Re: [PATCH v2 0/2] xennet/xenvif: fix TX doorbell coalescing and improve RX NBL batching

[Tu Dinh <ngoc-tu.dinh@xxxxxxxxxx>]

On 28/09/2026 12:42, david ambu wrote:
> Changes in v2:
> - Replace Co-Authored-By trailer with Assisted-by (no email address)
> - xennet: skip EtherType matching for LLC/SNAP frames (TypeOrLength <= 
> ETHERNET_MTU),
>    zeroing EtherType so those frames are excluded from 
> NDIS_RECEIVE_FLAGS_SINGLE_ETHER_TYPE
>    batching, matching the behaviour of xenvif!__ParseEthernetHeader

Both patches:

Reviewed-by: Tu Dinh <ngoc-tu.dinh@xxxxxxxxxx>

> 
> Background
> ----------
> NDIS passes packets to the miniport as chains of Net Buffer Lists (NBLs),
> each NBL containing one or more Net Buffers (NBs). The xennet transmitter
> signals xenvif's TX ring via a "More" flag: More=TRUE defers the ring
> doorbell (kick), More=FALSE fires it.
> 
> The intent is to batch all NBs within one NBL with a single kick at the
> NBL boundary. Each NBL carries a single hash, so all its NBs map to the
> same TX ring - making per-NBL kicking both correct and sufficient.
> 
> Patch 1 - xenvif (TX fix)
> --------------------------
> TransmitterQueuePacket was overriding More=FALSE unconditionally in the
> XENVIF_PACKET_HASH_ALGORITHM_NONE case, causing a ring kick on every
> individual NB instead of once per NBL. Removing the override lets the
> More flag from xennet flow through correctly, giving ALGORITHM_NONE the
> same per-NBL kick behaviour as ALGORITHM_TOEPLITZ.
> 
> Patch 2 - xennet (RX batching)
> -------------------------------
> Three improvements to the RX receive-indicate path:
> 
>    - Raise NblBatchMax default to 32 (configurable 1-64 via registry)
>      so NDIS receives larger batches in a single
>      NdisMIndicateReceiveNetBufferLists call, reducing per-indication
>      overhead.
> 
>    - Split batches at EtherType boundaries and set
>      NDIS_RECEIVE_FLAGS_SINGLE_ETHER_TYPE per batch, enabling NDIS to
>      apply protocol-layer optimisations.
> 
>    - In __ReceiverPushPackets: sample Returned before the InterlockedAdd
>      so it is read prior to the implicit full memory barrier that
>      InterlockedAdd provides; use the return value directly as Indicated
>      and remove the now-redundant KeMemoryBarrier.
> 
> Performance (inter-host, Windows desktop VM, 10G, 3 iterations x 60s)
> -----------------------------------------------------------------------
> 
>                       Without patch    With patch (32 NBL default)
> Single-stream TCP     4525 Mbits/s     5550 Mbits/s   (+23%)
> Single-stream UDP      372 Mbits/s      441 Mbits/s   (+19%)
> 8-stream TCP         11931 Mbits/s    17017 Mbits/s   (+43%)
> 8-stream UDP            87 Mbits/s      120 Mbits/s   (+38%)
> TCP latency (psping)   0.8 ms           0.9 ms        (no regression)
> ICMP latency (ping)    <1 ms            <1 ms         (no regression)
> 
> Tested at NblBatchMax = 8, 16, 32, 64. Peak throughput at 8-16;
> larger values show marginal returns. Default of 32 is a conservative
> mid-point. Latency is unaffected across all batch sizes.
> 
> david ambu (2):
>    xenvif: fix TX doorbell coalescing for ALGORITHM_NONE traffic
>    xennet: improve NBL batching receive path
> 
>   src/xennet.inf           |   6 +++
>   src/xennet/adapter.c     |  13 ++++++
>   src/xennet/adapter.h     |   8 ++++
>   src/xennet/receiver.c    | 110 +++++++++++++++++++++++++++++++++++----
>   src/xenvif/transmitter.c |   1 -
>   5 files changed, 107 insertions(+), 31 deletions(-)
> 
> --
> 2.51.0.windows.1
> 



--
Ngoc Tu Dinh | Vates XCP-ng Developer

XCP-ng & Xen Orchestra - Vates solutions

web: https://vates.tech