On Fri, Aug 21, 2020 at 10:59:19AM +0200, Bartosz Golaszewski wrote:
> On Fri, Aug 21, 2020 at 10:35 AM Andy Shevchenko
> <andriy.shevchenko@xxxxxxxxxxxxxxx> wrote:
> > On Thu, Aug 20, 2020 at 08:51:08PM +0200, Bartosz Golaszewski wrote:
> > > From: Bartosz Golaszewski <bgolaszewski@xxxxxxxxxxxx>
...
> > > +static struct devres *to_devres(void *data)
> > > +{
> > > + return data - ALIGN(sizeof(struct devres), ARCH_KMALLOC_MINALIGN);
> > > +}
> > > +
> > > +static size_t devres_data_size(size_t total_size)
> > > +{
> > > + return total_size - ALIGN(sizeof(struct devres), ARCH_KMALLOC_MINALIGN);
> > > +}
> >
> > I'm fine with above, but here is a side note, perhaps
> >
> > offsetof(struct devres, data)
> >
> > will be more practical (no duplication of alignment and hence slightly better
> > maintenance)? (Note, I didn't check if it provides the correct result)
> >
>
> Hi Andy,
>
> The data pointer in struct devres is defined as:
>
> u8 __aligned(ARCH_KMALLOC_MINALIGN) data[];
>
> And this value (assigned the value of ARCH_DMA_MINALIGN) varies from
> one arch to another. I wasn't really sure if offsetof() would work for
> every case so I went with something very explicit.
I have checked with a small program simulating to_devres() with your variant,
offsetof() and container_of().
The result is this: if MINALIGN < sizeof(long) and since struct is unpacked the
offsetof(), and thus container_of(), gives correct result, while ALIGN()
approach mistakenly moves pointer too back.
Test results
~~~~~~~~~~~~
sizeof(devres), ALIGN(data), resulting 3 pointers against NULL followed by
actual pointer of allocated struct and 3 pointers against it.
% for i in 1 2 4; do gcc -O2 -Wall -DSZ=$i -o test test.c && ./test | head -n1; done
szof: 24 a: 24 0xffffffffffffffe8 0xffffffffffffffef 0xffffffffffffffef, 0x55a3aa91e2a0: 0x55a3aa91e299 0x55a3aa91e2a0 0x55a3aa91e2a0
szof: 24 a: 24 0xffffffffffffffe8 0xffffffffffffffee 0xffffffffffffffee, 0x563d7b88b2a0: 0x563d7b88b29a 0x563d7b88b2a0 0x563d7b88b2a0
szof: 24 a: 24 0xffffffffffffffe8 0xffffffffffffffec 0xffffffffffffffec, 0x557d08cf82a0: 0x557d08cf829c 0x557d08cf82a0 0x557d08cf82a0
% for i in 1 2 4; do gcc -m32 -O2 -Wall -DSZ=$i -o test test.c && ./test | head -n1; done
szof: 12 a: 12 0xfffffff4 0xfffffff7 0xfffffff7, 0x584301a0: 0x5843019d 0x584301a0 0x584301a0
szof: 12 a: 12 0xfffffff4 0xfffffff6 0xfffffff6, 0x57bd61a0: 0x57bd619e 0x57bd61a0 0x57bd61a0
szof: 12 a: 12 0xfffffff4 0xfffffff4 0xfffffff4, 0x56b491a0: 0x56b491a0 0x56b491a0 0x56b491a0
I think you need to change this to use container_of() and offsetof().
> > Another side note: do we have existing users of these helpers?
>
> Which ones? Because I assume you're not referring to the ones I'm
> adding in this patch. :)
Opposite, the ones you are introduced here. Meaning that we might convert
existing user(s) in the separate change(s) later on.
--
With Best Regards,
Andy Shevchenko
