Re: [PATCH] for glibc-2.2 linux-specific getdents code

[Trond Myklebust <trond dot myklebust at fys dot uio dot no>]

>>>>> " " == Ulrich Drepper <drepper@redhat.com> writes:

     > Trond Myklebust <trond.myklebust@fys.uio.no> writes:
    >> No. NFS directory offsets have never been signed. They never
    >> were in 2.2.x, and they aren't in 2.4.x.

     > It doesn't matter what NFS wants or does, I'm using the kernel
     > interface.  It is supposed to be filesystem independent.

No you're not. You're breaking a kernel interface that dates from
before linux-2.2.x for which such unsigned directory offsets were
allowed.
The fact that the kernel defintion of the struct dirent(|64) has an
entry d_off of type (|l)off_t is irrelevant.

The kernel dirent structures of most UNIX systems are defined in
exactly the same way. However those same UNIX kernels are required to
support the existence of non-POSIX filesystems (of which NFS is one)
and therefore accept the fact that d_off is opaque. So has Linux up to
now.


For instance from the SunOS-5.7 manpage for dirent:

     The  dirent structure is defined:

     struct  dirent {
             ino_t           d_ino;
             off_t           d_off;
             unsigned short  d_reclen;
             char            d_name[1];
     };

     The d_ino is a number which is  unique for each file in  the
     file  system.  The   d_off  entry  contains a value which is
     interpretable only by  the filesystem that generated it.

The last line being the key here.

Cheers,
  Trond