Commit 07761b30 authored by Junxian Huang's avatar Junxian Huang Committed by Chengchang Tang
Browse files

RDMA/hns: Fix bonding failure due to wrong reset_state 

driver inclusion
category: feature
bugzilla: https://gitee.com/openeuler/kernel/issues/IB3K00



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

When roce driver is removed during reset, the reset flow of roce may
not be fully completed. This may lead to the reset_state of roce
handler stored in nic driver remaining in a middle state, such as
HNS_ROCE_STATE_RST_DOWN or HNS_ROCE_STATE_RST_UNINIT.

The reset_state won't be cleared even if roce driver is re-inited.
This cause that roce bonding which currently relies on reset_state
fails in this case.

Replace the reset detection for bonding with nic APIs (.ae_dev_resetting()
and .get_hw_reset_stat()), just like the reset detection elsewhere in
roce driver.

Fixes: 26d71e7c ("RDMA/hns: Fix the concurrency error between bond and reset.")
Signed-off-by: default avatarJunxian Huang <huangjunxian6@hisilicon.com>
parent 8b936998

drivers/infiniband/hw/hns/hns_roce_bond.c

+6 −3
Original line number Diff line number Diff line
@@ -543,6 +543,7 @@ static void hns_roce_do_bond(struct hns_roce_bond_group *bond_grp) 


bool is_bond_slave_in_reset(struct hns_roce_bond_group *bond_grp)

{

	const struct hnae3_ae_ops *ops;

	struct hnae3_handle *handle;

	struct net_device *net_dev;

	int i;
@@ -550,9 +551,11 @@ bool is_bond_slave_in_reset(struct hns_roce_bond_group *bond_grp) 
	for (i = 0; i < ROCE_BOND_FUNC_MAX; i++) {

		net_dev = bond_grp->bond_func_info[i].net_dev;

		handle = bond_grp->bond_func_info[i].handle;

		if (net_dev && handle &&

		    handle->rinfo.reset_state != HNS_ROCE_STATE_NON_RST &&

		    handle->rinfo.reset_state != HNS_ROCE_STATE_RST_INITED)

		if (!net_dev || !handle)

			continue;

		ops = handle->ae_algo->ops;

		if (ops->ae_dev_resetting(handle) ||

		    ops->get_hw_reset_stat(handle))

			return true;

	}