Re: [RFR] Store tearing

[Arnd Bergmann]

On Mon, Oct 29, 2018 at 2:21 AM Paul E. McKenney <paulmck@xxxxxxxxxxxxx> wrote:
>
> On Mon, Oct 29, 2018 at 12:10:03AM +0100, Andrea Parri wrote:
> > Hopefully, with Paul's proper email address this time,
> >
> >   Andrea
> >
> > On Mon, Oct 29, 2018 at 12:06:27AM +0100, Andrea Parri wrote:
> > > Hi,
> > >
> > > memory-barriers.txt says:
> > >
> > >   [on "store tearing"]
> > >
> > >   "In fact, a recent bug (since fixed) caused GCC to incorrectly use
> > >    this optimization in a volatile store.".
> > >
> > > I was wondering if you could help me retrieve some reference/discussions
> > > about this?
>
> This was quite some time ago, but it involved a 32-bit volatile store
> of a constant such as 0x10001.  The machine in question had a narrow
> store-immediate instruction, so the compiler emitted  a pair of 16-bit
> store-immediate instructions.  This bug was fixed, though only after
> significant screaming and shouting.

A related issue I remember was on ARMv5 (an architecture without
unaligned access) where a function like )not sure if this specific
one triggers it, but something like it did)

struct my_registers {
     u32 a;
     u32 b;
     u32 c;
} __attribute__((packed));
#define __raw_writel(p, v) do { (volatile u32 __iomem *)(p) = (v); } while (0)
void my_write_a(struct my_registers __iomem *r, u32 val)
{
       __raw_writel(&r->a, val);
}

The above is undefined behavior because we cast from an unaligned
data type to a 32-bit aligned type, and gcc resolved this by turning the
intended 32-bit store into a set of 8 bit stores. We worked around this
by changing __raw_writel() into a inline assembly that always uses a
32-bit store.

      Arnd