[PATCH] Support FreeBSD jails for testing

[Steven Hartland]

On 19/10/2015 18:57, Maxim Dounin wrote:
> Hello!
>
> On Fri, Oct 16, 2015 at 06:24:11PM +0100, Steven Hartland wrote:
>
>> On 16/10/2015 13:05, Maxim Dounin wrote:
>>> Hello!
>>>
>>> On Fri, Oct 16, 2015 at 12:09:49AM +0000, Steven Hartland wrote:
>>>
>>>> # HG changeset patch
>>>> # User Steven Hartland <steven.hartland at multiplay.co.uk>
>>>> # Date 1444954080 0
>>>> #      Fri Oct 16 00:08:00 2015 +0000
>>>> # Node ID c22d8299e7040e0de6f85b4e96d0dd953f7af644
>>>> # Parent  78b4e12e6efe642aff591234db0f0b040cae9b5e
>>>> Support FreeBSD jails for testing
>>>>
>>>> Ensure the test directory is read and writable to the test user.
>>>>
>>>> If you request 127.0.0.1 in a FreeBSD jail without specific access
>>>> to 127.0.0.1 then the socket binds to the interface address to
>>>> maintain compatibility. This results in the log entries being
>>> >from the bound interface address. To prevent failure compare
>>>> with the bound IP when requesting 127.0.0.1 in combined test.
>>> This jails behaviour is known to cause many problems, in
>>> particular, it makes impossible nginx binary upgrades unless all
>>> listen sockets are explicitly bound to jail's IP address.
>>>
>>> Fortunately, this was resolved several years ago by introduction
>>> of multi-IP jails.  You may try to use them for tests instead.
>>>
>>> Adding quirks everywhere to support this brain-damaged "no
>>> 127.0.0.1" case looks like a wrong way to go for me.  Especially
>>> given the fact that simple solution exists for years.
>>>
>>> [...]
>> That doesn't fix the directory permission issue which causes pretty much
>> every test to fail, so is this still an option for inclusion?
> Directory premissions may vary depending on umask used.  If your
> umask breaks tests for you - you may try changing it while running
> tests.  It also shouldn't be important when running tests under
> non-privileged user.
Confirmed it doesn't effect a non-privileged user.

The default umask is 022 hence the issue if its run as privileged user 
with the default user as nobody which will have no access to the created 
directory.
> I don't think this change should be added at all, and don't
> see how it's related to FreeBSD jails your patch says it's about.
Fair comment with regards relevance, it clearly would prevent the tests 
being run as a privileged user on any Unix platform which has a 022 
umask not just a FreeBSD jail.

Something to mention in the README, if you don't like the chmod?
>
>> With regards to binding 127.0.0.1, yes its possible to bind it using multi
>> IP, but doing so breaks security if you share it with the host, so its only
>> possible in some situations and usually only for a proper loop back address
>> which wouldn't be 127.0.0.1 just in that /24.
> AFAIR, multi-IP jails used to provide per-jail loopback
> addresses.  But may be I'm wrong here and mistaken with wildcard
> addresses.
Not aware of that, you can assign any address you like but it still 
needs to consciously done and any sharing with the host has the 
potential to cause issues.

Maybe you're thinking of vnet support which provides a jail with their 
own virtual network stack, however that requires a kernel compiled with 
the VIMAGE option.
>
>> I do agree quirks aren't ideal, but as its only the one test I thought it
>> would be nice to have given there's a simple and reliable way to correct
>> said test.
>>
>> With this in mind would you be up for making an exception in this case?
> As long as it's the only test affected we may consider it.
> Sergey, could you please take a look?
>