Re: [PATCH] mm: memcontrol: fix data race in mem_cgroup_select_victim_node

[Johannes Weiner]

On Tue, Oct 29, 2019 at 11:09:29AM -0700, Shakeel Butt wrote:
> +Marco
> 
> On Tue, Oct 29, 2019 at 2:03 AM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> >
> > On Mon 28-10-19 17:54:05, Shakeel Butt wrote:
> > > Syzbot reported the following bug:
> > >
> > > BUG: KCSAN: data-race in mem_cgroup_select_victim_node / mem_cgroup_select_victim_node
> > >
> > > write to 0xffff88809fade9b0 of 4 bytes by task 8603 on cpu 0:
> > >  mem_cgroup_select_victim_node+0xb5/0x3d0 mm/memcontrol.c:1686
> > >  try_to_free_mem_cgroup_pages+0x175/0x4c0 mm/vmscan.c:3376
> > >  reclaim_high.constprop.0+0xf7/0x140 mm/memcontrol.c:2349
> > >  mem_cgroup_handle_over_high+0x96/0x180 mm/memcontrol.c:2430
> > >  tracehook_notify_resume include/linux/tracehook.h:197 [inline]
> > >  exit_to_usermode_loop+0x20c/0x2c0 arch/x86/entry/common.c:163
> > >  prepare_exit_to_usermode+0x180/0x1a0 arch/x86/entry/common.c:194
> > >  swapgs_restore_regs_and_return_to_usermode+0x0/0x40
> > >
> > > read to 0xffff88809fade9b0 of 4 bytes by task 7290 on cpu 1:
> > >  mem_cgroup_select_victim_node+0x92/0x3d0 mm/memcontrol.c:1675
> > >  try_to_free_mem_cgroup_pages+0x175/0x4c0 mm/vmscan.c:3376
> > >  reclaim_high.constprop.0+0xf7/0x140 mm/memcontrol.c:2349
> > >  mem_cgroup_handle_over_high+0x96/0x180 mm/memcontrol.c:2430
> > >  tracehook_notify_resume include/linux/tracehook.h:197 [inline]
> > >  exit_to_usermode_loop+0x20c/0x2c0 arch/x86/entry/common.c:163
> > >  prepare_exit_to_usermode+0x180/0x1a0 arch/x86/entry/common.c:194
> > >  swapgs_restore_regs_and_return_to_usermode+0x0/0x40
> > >
> > > mem_cgroup_select_victim_node() can be called concurrently which reads
> > > and modifies memcg->last_scanned_node without any synchrnonization. So,
> > > read and modify memcg->last_scanned_node with READ_ONCE()/WRITE_ONCE()
> > > to stop potential reordering.
> >
> > I am sorry but I do not understand the problem and the fix. Why does the
> > race happen and why does _ONCE fixes it? There is still no
> > synchronization. Do you want to prevent from memcg->last_scanned_node
> > reloading?
> >
> 
> The problem is memcg->last_scanned_node can read and modified
> concurrently. Though to me it seems like a tolerable race and not
> worth to add an explicit lock. My aim was to make KCSAN happy here to
> look elsewhere for the concurrency bugs. However I see that it might
> complain next on memcg->scan_nodes.
> 
> Now taking a step back, I am questioning the whole motivation behind
> mem_cgroup_select_victim_node(). Since we pass ZONELIST_FALLBACK
> zonelist to the reclaimer, the shrink_node will be called for all
> potential nodes. Also we don't short circuit the traversal of
> shrink_node for all nodes on nr_reclaimed and we scan (size_on_node >>
> priority) for all nodes, I don't see the reason behind having round
> robin order of node traversal.

It's actually only very recently that we don't bail out of the reclaim
loop anymore - if I'm not missing anything, it was only 1ba6fc9af35b
("mm: vmscan: do not share cgroup iteration between reclaimers") that
removed the last bailout condition on sc->nr_reclaimed.

> I am thinking of removing the whole mem_cgroup_select_victim_node()
> heuristic. Please let me know if there are any objections.

In the current state, I don't see any reason to keep it, either. We
can always just start the zonelist walk from the current node.

A nice cleanup, actually. Good catch!