[MODERATED] Re: ***UNCHECKED*** Re: [PATCH v4 03/10] TAAv4 3

On Wed, Sep 04, 2019 at 01:27:58PM +0200, speck for Michal Hocko wrote:
> On Wed 04-09-19 10:43:06, speck for Greg KH wrote:
> > On Wed, Sep 04, 2019 at 12:58:47AM -0700, speck for Pawan Gupta wrote:
> > > On Wed, Sep 04, 2019 at 08:11:55AM +0200, speck for Greg KH wrote:
> > > > On Wed, Sep 04, 2019 at 08:01:47AM +0200, speck for Jiri Kosina wrote:
> > > > > On Wed, 4 Sep 2019, speck for Greg KH wrote:
> > > > > 
> > > > > > Did we ever have a reason to enable TSX?  I thought no one used it as it 
> > > > > > just didn't make any sense (as per Linus's old email about this)
> > > > > > 
> > > > > > Who will be turning this on?
> > > > > 
> > > > > We know about certain database vendor(s) who are somehow making use of TSX 
> > > > > for whatever reason, so having such a knob would be nice of us.
> > > > 
> > > > So they now need to explicitly enable it?  Would that not break their
> > > > existing systems by forcing them to do an additional startup step, or is
> > > > that somehow ok here?
> > > 
> > > For cases when the additional startup step is a problem, there is also
> > > a sysfs interface to control TSX after boot.
> > 
> > And that extra step could be a problem.  You had a machine that had TSX
> > enabled and the program used it.  Upgrade your kernel and now you have
> > to add an additional step in order to use this (command line or sysfs).
> > 
> > I am asking if this is going to be an issue for those systems that are
> > expecting a working TSX.
> 
> Indeed. I was expecting auto|on|off semantic like other knobs with auto
> preserving the current semantic (disable on broken models) and off to
> override.

In the previous version of this patch set, Thomas requested the default
be 'tsx=off', which is what Pawan has done now in this version.

But I agree with Greg, that would break existing workloads.  So I think
the default needs to be 'tsx=on', along with 'taa=full'.

-- 
Josh