Re: [PATCH V2 rdma-next 2/3] RDMA/hns: Fix the chip hanging caused by sending mailbox&CMQ during reset

From: Wei Hu (Xavier)
Date: Mon Jan 28 2019 - 21:19:03 EST




On 2019/1/29 2:27, Jason Gunthorpe wrote:
> On Sat, Jan 26, 2019 at 09:47:42AM +0800, Wei Hu (Xavier) wrote:
>>
>> On 2019/1/26 5:50, Jason Gunthorpe wrote:
>>> On Fri, Jan 25, 2019 at 10:15:40AM +0800, Wei Hu (Xavier) wrote:
>>>> On 2019/1/25 2:31, Jason Gunthorpe wrote:
>>>>> On Thu, Jan 24, 2019 at 11:13:29AM +0800, Wei Hu (Xavier) wrote:
>>>>>> On 2019/1/24 6:40, Jason Gunthorpe wrote:
>>>>>>> On Sat, Jan 19, 2019 at 11:36:06AM +0800, Wei Hu (Xavier) wrote:
>>>>>>>
>>>>>>>> +static int hns_roce_v2_cmd_hw_resetting(struct hns_roce_dev *hr_dev,
>>>>>>>> + unsigned long instance_stage,
>>>>>>>> + unsigned long reset_stage)
>>>>>>>> +{
>>>>>>>> + struct hns_roce_v2_priv *priv = (struct hns_roce_v2_priv *)hr_dev->priv;
>>>>>>>> + struct hnae3_handle *handle = priv->handle;
>>>>>>>> + const struct hnae3_ae_ops *ops = handle->ae_algo->ops;
>>>>>>>> + unsigned long end;
>>>>>>>> +
>>>>>>>> + /* When hardware reset is detected, we should stop sending mailbox&cmq
>>>>>>>> + * to hardware, and wait until hardware reset finished. If now
>>>>>>>> + * in .init_instance() function, we should exit with error. If now at
>>>>>>>> + * HNAE3_INIT_CLIENT stage of soft reset process, we should exit with
>>>>>>>> + * error, and then HNAE3_INIT_CLIENT related process can rollback the
>>>>>>>> + * operation like notifing hardware to free resources, HNAE3_INIT_CLIENT
>>>>>>>> + * related process will exit with error to notify NIC driver to
>>>>>>>> + * reschedule soft reset process once again.
>>>>>>>> + */
>>>>>>>> + end = msecs_to_jiffies(HNS_ROCE_V2_HW_RST_TIMEOUT) + jiffies;
>>>>>>>> + while (ops->get_hw_reset_stat(handle) && time_before(jiffies, end))
>>>>>>>> + udelay(1);
>>>>>>> I thought you were getting rid of these loops?
>>>>>> Hi, Jason
>>>>>>
>>>>>> Upper applications maybe notify driver to issue mailbox or CMD
>>>>>> commands to hardware, some commands used to cancel resources,
>>>>>> destory bt/destory cq/unreg mr/destory qp etc. when such
>>>>>> commands are executed successfully, the hardware engine will
>>>>>> no longer access some memory registered by the driver.
>>>>>>
>>>>>> When reset occurs, it is possible for upper applications notify driver
>>>>>> to issue mailbox or CMD commands, we need to wait until hardware
>>>>>> reset finished to ensure that hardware no longer accesses related
>>>>>> memory.
>>>>> You should not wait for things using loops like the above.
>>>> Hi, Jason
>>>>
>>>> Are your comments foucsing on udelay? If not, thanks for your detail
>>>> information.
>>>> In hns3 RoCE driver, some CMQ/mailbox operation are called inside
>>>> the lock,
>>>> we can't use msleep in the lock, otherwise it will cause deadlock.
>>>> When reset occurs, RDMA service cannot be provided normally, I think
>>>> in this
>>>> case using udelay will not have a great impact.
>>> You should not use any kind of sleep call in a loop like this.
>> Hi, Jason
>>
>> OK, I got your opinion and will modify it in v3 patch as below:
>>
>> end = msecs_to_jiffies(HNS_ROCE_V2_HW_RST_TIMEOUT) + jiffies;
>> while (time_before(jiffies, end))
>> if (!ops->get_hw_reset_stat(handle))
>> break;
> You shouldn't be looping like this at all, a busy loop is worse, don't
> try and open code spinlocks.
Hi, Jason

OK, we will modify some places calling CMQ/mailbox operation,
replace spinlock with mutex, and add msleep here:

end = msecs_to_jiffies(HNS_ROCE_V2_HW_RST_TIMEOUT) + jiffies;
while (time_before(jiffies, end)) {
if (!ops->get_hw_reset_stat(handle))
break;
msleep(20);
}

Thanks for your comments.

Regards
Xavier
> Jason
>
> .
>