[OE-core] [PATCH 08/20] udev-cache: Create cache asynchronously

[Richard Tollerton]

Otavio Salvador <otavio at ossystems.com.br> writes:

> On Mon, Aug 4, 2014 at 6:18 PM, Richard Tollerton <rich.tollerton at ni.com> wrote:
>> Otavio Salvador <otavio at ossystems.com.br> writes:
>>
>>> I am not sure about this one. I see the value you are adding here but
>>> I worry how often something can be connected during this process and
>>> change the contents along the way. Did you see something as that
>>> during your tests?
>>
>> No, but point taken. I didn't really test "physical" hotplug during
>> bootup -- and in particular, for this to be a persisting issue, the
>> added/removed device must be removed/added during this boot -- but I
>> suppose that is totally possible.
>>
>> This reflects an underlying race condition, starting at the computation
>> of NEWDATA/OLDDATA in the udev initscript, and ending in the tarball
>> creation in udev-cache. I suppose that we can mitigate this by spinning
>> in udev-cache until the devcache stops changing...? Cheesy, untested,
>> dirty RFC follows:
>
> You need to check the list of files taken by 'tar' and compare. This
> mitigates it but does not really solves the issue.

Hmm, I'm confused. If we relied entirely on /dev contents, how would we
detect (and correct for) if a device was removed between the execution
of udev and of udev-cache?

> Maybe we could "copy" the nodes to a tmp area and use this as a
> snapshot? The copy should be fast and makes it mostly "atomic".

I agree that the copy of the contents of /dev needs to be as fast as
possible -- your about using tmpfs as a staging area is well taken. But
beyond that, it seems to me that the root issue you articulated is about
ensuring adequate synchronization of udev.cache and the tarball;
rereading CMP_FILE_LIST in udev-cache thus seems unavoidable.

> -- 
> Otavio Salvador                             O.S. Systems
> http://www.ossystems.com.br        http://code.ossystems.com.br
> Mobile: +55 (53) 9981-7854            Mobile: +1 (347) 903-9750

-- 
Richard Tollerton <rich.tollerton at ni.com>