Re: [PATCH] mm/page_alloc: Add a reason for reserved pages in has_unmovable_pages()

[Michal Hocko]

On Fri 04-10-19 09:56:00, Qian Cai wrote:
> On Fri, 2019-10-04 at 15:38 +0200, Michal Hocko wrote:
> > On Fri 04-10-19 09:30:39, Qian Cai wrote:
> > > On Fri, 2019-10-04 at 15:07 +0200, Michal Hocko wrote:
> > > > On Fri 04-10-19 08:56:16, Qian Cai wrote:
> > > > [...]
> > > > > It might be a good time to rethink if it is really a good idea to dump_page()
> > > > > at all inside has_unmovable_pages(). As it is right now, it is a a potential
> > > > > deadlock between console vs memory offline. More details are in this thread,
> > > > > 
> > > > > https://lore.kernel.org/lkml/1568817579.5576.172.camel@xxxxxx/
> > > > 
> > > > Huh. That would imply we cannot do any printk from that path, no?
> > > 
> > > Yes, or use something like printk_deferred()
> > 
> > This is just insane. The hotplug code is in no way special wrt printk.
> > It is never called from the printk code AFAIK and thus there is no real
> > reason why this particular code should be any special. Not to mention
> > it calls printk indirectly from a code that is shared with other code
> > paths.
> 
> Basically, printk() while holding the zone_lock will be problematic as console
> is doing the opposite as it always needs to allocate some memory. Then, it will
> always find some way to form this chain,
> 
> console_lock -> * -> zone_lock.

So this is not as much a hotplug specific problem but zone->lock ->
printk -> alloc chain that is a problem, right? Who is doing an
allocation from this atomic context? I do not see any atomic allocation
in kernel/printk/printk.c.
-- 
Michal Hocko
SUSE Labs