Re: [PATCH for-rc 2/9] RDMA/hns: Fix a long wait for cmdq event during reset

From: Junxian Huang
Date: Tue Jul 09 2024 - 02:21:55 EST



On 2024/7/8 19:16, Leon Romanovsky wrote:
> On Mon, Jul 08, 2024 at 05:30:58PM +0800, Junxian Huang wrote:
>>
>>
>> On 2024/7/8 16:59, Leon Romanovsky wrote:
>>> On Mon, Jul 08, 2024 at 04:45:34PM +0800, Junxian Huang wrote:
>>>>
>>>>
>>>> On 2024/7/8 16:27, Leon Romanovsky wrote:
>>>>> On Mon, Jul 08, 2024 at 03:46:26PM +0800, Junxian Huang wrote:
>>>>>>
>>>>>>
>>>>>> On 2024/7/8 15:33, Leon Romanovsky wrote:
>>>>>>> On Mon, Jul 08, 2024 at 02:50:50PM +0800, Junxian Huang wrote:
>>>>>>>>
>>>>>>>>
>>>>>>>> On 2024/7/8 13:38, Leon Romanovsky wrote:
>>>>>>>>> On Mon, Jul 08, 2024 at 10:29:54AM +0800, Junxian Huang wrote:
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> On 2024/7/7 16:30, Leon Romanovsky wrote:
>>>>>>>>>>> On Fri, Jul 05, 2024 at 04:59:30PM +0800, Junxian Huang wrote:
>>>>>>>>>>>> From: wenglianfa <wenglianfa@xxxxxxxxxx>
>>>>>>>>>>>>
>>>>>>>>>>>> During reset, cmdq events won't be reported, leading to a long and
>>>>>>>>>>>> unnecessary wait. Notify all the cmdqs to stop waiting at the beginning
>>>>>>>>>>>> of reset.
>>>>>>>>>>>>
>>>>>>>>>>>> Fixes: 9a4435375cd1 ("IB/hns: Add driver files for hns RoCE driver")
>>>>>>>>>>>> Signed-off-by: wenglianfa <wenglianfa@xxxxxxxxxx>
>>>>>>>>>>>> Signed-off-by: Junxian Huang <huangjunxian6@xxxxxxxxxxxxx>
>>>>>>>>>>>> ---
>>>>>>>>>>>> drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 18 ++++++++++++++++++
>>>>>>>>>>>> 1 file changed, 18 insertions(+)
>>>>>>>>>>>>
>>>>>>>>>>>> diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
>>>>>>>>>>>> index a5d746a5cc68..ff135df1a761 100644
>>>>>>>>>>>> --- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
>>>>>>>>>>>> +++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
>>>>>>>>>>>> @@ -6977,6 +6977,21 @@ static void hns_roce_hw_v2_uninit_instance(struct hnae3_handle *handle,
>>>>>>>>>>>>
>>>>>>>>>>>> handle->rinfo.instance_state = HNS_ROCE_STATE_NON_INIT;
>>>>>>>>>>>> }
>>>>>>>>>>>> +
>>>>>>>>>>>> +static void hns_roce_v2_reset_notify_cmd(struct hns_roce_dev *hr_dev)
>>>>>>>>>>>> +{
>>>>>>>>>>>> + struct hns_roce_cmdq *hr_cmd = &hr_dev->cmd;
>>>>>>>>>>>> + int i;
>>>>>>>>>>>> +
>>>>>>>>>>>> + if (!hr_dev->cmd_mod)
>>>>>>>>>>>
>>>>>>>>>>> What prevents cmd_mod from being changed?
>>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> It's set when the device is being initialized, and won't be changed after that.
>>>>>>>>>
>>>>>>>>> This is exactly the point, you are assuming that the device is already
>>>>>>>>> ininitialized or not initialized at all. What prevents hns_roce_v2_reset_notify_cmd()
>>>>>>>>> from being called in the middle of initialization?
>>>>>>>>>
>>>>>>>>> Thanks
>>>>>>>>>
>>>>>>>>
>>>>>>>> This is ensured by hns3 NIC driver.
>>>>>>>>
>>>>>>>> Initialization and reset of hns RoCE are both called by hns3. It will check the state
>>>>>>>> of RoCE device (see line 3798), and notify RoCE device to reset (hns_roce_v2_reset_notify_cmd()
>>>>>>>> is called) only if the RoCE device has been already initialized:
>>>>>>>
>>>>>>> So why do you have "if (!hr_dev->cmd_mod)" check in the code?
>>>>>>>
>>>>>>> Thanks
>>>>>>>
>>>>>>
>>>>>> cmd_mod indicates the mode of cmdq (0: poll mode, 1: event mode).
>>>>>> This patch only affects event mode because HW won't report events during reset.
>>>>>
>>>>> You set cmd_mod to 1 in hns_roce_hw_v2_init_instance() without any
>>>>> condition, I don't see when hns v2 IB device is created and continue
>>>>> to operate in polling mode.
>>>>>
>>>>> Thanks
>>>>>
>>>>
>>>> Event mode is the default. In hns_roce_cmd_use_events(), if kcalloc() fails
>>>> then it'll be set to polling mode instead.
>>>
>>> 1. Awesome, and we are returning back to the question. What prevents
>>> hns_roce_v2_reset_notify_cmd() from being called in the middle of
>>> changing cmd_mod from 1 to 0 and from 0 to 1?
>>
>> The changing of cmd_mod is during the initialization of a device. The call
>> of hns_roce_v2_reset_notify_cmd() is during reset. As I said previously,
>> the hns3 NIC driver ensures that there will be no concurrency between
>> initialization and reset, and therefore hns_roce_v2_reset_notify_cmd() won't
>> be called in the middle of changing cmd_mod.
>>
>>> 2. This cmd_mode swtich from 1 to 0 should be removed. Failure to
>>> allocate memory is not a reason to switch to polling mode. The reason
>>> can be HW limitation, but not OS limitation.
>>>
>>
>> But event mode relies on the allocated resource. If the allocation fails and
>> we don't switch to polling mode, the driver won't work any more. Are you suggesting
>> we should return an error and fail the initialization in this case?
>
> Yes, please.

