Re: [PATCH 2/4] mm: introduce external memory hinting API

[Christian Brauner]

On Tue, Jan 14, 2020 at 10:59:44AM -0800, Minchan Kim wrote:
> Hi Christian,
> 
> On Mon, Jan 13, 2020 at 08:10:47PM +0100, Christian Brauner wrote:
> > On Mon, Jan 13, 2020 at 10:44:08AM -0800, Minchan Kim wrote:
> > > On Mon, Jan 13, 2020 at 11:42:57AM +0100, Christian Brauner wrote:
> > > > On Mon, Jan 13, 2020 at 11:47:11AM +0300, Kirill Tkhai wrote:
> > > 
> > > < snip >
> > > 
> > > > > > +SYSCALL_DEFINE5(process_madvise, int, pidfd, unsigned long, start,
> > > > > > +		size_t, len_in, int, behavior, unsigned long, flags)
> > > > > 
> > > > > I don't like the interface. The fact we have pidfd does not mean,
> > > > > we have to use it for new syscalls always. A user may want to set
> > > > > madvise for specific pid from console and pass pid as argument.
> > > > > pidfd would be an overkill in this case.
> > > > > We usually call "kill -9 pid" from console. Why shouldn't process_madvise()
> > > > > allow this?
> > > > > 
> > > > > I suggent to extend first argument to work with both pid and pidfd.
> > > > > Look at what we have for waitid(idtype, id_t id, ...) for example:
> > > > > 
> > > > >        idtype == P_PID
> > > > >               Wait for the child whose process ID matches id.
> > > > > 
> > > > >        idtype == P_PIDFD (since Linux 5.4)
> > > > >               Wait for the child referred to by the PID file descriptor specified in id.  (See pidfd_open(2) for  further  information  on
> > > > >               PID file descriptors.)
> > > > > 
> > > > > We may use @flags argument for this.
> > > > 
> > > > Sorry for chiming in just a comment. Overall, I don't particularly care
> > > > how or if you integrate pidfd here. One thing I would like to point out
> > > > is that we're working on a patch to place new features under pidfd
> > > > specific flags. This e.g. means a pidfd would be only be able to be used
> > > > for madvise operations (or getfd operations) if it was created with that
> > > > specific flag set making it easier to share them with other processes.
> > > > So if you integrate them here I would be quite thankful if you target
> > > > the patchset for the v5.7 merge window, not for v5.6.
> > > 
> > > Hi Christian,
> > > Sorry but I couldn't understand your point.
> > > Could you clarify what you meant?
> > 
> > Hi Minchan,
> > 
> > Sure. When you create a pidfd, e.g. with clone3() and you'd wanted to
> > use it for madvise you'd need to set a flag like pidfd_cap_madvise or
> > pidfd_feature_madvise when you create the pidfd. Only if the pidfd was
> > created with that flag set could you use it with madvise (This does not
> > affect the permission checking you're performing here.). This has come
> > up a couple of times and becomes more relevant now that people keep
> > adding new features on top of pidfd and is similar to what we are now
> > doing with openat2().
> 
> Thanks for the explain. When I read discussion with you and Daniel, it's
> still vague for me that what's the outcome so that it could land onto
> v5.6.(If I miss something progress on other thread, sorry about that.)

I'll try to post patches soon.

> 
> I will keep Ccing you so that you may notice when this patchset could
> be merged(Please Cc me when you send your patchset for me to notice)
> So if we judge it's worth to integrate, maybe we could make a quick
> patch to use it or postpone a cycle to intergrate it if we have more
> time.

Yeah, that would be great!
It's unlikely that process_madvise() will land for v5.6 anyway since
it's quite late in the cycle, so we should have some time to coordinate.

Thanks!
Christian