Re: [PATCH v4 2/6] io_uring/zcrx: return error from io_zcrx_map_area_*

[Pavel Begunkov <asml.silence@gmail.com>]

On 7/2/25 23:27, David Wei wrote:
> On 2025-07-02 07:29, Pavel Begunkov wrote:
>> io_zcrx_map_area_*() helpers return the number of processed niovs, which
>> we use to unroll some of the mappings for user memory areas. It's
>> unhandy, and dmabuf doesn't care about it. Return an error code instead
>> and move failure partial unmapping into io_zcrx_map_area_umem().
>>
>> Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
>> ---
>>   io_uring/zcrx.c | 27 ++++++++++++++-------------
>>   1 file changed, 14 insertions(+), 13 deletions(-)
>>
>> diff --git a/io_uring/zcrx.c b/io_uring/zcrx.c
>> index 99a253c1c6c5..2cde88988260 100644
>> --- a/io_uring/zcrx.c
>> +++ b/io_uring/zcrx.c
> 
> ...
> 
>> @@ -254,29 +254,30 @@ static int io_zcrx_map_area_umem(struct io_zcrx_ifq *ifq, struct io_zcrx_area *a
>>               break;
>>           }
>>       }
>> -    return i;
>> +
>> +    if (i != area->nia.num_niovs) {
>> +        __io_zcrx_unmap_area(ifq, area, i);
>> +        return -EINVAL;
>> +    }
>> +    return 0;
>>   }
> 
> Does io_release_dmabuf() still need to be called in
> io_zcrx_map_area_dmabuf() in case of failure, if say
> net_mp_niov_set_dma_addr() fails in the loop?

It doesn't. The function only sets niov addresses, and it's fine
to leave garbage on failure. In contrast, the umem version sets
up dma mappings and needs to unmap if there is an error.

-- 
Pavel Begunkov