Re: [PATCH] drivers/input: Remove all strcpy() uses in favor of strscpy()

[Kees Cook]

On Sun, Aug 01, 2021 at 10:19:07AM -0700, Dmitry Torokhov wrote:
> On Sun, Aug 01, 2021 at 09:44:33AM -0700, Kees Cook wrote:
> > On Sun, Aug 01, 2021 at 05:57:32PM +0200, Len Baker wrote:
> > > Hi,
> > > 
> > > On Sun, Aug 01, 2021 at 04:00:00PM +0100, Russell King (Oracle) wrote:
> > > > On Sun, Aug 01, 2021 at 04:43:16PM +0200, Len Baker wrote:
> > > > > strcpy() performs no bounds checking on the destination buffer. This
> > > > > could result in linear overflows beyond the end of the buffer, leading
> > > > > to all kinds of misbehaviors. The safe replacement is strscpy().
> > > > >
> > > > > Signed-off-by: Len Baker <len.baker@xxxxxxx>
> > > > > ---
> > > > > This is a task of the KSPP [1]
> > > > >
> > > > > [1] https://github.com/KSPP/linux/issues/88
> > > > >
> > > > >  drivers/input/keyboard/locomokbd.c | 2 +-
> > > > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > > > >
> > > > > diff --git a/drivers/input/keyboard/locomokbd.c b/drivers/input/keyboard/locomokbd.c
> > > > > index dae053596572..dbb3dc48df12 100644
> > > > > --- a/drivers/input/keyboard/locomokbd.c
> > > > > +++ b/drivers/input/keyboard/locomokbd.c
> > > > > @@ -254,7 +254,7 @@ static int locomokbd_probe(struct locomo_dev *dev)
> > > > >  	locomokbd->suspend_jiffies = jiffies;
> > > > >
> > > > >  	locomokbd->input = input_dev;
> > > > > -	strcpy(locomokbd->phys, "locomokbd/input0");
> > > > > +	strscpy(locomokbd->phys, "locomokbd/input0", sizeof(locomokbd->phys));
> > > >
> > > > So if the string doesn't fit, it's fine to silently truncate it?
> > > 
> > > I think it is better than overflow :)
> > > 
> > > > Rather than converting every single strcpy() in the kernel to
> > > > strscpy(), maybe there should be some consideration given to how the
> > > > issue of a strcpy() that overflows the buffer should be handled.
> > > > E.g. in the case of a known string such as the above, if it's longer
> > > > than the destination, should we find a way to make the compiler issue
> > > > a warning at compile time?
> > > 
> > > Good point. I am a kernel newbie and have no experience. So this
> > > question should be answered by some kernel hacker :) But I agree
> > > with your proposals.
> > > 
> > > Kees and folks: Any comments?
> > > 
> > > Note: Kees is asked the same question in [2]
> > > 
> > > [2] https://lore.kernel.org/lkml/20210731135957.GB1979@titan/
> > 
> > Hi!
> > 
> > Sorry for the delay at looking into this. It didn't use to be a problem
> > (there would always have been a compile-time warning generated for
> > known-too-small cases), but that appears to have regressed when,
> > ironically, strscpy() coverage was added. I've detailed it in the bug
> > report:
> > https://github.com/KSPP/linux/issues/88
> > 
> > So, bottom line: we need to fix the missing compile-time warnings for
> > strcpy() and strscpy() under CONFIG_FORTIFY_SOURCE=y.
> 
> Is it possible to have them warn always? Or that would be too many false
> positives?

There are actually no false positives right now (they were all fixed
before FORTIFY_SOURCE landed). Enabling it by default would likely mean
splitting compile-time checks from run-time checks... I'm kind of
already doing this in the recent memcpy() strictness series[1], so ...
maybe?

I think I'd like to land the memcpy() series first, then we can revisit
making it always warn.

> > In the past we'd tried to add a stracpy()[1] that would only work with
> > const string sources. Linus got angry[2] about API explosion, though,
> > so we're mostly faced with doing the strscpy() replacements.
> 
> I would like to have an API that would do compile-time checks and
> BUILD_BUG_ON() for a few places in input drivers where we copy constant
> strings. There is no reason to encumber the code with runtime checks,
> and bombing out on compile instead of truncating would be nice.

In theory, this is provided by CONFIG_FORTIFY_SOURCE, though a recent
change broke a specific instance. I've added tests for this now to the
memcpy() series, and will get it fixed in there too.

> > Another idea might be to have strcpy() do the "constant strings only"
> > thing, leaving strscpy() for the dynamic lengths.
> > 
> > One thing is clear: replacing strlcpy() with strscpy() is probably the
> > easiest and best first step to cleaning up the proliferation of str*()
> > functions.
> 
> OK, so the consensus is that we set this patch aside as it does not
> really fix any issues (the strcpy() destination is 32 bytes and is big
> enough to hold the string being copied)?

I think that's fair.

-Kees

[1] https://lore.kernel.org/lkml/20210727205855.411487-1-keescook@xxxxxxxxxxxx/

-- 
Kees Cook