Re: [PATCH] fs: make d_path-like functions all have unsigned size

[Matthew Wilcox]

On Tue, Jul 27, 2021 at 02:50:19PM +0000, Al Viro wrote:
> On Tue, Jul 27, 2021 at 12:36:25PM +0200, Greg Kroah-Hartman wrote:
> > When running static analysis tools to find where signed values could
> > potentially wrap the family of d_path() functions turn out to trigger a
> > lot of mess.  In evaluating the code, all of these usages seem safe, but
> > pointer math is involved so if a negative number is ever somehow passed
> > into these functions, memory can be traversed backwards in ways not
> > intended.
> > 
> > Resolve all of the abuguity by just making "size" an unsigned value,
> > which takes the guesswork out of everything involved.
> 
> TBH, I'm not sure it's the right approach.  Huge argument passed to d_path()
> is a bad idea, no matter what.  Do you really want to have the damn thing
> try and fill 3Gb of buffer, all while holding rcu_read_lock() and a global
> spinlock or two?  Hell, s/3Gb/1Gb/ and it won't get any better...
> 
> 
> How about we do this instead:
> 
> d_path(const struct path *path, char *buf, int buflen)
> {
> 	if (unlikely((unsigned)buflen > 0x8000)) {
> 		buf += (unsigned)buflen - 0x8000;
> 		buflen = 0x8000;
> 	}
> 	as in mainline
> }
> 
> and take care of both issues?

umm ... what if someone passes in -ENOMEM as buflen?  Not saying we
have such a path right now, but I could imagine it happening.

	if (unlikely(buflen < 0))
		return ERR_PTR(buflen);
	if (unlikely(buflen > 0x8000)) {
		buf += buflen - 0x8000;
		buflen = 0x8000;
	}
	...