Re: [PATCH v5 8/9] PCI: Define scoped based management functions

[Dan Williams <dan.j.williams@xxxxxxxxx>]

Ard Biesheuvel wrote:
> On Thu, 4 Jan 2024 at 08:02, Lukas Wunner <lukas@xxxxxxxxx> wrote:
> >
> > On Wed, Jan 03, 2024 at 10:43:40PM -0800, Dan Williams wrote:
> > > Lukas Wunner wrote:
> > > > On Wed, Dec 20, 2023 at 04:17:35PM -0800, Ira Weiny wrote:
> > > > > --- a/include/linux/pci.h
> > > > > +++ b/include/linux/pci.h
> > > > > @@ -1170,6 +1170,7 @@ int pci_get_interrupt_pin(struct pci_dev *dev, struct pci_dev **bridge);
> > > > >  u8 pci_common_swizzle(struct pci_dev *dev, u8 *pinp);
> > > > >  struct pci_dev *pci_dev_get(struct pci_dev *dev);
> > > > >  void pci_dev_put(struct pci_dev *dev);
> > > > > +DEFINE_FREE(pci_dev_put, struct pci_dev *, if (_T) pci_dev_put(_T))
> > > >
> > > > pci_dev_put() already performs a NULL pointer check internally.
> > > > Why duplicate it here?
> > >
> > > Greg asked the same for the introduction of __free(kvfree), and Peter
> > > clarified:
> > >
> > > http://lore.kernel.org/r/20230814161731.GN776869@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
> > >
> > > Essentially, that check is more for build-time than runtime because when
> > > the macro is expanded the compiler can notice scenarios where @pdev is
> > > set to NULL (likely by no_free_ptr()) and skip the call to pci_dev_put()
> > > altogether. pci_dev_put() also happens to be out-of-line, so saving a
> > > call when @pdev is NULL a small win in that respect as well.
> >
> > Doubtful whether that's correct.  The kernel is compiled with
> > -fno-delete-null-pointer-checks since commit a3ca86aea507
> > ("Add '-fno-delete-null-pointer-checks' to gcc CFLAGS").
> >
> > So these NULL pointer checks are generally not optimized away.
> >
> > I've just responded to the discussion you've linked above:
> > https://lore.kernel.org/all/20240104065744.GA6055@xxxxxxxxx/
> >
> 
> AIUI, Peter is referring to constant propagation of compile time
> constant pointers here, not pointer variables where the NULL check is
> elided if the variable has already been dereferenced.
> 

No, it is for auto (on stack) pointer variables. Consider this sequence:

	struct pci_dev *pdev __free(pci_dev_put) = pci_get_domain_bus_and_slot(...);

	if (!pdev)
		return NULL;

	if (!check_pdev(pdev))
		return NULL;

	return no_free_ptr(pdev);

...that expands at compile time to a first pass of:

	struct pci_dev *pdev = pci_get_domain_bus_and_slot(...);

	if (!pdev) {
		if (pdev)
			pci_dev_put(pdev);
		return NULL;
	}

	if (!check_pdev(pdev)) {
		if (pdev)
			pci_dev_put(pdev);
		return NULL;
	}

	struct pci_dev *tmp = pdev;
	pdev = NULL;
	if (pdev)
		pci_dev_put(pdev);
	return tmp;

...the compiler can then optimize this on a second pass to:

	if (!pdev)
		return NULL;

	if (!check_pdev(pdev)) {
		pci_dev_put(pdev);
		return NULL;
	}

	return pdev;

...if the NULL check is dropped from DEFINE_FREE(pci_dev_put...) then
this becomes unoptimizable by the compiler without
link-time-optimization (LTO) to see that pci_dev_put() has an internal
NULL check:

	struct pci_dev *pdev = pci_get_domain_bus_and_slot(...);

	if (!pdev) {
		pci_dev_put(pdev);
		return NULL;
	}

	if (!check_pdev(pdev)) {
		pci_dev_put(pdev);
		return NULL;
	}

	struct pci_dev *tmp = pdev;
	pdev = NULL;
	pci_dev_put(pdev);
	return tmp;

Now, if pci_dev_put() would become a static inline the compiler could
again do the optimization, but it is otherwise free (post compiler
optimization) to keep a conditional in these DEFINE_FREE() instances and
not worry about whether the actual free routine is inline, out-of-line,
or has its own NULL check.