Re: [RFC PATCH 09/35] ovl: stack file ops

[Miklos Szeredi <miklos@szeredi.hu>]

On Thu, Apr 26, 2018 at 4:56 PM, Vivek Goyal <vgoyal@redhat.com> wrote:
> On Thu, Apr 26, 2018 at 04:43:53PM +0200, Miklos Szeredi wrote:
>> On Thu, Apr 26, 2018 at 4:13 PM, Vivek Goyal <vgoyal@redhat.com> wrote:
>> > On Thu, Apr 12, 2018 at 05:08:00PM +0200, Miklos Szeredi wrote:
>> >
>> > [..]
>> >> diff --git a/fs/overlayfs/file.c b/fs/overlayfs/file.c
>> >> new file mode 100644
>> >> index 000000000000..a0b606885c41
>> >> --- /dev/null
>> >> +++ b/fs/overlayfs/file.c
>> >> @@ -0,0 +1,76 @@
>> >> +/*
>> >> + * Copyright (C) 2017 Red Hat, Inc.
>> >> + *
>> >> + * This program is free software; you can redistribute it and/or modify it
>> >> + * under the terms of the GNU General Public License version 2 as published by
>> >> + * the Free Software Foundation.
>> >> + */
>> >> +
>> >> +#include <linux/cred.h>
>> >> +#include <linux/file.h>
>> >> +#include <linux/xattr.h>
>> >> +#include "overlayfs.h"
>> >> +
>> >> +static struct file *ovl_open_realfile(const struct file *file)
>> >> +{
>> >> +     struct inode *inode = file_inode(file);
>> >> +     struct inode *upperinode = ovl_inode_upper(inode);
>> >> +     struct inode *realinode = upperinode ?: ovl_inode_lower(inode);
>> >> +     struct file *realfile;
>> >> +     const struct cred *old_cred;
>> >> +
>> >> +     old_cred = ovl_override_creds(inode->i_sb);
>> >> +     realfile = path_open(&file->f_path, file->f_flags | O_NOATIME,
>> >> +                          realinode, current_cred(), false);
>> >> +     revert_creds(old_cred);
>> >> +
>> >> +     pr_debug("open(%p[%pD2/%c], 0%o) -> (%p, 0%o)\n",
>> >> +              file, file, upperinode ? 'u' : 'l', file->f_flags,
>> >> +              realfile, IS_ERR(realfile) ? 0 : realfile->f_flags);
>> >> +
>> >> +     return realfile;
>> >> +}
>> >> +
>> >> +static int ovl_open(struct inode *inode, struct file *file)
>> >> +{
>> >> +     struct dentry *dentry = file_dentry(file);
>> >
>> > Hi Miklos,
>> >
>> > There is one thing I can't wrap my head around, so I better ask.
>> >
>> > file_dentry() will call ovl_d_real() and try to find dentry based on
>> > inode installed in f->f_inode. If ovl_d_real() can't find inode dentry
>> > matching the passed in inode, it warns.
>> >
>> > Assume, I have a stacked overlay configuration. Let me call top level
>> > overlay layer ovl1 and lower level overlay layer ovl2. Say I open a
>> > file foo.txt. Now ovl_open() in ovl1 decides that realinode is a lower
>> > inode and installs that inode f->f_inode of realfile. (This should be
>> > ovl2 layer inode, let me call it ovl2_inode). Now ovl_open() of ovl2 layer
>> > will be called and it will call file_dentry() and will look for dentry
>> > corresponding to ovl2_inode. I am wondering what if a copy up of foo.txt
>> > was triggered in ovl1 and by the time we called ovl_d_real(dentry,
>> > ovl2_inode), it will start comparing with inode of ovl1_upper and never
>> > find ovl2_inode.
>>
>> Okay, so we've modified ovl_d_real() to allow returning the overlay
>> dentry itself.  This is important: when we fail to match ovl1_upper
>> with ovl2_inode, well go on to get ovl2_dentry and call d_real()
>> recursively.  That recursive call should match the inode, return it to
>> outer ovl_d_real(), which again will match the inode and return
>> without warning.
>
> So current code does following.
>
> ovl_d_real() {
>         ...
>         ...
>
>         real = ovl_dentry_real(dentry);
>         if (inode == d_inode(real))
>                 return real;
>
>         /* Handle recursion */
>         if (unlikely(real->d_flags & DCACHE_OP_REAL))
>                 return real->d_op->d_real(real, inode);
> }
>
> If file got copied up in ovl1, then "real" will be ovl1_upper dentry. And
> upper is regular fs (only ovl1 lower is overlay), then it should not have
> DCACHE_OP_REAL set and that means we will not recurse further and not
> find ovl2 dentry matching ovl2_inode and print warning and return
> ovl1 dentry.
>
> What am I missing.

Ah,  that's indeed buggy.  The bug is in "[RFC PATCH 34/35] vfs:
simplify d_op->d_real()".

I've already reverted that (due to d_real_inode() acquiring a new
user) and the old code should be good (AFAICS).

Thanks,
Miklos