On 18/03/2021 00:15, Stefan Metzmacher wrote:
> Hi Pavel,
>
>>>>>> here're patches which fix linking of send[msg]()/recv[msg]() calls
>>>>>> and make sure io_uring_enter() never generate a SIGPIPE.
>>>>
>>>> 1/2 breaks userspace.
>>>
>>> Can you explain that a bit please, how could some application ever
>>> have a useful use of IOSQE_IO_LINK with these socket calls?
>>
>> Packet delivery of variable size, i.e. recv(max_size). Byte stream
>> that consumes whatever you've got and links something (e.g. notification
>> delivery, or poll). Not sure about netlink, but maybe. Or some
>> "create a file via send" crap, or some made-up custom protocols
>
> Ok, then we need a flag or a new opcode to provide that behavior?
>
> For recv() and recvmsg() MSG_WAITALL might be usable.
Hmm, unrelated, but there is a good chance MSG_WAITALL with io_uring
is broken because of our first MSG_DONTWAIT attempt.
> It's not defined in 'man 2 sendmsg', but should we use it anyway
> for IORING_OP_SEND[MSG] in order to activate the short send check
> as the low level sock_sendmsg() call seem to ignore unused flags,
> which seems to be the reason for the following logic in tcp_sendmsg_locked:
>
> if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) {
Yep, it maintains compatibility because of unchecked unsupported flags.
Alleviating an old design problem, IIRC.
>
> You need to set SOCK_ZEROCOPY in the socket in order to give a meaning
> to MSG_ZEROCOPY.
>
> Should I prepare an add-on patch to make the short send/recv logic depend
> on MSG_WAITALL?
IMHO, conceptually it would make much more sense with MSG_WAITALL.
>
> I'm cc'ing netdev@xxxxxxxxxxxxxxx in order to more feedback of
> MSG_WAITALL can be passed to sendmsg without fear to trigger
> -EINVAL.
>
> The example for io_sendmsg() would look like this:
>
> --- a/fs/io_uring.c
> +++ b/fs/io_uring.c
> @@ -4383,7 +4383,7 @@ static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
> struct io_async_msghdr iomsg, *kmsg;
> struct socket *sock;
> unsigned flags;
> - int expected_ret;
> + int min_ret = 0;
> int ret;
>
> sock = sock_from_file(req->file);
> @@ -4404,9 +4404,11 @@ static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
> else if (issue_flags & IO_URING_F_NONBLOCK)
> flags |= MSG_DONTWAIT;
>
> - expected_ret = iov_iter_count(&kmsg->msg.msg_iter);
> - if (unlikely(expected_ret == MAX_RW_COUNT))
> - expected_ret += 1;
> + if (flags & MSG_WAITALL) {
> + min_ret = iov_iter_count(&kmsg->msg.msg_iter);
> + if (unlikely(min_ret == MAX_RW_COUNT))
> + min_ret += 1;
> + }
> ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
> if ((issue_flags & IO_URING_F_NONBLOCK) && ret == -EAGAIN)
> return io_setup_async_msg(req, kmsg);
> @@ -4417,7 +4419,7 @@ static int io_sendmsg(struct io_kiocb *req, unsigned int issue_flags)
> if (kmsg->free_iov)
> kfree(kmsg->free_iov);
> req->flags &= ~REQ_F_NEED_CLEANUP;
> - if (ret != expected_ret)
> + if (ret < min_ret)
> req_set_fail_links(req);
> __io_req_complete(req, issue_flags, ret, 0);
> return 0;
>
> Which means the default of min_ret = 0 would result in:
>
> if (ret < 0)
> req_set_fail_links(req);
>
> again...
>
>>>> Sounds like 2/2 might too, does it?
>>>
>>> Do you think any application really expects to get a SIGPIPE
>>> when calling io_uring_enter()?
>>
>> If it was about what I think I would remove lots of old garbage :)
>> I doubt it wasn't working well before, e.g. because of iowq, but
>> who knows
>
> Yes, it was inconsistent before and now it's reliable.
--
Pavel Begunkov
