Re: [Qemu-devel] [big lock] Discussion about the convention of device's DMA each other after breaking down biglock

[Avi Kivity]

On 09/20/2012 10:51 AM, liu ping fan wrote:
> On Wed, Sep 19, 2012 at 5:23 PM, Avi Kivity <address@hidden> wrote:
>> On 09/19/2012 12:19 PM, liu ping fan wrote:
>>> On Wed, Sep 19, 2012 at 5:14 PM, Paolo Bonzini <address@hidden> wrote:
>>>> Il 19/09/2012 11:11, liu ping fan ha scritto:
>>>>>> > Why not? devA will drop its local lock, devX will retake the big lock
>>>>>> > recursively, devB will take its local lock.  In the end, we have 
>>>>>> > biglock
>>>>>> > -> devB.
>>>>>> >
>>>>> But when adopting local lock, we assume take local lock, then biglock.
>>>>
>>>> No, because the local lock will be dropped before taking the biglock.
>>>> The order must always be coarse->fine.
>>>>
>>> But if we takes coarse firstly, then the mmio-dispatcher will still
>>> contend for the big lock against each other.
>>
>> Can you detail the sequence?
>>
> LOCK(local lock)
> .......................
> LOCK(big lock)
> Access timer/block/network subsystem
> UNLOCK(big lock)
> .....................
> UNLOCK(local lock)

This is an invalid sequence.  Either the subsystem has to be fine-grain
locked, or the lock order has to be reversed.

Before we finish subsystem conversion, an mmio dispatcher may look like:

dev_write(...)
{
    lock(s->lock)
    switch (addr) {
        case REGA:
            ...
        case REGB:
            ...
        case REGC:
            unlock(s->lock)
            lock(big lock)
            lock(s->lock)
            qemu_mod_timer()
            unlock(bit lock)
            break;
        ...
    }
    unlock(s->lock)
}



-- 
error compiling committee.c: too many arguments to function