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

[Sergey Senozhatsky]

On (10/10/19 09:40), Michal Hocko wrote:
[..]
> > > Considering that console.write is called from essentially arbitrary code
> > > path IIUC then all the locks used in this path should be pretty much
> > > tail locks or console internal ones without external dependencies.
> > 
> > That's a good expectation, but I guess it's not always the case.
> > 
> > One example might be NET console - net subsystem locks, net device
> > drivers locks, maybe even some MM locks (skb allocations?).
> 
> I am not familiar with the netconsole code TBH. If there is absolutely
> no way around that then we might have to bite a bullet and consider some
> of MM locks a land of no printk.

So this is what netconsole does (before we pass on udp to net device
driver code, which *may be* can do more allocations, I don't know):

write_msg()
 netpoll_send_udp()
  find_skb()
   alloc_skb(len, GFP_ATOMIC)
    kmem_cache_alloc_node()

You are the right person to ask this question to - how many MM
locks are involved when we do GFP_ATOMIC kmem_cache allocation?
Is there anything to be concerned about?

> I have already said that in this thread. I am mostly pushing back
> on "let's just go the simplest way" approach.

I understand this. And don't have objections. Merely trying to get
better understanding.

> > But even more "commonly used" consoles sometimes break that
> > expectation. E.g. 8250
> >
> > serial8250_console_write()
> >  serial8250_modem_status()
> >   wake_up_interruptible()
>
> By that expectation you mean they are using external locks or that they
> really _need_ to allocate.

Sort of both. netconsole does allocations and has external lock
dependencies. Some of serial console drivers have external lock
dependencies (but don't do allocations, as far as I'm concerned).

> Because if you are pointing to wake_up_interruptible and therefore the rq
> then this is a well known thing and I was under impression even documented
> but I can only see LOGLEVEL_SCHED that is arguably a very obscure way to
> document the fact.

That's right. All I was commenting on is that console.write() callbacks
have external lock dependencies.

	-ss