Re: [PATCH] NFS: Replace null dentries that appear in readdir's list

[David Howells <dhowells@redhat.com>]

Andrew Morton <akpm@osdl.org> wrote:

> VFS: Busy inodes after unmount of 0:15. Self-destruct in 5 seconds.  Have a
> nice day...

Does the same happen with the first patch I sent you (see attached)?  I know
Trond doesn't like it, but it would be a useful data point if you could try
it.

David

---
NFS: Revalidate on readdir referring to null dentry

From: David Howells <dhowells@redhat.com>

Have nfs_readdir_lookup() force dentry revalidation when it comes across a
name specified by a READDIR RPC call that corresponds to a negative dentry in
the dcache.

This can be caused by an optimisation in nfs_lookup() that causes a dentry to
be incorrectly left as negative when mkdir() or similar is aborted by SELinux
mid-procedure.

This can be triggered by mounting through autofs4 a server:/ NFS share for
which there are other exports available on that server.  SELinux also has to
be turned on in enforcing mode to abort mid-flow the mkdir operation performed
by autofs4.

The problematic sequence of events is this:

 (1) nfs_lookup() is called by sys_mkdirat() -> lookup_create() ->
     __lookup_hash() with intent to create exclusively set in the nameidata:

	nd->flags == LOOKUP_CREATE
	nd->intent.open.flags == O_EXCL

 (2) nfs_lookup() has an optimisation to avoid going to the server in this
     case, presumably since the nfs_mkdir() op or whatever will deal with the
     conflict.

 (3) nfs_lookup() returns successfully, leaving the dentry in a negative state,
     but attached to the parent directory.

 (4) sys_mkdirat() calls vfs_mkdir() which calls may_create().  may_create()
     checks that the directory has MAY_WRITE and MAY_EXEC permissions.

 (5) may_create() calls nfs_permission(), which grants permission.

 (6) may_create() calls security_inode_permission(), which calls SELinux, which
     then _DENIES_ permission.

 (7) may_create() fails, and vfs_mkdir() then fails and sys_mkdirat() then
     fails (as does sys_mkdir).

     _However_, the new dentry is left in the negative state, with no
     consultation of the server.

 (8) The parent directory is listed, and the name of the new dentry is
     returned.

 (9) stat on the new dentry fails (because it's negative), and "ls -l" returns
     "?---------" as the file type and mode.

This fix makes step (8) cause a revalidation to occur on the dentry at the
start of step (9).

Signed-Off-By: David Howells <dhowells@redhat.com>
---

 fs/nfs/dir.c           |   16 +++++++++++++++-
 include/linux/dcache.h |    5 +++++
 2 files changed, 20 insertions(+), 1 deletions(-)

diff --git a/fs/nfs/dir.c b/fs/nfs/dir.c
index e746ed1..901b382 100644
--- a/fs/nfs/dir.c
+++ b/fs/nfs/dir.c
@@ -745,6 +745,13 @@ static int nfs_lookup_revalidate(struct 
 	nfs_inc_stats(dir, NFSIOS_DENTRYREVALIDATE);
 	inode = dentry->d_inode;
 
+	if (dentry->d_flags & DCACHE_NEED_REVALIDATE) {
+		spin_lock(&dentry->d_lock);
+		dentry->d_flags &= ~DCACHE_NEED_REVALIDATE;
+		spin_unlock(&dentry->d_lock);
+		goto out_bad;
+	}
+
 	if (!inode) {
 		if (nfs_neg_need_reval(dir, dentry, nd))
 			goto out_bad;
@@ -1105,8 +1112,15 @@ static struct dentry *nfs_readdir_lookup
 	}
 	name.hash = full_name_hash(name.name, name.len);
 	dentry = d_lookup(parent, &name);
-	if (dentry != NULL)
+	if (dentry != NULL) {
+		/* negative dentries must be reconsidered */
+		if (!IS_ERR(dentry) && !dentry->d_inode) {
+			spin_lock(&dentry->d_lock);
+			dentry->d_flags |= DCACHE_NEED_REVALIDATE;
+			spin_unlock(&dentry->d_lock);
+		}
 		return dentry;
+	}
 	if (!desc->plus || !(entry->fattr->valid & NFS_ATTR_FATTR))
 		return NULL;
 	/* Note: caller is already holding the dir->i_mutex! */
diff --git a/include/linux/dcache.h b/include/linux/dcache.h
index 63f64a9..c401a7d 100644
--- a/include/linux/dcache.h
+++ b/include/linux/dcache.h
@@ -176,6 +176,11 @@ #define DCACHE_UNHASHED		0x0010	
 
 #define DCACHE_INOTIFY_PARENT_WATCHED	0x0020 /* Parent inode is watched */
 
+#define DCACHE_NEED_REVALIDATE	0x0040
+	/* Dentry needs revalidation by filesystem.  Set by NFS, for example,
+	 * when we see in a directory listing a file for which we have a
+	 * negative dentry */
+
 extern spinlock_t dcache_lock;
 
 /**

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/