[PATCH 4.9 210/329] RDMA/qedr: fix QPs ack timeout configuration

From: Greg Kroah-Hartman
Date: Mon May 28 2018 - 06:35:51 EST


4.9-stable review patch. If anyone has any objections, please let me know.

------------------

From: "Kalderon, Michal" <Michal.Kalderon@xxxxxxxxxx>

[ Upstream commit c3594f22302cca5e924e47ec1cc8edd265708f41 ]

QPs that were configured with ack timeout value lower than 1
msec will not implement re-transmission timeout.
This means that if a packet / ACK were dropped, the QP
will not retransmit this packet.

This can lead to an application hang.

Fixes: cecbcddf6 ("qedr: Add support for QP verbs")
Signed-off-by: Michal Kalderon <Michal.Kalderon@xxxxxxxxxx>
Signed-off-by: Ariel Elior <Ariel.Elior@xxxxxxxxxx>
Signed-off-by: Jason Gunthorpe <jgg@xxxxxxxxxxxx>
Signed-off-by: Sasha Levin <alexander.levin@xxxxxxxxxxxxx>
Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
---
drivers/infiniband/hw/qedr/verbs.c | 25 +++++++++++++++----------
1 file changed, 15 insertions(+), 10 deletions(-)

--- a/drivers/infiniband/hw/qedr/verbs.c
+++ b/drivers/infiniband/hw/qedr/verbs.c
@@ -1888,18 +1888,23 @@ int qedr_modify_qp(struct ib_qp *ibqp, s
SET_FIELD(qp_params.modify_flags,
QED_ROCE_MODIFY_QP_VALID_ACK_TIMEOUT, 1);

- qp_params.ack_timeout = attr->timeout;
- if (attr->timeout) {
- u32 temp;
-
- temp = 4096 * (1UL << attr->timeout) / 1000 / 1000;
- /* FW requires [msec] */
- qp_params.ack_timeout = temp;
- } else {
- /* Infinite */
+ /* The received timeout value is an exponent used like this:
+ * "12.7.34 LOCAL ACK TIMEOUT
+ * Value representing the transport (ACK) timeout for use by
+ * the remote, expressed as: 4.096 * 2^timeout [usec]"
+ * The FW expects timeout in msec so we need to divide the usec
+ * result by 1000. We'll approximate 1000~2^10, and 4.096 ~ 2^2,
+ * so we get: 2^2 * 2^timeout / 2^10 = 2^(timeout - 8).
+ * The value of zero means infinite so we use a 'max_t' to make
+ * sure that sub 1 msec values will be configured as 1 msec.
+ */
+ if (attr->timeout)
+ qp_params.ack_timeout =
+ 1 << max_t(int, attr->timeout - 8, 0);
+ else
qp_params.ack_timeout = 0;
- }
}
+
if (attr_mask & IB_QP_RETRY_CNT) {
SET_FIELD(qp_params.modify_flags,
QED_ROCE_MODIFY_QP_VALID_RETRY_CNT, 1);