kern/52301: lfs deadlock between lfs_fsync and lfs_create

[dholland%NetBSD.org@localhost]

>Number:         52301
>Category:       kern
>Synopsis:       lfs deadlock between lfs_fsync and lfs_create
>Confidential:   no
>Severity:       critical
>Priority:       medium
>Responsible:    kern-bug-people
>State:          open
>Class:          sw-bug
>Submitter-Id:   net
>Arrival-Date:   Thu Jun 15 06:15:00 +0000 2017
>Originator:     David A. Holland
>Release:        NetBSD 7.99.75 (20170603)
>Organization:
>Environment:
System: NetBSD loggy 7.99.75 NetBSD 7.99.75 (GENERIC) #66: Sat Jun 10 05:26:28
+IDT 2017  fly@loggy:/home/fly/obj/sys/arch/amd64/compile/GENERIC amd64
Architecture: x86_64
Machine: amd64
>Description:

lfs can deadlock as follows:

1. process A calls lfs_fsync
   - which calls lfs_vflush
   - which takes the seglock and calls lfs_segwrite
2. meanwhile process B calls lfs_create
   - which starts a dirop
   - and then down inside lfs_valloc blocks waiting for the seglock.
3. now process A in lfs_segwrite calls lfs_writer_enter
   - which blocks waiting for dirops to clear
   - but they can't.

We found this by analyzing a crashdump, and confirmed it in the source
of a 7.99.75 tree from 20170603.

>How-To-Repeat:

See above.

>Fix:

Not obvious.

I wonder if this was introduced by the vnode cache changes... but even
if so reverting those isn't the answer.