[PATCH net V4] veth: Fix TXQ stall race condition

[Jesper Dangaard Brouer <hawk@kernel.org>]

This patchset addresses a race condition introduced in commit dc82a33297fc
("veth: apply qdisc backpressure on full ptr_ring to reduce TX drops"). In
production, this has been observed to cause a permanently stalled transmit
queue (TXQ) on ARM64 (Ampere Altra Max) systems, leading to a "lost wakeup"
scenario where the TXQ remains in the QUEUE_STATE_DRV_XOFF state and traffic
halts.

The root cause is a racy use of the__ptr_ring_empty() API from the producer
side (veth_xmit). The producer stops the queue and then checks the ptr_ring
consumer's head, but this is not guaranteed to be correct, when observed from
the producer side, when the NAPI consumer on another CPU has just finished
consuming.

This series fixes the race bug, making the driver more resilient to recover is
postponed to net-next as maintainers don't see this as an actual fix.

V4:
 - Focus on race fix for stable net-tree
 - Watchdog recovery patch is postponed to net-next tree

V3: https://lore.kernel.org/all/176236363962.30034.10275956147958212569.stgit@firesoul/
 - Don't keep NAPI running when detecting race, because end of veth_poll will
   see TXQ is stopped anyway and wake queue, making it responsibility of the
   producer veth_xmit to do a "flush" that restarts NAPI.

V2: https://lore.kernel.org/all/176159549627.5396.15971398227283515867.stgit@firesoul/
 - Drop patch that changed up/down NDOs
 - For race fix add a smb_rmb and improve commit message reasoning for race cases

V1: https://lore.kernel.org/all/176123150256.2281302.7000617032469740443.stgit@firesoul/

---

Jesper Dangaard Brouer (1):
      veth: more robust handing of race to avoid txq getting stuck


 drivers/net/veth.c | 38 ++++++++++++++++++++------------------
 1 file changed, 20 insertions(+), 18 deletions(-)

--