On Mon, Mar 23, 2020 at 02:43:18PM -0700, Sean Christopherson wrote: > On Mon, Mar 23, 2020 at 03:06:36PM -0400, Peter Xu wrote: > > On Fri, Mar 20, 2020 at 01:55:46PM -0700, Sean Christopherson wrote: > > > + /* > > > + * Spin until the memory region is moved to a misaligned address. This > > > + * may or may not trigger MMIO, as the window where the memslot is > > > + * invalid is quite small. > > > + */ > > > + val = guest_spin_on_val(0); > > > + GUEST_ASSERT(val == 1 || val == MMIO_VAL); > > > + > > > + /* Spin until the memory region is realigned. */ > > > + GUEST_ASSERT(guest_spin_on_val(MMIO_VAL) == 1); > > > > IIUC ideally we should do GUEST_SYNC() after each GUEST_ASSERT() to > > make sure the two threads are in sync. Otherwise e.g. there's no > > guarantee that the main thread won't run too fast to quickly remove > > the memslot and re-add it back before the guest_spin_on_val() starts > > above, then the assert could trigger when it reads the value as zero. > > Hrm, I was thinking ucall wasn't available across pthreads, but it's just > dumped into a global variable. I'll rework this to replace the udelay() > hacks with proper synchronization. I think ucall should work for pthread (shared address space of either kvm_run or guest memories), however my thought was even simpler than that, something like: - in guest code: do GUEST_SYNC after each GUEST_ASSERT - introduce a global_sem - in vcpu thread: when receive GUEST_SYNC, do "sem_post(&global_sem)" - in main thread: replace all usleep() with "sem_wait(&global_sem)" -- Peter Xu
