Re: [PATCH] Allow preemption during lazy mmu updates

[Avi Kivity <avi@redhat.com>]

Jeremy Fitzhardinge wrote:
> Avi Kivity wrote:
>> What are the slight differences in requirements?
>>
>> KVM wants to run in non-preemptible, interrupts-enabled context.
>
> There are two hooks: arch_start_context_switch() in 
> kernel/sched:context_switch(), and arch_end_context_switch() in 
> arch/x86/kernel/process_(32|64).c.  They bound the heart of the 
> context switch in which various bits of core state is changes, like 
> the cr3 reload, fpu TS flag, iopl, tls slots, etc.  All things which 
> require a hypercall in a paravirtualized environment, and so can be 
> batched together into a multicall (or whatever) to minimize the number 
> of context-switch time hypercalls.  The placement of 
> end_context_switch is particularly sensitive because it needs to be in 
> the right place relative to fpu context reload and segment register 
> reloading.
>
> Preemption is definitely disabled, and interrupts as well, I think.  
> So perhaps these won't work for you.
>
> However, looking at the fire_sched_out_preempt_notifiers() is almost 
> in the same position as arch_start_context_switch(), so I think they 
> could be unified one way or the other.  The sched_in notifier happens 
> way too late though.  Does KVM just use both in and out, or just one?

Both.  sched_out loads the host MSR_STAR and friends, sched_in loads the 
guest values.


-- 
Do not meddle in the internals of kernels, for they are subtle and quick to panic.