lkml.org 
[lkml]   [2012]   [Mar]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] hugetlbfs: lockdep annotate root inode properly
On Thu, Mar 08, 2012 at 01:02:56PM -0800, Andrew Morton wrote:
> > This fix the below lockdep warning
>
> OK, what's going on here.

Deadlock in hugetlbfs mmap getting misreported.

One last time: ->mmap_sem nests inside ->i_mutex. Both for regular
files and for directories. Always had.

For directories there's copy_to_user() from ->readdir() done under ->i_mutex.
For regular files there's copy_from_user() from ->write(), usually done under
->i_mutex. On hugetlbfs there's copy_to_user() from ->read() done under
->i_mutex.

It had not changed at all. Lockdep sees both call chains; the only question
is which chain is seen first. And usually reading a directory happens earlier
in the boot than writing into a file. That's all there is to it.

Unfortunately, the fact that call chain being reported is obviously about
directories leads to false hopes that deadlock doesn't exist - mmap()
obviously can't happen to a directory inode, so people hope that it's a
false positive. It isn't.

Patch separating directory and non-directory ->i_mutex into different classes
went in at some point, precisely due to those hopes. It had a braino that
made it useless. Fix for that braino had been posted and sits my queue; I'll
push it to Linus along with other pending fixes tonight.

It will *not* eliminate the (very real) deadlock. It might make the warning
go away, but only if read() on hugetlbfs files doesn't happen during boot.

I suspect that they right thing would be to have a way to set explicit
nesting rules, not tied to speficic call trace. I hadn't looked into
lockdep guts, so no idea how much will that hurt to implement. As in
lockdep_lock_nests(class_outer, class_inner, message), acting as if
there had been a call chain where class_outer had been taken before
class_inner, with message going in place of call trace for that chain
when we run into a conflict...


\
 
 \ /
  Last update: 2012-03-08 22:47    [from the cache]
©2003-2011 Jasper Spaans