Re: [Qemu-devel] Should we auto-generate IDs?

[Jeff Cody]

(Added Eric back in to the CC list.  Looks like he got dropped
somewhere along the way)

On Wed, Aug 26, 2015 at 11:22:08PM -0400, Programmingkid wrote:
> 
> On Aug 26, 2015, at 6:01 PM, Jeff Cody wrote:
> 
> > On Wed, Aug 26, 2015 at 02:17:17PM -0400, Programmingkid wrote:
> >> 
> >> On Aug 26, 2015, at 2:08 PM, Jeff Cody wrote:
> >> 
> >>> On Wed, Aug 26, 2015 at 01:29:04PM -0400, Programmingkid wrote:
> >>>> 
> >>>> On Aug 26, 2015, at 1:25 PM, Jeff Cody wrote:
> >>>> 
> >>>>> On Wed, Aug 26, 2015 at 06:31:57PM +0200, Markus Armbruster wrote:
> >>>>>> Did you drop cc's intentionally?  I put them right back.
> >>>>>> 
> >>>>>> Programmingkid <address@hidden> writes:
> >>>>>> 
> >>>>>>> On Aug 25, 2015, at 8:38 AM, Markus Armbruster wrote:
> >>>>>>> 
> >>>>>>>> You're proposing to revise a qdev design decision, namely the 
> >>>>>>>> purpose of
> >>>>>>>> IDs.  This has been discussed before, and IDs remained unchanged.
> >>>>>>>> Perhaps it's time to revisit this issue.  Cc'ing a few more people.
> >>>>>>>> 
> >>>>>>>> Relevant prior threads:
> >>>>>>>> * [PATCH] qdev: Reject duplicate and anti-social device IDs
> >>>>>>>> http://thread.gmane.org/gmane.comp.emulators.qemu/71230/focus=72272
> >>>>>>>> * [PATCH 6/6] qdev: Generate IDs for anonymous devices
> >>>>>>>> http://thread.gmane.org/gmane.comp.emulators.qemu/114853/focus=114858
> >>>>>>>> * [PATCH] qdev: Assign a default device ID when none is provided.
> >>>>>>>> http://thread.gmane.org/gmane.comp.emulators.qemu/249702
> >>>>>>>> * IDs in QOM (was: [PATCH] util: Emancipate id_wellformed() from 
> >>>>>>>> QemuOpt
> >>>>>>>> http://thread.gmane.org/gmane.comp.emulators.qemu/299945/focus=300381
> >>>>>>>> 
> >>>>>>> 
> >>>>>>> After reading all the threads, I realize why all the attempts to
> >>>>>>> accept a device ID patch failed.
> >>>>>>> It is because it was assumed everyone would agree on one patch to
> >>>>>>> accept. This is
> >>>>>>> very unlikely. It would take someone in a leadership position to
> >>>>>>> decide which patch
> >>>>>>> should be accepted. From one of the threads above, I saw Anthony
> >>>>>>> Liguori participate.
> >>>>>>> He was in the perfect position to make the choice. The person who is
> >>>>>>> in his position now
> >>>>>>> is Peter Maydell. Maybe we should just ask him to look at all the
> >>>>>>> candidate patches and
> >>>>>>> have him pick one to use. 
> >>>>>> 
> >>>>>> Yes, when no consensus emerges, problems tend to go unsolved.
> >>>>>> 
> >>>>>> Before we appeal to authority to break the deadlock, we should make
> >>>>>> another attempt at finding consensus.
> >>>>>> 
> >>>>>> I know that we've entertained the idea of automatically generated IDs
> >>>>>> for block layer objects (that's why I cc'ed some block guys).
> >>>>> 
> >>>>> Yeah, I was one of the ones that proposed some auto-generated IDs for
> >>>>> the block layer, specifically for BlockDriverState, making use of the
> >>>>> node-name field that Benoit introduced a while ago.  Here is my patch
> >>>>> (not sure if this is the latest version, but it is sufficient for this
> >>>>> discussion):
> >>>>> 
> >>>>> http://patchwork.ozlabs.org/patch/355990/
> >>>>> 
> >>>>> I'm not sure about the requirements needed by device ID names, and
> >>>>> they may of course differ from what I was thinking for BDS entries.
> >>>>> 
> >>>>> Here is what I was after with my patch for node-name auto-generation:
> >>>>> 
> >>>>> * Identifiable as QEMU generated / reserved namespace
> >>>>> 
> >>>>> * Guaranteed uniqueness
> >>>>> 
> >>>>> * Non-predictable (don't want users trying to guess / assume
> >>>>> generated node-names)
> >>>>> 
