The rx_buf_len parameter was recently added to the Rx zero-copy implementation. The expectation is that when not set system will maintain previous behavior and use the default buffer size (PAGE_SIZE). This works correctly at the iouring level, but we don't preserve the same "zero means default" semantics when registering the memory provider on the netdev. mp_param.rx_page_size is unconditionally set to PAGE_SIZE. This causes __net_mp_open_rxq() to check for QCFG_RX_PAGE_SIZE support in the driver, and return -EOPNOTSUPP for drivers that don't advertise it -- even though the user never asked for large buffers. Only set mp_param.rx_page_size when rx_buf_len was explicitly provided, so that the default page size path works on all zcrx-capable drivers. mlx5 and fbnic only support 4kB pages in the current release. Fixes: 795663b4d160 ("io_uring/zcrx: implement large rx buffer support") Signed-off-by: Jakub Kicinski --- CC: asml.silence@gmail.com CC: axboe@kernel.dk CC: io-uring@vger.kernel.org CC: netdev@vger.kernel.org --- io_uring/zcrx.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/io_uring/zcrx.c b/io_uring/zcrx.c index 28150c6578e3..594220c8eb7e 100644 --- a/io_uring/zcrx.c +++ b/io_uring/zcrx.c @@ -833,7 +833,8 @@ int io_register_zcrx_ifq(struct io_ring_ctx *ctx, if (ret) goto netdev_put_unlock; - mp_param.rx_page_size = 1U << ifq->niov_shift; + if (reg.rx_buf_len) + mp_param.rx_page_size = 1U << ifq->niov_shift; mp_param.mp_ops = &io_uring_pp_zc_ops; mp_param.mp_priv = ifq; ret = __net_mp_open_rxq(ifq->netdev, reg.if_rxq, &mp_param, NULL); -- 2.53.0