Re: [PATCH 1/4] sunrpc: add a rcu_head to svc_rqst and use kfree_rcu to free it

["J. Bruce Fields" <bfields@fieldses.org>]

On Mon, Dec 01, 2014 at 07:52:20PM -0500, Trond Myklebust wrote:
> On Mon, Dec 1, 2014 at 7:29 PM, Jeff Layton <jeff.layton@primarydata.com> wrote:
> > There are two atomicity "concerns" here...
> >
> > The main thing is to ensure that we use set_bit or test_and_set_bit to
> > set the flag. What we *can't* use is __set_bit which is non-atomic
> > or we'd end up hitting the exact problem you're talking about (possibly
> > changing an unrelated flag in the field that happened to flip at nearly
> > the same time).
> >
> > What's not necessary here is to use test_and_set_bit since all of this
> > is done under spinlock anyway. In principle, we could do a test_bit and
> > then follow that up with a set_bit if it's clear. But, I don't think
> > that really buys us much, and tend to find the test_and_set_bit to be
> > clearer when reading the code.
> 
> Fair enough. I too would be surprised if you could actually measure
> that performance difference in the thread kill code.

Yeah, please don't bother, I was just thinking aloud here.

--b.