On 06/03/2018 08:47, Alexey Kardashevskiy wrote:
> On 05/03/18 19:36, Paolo Bonzini wrote:
>> I noticed that the introduction of flatview_{read,write} placed
>> address_space_to_flatview outside the RCU lock. This is wrong and has
>> to be fixed, because address_space_to_flatview does an atomic_rcu_read.
>> These patches fix this one function at a time.
>
> out of curiosity - has this caused any actual bug? should be hard to
> reproduce, I suppose...
No, I just noticed it by code inspection, while looking at a
reimplementation of MemoryRegionCache (reverted in 2.9). One of the
differences between address_space_read and address_space_read_cached was
that the latter expects to be called within rcu_read_lock, so it was
surprising that address_space_read was expecting the same. :)
Paolo