kernfs: invoke kernfs_unmap_bin_file() directly from __kernfs_remove()
authorTejun Heo <tj@kernel.org>
Fri, 10 Jan 2014 13:57:23 +0000 (08:57 -0500)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Fri, 10 Jan 2014 21:48:08 +0000 (13:48 -0800)
kernfs_unmap_bin_file() is supposed to unmap all memory mappings of
the target file before kernfs_remove() finishes; however, it currently
is being called from kernfs_addrm_finish() and has the same race
problem as the original implementation of deactivation when there are
multiple removers - only the remover which snatches the node to its
addrm_cxt->removed list is guaranteed to wait for its completion
before returning.

It can be fixed by moving kernfs_unmap_bin_file() invocation from
kernfs_addrm_finish() to __kernfs_remove().  The function may be
called multiple times but that shouldn't do any harm.

We end up dropping kernfs_mutex in the removal loop and the node may
be removed inbetween by someone else.  kernfs_unlink_sibling() is
updated to test whether the node has already been removed and return
accordingly.  __kernfs_remove() in turn performs post-unlinking
cleanup only if it actually unlinked the node.

KERNFS_HAS_MMAP test is moved out of the unmap function into
__kernfs_remove() so that we don't unlock kernfs_mutex unnecessarily.
While at it, drop the now meaningless "bin" qualifier from the
function name.

v2: Rewritten to fit the v2 restructuring of removal path.  HAS_MMAP
    test relocated.

Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
fs/kernfs/dir.c
fs/kernfs/file.c
fs/kernfs/kernfs-internal.h

index e565ec096ae997e16369ec7a698d1a43ae1edce3..f878e4f2efe7681ad35de92a74c1a1a60c894cac 100644 (file)
@@ -121,13 +121,17 @@ static int kernfs_link_sibling(struct kernfs_node *kn)
  *     Locking:
  *     mutex_lock(kernfs_mutex)
  */
-static void kernfs_unlink_sibling(struct kernfs_node *kn)
+static bool kernfs_unlink_sibling(struct kernfs_node *kn)
 {
+       if (RB_EMPTY_NODE(&kn->rb))
+               return false;
+
        if (kernfs_type(kn) == KERNFS_DIR)
                kn->parent->dir.subdirs--;
 
        rb_erase(&kn->rb, &kn->parent->dir.children);
        RB_CLEAR_NODE(&kn->rb);
+       return true;
 }
 
 /**
@@ -496,7 +500,6 @@ void kernfs_addrm_finish(struct kernfs_addrm_cxt *acxt)
 
                acxt->removed = kn->u.removed_list;
 
-               kernfs_unmap_bin_file(kn);
                kernfs_put(kn);
        }
 }
@@ -854,23 +857,44 @@ static void __kernfs_remove(struct kernfs_addrm_cxt *acxt,
 
        /* unlink the subtree node-by-node */
        do {
-               struct kernfs_iattrs *ps_iattr;
-
                pos = kernfs_leftmost_descendant(kn);
 
-               if (pos->parent) {
-                       kernfs_unlink_sibling(pos);
+               /*
+                * We're gonna release kernfs_mutex to unmap bin files,
+                * Make sure @pos doesn't go away inbetween.
+                */
+               kernfs_get(pos);
+
+               /*
+                * This must be come before unlinking; otherwise, when
+                * there are multiple removers, some may finish before
+                * unmapping is complete.
+                */
+               if (pos->flags & KERNFS_HAS_MMAP) {
+                       mutex_unlock(&kernfs_mutex);
+                       kernfs_unmap_file(pos);
+                       mutex_lock(&kernfs_mutex);
+               }
+
+               /*
+                * kernfs_unlink_sibling() succeeds once per node.  Use it
+                * to decide who's responsible for cleanups.
+                */
+               if (!pos->parent || kernfs_unlink_sibling(pos)) {
+                       struct kernfs_iattrs *ps_iattr =
+                               pos->parent ? pos->parent->iattr : NULL;
 
                        /* update timestamps on the parent */
-                       ps_iattr = pos->parent->iattr;
                        if (ps_iattr) {
                                ps_iattr->ia_iattr.ia_ctime = CURRENT_TIME;
                                ps_iattr->ia_iattr.ia_mtime = CURRENT_TIME;
                        }
+
+                       pos->u.removed_list = acxt->removed;
+                       acxt->removed = pos;
                }
 
-               pos->u.removed_list = acxt->removed;
-               acxt->removed = pos;
+               kernfs_put(pos);
        } while (pos != kn);
 }
 
index 231a171f48b6f107f729594a82165f4973e56073..404ffd2f27bc0bbbcfa7e45f7d812c7fe99495e9 100644 (file)
@@ -700,14 +700,11 @@ static int kernfs_fop_release(struct inode *inode, struct file *filp)
        return 0;
 }
 
-void kernfs_unmap_bin_file(struct kernfs_node *kn)
+void kernfs_unmap_file(struct kernfs_node *kn)
 {
        struct kernfs_open_node *on;
        struct kernfs_open_file *of;
 
-       if (!(kn->flags & KERNFS_HAS_MMAP))
-               return;
-
        spin_lock_irq(&kernfs_open_node_lock);
        on = kn->attr.open;
        if (on)
index 57a93f4d645ca708fad37d0aafecd3bdefbe3889..e9ec38c86074fb79dfe4552e8a17680400ed1f4a 100644 (file)
@@ -113,7 +113,7 @@ struct kernfs_node *kernfs_new_node(struct kernfs_root *root, const char *name,
  */
 extern const struct file_operations kernfs_file_fops;
 
-void kernfs_unmap_bin_file(struct kernfs_node *kn);
+void kernfs_unmap_file(struct kernfs_node *kn);
 
 /*
  * symlink.c