Re: [Qemu-devel] [PATCH v6 01/13] block: Add op blocker type "device IO"

[Fam Zheng]

On Thu, 05/28 13:24, Kevin Wolf wrote:
> Am 28.05.2015 um 13:00 hat Paolo Bonzini geschrieben:
> > On 28/05/2015 12:55, Fam Zheng wrote:
> > > > Indeed. blk_pause/resume would handle everything in one central place
> > > > in the block layer instead of spreading the logic across all the block
> > > > layer users.
> > >
> > > Sorry, I'm confused. Do you mean there is a way to implement blk_pause
> > > completely in block layer, without the necessity of various notifier 
> > > handlers
> > > in device models?
> > 
> > How would you do that?  Do you have to keep a queue of pending requests
> > in the BlockBackend?  Since bdrv_drain_all may never return (e.g. stuck
> > NFS connection with nfs=hard), the guest can force unbounded allocation
> > in the host, which is bad.
> 
> We already queue requests for things like I/O throttling or
> serialisation. Why would this be any worse?
> 
> > In addition, the BDS doesn't have a list of BlockBackends attached to
> > it.  So you need the BlockBackends to register themselves for
> > pause/resume in some way---for example with a notifier list.
> > 
> > Then it's irrelevant whether it's the device model or the BB that
> > attaches itself to the notifier list.  You can start with doing it in
> > the device models (those that use ioeventfd), and later it can be moved
> > to the BB.  The low-level implementation remains the same.
> 
> The reason for doing it in the block layer is that it's in one place and
> we can be sure that it's applied. We can still in addition modify
> specific users to avoid even trying to send requests, but I think it's
> good to have the single place that always ensures correct functionality
> of the drain instead of making it dependent on the user.
> 

How to do that for the synchronous blk_write callers that don't run in
a coroutine?

Fam