On Fri, Apr 06, 2018 at 07:54:02PM +0300, Yury Norov wrote:
> In general, kick_all_cpus_sync() is needed to switch contexts. But exit from
> userspace is anyway the switch of context. And while in userspace, we cannot
> do something wrong on kernel side. For me it means that we can safely drop
> IPI for all userspace modes - both normal and nohz_full.
This *may* be true, but only if we never have to patch text in the
* between exception entry and eqs_exit()
* between eqs_enter() and exception return
* between eqs_enter() and eqs_exit() in the idle loop.
If it's possible that we need to execute patched text in any of those
paths, we must IPI all CPUs in order to correctly serialize things.
Digging a bit, I also thing that our ct_user_exit and ct_user_enter
usage is on dodgy ground today.
For example, in el0_dbg we call do_debug_exception() *before* calling
ct_user_exit. Which I believe means we'd use RCU while supposedly in an
extended quiescent period, which would be bad.
In other paths, we unmask all DAIF bits before calling ct_user_exit, so
we could similarly take an EL1 debug exception without having exited the
extended quiescent period.
I think similar applies to SDEI; we don't negotiate with RCU prior to
invoking handlers, which might need RCU.
> If it's correct, for v3 I would suggest:
> - in kick_all_cpus_sync() mask all is_idle_task() and user_mode() CPUs;
> - add isb() for arm64 in do_idle() path only - this path doesn't imply
> context switch.
As mentioned in my other reply, I don't think the ISB in do_idle()
makes sense, unless that occurs *after* we exit the extended quiescent