On Sat, 15 Jun 2019 at 01:25, William Tu <u9012063@xxxxxxxxx> wrote:
>
> Hi,
>
> In my implementation[1], I'm doing s.t like
> tx_done = xsk_ring_cons__peek(&xsk->umem->cq, BATCH_SIZE, &idx_cq);
> if (tx_done > 0) {
> xsk_ring_cons__release(&xsk->umem->cq, tx_done);
> xsk->outstanding_tx -= tx_done;
> }
>
> I expect if tx_done returns 32, then after calling xsk_ring_cons__release,
> the next time I call xsk_ring_cons__peek, I should get idx_cq + 32.
> However, sometimes I see the same value of idx_cq is returned as previous, even
> when tx_done > 0. Is this the expected behavior?
>
> I experiment with xdpsock_user.c with the patch below and run
> ~/bpf-next/samples/bpf# ./xdpsock -l -N -z -i eth3
> using bpf-next commit 5e2ac390fbd08b2a462db66cef2663e4db0d5191
>
Will, apologies for the slow response.
Hmm, it looks like a bug. We'll take a look!
Thanks,
Björn
> --- a/samples/bpf/xdpsock_user.c
> +++ b/samples/bpf/xdpsock_user.c
> @@ -444,6 +443,8 @@ static void kick_tx(struct xsk_socket_info *xsk)
> exit_with_error(errno);
> }
>
> +static int prev_idx_cq;
> +
> static inline void complete_tx_l2fwd(struct xsk_socket_info *xsk)
> {
> u32 idx_cq = 0, idx_fq = 0;
> @@ -463,6 +464,15 @@ static inline void complete_tx_l2fwd(struct
> xsk_socket_info *xsk)
> unsigned int i;
> int ret;
>
> + if (idx_cq == prev_idx_cq) {
> + printf("\n\n ERROR \n\n");
> + }
> + if (idx_cq - prev_idx_cq != rcvd) {
> + printf("\n\n ERROR request %d return %d idx_cq %x prev_cq
> %x diff %d\n",
> + ndescs, rcvd, idx_cq, prev_idx_cq, idx_cq - prev_idx_cq);
> + }
> + prev_idx_cq = idx_cq;
> +
> ret = xsk_ring_prod__reserve(&xsk->umem->fq, rcvd, &idx_fq);
> while (ret != rcvd) {
> if (ret < 0)
> @@ -599,7 +609,7 @@ static void tx_only(struct xsk_socket_info *xsk)
> }
> }
> Most of the time idx_cq - prev_idx_cq == rcvd, but sometimes I got:
> ERROR request 64 return 64 idx_cq 107f54 prev_cq 107f50 diff 4
> ERROR request 64 return 4 idx_cq 2df794 prev_cq 2df754 diff 64
> ERROR request 60 return 60 idx_cq 2df798 prev_cq 2df794 diff 4
> ERROR request 64 return 64 idx_cq 2df7d4 prev_cq 2df798 diff 60
> I wonder if it is a bug?
> [1]https://github.com/williamtu/ovs-ebpf/blob/afxdp-v11/lib/netdev-afxdp.c#L719
>
> Thanks
> William
|