Re: [PATCH v2 1/4] vfs, fdtable: Add get_task_file helper

[Oleg Nesterov <oleg@redhat.com>]

On 12/09, Sargun Dhillon wrote:
>
> +struct file *get_task_file(struct task_struct *task, unsigned int fd)
> +{
> +	struct file *file = NULL;
> +
> +	task_lock(task);
> +	rcu_read_lock();
> +
> +	if (task->files) {
> +		file = fcheck_files(task->files, fd);
> +		if (file && !get_file_rcu(file))
> +			file = NULL;
> +	}

On second thought this is not exactly right, get_file_rcu() can fail if
get_task_file() races with dup2(), in this case we need to do fcheck_files()
again. And this is what __fget() already does, so may be the patch below
makes more sense?

I will leave this to other reviewers, but suddenly I recall that I have
already sent the patch which adds a similar helper a while ago.

See https://lore.kernel.org/lkml/20180915160423.GA31461@redhat.com/

In short, get_files_struct() should be avoided because it can race with
exec() and break POSIX locks which use ->fl_owner = files_struct.

Oleg.

--- x/fs/file.c
+++ x/fs/file.c
@@ -706,9 +706,9 @@ void do_close_on_exec(struct files_struc
 	spin_unlock(&files->file_lock);
 }
 
-static struct file *__fget(unsigned int fd, fmode_t mask, unsigned int refs)
+static struct file *__fget_files(struct files_struct *files, unsigned int fd,
+					fmode_t mask, unsigned int refs)
 {
-	struct files_struct *files = current->files;
 	struct file *file;
 
 	rcu_read_lock();
@@ -729,6 +729,23 @@ loop:
 	return file;
 }
 
+struct file *fget_task(struct task_struct *task, unsigned int fd)
+{
+	struct file *file;
+
+	task_lock(task);
+	if (task->files)
+		file = __fget_files(task->files, fd, 0, 1);
+	task_unlock(task);
+
+	return file;
+}
+
+static struct file *__fget(unsigned int fd, fmode_t mask, unsigned int refs)
+{
+	return __fget_files(current->files, fd, mask, refs);
+}
+
 struct file *fget_many(unsigned int fd, unsigned int refs)
 {
 	return __fget(fd, FMODE_PATH, refs);