[PATCH net 0/3] igb: fix TX stall during XDP teardown with AF_XDP zero-copy

[Alex Dvoretsky <advoretsky@gmail.com>]

When an AF_XDP zero-copy application exits while an XDP program remains
attached, igb can permanently stall a TX queue associated with the
AF_XDP socket. The interface stops forwarding traffic and typically
requires a driver reload to recover.

Reproducer:

  1. Attach an XDP program to igb
  2. Run an AF_XDP zero-copy application
  3. kill -9 the application

The TX watchdog eventually fires and the interface becomes
unresponsive. Reproduced on Intel I210 with Linux 6.17.

igb_clean_rx_irq_zc() lacks a __IGB_DOWN guard. When the AF_XDP process
exits the XSK pool is destroyed, but NAPI continues polling. The
function then repeatedly returns the full budget, which prevents
napi_complete_done() from completing. As a result igb_down() blocks in
napi_synchronize() and TX completions stop being processed, eventually
triggering the TX watchdog.

Patch 1 adds a __IGB_DOWN guard to igb_clean_rx_irq_zc() to break the
infinite NAPI poll loop.

Patch 2 prevents igb_tx_timeout() from scheduling reset_task during XDP
transitions when the device is shutting down.

Patch 3 adds synchronization in igb_xdp_setup() to ensure that pending
ndo_xsk_wakeup() calls complete before the teardown continues, and
refreshes trans_start after igb_open() to prevent false TX timeouts.

igc handles a similar stale trans_start situation via
txq_trans_cond_update() (commit 86ea56c5b0c7). This patch adds
equivalent protection for igb during XDP transitions.

Tested on Intel I210:

  - AF_XDP ZC app exit with XDP attached
  - XDP detach while AF_XDP running
  - repeated XDP attach/detach cycles

Alex Dvoretsky (3):
  igb: check __IGB_DOWN in igb_clean_rx_irq_zc()
  igb: skip reset in igb_tx_timeout() during XDP transition
  igb: add XDP transition guards in igb_xdp_setup()

 drivers/net/ethernet/intel/igb/igb_main.c | 15 +++++++++++++++
 drivers/net/ethernet/intel/igb/igb_xsk.c  |  3 +++
 2 files changed, 18 insertions(+)

--
2.51.0