Skip to contentSkip to conversation

random: Drop ARCH limitations for CONFIG_RANDOM_TRUST_CPU

PATCH In review Alexander Sverdlin 3 participants 4 messages Mar 29, 2020 – Mar 30, 2020 Char and misc drivers

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 random: Drop ARCH limitations for CONFIG_RANDOM_TRUST_CPU

The option itself looks attractive for the embedded devices which often have HWRNG but less entropy from user-input. And these devices are often ARM/ARM64 or MIPS. The reason to limit it to X86/S390/PPC is not obvious.

---
 drivers/char/Kconfig | 1 -
 1 file changed, 1 deletion(-)
drivers/char/Kconfig +0 -1
index 26956c006987..abc874722251 100644
@@ -539,7 +539,6 @@ endmenu
 
 config RANDOM_TRUST_CPU
 	bool "Trust the CPU manufacturer to initialize Linux's CRNG"
-	depends on X86 || S390 || PPC
 	default n
 	help
 	Assume that CPU manufacturer (e.g., Intel or AMD for RDSEED or
Signature · 1 line
2.25.1
#1 lore raw
Quoted 5 lines · Alexander Sverdlin

The option itself looks attractive for the embedded devices which often have HWRNG but less entropy from user-input. And these devices are often ARM/ARM64 or MIPS. The reason to limit it to X86/S390/PPC is not obvious.

This feature is *only* applicable if the CPU supports a arch_get_random_seed_long() or arch_get_random_long(). I believe there are some server-class ARM64 CPU's that support such an instruction, but I don't believe any of the embedded arm64 --- and certainly non of the embedded arm --- SOC's support arch_get_random_long().

The reason why we limited it to X86/S390/PPC is because those were the platforms which supported an RDRAND-like instruction at the time. Richard Henderson added support for ARM64 in commit 1a50ec0b3b2e ("arm64: Implement archrandom.h for ARMv8.5-RNG") in late January 2020.

So we should either add ARM64 to the dependency list, or we could, as you suggest, simply remove the dependency altogether. The tradeoff is that it will cause an extra CONFIG prompt on a number of platforms (mips, arm, sparc, etc.) where it will be utterly pointless since those architectures have no chance of support a RDRAND-like instruction.

Cheers,

- Ted

#2 lore raw

Hello Ted,

Quoted 5 lines · On Sun, 29 Mar 2020 12:56:24 -0400 "Theodore Y. Ts'o" <[email protected]> wrote:
Quoted 4 lines · Alexander Sverdlin

The option itself looks attractive for the embedded devices which often have HWRNG but less entropy from user-input. And these devices are often ARM/ARM64 or MIPS. The reason to limit it to X86/S390/PPC is not obvious.

Quoted 8 lines
Quoted 1 line

This feature is *only* applicable if the CPU supports a arch_get_random_seed_long() or arch_get_random_long(). I believe there are some server-class ARM64 CPU's that support such an instruction, but I don't believe any of the embedded arm64 --- and certainly non of the embedded arm --- SOC's support arch_get_random_long().

you are right! Thank you for the explanation!
I totally missed the fact that it's not connected to hwrng drivers... Please ignore the patch.

[...]

Quoted 6 lines

So we should either add ARM64 to the dependency list, or we could, as you suggest, simply remove the dependency altogether. The tradeoff is that it will cause an extra CONFIG prompt on a number of platforms (mips, arm, sparc, etc.) where it will be utterly pointless since those architectures have no chance of support a RDRAND-like instruction.

Signature · 1 line
Alexander Sverdlin.
#3 lore raw
In reply to Theodore Y. Ts'o
Quoted 19 lines · Theodore Y. Ts'o
Quoted 5 lines · Alexander Sverdlin

The option itself looks attractive for the embedded devices which often have HWRNG but less entropy from user-input. And these devices are often ARM/ARM64 or MIPS. The reason to limit it to X86/S390/PPC is not obvious.

This feature is *only* applicable if the CPU supports a arch_get_random_seed_long() or arch_get_random_long(). I believe there are some server-class ARM64 CPU's that support such an instruction, but I don't believe any of the embedded arm64 --- and certainly non of the embedded arm --- SOC's support arch_get_random_long().

The reason why we limited it to X86/S390/PPC is because those were the platforms which supported an RDRAND-like instruction at the time. Richard Henderson added support for ARM64 in commit 1a50ec0b3b2e ("arm64: Implement archrandom.h for ARMv8.5-RNG") in late January 2020.

Quoted 6 lines

So we should either add ARM64 to the dependency list, or we could, as you suggest, simply remove the dependency altogether. The tradeoff is that it will cause an extra CONFIG prompt on a number of platforms (mips, arm, sparc, etc.) where it will be utterly pointless since those architectures have no chance of support a RDRAND-like instruction.

Just for anyone watching, the dependency rework is already handled in linux-next by commit:

23ae0c17b89cfeb5 ("random: Make RANDOM_TRUST_CPU depend on ARCH_RANDOM")

... where x86, s390, ppc, and arm64 all select ARCH_RANDOM.

Mark.