TF-RMM/tf-rmm

get_random_seed in attestation_rnd.c could fail depending on Arm Architecture Reference Manual

Closed this issue · 3 comments

Yiwei-Zhuang commented

During cold boot, function calls attestaion_init->attest_rnd_prng_init->get_random_seed->arch_collect_entropy in order. The arch_collect_entropy return false when RNDR sets NZCV to 0b0100 and it causes the whole attestation_init process to fail.

static inline bool arch_collect_entropy(uint64_t *random)
{
	unsigned long rc;
	uint64_t val;

	asm volatile(
	"	mrs  %[val], " __XSTRING(RNDR) "\n"
	"	str  %[val], %[random_ptr]\n"
	"	cset %[rc], ne\n" /* RNDR sets NZCV to 0b0100 on failure */
	: [random_ptr] "=m" (*random),
	  [rc] "=r" (rc),
	  [val] "=r" (val)
	:
	: "cc"
	);
	return (rc == 1UL);
}

Related description can be found in the Arm Reference Manual in C6.1.4 under "Effect of random number generation instructions on Condition flags". However, I notice there is note just below it:

The definition of "reasonable period of time" is IMPLEMENTATION DEFINED. The expectation is that software might use this as an opportunity to reschedule or run a different routine, perhaps after a small number of retries have failed to return a valid value.

Does it mean the RNDR instruction may fail and software should retry it a few times after a failure?

shrutiarm commented

Hi Yiwei,
Thanks for reporting the issue.
Yes SW should retry RNDR instruction if it fails,
we have similar logic in pauth_init_enable_el2, which can be replicated in get_random_seed.
I will be making the changes.

Thanks

  • 👍1