Date:   Mon, 9 Mar 2020 14:11:17 +0100
From:   Oleksandr Natalenko <oleksandr@...hat.com>
To:     Vlastimil Babka <vbabka@...e.cz>
Cc:     Minchan Kim <minchan@...nel.org>,
        Andrew Morton <akpm@...ux-foundation.org>,
        LKML <linux-kernel@...r.kernel.org>,
        linux-mm <linux-mm@...ck.org>, linux-api@...r.kernel.org,
        Suren Baghdasaryan <surenb@...gle.com>,
        Tim Murray <timmurray@...gle.com>,
        Daniel Colascione <dancol@...gle.com>,
        Sandeep Patil <sspatil@...gle.com>,
        Sonny Rao <sonnyrao@...gle.com>,
        Brian Geffon <bgeffon@...gle.com>,
        Michal Hocko <mhocko@...e.com>,
        Johannes Weiner <hannes@...xchg.org>,
        Shakeel Butt <shakeelb@...gle.com>,
        John Dias <joaodias@...gle.com>,
        Joel Fernandes <joel@...lfernandes.org>,
        Jann Horn <jannh@...gle.com>,
        alexander.h.duyck@...ux.intel.com, sj38.park@...il.com,
        SeongJae Park <sjpark@...zon.de>
Subject: Re: [PATCH v7 7/7] mm/madvise: allow KSM hints for remote API

On Fri, Mar 06, 2020 at 05:08:18PM +0100, Vlastimil Babka wrote:
> On 3/6/20 2:41 PM, Oleksandr Natalenko wrote:
> > On Fri, Mar 06, 2020 at 02:13:49PM +0100, Vlastimil Babka wrote:
> >> On 3/2/20 8:36 PM, Minchan Kim wrote:
> >> > From: Oleksandr Natalenko <oleksandr@...hat.com>
> >> > 
> >> > It all began with the fact that KSM works only on memory that is marked
> >> > by madvise(). And the only way to get around that is to either:
> >> > 
> >> >   * use LD_PRELOAD; or
> >> >   * patch the kernel with something like UKSM or PKSM.
> >> > 
> >> > (i skip ptrace can of worms here intentionally)
> >> > 
> >> > To overcome this restriction, lets employ a new remote madvise API. This
> >> > can be used by some small userspace helper daemon that will do auto-KSM
> >> > job for us.
> >> > 
> >> > I think of two major consumers of remote KSM hints:
> >> > 
> >> >   * hosts, that run containers, especially similar ones and especially in
> >> >     a trusted environment, sharing the same runtime like Node.js;
> 
> Ah, I forgot to ask, given the discussion of races in patch 2 (Question 2),
> where android can stop the tasks to apply the madvise hints in a race-free
> manner, how does that work for remote KSM hints in your scenarios, especially
> the one above?

We have cgroup.freeze for that.

> 
> >> > 
> >> >   * heavy applications, that can be run in multiple instances, not
> >> >     limited to opensource ones like Firefox, but also those that cannot be
> >> >     modified since they are binary-only and, maybe, statically linked.
> >> > 
> >> > Speaking of statistics, more numbers can be found in the very first
> >> > submission, that is related to this one [1]. For my current setup with
> >> > two Firefox instances I get 100 to 200 MiB saved for the second instance
> >> > depending on the amount of tabs.
> >> > 
> >> > 1 FF instance with 15 tabs:
> >> > 
> >> >    $ echo "$(cat /sys/kernel/mm/ksm/pages_sharing) * 4 / 1024" | bc
> >> >    410
> >> > 
> >> > 2 FF instances, second one has 12 tabs (all the tabs are different):
> >> > 
> >> >    $ echo "$(cat /sys/kernel/mm/ksm/pages_sharing) * 4 / 1024" | bc
> >> >    592
> >> > 
> >> > At the very moment I do not have specific numbers for containerised
> >> > workload, but those should be comparable in case the containers share
> >> > similar/same runtime.
> >> > 
> >> > [1] https://lore.kernel.org/patchwork/patch/1012142/
> >> > 
> >> > Reviewed-by: SeongJae Park <sjpark@...zon.de>
> >> > Signed-off-by: Oleksandr Natalenko <oleksandr@...hat.com>
> >> > Signed-off-by: Minchan Kim <minchan@...nel.org>
> >> 
> >> This will lead to one process calling unmerge_ksm_pages() of another. There's a
> >> (signal_pending(current)) test there, should it check also the other task,
> >> analogically to task 3?
> > 
> > Do we care about current there then? Shall we just pass mm into unmerge_ksm_pages and check the signals of the target task only, be it current or something else?
> 
> Dunno, it's nice to react to signals quickly, for any proces that gets them, no?

So, do you mean something like this?

===
diff --git a/mm/ksm.c b/mm/ksm.c
index 363ec8189561..b39c237cfcf4 100644
--- a/mm/ksm.c
+++ b/mm/ksm.c
@@ -849,7 +849,8 @@ static int unmerge_ksm_pages(struct vm_area_struct *vma,
 	for (addr = start; addr < end && !err; addr += PAGE_SIZE) {
 		if (ksm_test_exit(vma->vm_mm))
 			break;
-		if (signal_pending(current))
+		if (signal_pending(current) ||
+		    signal_pending(rcu_dereference(vma->vm_mm->owner)))
 			err = -ERESTARTSYS;
 		else
 			err = break_ksm(vma, addr);
===

BTW, this won't work with !CONFIG_MEMCG, so probably task_struct should be
passed through instead. IIUC, this would also require amending struct
mm_slot in order to share the same code path with ksmd.

I'm not sure I've seen such a culprit anywhere else, so I'm in doubt
this would be a correct thing to do.

Ideas?

> 
> >> Then break_ksm() is fine as it is, as ksmd also calls it, right?
> > 
> > I think break_ksm() cares only about mmap_sem protection, so we should
> > be fine here.
> > 
> >> 
> >> > ---
> >> >  mm/madvise.c | 4 ++++
> >> >  1 file changed, 4 insertions(+)
> >> > 
> >> > diff --git a/mm/madvise.c b/mm/madvise.c
> >> > index e77c6c1fad34..f4fa962ee74d 100644
> >> > --- a/mm/madvise.c
> >> > +++ b/mm/madvise.c
> >> > @@ -1005,6 +1005,10 @@ process_madvise_behavior_valid(int behavior)
> >> >  	switch (behavior) {
> >> >  	case MADV_COLD:
> >> >  	case MADV_PAGEOUT:
> >> > +#ifdef CONFIG_KSM
> >> > +	case MADV_MERGEABLE:
> >> > +	case MADV_UNMERGEABLE:
> >> > +#endif
> >> >  		return true;
> >> >  	default:
> >> >  		return false;
> >> > 
> >> 
> > 
> 

-- 
  Best regards,
    Oleksandr Natalenko (post-factum)
    Principal Software Maintenance Engineer

Powered by blists - more mailing lists