MFC r326597: vnic: apply hardware L3 checksum only for IPv4

Previously we set the csum_l3 flag for IPv4 and IPv6, but only IPv4
should have header checksumming applied.

Prompted by Linux commit fa6d7cb5d76cf0467c61420fc9238045aedfd379.

Sponsored by: The FreeBSD Foundation

MFC r325813 (bz): Unbreak IPv6.

No longer return ENXIO when trying to send an IPv6 packet in
nicvf_sq_add_hdr_subdesc().
Restructure the code so that the upper layer protocol parts are
agnostic of the L3 protocol (and no longer specific to IPv4).
With this basic IPv6 packets go through. We are still seeing
weird behaviour which needs further diagnosis.

PR: 223669

MFC r325683: vnic: apply BPF tap before passing packet to hardware

Previously we passed tx packets to hardware via nicvf_tx_mbuf_locked
and then to the BPF tap, with a possibly invalid mbuf which would result
in a panic.

PR: 223600
Discussed with: bz
Sponsored by: The FreeBSD Foundation, Packet.net (hardware)

Copy head@r302406 to stable/11 as part of the 11.0-RELEASE cycle.
Prune svn:mergeinfo from the new branch, as nothing has been merged
here.

Additional commits post-branch will follow.

Approved by: re (implicit)
Sponsored by: The FreeBSD Foundation

Fix deadlock in VNIC when using single CPU only

Number of free Tx descriptors does not need to be locked since
it can be modified atomically between SND and CQ tasks.
It will also block Tx routine from sending packets while CQ will not
be able to free descriptors.

Obtained from: Semihalf
Sponsored by: Cavium
Differential Revision: https://reviews.freebsd.org/D6266

Add HW RSS support to VNIC driver

Based on v1.0 driver provided by Cavium under BSD license.
Support in-hardware RSS to distribute IP, UDP and TCP traffic
among available RX Queues and hence multiple CPUs.

Reviewed by: wma
Obtained from: Semihalf
Sponsored by: Cavium
Differential Revision: https://reviews.freebsd.org/D6230

Fix sending TSO packets larger than single DMA segment on VNIC

Due to the bug in the number of 'GATHER' subdescriptors for TSO
packets, VNIC was not able to transmit more than one DMA segment
with TSO enabled.

Obtained from: Semihalf
Sponsored by: Cavium

tcp/lro: Use tcp_lro_flush_all in device drivers to avoid code duplication

And factor out tcp_lro_rx_done, which deduplicates the same logic with
netinet/tcp_lro.c

Reviewed by: gallatin (1st version), hps, zbb, np, Dexuan Cui <decui microsoft com>
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5725

Improve TX path of the VNIC driver

- Avoid memory leak when nicvf_tx_mbuf_locked() fails
- Introduce nicvf_xmit_locked() routine that uses drbr_peek(),
drbr_advance() or drbr_putback() for a specific ifnet.
This gives more clear and efficient design as well as
prevents from dropping mbufs that where not sent due to temporary
lack of descriptors.
- Add missing ETHER_BPF_MTAP() hook

Pointed out by: yongari
Reviewed by: wma
Obtained from: Semihalf
Sponsored by: Cavium
Differential Revision: https://reviews.freebsd.org/D5534

Improve HW checksums support in VNIC

- Do not mark CSUM_IP_CHECKED and CSUM_IP_VALID on IPv6 packets.
IPv6 does not have checksums by definition.
- Set SCTP packets csum_flags CSUM_SCTP_VALID instead of
CSUM_DATA_VALID and skip csum_data
- Set csum_data simply as 0xffff without byteswap

Pointed out by: yongari
Reviewed by: yongari, wma
Obtained from: Semihalf
Sponsored by: Cavium
Differential Revision: https://reviews.freebsd.org/D5537

Don't manage free SQ entry index by the atomic operations

It is not necessary as entries are being manipulated under lock.

