Re: [RFC] readdir mess

[Al Viro <viro@ZenIV.linux.org.uk>]

On Tue, Aug 12, 2008 at 07:22:40PM +0100, Al Viro wrote:

> PS: we might get away with both, if we used _positive_ values as well.
> E.g. have getdents() filldir return 1 if we are out of buffer *and*
> have ->previous != NULL (and -EINVAL if we are out of buffer on the
> first call)...  And have some other positive constant for "->readdir()
> didn't feel like going all the way to the end of directory".

FWIW, how about that sequence:

Patch 1:
Turn all filldir(...) < 0 into filldir() != 0 in ->readdir() instances,
no changes other than that.  Everything should keep working as-is.

Patch 2:
Make fillonedir() return 1 on the second call; make filldir() et.al.
return 1 instead of -EINVAL if we have ->previous != NULL.  Again,
should be no breakage.

Patch 3: switch ->readdir() to your "return anything non-null we got from
callback".  AFAICS, main callers will see no breakage, but in any case
we have few enough of those to adjust them as needed first.

Patch 4: get rid of ->error and its ilk; adjust callers in obvious ways
(e.g. sys_gtedents() would bail out on negative from vfs_readdir() as
it does now and treat 0 and 1 in the same way - put_user() ? -EFAULT : <how
much did we copy).  Callers can be taken care one by one.  Again, no breakage
and everything's bisectable.

Patch 5 (maybe):
#define READDIR_MORE INT_MAX
Have ->readdir() instances that decide to stop once they'd done several
filldir calls return it if there's still more left.
Have vfs_readdir() loop calling ->readdir() as long as it gets READDIR_MORE.
Get rid of weird loops in callers.

I'm not sure that the last one is needed - we might be better off just by
making the such instances loop themselves.  In any case, loops in callers
(nfsd, etc.) are begging for trouble...