Re: blk, mm: lockdep irq lock inversion in linux-next

[Hugh Dickins]

On Mon, 17 Sep 2012, Andrew Morton wrote:
> On Sat, 15 Sep 2012 15:50:07 +0200
> Sasha Levin <levinsasha928@xxxxxxxxx> wrote:
> 
> > Hi all,
> > 
> > While fuzzing with trinity within a KVM tools guest on a linux-next kernel, I
> > got the lockdep warning at the bottom of this mail.
> > 
> > I've tried figuring out where it was introduced, but haven't found any sign that
> > any of the code in that area changed recently, so I'm probably missing something...
> > 
> > 
> > [ 157.966399] =========================================================
> > [ 157.968523] [ INFO: possible irq lock inversion dependency detected ]
> > [ 157.970029] 3.6.0-rc5-next-20120914-sasha-00001-g802bf6c-dirty #340 Tainted: G W
> > [ 157.970029] ---------------------------------------------------------
> > [ 157.970029] trinity-child38/6642 just changed the state of lock:
> > [ 157.970029] (&(&mapping->tree_lock)->rlock){+.+...}, at: [<ffffffff8120cafc>]
> > invalidate_inode_pages2_range+0x20c/0x3c0
> > [ 157.970029] but this lock was taken by another, SOFTIRQ-safe lock in the past:
> > [ 157.970029] (&(&new->queue_lock)->rlock){..-...}
> > 
> > [snippage]
> 
> gack, what a mess.  Thanks for the report.  AFAICT, what has happened is:
> 
> invalidate_complete_page2()
> ->spin_lock_irq(&mapping->tree_lock)
> ->clear_page_mlock()
>   __clear_page_mlock()
>   ->isolate_lru_page()
>     ->spin_lock_irq(&zone->lru_lock)
>     ->spin_unlock_irq(&zone->lru_lock)
> 
> whoops.  isolate_lru_page() just enabled local interrupts while we're
> holding ->tree_lock, which is supposed to be an irq-save lock.  And in
> a rather obscure way, lockdep caught it.

Congratulations on deciphering the lockdep report, I soon gave up.

But it looks like a bigger problem than your patch addresses:
both filemap.c and rmap.c document tree_lock as nesting within
lru_lock; and although it's possible that time has changed that,
I doubt it.

I think invalidate_complete_page2() is simply wrong to be calling
clear_page_mlock() while holding mapping->tree_lock (other callsites
avoid doing so).  Maybe it should do a preliminary PageDirty test,
then clear_page_mlock(), then take mapping->tree_lock, then repeat
PageDirty test, without worrying about the odd case when it might
clear mlock but then decide to back off the page.

Oh, hold on, that reminds me: a few months ago I was putting together
a tidy-up patch near there, and it seemed to me inappropriate to be
clearing mlock down in truncate/invalidate, that belongs better to
when unmapping the page, doesn't it?

I'll look that out and try to finish it off.

Hugh

> 
> Problem is, I cannot find any recent change which might have triggered
> this.
> 
> I don't know how repeatable this is for you (not very at all, I
> suspect).  This?
> 
> 
> From: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> Subject: mm: isolate_lru_page(): don't enable local interrupts
> 
> isolate_lru_page() is called with local interrupts disabled, via
> 
> invalidate_complete_page2()
> ->spin_lock_irq(&mapping->tree_lock)
> ->clear_page_mlock()
>   __clear_page_mlock()
>   ->isolate_lru_page()
> 
> so it should not unconditionally enable local interrupts.
> 
> Sasha hit a lockdep warning when running Trinity as a result of this.
> 
> Reported-by: Sasha Levin <levinsasha928@xxxxxxxxx>
> Cc: Mel Gorman <mel@xxxxxxxxx>
> Signed-off-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> ---
> 
>  mm/vmscan.c |    5 +++--
>  1 file changed, 3 insertions(+), 2 deletions(-)
> 
> diff -puN mm/vmscan.c~mm-isolate_lru_page-dont-enable-local-interrupts mm/vmscan.c
> --- a/mm/vmscan.c~mm-isolate_lru_page-dont-enable-local-interrupts
> +++ a/mm/vmscan.c
> @@ -1161,8 +1161,9 @@ int isolate_lru_page(struct page *page)
>  	if (PageLRU(page)) {
>  		struct zone *zone = page_zone(page);
>  		struct lruvec *lruvec;
> +		unsigned long flags;
>  
> -		spin_lock_irq(&zone->lru_lock);
> +		spin_lock_irqsave(&zone->lru_lock, flags);
>  		lruvec = mem_cgroup_page_lruvec(page, zone);
>  		if (PageLRU(page)) {
>  			int lru = page_lru(page);
> @@ -1171,7 +1172,7 @@ int isolate_lru_page(struct page *page)
>  			del_page_from_lru_list(page, lruvec, lru);
>  			ret = 0;
>  		}
> -		spin_unlock_irq(&zone->lru_lock);
> +		spin_unlock_irqrestore(&zone->lru_lock, flags);
>  	}
>  	return ret;
>  }
> _
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/