Re: Memory hotplug regression in 4.13

[Michal Hocko]

On Fri 01-12-17 08:23:27, Seth Forshee wrote:
> On Mon, Sep 25, 2017 at 02:58:25PM +0200, Michal Hocko wrote:
> > On Thu 21-09-17 00:40:34, Seth Forshee wrote:
[...]
> > > It seems I don't have that kernel anymore, but I've got a 4.14-rc1 build
> > > and the problem still occurs there. It's pointing to the call to
> > > __builtin_memcpy in memcpy (include/linux/string.h line 340), which we
> > > get to via wp_page_copy -> cow_user_page -> copy_user_highpage.
> > 
> > Hmm, this is interesting. That would mean that we have successfully
> > mapped the destination page but its memory is still not accessible.
> > 
> > Right now I do not see how the patch you have bisected to could make any
> > difference because it only postponed the onlining to be independent but
> > your config simply onlines automatically so there shouldn't be any
> > semantic change. Maybe there is some sort of off-by-one or something.
> > 
> > I will try to investigate some more. Do you think it would be possible
> > to configure kdump on your system and provide me with the vmcore in some
> > way?
> 
> Sorry, I got busy with other stuff and this kind of fell off my radar.
> It came to my attention again recently though.

Apology on my side. This has completely fall of my radar.

> I was looking through the hotplug rework changes, and I noticed that
> 32-bit x86 previously was using ZONE_HIGHMEM as a default but after the
> rework it doesn't look like it's possible for memory to be associated
> with ZONE_HIGHMEM when onlining. So I made the change below against 4.14
> and am now no longer seeing the oopses.

Thanks a lot for debugging! Do I read the above correctly that the
current code simply returns ZONE_NORMAL and maps an unrelated pfn into
this zone and that leads to later blowups? Could you attach the fresh
boot dmesg output please?

> I'm sure this isn't the correct fix, but I think it does confirm that
> the problem is that the memory should be associated with ZONE_HIGHMEM
> but is not.


Yes, the fix is not quite right. HIGHMEM is not a _kernel_ memory
zone. The kernel cannot access that memory directly. It is essentially a
movable zone from the hotplug API POV. We simply do not have any way to
tell into which zone we want to online this memory range in.
Unfortunately both zones _can_ be present. It would require an explicit
configuration (movable_node and a NUMA hoptlugable nodes running in 32b
or and movable memory configured explicitly on the kernel command line).

The below patch is not really complete but I would rather start simple.
Maybe we do not even have to care as most 32b users will never use both
zones at the same time. I've placed a warning to learn about those.

Does this pass your testing?
---
diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
index 262bfd26baf9..18fec18bdb60 100644
--- a/mm/memory_hotplug.c
+++ b/mm/memory_hotplug.c
@@ -855,12 +855,29 @@ static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn
 	return &pgdat->node_zones[ZONE_NORMAL];
 }
 
+static struct zone *default_movable_zone_for_pfn(int nid)
+{
+	/*
+	 * Please note that 32b HIGHMEM systems might have 2 movable zones
+	 * actually so we have to check for both. This is rather ugly hack
+	 * to enforce using Highmem on those systems but we do not have a
+	 * good user API to tell into which movable zone we should online.
+	 * WARN if we have a movable zone which is not highmem.
+	 */
+#ifdef CONFIG_HIGHMEM
+	WARN_ON_ONCE(!zone_movable_is_highmem());
+	return &NODE_DATA(nid)->node_zones[ZONE_HIGHMEM];
+#else
+	return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
+#endif
+}
+
 static inline struct zone *default_zone_for_pfn(int nid, unsigned long start_pfn,
 		unsigned long nr_pages)
 {
 	struct zone *kernel_zone = default_kernel_zone_for_pfn(nid, start_pfn,
 			nr_pages);
-	struct zone *movable_zone = &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
+	struct zone *movable_zone = default_movable_zone_for_pfn(nid);
 	bool in_kernel = zone_intersects(kernel_zone, start_pfn, nr_pages);
 	bool in_movable = zone_intersects(movable_zone, start_pfn, nr_pages);
 
@@ -886,7 +903,7 @@ struct zone * zone_for_pfn_range(int online_type, int nid, unsigned start_pfn,
 		return default_kernel_zone_for_pfn(nid, start_pfn, nr_pages);
 
 	if (online_type == MMOP_ONLINE_MOVABLE)
-		return &NODE_DATA(nid)->node_zones[ZONE_MOVABLE];
+		return default_movable_zone_for_pfn(nid);
 
 	return default_zone_for_pfn(nid, start_pfn, nr_pages);
 }
-- 
Michal Hocko
SUSE Labs