEV_TX_OFFLOAD_TCP_TSO;
> +        conf.txmode.offloads |= DEV_TX_OFFLOAD_TCP_CKSUM;
> +        conf.txmode.offloads |= DEV_TX_OFFLOAD_IPV4_CKSUM;
> +    }
> +
>      /* Limit configured rss hash functions to only those supported
>       * by the eth device. */
>      conf.rx_adv_conf.rss_conf.rss_hf &= info.flow_type_rss_offloads;
> @@ -1093,6 +1103,9 @@ dpdk_eth_dev_init(struct netdev_dpdk *dev)
>      uint32_t rx_chksm_offload_capa = DEV_RX_OFFLOAD_UDP_CKSUM |
>                                       DEV_RX_OFFLOAD_TCP_CKSUM |
>                                       DEV_RX_OFFLOAD_IPV4_CKSUM;
> +    uint32_t tx_tso_offload_capa = DEV_TX_OFFLOAD_TCP_TSO |
> +                                   DEV_TX_OFFLOAD_TCP_CKSUM |
> +                                   DEV_TX_OFFLOAD_IPV4_CKSUM;
>  
>      rte_eth_dev_info_get(dev->port_id, &info);
>  
> @@ -1119,6 +1132,14 @@ dpdk_eth_dev_init(struct netdev_dpdk *dev)
>          dev->hw_ol_features &= ~NETDEV_RX_HW_SCATTER;
>      }
>  
> +    if (info.tx_offload_capa & tx_tso_offload_capa) {
> +        dev->hw_ol_features |= NETDEV_TX_TSO_OFFLOAD;
> +    } else {
> +        dev->hw_ol_features &= ~NETDEV_TX_TSO_OFFLOAD;
> +        VLOG_WARN("Tx TSO offload is not supported on %s port "
> +                  DPDK_PORT_ID_FMT, netdev_get_name(&dev->up), dev->port_id);
> +    }
> +
>      n_rxq = MIN(info.max_rx_queues, dev->up.n_rxq);
>      n_txq = MIN(info.max_tx_queues, dev->up.n_txq);
>  
> @@ -1369,14 +1390,16 @@ netdev_dpdk_vhost_construct(struct netdev *netdev)
>          goto out;
>      }
>  
> -    err = rte_vhost_driver_disable_features(dev->vhost_id,
> -                                1ULL << VIRTIO_NET_F_HOST_TSO4
> -                                | 1ULL << VIRTIO_NET_F_HOST_TSO6
> -                                | 1ULL << VIRTIO_NET_F_CSUM);
> -    if (err) {
> -        VLOG_ERR("rte_vhost_driver_disable_features failed for vhost user "
> -                 "port: %s\n", name);
> -        goto out;
> +    if (!userspace_tso_enabled()) {
> +        err = rte_vhost_driver_disable_features(dev->vhost_id,
> +                                    1ULL << VIRTIO_NET_F_HOST_TSO4
> +                                    | 1ULL << VIRTIO_NET_F_HOST_TSO6
> +                                    | 1ULL << VIRTIO_NET_F_CSUM);
> +        if (err) {
> +            VLOG_ERR("rte_vhost_driver_disable_features failed for vhost user "
> +                     "port: %s\n", name);
> +            goto out;
> +        }
>      }
>  
>      err = rte_vhost_driver_start(dev->vhost_id);
> @@ -1711,6 +1734,11 @@ netdev_dpdk_get_config(const struct netdev *netdev, struct smap *args)
>          } else {
>              smap_add(args, "rx_csum_offload", "false");
>          }
> +        if (dev->hw_ol_features & NETDEV_TX_TSO_OFFLOAD) {
> +            smap_add(args, "tx_tso_offload", "true");
> +        } else {
> +            smap_add(args, "tx_tso_offload", "false");
> +        }
>          smap_add(args, "lsc_interrupt_mode",
>                   dev->lsc_interrupt_mode ? "true" : "false");
>      }
> @@ -2138,6 +2166,67 @@ netdev_dpdk_rxq_dealloc(struct netdev_rxq *rxq)
>      rte_free(rx);
>  }
>  
> +/* Prepare the packet for HWOL.
> + * Return True if the packet is OK to continue. */
> +static bool
> +netdev_dpdk_prep_hwol_packet(struct netdev_dpdk *dev, struct rte_mbuf *mbuf)
> +{
> +    struct dp_packet *pkt = CONTAINER_OF(mbuf, struct dp_packet, mbuf);
> +
> +    if (mbuf->ol_flags & PKT_TX_L4_MASK) {
> +        mbuf->l2_len = (char *)dp_packet_l3(pkt) - (char *)dp_packet_eth(pkt);
> +        mbuf->l3_len = (char *)dp_packet_l4(pkt) - (char *)dp_packet_l3(pkt);
> +        mbuf->outer_l2_len = 0;
> +        mbuf->outer_l3_len = 0;
> +    }
> +
> +    if (mbuf->ol_flags & PKT_TX_TCP_SEG) {
> +        struct tcp_header *th = dp_packet_l4(pkt);
> +
> +        if (!th) {
> +            VLOG_WARN_RL(&rl, "%s: TCP Segmentation without L4 header"
> +                         " pkt len: %"PRIu32"", dev->up.name, mbuf->pkt_len);
> +            return false;
> +        }
> +
> +        mbuf->l4_len = TCP_OFFSET(th->tcp_ctl) * 4;
> +        mbuf->ol_flags |= PKT_TX_TCP_CKSUM;
> +        mbuf->tso_segsz = dev->mtu - mbuf->l3_len - mbuf->l4_len;
> +
> +        if (mbuf->ol_flags & PKT_TX_IPV4) {
> +            mbuf->ol_flags |= PKT_TX_IP_CKSUM;
> +        }
> +    }
> +    return true;
> +}
> +
> +/* Prepare a batch for HWOL.
> + * Return the number of good packets in the batch. */
> +static int
> +netdev_dpdk_prep_hwol_batch(struct netdev_dpdk *dev, struct rte_mbuf **pkts,
> +                            int pkt_cnt)
> +{
> +    int i = 0;
> +    int cnt = 0;
> +    struct rte_mbuf *pkt;
> +
> +    /* Prepare and filter bad HWOL packets. */
> +    for (i = 0; i < pkt_cnt; i++) {
> +        pkt = pkts[i];
> +        if (!netdev_dpdk_prep_hwol_packet(dev, pkt)) {
> +            rte_pktmbuf_free(pkt);
> +            continue;
> +        }
> +
> +        if (OVS_UNLIKELY(i != cnt)) {
> +            pkts[cnt] = pkt;
> +        }
> +        cnt++;
> +    }
> +
> +    return cnt;
> +}
> +
>  /* Tries to transmit 'pkts' to txq 'qid' of device 'dev'.  Takes ownership of
>   * 'pkts', even in case of failure.
>   *
> @@ -2147,11 +2236,22 @@ netdev_dpdk_eth_tx_burst(struct netdev_dpdk *dev, int qid,
>                           struct rte_mbuf **pkts, int cnt)
>  {
>      uint32_t nb_tx = 0;
> +    uint16_t nb_tx_prep = cnt;
> +
> +    if (userspace_tso_enabled()) {
> +        nb_tx_prep = rte_eth_tx_prepare(dev->port_id, qid, pkts, cnt);
> +        if (nb_tx_prep != cnt) {
> +            VLOG_WARN_RL(&rl, "%s: Output batch contains invalid packets. "
> +                         "Only %u/%u are valid: %s", dev->up.name, nb_tx_prep,
> +                         cnt, rte_strerror(rte_errno));
> +        }
> +    }
>  
> -    while (nb_tx != cnt) {
> +    while (nb_tx != nb_tx_prep) {
>          uint32_t ret;
>  
> -        ret = rte_eth_tx_burst(dev->port_id, qid, pkts + nb_tx, cnt - nb_tx);
> +        ret = rte_eth_tx_burst(dev->port_id, qid, pkts + nb_tx,
> +                               nb_tx_prep - nb_tx);
>          if (!ret) {
>              break;
>          }
> @@ -2437,11 +2537,14 @@ netdev_dpdk_filter_packet_len(struct netdev_dpdk *dev, struct rte_mbuf **pkts,
>      int cnt = 0;
>      struct rte_mbuf *pkt;
>  
> +    /* Filter oversized packets, unless are marked for TSO. */
>      for (i = 0; i < pkt_cnt; i++) {
>          pkt = pkts[i];
> -        if (OVS_UNLIKELY(pkt->pkt_len > dev->max_packet_len)) {
> -            VLOG_WARN_RL(&rl, "%s: Too big size %" PRIu32 " max_packet_len %d",
> -                         dev->up.name, pkt->pkt_len, dev->max_packet_len);
> +        if (OVS_UNLIKELY((pkt->pkt_len > dev->max_packet_len)
> +            && !(pkt->ol_flags & PKT_TX_TCP_SEG))) {
> +            VLOG_WARN_RL(&rl, "%s: Too big size %" PRIu32 " "
> +                         "max_packet_len %d", dev->up.name, pkt->pkt_len,
> +                         dev->max_packet_len);
>              rte_pktmbuf_free(pkt);
>              continue;
>          }
> @@ -2463,7 +2566,8 @@ netdev_dpdk_vhost_update_tx_counters(struct netdev_dpdk *dev,
>  {
>      int dropped = sw_stats_add->tx_mtu_exceeded_drops +
>                    sw_stats_add->tx_qos_drops +
> -                  sw_stats_add->tx_failure_drops;
> +                  sw_stats_add->tx_failure_drops +
> +                  sw_stats_add->tx_invalid_hwol_drops;
>      struct netdev_stats *stats = &dev->stats;
>      int sent = attempted - dropped;
>      int i;
> @@ -2482,6 +2586,7 @@ netdev_dpdk_vhost_update_tx_counters(struct netdev_dpdk *dev,
>          sw_stats->tx_failure_drops      += sw_stats_add->tx_failure_drops;
>          sw_stats->tx_mtu_exceeded_drops += sw_stats_add->tx_mtu_exceeded_drops;
>          sw_stats->tx_qos_drops          += sw_stats_add->tx_qos_drops;
> +        sw_stats->tx_invalid_hwol_drops += sw_stats_add->tx_invalid_hwol_drops;
>      }
>  }
>  
> @@ -2513,8 +2618,15 @@ __netdev_dpdk_vhost_send(struct netdev *netdev, int qid,
>          rte_spinlock_lock(&dev->tx_q[qid].tx_lock);
>      }
>  
> +    sw_stats_add.tx_invalid_hwol_drops = cnt;
> +    if (userspace_tso_enabled()) {
> +        cnt = netdev_dpdk_prep_hwol_batch(dev, cur_pkts, cnt);
> +    }
> +
> +    sw_stats_add.tx_invalid_hwol_drops -= cnt;
> +    sw_stats_add.tx_mtu_exceeded_drops = cnt;
>      cnt = netdev_dpdk_filter_packet_len(dev, cur_pkts, cnt);
> -    sw_stats_add.tx_mtu_exceeded_drops = total_packets - cnt;
> +    sw_stats_add.tx_mtu_exceeded_drops -= cnt;
>  
>      /* Check has QoS has been configured for the netdev */
>      sw_stats_add.tx_qos_drops = cnt;
> @@ -2562,6 +2674,121 @@ out:
>      }
>  }
>  
> +static void
> +netdev_dpdk_extbuf_free(void *addr OVS_UNUSED, void *opaque)
> +{
> +    rte_free(opaque);
> +}
> +
> +static struct rte_mbuf *
> +dpdk_pktmbuf_attach_extbuf(struct rte_mbuf *pkt, uint32_t data_len)
> +{
> +    uint32_t total_len = RTE_PKTMBUF_HEADROOM + data_len;
> +    struct rte_mbuf_ext_shared_info *shinfo = NULL;
> +    uint16_t buf_len;
> +    void *buf;
> +
> +    if (rte_pktmbuf_tailroom(pkt) >= sizeof(*shinfo)) {

