This series addresses TX drops observed in production when using veth
devices with threaded NAPI, and refactors a common qdisc check into a
shared helper.
In threaded NAPI mode, packet drops can occur when the ptr_ring backing
the veth peer fills up. This is typically due to a combination of
scheduling delays and the consumer (NAPI thread) being slower than the
producer. When the ring overflows, packets are dropped in veth_xmit().
Patch 1 introduces a backpressure mechanism: when the ring is full, the
driver returns NETDEV_TX_BUSY, signaling the qdisc layer to requeue the
packet. This allows Active Queue Management (AQM) - such as fq or sfq -
to spread traffic more fairly across flows and reduce damage from
elephant flows.
To minimize invasiveness, this backpressure behavior is only enabled when
a qdisc is attached. If no qdisc is present, the driver retains its
original behavior (dropping packets on a full ring), avoiding behavior
changes for configurations without a qdisc.
Detecting the presence of a "real" qdisc relies on a check that is
already duplicated across multiple drivers (e.g., veth, vrf). Patch-2
consolidates this logic into a new helper, qdisc_txq_is_noop(), to avoid
duplication and clarify intent.
---
Jesper Dangaard Brouer (2):
veth: apply qdisc backpressure on full ptr_ring to reduce TX drops
net: sched: generalize check for no-op qdisc
drivers/net/veth.c | 49 ++++++++++++++++++++++++++++++++-------
drivers/net/vrf.c | 3 +--
include/net/sch_generic.h | 7 +++++-
3 files changed, 48 insertions(+), 11 deletions(-)
--
