On 05/21/2015 10:45 PM, Herbert Xu wrote:
>> Setting LC_ALL has the nice property that LC_COLLATE and LC_CTYPE are
>> guaranteed to be compatible; if you just set LC_COLLATE but leave
>> LC_CTYPE unchanged and unset LC_ALL, it is possible to attempt a
>> collation that assumes one character set while still living in a ctype
>> that assumes another, and get garbled results.
> 
> Show me an actual pair of values for these two that produce
> incorrect results for mkbuiltins and I'll happily change both.

'sort -b' uses isspace() to determine which characters to strip.  There
are locales with a larger set of characters where isspace() returns true
than for the LC_CTYPE=C locale.  Suppose that I can find a single-byte
locale where isblank('\xff') is true.  If that is the case, then the
input '\xffa\nb\n' will sort differently for 'LC_ALL=C sort -b' (output
'b\n'\xffa\n') than for 'LANG=C LC_CTYPE=$locale' (output '\xffa\nb\n')
because the change in CTYPE changes whether the \xff is ignored as a
blank or included as part of the name being sorted.

However, the man pages for 'locale(1)' and 'localedef(1)' did not make
it obvious for me how to perform a search that would easily find such a
locale, so I'm open to suggestions on how to prove my point via more
than just analysis.

And there's still the point that mkbuiltins is being run on controlled
input, where you are sticking only to a subset of characters that happen
to be portable (that is, you are unlikely to be tripped up by a locale
where \xff is a blank, since you are not using \xff in your input).

-- 
Eric Blake   eblake redhat com    +1-919-301-3266
Libvirt virtualization library http://libvirt.org