Re: [v2 PATCH] mm: introduce arg_lock to protect arg_start|end and env_start|end in mm_struct

[Cyrill Gorcunov]

On Tue, Mar 27, 2018 at 06:10:09AM +0900, Tetsuo Handa wrote:
> On 2018/03/27 4:21, Cyrill Gorcunov wrote:
> > That said I think using read-lock here would be a bug.
> 
> If I understand correctly, the caller can't set both fields atomically, for
> prctl() does not receive both fields at one call.
> 
>   prctl(PR_SET_MM, PR_SET_MM_ARG_START xor PR_SET_MM_ARG_END xor PR_SET_MM_ENV_START xor PR_SET_MM_ENV_END, new value, 0, 0);
> 

True, but the key moment is that two/three/four system calls can
run simultaneously. And while previously they are ordered by "write",
with read lock they are completely unordered and this is really
worries me. To be fair I would prefer to drop this old per-field
interface completely. This per-field interface was rather an ugly
solution from my side.

> Then, I wonder whether reading arg_start|end and env_start|end atomically makes
> sense. Just retry reading if arg_start > env_end or env_start > env_end is fine?

Tetsuo, let me re-read this code tomorrow, maybe I miss something obvious.