linux-kernel - Re: [PATCH 2/2] tools/nolibc: implement strtol() and friends

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite for Android: free password hash cracker in your pocket

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date: Thu, 25 Apr 2024 18:29:18 +0200
From: Willy Tarreau <w@....eu>
To: Thomas Weißschuh <linux@...ssschuh.net>
Cc: Shuah Khan <shuah@...nel.org>, linux-kernel@...r.kernel.org,
        linux-kselftest@...r.kernel.org
Subject: Re: [PATCH 2/2] tools/nolibc: implement strtol() and friends

Hi Thomas,

On Thu, Apr 25, 2024 at 06:09:27PM +0200, Thomas Weißschuh wrote:
> The implementation always works on uintmax_t values.
> 
> This is inefficient when only 32bit are needed.
> However for all functions this only happens for strtol() on 32bit
> platforms.

That's indeed very useful! I think there's two small bugs below where
the second one hides the first one:

> +static __attribute__((unused))
> +uintmax_t __strtox(const char *nptr, char **endptr, int base, intmax_t lower_limit, uintmax_t upper_limit)
> +{
> +	const char signed_ = lower_limit != 0;
> +	unsigned char neg = 0, overflow = 0;
> +	uintmax_t val = 0, limit, old_val;
> +	char c;
> +
> +	if (base < 0 || base > 35) {
                        ^^^^^^^^^
should be 36 otherwise you won't support [0-9a-z].

> +		SET_ERRNO(EINVAL);
> +		goto out;
> +	}
(...)
> +		if (c > base)
> +			goto out;

This should be "c >= base" otherwise 'z' is accepted in base 35 for
example. I think it could be useful to add one more test covering base
36 to make sure all chars pass ?

> +	if (endptr)
> +		*endptr = (char *)nptr;
> +	return (neg ? -1 : 1) * val;

I just checked to see what the compiler does on this and quite frequently
it emits a multiply while the other approach involving only a negation is
always at least as short:

	return neg ? -val : val;

E.g. here's the test code:

  long fct1(long neg, long val)
  {
        return (neg ? -1 : 1) * val;
  }

  long fct2(long neg, long val)
  {
        return neg ? -val : val;
  }

- on x86_64 with gcc-13.2 -Os:

  0000000000000000 <fct1>:
     0:   f7 df                   neg    %edi
     2:   48 19 c0                sbb    %rax,%rax
     5:   48 83 c8 01             or     $0x1,%rax
     9:   48 0f af c6             imul   %rsi,%rax
     d:   c3                      ret    
  
  000000000000000e <fct2>:
     e:   48 89 f0                mov    %rsi,%rax
    11:   85 ff                   test   %edi,%edi
    13:   74 03                   je     18 <fct2+0xa>
    15:   48 f7 d8                neg    %rax
    18:   c3                      ret    

- on riscv64 with 13.2 -Os:

  0000000000000000 <fct1>:
     0:   c509                    beqz    a0,a 
     2:   557d                    li      a0,-1
     4:   02b50533                mul     a0,a0,a1
     8:   8082                    ret
     a:   4505                    li      a0,1
     c:   bfe5                    j       4
  
  000000000000000e <fct2>:
     e:   c119                    beqz    a0,14
    10:   40b005b3                neg     a1,a1
    14:   852e                    mv      a0,a1
    16:   8082                    ret

So IMHO it would be better to go the simpler way even if these are just a
few bytes (and possibly ones less mul on some slow archs).

Thanks!
Willy