Eric Blake wrote:
> On 04/27/2011 01:37 AM, Jim Meyering wrote:
>> Eric Blake wrote:
>>> save-cwd needs either a working fchdir or a working getcwd(NULL).
>>> The only platform where getcwd fixes problems not solved by
>>> getcwd-lgpl is Linux, but fchdir works there.
>>>
>> Hi Eric,
>>
>> Unfortunately, that change would introduce a bug. save_cwd's
>> open (".", O_SEARCH) may fail in a directory where getcwd's
>> function would succeed, while getcwd-lgpl's would not.
>> Thus, we have to use the robust getcwd module there.
>
> Hmm, you're right. I'll go with Bruno's suggestion, then, of
> documenting that save-cwd "works" with getcwd-lgpl, but is more robust
> with getcwd. Coreutils, of course, uses getcwd, so there is no
> regression in behavior there, even if the save-cwd module relaxes its
> default dependency.
Reducing robustness solely to reduce dependencies...
That could introduce a subtle flaw in a package that depends on the
robustness -- if the package maintainer doesn't read gnulib's NEWS.
Not so sure I like that, but I do recognized the value.
I guess it's ok if it's well documented.
Be sure to mention in NEWS that by default (without explicit
inclusion of the getcwd module), the save-cwd module is no longer robust.
>> This condition will persist as long as we have to worry about
>> O_SEARCH being defined to O_RDONLY.
>
> I wonder if anyone on the kernel is working on this.
