Re: Re: [PATCH 2/2] xen/gnt{dev, alloc}: reserve event channels for notify

[Daniel De Graaf <dgdegra@tycho.nsa.gov>]

On 10/25/2011 04:27 PM, Ian Campbell wrote:
> On Tue, 2011-10-25 at 20:41 +0100, Daniel De Graaf wrote:
>>
>>>> Hmm. Perhaps have a magic value for refcount (-1?) that indicates
>> evtchn_get is not
>>>> available. That would become the default value of refcnt, and
>> evtchn.c would then
>>>> use evtchn_make_refcounted() to change the refcount to 1 and allow
>> _get/_put to work.
>>>
>>> How would that work when the IRQ subsystem (so everything is setup
>> in the kernel)
>>> gets an event? Would the refcount be for that -1.. oh. You would
>> only set
>>> the refcnt when the _get/_put calls are made and not when in-kernel
>> calls to setup> IRQs are done?
>>>
>>
>> Right. The reference count would be a dual-purpose field indicating if
>> the event channel is kernel-internal (value -1) or userspace-visible
>> (reference count > 0). New event channels would start out at -1, and
>> evtchn.c would change them to 1. 
> 
> Is there any way that the reference count could be made part of the
> datastructures associated with the /dev/xen/evtchn driver instead of the
> core evtchn.c stuff? That wouldreduce the chance of current or futures
> users getting something wrong.
> 
> Ian.

This would require that the gntdev and gntalloc modules have a dependency
on the evtchn module, with the evtchn_{get,put} functions moved into that
module. The only other way would be to add a refcount maintenance
function pointer to the IRQ data structure, which seems like it would lead
to more problems than a simple reference count.

-- 
Daniel De Graaf
National Security Agency