Re: New Lock Order Reversal in 12.0?

From: Benjamin Kaduk <kaduk_at_mit.edu>
Date: Mon, 9 Jan 2017 15:53:42 -0600
On Mon, Jan 09, 2017 at 02:31:39AM +0000, Anindya Mukherjee wrote:
> Hi Ben,
> 
> Thanks for your reply, and yes, I did notice #238. I should say at this point that I'm a newbie when it comes to kernel code and am trying to learn about it (hence current).

Understood.  It's good that you are ready to try!

> The entry you refer to does look a bit like the one I've got, but I'm not totally sure if the same code path is being followed to arrive at this LOR. An inode is being created in my case, vs a directory creation (entry + inode probably) in #238, and then a sync is being attempted, which causes locks to activate in the softdep code. I've read a bit about this from McCusick's book but the details are still fuzzy.
> 
> Perhaps you are trying to tell me that it's benign? I know that WITNESS has false positives, an example being #236 where due to shared vs exclusive vnode locks required on the two ways to lock bufwait and dirhash a deadlock never happens. 

Well, it's hard to say for certain, but there are a few ufs/bufwait/ufs
LORs that are very commonly reported as WITNESS output but have never (IIRC)
been identified as causing a deadlock.  So, it seems pretty likely that this
one is similarly benign -- I, for one, do not plan to put in time trying
to analyze it until there is a hang or deadlock associated with it.

-Ben
Received on Mon Jan 09 2017 - 20:59:00 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:41:09 UTC