> [ kernel people added to Cc ]
>
> On Tue, Sep 16, 2008 at 02:27:36PM +0200, Daniel Gollub wrote:
>> Hi,
>>
>> with 2.6.27-rc1(?) linux/dirent.h got removed Kbuild headers_install targe and
>> dirent struct got removed:
>>
>>
>> commit cf6ae8b50e0ee3f764392dadd1970e3f03c40773
>> Author: Adrian Bunk <bunk@kernel.org>
>> Date: Fri Jul 25 01:46:46 2008 -0700
>>
>> remove the in-kernel struct dirent{,64}
>>
>> The kernel struct dirent{,64} were different from the ones in
>> userspace.
>>
>> Even worse, we exported the kernel ones to userspace.
>>
>> But after the fat usages are fixed we can remove the conflicting
>> kernel versions.
>>
>>
>> This broke the build of some LTP syscalls testcases for "getdents", since the
>> <linux/dirent.h> include is missing and the "internal" dirent struct get/got
>> used in:
>>
>> testcases/kernel/syscalls/getdents/getdents01.c
>> testcases/kernel/syscalls/getdents/getdents03.c
>> testcases/kernel/syscalls/getdents/getdents02.c
>> testcases/kernel/syscalls/getdents/getdents04.c
>>
>>
http://ltp.cvs.sourceforge.net/ltp/ltp/testcases/kernel/syscalls/getdents/
>>
>> Is it safe to use the userspace dirent struct from <dirent.h> to fix the build
>> of this? Or what should be used for testing the getdents interface instead?
>
> Thanks for your report.
>
> I missed this since the kernel sys_getdents() did not use the
> "struct dirent" from the header but an own "struct linux_dirent".
>
> It might be safe since the LTP tests don't seem to care whether
> getdents() returns nonsense.
>
> But it is not correct since the reason why I removed this struct from
> the userspace headers was that the kernel and libc structs are
> *different*.
>
> E.g. when I enable the commented out code in getdents01.c I get the
> expected result that using <dirent.h> instead of <linux/dirent.h>
> changes a PASS into a FAIL. [1]
>
> IMHO the solution is something like:
> - move "struct linux_dirent" from fs/readdir.c to include/linux/dirent.h
> - re-export linux/dirent.h to userspace
> - change getdents(2) to consistently use "struct linux_dirent"
>
> Questions to fsdevel:
> - Do we need to preserve API compatibility with compatibility #define's
> for struct dirent{,64} here?
> - LTP does malloc(sizeof(struct dirent)).
> Is this considered a valid API usage that must continue to work?
>
>> best regards,
>> Daniel
