Re: [PATCH 02/10] Drivers: hv: utils: run polling callback always in interrupt context

[Olaf Hering <olaf@aepfle.de>]

On Fri, Oct 09, Vitaly Kuznetsov wrote:

> Olaf Hering <olaf@aepfle.de> writes:
> 
> > On Thu, Oct 08, KY Srinivasan wrote:
> >
> >> > yes, but after doing fcopy_respond_to_host(). I'd suggest we leave the
> >> > check in place, better safe than sorry.
> >> 
> >> Agreed; Olaf, if it is ok with you, I can fix it up and send.
> >
> > I will retest with this part reverted. I think without two code paths
> > entering hv_fcopy_callback it should be ok to leave this check in.
> 
> I think hv_fcopy_callback() is not involved here: we call fcopy_on_msg()
> every time userspace daemon writes to the device and it is not anyhow
> synchronized with host-guest communication. 

An earlier variant of this patch used locks around the vmbus_recvpacket
and the result was used to decide which thread of execution notifies the
daemon. I think if the interrupt ran earlier than the daemon did the
write then the state expected in fcopy_on_msg would obviously be wrong.
As a result the daemon will just terminate with EFAULT. With the check
removed it would proceed, and either not chancel the timeout or
vmbus_recvpacket reads nothing.

But now that it is single threaded the state in fcopy_on_msg should be
as expected. As said, will retest. Either later today or on Monday.

Olaf