Re: [Patch v4 17/18] x86/speculation: Update SPEC_CTRL MSRs of remote CPUs

[Thomas Gleixner]

Tim,

On Tue, 30 Oct 2018, Tim Chen wrote:
>  void arch_set_security(struct task_struct *tsk, unsigned int value)
> diff --git a/arch/x86/kernel/process.c b/arch/x86/kernel/process.c
> index 943e90d..048b7f4b 100644
> --- a/arch/x86/kernel/process.c
> +++ b/arch/x86/kernel/process.c
> @@ -426,7 +426,19 @@ static __always_inline void spec_ctrl_update_msr(unsigned long tifn)
>  static __always_inline void __speculation_ctrl_update(unsigned long tifp,
>  						      unsigned long tifn)
>  {
> -	bool updmsr = !!((tifp ^ tifn) & _TIF_STIBP);
> +	/*
> +	 * If TIF_UPDATE_SPEC_CTRL bit is set in tifp, speculation related
> +	 * TIF flags have changed when previous task was running, but
> +	 * SPEC_CTRL MSR has not been synchronized with TIF flag changes.
> +	 * SPEC_CTRL MSR value can be out of date.
> +	 *
> +	 * Need to force update SPEC_CTRL MSR if TIF_UPDATE_SPEC_CTRL
> +	 * bit in tifp is set.
> +	 *
> +	 * The TIF_UPDATE_SPEC_CTRL bit in tifn was cleared before calling
> +	 * this function.
> +	 */
> +	bool updmsr = !!((tifp ^ tifn) & (_TIF_STIBP|_TIF_UPDATE_SPEC_CTRL));
>  
>  	/* If TIF_SSBD is different, select the proper mitigation method */
>  	if ((tifp ^ tifn) & _TIF_SSBD) {
> @@ -482,6 +494,14 @@ void __switch_to_xtra(struct task_struct *prev_p, struct task_struct *next_p,
>  	if ((tifp ^ tifn) & _TIF_NOCPUID)
>  		set_cpuid_faulting(!!(tifn & _TIF_NOCPUID));
>  
> +	if (tifp & _TIF_UPDATE_SPEC_CTRL)
> +		clear_tsk_thread_flag(prev_p, TIF_UPDATE_SPEC_CTRL);
> +
> +	if (tifn & _TIF_UPDATE_SPEC_CTRL) {
> +		clear_tsk_thread_flag(next_p, TIF_UPDATE_SPEC_CTRL);
> +		tifn &= ~_TIF_UPDATE_SPEC_CTRL;
> +	}

I'm really unhappy about adding yet more conditionals into this code
path. We really need to find some better solution for that.

There are basically two options:

 1) Restrict the PRCTL control so it is only possible to modify it at the
    point where the application is still single threaded.

 2) Add _TIF_UPDATE_SPEC_CTRL to the SYSCALL_EXIT_WORK_FLAGS and handle it
    in the slow work path.

    The KVM side can be handled in x86_virt_spec_ctrl().

Thanks,

	tglx