When running the ib_send_bw test bidirectional, there can be ctx->rx_depth + ctx->tx_depth/2 number of outstanding posted work queue entries but the (shared) completion queue is only size ctx->rx_depth. This can lead to completions being lost and the test hanging. Since memory isn't limited in the microbenchmark test and I think the ctx->tx_depth/2 send limit should probably be ctx->tx_depth, make the CQ size be ctx->rx_depth + ctx->tx_depth.
Signed-off-by: Ralph Campbell <[email protected]> diff --git a/send_bw.c b/send_bw.c index f842fb9..9624cfe 100644 --- a/send_bw.c +++ b/send_bw.c @@ -489,7 +489,8 @@ static struct pingpong_context *pp_init_ctx(struct ibv_device *ib_dev, } } - ctx->cq = ibv_create_cq(ctx->context, ctx->rx_depth, NULL, ctx->channel, 0); + ctx->cq = ibv_create_cq(ctx->context, ctx->rx_depth + ctx->tx_depth, + NULL, ctx->channel, 0); if (!ctx->cq) { fprintf(stderr, "Couldn't create CQ\n"); return NULL; _______________________________________________ ewg mailing list [email protected] http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ewg