Please, don't parenthesize argument of sizeof if it's a variable.

> +        shinfo = rte_pktmbuf_mtod(pkt, struct rte_mbuf_ext_shared_info *);
> +    } else {
> +        total_len += sizeof(*shinfo) + sizeof(uintptr_t);

Ditto.

> +        total_len = RTE_ALIGN_CEIL(total_len, sizeof(uintptr_t));
> +    }
> +
> +    if (unlikely(total_len > UINT16_MAX)) {
> +        VLOG_ERR("Can't copy packet: too big %u", total_len);
> +        return NULL;
> +    }
> +
> +    buf_len = total_len;
> +    buf = rte_malloc(NULL, buf_len, RTE_CACHE_LINE_SIZE);
> +    if (unlikely(buf == NULL)) {

OVS_UNLIKELY

> +        VLOG_ERR("Failed to allocate memory using rte_malloc: %u", buf_len);
> +        return NULL;
> +    }
> +
> +    /* Initialize shinfo. */
> +    if (shinfo) {
> +        shinfo->free_cb = netdev_dpdk_extbuf_free;
> +        shinfo->fcb_opaque = buf;
> +        rte_mbuf_ext_refcnt_set(shinfo, 1);
> +    } else {
> +        shinfo = rte_pktmbuf_ext_shinfo_init_helper(buf, &buf_len,
> +                                                    netdev_dpdk_extbuf_free,
> +                                                    buf);
> +        if (unlikely(shinfo == NULL)) {

OVS_UNLIKELY

> +            rte_free(buf);
> +            VLOG_ERR("Failed to initialize shared info for mbuf while "
> +                     "attempting to attach an external buffer.");
> +            return NULL;
> +        }
> +    }
> +
> +    rte_pktmbuf_attach_extbuf(pkt, buf, rte_malloc_virt2iova(buf), buf_len,
> +                              shinfo);
> +    rte_pktmbuf_reset_headroom(pkt);
> +
> +    return pkt;
> +}
> +
> +static struct rte_mbuf *
> +dpdk_pktmbuf_alloc(struct rte_mempool *mp, uint32_t data_len)
> +{
> +    struct rte_mbuf *pkt = rte_pktmbuf_alloc(mp);
> +
> +    if (OVS_UNLIKELY(!pkt)) {
> +        return NULL;
> +    }
> +
> +    dp_packet_init_specific((struct dp_packet *)pkt);

Why this needed?  rte_pktmbuf_alloc() always resets mbuf clearing
all the required fields.

> +    if (rte_pktmbuf_tailroom(pkt) >= data_len) {
> +        return pkt;
> +    }
> +
> +    if (dpdk_pktmbuf_attach_extbuf(pkt, data_len)) {
> +        return pkt;
> +    }
> +
> +    rte_pktmbuf_free(pkt);
> +
> +    return NULL;
> +}
> +
> +static struct dp_packet *
> +dpdk_copy_dp_packet_to_mbuf(struct rte_mempool *mp, struct dp_packet *pkt_orig)
> +{
> +    struct rte_mbuf *mbuf_dest;
> +    struct dp_packet *pkt_dest;
> +    uint32_t pkt_len;
> +
> +    pkt_len = dp_packet_size(pkt_orig);
> +    mbuf_dest = dpdk_pktmbuf_alloc(mp, pkt_len);
> +    if (OVS_UNLIKELY(mbuf_dest == NULL)) {
> +            return NULL;
> +    }
> +
> +    pkt_dest = CONTAINER_OF(mbuf_dest, struct dp_packet, mbuf);
> +    memcpy(dp_packet_data(pkt_dest), dp_packet_data(pkt_orig), pkt_len);
> +    dp_packet_set_size(pkt_dest, pkt_len);
> +
> +    mbuf_dest->tx_offload = pkt_orig->mbuf.tx_offload;
> +    mbuf_dest->packet_type = pkt_orig->mbuf.packet_type;
> +    mbuf_dest->ol_flags |= (pkt_orig->mbuf.ol_flags &
> +                            ~(EXT_ATTACHED_MBUF | IND_ATTACHED_MBUF));
> +
> +    memcpy(&pkt_dest->l2_pad_size, &pkt_orig->l2_pad_size,
> +           sizeof(struct dp_packet) - offsetof(struct dp_packet, l2_pad_size));


Above looks like a code duplication with dp_packet_clone_with_headroom().
It might be better to strip common parts of dp_packet_clone_with_headroom()
to a separate function and call it here.  However this might be done as
a separate patch later.

