Received-SPF: pass (google.com: domain of linux-crypto+bounces-4183-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:45e3:2400::1 as permitted sender) client-ip=2604:1380:45e3:2400::1;
Message-ID: <7eb6bf4b-5510-48fe-aa6c-ac5207d5a2c1@cryptogams.org>
Date: Wed, 15 May 2024 15:33:06 +0200
Precedence: bulk
MIME-Version: 1.0
User-Agent: Mozilla Thunderbird
Subject: Re: [PATCH 2/3] crypto: X25519 core functions for ppc64le
From: Andy Polyakov <appro@cryptogams.org>
To: Danny Tsen <dtsen@linux.ibm.com>, linux-crypto@vger.kernel.org
Cc: herbert@gondor.apana.org.au, leitao@debian.org, nayna@linux.ibm.com,
 linux-kernel@vger.kernel.org, linuxppc-dev@lists.ozlabs.org,
 mpe@ellerman.id.au, ltcgcw@linux.vnet.ibm.com, dtsen@us.ibm.com
References: <20240514173835.4814-1-dtsen@linux.ibm.com>
 <20240514173835.4814-3-dtsen@linux.ibm.com>
 <847f2e4f-ace1-415d-b129-ed2751429eec@cryptogams.org>
Content-Language: en-US
In-Reply-To: <847f2e4f-ace1-415d-b129-ed2751429eec@cryptogams.org>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 8bit

>> +static void cswap(fe51 p, fe51 q, unsigned int bit)
>> +{
>> +    u64 t, i;
>> +    u64 c = 0 - (u64) bit;
>> +
>> +    for (i = 0; i < 5; ++i) {
>> +        t = c & (p[i] ^ q[i]);
>> +        p[i] ^= t;
>> +        q[i] ^= t;
>> +    }
>> +}
> 
> The "c" in cswap stands for "constant-time," and the problem is that 
> contemporary compilers have exhibited the ability to produce 
> non-constant-time machine code as result of compilation of the above 
> kind of technique. The outcome is platform-specific and ironically some 
> of PPC code generators were observed to generate "most" 
> non-constant-time code. "Most" in sense that execution time variations 
> would be most easy to catch.

Just to substantiate the point, consider 
https://godbolt.org/z/faYnEcPT7, and note the conditional branch in the 
middle of the loop, which flies in the face of constant-time-ness. In 
case you object 'bit &= 1' on line 7 in the C code. Indeed, if you 
comment it out, the generated code will be fine. But the point is that 
the compiler is capable of and was in fact observed to figure out that 
the caller passes either one or zero and generate the machine code in 
the assembly window. In other words 'bit &= 1' is just a reflection of 
what the caller does.

> ... the permanent solution is to do it 
> in assembly. I can put together something...

Though you should be able to do this just as well :-) So should I or 
would you?

Cheers.