Messages in this thread | | | Date | Mon, 15 Sep 2008 00:02:08 +0200 | From | Nick Piggin <> | Subject | Re: [PATCH] sysfs: fix deadlock |
| |
On Fri, Sep 12, 2008 at 11:24:11AM +0200, Ingo Molnar wrote: > > [ Greg, please see the sysfs fix further below. ] > > * Nick Piggin <npiggin@suse.de> wrote: > > > > - moved the might_sleep() check outside the in_atomic() check, > > > > Hmm... but then it has the same failure case again in the is_preempt() > > code, does it not? > > > > I guess we should just convert that guy to either use get_user_atomic, > > (which would mean implementing that for x86), or use > > copy_from_user_inatomic. > > i've done the v3 patch below - that seems to have passed all my testing > without any new bugs found. I've reinstated your the clear_user() > might_fault() check, plus i removed it from __[get|put]_user_size, which > the _inatomic() API variants use. That enabled me to utilize the > _inatomic() API in probe_kernel_address(). > > we still have the checks in put_user()/get_user() and in all the > copy_*_user() APIs, which should be strong enough. [ I havent fully > checked whether __get_user_size() might be used by some less frequent > API - if it is then that API should grow a might_fault() check. ] > > > > i've attached the config. > > > > > > at first sight it looks like a genuine bug in fs/sysfs/bin.c? > > > > Yes, it is a real bug by the looks. bin.c takes bb->mutex under > > mmap_sem when it is mmapped, and then does its copy_*_user under > > bb->mutex too. > > ok - second patch attached below, Greg, could you please apply? This is > for v2.6.27 too i think. > > > > i.e. your patches are working as expected and the extended > > > validation mechanism is finding real bugs :-) > > > > Yeah it's nice. I'm just hoping we don't come across one that is as > > difficult to fix as prepare_write/commit_write were ;) > > > > Here is a basic fix for the sysfs lor. > > and that did the trick here - the patch with a tidied up changelog is > attached further below. [ the second patch is standalone and does not > need the first patch which is relative to tip/master ] > > thanks Nick, i think this is a great addition to lockdep! It already > found two real locking bugs within a day. If you can think of any other > proactive methods to widen our lock hierarchy knowledge that would be > great to add. I think what we want is to insert knowledge about other > unlikely lock acquire events, for locks that have a historic pattern of > producing regular locking bugs.
Well thanks to Peter as well. Actually I don't suppose this will throw off the lockstat statistics a bit? (although I guess serious lockstat profiling might not have prove locking turned on?).
The user fault I guess is the main thing like this in the VM that I can think of.
The user fault I guess is the main thing like this in the VM that I can think of.
| |