Re: [PATCH v4] ixgbe: let the xdpdrv work with more than 64 cpus
From: Jesse Brandeburg <hidden>
Date: 2021-08-26 15:23:55
Also in:
bpf, intel-wired-lan, lkml
On 8/26/2021 7:16 AM, kerneljasonxing@gmail.com wrote:
From: Jason Xing <redacted> Originally, ixgbe driver doesn't allow the mounting of xdpdrv if the server is equipped with more than 64 cpus online. So it turns out that the loading of xdpdrv causes the "NOMEM" failure. Actually, we can adjust the algorithm and then make it work through mapping the current cpu to some xdp ring with the protect of @tx_lock.
Thank you very much for working on this! you should put your sign off block here, and then end with a triple-dash "---" then have your vN: updates below that, so they will be dropped from final git apply. It's ok to have more than one triple-dash.
v4:
- Update the wrong commit messages. (Jason)
v3:
- Change nr_cpu_ids to num_online_cpus() (Maciej)
- Rename MAX_XDP_QUEUES to IXGBE_MAX_XDP_QS (Maciej)
- Rename ixgbe_determine_xdp_cpu() to ixgbe_determine_xdp_q_idx() (Maciej)
- Wrap ixgbe_xdp_ring_update_tail() with lock into one function (Maciej)
v2:
- Adjust cpu id in ixgbe_xdp_xmit(). (Jesper)
- Add a fallback path. (Maciej)
- Adjust other parts related to xdp ring.
Fixes: 33fdc82f08 ("ixgbe: add support for XDP_TX action")
Co-developed-by: Shujin Li <redacted>
Signed-off-by: Shujin Li <redacted>
Signed-off-by: Jason Xing <redacted>
---
drivers/net/ethernet/intel/ixgbe/ixgbe.h | 15 ++++-
drivers/net/ethernet/intel/ixgbe/ixgbe_lib.c | 9 ++-
drivers/net/ethernet/intel/ixgbe/ixgbe_main.c | 64 ++++++++++++++++------
.../net/ethernet/intel/ixgbe/ixgbe_txrx_common.h | 1 +
drivers/net/ethernet/intel/ixgbe/ixgbe_xsk.c | 9 +--
5 files changed, 73 insertions(+), 25 deletions(-)...
quoted hunk ↗ jump to hunk
@@ -8539,21 +8539,32 @@ static u16 ixgbe_select_queue(struct net_device *dev, struct sk_buff *skb, int ixgbe_xmit_xdp_ring(struct ixgbe_adapter *adapter, struct xdp_frame *xdpf) { - struct ixgbe_ring *ring = adapter->xdp_ring[smp_processor_id()]; struct ixgbe_tx_buffer *tx_buffer; union ixgbe_adv_tx_desc *tx_desc; + struct ixgbe_ring *ring; u32 len, cmd_type; dma_addr_t dma; + int index, ret; u16 i; len = xdpf->len; - if (unlikely(!ixgbe_desc_unused(ring))) - return IXGBE_XDP_CONSUMED; + index = ixgbe_determine_xdp_q_idx(smp_processor_id()); + ring = adapter->xdp_ring[index]; + + if (static_branch_unlikely(&ixgbe_xdp_locking_key)) + spin_lock(&ring->tx_lock); + + if (unlikely(!ixgbe_desc_unused(ring))) { + ret = IXGBE_XDP_CONSUMED; + goto out; + }
This static key stuff is tricky code, but I guess if it works, then it's better than nothing. As Maciej also commented, I'd like to see some before/after numbers for some of the xdp sample programs to make sure this doesn't cause a huge regression on the xdp transmit path. A small regression would be ok, since this *is* adding overhead. Jesse