[PATCH STABLE 4.9] x86, mm, gup: prevent get_page() race with munmap in paravirt guest

[Vlastimil Babka]

The x86 version of get_user_pages_fast() relies on disabled interrupts to
synchronize gup_pte_range() between gup_get_pte(ptep); and get_page() against
a parallel munmap. The munmap side nulls the pte, then flushes TLBs, then
releases the page. As TLB flush is done synchronously via IPI disabling
interrupts blocks the page release, and get_page(), which assumes existing
reference on page, is thus safe.
However when TLB flush is done by a hypercall, e.g. in a Xen PV guest, there is
no blocking thanks to disabled interrupts, and get_page() can succeed on a page
that was already freed or even reused.

We have recently seen this happen with our 4.4 and 4.12 based kernels, with
userspace (java) that exits a thread, where mm_release() performs a futex_wake()
on tsk->clear_child_tid, and another thread in parallel unmaps the page where
tsk->clear_child_tid points to. The spurious get_page() succeeds, but futex code
immediately releases the page again, while it's already on a freelist. Symptoms
include a bad page state warning, general protection faults acessing a poisoned
list prev/next pointer in the freelist, or free page pcplists of two cpus joined
together in a single list. Oscar has also reproduced this scenario, with a
patch inserting delays before the get_page() to make the race window larger.

Fix this by removing the dependency on TLB flush interrupts the same way as the
generic get_user_pages_fast() code by using page_cache_add_speculative() and
revalidating the PTE contents after pinning the page. Mainline is safe since
4.13 where the x86 gup code was removed in favor of the common code. Accessing
the page table itself safely also relies on disabled interrupts and TLB flush
IPIs that don't happen with hypercalls, which was acknowledged in commit
9e52fc2b50de ("x86/mm: Enable RCU based page table freeing
(CONFIG_HAVE_RCU_TABLE_FREE=y)"). That commit with follups should also be
backported for full safety, although our reproducer didn't hit a problem
without that backport.

Reproduced-by: Oscar Salvador <osalvador@xxxxxxx>
Signed-off-by: Vlastimil Babka <vbabka@xxxxxxx>
Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
Cc: Ingo Molnar <mingo@xxxxxxxxxx>
Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Cc: Juergen Gross <jgross@xxxxxxxx>
Cc: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
Cc: Vitaly Kuznetsov <vkuznets@xxxxxxxxxx>
Cc: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
Cc: Borislav Petkov <bp@xxxxxxxxx>
Cc: Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx>
Cc: Andy Lutomirski <luto@xxxxxxxxxx>
---

Hi, I'm sending this stable-only patch for consideration because it's probably
unrealistic to backport the 4.13 switch to generic GUP. I can look at 4.4 and
3.16 if accepted. The RCU page table freeing could be also considered.
Note the patch also includes page refcount protection. I found out that
8fde12ca79af ("mm: prevent get_user_pages() from overflowing page refcount")
backport to 4.9 missed the arch-specific gup implementations:
https://lore.kernel.org/lkml/6650323f-dbc9-f069-000b-f6b0f941a065@xxxxxxx/

 arch/x86/mm/gup.c | 32 ++++++++++++++++++++++++++++++--
 1 file changed, 30 insertions(+), 2 deletions(-)

diff --git a/arch/x86/mm/gup.c b/arch/x86/mm/gup.c
index 1680768d392c..d7db45bdfb3b 100644
--- a/arch/x86/mm/gup.c
+++ b/arch/x86/mm/gup.c
@@ -97,6 +97,20 @@ static inline int pte_allows_gup(unsigned long pteval, int write)
 	return 1;
 }
 
+/*
+ * Return the compund head page with ref appropriately incremented,
+ * or NULL if that failed.
+ */
+static inline struct page *try_get_compound_head(struct page *page, int refs)
+{
+	struct page *head = compound_head(page);
+	if (WARN_ON_ONCE(page_ref_count(head) < 0))
+		return NULL;
+	if (unlikely(!page_cache_add_speculative(head, refs)))
+		return NULL;
+	return head;
+}
+
 /*
  * The performance critical leaf functions are made noinline otherwise gcc
  * inlines everything into a single function which results in too much
@@ -112,7 +126,7 @@ static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
 	ptep = pte_offset_map(&pmd, addr);
 	do {
 		pte_t pte = gup_get_pte(ptep);
-		struct page *page;
+		struct page *head, *page;
 
 		/* Similar to the PMD case, NUMA hinting must take slow path */
 		if (pte_protnone(pte)) {
@@ -138,7 +152,21 @@ static noinline int gup_pte_range(pmd_t pmd, unsigned long addr,
 		}
 		VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
 		page = pte_page(pte);
-		get_page(page);
+
+		head = try_get_compound_head(page, 1);
+		if (!head) {
+			put_dev_pagemap(pgmap);
+			pte_unmap(ptep);
+			return 0;
+		}
+
+		if (unlikely(pte_val(pte) != pte_val(*ptep))) {
+			put_page(head);
+			put_dev_pagemap(pgmap);
+			pte_unmap(ptep);
+			return 0;
+		}
+
 		put_dev_pagemap(pgmap);
 		SetPageReferenced(page);
 		pages[*nr] = page;
-- 
2.22.0