2020-03-28 16:57:46

by Maciej Żenczykowski

[permalink] [raw]
Subject: Re: [RFC PATCH v1 18/50] net/ipv6/addrconf.c: Use prandom_u32_max for rfc3315 backoff time computation

> /* multiply 'initial retransmission time' by 0.9 .. 1.1 */
> - u64 tmp = (900000 + prandom_u32() % 200001) * (u64)irt;
> - do_div(tmp, 1000000);
> - return (s32)tmp;
> + s32 range = irt / 5;
> + return irt - (s32)(range/2) + (s32)prandom_u32_max(range);

The cast on range/2 looks entirely spurious

> /* multiply 'retransmission timeout' by 1.9 .. 2.1 */
> - u64 tmp = (1900000 + prandom_u32() % 200001) * (u64)rt;
> - do_div(tmp, 1000000);
> - if ((s32)tmp > mrt) {
> + s32 range = rt / 5;
> + s32 tmp = 2*rt - (s32)(range/2) + (s32)prandom_u32_max(range);

Here as well. Honestly the cast on prandom might also not be
necessary, but that at least has a reason.

> + if (tmp > mrt) {
> /* multiply 'maximum retransmission time' by 0.9 .. 1.1 */
> - tmp = (900000 + prandom_u32() % 200001) * (u64)mrt;
> - do_div(tmp, 1000000);
> + range = mrt / 5;
> + tmp = mrt - (s32)(range/2) + (s32)prandom_u32_max(range);

Ditto.

> }
> return (s32)tmp;
> }

- Maciej


2020-03-28 17:38:12

by George Spelvin

[permalink] [raw]
Subject: Re: [RFC PATCH v1 18/50] net/ipv6/addrconf.c: Use prandom_u32_max for rfc3315 backoff time computation

On Sat, Mar 28, 2020 at 09:56:58AM -0700, Maciej ?enczykowski wrote:
>> /* multiply 'initial retransmission time' by 0.9 .. 1.1 */
>> - u64 tmp = (900000 + prandom_u32() % 200001) * (u64)irt;
>> - do_div(tmp, 1000000);
>> - return (s32)tmp;
>> + s32 range = irt / 5;
>> + return irt - (s32)(range/2) + (s32)prandom_u32_max(range);
>
> The cast on range/2 looks entirely spurious

You're absolutely right; sorry about that. I was trying to
preserve the previous code's mixture of signed and unsigned types
and managed to confuse myself.

(I think I got distracted researching whether the inputs could be
negative.)

>> /* multiply 'retransmission timeout' by 1.9 .. 2.1 */
>> - u64 tmp = (1900000 + prandom_u32() % 200001) * (u64)rt;
>> - do_div(tmp, 1000000);
>> - if ((s32)tmp > mrt) {
>> + s32 range = rt / 5;
>> + s32 tmp = 2*rt - (s32)(ran ge/2) + (s32)prandom_u32_max(range);
>
> Here as well. Honestly the cast on prandom might also not be
> necessary, but that at least has a reason.

The whole thing should go. How about just doing it all in unsigned:

static inline s32 rfc3315_s14_backoff_init(s32 irt)
{
/* multiply 'initial retransmission time' by 0.9 .. 1.1 */
u32 range = irt / 5u;
return irt - range/2 + prandom_u32_max(range);
}

static inline s32 rfc3315_s14_backoff_update(s32 rt, s32 mrt)
{
/* multiply 'retransmission timeout' by 1.9 .. 2.1 */
u32 range = rt / 5u;
u32 tmp = 2u*rt - range/2 + prandom_u32_max(range);
if (tmp > mrt) {
/* multiply 'maximum retransmission time' by 0.9 .. 1.1 */
range = mrt / 5u;
tmp = mrt - range/2 + prandom_u32_max(range);
}
return tmp;
}

That lets "range/2" be implemented as a 1-bit shift.

An interesting question for the latter is whether
"prandom_u32_max(range) - range/2" can be considered a common
subexpression, or is they have to be *independent* random values.