Il 06/05/2013 13:39, Jan Kiszka ha scritto:
> On 2013-05-06 13:28, Paolo Bonzini wrote:
>> Il 06/05/2013 13:11, Jan Kiszka ha scritto:
>>> On 2013-05-06 12:58, Paolo Bonzini wrote:
>>>> Il 06/05/2013 12:56, Jan Kiszka ha scritto:
>>>>>> The problem is that even if I/O for a region is supposed to happen
>>>>>> within the BQL, lookup can happen outside the BQL. Lookup will use the
>>>>>> region even if it is just to discard it:
>>>>>>
>>>>>> VCPU thread (under BQL) device thread
>>>>>>
>>>>>> --------------------------------------------------------------------------------------
>>>>>> flatview_ref
>>>>>> memory_region_find
>>>>>> returns d->mr
>>>>>> memory_region_ref(d->mr)
>>>>>> /* nop */
>>>>>> qdev_free(d)
>>>>>> object_unparent(d)
>>>>>> unrealize(d)
>>>>>> memory_region_del_subregion(d->mr)
>>>>>> FlatView updated, d->mr not in the new view
>>>>>>
>>>>>> flatview_unref
>>>>>>
>>>>>> memory_region_unref(d->mr)
>>>>>> object_unref(d)
>>>>>> free(d)
>>>>>> if (!d->mr->is_ram) {
>>>>>> /* BAD! */
>>>>>>
>>>>>> memory_region_unref(d->mr) /* nop */
>>>>>> return error
>>>>>> }
>>>>>>
>>>>>>
>>>>>> Here, the memory region is dereferenced *before* we know that it is
>>>>>> BQL-free
>>>>>> (in fact, exactly to ascertain whether it is BQL-free).
>>>>>
>>>>> Both flatview update and lookup *plus* locking type evaluation (i.e.
>>>>> memory region dereferencing) always happen under the address space lock.
>>>>> See Pingfan's patch.
>>>>
>>>> That's true of address_space_rw/map, but I don't think it holds for
>>>> memory_region_find.
>>>
>>> It has to, or it would be broken: Either it is called on a region that
>>> supports reference counting
>>
>> You cannot know that in advance, can you? The address is decided by the
>> guest.
>
> Need to help me again to get the context: In which case is this a
> hot-path that we want to keep BQL-free? Current users of
> memory_region_find appear to be all relatively slow paths, thus are fine
> with staying under BQL.
virtio-blk-dataplane is basically redoing memory_region_find with a
separate data structure, exactly so that it can run outside the BQL
before we get BQL-free MMIO dispatch.
I can try to post patches later today that actually use
memory_region_find instead.
>>> and, thus, increments the counter before
>>> returning, or it has to be called with the BQL held.
>>
>> ... or we need to support reference counting on all regions, so that the
>> other possibility is automatically true.
>>
>> Strictly speaking, only regions that can be unplugged need to support
>> reference counting.
>
> That should make the conversion, if actually required, more bearable.
> Having to assign an owner to every region around as a precondition to
> introduce a new concept with initially less than a handful of users
> would be too much, I suppose.
I agree. Though I have some hope that it could be scripted, this
exception would make the conversion possible for non-qdevified devices.
Paolo
