Re: Regression with orderly_poweroff()

[Oleg Nesterov <oleg@redhat.com>]

On 03/12, Lucas De Marchi wrote:
>
> I have some pending patches on LKML to remove
> call_usermodehelper_fns() and export call_usermodehelper_{setup,exec}.
>  Doing this we can separate the allocation part using GFP_ATOMIC

But this is pointless. Contrary, we should change __orderly_poweroff()
to do not use GFP_ATOMIC for argv_split(), see below.

> --- a/kernel/sys.c
> +++ b/kernel/sys.c
> @@ -2194,7 +2194,8 @@ static int __orderly_poweroff(void)
>   "PATH=/sbin:/bin:/usr/sbin:/usr/bin",
>   NULL
>   };
> - int ret;
> + struct subprocess_info *info;
> + int ret = -ENOMEM;
> 
>   argv = argv_split(GFP_ATOMIC, poweroff_cmd, &argc);
>   if (argv == NULL) {
> @@ -2203,7 +2204,10 @@ static int __orderly_poweroff(void)
>   return -ENOMEM;
>   }
> 
> - ret = call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC);
> + info = call_usermodehelper_setup(argv[0], argv, envp, GFP_ATOMIC,
> + NULL, NULL, NULL);
> + if (info)
> + ret = call_usermodehelper_exec(info, UMH_WAIT_EXEC);

And how this can help? The real problem is not GFP_KERNEL.
call_usermodehelper_exec(UMH_WAIT_EXEC) will block.

Oleg.