svn commit: r212506 - head/sys/nfsclient

[Robert Watson]

On Sun, 12 Sep 2010, Kostik Belousov wrote:

> On Sun, Sep 12, 2010 at 09:41:46PM +0100, Robert Watson wrote:
>>
>> On Sun, 12 Sep 2010, Konstantin Belousov wrote:
>>
>>> Do not fork nfsiod directly from the vop methods. This causes LORs between 
>>> vnode lock and several locks needed during fork, like fd lock.
>>>
>>> Instead, schedule the task to be executed in the taskqueue context. We 
>>> still waiting for the fork to finish, but the context of the thread 
>>> executing the task does not make real LORs with our vnode lock.
>>
>> Does this actually functionally improve things, or is all this complexity 
>> about suppressing the lock order reversal?  If we're waiting synchronously 
>> for the other thread to launch from the task queue, then the lock order 
>> reversal still exists, it's just not visible to WITNESS...  If we had a 
>> static analysis tool that could run on lock and sleep/wakeup traces, it 
>> would still show a deadlock opportunity.
>
> As I said in commit message, the deadlock should be fixed, because the taskq 
> thread is executed in the kernel process that should even not have file 
> descriptors at all.

Ah, I think I see where my misunderstanding arose: the behavior of kernel 
process fork is changed as a result of running in a different (specific) 
context, and hence never acquires the file descriptor lock class?

Robert