Re: [PATCH v6] arm64: Add support for new control bits CTR_EL0.DIC and CTR_EL0.IDC

[Will Deacon]

Hi Shanker,

On Wed, Feb 28, 2018 at 10:14:00PM -0600, Shanker Donthineni wrote:
> The DCache clean & ICache invalidation requirements for instructions
> to be data coherence are discoverable through new fields in CTR_EL0.
> The following two control bits DIC and IDC were defined for this
> purpose. No need to perform point of unification cache maintenance
> operations from software on systems where CPU caches are transparent.
> 
> This patch optimize the three functions __flush_cache_user_range(),
> clean_dcache_area_pou() and invalidate_icache_range() if the hardware
> reports CTR_EL0.IDC and/or CTR_EL0.IDC. Basically it skips the two
> instructions 'DC CVAU' and 'IC IVAU', and the associated loop logic
> in order to avoid the unnecessary overhead.
> 
> CTR_EL0.DIC: Instruction cache invalidation requirements for
>  instruction to data coherence. The meaning of this bit[29].
>   0: Instruction cache invalidation to the point of unification
>      is required for instruction to data coherence.
>   1: Instruction cache cleaning to the point of unification is
>       not required for instruction to data coherence.
> 
> CTR_EL0.IDC: Data cache clean requirements for instruction to data
>  coherence. The meaning of this bit[28].
>   0: Data cache clean to the point of unification is required for
>      instruction to data coherence, unless CLIDR_EL1.LoC == 0b000
>      or (CLIDR_EL1.LoUIS == 0b000 && CLIDR_EL1.LoUU == 0b000).
>   1: Data cache clean to the point of unification is not required
>      for instruction to data coherence.
> 
> Co-authored-by: Philip Elcan <pelcan@xxxxxxxxxxxxxx>
> Signed-off-by: Shanker Donthineni <shankerd@xxxxxxxxxxxxxx>
> ---
> Changes since v5:
>   -Addressed Mark's review comments.

This mostly looks good now. Just a few comments inline.

> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> index 7381eeb..41af850 100644
> --- a/arch/arm64/Kconfig
> +++ b/arch/arm64/Kconfig
> @@ -1091,6 +1091,18 @@ config ARM64_RAS_EXTN
>  	  and access the new registers if the system supports the extension.
>  	  Platform RAS features may additionally depend on firmware support.
>  
> +config ARM64_SKIP_CACHE_POU
> +	bool "Enable support to skip cache PoU operations"
> +	default y
> +	help
> +	  Explicit point of unification cache operations can be eliminated
> +	  in software if the hardware handles transparently. The new bits in
> +	  CTR_EL0, CTR_EL0.DIC and CTR_EL0.IDC indicates the hardware
> +	  capabilities of ICache and DCache PoU requirements.
> +
> +	  Selecting this feature will allow the kernel to optimize cache
> +	  maintenance to the PoU.
> +
>  endmenu

Let's not bother with a Kconfig option. I think the extra couple of NOPs
this introduces for CPUs that don't implement the new features isn't going
to hurt anybody.

> diff --git a/arch/arm64/include/asm/assembler.h b/arch/arm64/include/asm/assembler.h
> index 3c78835..39f2274 100644
> --- a/arch/arm64/include/asm/assembler.h
> +++ b/arch/arm64/include/asm/assembler.h
> @@ -444,6 +444,11 @@
>   * 	Corrupts:	tmp1, tmp2
>   */
>  	.macro invalidate_icache_by_line start, end, tmp1, tmp2, label
> +#ifdef CONFIG_ARM64_SKIP_CACHE_POU
> +alternative_if ARM64_HAS_CACHE_DIC
> +	b	9996f
> +alternative_else_nop_endif
> +#endif
>  	icache_line_size \tmp1, \tmp2
>  	sub	\tmp2, \tmp1, #1
>  	bic	\tmp2, \start, \tmp2
> @@ -453,6 +458,7 @@
>  	cmp	\tmp2, \end
>  	b.lo	9997b
>  	dsb	ish
> +9996:
>  	isb
>  	.endm
>  
> diff --git a/arch/arm64/include/asm/cache.h b/arch/arm64/include/asm/cache.h
> index ea9bb4e..d460e9f 100644
> --- a/arch/arm64/include/asm/cache.h
> +++ b/arch/arm64/include/asm/cache.h
> @@ -20,8 +20,12 @@
>  
>  #define CTR_L1IP_SHIFT		14
>  #define CTR_L1IP_MASK		3
> +#define CTR_DMLINE_SHIFT	16

