mm/hugetlb_vmemmap: map the shared tail page writable

page_ref_add_unless_zero() reads the refcount and then updates it with a
cmpxchg.  A lockless get, such as a page cache lookup or a PFN scanner,
can read a non-zero refcount from a page that is then freed and reused as
a hugetlb tail page.  If HVO remaps its struct page to the shared tail
page in between, the cmpxchg writes to read-only vmemmap:

    CPU0 (lockless get)               CPU1
    c = atomic_read(&page->_refcount)
                                      page freed
                                      page allocated as a hugetlb tail
                                      HVO remaps the struct page RO
    try_cmpxchg(&page->_refcount, &c, c + 1)
      -> write fault on read-only vmemmap

The refcount is zero by then and the compare fails, but a compare-and-swap
instruction needs write access whether or not it succeeds.  The kernel
oopses when hugetlb pages are allocated at runtime with HVO enabled while
a lockless get races with it.

Commit bd225530a4 ("mm/hugetlb_vmemmap: fix race with speculative PFN
walkers") closed the window with RCU on both sides.  The RCU was removed
together with fake heads, on the wrong belief that it was only needed for
them.

Map the shared tail page writable instead of restoring the RCU.  Every
struct page on it is a tail with a zero refcount, so a late cmpxchg
fails and leaves the page unchanged.  RCU would put rcu_read_lock()
back around every get and synchronize_rcu() back into hugetlb
allocation and freeing.

The cost is the protection against stray writes to tail struct pages.  The
shared page is shared by all hugetlb folios of the same order in the zone,
so such a write now corrupts all of them instead of oopsing.

Bootmem folios optimized when the memmap is set up keep their tails
read-only.  Nothing does a lockless get on them before hugetlb takes them
over.

Link: https://lore.kernel.org/20261028132623.2560146-1-kirill@shutemov.name
Fixes: 32c440d67e ("mm: drop fake head checks")
Fixes: 01b1d0ffb6 ("hugetlb: remove VMEMMAP_SYNCHRONIZE_RCU")
Signed-off-by: Kiryl Shutsemau (Meta) <kas@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reported-by: David Hildenbrand (Arm) <david@kernel.org>
Closes: https://lore.kernel.org/r/ebc7bde1-f52e-4da8-b16f-822d6c0a4d74@kernel.org
Acked-by: Muchun Song <muchun.song@linux.dev>
Reviewed-by: Oscar Salvador <osalvador@suse.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Frank van der Linden <fvdl@google.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Lorenzo Stoakes <ljs@kernel.org>
Cc: Liam R. Howlett <liam@infradead.org>
Cc: Vlastimil Babka <vbabka@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Shuah Khan <skhan@linuxfoundation.org>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: <stable@vger.kernel.org>

Documentation/mm/vmemmap_dedup.rst

@@ -130,8 +130,19 @@ pages of ``struct page`` (page 1 to page 7) are tail pages.
 The optimization is only applied when the size of the struct page is a power
 of 2. In this case, all tail pages of the same order are identical. See
 compound_head(). This allows us to remap the tail pages of the vmemmap to a
-shared, read-only page. The head page is also remapped to a new page. This
-allows the original vmemmap pages to be freed.
+shared page. The head page is also remapped to a new page. This allows the
+original vmemmap pages to be freed.
+
+The shared page is mapped writable. A lockless get of a page reference can
+read a non-zero refcount before the remap and attempt its atomic update
+after it. The refcount of a tail page is zero, so the update fails without
+changing the page. Nothing else may write to the shared page: it is shared
+by all HugeTLB pages of the same size in the zone, and a stray write would
+corrupt all of them.
+
+HugeTLB pages whose vmemmap is optimized when the memmap is first set up at
+boot keep a read-only mapping of the shared page. Nothing can do a lockless
+get on them before they are handed to HugeTLB.
 
 Here is how things look after remapping::
 

mm/hugetlb_vmemmap.c

@@ -222,10 +222,12 @@ static void vmemmap_remap_pte(pte_t *pte, unsigned long addr,
 		VM_WARN_ON_ONCE(!PageTail((const struct page *)addr));
 
 		/*
-		 * Remap the tail pages as read-only to catch illegal write
-		 * operation to the tail pages.
+		 * Map the shared tail page writable. A lockless get that read
+		 * a non-zero refcount before the remap can still do its
+		 * cmpxchg after it. The refcount is zero, so the cmpxchg
+		 * fails and leaves the page unchanged.
 		 */
-		entry = mk_pte(walk->vmemmap_tail, PAGE_KERNEL_RO);
+		entry = mk_pte(walk->vmemmap_tail, PAGE_KERNEL);
 	}
 
 	list_add(&page->lru, walk->vmemmap_pages);