> +
> +    if (mbuf_dest->ol_flags & PKT_TX_L4_MASK) {
> +        mbuf_dest->l2_len = (char *)dp_packet_l3(pkt_dest)
> +                                - (char *)dp_packet_eth(pkt_dest);
> +        mbuf_dest->l3_len = (char *)dp_packet_l4(pkt_dest)
> +                                - (char *) dp_packet_l3(pkt_dest);
> +    }
> +
> +    return pkt_dest;
> +}
> +
>  /* Tx function. Transmit packets indefinitely */
>  static void
>  dpdk_do_tx_copy(struct netdev *netdev, int qid, struct dp_packet_batch *batch)
> @@ -2575,7 +2802,7 @@ dpdk_do_tx_copy(struct netdev *netdev, int qid, struct dp_packet_batch *batch)
>      enum { PKT_ARRAY_SIZE = NETDEV_MAX_BURST };
>  #endif
>      struct netdev_dpdk *dev = netdev_dpdk_cast(netdev);
> -    struct rte_mbuf *pkts[PKT_ARRAY_SIZE];
> +    struct dp_packet *pkts[PKT_ARRAY_SIZE];
>      struct netdev_dpdk_sw_stats *sw_stats = dev->sw_stats;
>      uint32_t cnt = batch_cnt;
>      uint32_t dropped = 0;
> @@ -2596,34 +2823,30 @@ dpdk_do_tx_copy(struct netdev *netdev, int qid, struct dp_packet_batch *batch)
>          struct dp_packet *packet = batch->packets[i];
>          uint32_t size = dp_packet_size(packet);
>  
> -        if (OVS_UNLIKELY(size > dev->max_packet_len)) {
> -            VLOG_WARN_RL(&rl, "Too big size %u max_packet_len %d",
> -                         size, dev->max_packet_len);
> -
> +        if (size > dev->max_packet_len
> +            && !(packet->mbuf.ol_flags & PKT_TX_TCP_SEG)) {
> +            VLOG_WARN_RL(&rl, "Too big size %u max_packet_len %d", size,
> +                         dev->max_packet_len);
>              mtu_drops++;
>              continue;
>          }
>  
> -        pkts[txcnt] = rte_pktmbuf_alloc(dev->dpdk_mp->mp);
> +        pkts[txcnt] = dpdk_copy_dp_packet_to_mbuf(dev->dpdk_mp->mp, packet);
>          if (OVS_UNLIKELY(!pkts[txcnt])) {
>              dropped = cnt - i;
>              break;
>          }
>  
> -        /* We have to do a copy for now */
> -        memcpy(rte_pktmbuf_mtod(pkts[txcnt], void *),
> -               dp_packet_data(packet), size);
> -        dp_packet_set_size((struct dp_packet *)pkts[txcnt], size);
> -
>          txcnt++;
>      }
>  
>      if (OVS_LIKELY(txcnt)) {
>          if (dev->type == DPDK_DEV_VHOST) {
> -            __netdev_dpdk_vhost_send(netdev, qid, (struct dp_packet **) pkts,
> -                                     txcnt);
> +            __netdev_dpdk_vhost_send(netdev, qid, pkts, txcnt);
>          } else {
> -            tx_failure = netdev_dpdk_eth_tx_burst(dev, qid, pkts, txcnt);
> +            tx_failure += netdev_dpdk_eth_tx_burst(dev, qid,
> +                                                   (struct rte_mbuf **)pkts,
> +                                                   txcnt);
>          }
>      }
>  
> @@ -2676,26 +2899,33 @@ netdev_dpdk_send__(struct netdev_dpdk *dev, int qid,
>          dp_packet_delete_batch(batch, true);
>      } else {
>          struct netdev_dpdk_sw_stats *sw_stats = dev->sw_stats;
> -        int tx_cnt, dropped;
> -        int tx_failure, mtu_drops, qos_drops;
> +        int dropped;
> +        int tx_failure, mtu_drops, qos_drops, hwol_drops;
>          int batch_cnt = dp_packet_batch_size(batch);
>          struct rte_mbuf **pkts = (struct rte_mbuf **) batch->packets;
>  
> -        tx_cnt = netdev_dpdk_filter_packet_len(dev, pkts, batch_cnt);
> -        mtu_drops = batch_cnt - tx_cnt;
> -        qos_drops = tx_cnt;
> -        tx_cnt = netdev_dpdk_qos_run(dev, pkts, tx_cnt, true);
> -        qos_drops -= tx_cnt;
> +        hwol_drops = batch_cnt;
> +        if (userspace_tso_enabled()) {
> +            batch_cnt = netdev_dpdk_prep_hwol_batch(dev, pkts, batch_cnt);
> +        }
> +        hwol_drops -= batch_cnt;
> +        mtu_drops = batch_cnt;
> +        batch_cnt = netdev_dpdk_filter_packet_len(dev, pkts, batch_cnt);
> +        mtu_drops -= batch_cnt;
> +        qos_drops = batch_cnt;
> +        batch_cnt = netdev_dpdk_qos_run(dev, pkts, batch_cnt, true);
> +        qos_drops -= batch_cnt;
>  
> -        tx_failure = netdev_dpdk_eth_tx_burst(dev, qid, pkts, tx_cnt);
> +        tx_failure = netdev_dpdk_eth_tx_burst(dev, qid, pkts, batch_cnt);
>  
> -        dropped = tx_failure + mtu_drops + qos_drops;
> +        dropped = tx_failure + mtu_drops + qos_drops + hwol_drops;
>          if (OVS_UNLIKELY(dropped)) {
>              rte_spinlock_lock(&dev->stats_lock);
>              dev->stats.tx_dropped += dropped;
>              sw_stats->tx_failure_drops += tx_failure;
>              sw_stats->tx_mtu_exceeded_drops += mtu_drops;
>              sw_stats->tx_qos_drops += qos_drops;
> +            sw_stats->tx_invalid_hwol_drops += hwol_drops;
>              rte_spinlock_unlock(&dev->stats_lock);
>          }
>      }
> @@ -3011,7 +3241,8 @@ netdev_dpdk_get_sw_custom_stats(const struct netdev *netdev,
>      SW_CSTAT(tx_failure_drops)       \
>      SW_CSTAT(tx_mtu_exceeded_drops)  \
>      SW_CSTAT(tx_qos_drops)           \
> -    SW_CSTAT(rx_qos_drops)
> +    SW_CSTAT(rx_qos_drops)           \
> +    SW_CSTAT(tx_invalid_hwol_drops)
>  
>  #define SW_CSTAT(NAME) + 1
>      custom_stats->size = SW_CSTATS;
> @@ -4874,6 +5105,12 @@ netdev_dpdk_reconfigure(struct netdev *netdev)
>  
>      rte_free(dev->tx_q);
>      err = dpdk_eth_dev_init(dev);
> +    if (dev->hw_ol_features & NETDEV_TX_TSO_OFFLOAD) {
> +        netdev->ol_flags |= NETDEV_TX_OFFLOAD_TCP_TSO;
> +        netdev->ol_flags |= NETDEV_TX_OFFLOAD_TCP_CKSUM;
> +        netdev->ol_flags |= NETDEV_TX_OFFLOAD_IPV4_CKSUM;
> +    }
> +
>      dev->tx_q = netdev_dpdk_alloc_txq(netdev->n_txq);
>      if (!dev->tx_q) {
>          err = ENOMEM;
> @@ -4903,6 +5140,11 @@ dpdk_vhost_reconfigure_helper(struct netdev_dpdk *dev)
>          dev->tx_q[0].map = 0;
>      }
>  
> +    if (userspace_tso_enabled()) {
> +        dev->hw_ol_features |= NETDEV_TX_TSO_OFFLOAD;
> +        VLOG_DBG("%s: TSO enabled on vhost port", netdev_get_name(&dev->up));
> +    }
> +
>      netdev_dpdk_remap_txqs(dev);
>  
>      err = netdev_dpdk_mempool_configure(dev);
> @@ -4975,6 +5217,11 @@ netdev_dpdk_vhost_client_reconfigure(struct netdev *netdev)
>              vhost_flags |= RTE_VHOST_USER_DEQUEUE_ZERO_COPY;
>          }
>  
> +        /* Enable External Buffers if TCP Segmentation Offload is enabled. */
> +        if (userspace_tso_enabled()) {
> +            vhost_flags |= RTE_VHOST_USER_EXTBUF_SUPPORT;
> +        }
> +
>          err = rte_vhost_driver_register(dev->vhost_id, vhost_flags);
>          if (err) {
>              VLOG_ERR("vhost-user device setup failure for device %s\n",
> @@ -4999,14 +5246,20 @@ netdev_dpdk_vhost_client_reconfigure(struct netdev *netdev)
>              goto unlock;
>          }
>  
> -        err = rte_vhost_driver_disable_features(dev->vhost_id,
> -                                    1ULL << VIRTIO_NET_F_HOST_TSO4
> -                                    | 1ULL << VIRTIO_NET_F_HOST_TSO6
> -                                    | 1ULL << VIRTIO_NET_F_CSUM);
> -        if (err) {
> -            VLOG_ERR("rte_vhost_driver_disable_features failed for vhost user "
> -                     "client port: %s\n", dev->up.name);
> -            goto unlock;
> +        if (userspace_tso_enabled()) {
> +            netdev->ol_flags |= NETDEV_TX_OFFLOAD_TCP_TSO;
> +            netdev->ol_flags |= NETDEV_TX_OFFLOAD_TCP_CKSUM;
> +            netdev->ol_flags |= NETDEV_TX_OFFLOAD_IPV4_CKSUM;
> +        } else {
> +            err = rte_vhost_driver_disable_features(dev->vhost_id,
> +                                        1ULL << VIRTIO_NET_F_HOST_TSO4
> +                                        | 1ULL << VIRTIO_NET_F_HOST_TSO6
> +                                        | 1ULL << VIRTIO_NET_F_CSUM);
> +            if (err) {
> +                VLOG_ERR("rte_vhost_driver_disable_features failed for "
> +                         "vhost user client port: %s\n", dev->up.name);
> +                goto unlock;
> +            }
>          }
>  
>          err = rte_vhost_driver_start(dev->vhost_id);
> diff --git a/lib/netdev-linux-private.h b/lib/netdev-linux-private.h
> index f08159aa7..9dbc67658 100644
> --- a/lib/netdev-linux-private.h
> +++ b/lib/netdev-linux-private.h
> @@ -27,6 +27,7 @@
>  #include <stdint.h>
>  #include <stdbool.h>
>  
> +#include "dp-packet.h"
>  #include "netdev-afxdp.h"
>  #include "netdev-afxdp-pool.h"
>  #include "netdev-provider.h"
> @@ -37,10 +38,13 @@
>  
>  struct netdev;
>  
> +#define LINUX_RXQ_TSO_MAX_LEN 65536
> +
>  struct netdev_rxq_linux {
>      struct netdev_rxq up;
>      bool is_tap;
>      int fd;
> +    char *aux_bufs[NETDEV_MAX_BURST]; /* Batch of preallocated TSO buffers. */
>  };
>  
>  int netdev_linux_construct(struct netdev *);
> @@ -92,6 +96,7 @@ struct netdev_linux {
>      int tap_fd;
>      bool present;               /* If the device is present in the namespace */
>      uint64_t tx_dropped;        /* tap device can drop if the iface is down */
> +    uint64_t rx_dropped;        /* Packets dropped while recv from kernel. */
>  
>      /* LAG information. */
>      bool is_lag_master;         /* True if the netdev is a LAG master. */
> diff --git a/lib/netdev-linux.c b/lib/netdev-linux.c
> index 41d1e9273..c308abf54 100644
> --- a/lib/netdev-linux.c
> +++ b/lib/netdev-linux.c
> @@ -29,16 +29,18 @@
>  #include <linux/filter.h>
>  #include <linux/gen_stats.h>
>  #include <linux/if_ether.h>
> +#include <linux/if_packet.h>
>  #include <linux/if_tun.h>
>  #include <linux/types.h>
>  #include <linux/ethtool.h>
>  #include <linux/mii.h>
>  #include <linux/rtnetlink.h>
>  #include <linux/sockios.h>
> +#include <linux/virtio_net.h>
>  #include <sys/ioctl.h>
>  #include <sys/socket.h>
> +#include <sys/uio.h>
>  #include <sys/utsname.h>
> -#include <netpacket/packet.h>
>  #include <net/if.h>
>  #include <net/if_arp.h>
>  #include <net/route.h>
> @@ -72,6 +74,7 @@
>  #include "socket-util.h"
>  #include "sset.h"
>  #include "tc.h"
> +#include "userspace-tso.h"