Reviewed by: wma
Obtained from: Semihalf
Sponsored by: Cavium
Differential Revision: https://reviews.freebsd.org/D5536

Fix bug in VNIC causing phony number of available TX descriptors

TSO packets will signal segments TX completion in the separate CQ
descriptors. Each CQ descriptor for HW TSO will point to the same
SQ entry.
Do not invoke nicvf_put_sq_desc() for secondary segments to avoid
free_cnt corruption and eventually integer overflow that will result
in the negative free_cnt value and hence impossibility of further
transmission.

Reviewed by: wma
Obtained from: Semihalf
Sponsored by: Cavium
Differential Revision: https://reviews.freebsd.org/D5535

Fix "received NULL mbuf" bug in VNIC

Do not modify NIC_QSET_CQ_0_7_HEAD manually, especially
in non-atomic context.
It doesn't seem to be necessary to recreate CQ head after
interrupt clearing too.

Reviewed by: wma
Obtained from: Semihalf
Sponsored by: Cavium
Differential Revision: https://reviews.freebsd.org/D5533

Introduce HW TSO support for VNIC

This feature was added in Pass2.0.
Significantly improves VNIC's TCP performance on Tx.

Reviewed by: wma
Obtained from: Semihalf
Sponsored by: Cavium
Differential Revision: https://reviews.freebsd.org/D5424

Enable CQ count threshold interrupt on VNIC Pass2.0

On Pass2.0 can trigger interrupt on both timer and CQ count.

Reviewed by: wma
Obtained from: Semihalf
Sponsored by: Cavium
Differential Revision: https://reviews.freebsd.org/D5423

Improve VNIC performance on Tx path by immediate packet transmission

Don't postpone Tx if the Tx lock can be acquired now.
This gives 3x better performance on egress.

Reviewed by: wma
Obtained from: Semihalf
Sponsored by: Cavium
Differential Revision: https://reviews.freebsd.org/D5325

Prefetch next CQ descriptor in Completion Queue handling loop of VNIC

Reviewed by: wma
Obtained from: Semihalf
Sponsored by: Cavium
Differential Revision: https://reviews.freebsd.org/D5322

Enable LRO support for VNIC driver

Support for software LRO when enabled in the capabilities

Reviewed by: wma
Obtained from: Semihalf
Sponsored by: Cavium
Differential Revision: https://reviews.freebsd.org/D5321

Add support for hardware Tx and Rx checksums to VNIC driver

- The network controller verifies Rx TCP/UDP/SCTP checksums by default.
Communicate this to the stack when the packet is not marked as erroneous
to avoid redundant checksum calculation in kernel.
- It is not uncommon to get the mbuf with m_len that is less than
the minimal size for the IP, TCP, UDP, etc. when HW checsumming
is enabled. To avoid data corruption performed by the HW that is
intended to write IP and TCP/UDP/SCTP checksums to the data segment,
the mbuf needs to be pulled up by the required number of bytes.
- Make sure that one can modify the mbufs that require checsum calculation
rather than check for NULL mbuf on each transmission.

Reviewed by: wma
Obtained from: Semihalf
Sponsored by: Cavium
Differential Revision: https://reviews.freebsd.org/D5320

Introduce initial support for Cavium's ThunderX networking interface

- The driver consists of three main componens: PF, VF, BGX
- Requires appropriate entries in DTS and MDIO driver
- Supports only FDT configuration
- Multiple Tx queues and single Rx queue supported
- No RSS, HW checksum and TSO support
- No more than 8 queues per-IF (only one Queue Set per IF)
- HW statistics enabled
- Works in all available MAC modes (1,10,20,40G)
- Style converted to BSD according to style(9)
- The code brings lmac_if interface used by the BGX driver to
update its logical MACs state.

Obtained from: Semihalf
Sponsored by: The FreeBSD Foundation

Raw import of ThunderX VNIC networking driver components

This import brings following components of the Linux driver:
- Thunder BGX (programmable MAC)
- Physical Function driver
- Virtual Function driver
- Headers

Revision: 1.0
Obtained from: Cavium
License information: Cavium provided these files under BSD license