RE: [PATCH 2/3]: Staging: hv: Use native wait primitives

[KY Srinivasan <kys@microsoft.com>]

> -----Original Message-----
> From: Greg KH [mailto:gregkh@suse.de]
> Sent: Tuesday, February 15, 2011 11:30 AM
> To: KY Srinivasan
> Cc: Jiri Slaby; linux-kernel@vger.kernel.org; devel@linuxdriverproject.org;
> virtualization@lists.osdl.org
> Subject: Re: [PATCH 2/3]: Staging: hv: Use native wait primitives
> 
> On Tue, Feb 15, 2011 at 04:22:20PM +0000, KY Srinivasan wrote:
> >
> >
> > > -----Original Message-----
> > > From: Greg KH [mailto:gregkh@suse.de]
> > > Sent: Tuesday, February 15, 2011 9:03 AM
> > > To: KY Srinivasan
> > > Cc: Jiri Slaby; linux-kernel@vger.kernel.org; devel@linuxdriverproject.org;
> > > virtualization@lists.osdl.org
> > > Subject: Re: [PATCH 2/3]: Staging: hv: Use native wait primitives
> > >
> > > On Tue, Feb 15, 2011 at 01:35:56PM +0000, KY Srinivasan wrote:
> > > >
> > > >
> > > > > -----Original Message-----
> > > > > From: Jiri Slaby [mailto:jirislaby@gmail.com]
> > > > > Sent: Tuesday, February 15, 2011 4:21 AM
> > > > > To: KY Srinivasan
> > > > > Cc: gregkh@suse.de; linux-kernel@vger.kernel.org;
> > > > > devel@linuxdriverproject.org; virtualization@lists.osdl.org
> > > > > Subject: Re: [PATCH 2/3]: Staging: hv: Use native wait primitives
> > > > >
> > > > > On 02/11/2011 06:59 PM, K. Y. Srinivasan wrote:
> > > > > > In preperation for getting rid of the osd layer; change
> > > > > > the code to use native wait interfaces. As part of this,
> > > > > > fixed the buggy implementation in the osd_wait_primitive
> > > > > > where the condition was cleared potentially after the
> > > > > > condition was signalled.
> > > > > ...
> > > > > > @@ -566,7 +567,11 @@ int vmbus_establish_gpadl(struct
> vmbus_channel
> > > > > *channel, void *kbuffer,
> > > > > >
> > > > > >  		}
> > > > > >  	}
> > > > > > -	osd_waitevent_wait(msginfo->waitevent);
> > > > > > +	wait_event_timeout(msginfo->waitevent,
> > > > > > +				msginfo->wait_condition,
> > > > > > +				msecs_to_jiffies(1000));
> > > > > > +	BUG_ON(msginfo->wait_condition == 0);
> > > > >
> > > > > The added BUG_ONs all over the code look scary. These shouldn't be
> > > > > BUG_ONs at all. You should maybe warn and bail out, but not kill the
> > > > > whole machine.
> > > >
> > > > This is Linux code running as a guest on a Windows host; and so the guest
> > > cannot
> > > > tolerate a failure of the host. In the cases where I have chosen to BUG_ON,
> > > there
> > > > is no reasonable recovery possible when the host is non-functional (as
> > > determined
> > > > by a non-responsive host).
> > >
> > > If you have a non-responsive host, wouldn't that imply that this guest
> > > code wouldn't run at all?  :)
> >
> > The fact  that on a particular transaction the host has not responded within an
> expected
> > time interval does not necessarily  mean that the guest code would not be
> running. There may be
> > issues on the host side that may be either transient or permanent that may
> cause problems like
> > this. Keep in mind, HyperV is a type 1 hypervisor that would schedule all VMs
> including the host
> > and so, guest would get scheduled.
> >
> > >
> > > Having BUG_ON() in drivers is not a good idea either way.  Please remove
> > > these in future patches.
> >
> > In situations where there is not a reasonable rollback strategy (for
> > instance in one of the cases, we are granting access to the guest
> > physical pages to the host) we really have only 2 options:
> >
> > 1) Wait until the host responds. This wait could potentially be unbounded
> > and in fact this  was the way the code was to begin with. One of the reviewers
> > had suggested that unbounded wait was to be corrected.
> > 2) Wait for a specific period and if the host does not respond
> > within a reasonable period, kill the guest since there is no recovery
> > possible.
> 
> Killing the guest is a very serious thing, causing all sorts of possible
> problems with it, right?
If there was a reasonable rollback strategy, I would not
be killing the guest.
> 
> > I chose option 2, as part of addressing some of the prior review
> > comments. If the consensus now is to go back to option 1, I am fine with that;
> 
> Unbounded waits aren't ok either, you need some sort of timeout.
> 
> But, as this is a bit preferable to dieing, I suggest doing this, and
> comment the heck out of it to explain all of this for anyone who reads
> it.

If I understand you correctly, you would be prefer to have unbounded waiting with comments 
justifying why we cannot have timeouts. I will roll out a patch once the tree stabilizes.

Regards,

K. Y 
> 
> thanks,
> 
> greg k-h