Alphabetical order.

>  #include "timer.h"
>  #include "unaligned.h"
>  #include "openvswitch/vlog.h"
> @@ -237,6 +240,16 @@ enum {
>      VALID_DRVINFO           = 1 << 6,
>      VALID_FEATURES          = 1 << 7,
>  };
> +
> +/* Use one for the packet buffer and another for the aux buffer to receive
> + * TSO packets. */
> +#define IOV_STD_SIZE 1
> +#define IOV_TSO_SIZE 2
> +
> +enum {
> +    IOV_PACKET = 0,
> +    IOV_AUXBUF = 1,
> +};
>  

>  struct linux_lag_slave {
>     uint32_t block_id;
> @@ -501,6 +514,8 @@ static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20);
>   * changes in the device miimon status, so we can use atomic_count. */
>  static atomic_count miimon_cnt = ATOMIC_COUNT_INIT(0);
>  
> +static int netdev_linux_parse_vnet_hdr(struct dp_packet *b);
> +static void netdev_linux_prepend_vnet_hdr(struct dp_packet *b, int mtu);
>  static int netdev_linux_do_ethtool(const char *name, struct ethtool_cmd *,
>                                     int cmd, const char *cmd_name);
>  static int get_flags(const struct netdev *, unsigned int *flags);
> @@ -902,6 +917,13 @@ netdev_linux_common_construct(struct netdev *netdev_)
>      /* The device could be in the same network namespace or in another one. */
>      netnsid_unset(&netdev->netnsid);
>      ovs_mutex_init(&netdev->mutex);
> +
> +    if (userspace_tso_enabled()) {
> +        netdev_->ol_flags |= NETDEV_TX_OFFLOAD_TCP_TSO;
> +        netdev_->ol_flags |= NETDEV_TX_OFFLOAD_TCP_CKSUM;
> +        netdev_->ol_flags |= NETDEV_TX_OFFLOAD_IPV4_CKSUM;
> +    }
> +
>      return 0;
>  }
>  
> @@ -961,6 +983,10 @@ netdev_linux_construct_tap(struct netdev *netdev_)
>      /* Create tap device. */
>      get_flags(&netdev->up, &netdev->ifi_flags);
>      ifr.ifr_flags = IFF_TAP | IFF_NO_PI;
> +    if (userspace_tso_enabled()) {
> +        ifr.ifr_flags |= IFF_VNET_HDR;
> +    }
> +
>      ovs_strzcpy(ifr.ifr_name, name, sizeof ifr.ifr_name);
>      if (ioctl(netdev->tap_fd, TUNSETIFF, &ifr) == -1) {
>          VLOG_WARN("%s: creating tap device failed: %s", name,
> @@ -1024,6 +1050,15 @@ static struct netdev_rxq *
>  netdev_linux_rxq_alloc(void)
>  {
>      struct netdev_rxq_linux *rx = xzalloc(sizeof *rx);
> +    if (userspace_tso_enabled()) {
> +        int i;
> +
> +        /* Allocate auxiliay buffers to receive TSO packets */

Period at the end of comment.

