[libvirt] Cannot write data: Broken pipe when accesing libvirt from a forked process

[Cole Robinson]

On 09/30/2015 10:24 AM, Daniel P. Berrange wrote:
> On Wed, Sep 30, 2015 at 01:06:48PM +0200, Josef Stribny wrote:
>> Hi all,
>>
>> I am trying to add vagrant-libvirt support for landrush plugin[0] and found
>> out that libvirt fails for me with:
>>
>>
>> Call to virConnectNumOfNetworks failed: Cannot write data: Broken pipe
>> (Libvirt::RetrieveError)
>>
>> I created a minimal reproducer that causes this:
>>
>> ```
>> #!/usr/bin/ruby
>> require 'libvirt'
>>
>> conn = Libvirt::open("qemu:///system")
>>
>> fork do
>>  puts conn.list_networks
>> end
>>
>> puts conn.list_networks
>> ```
>>
>> This works just fine on my host, but fails on my virtualized guest (when
>> using nested KVM). The journal shows
>> the following lines in logs:
>>
>> Failed to acquire pid file '/run/user/1001/libvirt/libvirtd.pid': Resource
>> temporarily unavailable
>>
>> So the forked process here tries to access not-existing libvirtd pid file
>> for root (1001 on the VM).
>>
>> Is this a feature/bug? How can one avoid it?
> 
> It is a bug in your code. You open a connection to libvirt, then fork
> and then use that single socket connection from two processes. The
> data sent over the socket will be arbitrary interleaved from the
> two processes causing corruption of the protocol, so libvirtd
> kills your connection.
> 

Connections can be shared across threads though, so either 1) share the
connection object across threads instead of separate processes, 2) open a new
connection handle for each separate process

- Cole