Re: [PATCH for-8.0 07/29] accel/tcg: Honor atomicity of loads

From: Peter Maydell <peter.maydell@linaro.org>
To: Richard Henderson <richard.henderson@linaro.org>
Cc: qemu-devel@nongnu.org
Subject: Re: [PATCH for-8.0 07/29] accel/tcg: Honor atomicity of loads
Date: Tue, 22 Nov 2022 14:35:49 +0000	[thread overview]
Message-ID: <CAFEAcA_6PHdEPBFqyHC+z8Xz8FcJw1wgDu-R+ynm+Qv04WS1Xg@mail.gmail.com> (raw)
In-Reply-To: <20221118094754.242910-8-richard.henderson@linaro.org>

On Fri, 18 Nov 2022 at 09:50, Richard Henderson
<richard.henderson@linaro.org> wrote:
>
> Create ldst_atomicity.c.inc.
>
> Not required for user-only code loads, because we've ensured that
> the page is read-only before beginning to translate code.
>
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>

> +/**
> + * required_atomicity:
> + *
> + * Return the lg2 bytes of atomicity required by @memop for @p.
> + * If the operation must be split into two operations to be
> + * examined separately for atomicity, return -lg2.
> + */
> +static int required_atomicity(CPUArchState *env, uintptr_t p, MemOp memop)
> +{
> +    int atmax = memop & MO_ATMAX_MASK;
> +    int size = memop & MO_SIZE;
> +    unsigned tmp;
> +
> +    if (atmax == MO_ATMAX_SIZE) {
> +        atmax = size;
> +    } else {
> +        atmax >>= MO_ATMAX_SHIFT;
> +    }
> +
> +    switch (memop & MO_ATOM_MASK) {
> +    case MO_ATOM_IFALIGN:
> +        tmp = (1 << atmax) - 1;
> +        if (p & tmp) {
> +            return MO_8;
> +        }
> +        break;
> +    case MO_ATOM_NONE:
> +        return MO_8;
> +    case MO_ATOM_SUBALIGN:
> +        tmp = p & -p;
> +        if (tmp != 0 && tmp < atmax) {
> +            atmax = tmp;
> +        }
> +        break;
> +    case MO_ATOM_WITHIN16:
> +        tmp = p & 15;
> +        if (tmp + (1 << size) <= 16) {
> +            atmax = size;
> +        } else if (atmax < size && tmp + (1 << atmax) != 16) {
> +            /*
> +             * Paired load/store, where the pairs aren't aligned.
> +             * One of the two must still be handled atomically.
> +             */
> +            atmax = -atmax;
> +        }
> +        break;
> +    default:
> +        g_assert_not_reached();
> +    }
> +
> +    /*
> +     * Here we have the architectural atomicity of the operation.
> +     * However, when executing in a serial context, we need no extra
> +     * host atomicity in order to avoid racing.  This reduction
> +     * avoids looping with cpu_loop_exit_atomic.
> +     */
> +    if (cpu_in_serial_context(env_cpu(env))) {

Is it OK to use cpu_in_serial_context() here ? Even if
there's no other vCPU executing in parallel, there might
be device model code doing a memory write in the iothread,
I think.

> +        return MO_8;
> +    }
> +    return atmax;
> +}

thanks
-- PMM