Re: [Xen-devel] Rats nest with domain pirq initialisation

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 05/09/18 08:24, Jan Beulich wrote:
>>>> On 04.09.18 at 20:44, <andrew.cooper3@xxxxxxxxxx> wrote:
>> On 13/08/18 11:01, Andrew Cooper wrote:
>>> This is in preparation to set up d->max_cpus and d->vcpu[] in 
>>> domain_create(),
>>> and allow later parts of domain construction to have access to the values.
>>>
>>> Signed-off-by: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>
>>> ---
>>> CC: Jan Beulich <JBeulich@xxxxxxxx>
>>> CC: Stefano Stabellini <sstabellini@xxxxxxxxxx>
>>> CC: Julien Grall <julien.grall@xxxxxxx>
>>> CC: Wei Liu <wei.liu2@xxxxxxxxxx>
>>> ---
>>>  xen/common/domain.c | 34 +++++++++++++++++-----------------
>>>  1 file changed, 17 insertions(+), 17 deletions(-)
>>>
>>> diff --git a/xen/common/domain.c b/xen/common/domain.c
>>> index be51426..0c44f27 100644
>>> --- a/xen/common/domain.c
>>> +++ b/xen/common/domain.c
>>> @@ -322,6 +322,23 @@ struct domain *domain_create(domid_t domid,
>>>          else
>>>              d->guest_type = guest_type_pv;
>>>  
>>> +        if ( !is_hardware_domain(d) )
>>> +            d->nr_pirqs = nr_static_irqs + extra_domU_irqs;
>>> +        else
>>> +            d->nr_pirqs = extra_hwdom_irqs ? nr_static_irqs + 
>>> extra_hwdom_irqs
>>> +                                           : arch_hwdom_irqs(domid);
>>> +        if ( d->nr_pirqs > nr_irqs )
>>> +            d->nr_pirqs = nr_irqs;
>>> +
>>> +        radix_tree_init(&d->pirq_tree);
>>> +    }
>>> +
>>> +    if ( (err = arch_domain_create(d, config)) != 0 )
>>> +        goto fail;
>>> +    init_status |= INIT_arch;
>>> +
>>> +    if ( !is_idle_domain(d) )
>>> +    {
>>>          watchdog_domain_init(d);
>>>          init_status |= INIT_watchdog;
>>>  
>>> @@ -352,16 +369,6 @@ struct domain *domain_create(domid_t domid,
>> Between these two hunks is:
>>
>>         d->iomem_caps = rangeset_new(d, "I/O Memory", 
>> RANGESETF_prettyprint_hex);
>>         d->irq_caps   = rangeset_new(d, "Interrupts", 0);
>>
>> which is important, because it turns out that x86's
>> arch_domain_destroy() depends on d->irq_caps already being initialised.
> Moving this up looks reasonable to me. "Simple" initialization can
> certainly be done early (i.e. before arch_domain_create()), don't
> you think?

No - that defeats the purpose of making the destroy path idempotent. 
For us to remove the max_vcpus hypercall, _domain_destroy() must be
capable of correctly cleaning up a domain from any state of
initialisation, including if the relevant init calls haven't been made yet.

These rangeset_new() calls cannot move earlier than the first action
which might fail (which is the XSM init call to get the security label
correct).

>
>> The path which blows up is:
>>
>> arch_domain_destroy()
>>   free_domain_pirqs()
>>     unmap_domain_pirq()
>>       irq_deny_access()
>>         rangeset_remove_singleton((d)->irq_caps, i)
> But what IRQ do we find to unmap here? There can't be any that have
> been mapped, when ->irq_caps is still NULL. IOW I don't currently see
> how domain_pirq_to_irq() would legitimately return a positive value at
> this point in time, yet that's what guards the calls to unmap_domain_pirq().

It is pirq 2 which explodes, which is the first of the redundant pirq
structures allocated for legacy routing.

I'm not sure I understand this code well enough to comment on why
domain_pirq_to_irq() returns a positive value at this point, but I'm
going to go out on a limb and suggest it might be related to our
unnecessary(?) preallocation.

>
>> Unlike the boolean-nature rangeset_contains_*() helpers, I don't think
>> it is reasonable to make rangeset_remove_*() tolerate a NULL rangeset.
> +1
>
>> The behaviour of automatically revoking irq access is dubious at best. 
>> It is asymmetric with the XEN_DOMCTL_irq_permission, and a caller would
>> reasonably expect not to have to re-grant identical permissions as the
>> irq is mapped/unmapped.  Does anyone know why we have this suspect
>> behaviour in the first place?
> Wasn't it that it was symmetric originally, and the grant/map side has been
> split perhaps a couple of years ago? If so, the unmap side splitting was
> perhaps simply missed?

Perhaps?  I don't know the answers to these.

>
>> One way or another, this path needs to become idempotent, but simply
>> throwing some NULL pointer checks into unmap_domain_pirq() doesn't feel
>> like the right thing to do.
> As per above - I think either free_domain_pirqs() should gain a single
> such NULL check, or domain_pirq_to_irq() should be made sure doesn't
> return positive values prior to ->irq_caps having been set up.
>
>> A separate mess is that we appear to allocate full pirq structures for
>> all legacy irqs for every single domain, in init_domain_irq_mapping(). 
>> At the very least, this is wasteful as very few domains get access to
>> real hardware in the first place.
> I vaguely recall there was some hope to get rid of this, but I don't
> recall the prereqs necessary.

I'm beginning to regret looking at this code.  Whatever is going on, it
looks like it is far more complicated than it needs to be.

It would help if there were even some comments...

>> The other thing I notice is that alloc_pirq_struct() is downright
>> dangerous, as it deliberately tries to allocate half a struct pirq for
>> the !hvm case.  I can only assume this is a space saving measure, but
>> there is absolutely no help in the commit message which introduced it
>> (c/s c24536b636f).
> Space saving, yes. Just like it is forbidden to access d->arch.hvm
> for a PV d, accessing pirq->arch.hvm is forbidden to access for a
> PV domain's pirq. What point is there to allocate the space then?

Because when the code inevitably gets things wrong, you only
read/corrupt your own pirq structure, rather than whichever object
happens to be allocated adjacently.  Most likely, this will be tlsf
metadata.

~Andrew

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel