[PATCH] x86/mm: Set NX bit when making pages present

[Dave Hansen <dave.hansen@linux.intel.com>]

The x86 mm code now actively refuses to create writable, executable
mappings and warns when there is an attempt to create one.

0day ran across a case triggered by module unloading, but that looks
to be a generic problem.  It presumably goes like this:

	1. Load module with direct map, P=1,W=1,NX=1
	2. Map module executable, set P=1,W=0,NX=0
	3. Free module, land in vfree()->vm_remove_mappings()
	4. Set P=0 during alias processing, P=0,W=0,NX=0
	5. Restore kernel mapping via set_direct_map_default_noflush(),
	   set P=1,W=1, resulting in P=1,W=1,NX=0

That's clearly a writable, executable mapping which is a no-no.  The
new W^X code is clearly doing its job.

Fix it by actively setting _PAGE_NX when creating writable mappings.

One concern: I haven't been able to actually reproduce this, even by
loading and unloading the module that 0day hit it with.  I'd like to
be able to reproduce this before committing a fix.

Reported-by: kernel test robot <yujie.liu@intel.com>
Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Borislav Petkov <bp@alien8.de>
Cc: x86@kernel.org
Cc: "H. Peter Anvin" <hpa@zytor.com>
Link: https://lore.kernel.org/all/fcf89147-440b-e478-40c9-228c9fe56691@intel.com/

--

0day folks, please do share these as they come up.  We want to keep
fixing them.
---
 arch/x86/mm/pat/set_memory.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/arch/x86/mm/pat/set_memory.c b/arch/x86/mm/pat/set_memory.c
index 1a2d6376251c..5fb5874ea2c6 100644
--- a/arch/x86/mm/pat/set_memory.c
+++ b/arch/x86/mm/pat/set_memory.c
@@ -2247,6 +2247,12 @@ static int __set_pages_p(struct page *page, int numpages)
 				.mask_clr = __pgprot(0),
 				.flags = 0};
 
+	/*
+	 * Avoid W^X mappings that occur if the old
+	 * mapping was !_PAGE_RW and !_PAGE_NX.
+	 */
+	pgprot_val(cpa.mask_set) |= __supported_pte_mask & _PAGE_NX;
+
 	/*
 	 * No alias checking needed for setting present flag. otherwise,
 	 * we may need to break large pages for 64-bit kernel text
-- 
2.34.1