Re: [PATCH v3 4/5] evtchn: convert domain event lock to an r/w one

[Jan Beulich <jbeulich@suse.com>]

On 21.12.2020 18:45, Julien Grall wrote:
> Hi Jan,
> 
> On 14/12/2020 09:40, Jan Beulich wrote:
>> On 11.12.2020 11:57, Julien Grall wrote:
>>> On 11/12/2020 10:32, Jan Beulich wrote:
>>>> On 09.12.2020 12:54, Julien Grall wrote:
>>>>> On 23/11/2020 13:29, Jan Beulich wrote:
>>>>>> @@ -620,7 +620,7 @@ int evtchn_close(struct domain *d1, int
>>>>>>         long           rc = 0;
>>>>>>     
>>>>>>      again:
>>>>>> -    spin_lock(&d1->event_lock);
>>>>>> +    write_lock(&d1->event_lock);
>>>>>>     
>>>>>>         if ( !port_is_valid(d1, port1) )
>>>>>>         {
>>>>>> @@ -690,13 +690,11 @@ int evtchn_close(struct domain *d1, int
>>>>>>                     BUG();
>>>>>>     
>>>>>>                 if ( d1 < d2 )
>>>>>> -            {
>>>>>> -                spin_lock(&d2->event_lock);
>>>>>> -            }
>>>>>> +                read_lock(&d2->event_lock);
>>>>>
>>>>> This change made me realized that I don't quite understand how the
>>>>> rwlock is meant to work for event_lock. I was actually expecting this to
>>>>> be a write_lock() given there are state changed in the d2 events.
>>>>
>>>> Well, the protection needs to be against racing changes, i.e.
>>>> parallel invocations of this same function, or evtchn_close().
>>>> It is debatable whether evtchn_status() and
>>>> domain_dump_evtchn_info() would better also be locked out
>>>> (other read_lock() uses aren't applicable to interdomain
>>>> channels).
>>>>
>>>>> Could you outline how a developper can find out whether he/she should
>>>>> use read_lock or write_lock?
>>>>
>>>> I could try to, but it would again be a port type dependent
>>>> model, just like for the per-channel locks.
>>>
>>> It is quite important to have clear locking strategy (in particular
>>> rwlock) so we can make correct decision when to use read_lock or write_lock.
>>>
>>>> So I'd like it to
>>>> be clarified first whether you aren't instead indirectly
>>>> asking for these to become write_lock()
>>>
>>> Well, I don't understand why this is a read_lock() (even with your
>>> previous explanation). I am not suggesting to switch to a write_lock(),
>>> but instead asking for the reasoning behind the decision.
>>
>> So if what I've said in my previous reply isn't enough (including the
>> argument towards using two write_lock() here), I'm struggling to
>> figure what else to say. The primary goal is to exclude changes to
>> the same ports. For this it is sufficient to hold just one of the two
>> locks in writer mode, as the other (racing) one will acquire that
>> same lock for at least reading. The question whether both need to use
>> writer mode can only be decided when looking at the sites acquiring
>> just one of the locks in reader mode (hence the reference to
>> evtchn_status() and domain_dump_evtchn_info()) - if races with them
>> are deemed to be a problem, switching to both-writers will be needed.
> 
> I had another look at the code based on your explanation. I don't think 
> it is fine to allow evtchn_status() to be concurrently called with 
> evtchn_close().
> 
> evtchn_close() contains the following code:
> 
>    chn2->state = ECS_UNBOUND;
>    chn2->u.unbound.remote_domid = d1->domain_id;
> 
> Where chn2 is a event channel of the remote domain (d2). Your patch will 
> only held the read lock for d2.
> 
> However evtchn_status() expects the event channel state to not change 
> behind its back. This assumption doesn't hold for d2, and you could 
> possibly end up to see the new value of chn2->state after the new 
> chn2->u.unbound.remote_domid.
> 
> Thanksfully, it doesn't look like chn2->u.interdomain.remote_domain 
> would be overwritten. Otherwise, this would be a straight dereference of 
> an invalid pointer.
> 
> So I think, we need to held the write event lock for both domain.

Well, okay. Three considerations though:

1) Neither evtchn_status() nor domain_dump_evtchn_info() appear to
have a real need to acquire the per-domain lock. They could as well
acquire the per-channel ones. (In the latter case this will then
also allow inserting the so far missing process_pending_softirqs()
call; it shouldn't be made with a lock held.)

2) With the double-locking changed and with 1) addressed, there's
going to be almost no read_lock() left. hvm_migrate_pirqs() and
do_physdev_op()'s PHYSDEVOP_eoi handling, evtchn_move_pirqs(), and
hvm_dpci_msi_eoi(). While for these it may still be helpful to be
possible to run in parallel, I then nevertheless wonder whether the
change as a whole is still worthwhile.

3) With the per-channel double locking and with 1) addressed I
can't really see the need for the double per-domain locking in
evtchn_bind_interdomain() and evtchn_close(). The write lock is
needed for the domain allocating a new port or freeing one. But why
is there any need for holding the remote domain's lock, when its
side of the channel gets guarded by the per-channel lock anyway?
Granted the per-channel locks may then need acquiring a little
earlier, before checking the remote channel's state. But this
shouldn't be an issue.

I guess I'll make addressing 1) and 3) prereq patches to this one,
unless I learn of reasons why things need to remain the way they
are.

Jan