Re: [PATCH v6 1/3] evtchn: slightly defer lock acquire where possible

[Jan Beulich <jbeulich@xxxxxxxx>]

On 28.05.2021 10:30, Roger Pau Monné wrote:
> On Thu, May 27, 2021 at 07:48:41PM +0100, Julien Grall wrote:
>> On 27/05/2021 12:28, Jan Beulich wrote:
>>> port_is_valid() and evtchn_from_port() are fine to use without holding
>>> any locks. Accordingly acquire the per-domain lock slightly later in
>>> evtchn_close() and evtchn_bind_vcpu().
>>
>> So I agree that port_is_valid() and evtchn_from_port() are fine to use
>> without holding any locks in evtchn_bind_vcpu(). However, this is misleading
>> to say there is no problem with evtchn_close().
>>
>> evtchn_close() can be called with current != d and therefore, there is a
> 
> The only instances evtchn_close is called with current != d and the
> domain could be unpaused is in free_xen_event_channel AFAICT.

As long as the domain is not paused, ->valid_evtchns can't ever
decrease: The only point where this gets done is in evtchn_destroy().
Hence ...

>> risk that port_is_valid() may be valid and then invalid because
>> d->valid_evtchns is decremented in evtchn_destroy().
> 
> Hm, I guess you could indeed have parallel calls to
> free_xen_event_channel and evtchn_destroy in a way that
> free_xen_event_channel could race with valid_evtchns getting
> decreased?

... I don't see this as relevant.

>> Thankfully the memory is still there. So the current code is okayish and I
>> could reluctantly accept this behavior to be spread. However, I don't think
>> this should be left uncommented in both the code (maybe on top of
>> port_is_valid()?) and the commit message.
> 
> Indeed, I think we need some expansion of the comment in port_is_valid
> to clarify all this. I'm not sure I understand it properly myself when
> it's fine to use port_is_valid without holding the per domain event
> lock.

Because of the above property plus the fact that even if
->valid_evtchns decreases, the underlying struct evtchn instance
will remain valid (i.e. won't get de-allocated, which happens only
in evtchn_destroy_final()), it is always fine to use it without
lock. With this I'm having trouble seeing what would need adding
to port_is_valid()'s commentary.

The only thing which shouldn't happen anywhere is following a
port_is_valid() check which has returned false by code assuming
the port is going to remain invalid. But I think that's obvious
when you don't hold any suitable lock.

I do intend to follow Julien's request to explain things more for
evtchn_close().

Jan