> +        for (i = 0; i < NETDEV_MAX_BURST; i++) {
> +            rx->aux_bufs[i] = xmalloc(LINUX_RXQ_TSO_MAX_LEN);
> +        }
> +    }
> +
>      return &rx->up;
>  }
>  
> @@ -1069,6 +1104,15 @@ netdev_linux_rxq_construct(struct netdev_rxq *rxq_)
>              goto error;
>          }
>  
> +        if (userspace_tso_enabled()
> +            && setsockopt(rx->fd, SOL_PACKET, PACKET_VNET_HDR, &val,
> +                          sizeof val)) {
> +            error = errno;
> +            VLOG_ERR("%s: failed to enable vnet hdr in txq raw socket: %s",
> +                     netdev_get_name(netdev_), ovs_strerror(errno));
> +            goto error;
> +        }
> +
>          /* Set non-blocking mode. */
>          error = set_nonblocking(rx->fd);
>          if (error) {
> @@ -1119,10 +1163,15 @@ static void
>  netdev_linux_rxq_destruct(struct netdev_rxq *rxq_)
>  {
>      struct netdev_rxq_linux *rx = netdev_rxq_linux_cast(rxq_);
> +    int i;
>  
>      if (!rx->is_tap) {
>          close(rx->fd);
>      }
> +
> +    for (i = 0; i < NETDEV_MAX_BURST; i++) {
> +        free(rx->aux_bufs[i]);
> +    }
>  }
>  
>  static void
> @@ -1159,12 +1208,14 @@ auxdata_has_vlan_tci(const struct tpacket_auxdata *aux)
>   * It also used recvmmsg to reduce multiple syscalls overhead;
>   */
>  static int
> -netdev_linux_batch_rxq_recv_sock(int fd, int mtu,
> +netdev_linux_batch_rxq_recv_sock(struct netdev_rxq_linux *rx, int mtu,
>                                   struct dp_packet_batch *batch)
>  {
> -    size_t size;
> +    int iovlen;
> +    size_t std_len;
>      ssize_t retval;
> -    struct iovec iovs[NETDEV_MAX_BURST];
> +    int virtio_net_hdr_size;
> +    struct iovec iovs[NETDEV_MAX_BURST][IOV_TSO_SIZE];
>      struct cmsghdr *cmsg;
>      union {
>          struct cmsghdr cmsg;
> @@ -1174,41 +1225,87 @@ netdev_linux_batch_rxq_recv_sock(int fd, int mtu,
>      struct dp_packet *buffers[NETDEV_MAX_BURST];
>      int i;
>  
> +    if (userspace_tso_enabled()) {
> +        /* Use the buffer from the allocated packet below to receive MTU
> +         * sized packets and an aux_buf for extra TSO data. */
> +        iovlen = IOV_TSO_SIZE;
> +        virtio_net_hdr_size = sizeof(struct virtio_net_hdr);
> +    } else {
> +        /* Use only the buffer from the allocated packet. */
> +        iovlen = IOV_STD_SIZE;
> +        virtio_net_hdr_size = 0;
> +    }
> +
> +    std_len = VLAN_ETH_HEADER_LEN + mtu + virtio_net_hdr_size;
>      for (i = 0; i < NETDEV_MAX_BURST; i++) {
> -         buffers[i] = dp_packet_new_with_headroom(VLAN_ETH_HEADER_LEN + mtu,
> -                                                  DP_NETDEV_HEADROOM);
> -         /* Reserve headroom for a single VLAN tag */
> -         dp_packet_reserve(buffers[i], VLAN_HEADER_LEN);
> -         size = dp_packet_tailroom(buffers[i]);
> -         iovs[i].iov_base = dp_packet_data(buffers[i]);
> -         iovs[i].iov_len = size;
> +         buffers[i] = dp_packet_new_with_headroom(std_len, DP_NETDEV_HEADROOM);
> +         iovs[i][IOV_PACKET].iov_base = dp_packet_data(buffers[i]);
> +         iovs[i][IOV_PACKET].iov_len = std_len;
> +         iovs[i][IOV_AUXBUF].iov_base = rx->aux_bufs[i];
> +         iovs[i][IOV_AUXBUF].iov_len = LINUX_RXQ_TSO_MAX_LEN;
>           mmsgs[i].msg_hdr.msg_name = NULL;
>           mmsgs[i].msg_hdr.msg_namelen = 0;
> -         mmsgs[i].msg_hdr.msg_iov = &iovs[i];
> -         mmsgs[i].msg_hdr.msg_iovlen = 1;
> +         mmsgs[i].msg_hdr.msg_iov = iovs[i];
> +         mmsgs[i].msg_hdr.msg_iovlen = iovlen;
>           mmsgs[i].msg_hdr.msg_control = &cmsg_buffers[i];
>           mmsgs[i].msg_hdr.msg_controllen = sizeof cmsg_buffers[i];
>           mmsgs[i].msg_hdr.msg_flags = 0;
>      }
>  
>      do {
> -        retval = recvmmsg(fd, mmsgs, NETDEV_MAX_BURST, MSG_TRUNC, NULL);
> +        retval = recvmmsg(rx->fd, mmsgs, NETDEV_MAX_BURST, MSG_TRUNC, NULL);
>      } while (retval < 0 && errno == EINTR);
>  
>      if (retval < 0) {
> -        /* Save -errno to retval temporarily */
> -        retval = -errno;
> -        i = 0;
> -        goto free_buffers;
> +        retval = errno;
> +        for (i = 0; i < NETDEV_MAX_BURST; i++) {
> +            dp_packet_delete(buffers[i]);
> +        }
> +
> +        return retval;
>      }
>  
>      for (i = 0; i < retval; i++) {
>          if (mmsgs[i].msg_len < ETH_HEADER_LEN) {
> -            break;
> +            struct netdev *netdev_ = netdev_rxq_get_netdev(&rx->up);
> +            struct netdev_linux *netdev = netdev_linux_cast(netdev_);
> +
> +            dp_packet_delete(buffers[i]);
> +            netdev->rx_dropped += 1;
> +            VLOG_WARN_RL(&rl, "%s: Dropped packet: less than ether hdr size",
> +                         netdev_get_name(netdev_));
> +            continue;
> +        }
> +
> +        if (mmsgs[i].msg_len > std_len) {
> +            /* Build a single linear TSO packet by expanding the current packet
> +             * to append the data received in the aux_buf. */
> +            size_t extra_len = mmsgs[i].msg_len - std_len;
> +
> +            dp_packet_set_size(buffers[i], dp_packet_size(buffers[i])
> +                               + std_len);
> +            dp_packet_prealloc_tailroom(buffers[i], extra_len);
> +            memcpy(dp_packet_tail(buffers[i]), rx->aux_bufs[i], extra_len);
> +            dp_packet_set_size(buffers[i], dp_packet_size(buffers[i])
> +                               + extra_len);
> +        } else {
> +            dp_packet_set_size(buffers[i], dp_packet_size(buffers[i])
> +                               + mmsgs[i].msg_len);
>          }
>  
> -        dp_packet_set_size(buffers[i],
> -                           dp_packet_size(buffers[i]) + mmsgs[i].msg_len);
> +        if (virtio_net_hdr_size && netdev_linux_parse_vnet_hdr(buffers[i])) {
> +            struct netdev *netdev_ = netdev_rxq_get_netdev(&rx->up);
> +            struct netdev_linux *netdev = netdev_linux_cast(netdev_);
> +
> +            /* Unexpected error situation: the virtio header is not present
> +             * or corrupted. Drop the packet but continue in case next ones
> +             * are correct. */
> +            dp_packet_delete(buffers[i]);
> +            netdev->rx_dropped += 1;
> +            VLOG_WARN_RL(&rl, "%s: Dropped packet: Invalid virtio net header",
> +                         netdev_get_name(netdev_));
> +            continue;
> +        }
>  
>          for (cmsg = CMSG_FIRSTHDR(&mmsgs[i].msg_hdr); cmsg;
>                   cmsg = CMSG_NXTHDR(&mmsgs[i].msg_hdr, cmsg)) {
> @@ -1238,22 +1335,11 @@ netdev_linux_batch_rxq_recv_sock(int fd, int mtu,
>          dp_packet_batch_add(batch, buffers[i]);
>      }
>  
> -free_buffers:
> -    /* Free unused buffers, including buffers whose size is less than
> -     * ETH_HEADER_LEN.
> -     *
> -     * Note: i has been set correctly by the above for loop, so don't
> -     * try to re-initialize it.
> -     */
> +    /* Delete unused buffers */

Period at the end of comment.

