Re: [PATCH] arm64: archrandom: avoid trapping ID register read in __cpu_has_rng()

[Will Deacon <will@kernel.org>]

Hi Aman,

On Mon, Jul 20, 2026 at 03:06:15PM +0100, Aman Priyadarshi wrote:
> __cpu_has_rng() has an early-boot fallback, taken before the
> ARM64_HAS_RNG alternative is patched, that calls
> this_cpu_has_cap(ARM64_HAS_RNG). With SCOPE_LOCAL_CPU that resolves the
> capability by reading ID_AA64ISAR0_EL1 directly from hardware via
> __read_sysreg_by_encoding(), on every invocation.
> 
> Until the CRNG is seeded, crng_make_state() routes every get_random_*()
> through extract_entropy(), which drains architectural entropy via
> arch_get_random_seed_longs()/arch_get_random_longs() and so calls
> __cpu_has_rng() several times per request. On a direct (non-EFI) boot
> there is no bootloader seed, so the CRNG stays unseeded for much of
> boot and essentially every early randomness consumer takes this path.
> 
> Under virtualization this is costly: the hypervisor traps guest
> accesses to the ID registers (HCR_EL2.TID3), making each read a vmexit,
> producing ~200k trapped ID_AA64ISAR0_EL1 reads during boot.
> 
> The register value is invariant, so read the sanitised feature register
> instead. read_sanitised_ftr_reg() returns the cached value from
> arm64_ftr_regs[] with no sysreg access, and hence no trap. That array
> is populated by cpuinfo_store_boot_cpu() in smp_prepare_boot_cpu(),
> before the first early RNG use in random_init_early(), so it is always
> valid here.
> 
> With this change the trapped reads drop from ~200k to handful number of
> times, and the boot time drops roughly by 6.3% in the test environment.

Yikes, that's quite a compelling performance improvement.

> diff --git a/arch/arm64/include/asm/archrandom.h b/arch/arm64/include/asm/archrandom.h
> index 8babfbe31f95..8067e9a35641 100644
> --- a/arch/arm64/include/asm/archrandom.h
> +++ b/arch/arm64/include/asm/archrandom.h
> @@ -61,8 +61,22 @@ static inline bool __arm64_rndrrs(unsigned long *v)
>  
>  static __always_inline bool __cpu_has_rng(void)
>  {
> -	if (unlikely(!system_capabilities_finalized() && !preemptible()))
> -		return this_cpu_has_cap(ARM64_HAS_RNG);
> +	if (unlikely(!system_capabilities_finalized() && !preemptible())) {
> +		/*
> +		 * Until the ARM64_HAS_RNG alternative is patched we can't use
> +		 * the static-branch form, so consult the feature register
> +		 * directly. Don't use this_cpu_has_cap() here: it reads
> +		 * ID_AA64ISAR0_EL1 from hardware on every call, under
> +		 * virtualization each ID register read traps to the hypervisor
> +		 * (HCR_EL2.TID3) -- producing a storm of vmexits during boot.
> +		 * The sanitised value is cached in memory.
> +		 */
> +		u64 isar0 = read_sanitised_ftr_reg(SYS_ID_AA64ISAR0_EL1);
> +
> +		return cpuid_feature_extract_unsigned_field(isar0,
> +				ID_AA64ISAR0_EL1_RNDR_SHIFT) >=
> +			ID_AA64ISAR0_EL1_RNDR_IMP;
> +	}

You can probably rewrite this a little more cleanly along the lines of
the (not even compile-tested) diff below. I was about to do that, but
then I got a bit confused by the whole thing. The preemptible() check is
presumably not needed if we're accessing the in-memory feature registers
rather than the per-CPU id registers, but then how do you handle races
with concurrent updates to the "safe value" made by CPUs concurrently
coming online?

Will

--->8

diff --git a/arch/arm64/include/asm/archrandom.h b/arch/arm64/include/asm/archrandom.h
index 8babfbe31f95..1c6ccc1776cd 100644
--- a/arch/arm64/include/asm/archrandom.h
+++ b/arch/arm64/include/asm/archrandom.h
@@ -61,8 +61,17 @@ static inline bool __arm64_rndrrs(unsigned long *v)

 static __always_inline bool __cpu_has_rng(void)
 {
-       if (unlikely(!system_capabilities_finalized() && !preemptible()))
-               return this_cpu_has_cap(ARM64_HAS_RNG);
+       if (unlikely(!system_capabilities_finalized() && !preemptible())) {
+               /*
+                * Query the in-memory sanitised value to avoid a potential
+                * trap when accessing the ID register under a hypervisor.
+                */
+               u64 isar0 = read_sanitised_ftr_reg(SYS_ID_AA64ISAR0_EL1);
+               u64 rndr = SYS_FIELD_GET(ID_AA64ISAR0_EL1, RNDR, isar0);
+
+               return rndr >= ID_AA64ISAR0_EL1_RNDR_IMP;
+       }
+
        return alternative_has_cap_unlikely(ARM64_HAS_RNG);
 }