Re: [PATCH] mm, memory_failure: only send BUS_MCEERR_AO to early-kill process

["HORIGUCHI NAOYA(堀口 直也)" <naoya.horiguchi@nec.com>]

On Thu, May 28, 2020 at 02:50:09PM +0800, wetp wrote:
> 
> On 2020/5/28 上午10:22, HORIGUCHI NAOYA(堀口 直也) wrote:
> > Hi Zhang,
> > 
> > Sorry for my late response.
> > 
> > On Tue, May 26, 2020 at 03:06:41PM +0800, Wetp Zhang wrote:
> > > From: Zhang Yi <wetpzy@gmail.com>
> > > 
> > > If a process don't need early-kill, it may not care the BUS_MCEERR_AO.
> > > Let the process to be killed when it really access the corrupted memory.
> > > 
> > > Signed-off-by: Zhang Yi <wetpzy@gmail.com>
> > Thank you for pointing this. This looks to me a bug (per-process flag
> > is ignored when system-wide flag is set).
> 
> The flag is not problem for me.
> 
> In my case, two processes share memory with no any flag setting, both will
> be killed when only one
> 
> access the fail memory.

Thanks, now your problem seems clearer.

It seems that this happens because in "Action Required" case kill_proc()
takes the first branch for current process, while it takes the else branch
for other affected processes:

    static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags)
    {
            ...
    
            if ((flags & MF_ACTION_REQUIRED) && t->mm == current->mm) {
                    ret = force_sig_mceerr(BUS_MCEERR_AR, (void __user *)tk->addr,
                                           addr_lsb);
            } else {
                    /*
                     * Don't use force here, it's convenient if the signal
                     * can be temporarily blocked.
                     * 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? */
            }

Sending SIGBUS with BUS_MCEERR_AO for action optional error is strange, so
maybe this logic should be like this:


            if (flags & MF_ACTION_REQUIRED) {
                    if (t->mm == current->mm)
                            ret = force_sig_mceerr(BUS_MCEERR_AR, (void __user *)tk->addr,
                                           addr_lsb);
                    /* send no signal to non-current processes */
            } else {
                    /*
                     * Don't use force here, it's convenient if the signal
                     * can be temporarily blocked.
                     * 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? */
            }

> 
> > > ---
> > >   mm/memory-failure.c | 7 ++++---
> > >   1 file changed, 4 insertions(+), 3 deletions(-)
> > > 
> > > diff --git a/mm/memory-failure.c b/mm/memory-failure.c
> > > index a96364be8ab4..2db13d48865c 100644
> > > --- a/mm/memory-failure.c
> > > +++ b/mm/memory-failure.c
> > > @@ -210,7 +210,7 @@ static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags)
> > >   {
> > >   	struct task_struct *t = tk->tsk;
> > >   	short addr_lsb = tk->size_shift;
> > > -	int ret;
> > > +	int ret = 0;
> > > 
> > >   	pr_err("Memory failure: %#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n",
> > >   		pfn, t->comm, t->pid);
> > > @@ -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,
Naoya Horiguchi