>      for (; i < NETDEV_MAX_BURST; i++) {
>          dp_packet_delete(buffers[i]);
>      }
>  
> -    /* netdev_linux_rxq_recv needs it to return 0 or positive errno */
> -    if (retval < 0) {
> -        return -retval;
> -    }> -
>      return 0;
>  }
>  
> @@ -1263,20 +1349,40 @@ free_buffers:
>   * packets are added into *batch. The return value is 0 or errno.
>   */
>  static int
> -netdev_linux_batch_rxq_recv_tap(int fd, int mtu, struct dp_packet_batch *batch)
> +netdev_linux_batch_rxq_recv_tap(struct netdev_rxq_linux *rx, int mtu,
> +                                struct dp_packet_batch *batch)
>  {
>      struct dp_packet *buffer;
> +    int virtio_net_hdr_size;
>      ssize_t retval;
> -    size_t size;
> +    size_t std_len;
> +    int iovlen;
>      int i;
>  
> +    if (userspace_tso_enabled()) {
> +        /* Use the buffer from the allocated packet below to receive MTU
> +         * sized packets and an aux_buf for extra TSO data. */
> +        iovlen = IOV_TSO_SIZE;
> +        virtio_net_hdr_size = sizeof(struct virtio_net_hdr);
> +    } else {
> +        /* Use only the buffer from the allocated packet. */
> +        iovlen = IOV_STD_SIZE;
> +        virtio_net_hdr_size = 0;
> +    }
> +
> +    std_len = VLAN_ETH_HEADER_LEN + mtu + virtio_net_hdr_size;
>      for (i = 0; i < NETDEV_MAX_BURST; i++) {
> +        struct iovec iov[IOV_TSO_SIZE];
> +
>          /* Assume Ethernet port. No need to set packet_type. */
> -        buffer = dp_packet_new_with_headroom(VLAN_ETH_HEADER_LEN + mtu,
> -                                             DP_NETDEV_HEADROOM);
> -        size = dp_packet_tailroom(buffer);
> +        buffer = dp_packet_new_with_headroom(std_len, DP_NETDEV_HEADROOM);
> +        iov[IOV_PACKET].iov_base = dp_packet_data(buffer);
> +        iov[IOV_PACKET].iov_len = std_len;
> +        iov[IOV_AUXBUF].iov_base = rx->aux_bufs[i];
> +        iov[IOV_AUXBUF].iov_len = LINUX_RXQ_TSO_MAX_LEN;
> +
>          do {
> -            retval = read(fd, dp_packet_data(buffer), size);
> +            retval = readv(rx->fd, iov, iovlen);
>          } while (retval < 0 && errno == EINTR);
>  
>          if (retval < 0) {
> @@ -1284,7 +1390,33 @@ netdev_linux_batch_rxq_recv_tap(int fd, int mtu, struct dp_packet_batch *batch)
>              break;
>          }
>  
> -        dp_packet_set_size(buffer, dp_packet_size(buffer) + retval);
> +        if (retval > std_len) {
> +            /* Build a single linear TSO packet by expanding the current packet
> +             * to append the data received in the aux_buf. */
> +            size_t extra_len = retval - std_len;
> +
> +            dp_packet_set_size(buffer, dp_packet_size(buffer) + std_len);
> +            dp_packet_prealloc_tailroom(buffer, extra_len);
> +            memcpy(dp_packet_tail(buffer), rx->aux_bufs[i], extra_len);
> +            dp_packet_set_size(buffer, dp_packet_size(buffer) + extra_len);
> +        } else {
> +            dp_packet_set_size(buffer, dp_packet_size(buffer) + retval);
> +        }
> +
> +        if (virtio_net_hdr_size && netdev_linux_parse_vnet_hdr(buffer)) {
> +            struct netdev *netdev_ = netdev_rxq_get_netdev(&rx->up);
> +            struct netdev_linux *netdev = netdev_linux_cast(netdev_);
> +
> +            /* Unexpected error situation: the virtio header is not present
> +             * or corrupted. Drop the packet but continue in case next ones
> +             * are correct. */
> +            dp_packet_delete(buffer);
> +            netdev->rx_dropped += 1;
> +            VLOG_WARN_RL(&rl, "%s: Dropped packet: Invalid virtio net header",
> +                         netdev_get_name(netdev_));
> +            continue;
> +        }
> +
>          dp_packet_batch_add(batch, buffer);
>      }
>  
> @@ -1310,8 +1442,8 @@ netdev_linux_rxq_recv(struct netdev_rxq *rxq_, struct dp_packet_batch *batch,
>  
>      dp_packet_batch_init(batch);
>      retval = (rx->is_tap
> -              ? netdev_linux_batch_rxq_recv_tap(rx->fd, mtu, batch)
> -              : netdev_linux_batch_rxq_recv_sock(rx->fd, mtu, batch));
> +              ? netdev_linux_batch_rxq_recv_tap(rx, mtu, batch)
> +              : netdev_linux_batch_rxq_recv_sock(rx, mtu, batch));
>  
>      if (retval) {
>          if (retval != EAGAIN && retval != EMSGSIZE) {
> @@ -1353,7 +1485,7 @@ netdev_linux_rxq_drain(struct netdev_rxq *rxq_)
>  }
>  
>  static int
> -netdev_linux_sock_batch_send(int sock, int ifindex,
> +netdev_linux_sock_batch_send(int sock, int ifindex, bool tso, int mtu,
>                               struct dp_packet_batch *batch)
>  {
>      const size_t size = dp_packet_batch_size(batch);
> @@ -1367,6 +1499,10 @@ netdev_linux_sock_batch_send(int sock, int ifindex,
>  
>      struct dp_packet *packet;
>      DP_PACKET_BATCH_FOR_EACH (i, packet, batch) {
> +        if (tso) {
> +            netdev_linux_prepend_vnet_hdr(packet, mtu);
> +        }
> +
>          iov[i].iov_base = dp_packet_data(packet);
>          iov[i].iov_len = dp_packet_size(packet);
>          mmsg[i].msg_hdr = (struct msghdr) { .msg_name = &sll,
> @@ -1399,7 +1535,7 @@ netdev_linux_sock_batch_send(int sock, int ifindex,
>   * on other interface types because we attach a socket filter to the rx
>   * socket. */
>  static int
> -netdev_linux_tap_batch_send(struct netdev *netdev_,
> +netdev_linux_tap_batch_send(struct netdev *netdev_, bool tso, int mtu,
>                              struct dp_packet_batch *batch)
>  {
>      struct netdev_linux *netdev = netdev_linux_cast(netdev_);
> @@ -1416,10 +1552,15 @@ netdev_linux_tap_batch_send(struct netdev *netdev_,
>      }
>  
>      DP_PACKET_BATCH_FOR_EACH (i, packet, batch) {
> -        size_t size = dp_packet_size(packet);
> +        size_t size;
>          ssize_t retval;
>          int error;
>  
> +        if (tso) {
> +            netdev_linux_prepend_vnet_hdr(packet, mtu);
> +        }
> +
> +        size = dp_packet_size(packet);
>          do {
>              retval = write(netdev->tap_fd, dp_packet_data(packet), size);
>              error = retval < 0 ? errno : 0;
> @@ -1454,9 +1595,15 @@ netdev_linux_send(struct netdev *netdev_, int qid OVS_UNUSED,
>                    struct dp_packet_batch *batch,
>                    bool concurrent_txq OVS_UNUSED)
>  {
> +    bool tso = userspace_tso_enabled();
> +    int mtu = ETH_PAYLOAD_MAX;
>      int error = 0;
>      int sock = 0;
>  
> +    if (tso) {
> +        netdev_linux_get_mtu__(netdev_linux_cast(netdev_), &mtu);

netdev_linux_get_mtu__() could fail.  This needs to be handled.

> +    }
> +
>      if (!is_tap_netdev(netdev_)) {
>          if (netdev_linux_netnsid_is_remote(netdev_linux_cast(netdev_))) {
>              error = EOPNOTSUPP;
> @@ -1475,9 +1622,9 @@ netdev_linux_send(struct netdev *netdev_, int qid OVS_UNUSED,
>              goto free_batch;
>          }
>  
> -        error = netdev_linux_sock_batch_send(sock, ifindex, batch);
> +        error = netdev_linux_sock_batch_send(sock, ifindex, tso, mtu, batch);
>      } else {
> -        error = netdev_linux_tap_batch_send(netdev_, batch);
> +        error = netdev_linux_tap_batch_send(netdev_, tso, mtu, batch);
>      }
>      if (error) {
>          if (error == ENOBUFS) {
> @@ -2045,6 +2192,7 @@ netdev_tap_get_stats(const struct netdev *netdev_, struct netdev_stats *stats)
>          stats->collisions          += dev_stats.collisions;
>      }
>      stats->tx_dropped += netdev->tx_dropped;
> +    stats->rx_dropped += netdev->rx_dropped;
>      ovs_mutex_unlock(&netdev->mutex);
>  
>      return error;
> @@ -6223,6 +6371,17 @@ af_packet_sock(void)
>              if (error) {
>                  close(sock);
>                  sock = -error;
> +            } else if (userspace_tso_enabled()) {
> +                int val = 1;
> +                error = setsockopt(sock, SOL_PACKET, PACKET_VNET_HDR, &val,
> +                                   sizeof val);
> +                if (error) {
> +                    error = errno;
> +                    VLOG_ERR("failed to enable vnet hdr in raw socket: %s",
> +                             ovs_strerror(errno));
> +                    close(sock);
> +                    sock = -error;
> +                }
>              }
>          } else {
>              sock = -errno;
> @@ -6234,3 +6393,136 @@ af_packet_sock(void)
>  
>      return sock;
>  }
> +
> +static int
> +netdev_linux_parse_l2(struct dp_packet *b, uint16_t *l4proto)

I'm not generally a fan of parsing packets here on receive especially
because we will parse it with miniflow_extract() later, but I'm not
sure how to avoid that.

