On 18/04/2024 11:35 am, Sean Christopherson wrote:
On Thu, Apr 18, 2024, Kai Huang wrote:
On 18/04/2024 2:40 am, Sean Christopherson wrote:
This way, architectures that aren't saddled with out-of-tree hypervisors can do
the dead simple thing of enabling hardware during their initialization sequence,
and the TDX code is much more sane, e.g. invoke kvm_x86_enable_virtualization()
during late_hardware_setup(), and kvm_x86_disable_virtualization() during module
exit (presumably).
Fine to me, given I am not familiar with other ARCHs, assuming always enable
virtualization when KVM present is fine to them. :-)
Two questions below:
+int kvm_x86_enable_virtualization(void)
+{
+ int r;
+
+ guard(mutex)(&vendor_module_lock);
It's a little bit odd to take the vendor_module_lock mutex.
It is called by kvm_arch_init_vm(), so more reasonablly we should still use
kvm_lock?
I think this should take an x86-specific lock, since it's guarding x86-specific
data.
OK. This makes sense.
And vendor_module_lock fits the bill perfectly. Well, except for the
name, and I definitely have no objection to renaming it.
No opinion on renaming. Personally I wouldn't bother to rename. We can
add a comment in kvm_x86_enable_virtualization() to explain. Perhaps in
the future we just want to change to always enable virtualization for
x86 too..
Also, if we invoke kvm_x86_enable_virtualization() from
kvm_x86_ops->late_hardware_setup(), then IIUC we will deadlock here because
kvm_x86_vendor_init() already takes the vendor_module_lock?
Ah, yeah. Oh, duh. I think the reason I didn't initially suggest late_hardware_setup()
is that I was assuming/hoping TDX setup could be done after kvm_x86_vendor_exit().
E.g. in vt_init() or whatever it gets called:
r = kvm_x86_vendor_exit(...);
if (r)
return r;
if (enable_tdx) {
r = tdx_blah_blah_blah();
if (r)
goto vendor_exit;
}
I assume the reason you introduced the late_hardware_setup() is purely
because you want to do:
cpu_emergency_register_virt_callback(kvm_x86_ops.emergency_enable);
after
kvm_ops_update()?
Anyway, we can also do 'enable_tdx' outside of kvm_x86_vendor_init() as
above, given it cannot be done in hardware_setup() anyway.
If we do 'enable_tdx' in late_hardware_setup(), we will need a
kvm_x86_enable_virtualization_nolock(), but that's also not a problem to me.
So which way do you prefer?
Btw, with kvm_x86_virtualization_enable(), it seems the compatibility
check is lost, which I assume is OK?
Btw2, currently tdx_enable() requires cpus_read_lock() must be called
prior. If we do unconditional tdx_cpu_enable() in vt_hardware_enable(),
then with your proposal IIUC there's no such requirement anymore,
because no task will be scheduled to the new CPU before it reaches
CPUHP_AP_ACTIVE. But now calling cpus_read_lock()/unlock() around
tdx_enable() also acceptable to me.
[...]
+int kvm_enable_virtualization(void)
{
+ int r;
+
+ r = cpuhp_setup_state(CPUHP_AP_KVM_ONLINE, "kvm/cpu:online",
+ kvm_online_cpu, kvm_offline_cpu);
+ if (r)
+ return r;
+
+ register_syscore_ops(&kvm_syscore_ops);
+
+ /*
+ * Manually undo virtualization enabling if the system is going down.
+ * If userspace initiated a forced reboot, e.g. reboot -f, then it's
+ * possible for an in-flight module load to enable virtualization
+ * after syscore_shutdown() is called, i.e. without kvm_shutdown()
+ * being invoked. Note, this relies on system_state being set _before_
+ * kvm_shutdown(), e.g. to ensure either kvm_shutdown() is invoked
+ * or this CPU observes the impedning shutdown. Which is why KVM uses
+ * a syscore ops hook instead of registering a dedicated reboot
+ * notifier (the latter runs before system_state is updated).
+ */
+ if (system_state == SYSTEM_HALT || system_state == SYSTEM_POWER_OFF ||
+ system_state == SYSTEM_RESTART) {
+ unregister_syscore_ops(&kvm_syscore_ops);
+ cpuhp_remove_state(CPUHP_AP_KVM_ONLINE);
+ return -EBUSY;
+ }
+
Aren't we also supposed to do:
on_each_cpu(__kvm_enable_virtualization, NULL, 1);
here?
No, cpuhp_setup_state() invokes the callback, kvm_online_cpu(), on each CPU.
I.e. KVM has been doing things the hard way by using cpuhp_setup_state_nocalls().
That's part of the complexity I would like to get rid of.
Ah, right :-)
Btw, why couldn't we do the 'system_state' check at the very beginning
of this function?
