Re: [PATCH v3] ovl: use a dedicated semaphore for dir upperfile caching

[Miklos Szeredi]

On Tue, Jan 05, 2021 at 08:47:41AM +0200, Amir Goldstein wrote:
> On Tue, Jan 5, 2021 at 2:36 AM Icenowy Zheng <icenowy@xxxxxxx> wrote:
> >
> > The function ovl_dir_real_file() currently uses the semaphore of the
> > inode to synchronize write to the upperfile cache field.
> 
> Although the inode lock is a rw_sem it is referred to as the "inode lock"
> and you also left semaphore in the commit subject.
> No need to re-post. This can be fixed on commit.
> 
> >
> > However, this function will get called by ovl_ioctl_set_flags(), which
> > utilizes the inode semaphore too. In this case ovl_dir_real_file() will
> > try to claim a lock that is owned by a function in its call stack, which
> > won't get released before ovl_dir_real_file() returns.
> >
> > Define a dedicated semaphore for the upperfile cache, so that the
> > deadlock won't happen.
> >
> > Fixes: 61536bed2149 ("ovl: support [S|G]ETFLAGS and FS[S|G]ETXATTR ioctls for directories")
> > Cc: stable@xxxxxxxxxxxxxxx # v5.10
> > Signed-off-by: Icenowy Zheng <icenowy@xxxxxxx>
> > ---
> > Changes in v2:
> > - Fixed missing replacement in error handling path.
> > Changes in v3:
> > - Use mutex instead of semaphore.
> >
> >  fs/overlayfs/readdir.c | 10 +++++-----
> >  1 file changed, 5 insertions(+), 5 deletions(-)
> >
> > diff --git a/fs/overlayfs/readdir.c b/fs/overlayfs/readdir.c
> > index 01620ebae1bd..3980f9982f34 100644
> > --- a/fs/overlayfs/readdir.c
> > +++ b/fs/overlayfs/readdir.c
> > @@ -56,6 +56,7 @@ struct ovl_dir_file {
> >         struct list_head *cursor;
> >         struct file *realfile;
> >         struct file *upperfile;
> > +       struct mutex upperfile_mutex;
> 
> That's a very specific name.
> This mutex protects members of struct ovl_dir_file, which could evolve
> into struct ovl_file one day (because no reason to cache only dir upper file),
> so I would go with a more generic name, but let's leave it to Miklos to decide.
> 
> He could have a different idea altogether for fixing this bug.

How about this (untested) patch?

It's a cleanup as well as a fix, but maybe we should separate the cleanup from
the fix...

Thanks,
Miklos
---

 fs/overlayfs/readdir.c |   23 +++++++----------------
 1 file changed, 7 insertions(+), 16 deletions(-)

--- a/fs/overlayfs/readdir.c
+++ b/fs/overlayfs/readdir.c
@@ -865,7 +865,7 @@ struct file *ovl_dir_real_file(const str
 
 	struct ovl_dir_file *od = file->private_data;
 	struct dentry *dentry = file->f_path.dentry;
-	struct file *realfile = od->realfile;
+	struct file *old, *realfile = od->realfile;
 
 	if (!OVL_TYPE_UPPER(ovl_path_type(dentry)))
 		return want_upper ? NULL : realfile;
@@ -874,29 +874,20 @@ struct file *ovl_dir_real_file(const str
 	 * Need to check if we started out being a lower dir, but got copied up
 	 */
 	if (!od->is_upper) {
-		struct inode *inode = file_inode(file);
-
 		realfile = READ_ONCE(od->upperfile);
 		if (!realfile) {
 			struct path upperpath;
 
 			ovl_path_upper(dentry, &upperpath);
 			realfile = ovl_dir_open_realfile(file, &upperpath);
+			if (IS_ERR(realfile))
+				return realfile;
 
-			inode_lock(inode);
-			if (!od->upperfile) {
-				if (IS_ERR(realfile)) {
-					inode_unlock(inode);
-					return realfile;
-				}
-				smp_store_release(&od->upperfile, realfile);
-			} else {
-				/* somebody has beaten us to it */
-				if (!IS_ERR(realfile))
-					fput(realfile);
-				realfile = od->upperfile;
+			old = cmpxchg_release(&od->upperfile, NULL, realfile);
+			if (old) {
+				fput(realfile);
+				realfile = old;
 			}
-			inode_unlock(inode);
 		}
 	}