Re: [PATCH] arm64: turn off xgene branch prediction while in kernel space

[Mark Salter <msalter@redhat.com>]

On Wed, 2018-01-24 at 10:58 +0000, Marc Zyngier wrote:
> Khuong,
> 
> On 24/01/18 02:13, Khuong Dinh wrote:
> > Aliasing attacks against CPU branch predictors can allow an attacker to
> > redirect speculative control flow on some CPUs and potentially divulge
> > information from one context to another.
> > 
> > This patch only supports for XGene processors.
> > 
> > Signed-off-by: Mark Salter <msalter@redhat.com>
> > Signed-off-by: Khuong Dinh <kdinh@apm.com>
> > ---
> >  arch/arm64/include/asm/cpucaps.h |    3 ++-
> >  arch/arm64/include/asm/fixmap.h  |    4 ++++
> >  arch/arm64/kernel/cpu_errata.c   |   18 ++++++++++++++++++
> >  arch/arm64/kernel/entry.S        |   28 ++++++++++++++++++++++++++++
> >  arch/arm64/kernel/smp.c          |   34 ++++++++++++++++++++++++++++++++++
> >  5 files changed, 86 insertions(+), 1 deletions(-)
> > 
> > diff --git a/arch/arm64/include/asm/cpucaps.h b/arch/arm64/include/asm/cpucaps.h
> > index bb26382..dc9ada1 100644
> > --- a/arch/arm64/include/asm/cpucaps.h
> > +++ b/arch/arm64/include/asm/cpucaps.h
> > @@ -45,7 +45,8 @@
> >  #define ARM64_HARDEN_BRANCH_PREDICTOR		24
> >  #define ARM64_HARDEN_BP_POST_GUEST_EXIT		25
> >  #define ARM64_HAS_RAS_EXTN			26
> > +#define ARM64_XGENE_HARDEN_BRANCH_PREDICTOR	27
> >  
> 
> Why isn't this using the infrastructure that is already in place?

That infrastructure relies on a cpu-specific flush of the branch
predictor. XGene does not have the ability to flush the branch
predictor. It can only turn it on or off.

