On Mon, Mar 23, 2020 at 03:51:06PM +0100, Michal Hocko wrote:
> On Mon 23-03-20 10:42:40, Rafael Aquini wrote:
> > On Mon, Mar 23, 2020 at 08:52:08AM +0100, Michal Hocko wrote:
> > > On Sun 22-03-20 09:36:49, Shakeel Butt wrote:
> > > > On Sat, Mar 21, 2020 at 9:31 PM Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
> > > > >
> > > > > On Sat, 21 Mar 2020 22:03:26 -0400 Rafael Aquini <aquini@xxxxxxxxxx> wrote:
> > > > >
> > > > > > > > + * In order to sort out that race, and get the after fault checks consistent,
> > > > > > > > + * the "quick and dirty" trick below is required in order to force a call to
> > > > > > > > + * lru_add_drain_all() to get the recently MLOCK_ONFAULT pages moved to
> > > > > > > > + * the unevictable LRU, as expected by the checks in this selftest.
> > > > > > > > + */
> > > > > > > > +static void force_lru_add_drain_all(void)
> > > > > > > > +{
> > > > > > > > + sched_yield();
> > > > > > > > + system("echo 1 > /proc/sys/vm/compact_memory");
> > > > > > > > +}
> > > > > > >
> > > > > > > What is the sched_yield() for?
> > > > > > >
> > > > > >
> > > > > > Mostly it's there to provide a sleeping gap after the fault, whithout
> > > > > > actually adding an arbitrary value with usleep().
> > > > > >
> > > > > > It's not a hard requirement, but, in some of the tests I performed
> > > > > > (whithout that sleeping gap) I would still see around 1% chance
> > > > > > of hitting the false-negative. After adding it I could not hit
> > > > > > the issue anymore.
> > > > >
> > > > > It's concerning that such deep machinery as pagevec draining is visible
> > > > > to userspace.
> > > > >
> > > >
> > > > We already have other examples like memcg stats where the
> > > > optimizations like batching per-cpu stats collection exposes
> > > > differences to the userspace. I would not be that worried here.
> > >
> > > Agreed! Tests should be more tolerant for counters imprecision.
> > > Unevictable LRU is an optimization and transition to that list is a
> > > matter of an internal implementation detail.
> > >
> > > > > I suppose that for consistency and correctness we should perform a
> > > > > drain prior to each read from /proc/*/pagemap. Presumably this would
> > > > > be far too expensive.
> > > > >
> > > > > Is there any other way? One such might be to make the MLOCK_ONFAULT
> > > > > pages bypass the lru_add_pvecs?
> > > > >
> > > >
> > > > I would rather prefer to have something similar to
> > > > /proc/sys/vm/stat_refresh which drains the pagevecs.
> > >
> > > No, please don't. Pagevecs draining is by far not the only batching
> > > scheme we use and an interface like this would promise users to
> > > effectivelly force flushing all of them.
> > >
> > > Can we simply update the test to be more tolerant to imprecisions
> > > instead?
> > >
> >
> > I don't think, thouhg, that this particular test case can be entirely
> > reduced as "counter imprecison".
> >
> > The reason I think this is a different beast, is that having the page
> > being flagged as PG_unevictable is expected part of the aftermath of
> > a mlock* call. This selftest is, IMO, correctly verifying that fact,
> > as it checks the functionality correctness.
> >
> > The problem boils down to the fact that the page would immediately
> > be flagged as PG_unevictable after the mlock (under MCL_FUTURE|MCL_ONFAULT
> > semantics) call, and the test was expecting it, and commit 9c4e6b1a7027f
> > changed that by "delaying" that flag setting.
>
> As I've tried to explain in other email in this email thread. The test
> was exploiting a certain user visible side effect. The unevictable flag
> or the placement on the unevictable LRU list is are not really needed
> for the user contract correctness. That means that the test is not
> really correct. Working around that by trying to enforce kernel to
> comply with the test expectations is just plain wrong at least for two
> reasons 1) you cannot expect or event do not want userspace to do the
> same because the behavior might change in the future 2) the test is not
> really testing for correctness in the first place.
>
Sorry, Michal, it seems we keep going back and forth (I just replied to
your comment on the other thread)
The selftest also checks the kernel visible effect, via
/proc/kpageflags, and that's where it fails after 9c4e6b1a7027f.
As I mentioned before, I think it is a reasonable check, given this
is a kernel selftest, although we need to compensate it for the
differences between its expectations and what the kernel is doing
currently.
-- Rafael
