Re: [PATCH] linux-user: Set CF_PARALLEL when mapping shared memory

[Laurent Vivier]

Le 12/06/2021 à 08:08, Richard Henderson a écrit :
> Signal the translator to use host atomic instructions for
> guest operations, insofar as it is possible.  This is the
> best we can do to allow the guest to interact atomically
> with other processes.
> 
> Resolves: https://gitlab.com/qemu-project/qemu/-/issues/121
> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
> ---
>  linux-user/mmap.c    | 14 ++++++++++++++
>  linux-user/syscall.c | 12 ++++++++++++
>  2 files changed, 26 insertions(+)
> 
> diff --git a/linux-user/mmap.c b/linux-user/mmap.c
> index 7e3b245036..0e103859fe 100644
> --- a/linux-user/mmap.c
> +++ b/linux-user/mmap.c
> @@ -451,6 +451,20 @@ abi_long target_mmap(abi_ulong start, abi_ulong len, int 
> target_prot,
>          goto fail;
>      }
>  
> +    /*
> +     * If we're mapping shared memory, ensure we generate code for parallel
> +     * execution and flush old translations.  This will work up to the level
> +     * supported by the host -- anything that requires EXCP_ATOMIC will not
> +     * be atomic with respect to an external process.
> +     */
> +    if (flags & MAP_SHARED) {
> +        CPUState *cpu = thread_cpu;
> +        if (!(cpu->tcg_cflags & CF_PARALLEL)) {
> +            cpu->tcg_cflags |= CF_PARALLEL;
> +            tb_flush(cpu);
> +        }
> +    }
> +
>      real_start = start & qemu_host_page_mask;
>      host_offset = offset & qemu_host_page_mask;
>  
> diff --git a/linux-user/syscall.c b/linux-user/syscall.c
> index 974dd46c9a..54037db8d6 100644
> --- a/linux-user/syscall.c
> +++ b/linux-user/syscall.c
> @@ -4603,6 +4603,7 @@ static inline abi_ulong target_shmlba(CPUArchState 
> *cpu_env)
>  static inline abi_ulong do_shmat(CPUArchState *cpu_env,
>                                   int shmid, abi_ulong shmaddr, int shmflg)
>  {
> +    CPUState *cpu = env_cpu(cpu_env);
>      abi_long raddr;
>      void *host_raddr;
>      struct shmid_ds shm_info;
> @@ -4633,6 +4634,17 @@ static inline abi_ulong do_shmat(CPUArchState *cpu_env,
>  
>      mmap_lock();
>  
> +    /*
> +     * We're mapping shared memory, so ensure we generate code for parallel
> +     * execution and flush old translations.  This will work up to the level
> +     * supported by the host -- anything that requires EXCP_ATOMIC will not
> +     * be atomic with respect to an external process.
> +     */
> +    if (!(cpu->tcg_cflags & CF_PARALLEL)) {
> +        cpu->tcg_cflags |= CF_PARALLEL;
> +        tb_flush(cpu);
> +    }
> +
>      if (shmaddr)
>          host_raddr = shmat(shmid, (void *)g2h_untagged(shmaddr), shmflg);
>      else {
> 

Applied to my linux-user-for-6.1 branch.

Thanks,
Laurent