On Thu, Feb 13, 2020 at 10:53:43AM +0100, Jan Beulich wrote: > On 12.02.2020 22:05, Julien Grall wrote: > > On 12/02/2020 17:49, Roger Pau Monne wrote: > >> Commit: > >> > >> 5500d265a2a8fa63d60c08beb549de8ec82ff7a5 > >> x86/smp: use APIC ALLBUT destination shorthand when possible > > > > There is a more subtle problem introduced by this patch. I thought I > > would mention it here just in case this affect the approach you have > > chosen in this series. > > > > get_cpu_maps() is used by stop_machine_run() to serialize the callers. > > If the latter fails to acquire the lock, it will bail out. > > Unfortunately, rcu_barrier() is implemented using stop_machine_run() and > > will be turned to pretty much a NOP if the latter fails (e.g the lock > > cannot be acquired). > > > > This means that the rcu_barrier() will not do the expected job and > > potentially introduce unknown issues (e.g use-after-free...). > > > > Before your patch, it would have been pretty hard to hit the problem > > above. After, you can race more easily with rcu_barrier() as sending IPI > > is pretty common. > > > > Sadly, I don't have a suggestion yet how to fix this problem. > > A frequent use like on the IPI sending path suggests the lock may > want to become an r/w one, where both parties you mention want to > acquire it in read mode.
I'm already preparing a patch in this direction. Roger. _______________________________________________ Xen-devel mailing list Xen-devel@lists.xenproject.org https://lists.xenproject.org/mailman/listinfo/xen-devel