The reason of switching cmd_mod is that we try to keep the driver available,
even if the allocation of event mode resources fails. We don't consider this
as a critical error that should lead to an initialization failure. The driver
can still post mailbox and provide normal functionality in this case.

Our discussion seems to have strayed a bit away？This patch doesn't involve
polling mode.

Junxian

>
> Thanks
>
>>
>> Junxian
>>
>>> Thanks
>>>
>>>>
>>>> Junxian
>>>>
>>>>>>
>>>>>> Junxian
>>>>>>
>>>>>>>>
>>>>>>>> 3791 static int hclge_notify_roce_client(struct hclge_dev *hdev,
>>>>>>>> 3792 enum hnae3_reset_notify_type type)
>>>>>>>> 3793 {
>>>>>>>> 3794 struct hnae3_handle *handle = &hdev->vport[0].roce;
>>>>>>>> 3795 struct hnae3_client *client = hdev->roce_client;
>>>>>>>> 3796 int ret;
>>>>>>>> 3797
>>>>>>>> 3798 if (!test_bit(HCLGE_STATE_ROCE_REGISTERED, &hdev->state) || !client)
>>>>>>>> 3799 return 0;
>>>>>>>> 3800
>>>>>>>> 3801 if (!client->ops->reset_notify)
>>>>>>>> 3802 return -EOPNOTSUPP;
>>>>>>>> 3803
>>>>>>>> 3804 ret = client->ops->reset_notify(handle, type);
>>>>>>>> 3805 if (ret)
>>>>>>>> 3806 dev_err(&hdev->pdev->dev, "notify roce client failed %d(%d)",
>>>>>>>> 3807 type, ret);
>>>>>>>> 3808
>>>>>>>> 3809 return ret;
>>>>>>>> 3810 }
>>>>>>>>
>>>>>>>> And the bit is set (see line 11246) after the initialization has been done (line 11242):
>>>>>>>>
>>>>>>>> 11224 static int hclge_init_roce_client_instance(struct hnae3_ae_dev *ae_dev,
>>>>>>>> 11225 struct hclge_vport *vport)
>>>>>>>> 11226 {
>>>>>>>> 11227 struct hclge_dev *hdev = ae_dev->priv;
>>>>>>>> 11228 struct hnae3_client *client;
>>>>>>>> 11229 int rst_cnt;
>>>>>>>> 11230 int ret;
>>>>>>>> 11231
>>>>>>>> 11232 if (!hnae3_dev_roce_supported(hdev) || !hdev->roce_client ||
>>>>>>>> 11233 !hdev->nic_client)
>>>>>>>> 11234 return 0;
>>>>>>>> 11235
>>>>>>>> 11236 client = hdev->roce_client;
>>>>>>>> 11237 ret = hclge_init_roce_base_info(vport);
>>>>>>>> 11238 if (ret)
>>>>>>>> 11239 return ret;
>>>>>>>> 11240
>>>>>>>> 11241 rst_cnt = hdev->rst_stats.reset_cnt;
>>>>>>>> 11242 ret = client->ops->init_instance(&vport->roce);
>>>>>>>> 11243 if (ret)
>>>>>>>> 11244 return ret;
>>>>>>>> 11245
>>>>>>>> 11246 set_bit(HCLGE_STATE_ROCE_REGISTERED, &hdev->state);
>>>>>>>> 11247 if (test_bit(HCLGE_STATE_RST_HANDLING, &hdev->state) ||
>>>>>>>> 11248 rst_cnt != hdev->rst_stats.reset_cnt) {
>>>>>>>> 11249 ret = -EBUSY;
>>>>>>>> 11250 goto init_roce_err;
>>>>>>>> 11251 }
>>>>>>>>
>>>>>>>> Junxian
>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> Junxian
>>>>>>>>>>
>>>>>>>>>>>> + return;
>>>>>>>>>>>> +
>>>>>>>>>>>> + for (i = 0; i < hr_cmd->max_cmds; i++) {
>>>>>>>>>>>> + hr_cmd->context[i].result = -EBUSY;
>>>>>>>>>>>> + complete(&hr_cmd->context[i].done);
>>>>>>>>>>>> + }
>>>>>>>>>>>> +}
>>>>>>>>>>>> +
>>>>>>>>>>>> static int hns_roce_hw_v2_reset_notify_down(struct hnae3_handle *handle)
>>>>>>>>>>>> {
>>>>>>>>>>>> struct hns_roce_dev *hr_dev;
>>>>>>>>>>>> @@ -6997,6 +7012,9 @@ static int hns_roce_hw_v2_reset_notify_down(struct hnae3_handle *handle)
>>>>>>>>>>>> hr_dev->dis_db = true;
>>>>>>>>>>>> hr_dev->state = HNS_ROCE_DEVICE_STATE_RST_DOWN;
>>>>>>>>>>>>
>>>>>>>>>>>> + /* Complete the CMDQ event in advance during the reset. */
>>>>>>>>>>>> + hns_roce_v2_reset_notify_cmd(hr_dev);
>>>>>>>>>>>> +
>>>>>>>>>>>> return 0;
>>>>>>>>>>>> }
>>>>>>>>>>>>
>>>>>>>>>>>> --
>>>>>>>>>>>> 2.33.0
>>>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>
>>>>
>>>
>>
>