Re: sysfs regression: wrong link counts

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

On Mon, Jan 30, 2012 at 10:43:50PM +0000, Al Viro wrote:
> On Mon, Jan 30, 2012 at 02:27:17PM -0800, Greg KH wrote:
> > On Mon, Jan 30, 2012 at 10:10:59PM +0000, Alan Cox wrote:
> > > > Isn't there some other "proper" way of doing this in userspace, or is
> > > > this really the correct way?
> > > 
> > > You can look at the S_IFMT bits and stuff however link count indicating
> > > number of subdirectories is a standard Unix thing and used by many quite
> > > mundane tools as an optimisation.
> > 
> > Ah, yeah, that is easier.
> > 
> > Eric, care to fix this or want me to revert it?
> 
> Fix _what_?  Userland shite quoted upthread?  Because that's where the bug
> is - the mundane tools mentioned by Alan treat 1 in st_nlink as "no
> information about the number of subdirectories".  And shite might be too
> mild a term for the little gem in question, really...

To repeat this piece of bogosity for those who might've missed it:

/* returns !0 if sysfs filesystem was found, 0 otherwise */
int sensors_init_sysfs(void)
{
        struct stat statbuf;

        snprintf(sensors_sysfs_mount, NAME_MAX, "%s", "/sys");
        if (stat(sensors_sysfs_mount, &statbuf) < 0
         || statbuf.st_nlink <= 2)      /* Empty directory */
                return 0;

        return 1;
}

which is completely bogus - contrary to what it says in comments, it does
*not* check anything about sysfs (or directories being empty).  Checking
that sysfs is mounted on /sys could be done by statfs(2) and checking
->f_type, or, considering what the code in lib/sysfs.c is doing, just
checking that /sys/class is readable and failing otherwise.

Sigh...