On Tue, Oct 01, 2024 at 09:02:02PM -0400, Mathieu Desnoyers wrote:
> Compiler CSE and SSA GVN optimizations can cause the address dependency
> of addresses returned by rcu_dereference to be lost when comparing those
> pointers with either constants or previously loaded pointers.
> 
> Introduce ptr_eq() to compare two addresses while preserving the address
> dependencies for later use of the address. It should be used when
> comparing an address returned by rcu_dereference().
> 
> This is needed to prevent the compiler CSE and SSA GVN optimizations
> from using @a (or @b) in places where the source refers to @b (or @a)
> based on the fact that after the comparison, the two are known to be
> equal, which does not preserve address dependencies and allows the
> following misordering speculations:
> 
> - If @b is a constant, the compiler can issue the loads which depend
>   on @a before loading @a.
> - If @b is a register populated by a prior load, weakly-ordered
>   CPUs can speculate loads which depend on @a before loading @a.
> 
> The same logic applies with @a and @b swapped.
> 
[...]
> +/*
> + * Compare two addresses while preserving the address dependencies for
> + * later use of the address. It should be used when comparing an address
> + * returned by rcu_dereference().
> + *
> + * This is needed to prevent the compiler CSE and SSA GVN optimizations
> + * from using @a (or @b) in places where the source refers to @b (or @a)
> + * based on the fact that after the comparison, the two are known to be
> + * equal, which does not preserve address dependencies and allows the
> + * following misordering speculations:
> + *
> + * - If @b is a constant, the compiler can issue the loads which depend
> + *   on @a before loading @a.
> + * - If @b is a register populated by a prior load, weakly-ordered
> + *   CPUs can speculate loads which depend on @a before loading @a.
> + *
> + * The same logic applies with @a and @b swapped.
> + *
> + * Return value: true if pointers are equal, false otherwise.
> + *
> + * The compiler barrier() is ineffective at fixing this issue. It does
> + * not prevent the compiler CSE from losing the address dependency:
> + *
> + * int fct_2_volatile_barriers(void)
> + * {
> + *     int *a, *b;
> + *
> + *     do {
> + *         a = READ_ONCE(p);
> + *         asm volatile ("" : : : "memory");
> + *         b = READ_ONCE(p);
> + *     } while (a != b);
> + *     asm volatile ("" : : : "memory");  <-- barrier()
> + *     return *b;
> + * }
> + *
> + * With gcc 14.2 (arm64):
> + *
> + * fct_2_volatile_barriers:
> + *         adrp    x0, .LANCHOR0
> + *         add     x0, x0, :lo12:.LANCHOR0
> + * .L2:
> + *         ldr     x1, [x0]  <-- x1 populated by first load.
> + *         ldr     x2, [x0]
> + *         cmp     x1, x2
> + *         bne     .L2
> + *         ldr     w0, [x1]  <-- x1 is used for access which should depend 
> on b.
> + *         ret
> + *

I could reproduce this in compiler explorer, but I'm curious what flags are
you using? For me it does a bunch of usage of the stack for temporary storage
(still incorrectly returns *a though as you pointed).

Interestingly, if I just move the comparison into an an __always_inline__
function like below, but without the optimizer hide stuff, gcc 14.2 on arm64
does generate the correct code:

static inline __attribute__((__always_inline__)) int ptr_eq(const volatile void 
*a, const volatile void *b)
{
    /* No OPTIMIZER_HIDE_VAR */
    return a == b;
}

volatile int *p = 0;

int fct_2_volatile_barriers()
{
    int *a, *b;

    do {
        a = READ_ONCE(p);
        asm volatile ("" : : : "memory");
        b = READ_ONCE(p);
    } while (!ptr_eq(a, b));
    asm volatile ("" : : : "memory");  // barrier()
    return *b;
}

But not sure if it fixes the speculation issue you referred to.

Putting back the OPTIMIZER_HIDE_VAR() then just seems to pass the a and b
stored on the stack through a washing machine:

        ldr     x0, [sp, 8]
        str     x0, [sp, 8]
        ldr     x0, [sp]
        str     x0, [sp]

And here I thought the "" in OPTIMIZER_HIDE_VAR was not supposed to generate
any code but I guess it is still a NOOP.

Anyway, as such this LGTM since whether OPTIMIZER_HIDE_VAR() used or not, it
does fix the problem.

Reviewed-by: Joel Fernandes (Google) <j...@joelfernandes.org>

thanks,

 - Joel


Reply via email to