[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-devel] [PATCH v7 0/7] push mmio dispatch out of big lock

From: Paolo Bonzini
Subject: Re: [Qemu-devel] [PATCH v7 0/7] push mmio dispatch out of big lock
Date: Mon, 06 May 2013 13:47:16 +0200
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:17.0) Gecko/20130311 Thunderbird/17.0.4

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.


reply via email to

[Prev in Thread] Current Thread [Next in Thread]