Re: [RFC PATCH 0/6] Removal of Aiocontext lock and usage of subtree drains in aborted transactions

[Kevin Wolf]

Am 15.12.2021 um 13:34 hat Hanna Reitz geschrieben:
> On 14.12.21 19:10, Emanuele Giuseppe Esposito wrote:
> > 
> > 
> > On 13/12/2021 15:52, Stefan Hajnoczi wrote:
> > > Off-topic: I don't understand the difference between the effects of
> > > bdrv_drained_begin() and bdrv_subtree_drained_begin(). Both call
> > > aio_disable_external(aio_context) and aio_poll(). bdrv_drained_begin()
> > > only polls parents and itself, while bdrv_subtree_drained_begin() also
> > > polls children. But why does that distinction matter? I wouldn't know
> > > when to use one over the other.
> > 
> > Good point. Now I am wondering the same, so it would be great if anyone
> > could clarify it.
> 
> As far as I understand, bdrv_drained_begin() is used to drain and stop
> requests on a single BDS, whereas bdrv_subtree_drained_begin() drains the
> BDS and all of its children.  So when you don’t care about lingering
> requests in child nodes, then bdrv_drained_begin() suffices.

Right. This is different in practice when a child node has multiple
parents. Usually, when you want to quiesce one parent, the other parent
can keep using the child without being in the way.

For example, two qcow2 overlays based on a single template:

    vda             vdb
     |               |
     v               v
   qcow2           qcow2
(vda.qcow2)     (vdb.qcow2)
     |               |
     +-----+   +-----+
           |   |
           v   v
           qcow2
      (template.qcow2)

If you drain vdb.qcow2 because you want to safely modify something in
its BlockDriverState, there is nothing that should stop vda.qcow2 from
processing requests.

If you're not sure which one to use, bdrv_drained_begin() is what you
want. If you want bdrv_subtree_drained_begin(), you'll know. (It's
currently only used by reopen and by drop_intermediates, which both
operate on more than one node.)

Kevin