Messages in this thread | | | Date | Thu, 8 Mar 2012 21:44:25 +0000 | From | Al Viro <> | Subject | Re: [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...
| |