On 2021/11/22 19:10, Leon Romanovsky wrote:
> On Mon, Nov 22, 2021 at 05:28:31PM +0800, Wenpeng Liang wrote:
>>
>>
>> On 2021/11/22 16:58, Leon Romanovsky wrote:
>>>> entry = to_hns_mmap(rdma_entry);
>>>> pfn = entry->address >> PAGE_SHIFT;
>>>> - prot = vma->vm_page_prot;
>>>>
>>>> - if (entry->mmap_type != HNS_ROCE_MMAP_TYPE_TPTR)
>>>> - prot = pgprot_noncached(prot);
>>>> + switch (entry->mmap_type) {
>>>> + case HNS_ROCE_MMAP_TYPE_DB:
>>>> + prot = pgprot_noncached(vma->vm_page_prot);
>>>> + break;
>>>> + case HNS_ROCE_MMAP_TYPE_TPTR:
>>>> + prot = vma->vm_page_prot;
>>>> + break;
>>>> + case HNS_ROCE_MMAP_TYPE_DWQE:
>>>> + prot = pgprot_device(vma->vm_page_prot);
>>> Everything fine, except this pgprot_device(). You probably need to check
>>> WC internally in your driver and use or pgprot_writecombine() or
>>> pgprot_noncached() explicitly.
>>>
>>> Thanks
>>> .
>>>
>>
>> This issue is also discussed in the v2 version, direct wqe uses
>> this prot on HIP09 can achieve better performance than NC.
>>
>> v2 link: https://patchwork.kernel.org/project/linux-rdma/patch/1622705834-19353-3-git-send-email-liweihang@xxxxxxxxxx/
>
> But isn't it specific to ARM model that behaves such? Will it be the case
> when you move to upgrade your ARM core?
>
> Thanks
>
Although the hns roce engine is a PCIe device, it is integrated into the SoC,
and using pgprot_device() will bring a higher performance improvement.
If the ARM core is upgraded, we will consider this issue again.
Thanks
Wenpeng
>>
>> Thanks
>> Wenpeng
> .
>
