Re: [PATCH v2] p54: Fix an error handling path in p54spi_probe()

[Christophe JAILLET]

Le 16/06/2022 à 17:19, Dan Carpenter a écrit :
> On Thu, Jun 16, 2022 at 03:13:26PM +0200, Christian Lamparter wrote:
> > On 16/06/2022 12:36, Dan Carpenter wrote:
> > > > > If it deserves a v3 to axe some lines of code, I can do it but, as said
> > > > > previously,
> > > > > v1 is for me the cleaner and more future proof.
> > > >
> > > > Gee, that last sentence about "future proof" is daring.
> > >
> > > The future is vast and unknowable but one thing which is pretty likely
> > > is that Christophe's patch will introduce a static checker warning.  We
> > > really would have expected a to find a release_firmware() in the place
> > > where it was in the original code.  There is a comment there now so no
> > > one is going to re-add the release_firmware() but that's been an issue
> > > in the past.
> > >
> > > I'm sort of surprised that it wasn't a static checker warning already.
> > > Anyway, I'll add this to Smatch check_unwind.c
> > >
> > > +         { "request_firmware", ALLOC, 0, "*$", &int_zero, &int_zero},
> > > +         { "release_firmware", RELEASE, 0, "$"},
> >
> > hmm? I don't follow you there. Why should there be a warning "now"?
> > (I assume you mean with v2 but not with v1?).
>
> Yep.  Generally, static checkers assume that functions clean up after
> themselves on error paths so there would be a warning in
> p54spi_request_firmware().  This is the easiest kind of static analysis
> to implement and it's the way most kernel error handling is written.
>
> > If it's because the static
> > checker can't look beyond the function scope then this would be bad news
> > since on the "success" path the firmware will stick around until
> > p54spi_remove().
>
> Presumably Christophe found this bug with static analysis already but

True, I use a coccinelle script that looks at functions called in 
.remove() functions that are not called in what looks like an error 
handling path in the corresponding probe.

> my guess is that it has a lot of false positives?

This is SOOOO true !
The output is 23k LoC, mostly false positive!

In fact I only checks the diff between the outputs of my coccinelle 
script from time to time.

Looking at only the diff, most of the false positives get ignored and I 
manage to spot ~5-10 issues of this kind in each dev cycle in new code.

CJ

> Eventually the leak in the probe function would be found with static
> analysis as well.  The truth is that there are a lot of leaks so I'm
> already a bit overwhelmed fixing the ones that I know about.
>
> It would be fairly simple to make a high quality resource leak checker
> which is specific to probe functions.  But the thing is that leaks in
> probe functions are not really exploitable.  Also some devices are
> needed for the system to boot so often the devs don't care about about
> cleaning up...  My motivation is low.
>
> regards,
> dan carpenter