Re: [PATCH 28/35] x86/idle: Control Indirect Branch Speculation in idle

[Andrew Cooper <andrew.cooper3@citrix.com>]

[-- Attachment #1: Type: text/plain, Size: 2667 bytes --]

On 18/01/18 13:48, Peter Zijlstra wrote:
> From: Thomas Gleixner <tglx@linutronix.de>
>
> Indirect Branch Speculation (IBS) is controlled per physical core. If one
> thread disables it then it's disabled for the core. If a thread enters idle
> it makes sense to reenable IBS so the sibling thread can run with full
> speculation enabled in user space.
>
> This makes only sense in mwait_idle_with_hints() because mwait_idle() can
> serve an interrupt immediately before speculation can be stopped again. SKL
> which requires IBRS should use mwait_idle_with_hints() so this is a non
> issue and in the worst case a missed optimization.
>
> [peterz: fixed stop_indirect_branch_speculation placement]
>
> Originally-by: Tim Chen <tim.c.chen@linux.intel.com>
> Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
> ---
>  arch/x86/include/asm/mwait.h |   14 ++++++++++++++
>  arch/x86/kernel/process.c    |   14 ++++++++++++++
>  2 files changed, 28 insertions(+)
>
> --- a/arch/x86/include/asm/mwait.h
> +++ b/arch/x86/include/asm/mwait.h
> @@ -6,6 +6,7 @@
>  #include <linux/sched/idle.h>
>  
>  #include <asm/cpufeature.h>
> +#include <asm/nospec-branch.h>
>  
>  #define MWAIT_SUBSTATE_MASK		0xf
>  #define MWAIT_CSTATE_MASK		0xf
> @@ -106,9 +107,22 @@ static inline void mwait_idle_with_hints
>  			mb();
>  		}
>  
> +		/*
> +		 * Indirect Branch Speculation (IBS) is controlled per
> +		 * physical core. If one thread disables it, then it's
> +		 * disabled on all threads of the core. The kernel disables
> +		 * it on entry from user space. Reenable it on the thread
> +		 * which goes idle so the other thread has a chance to run
> +		 * with full speculation enabled in userspace.
> +		 */
> +		restart_indirect_branch_speculation();
>  		__monitor((void *)&current_thread_info()->flags, 0, 0);
>  		if (!need_resched())
>  			__mwait(eax, ecx);
> +		/*
> +		 * Stop IBS again to protect kernel execution.
> +		 */
> +		stop_indirect_branch_speculation();

Be very careful.  The safety of this on Skylake+ depends on there not
being a ret instruction which can be speculatively reached between the
two WRMSRs used to play with MSR_SPEC_CTRL.

You need to guarantee that the net call tree of these five functions is
forced always inline.  A plain "static inline" function is not good
enough (and there are several here), if the compiler chooses to
out-of-line it, and a real function call is definitely a problem.

I accidentally introduced a vulnerability into Xen in my first attempt
at this.  (I'm still trying to decide whether reimplementing it in asm
would be better long term).

~Andrew

[-- Attachment #2: Type: text/html, Size: 3599 bytes --]