On Wed, Feb 01, 2012 at 02:44:32PM -0800, Eric W. Biederman wrote:
> Greg Kroah-Hartman <gregkh@linuxfoundation.org> writes:
> 
> > On Wed, Feb 01, 2012 at 02:21:59PM -0800, Eric W. Biederman wrote:
> >> 
> >> lm_sensors and possibly other applications get confused if all sysfs
> >> directories return nlink == 1.  The lm_sensors code that got confused
> >> was just wrong and a fixed version of lm_sensors should be released
> >> shortly.
> >> 
> >> There may be other applications that have problems with sysfs return
> >> nlink == 1 for directories.  To allow people to continue to use old
> >> versions of userspace with new kernels add to sysfs a compile time
> >> option to maintain mostly precise directory counts for those people who
> >> don't mind the cost.
> >> 
> >> I have moved where we keep nlink in sysfs_dirent as compared to previous
> >> versions of subdirectory counting to a location that packs better.
> >> 
> >> Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
> >> ---
> >>  fs/sysfs/Kconfig |   15 +++++++++++++++
> >>  fs/sysfs/dir.c   |    8 ++++++++
> >>  fs/sysfs/inode.c |    2 ++
> >>  fs/sysfs/sysfs.h |   38 ++++++++++++++++++++++++++++++++++++++
> >>  4 files changed, 63 insertions(+), 0 deletions(-)
> >> 
> >> diff --git a/fs/sysfs/Kconfig b/fs/sysfs/Kconfig
> >> index 8c41fea..9b403e9 100644
> >> --- a/fs/sysfs/Kconfig
> >> +++ b/fs/sysfs/Kconfig
> >> @@ -21,3 +21,18 @@ config SYSFS
> >>  	example, "root=03:01" for /dev/hda1.
> >> 
> >>  	Designers of embedded systems may wish to say N here to conserve space.
> >> +
> >> +config SYSFS_COUNT_LINKS
> >> +	bool "sysfs count subdirectoires to support buggy applications"
> >> +	default n
> >
> > As we don't want to break things, this should be default y, right?
> 
> The new behavior is backwards compatible.  What the new behavior is not
> is bug compatible.  So nothing *should* break.

"should", but you really don't know, as all we have is one report so
far.

> Furthermore the breaking we have seen so far is limited to just
> lm_sensors.  That is exactly one program that is not a server failing to
> start.  That seems pretty minor in the worst case.
> 
> So I really don't expect anyone who ships 3.4 to enable this option.

What about users who use a new kernel on old userspace, which happens
all the time (i.e. all the kernel developers themselves)?

> I have written the option solely so that in case my assessment turns out
> to be wrong there is already a tested solution.  I have been through the
> pain of not being able to upgrade/test a new kernel because of a
> backwards incompatible change and it can be very unpleasant.

I'd really prefer this to be default 'y', and if a distro knows it can
turn it off to save time/space, it can.

thanks,

greg k-h