Re: [PATCH -next] mm/hotplug: silence a lockdep splat with printk()

[Qian Cai <cai@lca.pw>]

> On Jan 14, 2020, at 6:53 PM, Andrew Morton <akpm@linux-foundation.org> wrote:
> 
>> On Jan 14, 2020, at 4:02 PM, Michal Hocko <mhocko@kernel.org> wrote:
>>> 
>>> Yeah, that was a long discussion with a lot of lockdep false positives.
>>> I believe I have made it clear that the console code shouldn't depend on
>>> memory allocation because that is just too fragile. If that is not
>>> possible for some reason then it has to be mentioned in the changelog.
>>> I really do not want us to add kludges to the MM code just because of
>>> printk deficiencies unless that is absolutely inevitable.
>> 
>> I don't know how to convince you, but both random number generator and
>> printk() maintainers agreed to get ride of printk() with zone->lock
>> held as you can see in the approved commit mentioned in this patch
>> description because it is a whac-a-mole to fix other places.  In other
>> word, the patch alone fixes quite a few false positives and potential
>> real deadlocks.  Maybe Andrew please has a look at this directly?
>> 
> 
> Well, a few things.
> 
> The changelog is quite poor.  It doesn't describe the problem (console
> drivers allocating memory) not does it describe the solution
> (deferring the dump_page() until after release of zone->lock).
> 
> So I changed it to this:
> 
> : Some console drivers can perform memory allocation at inappropriate times,
> : which can result in lockdep warnings (and presumably deadlocks) if printk
> : is called with zone->lock held.
> : 
> : By far the best fix is to reeducate those console drivers to not perform
> : these allocations, but this is proving difficult.

… but this is proving difficult because even if we fixed that directly, lockdep
Is still able to find an indirect dependency chain, for example [1]

CPU1: console_owner —> port_lock_key
CPU2: port_lock_key —> (&port->lock)->rlock
CPU3: (&port->lock)->rlock —> zone->lock

which will trigger a splat with

zone->lock —> console_owner

[1] https://lore.kernel.org/linux-mm/1570460350.5576.290.camel@lca.pw/

> : 
> : Another but poorer approach is to call printk_deferred() when holding
> : zone->lock, but memory offline will call dump_page() which needs to defer
> : after the lock.
> : 
> : So change has_unmovable_pages() so that it no longer calls dump_page()
> : itself - instead it passes the page's descripton (as a string) back to the
> : caller so that in the case of a has_unmovable_pages() failure, the caller
> : can call dump_page() after releasing zone->lock.
> : 
> : While at it, remove a similar but unnecessary debug printk() as well.
> 
> But I see a couple of other issues.
> 
>> @@ -8290,8 +8290,10 @@ bool has_unmovable_pages(struct zone *zo
>> 	return false;
>> unmovable:
>> 	WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
>> -	if (flags & REPORT_FAILURE)
>> -		dump_page(pfn_to_page(pfn + iter), reason);
>> +	if (flags & REPORT_FAILURE) {
>> +		page = pfn_to_page(pfn + iter);
> 
> This statement appears to be unnecessary.

dump_page() in set_migratetype_isolate() needs that “page”.

> 
>> +		strscpy(dump, reason, 64);
>> +	}
> 
> 
> Also, that whole `reason' thing in has_unmovable_pages() is just there
> to tell us whether it was an "unmovable page" or a "CMA page".  This
> doesn't seem terribly useful to me.  Also, I expect that the
> dump_page() output will permit the user to determine that it was a CMA
> page anyway.  If not, we can change dump_page() to add that info.
> 
> So how about we remove that whole `reason' thing and possibly enhance
> dump_page()?  The patch then becomes much simpler.

Sounds like a good idea. I’ll send a v2.