> 
> > -#define ARM64_NCAPS				27
> > +#define ARM64_NCAPS				28
> >  
> >  #endif /* __ASM_CPUCAPS_H */
> > diff --git a/arch/arm64/include/asm/fixmap.h b/arch/arm64/include/asm/fixmap.h
> > index ec1e6d6..d5400ca 100644
> > --- a/arch/arm64/include/asm/fixmap.h
> > +++ b/arch/arm64/include/asm/fixmap.h
> > @@ -63,6 +63,10 @@ enum fixed_addresses {
> >  	FIX_ENTRY_TRAMP_TEXT,
> >  #define TRAMP_VALIAS		(__fix_to_virt(FIX_ENTRY_TRAMP_TEXT))
> >  #endif /* CONFIG_UNMAP_KERNEL_AT_EL0 */
> > +
> > +#ifdef CONFIG_HARDEN_BRANCH_PREDICTOR
> > +	FIX_BOOT_CPU_BP_CTLREG,
> > +#endif /* CONFIG_HARDEN_BRANCH_PREDICTOR */
> >  	__end_of_permanent_fixed_addresses,
> >  
> >  	/*
> > diff --git a/arch/arm64/kernel/cpu_errata.c b/arch/arm64/kernel/cpu_errata.c
> > index ed68818..1554014 100644
> > --- a/arch/arm64/kernel/cpu_errata.c
> > +++ b/arch/arm64/kernel/cpu_errata.c
> > @@ -53,6 +53,18 @@
> >  		(arm64_ftr_reg_ctrel0.sys_val & arm64_ftr_reg_ctrel0.strict_mask);
> >  }
> >  
> > +#ifdef CONFIG_HARDEN_BRANCH_PREDICTOR
> > +static bool is_xgene_cpu(const struct arm64_cpu_capabilities *entry, int scope)
> > +{
> > +	unsigned int midr = read_cpuid_id();
> > +	unsigned int variant = MIDR_VARIANT(midr);
> > +
> > +	WARN_ON(scope != SCOPE_LOCAL_CPU);
> > +	return MIDR_IMPLEMENTOR(midr) == ARM_CPU_IMP_APM && (variant <= 3) &&
> > +					is_hyp_mode_available();
> > +}
> 
> So what happens in a guest? No BP invalidation whatsoever? I don't think
> that's acceptable.
> 
> > +#endif
> > +
> >  static int cpu_enable_trap_ctr_access(void *__unused)
> >  {
> >  	/* Clear SCTLR_EL1.UCT */
> > @@ -369,6 +381,12 @@ static int qcom_enable_link_stack_sanitization(void *data)
> >  		MIDR_ALL_VERSIONS(MIDR_CAVIUM_THUNDERX2),
> >  		.enable = enable_psci_bp_hardening,
> >  	},
> > +	{
> > +		.desc = "ARM64 XGENE branch predictors control",
> > +		.capability = ARM64_XGENE_HARDEN_BRANCH_PREDICTOR,
> > +		.def_scope = SCOPE_LOCAL_CPU,
> > +		.matches = is_xgene_cpu,
> > +	},
> >  #endif
> >  	{
> >  	}
> > diff --git a/arch/arm64/kernel/entry.S b/arch/arm64/kernel/entry.S
> > index b34e717..8c7d98e 100644
> > --- a/arch/arm64/kernel/entry.S
> > +++ b/arch/arm64/kernel/entry.S
> > @@ -62,6 +62,32 @@
> >  #endif
> >  	.endm
> >  
> > +	.macro bp_disable, tmp1, tmp2
> > +#ifdef CONFIG_HARDEN_BRANCH_PREDICTOR
> > +	alternative_if ARM64_XGENE_HARDEN_BRANCH_PREDICTOR
> > +	adr_l	x\tmp1, bp_ctlreg
> > +	mrs	x\tmp2, tpidr_el1
> > +	ldr	x\tmp1, [x\tmp1, x\tmp2]
> > +	ldr	w\tmp2, [x\tmp1]
> > +	orr	w\tmp2, w\tmp2, #(1 << 25)
> > +	str	w\tmp2, [x\tmp1]
> > +	alternative_else_nop_endif
> > +#endif
> > +	.endm
> > +
> > +	.macro bp_enable, tmp1, tmp2
> > +#ifdef CONFIG_HARDEN_BRANCH_PREDICTOR
> > +	alternative_if ARM64_XGENE_HARDEN_BRANCH_PREDICTOR
> > +	adr_l	x\tmp1, bp_ctlreg
> > +	mrs	x\tmp2, tpidr_el1
> > +	ldr	x\tmp1, [x\tmp1, x\tmp2]
> > +	ldr	w\tmp2, [x\tmp1]
> > +	and	w\tmp2, w\tmp2, #~(1 << 25)
> > +	str	w\tmp2, [x\tmp1]
> > +	alternative_else_nop_endif
> > +#endif
> > +	.endm
> > +
> >  /*
> >   * Bad Abort numbers
> >   *-----------------
> > @@ -158,6 +184,7 @@ alternative_else_nop_endif
> >  	stp	x28, x29, [sp, #16 * 14]
> >  
> >  	.if	\el == 0
> > +	bp_disable 20, 21
> >  	mrs	x21, sp_el0
> >  	ldr_this_cpu	tsk, __entry_task, x20	// Ensure MDSCR_EL1.SS is clear,
> >  	ldr	x19, [tsk, #TSK_TI_FLAGS]	// since we can unmask debug
> > @@ -307,6 +334,7 @@ alternative_else_nop_endif
> >  
> >  	msr	elr_el1, x21			// set up the return data
> >  	msr	spsr_el1, x22
> > +	bp_enable 21, 22
> >  	ldp	x0, x1, [sp, #16 * 0]
> >  	ldp	x2, x3, [sp, #16 * 1]
> >  	ldp	x4, x5, [sp, #16 * 2]
> 
> This is not what we do on other cores. Why is XGene any different?
> 
> > diff --git a/arch/arm64/kernel/smp.c b/arch/arm64/kernel/smp.c
> > index 3b8ad7b..69646be 100644
> > --- a/arch/arm64/kernel/smp.c
> > +++ b/arch/arm64/kernel/smp.c
> > @@ -85,6 +85,38 @@ enum ipi_msg_type {
> >  	IPI_WAKEUP
> >  };
> >  
> > +#ifdef CONFIG_HARDEN_BRANCH_PREDICTOR
> > +DEFINE_PER_CPU_READ_MOSTLY(void __iomem *, bp_ctlreg);
> > +
> > +static void map_bp_ctlreg(void)
> > +{
> > +	if (cpus_have_const_cap(ARM64_XGENE_HARDEN_BRANCH_PREDICTOR)) {
> > +		u64 mpidr = read_cpuid_mpidr();
> > +		unsigned int idx;
> > +		void __iomem *p;
> > +		phys_addr_t pa;
> > +
> > +		idx =  (MPIDR_AFFINITY_LEVEL(mpidr, 1) << 1) +
> > +			MPIDR_AFFINITY_LEVEL(mpidr, 0);
> > +		pa = 0x7c0c0000ULL | (0x100000ULL * idx);
> > +		if (smp_processor_id())
> > +			p = ioremap(pa, PAGE_SIZE);
> > +		else {
> > +			/* boot processor uses fixmap */
> > +			set_fixmap_io(FIX_BOOT_CPU_BP_CTLREG, pa);
> > +			p = (void __iomem *)__fix_to_virt(
> > +						FIX_BOOT_CPU_BP_CTLREG);
> > +		}
> > +		__this_cpu_write(bp_ctlreg, p);
> > +
> > +		pr_debug("%s: cpu%d idx=%d pa=0x%llx %p", __func__,
> > +			 smp_processor_id(), idx, pa, p);
> > +	}
> > +}
> > +#else
> > +static inline void map_bp_ctlreg(void) {}
> > +#endif
> > +
> >  #ifdef CONFIG_ARM64_VHE
> >  
> >  /* Whether the boot CPU is running in HYP mode or not*/
> > @@ -224,6 +256,7 @@ asmlinkage void secondary_start_kernel(void)
> >  
> >  	cpu = task_cpu(current);
> >  	set_my_cpu_offset(per_cpu_offset(cpu));
> > +	map_bp_ctlreg();
> >  
> >  	/*
> >  	 * All kernel threads share the same mm context; grab a
> > @@ -454,6 +487,7 @@ void __init smp_prepare_boot_cpu(void)
> >  	 * cpuinfo_store_boot_cpu() above.
> >  	 */
> >  	update_cpu_errata_workarounds();
> > +	map_bp_ctlreg();
> >  }
> >  
> >  static u64 __init of_get_cpu_mpidr(struct device_node *dn)
> > 
> 
> Thanks,
> 
> 	M.