x86/ldt/64: Refresh DS and ES when modify_ldt changes an entry

On x86_32, modify_ldt() implicitly refreshes the cached DS and ES
segments because they are refreshed on return to usermode.

On x86_64, they're not refreshed on return to usermode.  To improve
determinism and match x86_32's behavior, refresh them when we update
the LDT.

This avoids a situation in which the DS points to a descriptor that is
changed but the old cached segment persists until the next reschedule.
If this happens, then the user-visible state will change
nondeterministically some time after modify_ldt() returns, which is
unfortunate.

Signed-off-by: Andy Lutomirski <luto@kernel.org>
Cc: Borislav Petkov <bpetkov@suse.de>
Cc: Chang Seok <chang.seok.bae@intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@kernel.org>

arch/x86/kernel/ldt.c

@@ -21,6 +21,25 @@
 #include <asm/mmu_context.h>
 #include <asm/syscalls.h>
 
+static void refresh_ldt_segments(void)
+{
+#ifdef CONFIG_X86_64
+	unsigned short sel;
+
+	/*
+	 * Make sure that the cached DS and ES descriptors match the updated
+	 * LDT.
+	 */
+	savesegment(ds, sel);
+	if ((sel & SEGMENT_TI_MASK) == SEGMENT_LDT)
+		loadsegment(ds, sel);
+
+	savesegment(es, sel);
+	if ((sel & SEGMENT_TI_MASK) == SEGMENT_LDT)
+		loadsegment(es, sel);
+#endif
+}
+
 /* context.lock is held for us, so we don't need any locking. */
 static void flush_ldt(void *__mm)
 {
@@ -32,6 +51,8 @@ static void flush_ldt(void *__mm)
 
 	pc = &mm->context;
 	set_ldt(pc->ldt->entries, pc->ldt->nr_entries);
+
+	refresh_ldt_segments();
 }
 
 /* The caller must call finalize_ldt_struct on the result. LDT starts zeroed. */