RE: [PATCH] printk: ringbuffer: Fix truncating buffer size min_t cast

[David Laight]

From: Petr Mladek
> Sent: 14 August 2023 13:56
> 
> On Mon 2023-08-14 10:42:26, David Laight wrote:
> > From: Kees Cook
> > > Sent: 11 August 2023 06:46
> > >
> > > If an output buffer size exceeded U16_MAX, the min_t(u16, ...) cast in
> > > copy_data() was causing writes to truncate. This manifested as output
> > > bytes being skipped, seen as %NUL bytes in pstore dumps when the available
> > > record size was larger than 65536. Fix the cast to no longer truncate
> > > the calculation.
> > >
> > ...
> > > diff --git a/kernel/printk/printk_ringbuffer.c b/kernel/printk/printk_ringbuffer.c
> > > index 2dc4d5a1f1ff..fde338606ce8 100644
> > > --- a/kernel/printk/printk_ringbuffer.c
> > > +++ b/kernel/printk/printk_ringbuffer.c
> > > @@ -1735,7 +1735,7 @@ static bool copy_data(struct prb_data_ring *data_ring,
> > >  	if (!buf || !buf_size)
> > >  		return true;
> > >
> > > -	data_size = min_t(u16, buf_size, len);
> > > +	data_size = min_t(unsigned int, buf_size, len);
> >
> > I'd noticed that during one of my test compiles while looking
> > at making min() less fussy.
> >
> > A better fix would be:
> > 	data_size = min(buf_size + 0u, len);
> 
> This looks like a magic to me. The types are:

Not quite the right magic though, needs to be 'len + 0u'.

> 
> 	unsigned int data_size;
> 	unsigned int buf_size;
> 	u16 len
> 
> I would naively expect that
> 
> 	data_size = min(buf_size, len);
> 
> would do the right job and expand @len to "unsigned int".
> 
> I do not remember why "min_t" was used. Was it an optimization?
> Did we miss the problem with casting "u32" down to "u16"?

The underlying problem is that (presumably) in order to stop
min(signed_a, unsigned_b) converting a negative value to a large
unsigned one (very nasty) min() contains (effectively) sizeof(&a == &b)
so barfs if the types differ at all.

I'm sure the intent was that the types would be fixed - in this case
chasing 'len' back all the way back and using 'unsigned int'.
(That probably generates better code as well.)

However everyone just uses min_t(type,a,b) if type is 32bit unsigned
they are mostly ok because the kernel only really deals in 'small'
unsigned values.
But, as in the case here, it is easy to pick a type that is too small.
Pretty much all the min_t() with u8/u16 are likely to be dubious.
I found an 'unsigned long' case in a filesystem where one value
was u64 - could be problematic for a large file on 32bit.
(The u64 definitely contained a 'file size' value.)

The patch set I proposed (see https://lore.kernel.org/lkml/01e3e09005e9434b8f558a893a47c053@xxxxxxxxxxxxxxxx/)
changes the basic test to (is_signed(a) == is_signed(b)) which will
never generate the 'nasty' conversion of -1 to 0xffffffffull.

Of course, it is never quite that simple :-)
Linus seems willing to accept min(unsigned_var, 20) but not
min(signed_var, 20u) - typically as min(signed_var, sizeof(type)).

...
> PS: I have already pushed the patch because it looked reasonable and
>     got testing. I have to admit that I am probably in a pre-vacation
>     hurry mode.

Don't worry it is now not any worse than the other 4500 min_t().
Much the same as the number of min().

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)