> >>>>> My approach was overkill in some ways (24 characters!).  But for
> >>>>> better or worse, what I had was:
> >>>>> 
> >>>>>              __qemu##00000000IAIYNXXR
> >>>>>              ^^^^^^^^
> >>>>> QEMU namespace ----|    ^^^^^^^^
> >>>>>                        |     ^^^^^^^^^
> >>>>> Increment counter, unique |         |
> >>>>>                                  |
> >>>>> Random string, to spoil prediction  |
> >>>> 
> >>>> Yikes! 24 characters long. That is a bit much to type. Thank you very 
> >>>> much
> >>>> for your effort.
> >>> 
> >>> IMO, the number of characters to type is pretty low on the list of
> >>> requirements, although it can still be addressed secondary to other
> >>> concerns.
> >>> 
> >>> I should have made this in reply to Markus' other email, because the
> >>> important part of this is try and address his point #2:
> >>> 
> >>> (from Markus' other email):
> >>>> 2. The ID must be well-formed.
> >>> 
> >>> To have a well-formed ID, we need to have know requirements of the ID
> >>> structure (i.e. the why and what of it all)
> >>> 
> >>> I don't know if the three requirements I had above apply to all areas
> >>> in QEMU, but I expect they do, in varying degrees of importance.  The
> >>> length itself can be tweaked.
> >>> 
> >>> Talking with John Snow over IRC (added to the CC), one thing he
> >>> suggested was adding in sub-domain spaces; e.g.:
> >>> 
> >>> __qemu#bn#00000000#IAIYNXXR
> >>> 
> >>> Where the 'bn' in this case would be for Block Nodes, etc..
> >>> 
> >>> This may make the scheme extensible through QEMU, where auto-generated
> >>> IDs are desired.
> >>> 
> >>> (sorry to say, this lengthens things, rather than shortening them!)
> >>> 
> >>> We can, of course, make the string shorter - if the random characters
> >>> are just there for spoiling predictability, then 2-3 should be
> >>> sufficient. We could then end up with something like this:
> >>> 
> >>> __qemu#bn#00000000#XR
> >>> 
> >>> The "__qemu" part of the namespace could be shortened as well, but it
> >>> would be nice if it was easy recognizable as being from QEMU.
> >> 
> >> If this ID format was supported, I'm thinking being able to copy and paste 
> >> from
> >> the monitor is a necessary feature. 
> >> 
> >> Any way it could be shorted? I was hoping no more than three characters 
> >> long. 
> >> 
> > 
> > Likely could be shorter, but something in the realm of three
> > characters doesn't seem very realistic.
> 
> Sure it is. Just set device id's like this: 0, 1, 2, 3, 4, 5, 6....

I'm not married to the ID generation scheme I proposed.  

What I am trying to do, however, is have a technical discussion on
generating an ID in a well-formed manner.  And hopefully, in a way
that is useful to all interested subsystems, if possible.

Do you disagree with the requirements I listed above?  If so, it would
be useful to begin the discussion around that.  For ease of
discussion, I'll list them again:

* Reserved namespaces
* Uniqueness
* Non-predictable (to avoid inadvertently creating a de facto ABI)


. . .

On the generation scheme proposed above:

I understand that something you desire is an ID that is easier to
type.

If we wanted to make it shorter, perhaps we could have the number
counter be variable length:

            qemu#ss#D#XY
              |   | | |
qemu reserved -   | | |
                  | | |
subsystem name ---| | |
                    | |
    counter --------| |
                      |
    2-digit random ---|


The counter would just grow to however many digits are needed.  There
is another benefit to growing that number as well - we can use
whatever integer size we think is adequate in the code, without
affecting the generation scheme.

-Jeff