Re: [PATCH v2] mm/page_isolation: fix a deadlock with printk()

From: Petr Mladek
Date: Mon Oct 07 2019 - 10:11:04 EST


On Mon 2019-10-07 09:07:02, Qian Cai wrote:
> On Mon, 2019-10-07 at 14:43 +0200, Michal Hocko wrote:
> > On Mon 07-10-19 08:11:44, Qian Cai wrote:
> > > On Mon, 2019-10-07 at 13:37 +0200, Michal Hocko wrote:
> > > > On Mon 07-10-19 07:04:00, Qian Cai wrote:
> > > > >
> > > > >
> > > > > > On Oct 7, 2019, at 4:07 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > > > > >
> > > > > > I do not think that removing the printk is the right long term solution.
> > > > > > While I do agree that removing the debugging printk __offline_isolated_pages
> > > > > > does make sense because it is essentially of a very limited use, this
> > > > > > doesn't really solve the underlying problem. There are likely other
> > > > > > printks from zone->lock. It would be much more saner to actually
> > > > > > disallow consoles to allocate any memory while printk is called from an
> > > > > > atomic context.
> > > > >
> > > > > No, there is only a handful of places called printk() from
> > > > > zone->lock. It is normal that the callers will quietly process
> > > > > âstruct zoneâ modification in a short section with zone->lock
> > > > > held.
> > > >
> > > > It is extremely error prone to have any zone->lock vs. printk
> > > > dependency. I do not want to play an endless whack a mole.
> > > >
> > > > > No, it is not about âallocate any memory while printk is called from an
> > > > > atomic contextâ. It is opposite lock chain from different processors which has the same effect. For example,
> > > > >
> > > > > CPU0: CPU1: CPU2:
> > > > > console_owner
> > > > > sclp_lock
> > > > > sclp_lock zone_lock
> > > > > zone_lock
> > > > > console_owner
> > > >
> > > > Why would sclp_lock ever take a zone->lock (apart from an allocation).
> > > > So really if sclp_lock is a lock that might be taken from many contexts
> > > > and generate very subtle lock dependencies then it should better be
> > > > really careful what it is calling into.
> > > >
> > > > In other words you are trying to fix a wrong end of the problem. Fix the
> > > > console to not allocate or depend on MM by other means.
> > >
> > > It looks there are way too many places that could generate those indirect lock
> > > chains that are hard to eliminate them all. Here is anther example, where it
> > > has,
> >
> > Yeah and I strongly suspect they are consoles which are broken and need
> > to be fixed rathert than the problem papered over.
> >
> > I do realize how tempting it is to remove all printks from the
> > zone->lock but do realize that as soon as the allocator starts using any
> > other locks then we are back to square one and the problem is there
> > again. We would have to drop _all_ printks from any locked section in
> > the allocator and I do not think this is viable.
> >
> > Really, the only way forward is to make these consoles be more careful
> > of external dependencies.
>
> Even with the new printk() Petr proposed. There is no guarantee it will fix it
> properly. It looks like just reduce the chance of this kind of deadlocks as it
> may or may not callÂwake_up_klogd() inÂvprintk_emit() depends on timing.

The chain below is wrong:

> zone->lock
> printk_deferred()
> vprintk_emit()
> wake_up_klogd()

wake_up_klogd() calls irq_work_queue(). It queues the work for
an interrupt handler and triggers the interrupt.

> wake_up_klogd_work_func()
> console_unlock()

The work is done in the interrupt context. The interrupt could
never be handled under zone->lock.

So, printk_deferred() would help. But I do not think that it is
really needed. I am going to answer the original mail with
all the full lockdep report.

Best Regards,
Petr