x86, traps: Fix ist_enter from userspace
authorAndy Lutomirski <luto@amacapital.net>
Sat, 31 Jan 2015 12:53:53 +0000 (04:53 -0800)
committerAndy Lutomirski <luto@amacapital.net>
Sun, 1 Feb 2015 12:02:53 +0000 (04:02 -0800)
context_tracking_user_exit() has no effect if in_interrupt() returns true,
so ist_enter() didn't work.  Fix it by calling exception_enter(), and thus
context_tracking_user_exit(), before incrementing the preempt count.

This also adds an assertion that will catch the problem reliably if
CONFIG_PROVE_RCU=y to help prevent the bug from being reintroduced.

Link: http://lkml.kernel.org/r/261ebee6aee55a4724746d0d7024697013c40a08.1422709102.git.luto@amacapital.net
Fixes: 959274753857 x86, traps: Track entry into and exit from IST context
Reported-and-tested-by: Sasha Levin <sasha.levin@oracle.com>
Signed-off-by: Andy Lutomirski <luto@amacapital.net>
arch/x86/kernel/traps.c

index 7176f84f95a4642c85b88ea0a886d91ccfff3522..c74f2f5652da7ad99f8d53e844c4646088d428e7 100644 (file)
@@ -110,15 +110,11 @@ static inline void preempt_conditional_cli(struct pt_regs *regs)
 
 enum ctx_state ist_enter(struct pt_regs *regs)
 {
-       /*
-        * We are atomic because we're on the IST stack (or we're on x86_32,
-        * in which case we still shouldn't schedule.
-        */
-       preempt_count_add(HARDIRQ_OFFSET);
+       enum ctx_state prev_state;
 
        if (user_mode_vm(regs)) {
                /* Other than that, we're just an exception. */
-               return exception_enter();
+               prev_state = exception_enter();
        } else {
                /*
                 * We might have interrupted pretty much anything.  In
@@ -127,12 +123,27 @@ enum ctx_state ist_enter(struct pt_regs *regs)
                 * but we need to notify RCU.
                 */
                rcu_nmi_enter();
-               return IN_KERNEL;  /* the value is irrelevant. */
+               prev_state = IN_KERNEL;  /* the value is irrelevant. */
        }
+
+       /*
+        * We are atomic because we're on the IST stack (or we're on x86_32,
+        * in which case we still shouldn't schedule).
+        *
+        * This must be after exception_enter(), because exception_enter()
+        * won't do anything if in_interrupt() returns true.
+        */
+       preempt_count_add(HARDIRQ_OFFSET);
+
+       /* This code is a bit fragile.  Test it. */
+       rcu_lockdep_assert(rcu_is_watching(), "ist_enter didn't work");
+
+       return prev_state;
 }
 
 void ist_exit(struct pt_regs *regs, enum ctx_state prev_state)
 {
+       /* Must be before exception_exit. */
        preempt_count_sub(HARDIRQ_OFFSET);
 
        if (user_mode_vm(regs))