[Qemu-devel] IOTEST 162

[Maxim Levitsky]

Hi!

I just had a very fun rabbit hole dive, and I want to share it with you.

I notice for some time that iotest 162 fails with that:

-qemu-img: Could not open 'json:{"driver": "nbd", "host": 42}': Failed to 
connect socket: Invalid argument
+qemu-img: Could not open 'json:{"driver": "nbd", "host": 42}': Failed to 
connect socket:
Connection timed out


I didn't bother to report it yet as it started failing more or less when qemu 
4.2 merge window opened,
so I thought that maybe something got broken, or maybe something broken in my 
environment 
(ahem, AF_UNIX path is too long, ahem)

I just didn't had enough time to pay attention to this. Until today.

So I asked Kevin Wolf today, just by the way why I see for some time that
iotest 162 fails. for me.


He told me that it works for him.
So I look at the test and I see that it just does

qemu-img info 'json:{"driver": "nbd", "host": 42}'

Supposedly it should fail because 42 is not quoted, and it 
doesn't look like a valid host name.

I try with disto's qemu-img (2.11) and I still see that I get the 'Connection 
timed out'

Then I ask him to try on his system with '42' quoted and it still passes.

All right - this means that this 42 is parsed. He also mentions that he uses 
fedora 29 and I still
use fedora 28. So I start a VM with fedora 30, and yep, there it does work 
correctly.

All right, that 42 must be parsed as a host name. Yep. 'ping 42' on my machine 
tries to ping 0.0.0.42,
and in VM exits immediately with 'Invalid argument'
All right, lets ping 0.0.0.42. Maybe something in hostname parsing changed, 
maybe something parses this on DNS level?
Nope, ping 0.0.0.42 works on my machine, fails with invalid argument in VM...

All right lets strace it

connect(5, {sa_family=AF_INET, sin_port=htons(1025), 
sin_addr=inet_addr("0.0.0.42")}, 16) = -1 EINVAL (Invalid argument)

Same thing passes on my machine....

Hmm... this is something in the kernel. Maybe something in iptables/etc. I 
don't yet know that
area that well to be honest.

So I googled a bit and look what I found:

https://lwn.net/Articles/791086/

And yes, while my machine runs fedora 28, it as usual runs vanilla git master
kernel, which I compile from time to time. Currently I am on 5.3.0-rc4+.

So I must say that nothing less but kernel 5.3.0, breaks iotest 162.

I''l prepare a patch soon to fix this.

Best regards,
        Maxim Levitsky