On 2024/9/23 17:02, Leon Romanovsky wrote:
> On Mon, Sep 23, 2024 at 02:17:40PM +0800, Junxian Huang wrote:
>>
>>
>> On 2024/9/20 20:47, Jason Gunthorpe wrote:
>>> On Fri, Sep 20, 2024 at 05:18:14PM +0800, Junxian Huang wrote:
>>>
>>>>>> diff --git a/drivers/infiniband/hw/hns/hns_roce_main.c b/drivers/infiniband/hw/hns/hns_roce_main.c
>>>>>> index 4cb0af733587..49315f39361d 100644
>>>>>> --- a/drivers/infiniband/hw/hns/hns_roce_main.c
>>>>>> +++ b/drivers/infiniband/hw/hns/hns_roce_main.c
>>>>>> @@ -466,6 +466,11 @@ static int hns_roce_mmap(struct ib_ucontext *uctx, struct vm_area_struct *vma)
>>>>>> pgprot_t prot;
>>>>>> int ret;
>>>>>>
>>>>>> + if (hr_dev->dis_db) {
>>>>>
>>>>> How do you clear dis_db after calling to hns_roce_hw_v2_reset_notify_down()? Does it have any locking protection?
>>>>>
>>>>
>>>> Sorry for the late response, I just came back from vacation.
>>>>
>>>> After calling hns_roce_hw_v2_reset_notify_down(), we will call ib_unregister_device()
>>>> and destory all HW resources eventually, so there is no need to clear dis_db.
>>>
>>> Why can't you do the unregister device sooner then and avoid all this
>>> special stuff?
>>>
>>
>> It's a limitation of HW. Resources such as QP/CQ/MR will be destoryed
>> during unregistering device. This is not allowed by HW until
>> hns_roce_hw_v2_reset_notify_uninit(), or it may lead to some HW errors.
>
> It is interested claim given the fact that you are changing original
> code from 2016.
>
Well, this isn't a new issue. We once sent a patch to try to address it
in 2019 [1], but that solution wasn't the right way since it relied on
userspace. We haven't come up with any new solutions since then, until
this series recently.
[1] https://lore.kernel.org/linux-rdma/20190812055220.GA8440@xxxxxxxxxxxxxxxxxx/
Junxian
> Thanks
>
>>
>>> I assumed you'd bring the same device back after completing the reset??
>>>
>>
>> Yes
>>
>>> Jason
>>
