Re: [PATCH -next] exec: Remove redundant check in do_open_execat/uselib

[Zhihao Cheng]

在 2022/5/19 3:17, Kees Cook 写道:

> > > WARNON(path_noexec(&file->f_path)) // path_noexec() checks fail
>
> Did you encounter this in the real world?
I found the problem by running fuzz test.(syzkaller)

Here is a brief reproducer.
1. Apply diff
2. Complie and run repo.c
diff
diff --git a/fs/exec.c b/fs/exec.c
index e3e55d5e0be1..388d38b87e9a 100644
--- a/fs/exec.c
+++ b/fs/exec.c
@@ -897,6 +897,7 @@ EXPORT_SYMBOL(transfer_args_to_stack);

 #endif /* CONFIG_MMU */

+#include <linux/delay.h>
 static struct file *do_open_execat(int fd, struct filename *name, int 
flags)
 {
 	struct file *file;
@@ -925,9 +926,15 @@ static struct file *do_open_execat(int fd, struct 
filename *name, int flags)
 	 * and check again at the very end too.
 	 */
 	err = -EACCES;
+	if (!strcmp(file->f_path.dentry->d_iname, "my_bin")) {
+		pr_err("wait ...\n");
+		msleep(3000);
+	}
 	if (WARN_ON_ONCE(!S_ISREG(file_inode(file)->i_mode) ||
-			 path_noexec(&file->f_path)))
+			 path_noexec(&file->f_path))) {
+		pr_err("exec %pd %d %d %s\n", file->f_path.dentry, 
file->f_path.mnt->mnt_flags & MNT_NOEXEC, 
file->f_path.mnt->mnt_sb->s_iflags & SB_I_NOEXEC, 
file->f_path.mnt->mnt_sb->s_type->name);
 		goto exit;
+	}

 	err = deny_write_access(file);
 	if (err)

repo.c
int main(void)
{
	int ret;

	system("umount temp 2>&1 > /dev/null");
	system("mount -t tmpfs none temp");
	system("echo 12312 > temp/my_bin && chmod +x temp/my_bin");
	ret = fork();
	if (ret < 0) {
		perror("fork fail");
		return 0;
	}
	if (ret == 0) {
		system("mount -oremount,noexec temp");
		exit(0);
	} else {
		execve("/root/temp/my_bin", NULL, 0);
		//syscall(__NR_uselib, "/root/temp/my_bin");
	}
	return 0;
}
> > You're saying this is a race condition?  A concurrent remount causes
> > this warning?
>
> It seems not an unreasonable thing to warn about. Perhaps since it's
> technically reachable from userspace, it could be downgraded to
> pr_warn(), but I certainly don't want to remove the checks.


> I'd like to leave this as-is, since we _do_ want to find the cases where
> we're about to allow an exec and a very important security check was NOT
> handled.
> I think removing redundant checking is okay,

do_open_execat/uselib has initialized the acc_mode and open_flag for 
exec file, the check is equivalent to check in may_open().

Remount(noexec) operations can alos happen after the latest check, 
double check has no means for the concurrent situation.

The MNT_NOEXEC flag only affects the open operation, it won't cause any 
problems that an opened bin file is executing in a non-exec mounted 
filesystem.