On 30/10/2023 20:40, Jason Gunthorpe wrote:
> On Mon, Oct 30, 2023 at 07:51:41AM +0000, Zhijian Li (Fujitsu) wrote:
>>
>>
>> On 27/10/2023 13:41, Li Zhijian wrote:
>>> mr->page_list only encodes *page without page offset, when
>>> page_size != PAGE_SIZE, we cannot restore the address with a wrong
>>> page_offset.
>>>
>>> Note that this patch will break some ULPs that try to register 4K
>>> MR when PAGE_SIZE is not 4K.
>>> SRP and nvme over RXE is known to be impacted.
>>>
>>> Signed-off-by: Li Zhijian <lizhijian@xxxxxxxxxxx>
>>> ---
>>> drivers/infiniband/sw/rxe/rxe_mr.c | 6 ++++++
>>> 1 file changed, 6 insertions(+)
>>>
>>> diff --git a/drivers/infiniband/sw/rxe/rxe_mr.c b/drivers/infiniband/sw/rxe/rxe_mr.c
>>> index f54042e9aeb2..61a136ea1d91 100644
>>> --- a/drivers/infiniband/sw/rxe/rxe_mr.c
>>> +++ b/drivers/infiniband/sw/rxe/rxe_mr.c
>>> @@ -234,6 +234,12 @@ int rxe_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sgl,
>>> struct rxe_mr *mr = to_rmr(ibmr);
>>> unsigned int page_size = mr_page_size(mr);
>>>
>>> + if (page_size != PAGE_SIZE) {
>>
>> It seems this condition is too strict, it should be:
>> if (!IS_ALIGNED(page_size, PAGE_SIZE))
>>
I have to say I retract this conclusion. It still misses something.
To support PAGE_SIZE aligned MR, we have to refactor rxe_map_mr_sg() or rxe_set_page()
Currently, rxe_set_page() will be called in the step of page_size, this doesn't split N*PAGE_SIZE memory into
N *page. So when we restore an iova from xarray, the array index is wrong as well.
So i'm going to refactor rxe_map_mr_sg() to iterate the sgl by myself in rxe_map_mr_sg() like SIW does.
Hope this refactor can help RXE to support SZ_4K when PAGE_SIZE!=4K as well.
Thanks
Zhijian
>> So that, page_size with (N * PAGE_SIZE) can work as previously.
>> Because the offset(mr.iova & page_mask) will get lost only when !IS_ALIGNED(page_size, PAGE_SIZE)
>
> That makes sense
>
> Jason
