Am 16.03.21 um 15:23 schrieb Jens Axboe:
> On 3/16/21 8:00 AM, Norman Maurer wrote:
>> Hi there,
>>
>> I think I found a bug in the current io_uring implementation. It seems
>> like recvmsg currently not respect when a fd is set to non-blocking.
>> At the moment recvmsg never returns in this case. I can work around
>> this by using MSG_DONTWAIT but I don’t think this should be needed.
>>
>> I am using the latest 5.12 code base atm.
>
> This is actually "by design" in that system calls that offer a "don't
> block for this operation" (like MSG_DONTWAIT here) will not be looking
> at the O_NONBLOCK flag. Though it is a bit confusing and potentially
> inconsistent, my argument here is that this is the case for system calls
> in general, where even O_NONBLOCK has very hazy semantics depending on
> what system call you are looking at.
>
> The issue is mostly around when to use -EAGAIN to arm async retry, and
> when to return -EAGAIN to the application.
>
> I'd like to hear from others here, but as far as io_uring is concerned,
> we _should_ be consistent in how we treat O_NONBLOCK _depending_ on if
> that system call allows a flags method of passing in nonblock behavior.
As ____sys_recvmsg() has this:
if (sock->file->f_flags & O_NONBLOCK)
flags |= MSG_DONTWAIT;
The difference is that we don't block in__sys_recvmsg() within a worker thread,
but instead handle it via io_arm_poll_handler()?
I think it should be documented, but I guess it might be useful to keep a way
to switch between the 3 available modes in order to find the one that
performances best depending on the workload.
metze
