Re: [PATCH 5/5] pstore: inode: Use cleanup.h for struct pstore_private

[Kees Cook]

On Sat, Dec 02, 2023 at 10:27:06PM +0000, Al Viro wrote:
> On Sat, Dec 02, 2023 at 01:22:15PM -0800, Kees Cook wrote:
> 
> >  static void *pstore_ftrace_seq_start(struct seq_file *s, loff_t *pos)
> >  {
> > @@ -338,9 +339,8 @@ int pstore_mkfile(struct dentry *root, struct pstore_record *record)
> >  {
> >  	struct dentry		*dentry;
> >  	struct inode		*inode __free(iput) = NULL;
> > -	int			rc = 0;
> >  	char			name[PSTORE_NAMELEN];
> > -	struct pstore_private	*private, *pos;
> > +	struct pstore_private	*private __free(pstore_private) = NULL, *pos;
> >  	size_t			size = record->size + record->ecc_notice_size;
> >  
> >  	if (WARN_ON(!inode_is_locked(d_inode(root))))
> > @@ -356,7 +356,6 @@ int pstore_mkfile(struct dentry *root, struct pstore_record *record)
> >  			return -EEXIST;
> >  	}
> >  
> > -	rc = -ENOMEM;
> >  	inode = pstore_get_inode(root->d_sb);
> >  	if (!inode)
> >  		return -ENOMEM;
> > @@ -373,7 +372,7 @@ int pstore_mkfile(struct dentry *root, struct pstore_record *record)
> >  
> >  	dentry = d_alloc_name(root, name);
> >  	if (!dentry)
> > -		goto fail_private;
> > +		return -ENOMEM;
> >  
> >  	private->dentry = dentry;
> >  	private->record = record;
> > @@ -386,13 +385,9 @@ int pstore_mkfile(struct dentry *root, struct pstore_record *record)
> >  
> >  	d_add(dentry, no_free_ptr(inode));
> >  
> > -	list_add(&private->list, &records_list);
> > +	list_add(&(no_free_ptr(private))->list, &records_list);
> 
> That's really brittle.  It critically depends upon having no failure
> exits past the assignment to ->i_private; once you've done that,
> you have transferred the ownership of that thing to the inode
> (look at your ->evict_inode()).

I guess so, but it was already that way, and it isn't assignment to
i_private until everything else is "done", apart from adding to the
dentry and the pstore internal list.

> But you can't say
>         inode->i_private = no_free_ptr(private);
> since you are using private past that point.

True. How about this reordering:

        if (record->time.tv_sec)
                inode_set_mtime_to_ts(inode,
                                      inode_set_ctime_to_ts(inode, record->time));

        list_add(&private->list, &records_list);
        inode->i_private = no_free_ptr(private);

        d_add(dentry, no_free_ptr(inode));


But none of this gets into how you'd like to see the iput handled. If
you'd prefer, I can just define a pstore_iput handler and you don't have
to look at it at all. :)

-Kees

-- 
Kees Cook