Jeff Layton wrote:
> On Mon, 9 Jun 2008 14:39:23 -0400
> "J. Bruce Fields" <[email protected]> wrote:
>
>
>> On Mon, Jun 09, 2008 at 09:19:48AM -0400, Jeff Layton wrote:
>>
>>> On Mon, 9 Jun 2008 09:08:30 -0400
>>> "J. Bruce Fields" <[email protected]> wrote:
>>>
>>>
>>>> On Fri, Jun 06, 2008 at 03:49:48PM -0400, Jeff Layton wrote:
>>>>
>>>>> On Fri, 6 Jun 2008 14:11:16 -0400
>>>>> Jeff Layton <[email protected]> wrote:
>>>>>
>>>>>
>>>>>> On Fri, 6 Jun 2008 13:24:31 -0400
>>>>>> "J. Bruce Fields" <[email protected]> wrote:
>>>>>>
>>>>>>
>>>>> ...
>>>>>
>>>>>>> How does the module refcounting work after this patch?
>>>>>>>
>>>>>>> --b.
>>>>>>>
>>>>>>>
>>>>>> I think I've goofed this part, actually. I was thinking that we didn't
>>>>>> need to bump the refcount here, and that the kernel would realize that
>>>>>> nfsd() hadn't returned and would prevent unloading until it had. This
>>>>>> doesn't seem to be the case. I'll need to go back and add refcounting
>>>>>> back in.
>>>>>>
>>>>>>
>>>>> Here's a respun patch that adds back in the module refcounts and also
>>>>> removes the unneeded "err = 0;" at the bottom of the loop. Thoughts?
>>>>>
>>>> Looks good to me. I'll apply all 5 (with this version of #4) if noone
>>>> catches something else.
>>>>
>>>> --b.
>>>>
>>>>
>>> Sounds good. My only concern here is whether moving the __module_get
>>> from the RPC layer to nfsd() itself is OK.
>>>
>> Oh, I see, I missed that.
>>
>>
>>> I *think* it is since the
>>> nfsctl and /proc/fs/nfsd routines are all part of the nfsd module, so
>>> we're guaranteed to have a reference there anyway, but if there are
>>> potential races then we may want to go back to the old way.
>>>
>> The vfs should take care that e.g. it gets a reference on the module
>> before creating an open file for the nfsd filesystem.
Yes, but indirectly; I believe the module refcount belongs to the nfsd
filesystem super_block, and having an open file on that filesystem keeps
a vfsmount refcount which keeps the super_block refcount which keeps the
module refcount.
>> But it don't see
>> how anything can guarantee that the __module_get() in the new nfsd
>> thread completes before whoever called svc_set_num_threads() returns and
>> drops their reference.
>>
>> So, yeah, I think it's not right.
>>
>>
Agreed.
>
> Ok, that makes sense. I'll need to respin the set then since we'll need
> the sv_module field in the svc_serv struct after all. Let me do that
> and get back to you...
>
>
Bother, I think that's right.
--
Greg Banks, P.Engineer, SGI Australian Software Group.
The cake is *not* a lie.
I don't speak for SGI.