kaiser: do not set _PAGE_NX on pgd_none

native_pgd_clear() uses native_set_pgd(), so native_set_pgd() must
avoid setting the _PAGE_NX bit on an otherwise pgd_none() entry:
usually that just generated a warning on exit, but sometimes
more mysterious and damaging failures (our production machines
could not complete booting).

The original fix to this just avoided adding _PAGE_NX to
an empty entry; but eventually more problems surfaced with kexec,
and EFI mapping expected to be a problem too.  So now instead
change native_set_pgd() to update shadow only if _PAGE_USER:

A few places (kernel/machine_kexec_64.c, platform/efi/efi_64.c for sure)
use set_pgd() to set up a temporary internal virtual address space, with
physical pages remapped at what Kaiser regards as userspace addresses:
Kaiser then assumes a shadow pgd follows, which it will try to corrupt.

This appears to be responsible for the recent kexec and kdump failures;
though it's unclear how those did not manifest as a problem before.
Ah, the shadow pgd will only be assumed to "follow" if the requested
pgd is on an even-numbered page: so I suppose it was going wrong 50%
of the time all along.

What we need is a flag to set_pgd(), to tell it we're dealing with
userspace.  Er, isn't that what the pgd's _PAGE_USER bit is saying?
Add a test for that.  But we cannot do the same for pgd_clear()
(which may be called to clear corrupted entries - set aside the
question of "corrupt in which pgd?" until later), so there just
rely on pgd_clear() not being called in the problematic cases -
with a WARN_ON_ONCE() which should fire half the time if it is.

But this is getting too big for an inline function: move it into
arch/x86/mm/kaiser.c (which then demands a boot/compressed mod);
and de-void and de-space native_get_shadow/normal_pgd() while here.

Also make an unnecessary change to KASLR's init_trampoline(): it was
using set_pgd() to assign a pgd-value to a global variable (not in a
pg directory page), which was rather scary given Kaiser's previous
set_pgd() implementation: not a problem now, but too scary to leave
as was, it could easily blow up if we have to change set_pgd() again.

Signed-off-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

4 files changed, 58 insertions, 40 deletions

diff --git a/arch/x86/boot/compressed/misc.h b/arch/x86/boot/compressed/misc.h
index 766a5211f827..80e2cce110b9 100644
--- a/arch/x86/boot/compressed/misc.h
+++ b/arch/x86/boot/compressed/misc.h
@@ -9,6 +9,7 @@
  */
 #undef CONFIG_PARAVIRT
 #undef CONFIG_PARAVIRT_SPINLOCKS
+#undef CONFIG_KAISER
 #undef CONFIG_KASAN
 
 #include <linux/linkage.h>
diff --git a/arch/x86/include/asm/pgtable_64.h b/arch/x86/include/asm/pgtable_64.h
index 000265c8b74e..177caf3f7ab1 100644
--- a/arch/x86/include/asm/pgtable_64.h
+++ b/arch/x86/include/asm/pgtable_64.h
@@ -107,61 +107,36 @@ static inline void native_pud_clear(pud_t *pud)
 }
 
 #ifdef CONFIG_KAISER
-static inline pgd_t * native_get_shadow_pgd(pgd_t *pgdp)
+extern pgd_t kaiser_set_shadow_pgd(pgd_t *pgdp, pgd_t pgd);
+
+static inline pgd_t *native_get_shadow_pgd(pgd_t *pgdp)
 {
-	return (pgd_t *)(void*)((unsigned long)(void*)pgdp | (unsigned long)PAGE_SIZE);
+	return (pgd_t *)((unsigned long)pgdp | (unsigned long)PAGE_SIZE);
 }
 
-static inline pgd_t * native_get_normal_pgd(pgd_t *pgdp)
+static inline pgd_t *native_get_normal_pgd(pgd_t *pgdp)
 {
-	return (pgd_t *)(void*)((unsigned long)(void*)pgdp &  ~(unsigned long)PAGE_SIZE);
+	return (pgd_t *)((unsigned long)pgdp & ~(unsigned long)PAGE_SIZE);
 }
 #else
-static inline pgd_t * native_get_shadow_pgd(pgd_t *pgdp)
+static inline pgd_t kaiser_set_shadow_pgd(pgd_t *pgdp, pgd_t pgd)
+{
+	return pgd;
+}
+static inline pgd_t *native_get_shadow_pgd(pgd_t *pgdp)
 {
 	BUILD_BUG_ON(1);
 	return NULL;
 }
-static inline pgd_t * native_get_normal_pgd(pgd_t *pgdp)
+static inline pgd_t *native_get_normal_pgd(pgd_t *pgdp)
 {
 	return pgdp;
 }
 #endif /* CONFIG_KAISER */
 
