Re: [PATCH v2 1/2] io_uring/zcrx: add single shot recvzc

[David Wei <dw@davidwei.uk>]

On 2025-02-21 16:40, Pavel Begunkov wrote:
> On 2/21/25 20:51, David Wei wrote:
>> Currently only multishot recvzc requests are supported, but sometimes
>> there is a need to do a single recv e.g. peeking at some data in the
>> socket. Add single shot recvzc requests where IORING_RECV_MULTISHOT is
>> _not_ set and the sqe->len field is set to the number of bytes to read
>> N.
> 
> There is no oneshot, we need to change the message.

I'll make commit subject/descriptions consistent. There is no single
shot, only multishot with and without limits.

> 
> 
>> There could be multiple completions containing data, like the multishot
>> case, since N bytes could be split across multiple frags. This is
>> followed by a final completion with res and cflags both set to 0 that
>> indicate the completion of the request, or a -res that indicate an
>> error.
>>
>> Signed-off-by: David Wei <dw@davidwei.uk>
>> ---
>>   io_uring/net.c  | 19 +++++++++++++++++--
>>   io_uring/zcrx.c | 17 ++++++++++++-----
>>   io_uring/zcrx.h |  2 +-
>>   3 files changed, 30 insertions(+), 8 deletions(-)
>>
>> diff --git a/io_uring/net.c b/io_uring/net.c
>> index 000dc70d08d0..cae34a24266c 100644
>> --- a/io_uring/net.c
>> +++ b/io_uring/net.c
>> @@ -94,6 +94,7 @@ struct io_recvzc {
>>       struct file            *file;
>>       unsigned            msg_flags;
>>       u16                flags;
>> +    u32                len;
>>       struct io_zcrx_ifq        *ifq;
>>   };
>>   @@ -1241,7 +1242,7 @@ int io_recvzc_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
>>       unsigned ifq_idx;
>>         if (unlikely(sqe->file_index || sqe->addr2 || sqe->addr ||
>> -             sqe->len || sqe->addr3))
>> +             sqe->addr3))
>>           return -EINVAL;
>>         ifq_idx = READ_ONCE(sqe->zcrx_ifq_idx);
>> @@ -1250,6 +1251,12 @@ int io_recvzc_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
>>       zc->ifq = req->ctx->ifq;
>>       if (!zc->ifq)
>>           return -EINVAL;
>> +    zc->len = READ_ONCE(sqe->len);
>> +    if (zc->len == UINT_MAX)
>> +        return -EINVAL;
> 
> The uapi gives u32, if we're using a special value it should
> match the type. ~(u32)0
> 
>> +    /* UINT_MAX means no limit on readlen */
>> +    if (!zc->len)
>> +        zc->len = UINT_MAX;
>>         zc->flags = READ_ONCE(sqe->ioprio);
>>       zc->msg_flags = READ_ONCE(sqe->msg_flags);
>> @@ -1269,6 +1276,7 @@ int io_recvzc_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
>>   int io_recvzc(struct io_kiocb *req, unsigned int issue_flags)
>>   {
>>       struct io_recvzc *zc = io_kiocb_to_cmd(req, struct io_recvzc);
>> +    bool limit = zc->len != UINT_MAX;
>>       struct socket *sock;
>>       int ret;
>>   @@ -1281,7 +1289,7 @@ int io_recvzc(struct io_kiocb *req, unsigned int issue_flags)
>>           return -ENOTSOCK;
>>         ret = io_zcrx_recv(req, zc->ifq, sock, zc->msg_flags | MSG_DONTWAIT,
>> -               issue_flags);
>> +               issue_flags, &zc->len);
>>       if (unlikely(ret <= 0) && ret != -EAGAIN) {
>>           if (ret == -ERESTARTSYS)
>>               ret = -EINTR;
>> @@ -1296,6 +1304,13 @@ int io_recvzc(struct io_kiocb *req, unsigned int issue_flags)
>>           return IOU_OK;
>>       }
>>   +    if (zc->len == 0) {
> 
> If len hits zero we should always complete it, regardless
> of errors the stack might have returned, so might be
> cleaner if you do that check right after io_zcrx_recv().

Sounds good.

> 
>> +        io_req_set_res(req, 0, 0);
>> +
>> +        if (issue_flags & IO_URING_F_MULTISHOT)
>> +            return IOU_STOP_MULTISHOT;
>> +        return IOU_OK;
>> +    }
>>       if (issue_flags & IO_URING_F_MULTISHOT)
>>           return IOU_ISSUE_SKIP_COMPLETE;
>>       return -EAGAIN;
>> diff --git a/io_uring/zcrx.c b/io_uring/zcrx.c
>> index f2d326e18e67..74bca4e471bc 100644
>> --- a/io_uring/zcrx.c
>> +++ b/io_uring/zcrx.c
>> @@ -817,6 +817,7 @@ io_zcrx_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
>>       int i, copy, end, off;
>>       int ret = 0;
>>   +    len = min_t(size_t, len, desc->count);
>>       if (unlikely(args->nr_skbs++ > IO_SKBS_PER_CALL_LIMIT))
>>           return -EAGAIN;
>>   @@ -894,26 +895,32 @@ io_zcrx_recv_skb(read_descriptor_t *desc, struct sk_buff *skb,
>>   out:
>>       if (offset == start_off)
>>           return ret;
>> +    if (desc->count != UINT_MAX)
>> +        desc->count -= (offset - start_off);
> 
> I'd say just set desc->count to it's max value (size_t), and
> never care about checking for limits after.

True, we're limited by IO_SKBS_PER_CALL_LIMIT.

> 
>>       return offset - start_off;
>>   }
>>     static int io_zcrx_tcp_recvmsg(struct io_kiocb *req, struct io_zcrx_ifq *ifq,
>>                   struct sock *sk, int flags,
>> -                unsigned issue_flags)
>> +                unsigned issue_flags, unsigned int *outlen)
>>   {
>> +    unsigned int len = *outlen;
>> +    bool limit = len != UINT_MAX;
>>       struct io_zcrx_args args = {
>>           .req = req,
>>           .ifq = ifq,
>>           .sock = sk->sk_socket,
>>       };
>>       read_descriptor_t rd_desc = {
>> -        .count = 1,
>> +        .count = len,
>>           .arg.data = &args,
>>       };
>>       int ret;
>>         lock_sock(sk);
>>       ret = tcp_read_sock(sk, &rd_desc, io_zcrx_recv_skb);
>> +    if (limit && ret)
>> +        *outlen = len - ret;
>>       if (ret <= 0) {
>>           if (ret < 0 || sock_flag(sk, SOCK_DONE))
>>               goto out;
>> @@ -930,7 +937,7 @@ static int io_zcrx_tcp_recvmsg(struct io_kiocb *req, struct io_zcrx_ifq *ifq,
>>           ret = IOU_REQUEUE;
>>       } else if (sock_flag(sk, SOCK_DONE)) {
>>           /* Make it to retry until it finally gets 0. */
>> -        if (issue_flags & IO_URING_F_MULTISHOT)
>> +        if (!limit && (issue_flags & IO_URING_F_MULTISHOT))
>>               ret = IOU_REQUEUE;
> 
> And with earlier len check in net.c you don't need this change,
> which feels wrong, as it's only here to circumvent some handling
> in net.c, I assume
> 

Yeah, I don't think this is needed anymore, will remove.