[LU-3484] Incorrect identification of anonymous dentry as root under SLES11SP2

• Type: Bug
• Resolution: Fixed
• Priority: Critical
• Fix Version/s: Lustre 2.5.0
• Affects Version/s: Lustre 2.4.0
• Labels: mn4 patch
• Severity: 3
• Assignee: Bob Glossman (Inactive)
• Reporter: Patrick Farrell

Description

When attempting to export Lustre 2.4 via NFS with SLES11SP2, we ran in to an issue (See LU-3483 as well for another) where Lustre identifies anonymous dentries as root dentries and bugs when calling ll_invalidate_aliases.

Here's the code in question (from ll_invalidate_aliases in lustre/llite/dcache.c):

if (dentry->d_name.len == 1 && dentry->d_name.name[0] == '/') {
                        CERROR("called on root (?) dentry=%p, inode=%p "
                               "ino=%lu\n", dentry, inode, inode->i_ino);
                        lustre_dump_dentry(dentry, 1);
                        libcfs_debug_dumpstack(NULL);
                }

The problem is a change in d_obtain_alias that's been made in SLES11SP2. Here's where anonymous dentries are named, this is from d_obtain_alias in CentOS6.4 or SLES11SP1:

        static const struct qstr anonstring = { .name = "" };
        :
        :
        tmp = d_alloc(NULL, &anonstring);

In SLES11SP2, it's this:

        static const struct qstr anonstring = { .name = "/", .len = 1 };

        tmp = d_alloc(NULL, &anonstring);

As you can see, that looks like a root dentry to the check being done by Lustre. There's a comment just above saying this was the intention:

 * This alias can sometimes be seen by prepend_path, so make it look
 * the same as a d_alloc_root alias.
 */

[The commit is recorded here: http://kernel.opensuse.org/cgit/kernel-source/commit/?id=f9a956fc1dc4b5271662736b5016324d43b1b99f
and appears to be SLES specific.]

Ideally, Lustre would find another way to identify the root dentry here. I have simply commented out this check while doing further testing, but I'd like to understand:
Is there a reason Lustre can't use the same method to see if a dentry is root as the kernel does?
IE:

#define IS_ROOT(x) ((x) == (x)->d_parent)

If so, I'd suggest that.