Re: [PATCH] docs/memory-barriers.txt: volatile is not a barrier() substitute

From: Ard Biesheuvel
Date: Tue Feb 01 2022 - 04:32:29 EST


On Mon, 31 Jan 2022 at 23:53, Nick Desaulniers <ndesaulniers@xxxxxxxxxx> wrote:
>
> Add text to memory-barriers.txt and deprecated.rst to denote that
> volatile-qualifying an asm statement is not a substitute for either a
> compiler barrier (``barrier();``) or a clobber list.
>
> This way we can point to this in code that strengthens existing
> volatile-qualified asm statements to use a compiler barrier.
>
> Suggested-by: Kees Cook <keescook@xxxxxxxxxxxx>
> Signed-off-by: Nick Desaulniers <ndesaulniers@xxxxxxxxxx>
> ---
> Example: https://godbolt.org/z/8PW549zz9
>
> Documentation/memory-barriers.txt | 24 ++++++++++++++++++++++++
> Documentation/process/deprecated.rst | 17 +++++++++++++++++
> 2 files changed, 41 insertions(+)
>
> diff --git a/Documentation/memory-barriers.txt b/Documentation/memory-barriers.txt
> index b12df9137e1c..f3908c0812da 100644
> --- a/Documentation/memory-barriers.txt
> +++ b/Documentation/memory-barriers.txt
> @@ -1726,6 +1726,30 @@ of optimizations:
> respect the order in which the READ_ONCE()s and WRITE_ONCE()s occur,
> though the CPU of course need not do so.
>
> + (*) Similarly, the compiler is within its rights to reorder instructions

Similar to what? Was this intended to be the second bullet point
rather than the first?

> + around an asm statement so long as clobbers are not violated. For example,
> +
> + asm volatile ("");
> + flag = true;
> +
> + May be modified by the compiler to:
> +
> + flag = true;
> + asm volatile ("");
> +
> + Marking an asm statement as volatile is not a substitute for barrier(),
> + and is implicit for asm goto statements and asm statements that do not
> + have outputs (like the above example). Prefer either:
> +
> + asm ("":::"memory");
> + flag = true;
> +
> + Or:
> +
> + asm ("");
> + barrier();
> + flag = true;
> +

I would expect the memory clobber to only hazard against the
assignment of flag if it results in a store, but looking at your
Godbolt example, this appears to apply even if flag is kept in a
register.

Is that behavior documented/codified anywhere? Or are we relying on
compiler implementation details here?


> (*) The compiler is within its rights to invent stores to a variable,
> as in the following example:
>
> diff --git a/Documentation/process/deprecated.rst b/Documentation/process/deprecated.rst
> index 388cb19f5dbb..432816e2f79e 100644
> --- a/Documentation/process/deprecated.rst
> +++ b/Documentation/process/deprecated.rst
> @@ -329,3 +329,20 @@ struct_size() and flex_array_size() helpers::
> instance->count = count;
>
> memcpy(instance->items, source, flex_array_size(instance, items, instance->count));
> +
> +Volatile Qualified asm Statements
> +=================================
> +
> +According to `the GCC docs on inline asm
> +https://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html#Volatile`_:
> +
> + asm statements that have no output operands and asm goto statements,
> + are implicitly volatile.
> +
> +For many uses of asm statements, that means adding a volatile qualifier won't
> +hurt (making the implicit explicit), but it will not strengthen the semantics
> +for such cases where it would have been implied. Care should be taken not to
> +confuse ``volatile`` with the kernel's ``barrier()`` macro or an explicit
> +clobber list. See [memory-barriers]_ for more info on ``barrier()``.
> +
> +.. [memory-barriers] Documentation/memory-barriers.txt
> --
> 2.35.0.rc2.247.g8bbb082509-goog
>