On Mon, Mar 26, 2018 at 05:59:49PM -0400, Yang Shi wrote:
> > Say we've two syscalls running prctl_set_mm_map in parallel, and imagine
> > one have @start_brk = 20 @brk = 10 and second caller has @start_brk = 30
> > and @brk = 20. Since now the call is guarded by _read_ the both calls
> > unlocked and due to OO engine it may happen then when both finish
> > we have @start_brk = 30 and @brk = 10. In turn "write" semaphore
> > has been take to have consistent data on exit, either you have [20;10]
> > or [30;20] assigned not something mixed.
> >
> > That said I think using read-lock here would be a bug.
>
> Yes it sounds so. However, it was down_read before
> ddf1d398e517e660207e2c807f76a90df543a217 ("prctl: take mmap sem for writing
> to protect against others"). And, that commit is for fixing the concurrent
> writing to arg_* and env_*. I just checked that commit, but omitted the brk
> part. The potential issue mentioned by you should exist before that commit,
> but might be just not discovered or very rare to hit.
>
> I will change it back to down_write.
down_read before was a bug ;) And it was not discovered earlier simply
because not that many users of this interface exist, namely only criu
as far as I know by now.
