On 2018/9/7 2:47, Jason Gunthorpe wrote:
> On Thu, Sep 06, 2018 at 04:49:40PM +0800, Yixian Liu wrote:
>> diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
>> index b16ad95..5b65f2b 100644
>> +++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
>> @@ -1255,6 +1255,10 @@ static int hns_roce_v2_profile(struct hns_roce_dev *hr_dev)
>> HNS_ROCE_CAP_FLAG_RQ_INLINE |
>> HNS_ROCE_CAP_FLAG_RECORD_DB |
>> HNS_ROCE_CAP_FLAG_SQ_RECORD_DB;
>> +
>> + if (hr_dev->pci_dev->revision == 0x21)
>> + caps->flags |= HNS_ROCE_CAP_FLAG_MW;
>
> No extra horizontal space
ok, will fix it next version.
>
>> diff --git a/drivers/infiniband/hw/hns/hns_roce_main.c b/drivers/infiniband/hw/hns/hns_roce_main.c
>> index c5cae9a..ff71d91 100644
>> +++ b/drivers/infiniband/hw/hns/hns_roce_main.c
>> @@ -584,6 +584,12 @@ static int hns_roce_register_device(struct hns_roce_dev *hr_dev)
>> ib_dev->uverbs_cmd_mask |= (1ULL << IB_USER_VERBS_CMD_REREG_MR);
>> }
>>
>> + /* MW */
>> + if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_MW) {
>> + ib_dev->alloc_mw = hns_roce_alloc_mw;
>> + ib_dev->uverbs_cmd_mask |= (1ULL << IB_USER_VERBS_CMD_ALLOC_MW);
>> + }
>
> No extra horizontal space
Ok, will fix it next version.
>
>> +struct ib_mw *hns_roce_alloc_mw(struct ib_pd *ib_pd, enum ib_mw_type type,
>> + struct ib_udata *udata)
>> +{
>> + struct hns_roce_dev *hr_dev = to_hr_dev(ib_pd->device);
>> + struct hns_roce_mw *mw;
>> + unsigned long index = 0;
>> + int ret;
>> +
>> + mw = kmalloc(sizeof(*mw), GFP_KERNEL);
>> + if (!mw)
>> + return ERR_PTR(-ENOMEM);
>> +
>> + /* Allocate a key for mw from bitmap */
>> + ret = hns_roce_bitmap_alloc(&hr_dev->mr_table.mtpt_bitmap, &index);
>> + if (ret)
>> + goto err_bitmap;
>> +
>> + mw->rkey = hw_index_to_key(index);
>> +
>> + if (ib_pd->uobject) {
>> + mw->ibmw.device = ib_pd->device;
>> + mw->ibmw.pd = ib_pd;
>> + mw->ibmw.uobject = ib_pd->uobject;
>
> What is this?
>
> ib_uverbs_alloc_mw does all of this:
Ok, I didn't notice ib_uverbs has done these things, will delete them next version.
>
> mw->device = pd->device;
> mw->pd = pd;
> mw->uobject = uobj;
>
> Why is there a user-space difference to this flow anyhow? very
> surprising.
>
Yes, it is a misunderstanding that the mw can be allocated from kernel before I
double checked that it is not real. I will fix it next version.
