Re: [PATCH] virtio-console: avoid DMA from vmalloc area

[Xianting Tian]

Arnd, thanks for your quick reply,

As we know put_chars() of virtio-console is registered to hvc framework.
I go throughed the code, actually there are totally three places that 
put_chars() is called in hvc driver,  but only 1 has issue which is 
fixed by commit c4baad5029.
So I think the scenario that the buf is from "ioremap(), kmap_atomic() , 
fixmap, loadable module" doesn't exist for virtio-console.
If there is something wrong about above description, please correct me, 
thanks.

Three places that put_chars() is called in hvc driver:
1， it is on stack buf,  it is not ok for dma
    hvc_console_print():
        char c[N_OUTBUF] __ALIGNED__;
        cons_ops[index]->put_chars(vtermnos[index], c, i);

2， just one byte, no issue for dma
    static void hvc_poll_put_char(struct tty_driver *driver, int line, 
char ch)
    {
        struct tty_struct *tty = driver->ttys[0];
        struct hvc_struct *hp = tty->driver_data;
        int n;

        do {
            n = hp->ops->put_chars(hp->vtermno, &ch, 1);
        } while (n <= 0);
    }
3,  hp->outbuf is allocated in hvc_alloc() via kzalloc(), no issue for dma
    static int hvc_push(struct hvc_struct *hp)
    {
        int n;

        n = hp->ops->put_chars(hp->vtermno, hp->outbuf, hp->n_outbuf);
        …
    }

在 2021/7/27 下午9:18, Arnd Bergmann 写道:
> On Tue, Jul 27, 2021 at 3:13 PM Xianting Tian
> <xianting.tian@xxxxxxxxxxxxxxxxx> wrote:
> > @@ -1127,13 +1128,18 @@ static int put_chars(u32 vtermno, const char *buf, int count)
> >          if (!port)
> >                  return -EPIPE;
> >
> > -       data = kmemdup(buf, count, GFP_ATOMIC);
> > -       if (!data)
> > -               return -ENOMEM;
> > +       if (is_vmalloc_addr(buf)) {
> > +               data = kmemdup(buf, count, GFP_ATOMIC);
> What about buffers in .data? If those are in a loadable module, I guess you have
> the same problem as with vmalloc() and vmap().
>
> is_vmalloc_or_module_addr() would take care of both, not sure if there are
> other examples that don't work. In theory it could be ioremap(), kmap_atomic()
> or fixmap as well, but those seem less likely to matter here.
>
>          Arnd