On Mon, Aug 24, 2026 at 12:17 AM mike <[email protected]> wrote: > > atl_xmit_cleanup() scans hw_ring descriptors for a set dd bit to > know how many are safe to reclaim, reading dd straight from > DMA-coherent memory the NIC concurrently writes, with no barrier. > > A torn or stale read here can make the scan believe nothing is > done when descriptors have genuinely completed, so to_clean comes > back 0 and cleanup is skipped. Since atl_xmit_pkts() only refuses > to send when tx_free is too low, and only cleanup ever grows > tx_free back, a missed cleanup pass can leave tx_free permanently > short: every later send call sees insufficient free descriptors > and transmits nothing, indefinitely, with no way to recover. > > Add a read barrier once the scan's dd checks are complete and > before acting on to_clean. > > Fixes: 2b1472d7150c ("net/atlantic: implement Tx path") > Cc: [email protected] > > Signed-off-by: Mike Murphy <[email protected]> > --- > drivers/net/atlantic/atl_rxtx.c | 6 ++++++ > 1 file changed, 6 insertions(+) > > diff --git a/drivers/net/atlantic/atl_rxtx.c b/drivers/net/atlantic/atl_rxtx.c > index 2d551a341f..34f53cd6cc 100644 > --- a/drivers/net/atlantic/atl_rxtx.c > +++ b/drivers/net/atlantic/atl_rxtx.c > @@ -1162,6 +1162,12 @@ atl_xmit_cleanup(struct atl_tx_queue *txq) > break; > } > > + /* > + * Same dd/DMA race as the Rx path; barrier before > + * acting on what was just read. > + */ > + rte_rmb();
Use rte_io_rmb() version to express as it is a barrier between Device and CPU. Same for patch 1/2 Since it is affecting fastpath performance, I will wait for ack from maintainers. > + > if (to_clean == 0) > return; > > -- > 2.34.1 > >
