Re: [Qemu-devel] [PATCH] Add a memory barrier to guest memory access functions

[Paolo Bonzini]

Il 21/05/2012 14:18, Michael S. Tsirkin ha scritto:
>> > Almost all our
>> > devices were written without any thought given to ordering, so they
>> > basically can and should be considered as all broken.
> Problem is, a lot of code is likely broken even after you sprinkle
> barriers around. For example qemu might write A then B where guest driver
> expects to see B written before A.

This would be a bug in the guest driver, and usually relatively easy to
reproduce.  The specs (I know of UHCI) should be very precise on this
for obvious reasons.

Paolo