Re: [PATCH 0/3] locking/static_key: improve rate limited labels
From: Peter Zijlstra
Date: Tue Apr 16 2019 - 13:39:00 EST
On Tue, Apr 16, 2019 at 09:33:15AM -0700, Jakub Kicinski wrote:
> On Thu, 4 Apr 2019 12:29:02 +0200, Peter Zijlstra wrote:
> > On Mon, Apr 01, 2019 at 11:21:53AM -0700, Jakub Kicinski wrote:
> > > On Fri, 29 Mar 2019 17:08:51 -0700, Jakub Kicinski wrote:
> > > > Hi!
> > > >
> > > > This will be used to fix the static branch disabling in the TLS
> > > > code. The net/tls/ code should be using the deferred static
> > > > branch type, because unprivileged users can flip the branch
> > > > on and off quite easily with CONFIG_TLS_DEVICE=y.
> > > >
> > > > Second of all we shouldn't take the jump label locks from
> > > > the RX path, when the socket is destroyed. This can be avoided
> > > > with some slight code refactoring in deferred static_key as
> > > > it already runs from a workqueue.
> > > >
> > > > This the series (and a simple tls patch which makes use of it)
> > > > applied opening 0.5M TLS connections to localhost (just calling
> > > > setsockopt, no data exchange) goes down from 37.9s to 12.4s.
> > >
> > > Once/if we get positive feedback from locking folks, would it be
> > > possible to merge these via net-next tree alongside the patch
> > > converting TLS to rate limited branches?
> > Looks good. If routed through the network tree because usage there:
> > Acked-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
> > Otherwise let me know and I'll carry them.
> Hi Peter! I was initially hoping that a1247d06d010
> ("locking/static_key: Fix false positive warnings on concurrent dec/inc")
> may go into 5.1, but it's not really a regression. It will conflict, so
> the net-next route won't work. Would you be able to carry this set
> after all?