> +{
> +    struct eth_header *eth_hdr;
> +    ovs_be16 eth_type;
> +    int l2_len;
> +
> +    eth_hdr = dp_packet_at(b, 0, ETH_HEADER_LEN);
> +    if (!eth_hdr) {
> +        return -EINVAL;
> +    }
> +
> +    l2_len = ETH_HEADER_LEN;
> +    eth_type = eth_hdr->eth_type;
> +    if (eth_type_vlan(eth_type)) {
> +        struct vlan_header *vlan = dp_packet_at(b, l2_len, VLAN_HEADER_LEN);
> +
> +        if (!vlan) {
> +            return -EINVAL;
> +        }
> +
> +        eth_type = vlan->vlan_next_type;
> +        l2_len += VLAN_HEADER_LEN;
> +    }
> +
> +    if (eth_type == htons(ETH_TYPE_IP)) {
> +        struct ip_header *ip_hdr = dp_packet_at(b, l2_len, IP_HEADER_LEN);
> +
> +        if (!ip_hdr) {
> +            return -EINVAL;
> +        }
> +
> +        *l4proto = ip_hdr->ip_proto;
> +        dp_packet_hwol_set_tx_ipv4(b);
> +    } else if (eth_type == htons(ETH_TYPE_IPV6)) {
> +        struct ovs_16aligned_ip6_hdr *nh6;
> +
> +        nh6 = dp_packet_at(b, l2_len, IPV6_HEADER_LEN);
> +        if (!nh6) {
> +            return -EINVAL;
> +        }
> +
> +        *l4proto = nh6->ip6_ctlun.ip6_un1.ip6_un1_nxt;
> +        dp_packet_hwol_set_tx_ipv6(b);
> +    }
> +
> +    return 0;
> +}
> +
> +static int
> +netdev_linux_parse_vnet_hdr(struct dp_packet *b)
> +{
> +    struct virtio_net_hdr *vnet = dp_packet_pull(b, sizeof *vnet);
> +    uint16_t l4proto = 0;
> +
> +    if (OVS_UNLIKELY(!vnet)) {
> +        return -EINVAL;
> +    }
> +
> +    if (vnet->flags == 0 && vnet->gso_type == VIRTIO_NET_HDR_GSO_NONE) {
> +        return 0;
> +    }
> +
> +    if (netdev_linux_parse_l2(b, &l4proto)) {
> +        return -EINVAL;
> +    }
> +
> +    if (vnet->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
> +        if (l4proto == IPPROTO_TCP) {
> +            dp_packet_hwol_set_csum_tcp(b);
> +        } else if (l4proto == IPPROTO_UDP) {
> +            dp_packet_hwol_set_csum_udp(b);
> +        } else if (l4proto == IPPROTO_SCTP) {
> +            dp_packet_hwol_set_csum_sctp(b);
> +        }
> +    }
> +
> +    if (l4proto && vnet->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
> +        uint8_t allowed_mask = VIRTIO_NET_HDR_GSO_TCPV4
> +                                | VIRTIO_NET_HDR_GSO_TCPV6
> +                                | VIRTIO_NET_HDR_GSO_UDP;
> +        uint8_t type = vnet->gso_type & allowed_mask;
> +
> +        if (type == VIRTIO_NET_HDR_GSO_TCPV4
> +            || type == VIRTIO_NET_HDR_GSO_TCPV6) {
> +            dp_packet_hwol_set_tcp_seg(b);
> +        }
> +    }
> +
> +    return 0;
> +}
> +
> +static void
> +netdev_linux_prepend_vnet_hdr(struct dp_packet *b, int mtu)
> +{
> +    struct virtio_net_hdr *vnet = dp_packet_push_zeros(b, sizeof *vnet);
> +
> +    if (dp_packet_hwol_is_tso(b)) {
> +        uint16_t hdr_len = ((char *)dp_packet_l4(b) - (char *)dp_packet_eth(b))
> +                            + TCP_HEADER_LEN;
> +
> +        vnet->hdr_len = (OVS_FORCE __virtio16)hdr_len;
> +        vnet->gso_size = (OVS_FORCE __virtio16)(mtu - hdr_len);
> +        if (dp_packet_hwol_is_ipv4(b)) {
> +            vnet->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
> +        } else {
> +            vnet->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
> +        }
> +
> +    } else {
> +        vnet->flags = VIRTIO_NET_HDR_GSO_NONE;
> +    }
> +
> +    if (dp_packet_hwol_l4_mask(b)) {
> +        vnet->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
> +        vnet->csum_start = (OVS_FORCE __virtio16)((char *)dp_packet_l4(b)
> +                                                  - (char *)dp_packet_eth(b));
> +
> +        if (dp_packet_hwol_l4_is_tcp(b)) {
> +            vnet->csum_offset = (OVS_FORCE __virtio16) __builtin_offsetof(
> +                                    struct tcp_header, tcp_csum);
> +        } else if (dp_packet_hwol_l4_is_udp(b)) {
> +            vnet->csum_offset = (OVS_FORCE __virtio16) __builtin_offsetof(
> +                                    struct udp_header, udp_csum);
> +        } else if (dp_packet_hwol_l4_is_sctp(b)) {
> +            vnet->csum_offset = (OVS_FORCE __virtio16) __builtin_offsetof(
> +                                    struct sctp_header, sctp_csum);
> +        } else {
> +            VLOG_WARN_RL(&rl, "Unsupported L4 protocol");
> +        }
> +    }
> +}
> diff --git a/lib/netdev-provider.h b/lib/netdev-provider.h
> index f109c4e66..22f4cde33 100644
> --- a/lib/netdev-provider.h
> +++ b/lib/netdev-provider.h
> @@ -37,6 +37,12 @@ extern "C" {
>  struct netdev_tnl_build_header_params;
>  #define NETDEV_NUMA_UNSPEC OVS_NUMA_UNSPEC
>  
> +enum netdev_ol_flags {
> +    NETDEV_TX_OFFLOAD_IPV4_CKSUM = 1 << 0,
> +    NETDEV_TX_OFFLOAD_TCP_CKSUM = 1 << 1,
> +    NETDEV_TX_OFFLOAD_TCP_TSO = 1 << 2,
> +};
> +
>  /* A network device (e.g. an Ethernet device).
>   *
>   * Network device implementations may read these members but should not modify
> @@ -51,6 +57,9 @@ struct netdev {
>       * opening this device, and therefore got assigned to the "system" class */
>      bool auto_classified;
>  
> +    /* This bitmask of the offloading features enabled by the netdev. */
> +    uint64_t ol_flags;
> +
>      /* If this is 'true', the user explicitly specified an MTU for this
>       * netdev.  Otherwise, Open vSwitch is allowed to override it. */
>      bool mtu_user_config;
> diff --git a/lib/netdev.c b/lib/netdev.c
> index 405c98c68..f95b19af4 100644
> --- a/lib/netdev.c
> +++ b/lib/netdev.c
> @@ -66,6 +66,8 @@ COVERAGE_DEFINE(netdev_received);
>  COVERAGE_DEFINE(netdev_sent);
>  COVERAGE_DEFINE(netdev_add_router);
>  COVERAGE_DEFINE(netdev_get_stats);
> +COVERAGE_DEFINE(netdev_send_prepare_drops);
> +COVERAGE_DEFINE(netdev_push_header_drops);
>  
>  struct netdev_saved_flags {
>      struct netdev *netdev;
> @@ -782,6 +784,54 @@ netdev_get_pt_mode(const struct netdev *netdev)
>              : NETDEV_PT_LEGACY_L2);
>  }
>  
> +/* Check if a 'packet' is compatible with 'netdev_flags'.
> + * If a packet is incompatible, return 'false' with the 'errormsg'
> + * pointing to a reason. */
> +static bool
> +netdev_send_prepare_packet(const uint64_t netdev_flags,
> +                           struct dp_packet *packet, char **errormsg)
> +{
> +    if (dp_packet_hwol_is_tso(packet)
> +        && !(netdev_flags & NETDEV_TX_OFFLOAD_TCP_TSO)) {
> +            /* Fall back to GSO in software. */
> +            VLOG_ERR_BUF(errormsg, "No TSO support");
> +            return false;
> +    }
> +
> +    if (dp_packet_hwol_l4_mask(packet)
> +        && !(netdev_flags & NETDEV_TX_OFFLOAD_TCP_CKSUM)) {
> +            /* Fall back to L4 csum in software. */
> +            VLOG_ERR_BUF(errormsg, "No L4 checksum support");
> +            return false;
> +    }
> +
> +    return true;
> +}
> +
> +/* Check if each packet in 'batch' is compatible with 'netdev' features,
> + * otherwise either fall back to software implementation or drop it. */
> +static void
> +netdev_send_prepare_batch(const struct netdev *netdev,
> +                          struct dp_packet_batch *batch)
> +{
> +    struct dp_packet *packet;
> +    size_t i, size = dp_packet_batch_size(batch);
> +
> +    DP_PACKET_BATCH_REFILL_FOR_EACH (i, size, packet, batch) {
> +        char *errormsg = NULL;
> +
> +        if (netdev_send_prepare_packet(netdev->ol_flags, packet, &errormsg)) {
> +            dp_packet_batch_refill(batch, packet, i);
> +        } else {
> +            dp_packet_delete(packet);
> +            COVERAGE_INC(netdev_send_prepare_drops);
> +            VLOG_WARN_RL(&rl, "%s: Packet dropped: %s",
> +                         netdev_get_name(netdev), errormsg);
> +            free(errormsg);
> +        }
> +    }
> +}
> +
>  /* Sends 'batch' on 'netdev'.  Returns 0 if successful (for every packet),
>   * otherwise a positive errno value.  Returns EAGAIN without blocking if
>   * at least one the packets cannot be queued immediately.  Returns EMSGSIZE
> @@ -811,8 +861,14 @@ int
>  netdev_send(struct netdev *netdev, int qid, struct dp_packet_batch *batch,
>              bool concurrent_txq)
>  {
> -    int error = netdev->netdev_class->send(netdev, qid, batch,
> -                                           concurrent_txq);
> +    int error;
> +
> +    netdev_send_prepare_batch(netdev, batch);
> +    if (OVS_UNLIKELY(dp_packet_batch_is_empty(batch))) {
> +        return 0;
> +    }
> +
> +    error = netdev->netdev_class->send(netdev, qid, batch, concurrent_txq);
>      if (!error) {
>          COVERAGE_INC(netdev_sent);
>      }
> @@ -878,9 +934,21 @@ netdev_push_header(const struct netdev *netdev,
>                     const struct ovs_action_push_tnl *data)
>  {
>      struct dp_packet *packet;
> -    DP_PACKET_BATCH_FOR_EACH (i, packet, batch) {
> -        netdev->netdev_class->push_header(netdev, packet, data);
> -        pkt_metadata_init(&packet->md, data->out_port);
> +    size_t i, size = dp_packet_batch_size(batch);
> +
> +    DP_PACKET_BATCH_REFILL_FOR_EACH (i, size, packet, batch) {
> +        if (OVS_UNLIKELY(dp_packet_hwol_is_tso(packet)
> +                         || dp_packet_hwol_l4_mask(packet))) {
> +            COVERAGE_INC(netdev_push_header_drops);
> +            dp_packet_delete(packet);
> +            VLOG_WARN_RL(&rl, "%s: Tunneling packets with HW offload flags is "
> +                         "not supported: packet dropped",
> +                         netdev_get_name(netdev));
> +        } else {
> +            netdev->netdev_class->push_header(netdev, packet, data);
> +            pkt_metadata_init(&packet->md, data->out_port);
> +            dp_packet_batch_refill(batch, packet, i);
> +        }
>      }
>  
>      return 0;
> diff --git a/lib/userspace-tso.c b/lib/userspace-tso.c
> new file mode 100644
> index 000000000..f843c2a76
> --- /dev/null
> +++ b/lib/userspace-tso.c
> @@ -0,0 +1,48 @@
> +/*
> + * Copyright (c) 2020 Red Hat, Inc.
> + *
> + * Licensed under the Apache License, Version 2.0 (the "License");
> + * you may not use this file except in compliance with the License.
> + * You may obtain a copy of the License at:
> + *
> + *     http://www.apache.org/licenses/LICENSE-2.0
> + *
> + * Unless required by applicable law or agreed to in writing, software
> + * distributed under the License is distributed on an "AS IS" BASIS,
> + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
> + * See the License for the specific language governing permissions and
> + * limitations under the License.
> + */
> +
> +#include <config.h>
> +
> +#include "smap.h"
> +#include "ovs-thread.h"
> +#include "openvswitch/vlog.h"
> +#include "dpdk.h"
> +#include "userspace-tso.h"
> +#include "vswitch-idl.h"
> +
> +VLOG_DEFINE_THIS_MODULE(userspace_tso);
> +
> +static bool userspace_tso = false;
> +
> +void
> +userspace_tso_init(const struct smap *ovs_other_config)
> +{
> +    if (smap_get_bool(ovs_other_config, "userspace-tso-enable", false)) {
> +        static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER;
> +
> +        if (ovsthread_once_start(&once)) {
> +            VLOG_INFO("Userspace TCP Segmentation Offloading support enabled");
> +            userspace_tso = true;