This should be "CTR_DMINLINE_SHIFT"

> +#define CTR_ERG_SHIFT		20
>  #define CTR_CWG_SHIFT		24
>  #define CTR_CWG_MASK		15
> +#define CTR_IDC_SHIFT		28
> +#define CTR_DIC_SHIFT		29
>  
>  #define CTR_L1IP(ctr)		(((ctr) >> CTR_L1IP_SHIFT) & CTR_L1IP_MASK)
>  
> diff --git a/arch/arm64/include/asm/cpucaps.h b/arch/arm64/include/asm/cpucaps.h
> index bb26382..8dd42ae 100644
> --- a/arch/arm64/include/asm/cpucaps.h
> +++ b/arch/arm64/include/asm/cpucaps.h
> @@ -45,7 +45,9 @@
>  #define ARM64_HARDEN_BRANCH_PREDICTOR		24
>  #define ARM64_HARDEN_BP_POST_GUEST_EXIT		25
>  #define ARM64_HAS_RAS_EXTN			26
> +#define ARM64_HAS_CACHE_IDC			27
> +#define ARM64_HAS_CACHE_DIC			28
>  
> -#define ARM64_NCAPS				27
> +#define ARM64_NCAPS				29
>  
>  #endif /* __ASM_CPUCAPS_H */
> diff --git a/arch/arm64/kernel/cpufeature.c b/arch/arm64/kernel/cpufeature.c
> index 2985a06..0b64b55 100644
> --- a/arch/arm64/kernel/cpufeature.c
> +++ b/arch/arm64/kernel/cpufeature.c
> @@ -199,12 +199,12 @@ static int __init register_cpu_hwcaps_dumper(void)
>  };
>  
>  static const struct arm64_ftr_bits ftr_ctr[] = {
> -	ARM64_FTR_BITS(FTR_VISIBLE, FTR_STRICT, FTR_EXACT, 31, 1, 1),		/* RES1 */
> -	ARM64_FTR_BITS(FTR_VISIBLE, FTR_STRICT, FTR_LOWER_SAFE, 29, 1, 1),	/* DIC */
> -	ARM64_FTR_BITS(FTR_VISIBLE, FTR_STRICT, FTR_LOWER_SAFE, 28, 1, 1),	/* IDC */
> -	ARM64_FTR_BITS(FTR_VISIBLE, FTR_STRICT, FTR_HIGHER_SAFE, 24, 4, 0),	/* CWG */
> -	ARM64_FTR_BITS(FTR_VISIBLE, FTR_STRICT, FTR_HIGHER_SAFE, 20, 4, 0),	/* ERG */
> -	ARM64_FTR_BITS(FTR_VISIBLE, FTR_STRICT, FTR_LOWER_SAFE, 16, 4, 1),	/* DminLine */
> +	ARM64_FTR_BITS(FTR_VISIBLE, FTR_STRICT, FTR_EXACT, 31, 1, 1),                    /* RES1 */
> +	ARM64_FTR_BITS(FTR_VISIBLE, FTR_STRICT, FTR_LOWER_SAFE, CTR_DIC_SHIFT, 1, 1),    /* DIC */
> +	ARM64_FTR_BITS(FTR_VISIBLE, FTR_STRICT, FTR_LOWER_SAFE, CTR_IDC_SHIFT, 1, 1),    /* IDC */
> +	ARM64_FTR_BITS(FTR_VISIBLE, FTR_STRICT, FTR_HIGHER_SAFE, CTR_CWG_SHIFT, 4, 0),   /* CWG */
> +	ARM64_FTR_BITS(FTR_VISIBLE, FTR_STRICT, FTR_HIGHER_SAFE, CTR_ERG_SHIFT, 4, 0),   /* ERG */
> +	ARM64_FTR_BITS(FTR_VISIBLE, FTR_STRICT, FTR_LOWER_SAFE, CTR_DMLINE_SHIFT, 4, 1), /* DminLine */
>  	/*
>  	 * Linux can handle differing I-cache policies. Userspace JITs will
>  	 * make use of *minLine.
> @@ -852,6 +852,20 @@ static bool has_no_fpsimd(const struct arm64_cpu_capabilities *entry, int __unus
>  					ID_AA64PFR0_FP_SHIFT) < 0;
>  }
>  
> +#ifdef CONFIG_ARM64_SKIP_CACHE_POU
> +static bool has_cache_idc(const struct arm64_cpu_capabilities *entry,
> +			  int __unused)
> +{
> +	return read_sanitised_ftr_reg(SYS_CTR_EL0) & BIT(CTR_IDC_SHIFT);
> +}
> +
> +static bool has_cache_dic(const struct arm64_cpu_capabilities *entry,
> +			  int __unused)
> +{
> +	return read_sanitised_ftr_reg(SYS_CTR_EL0) & BIT(CTR_DIC_SHIFT);
> +}
> +#endif
> +
>  #ifdef CONFIG_UNMAP_KERNEL_AT_EL0
>  static int __kpti_forced; /* 0: not forced, >0: forced on, <0: forced off */
>  
> @@ -1088,6 +1102,20 @@ static int cpu_copy_el2regs(void *__unused)
>  		.enable = cpu_clear_disr,
>  	},
>  #endif /* CONFIG_ARM64_RAS_EXTN */
> +#ifdef CONFIG_ARM64_SKIP_CACHE_POU
> +	{
> +		.desc = "Skip D-Cache maintenance 'DC CVAU' (CTR_EL0.IDC=1)",

