[Libguestfs] [nbdkit PATCH v3 15/15] RFC: plugins: Add back-compat for new plugin with old nbdkit

[Eric Blake]

On 03/20/2018 05:00 AM, Richard W.M. Jones wrote:
> On Thu, Mar 08, 2018 at 05:03:11PM -0600, Eric Blake wrote:
>> If we bump NBDKIT_API_VERSION, we have forcefully made older
>> nbdkit reject all plugins that opt to the newer API:
>>
>> $ nbdkit ./plugins/nbd/.libs/nbdkit-nbd-plugin.so --dump-plugin
>> nbdkit: ./plugins/nbd/.libs/nbdkit-nbd-plugin.so: plugin is incompatible with this version of nbdkit (_api_version = 2)
>>
>> But with a bit of finesse, we can make opting in to FUA support
>> orthogonal to NBDKIT_API_VERSION, by introducing a new witness
>> level that the user controls, and by providing sane fallbacks
>> so that newer plugins correctly populate the fields expected by
>> older nbdkit.
>>
>> ---
>> v3: rework entirely, add new variable instead of NBDKIT_API_VERSION
>>
>> This patch is still RFC; if you like it, it's probably better to
>> rebase portions of this patch into the rest of the series, rather
>> than churning NBDKIT_API_VERSION temporarily to 2 and now back to 1
> 
> I do wonder if this patch is really worthwhile.  Old nbdkit / new
> plugin is probably going to go untested in general, and therefore will
> just cause problems/complexity.

Fair enough. I'm happy to drop it (I had a tough enough time figuring 
out how to test it; my initial gut feel was that an old server + new 
plugin.so would die with an assertion failure when the old server 
asserted that .pread was non-NULL, and I was actually pleasantly 
surprised that the failure was graceful due to the incompatible version 
check that occurred prior to the .pread assertion.  Keeping a graceful 
failure is acceptable; I would have been pushing for something like this 
a lot harder if it had actually caused assertion failures).

New nbdkit / old plugin gets plenty of testing (most of our plugins 
still use the old interface, where only a few opt in to the new one, and 
since the testsuite still passes, we've proven that both interfaces are 
acceptable to the new binary).

So I'll go ahead and push 1-14, and then see if I can start playing with 
ways to expose FUA flags through the language bindings (I'll probably 
attempt python first).

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3266
Virtualization:  qemu.org | libvirt.org