Re: [PATCH iwl-net 1/2] igc: Fix Tx stop threshold to cover empty frame descriptors

[Avigail Dahan <Avigailx.dahan@intel.com>]

On 04/10/2026 18:28, Benoit DE RANCOURT wrote:
> Commit db0b124f02ba ("igc: Enhance Qbv scheduling by using first flag
> bit") raised the number of free descriptors that igc_xmit_frame_ring()
> requires before mapping a packet from count + 3 to count + 5, to leave
> room for the empty frame (one context and one data descriptor) that may
> be inserted ahead of a launch time packet. DESC_NEEDED was left at
> MAX_SKB_FRAGS + 4. igc_tx_map() uses it to stop the queue in advance,
> and TX_WAKE_THRESHOLD is derived from it.
> 
> A queue left running after a transmit therefore only guarantees
> MAX_SKB_FRAGS + 4 free descriptors, while the next skb with a linear
> part and MAX_SKB_FRAGS fragments, each fitting in one data descriptor,
> requires MAX_SKB_FRAGS + 6. igc_xmit_frame_ring() then returns
> NETDEV_TX_BUSY, which the queue stop logic is meant to prevent.
> 
> With TSO this happens routinely. On an I226-V (8086:125c) running 7.2.8
> with MAX_SKB_FRAGS = 17, routed and locally generated TCP traffic on a
> CPU-saturated router produced 10890 NETDEV_TX_BUSY returns in three
> runs of 180 seconds, all for TSO skbs with 16 or 17 fragments, with
> 21 or 22 free descriptors in all but 5 cases. Together with deferred
> tail writes (xmit_more), these returns led to 25 Tx timeouts and
> adapter resets in the same runs:
> 
>    igc 0000:08:00.0 terra: NETDEV WATCHDOG: CPU: 1: transmit queue 2 timed out 5353 ms
>    igc 0000:08:00.0 terra: Reset adapter
> 
> The following patch addresses the lost tail write itself.
> 
> Raise DESC_NEEDED to MAX_SKB_FRAGS + 6 to match the admission check.
> The additional reservation is currently unconditional, including when
> launch time is disabled; match that existing admission policy in the
> proactive stop threshold. With MAX_SKB_FRAGS = 17, the queue is now
> stopped below 23 free descriptors instead of 21. This also raises the
> completion-based wake threshold from 42 to 46 free descriptors,
> preserving the existing two-times-stop-threshold policy. Throughput
> and latency effects were not measured.
> 
> With this change alone, three runs of the same test on 7.2.8 produced
> no NETDEV_TX_BUSY and no Tx timeout over 8.8M NETDEV_TX_OK returns,
> including 613k TSO skbs with 17 fragments.
> 
> As before commit db0b124f02ba, an skb whose linear part or fragments
> exceed IGC_MAX_DATA_PER_TXD may still need more descriptors than
> DESC_NEEDED accounts for.
> 
> Fixes: db0b124f02ba ("igc: Enhance Qbv scheduling by using first flag bit")
> Cc: stable@vger.kernel.org
> Assisted-by: LLM bpftrace
> Signed-off-by: Benoit DE RANCOURT <b2rancourt@gmail.com>
> ---
>   drivers/net/ethernet/intel/igc/igc.h | 8 ++++++--
>   1 file changed, 6 insertions(+), 2 deletions(-)
> 

Tested-by: Avigail Dahan <avigailx.dahan@intel.com>