Can we stick a bit closer to the architectural text here, please? How about:

"Data cache clean to the PoU not required for I/D coherence"

> +		.capability = ARM64_HAS_CACHE_IDC,
> +		.def_scope = SCOPE_SYSTEM,
> +		.matches = has_cache_idc,
> +	},
> +	{
> +		.desc = "Skip I-Cache maintenance 'IC IVAU' (CTR_EL0.DIC=1)",

"Instruction cache invalidation not required for I/D coherence"

> +		.capability = ARM64_HAS_CACHE_DIC,
> +		.def_scope = SCOPE_SYSTEM,
> +		.matches = has_cache_dic,
> +	},
> +#endif /* CONFIG_ARM64_SKIP_CACHE_POU */
>  	{},
>  };
>  
> diff --git a/arch/arm64/mm/cache.S b/arch/arm64/mm/cache.S
> index 758bde7..d8d7a32 100644
> --- a/arch/arm64/mm/cache.S
> +++ b/arch/arm64/mm/cache.S
> @@ -50,6 +50,12 @@ ENTRY(flush_icache_range)
>   */
>  ENTRY(__flush_cache_user_range)
>  	uaccess_ttbr0_enable x2, x3, x4
> +#ifdef CONFIG_ARM64_SKIP_CACHE_POU
> +alternative_if ARM64_HAS_CACHE_IDC
> +	dsb	ishst
> +	b	8f
> +alternative_else_nop_endif
> +#endif
>  	dcache_line_size x2, x3
>  	sub	x3, x2, #1
>  	bic	x4, x0, x3
> @@ -60,6 +66,7 @@ user_alt 9f, "dc cvau, x4",  "dc civac, x4",  ARM64_WORKAROUND_CLEAN_CACHE
>  	b.lo	1b
>  	dsb	ish
>  
> +8:
>  	invalidate_icache_by_line x0, x1, x2, x3, 9f
>  	mov	x0, #0
>  1:
> @@ -116,6 +123,12 @@ ENDPIPROC(__flush_dcache_area)
>   *	- size    - size in question
>   */
>  ENTRY(__clean_dcache_area_pou)
> +#ifdef CONFIG_ARM64_SKIP_CACHE_POU
> +alternative_if ARM64_HAS_CACHE_IDC
> +	dsb	ishst
> +	ret
> +alternative_else_nop_endif

I think this is a slight asymmetry with the code for the I-side. On the
I-side, you hook into invalidate_icache_by_line, whereas on the D-side you
hook into the callers of dcache_by_line_op. Why is that?

I notice that the only user other than
flush_icache_range/__flush_cache_user_range or invalidate_icache_by_line
is in KVM, via invalidate_icache_range. If you want to hook in there, why
aren't you also patching __flush_icache_all? If so, I'd rather have the
I-side code consistent with the D-side code and do this in the handful of
callers. We might even be able to elide a branch or two that way.

I'm going to assume that I-cache aliases are all coherent if DIC=1, so it's
safe to elide our alias sync code.

Will