Re: [PATCH v25 01/12] Linux Random Number Generator

["Stephan Müller" <smueller@chronox.de>]

Am Samstag, 16. November 2019, 19:13:23 CET schrieb Nicolai Stange:

Hi Nicolai,

> Hi Stephan,
> 
> Stephan Müller <smueller@chronox.de> writes:
> > +/* Initialize the default DRNG during boot */
> 
> I think that this can get called a bit too early through the
> get_random_bytes() invoked from e.g. boot_init_stack_canary(): in
> start_kernel(), there is
> 
> 	boot_init_stack_canary();
> 
> 	time_init();
> 
> On ARM (at least with arm_arch_timer.c), get_cycles() would return 0
> until
> 
>   time_init() => timer_probe() => arch_timer_of_init() =>
>   arch_timer_common_init() => arch_timer_arch_init() =>
>   arch_timer_delay_timer_register() => register_current_timer_delay()
> 
> has executed and thus, ...
> 
> > +void lrng_drngs_init_cc20(void)
> > +{
> > +	unsigned long flags = 0;
> > +
> > +	if (lrng_get_available())
> > +		return;
> > +
> > +	lrng_sdrng_lock(&lrng_sdrng_init, &flags);
> > +	if (lrng_get_available()) {
> > +		lrng_sdrng_unlock(&lrng_sdrng_init, &flags);
> > +		return;
> > +	}
> > +
> > +	if (random_get_entropy() || random_get_entropy()) {
> > +		/*
> > +		 * As the highres timer is identified here, previous interrupts
> > +		 * obtained during boot time are treated like a lowres-timer
> > +		 * would have been present.
> > +		 */
> > +		lrng_pool_configure(true, LRNG_IRQ_ENTROPY_BITS);
> > +	} else {
> > +		lrng_health_disable();
> > +		lrng_pool_configure(false, LRNG_IRQ_ENTROPY_BITS *
> > +					   LRNG_IRQ_OVERSAMPLING_FACTOR);
> > +		pr_warn("operating without high-resolution timer and applying "
> > +			"IRQ oversampling factor %u\n",
> > +			LRNG_IRQ_OVERSAMPLING_FACTOR);
> 
> ... LRNG thinks that no high-res timer is available even though there
> is:
> 
> [    0.000000] lrng_sdrng: operating without high-resolution timer and
> applying IRQ oversampling factor 10 [    0.000000] lrng_chacha20: ChaCha20
> core initialized
> [    0.000000] lrng_chacha20: ChaCha20 core initialized
> [    0.000014] sched_clock: 32 bits at 1000kHz, resolution 1000ns, wraps
> every 2147483647500ns [    0.000036] clocksource: timer: mask: 0xffffffff
> max_cycles: 0xffffffff, max_idle_ns: 1911260446275 ns [    0.000114]
> bcm2835: system timer (irq = 27)
> [    0.000594] arch_timer: cp15 timer(s) running at 19.20MHz (phys).
> [    0.000613] clocksource: arch_sys_counter: mask: 0xffffffffffffff
> max_cycles: 0x46d987e47, max_idle_ns: 440795202767 ns [    0.000631]
> sched_clock: 56 bits at 19MHz, resolution 52ns, wraps every 4398046511078ns
> [    0.000645] Switching to timer-based delay loop, resolution 52ns
> 
> Note that this last line comes from aforementioned
> register_current_timer_delay().
> 
> Similarly, get_random_bytes() can get called quite early through
> WARN() => warn_slowpath_fmt() => __warn() => print_oops_end_marker() =>
> init_oops_id().
> 
> Perhaps it would make sense not to do the (pool + health test)
> initalization "on-demand", but rather make sure it happens at some
> well-defined point after time_init()? Or at least that the pool +
> the health tests get reconfigured eventually?


Thank you very much for testing this and reporting it.

I have extracted the initialization of the time source into its own function 
and execute it with core_initcall. With that, the DRNG is initialized at the 
time it needs initialization.

But the time source check is now done when time_init is completed as first 
time_init is called and then arch_call_rest_init -> rest_init -> kernel_init -
> kernel_init_freeable -> do_basic_setup -> do_initcalls where the initcalls 
registered with the core_initcall callback are now executed as the first 
batch.


> 
> 
> Thanks,
> 
> Nicolai
> 
> P.S: include/linux/lrng.h needs an #include <linux/errno.h> for
>      CONFIG_LRNG_DRNG_SWITCH=n

Thank you, added.
> 
> > +	}
> > +
> > +	lrng_sdrng_reset(&lrng_sdrng_init);
> > +	lrng_cc20_init_state(&secondary_chacha20);
> > +	lrng_state_init_seed_work();
> > +	lrng_sdrng_unlock(&lrng_sdrng_init, &flags);
> > +
> > +	lrng_sdrng_lock(&lrng_sdrng_atomic, &flags);
> > +	lrng_sdrng_reset(&lrng_sdrng_atomic);
> > +	/*
> > +	 * We do not initialize the state of the atomic DRNG as it is identical
> > +	 * to the secondary DRNG at this point.
> > +	 */
> > +	lrng_sdrng_unlock(&lrng_sdrng_atomic, &flags);
> > +
> > +	lrng_trng_init();
> > +
> > +	lrng_set_available();
> > +}
> > +
> > +/* Reset LRNG such that all existing entropy is gone */


Ciao
Stephan