[PATCH] mm/sparse: Fix race on mem_section->usage in pfn walkers

[Muchun Song <songmuchun@bytedance.com>]

When memory is hot-removed, section_deactivate() can tear down
mem_section->usage while concurrent pfn walkers still inspect the
subsection map via pfn_section_valid() or pfn_section_first_valid().

After commit 5ec8e8ea8b77 ("mm/sparsemem: fix race in accessing
memory_section->usage") converted the teardown to an RCU-based
scheme, the code still relies on SECTION_HAS_MEM_MAP becoming visible
to readers before ms->usage is cleared and queued for freeing.

That ordering is not guaranteed. section_deactivate() can clear
ms->usage and queue kfree_rcu() before another CPU observes the
SECTION_HAS_MEM_MAP clear. A concurrent pfn walker can therefore see
valid_section() return true, enter its sched-RCU read-side critical
section after kfree_rcu() has already been queued, and then dereference
a stale ms->usage pointer.

And pfn_to_online_page() can call pfn_section_valid() without its
own sched-RCU read-side critical section, which has similar problem.

The race looks like this:

  compact_zone()                    memunmap_pages
  ==============                    ==============
                                    __remove_pages()->
                                      sparse_remove_section()->
                                        section_deactivate():
                                          a) [ Clear SECTION_HAS_MEM_MAP
                                               is reordered to b) ]
                                          kfree_rcu(ms->usage)
      __pageblock_pfn_to_page
         ......
          pfn_valid():
            rcu_read_lock_sched()
            valid_section() // return true
            pfn_section_valid()
              [Access ms->usage which is UAF]
                                          WRITE_ONCE(ms->usage, NULL)
            rcu_read_unlock_sched()       b) Clear SECTION_HAS_MEM_MAP

Fix this by using rcu_replace_pointer() when clearing ms->usage in
section_deactivate(), then it does not rely on the order of clearing
of SECTION_HAS_MEM_MAP.

Fixes: 5ec8e8ea8b77 ("mm/sparsemem: fix race in accessing memory_section->usage")
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
---
This patch is focused on the ms->usage lifetime race only.

One open question is the interaction between pfn_to_online_page() and
vmemmap teardown during memory hot-remove.

Could pfn_to_online_page() still hand out a stale struct page here?

The new sched-RCU critical section ends before pfn_to_page(pfn), but
section_deactivate() can still tear the vmemmap down immediately afterwards:

mm/sparse-vmemmap.c:section_deactivate()
    ms->section_mem_map &= ~SECTION_HAS_MEM_MAP;
    usage = rcu_replace_pointer(ms->usage, NULL, true);
    kfree_rcu(usage, rcu);
    depopulate_section_memmap(...);

That looks like a reader can observe valid = true, drop sched-RCU, race with
section_deactivate(), and then execute pfn_to_page(pfn) after the backing
vmemmap was depopulated.

Callers such as mm/compaction.c:__reset_isolation_pfn(),
mm/page_idle.c:page_idle_get_folio(), and fs/proc/kcore.c:read_kcore_iter()
dereference the returned page immediately, and they do not appear to hold
get_online_mems() across the pfn_to_online_page() call.

I am not fully sure whether that reasoning is correct, or whether current
callers are expected to rely on additional hotplug serialization instead.
Comments on whether this is a real issue, and how the vmemmap lifetime is
expected to be handled here, would be very helpful.
---
 include/linux/mmzone.h | 6 +++---
 mm/memory_hotplug.c    | 6 +++++-
 mm/sparse-vmemmap.c    | 6 ++++--
 3 files changed, 12 insertions(+), 6 deletions(-)

diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
index 238bf2d35a54..0e850924cbeb 100644
--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h
@@ -2014,7 +2014,7 @@ struct mem_section {
 	 */
 	unsigned long section_mem_map;
 
-	struct mem_section_usage *usage;
+	struct mem_section_usage __rcu *usage;
 #ifdef CONFIG_PAGE_EXTENSION
 	/*
 	 * If SPARSEMEM, pgdat doesn't have page_ext pointer. We use
@@ -2178,14 +2178,14 @@ static inline int subsection_map_index(unsigned long pfn)
 static inline int pfn_section_valid(struct mem_section *ms, unsigned long pfn)
 {
 	int idx = subsection_map_index(pfn);
-	struct mem_section_usage *usage = READ_ONCE(ms->usage);
+	struct mem_section_usage *usage = rcu_dereference_sched(ms->usage);
 
 	return usage ? test_bit(idx, usage->subsection_map) : 0;
 }
 
 static inline bool pfn_section_first_valid(struct mem_section *ms, unsigned long *pfn)
 {
-	struct mem_section_usage *usage = READ_ONCE(ms->usage);
+	struct mem_section_usage *usage = rcu_dereference_sched(ms->usage);
 	int idx = subsection_map_index(*pfn);
 	unsigned long bit;
 
diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
index 0c1d3df3a296..335835abe74c 100644
--- a/mm/memory_hotplug.c
+++ b/mm/memory_hotplug.c
@@ -340,6 +340,7 @@ struct page *pfn_to_online_page(unsigned long pfn)
 	unsigned long nr = pfn_to_section_nr(pfn);
 	struct dev_pagemap *pgmap;
 	struct mem_section *ms;
+	bool valid;
 
 	if (nr >= NR_MEM_SECTIONS)
 		return NULL;
@@ -355,7 +356,10 @@ struct page *pfn_to_online_page(unsigned long pfn)
 	if (IS_ENABLED(CONFIG_HAVE_ARCH_PFN_VALID) && !pfn_valid(pfn))
 		return NULL;
 
-	if (!pfn_section_valid(ms, pfn))
+	rcu_read_lock_sched();
+	valid = pfn_section_valid(ms, pfn);
+	rcu_read_unlock_sched();
+	if (!valid)
 		return NULL;
 
 	if (!online_device_section(ms))
diff --git a/mm/sparse-vmemmap.c b/mm/sparse-vmemmap.c
index 717ac953bba2..05f68dcec0f8 100644
--- a/mm/sparse-vmemmap.c
+++ b/mm/sparse-vmemmap.c
@@ -601,8 +601,10 @@ static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
 		 * was allocated during boot.
 		 */
 		if (!PageReserved(virt_to_page(ms->usage))) {
-			kfree_rcu(ms->usage, rcu);
-			WRITE_ONCE(ms->usage, NULL);
+			struct mem_section_usage *usage;
+
+			usage = rcu_replace_pointer(ms->usage, NULL, true);
+			kfree_rcu(usage, rcu);
 		}
 		memmap = pfn_to_page(SECTION_ALIGN_DOWN(pfn));
 	}
-- 
2.20.1