On Fri, May 29, 2020 at 01:56:25PM +0800, wetp wrote:
> On 2020/5/29 上午10:12, HORIGUCHI NAOYA(堀口 直也) wrote:
...
> > > > > @@ -225,8 +225,9 @@ static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags)
> > > > > * This could cause a loop when the user sets SIGBUS
> > > > > * to SIG_IGN, but hopefully no one will do that?
> > > > > */
> > > > > - ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr,
> > > > > - addr_lsb, t); /* synchronous? */
> > > > > + if ((t->flags & PF_MCE_PROCESS) && (t->flags & PF_MCE_EARLY))
> > > > > + ret = send_sig_mceerr(BUS_MCEERR_AO,
> > > > > + (void __user *)tk->addr, addr_lsb, t);
> > > > kill_proc() could be called only for processes that are selected by
> > > > collect_procs() with task_early_kill(). So I think that we should fix
> > > > task_early_kill(), maybe by reordering sysctl_memory_failure_early_kill
> > > > check and find_early_kill_thread() check.
> > > >
> > > > static struct task_struct *task_early_kill(struct task_struct *tsk,
> > > > int force_early)
> > > > {
> > > > struct task_struct *t;
> > > > if (!tsk->mm)
> > > > return NULL;
> > > > if (force_early)
> > > > return tsk;
> > > The force_early is rely the flag MF_ACTION_REQUIRED, so it is always true
> > > when MCE occurs.
> > >
> > > This leads always sending SIGBUS to processes even if those are not current
> > > or no flag setting.
> > >
> > > I think it could keep the non-current processes which has no flag setting
> > > running.
> > >
> > >
> > > Besides, base on your recommendation I reorder the force_early check and
> > > find_early_kill_thread()
> > >
> > > check, to send the signal to the right thread.
> > Sorry, my previous comment around task_early_kill() is for a separate problem,
> > so I'll try some fix on this later.
> Thanks.
>
> Should me send the patch V2 for my problem alone? Or you will fix it with
> task_early_kill() together ？
You can send v2 if you like. Please note that patch description should be
descriptive enough for future reference.
I'm working on related testing now. I guess that action optional cases are
not tested well enough from current process context because there's no useful
injector.
Thanks,
Naoya Horiguchi
