2014-08-20 18:03+0200, Paolo Bonzini: > Il 20/08/2014 18:01, Radim Krčmář ha scritto: > > 2014-08-20 17:34+0200, Paolo Bonzini: > >> Il 20/08/2014 17:31, Radim Krčmář ha scritto: > >>> Btw. without extra code, we are still going to overflow on races when > >>> changing PW_grow, should they be covered as well? > >> > >> You mean because there is no spinlock or similar protecting the changes? > >> I guess you could use a seqlock. > > > > Yes, for example between a modification of ple_window > > new = min(old, PW_actual_max) * PW_grow > > which gets compiled into something like this: > > 1) tmp = min(old, PW_actual_max) > > 2) new = tmp * PW_grow > > and a write to increase PW_grow > > 3) PW_actual_max = min(PW_max / new_PW_grow, PW_actual_max) > > 4) PW_grow = new_PW_grow > > 5) PW_actual_max = PW_max / new_PW_grow > > > > 3 and 4 can exectute between 1 and 2, which could overflow. > > > > I don't think they are important enough to warrant a significant > > performance hit of locking. > > A seqlock just costs two memory accesses to the same (shared) cache line > as the PW data, and a non-taken branch. Oh, seqlock readers do not have to write to shared memory, so it is acceptable ... > I don't like code that is > unsafe by design... I wouldn't say it is unsafe, because VCPU's PW is always greater than module's PW. We are just going to PLE exit sooner than expected. -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
