Paolo Bonzini <pbonz...@redhat.com> writes: > On 14/06/2017 14:14, Alex Bennée wrote: >>> Then Emilio's patch, if a bit of a heavy hammer, is correct. After >>> aa64_daif_write needs you need an exit_tb so that arm_cpu_exec_interrupt >>> is executed again. >> >> This is a case of cpu->interrupt_request being pending but not having >> set cpu->icount_decr yet to signal the exit. > > Rather than "yet", "anymore". So far it has always been an invariant > that anything that re-enabled an interrupt had to do exit_tb. > >> Wouldn't another approach >> (that didn't involve futzing with each front-end) to be to check >> cpu->interrupt_request and force the exit in lookup_tb_ptr? > > That would cause an unnecessary slowdown in code that runs with > interrupts disabled but does a lot of indirect jumps... ppc's SLOF > firmware probably qualifies.
Really? I'd have to measure the change it makes. Is there a benchmark stanza for measuring the PPC slof firmware time? I have 3 patches now which all fix the same thing so we can pick and choose which we should apply. Patches incoming... -- Alex Bennée
