The cma is currently using a hard-coded value, CMA_IBOE_PACKET_LIFETIME,
for the PacketLifeTime, as it can not be determined from the network.
This value might not be optimal for all networks.
The cma module supports the function rdma_set_ack_timeout to set the
ACK timeout for a QP associated with a connection. As per IBTA 12.7.34
local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
Assuming a negligible local ACK delay, we can use
PacketLifeTime = local ACK timeout/2
as a reasonable approximation for RoCE networks.
Signed-off-by: Dag Moxnes <[email protected]>
---
drivers/infiniband/core/cma.c | 15 +++++++++++++--
1 file changed, 13 insertions(+), 2 deletions(-)
diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c
index c8566a4237..2c1b08bde2 100644
--- a/drivers/infiniband/core/cma.c
+++ b/drivers/infiniband/core/cma.c
@@ -2530,7 +2530,9 @@ EXPORT_SYMBOL(rdma_set_service_type);
* This function should be called before rdma_connect() on active side,
* and on passive side before rdma_accept(). It is applicable to primary
* path only. The timeout will affect the local side of the QP, it is not
- * negotiated with remote side and zero disables the timer.
+ * negotiated with remote side and zero disables the timer. In case it is
+ * set before rdma_resolve_route, the value will also be used to determine
+ * PacketLifeTime for RoCE.
*
* Return: 0 for success
*/
@@ -2939,7 +2941,16 @@ static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
route->path_rec->rate = iboe_get_rate(ndev);
dev_put(ndev);
route->path_rec->packet_life_time_selector = IB_SA_EQ;
- route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
+ /* In case ACK timeout is set, use this value to calculate
+ * PacketLifeTime. As per IBTA 12.7.34,
+ * local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
+ * Assuming a negligible local ACK delay, we can use
+ * PacketLifeTime = local ACK timeout/2
+ * as a reasonable approximation for RoCE networks.
+ */
+ route->path_rec->packet_life_time = id_priv->timeout_set ?
+ id_priv->timeout - 1 : CMA_IBOE_PACKET_LIFETIME;
+
if (!route->path_rec->mtu) {
ret = -EINVAL;
goto err2;
--
2.20.1
On Wed, Oct 30, 2019 at 01:44:00PM +0100, Dag Moxnes wrote:
> The cma is currently using a hard-coded value, CMA_IBOE_PACKET_LIFETIME,
> for the PacketLifeTime, as it can not be determined from the network.
> This value might not be optimal for all networks.
>
> The cma module supports the function rdma_set_ack_timeout to set the
> ACK timeout for a QP associated with a connection. As per IBTA 12.7.34
> local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
> Assuming a negligible local ACK delay, we can use
> PacketLifeTime = local ACK timeout/2
> as a reasonable approximation for RoCE networks.
>
> Signed-off-by: Dag Moxnes <[email protected]>
> ---
> drivers/infiniband/core/cma.c | 15 +++++++++++++--
> 1 file changed, 13 insertions(+), 2 deletions(-)
This seems like a reasonable thing to do, applied to for-next
Thanks,
Jason