-/*
- * Page table pages are page-aligned.  The lower half of the top
- * level is used for userspace and the top half for the kernel.
- * This returns true for user pages that need to get copied into
- * both the user and kernel copies of the page tables, and false
- * for kernel pages that should only be in the kernel copy.
- */
-static inline bool is_userspace_pgd(void *__ptr)
-{
-	unsigned long ptr = (unsigned long)__ptr;
-
-	return ((ptr % PAGE_SIZE) < (PAGE_SIZE / 2));
-}
-
 static inline void native_set_pgd(pgd_t *pgdp, pgd_t pgd)
 {
-#ifdef CONFIG_KAISER
-	pteval_t extra_kern_pgd_flags = 0;
-	/* Do we need to also populate the shadow pgd? */
-	if (is_userspace_pgd(pgdp)) {
-		native_get_shadow_pgd(pgdp)->pgd = pgd.pgd;
-		/*
-		 * Even if the entry is *mapping* userspace, ensure
-		 * that userspace can not use it.  This way, if we
-		 * get out to userspace running on the kernel CR3,
-		 * userspace will crash instead of running.
-		 */
-		extra_kern_pgd_flags = _PAGE_NX;
-	}
-	pgdp->pgd = pgd.pgd;
-	pgdp->pgd |= extra_kern_pgd_flags;
-#else /* CONFIG_KAISER */
-	*pgdp = pgd;
-#endif
+	*pgdp = kaiser_set_shadow_pgd(pgdp, pgd);
 }
 
 static inline void native_pgd_clear(pgd_t *pgd)
diff --git a/arch/x86/mm/kaiser.c b/arch/x86/mm/kaiser.c
index 7270a293175d..8d6061c2bc79 100644
--- a/arch/x86/mm/kaiser.c
+++ b/arch/x86/mm/kaiser.c
@@ -302,4 +302,46 @@ void kaiser_remove_mapping(unsigned long start, unsigned long size)
 		unmap_pud_range_nofree(pgd, addr, end);
 	}
 }
+
+/*
+ * Page table pages are page-aligned.  The lower half of the top
+ * level is used for userspace and the top half for the kernel.
+ * This returns true for user pages that need to get copied into
+ * both the user and kernel copies of the page tables, and false
+ * for kernel pages that should only be in the kernel copy.
+ */
+static inline bool is_userspace_pgd(pgd_t *pgdp)
+{
+	return ((unsigned long)pgdp % PAGE_SIZE) < (PAGE_SIZE / 2);
+}
+
+pgd_t kaiser_set_shadow_pgd(pgd_t *pgdp, pgd_t pgd)
+{
+	/*
+	 * Do we need to also populate the shadow pgd?  Check _PAGE_USER to
+	 * skip cases like kexec and EFI which make temporary low mappings.
+	 */
+	if (pgd.pgd & _PAGE_USER) {
+		if (is_userspace_pgd(pgdp)) {
+			native_get_shadow_pgd(pgdp)->pgd = pgd.pgd;
+			/*
+			 * Even if the entry is *mapping* userspace, ensure
+			 * that userspace can not use it.  This way, if we
+			 * get out to userspace running on the kernel CR3,
+			 * userspace will crash instead of running.
+			 */
+			pgd.pgd |= _PAGE_NX;
+		}
+	} else if (!pgd.pgd) {
+		/*
+		 * pgd_clear() cannot check _PAGE_USER, and is even used to
+		 * clear corrupted pgd entries: so just rely on cases like
+		 * kexec and EFI never to be using pgd_clear().
+		 */
+		if (!WARN_ON_ONCE((unsigned long)pgdp & PAGE_SIZE) &&
+		    is_userspace_pgd(pgdp))
+			native_get_shadow_pgd(pgdp)->pgd = pgd.pgd;
+	}
+	return pgd;
+}
 #endif /* CONFIG_KAISER */
diff --git a/arch/x86/mm/kaslr.c b/arch/x86/mm/kaslr.c
index aed206475aa7..9284ec1e8aa9 100644
--- a/arch/x86/mm/kaslr.c
+++ b/arch/x86/mm/kaslr.c
@@ -189,6 +189,6 @@ void __meminit init_trampoline(void)
 		*pud_tramp = *pud;
 	}
 
-	set_pgd(&trampoline_pgd_entry,
-		__pgd(_KERNPG_TABLE | __pa(pud_page_tramp)));
+	/* Avoid set_pgd(), in case it's complicated by CONFIG_KAISER */
+	trampoline_pgd_entry = __pgd(_KERNPG_TABLE | __pa(pud_page_tramp));
 }