Re: [RFC PATCH 05/16] x86/efi: Get entropy through EFI random number generator protocol

[joeyli <jlee@suse.com>]

On Fri, Jul 31, 2015 at 10:59:12PM +0800, joeyli wrote:
> On Thu, Jul 30, 2015 at 05:11:44PM +0100, Matt Fleming wrote:
> > On Thu, 2015-07-16 at 22:25 +0800, Lee, Chun-Yi wrote:
> > > To grab random numbers through EFI protocol as one of the entropies
> > > source of swsusp key, this patch adds the logic for accessing EFI RNG
> > > (random number generator) protocol that's introduced since UEFI 2.4.
> > > 
> > > Signed-off-by: Lee, Chun-Yi <jlee@suse.com>
> > > ---
> > >  arch/x86/boot/compressed/efi_random.c | 193 ++++++++++++++++++++++++++++++++++
> > >  include/linux/efi.h                   |  46 ++++++++
> > >  2 files changed, 239 insertions(+)
> > 
> > [...]
> > 
> > > @@ -2,6 +2,191 @@
> > >  
> > >  #include <linux/efi.h>
> > >  #include <asm/archrandom.h>
> > > +#include <asm/efi.h>
> > > +
> > > +static efi_status_t efi_locate_rng(efi_system_table_t *sys_table,
> > > +				   void ***rng_handle)
> > > +{
> > > +	efi_guid_t rng_proto = EFI_RNG_PROTOCOL_GUID;
> > > +	unsigned long size = 0;
> > > +	efi_status_t status;
> > > +
> > > +	status = efi_call_early(locate_handle,
> > > +				EFI_LOCATE_BY_PROTOCOL,
> > > +				&rng_proto, NULL, &size, *rng_handle);
> > > +
> > > +	if (status == EFI_BUFFER_TOO_SMALL) {
> > > +		status = efi_call_early(allocate_pool,
> > > +					EFI_LOADER_DATA,
> > > +					size, (void **)rng_handle);
> > > +
> > > +		if (status != EFI_SUCCESS) {
> > > +			efi_printk(sys_table, " Failed to alloc mem for rng_handle");
> > > +			return status;
> > > +		}
> > > +
> > > +		status = efi_call_early(locate_handle,
> > > +					EFI_LOCATE_BY_PROTOCOL, &rng_proto,
> > > +					NULL, &size, *rng_handle);
> > > +	}
> > > +
> > > +	if (status != EFI_SUCCESS) {
> > > +		efi_printk(sys_table, " Failed to locate EFI_RNG_PROTOCOL");
> > > +		goto free_handle;
> > > +	}
> > > +
> > > +	return EFI_SUCCESS;
> > > +
> > > +free_handle:
> > > +	efi_call_early(free_pool, *rng_handle);
> > > +
> > > +	return status;
> > > +}
> > 
> > I would suggest setting *rng_handle = NULL at the beginning of this
> > function just because if we ever forget to set it that way in the caller
> > this free_pool call might do screwy things.
> >
> 
> Thanks for your suggestion, I will set NULL to *rng_handle.
>  
> > 
> > > +static bool efi_rng_supported(efi_system_table_t *sys_table)
> > > +{
> > > +	const struct efi_config *efi_early = __efi_early();
> > > +	u32 random = 0;
> > > +	efi_status_t status;
> > > +	void **rng_handle = NULL;
> > > +
> > > +	status = efi_locate_rng(sys_table, &rng_handle);
> > > +	if (status != EFI_SUCCESS)
> > > +		return false;
> > > +
> > > +	if (efi_early->is64)
> > > +		random = efi_rng_supported64(sys_table, rng_handle);
> > > +	else
> > > +		random = efi_rng_supported32(sys_table, rng_handle);
> > > +
> > > +	efi_call_early(free_pool, rng_handle);
> > > +
> > > +	return random;
> > 
> > Oops, 'random' isn't a bool but it should be.
> > 
> 
> I will change type of random to boot.
> 
> > > @@ -51,6 +236,14 @@ static unsigned long get_random_long(unsigned long entropy,
> > >  		use_i8254 = false;
> > >  	}
> > >  
> > > +	if (efi_rng_supported(sys_table)) {
> > > +		efi_printk(sys_table, " EFI_RNG");
> > > +		raw = efi_get_rng(sys_table);
> > > +		if (raw)
> > > +			random ^= raw;
> > > +		use_i8254 = false;
> > > +	}
> > > +
> > >  	if (use_i8254) {
> > >  		efi_printk(sys_table, " i8254");
> > >  		random ^= i8254();
> > 
> > Have you looked at the tradeoff in terms of boot time for building a key
> > array in 'unsigned long' chunks as opposed to passing the array and size
> > directly for the RNG protocol?
> >
> 
> I didn't really measure the speed, but directly passing array and size to
> RNG protocol should a bit faster than calling the protocol a could of times.
> 
> But, the key generation process only in first time building or trigger by
> user raises the rebuild flag. So, it doesn't affect to every booting time.
> 
> Due to I want let the whole key array more random, so each unsigned long
> chunk was mixed(xor) by following entropy:
>  + random long from RDRAND
>  + RDTSC
>  + random long from EFI RNG protocol
>  + last unsigned long chunk
> 
> Another reason is voiding the result of EFI RNG protocol to get weight
		   ^^^^^^^^^ avoiding
Sorry for my typo!

> higher than other source, in case too trust EFI RNG.

Joey Lee