Re: [PATCH 16/36] tty/vt: consolemap: check put_user() in con_get_unimap()

[Jiri Slaby]

On 08. 06. 22, 10:11, Jiri Slaby wrote:
> On 08. 06. 22, 10:02, David Laight wrote:
> > From: Jiri Slaby
> > > Sent: 07 June 2022 11:49
> > >
> > > Only the return value of copy_to_user() is checked in con_get_unimap().
> > > Do the same for put_user() of the count too.
> > >
> > > Signed-off-by: Jiri Slaby <jslaby@xxxxxxx>
> > > ---
> > >   drivers/tty/vt/consolemap.c | 3 ++-
> > >   1 file changed, 2 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/drivers/tty/vt/consolemap.c b/drivers/tty/vt/consolemap.c
> > > index 831450f2bfd1..92b5dddb00d9 100644
> > > --- a/drivers/tty/vt/consolemap.c
> > > +++ b/drivers/tty/vt/consolemap.c
> > > @@ -813,7 +813,8 @@ int con_get_unimap(struct vc_data *vc, ushort ct, 
> > > ushort __user *uct,
> > >       console_unlock();
> > >       if (copy_to_user(list, unilist, min(ect, ct) * sizeof(*unilist)))
> > >           ret = -EFAULT;
> > > -    put_user(ect, uct);
> > > +    if (put_user(ect, uct))
> > > +        ret = -EFAULT;
> > >       kvfree(unilist);
> > >       return ret ? ret : (ect <= ct) ? 0 : -ENOMEM;
> > >   }
> >
> > How is the user expected to check the result of this code?
> >
> > AFAICT -ENOMEM is returned if either kmalloc() fails or
> > the user buffer is too short?
> > Looks pretty hard to detect which.
>
> Agreed. The code is far from perfect. We might try to return ENOSPC and 
> watch what breaks.

brltty and kbd (see below) would break at least:
https://sources.debian.org/src/brltty/6.4-6/Drivers/Screen/Linux/screen.c/#L875

brltty apparently relies exactly on ENOMEM, increases buffer if that 
error is returned, and retries.

So I don't think we can change that ENOMEM to anything else.

> > I've not looked at the effect of all the patches, but setting
> > 'ret = -ENOMEM' and breaking the loop when the array is too
> > small would simplify things.

That would break kbd for example:
https://sources.debian.org/src/kbd/2.3.0-3/src/libkfont/kdmapop.c/?hl=154#L159

GIO_UNIMAP is called with zero count to actually find out the count...

So apart from the original patch which checks the return value of 
put_user, we cannot do anything else. (Except decoupling the "?:" to 
make it more readable.)

thanks,
--
js
suse labs