Date: Mon, 20 Dec 2010 07:53:06 -0500 From: John Baldwin <jhb@freebsd.org> To: freebsd-fs@freebsd.org Subject: Re: debugging process in bovlbx state Message-ID: <201012200753.06804.jhb@freebsd.org> In-Reply-To: <alpine.GSO.1.10.1012191909420.640@multics.mit.edu> References: <alpine.GSO.1.10.1012191909420.640@multics.mit.edu>
next in thread | previous in thread | raw e-mail | index | archive | help
On Sunday, December 19, 2010 7:10:04 pm Benjamin Kaduk wrote:
> Hi all,
>
> I'm working on bringing the out-of-tree OpenAFS network filesystem
> up-to-date for FreeBSD 7.3-RELEASE, and I think I need some help to fix
> this bug.
> I should preface my discourse with the fact that there is a whole slow of
> lock order reversals that I haven't even tried to track down, but I do not
> believe that this hang is deadlock since 'show alllocks' in DDB does not
> show anything that seems interesting.
>
> Any pointers for things to look at would be appreciated; more details of
> the failing case below.
>
>
> In order to get the afs kernel module to load, I needed to tweak a few
> lines of code in getpages(), as I had previously cribbed a bunch of
> changes/updates from the experimental NFS client while getting AFS to work
> on current freebsd. In particular, vm_page_set_valid is not present in
> 7.3, so I am currently running with:
> --- a/src/afs/FBSD/osi_vnodeops.c
> +++ b/src/afs/FBSD/osi_vnodeops.c
> @@ -890,12 +890,8 @@ afs_vop_getpages(struct vop_getpages_args *ap)
> * Read operation filled a partial page.
> */
> m->valid = 0;
> - vm_page_set_valid(m, 0, size - toff);
> -#ifndef AFS_FBSD80_ENV
> - vm_page_undirty(m);
> -#else
> + vm_page_set_validclean(m, 0, size - toff);
> KASSERT(m->dirty == 0, ("afs_getpages: page %p is dirty", m));
> -#endif
> }
>
>
> But my knowledge of vm_page_* is approximately nil, so there's no reason
> to think everything was correct even before that patch.
>
> Anyway, my test case is running libarchive's configure script with source
> and destination directories in (different places in) AFS. It only gets
> twenty lines in, ending with:
> checking for gcc option to accept ISO C89... none needed
> checking for style of include used by make... GNU
> checking dependency style of gcc...
> ^Tload: 0.04 cmd: cp 1250 [bovlbx] 0.00u 0.00
>
> procstat -kk reports:
> mega-man# procstat -kk 1250
> PID TID COMM TDNAME KSTACK
> 1250 100060 cp - mi_switch+0x233
> sleepq_switch+0xe9 sleepq_wait+0x44 _sleep+0x3a0 vm_object_pip_wait+0x4e
> bufobj_invalbuf+0x10e afs_GetVCache+0x2f7
>
> The call to vinvalbuf in afs_GetVCache is here:
> 1646 iheldthelock = VOP_ISLOCKED(vp, curthread);
This is probably wrong. VOP_ISLOCKED() can return four different values:
- LK_SHARED: (someone, possibly curthread) holds a shared lock
- LK_EXCLUSIVE: curthread holds an exclusive lock
- LK_EXCLOTHER: some other thread holds an exclusive lock
- 0: no thread holds any lock.
This means if another thread has the vnode locked, you don't try to lock it.
:)
Do you actually know that this routine can be held without the vnode locked by
the current thread?
--
John Baldwin
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?201012200753.06804.jhb>