Since dp_packet functions has no implementation if OVS built without
DPDK support, I think, we need to restrict enabling the functionality.

#ifdef DPDK_NETDEV
            VLOG_INFO("Userspace TCP Segmentation Offloading support enabled");
            userspace_tso = true;
#else
            VLOG_WARN("Userspace TCP Segmentation Offloading can not be enabled"
                      "since OVS built without DPDK support.");
#endif

> +            ovsthread_once_done(&once);
> +        }
> +    }
> +}
> +
> +bool
> +userspace_tso_enabled(void)
> +{
> +    return userspace_tso;
> +}
> diff --git a/lib/userspace-tso.h b/lib/userspace-tso.h
> new file mode 100644
> index 000000000..0758274c0
> --- /dev/null
> +++ b/lib/userspace-tso.h
> @@ -0,0 +1,23 @@
> +/*
> + * Copyright (c) 2020 Red Hat Inc.
> + *
> + * Licensed under the Apache License, Version 2.0 (the "License");
> + * you may not use this file except in compliance with the License.
> + * You may obtain a copy of the License at:
> + *
> + *     http://www.apache.org/licenses/LICENSE-2.0
> + *
> + * Unless required by applicable law or agreed to in writing, software
> + * distributed under the License is distributed on an "AS IS" BASIS,
> + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
> + * See the License for the specific language governing permissions and
> + * limitations under the License.
> + */
> +
> +#ifndef USERSPACE_TSO_H
> +#define USERSPACE_TSO_H 1
> +
> +void userspace_tso_init(const struct smap *ovs_other_config);
> +bool userspace_tso_enabled(void);
> +
> +#endif /* userspace-tso.h */
> diff --git a/vswitchd/bridge.c b/vswitchd/bridge.c
> index 86c7b10a9..e591c26a6 100644
> --- a/vswitchd/bridge.c
> +++ b/vswitchd/bridge.c
> @@ -65,6 +65,7 @@
>  #include "system-stats.h"
>  #include "timeval.h"
>  #include "tnl-ports.h"
> +#include "userspace-tso.h"
>  #include "util.h"
>  #include "unixctl.h"
>  #include "lib/vswitch-idl.h"
> @@ -3285,6 +3286,7 @@ bridge_run(void)
>      if (cfg) {
>          netdev_set_flow_api_enabled(&cfg->other_config);
>          dpdk_init(&cfg->other_config);
> +        userspace_tso_init(&cfg->other_config);
>      }
>  
>      /* Initialize the ofproto library.  This only needs to run once, but
> diff --git a/vswitchd/vswitch.xml b/vswitchd/vswitch.xml
> index c43cb1aa4..a9efe71a5 100644
> --- a/vswitchd/vswitch.xml
> +++ b/vswitchd/vswitch.xml
> @@ -690,6 +690,23 @@
>           once in few hours or a day or a week.
>          </p>
>        </column>
> +      <column name="other_config" key="userspace-tso-enable"
> +              type='{"type": "boolean"}'>
> +        <p>
> +          Set this value to <code>true</code> to enable userspace support for
> +          TCP Segmentation Offloading (TSO). When it is enabled, the interfaces
> +          can provide an oversized TCP segment to the datapath and the datapath
> +          will offload the TCP segmentation and checksum calculation to the
> +          interfaces when necessary.
> +        </p>
> +        <p>
> +          The default value is <code>false</code>. Changing this value requires
> +          restarting the daemon.

Works only if OVS is built with DPDK support.

> +        </p>
> +        <p>
> +          The feature is considered experimental.
> +        </p>
> +      </column>
>      </group>
>      <group title="Status">
>        <column name="next_cfg">
> 


More information about the dev mailing list