RDMA/hns: Fix an cmd queue issue when resetting

If a IMP reset caused by some hardware errors and hns RoCE driver reset
occurred at the same time, there is a possiblity that the IMP will stop
dealing with command and users can't use the hardware. The logs are as
follows:

 hns3 0000:fd:00.1: cleaned 0, need to clean 1
 hns3 0000:fd:00.1: firmware version query failed -11
 hns3 0000:fd:00.1: Cmd queue init failed
 hns3 0000:fd:00.1: Upgrade reset level
 hns3 0000:fd:00.1: global reset interrupt

The hns NIC driver divides the reset process into 3 status:
initialization, hardware resetting and softwaring restting. RoCE driver
gets reset status by interfaces provided by NIC driver and commands will
not be sent to the IMP if the driver is in any above status. The main
reason for this issue is that there is a time gap between status 1 and 2,
if the RoCE driver sends commands to the IMP during this gap, the IMP will
stop working because it is not ready.

To eliminate the time gap, the hns NIC driver has added a new interface in
commit a4de02287a ("net: hns3: provide .get_cmdq_stat interface for the
client"), so RoCE driver can ensure that no commands will be sent during
resetting.

Link: https://lore.kernel.org/r/1592314778-52822-1-git-send-email-liweihang@huawei.com
Signed-off-by: Yangyang Li <liyangyang20@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
This commit is contained in:
Yangyang Li 2020-06-16 21:39:38 +08:00 committed by Jason Gunthorpe
parent 98a6151907
commit 3ec5f54f7a
1 changed files with 1 additions and 1 deletions

View File

@ -910,7 +910,7 @@ static int hns_roce_v2_rst_process_cmd(struct hns_roce_dev *hr_dev)
instance_stage = handle->rinfo.instance_state;
reset_stage = handle->rinfo.reset_state;
reset_cnt = ops->ae_dev_reset_cnt(handle);
hw_resetting = ops->get_hw_reset_stat(handle);
hw_resetting = ops->get_cmdq_stat(handle);
sw_resetting = ops->ae_dev_resetting(handle);
if (reset_cnt != hr_dev->reset_cnt)