1Using RCU to Protect Dynamic NMI Handlers
  2
  3
  4Although RCU is usually used to protect read-mostly data structures,
  5it is possible to use RCU to provide dynamic non-maskable interrupt
  6handlers, as well as dynamic irq handlers.  This document describes
  7how to do this, drawing loosely from Zwane Mwaikambo's NMI-timer
  8work in "arch/x86/oprofile/nmi_timer_int.c" and in
  9"arch/x86/kernel/traps.c".
 10
 11The relevant pieces of code are listed below, each followed by a
 12brief explanation.
 13
 14	static int dummy_nmi_callback(struct pt_regs *regs, int cpu)
 15	{
 16		return 0;
 17	}
 18
 19The dummy_nmi_callback() function is a "dummy" NMI handler that does
 20nothing, but returns zero, thus saying that it did nothing, allowing
 21the NMI handler to take the default machine-specific action.
 22
 23	static nmi_callback_t nmi_callback = dummy_nmi_callback;
 24
 25This nmi_callback variable is a global function pointer to the current
 26NMI handler.
 27
 28	void do_nmi(struct pt_regs * regs, long error_code)
 29	{
 30		int cpu;
 31
 32		nmi_enter();
 33
 34		cpu = smp_processor_id();
 35		++nmi_count(cpu);
 36
 37		if (!rcu_dereference_sched(nmi_callback)(regs, cpu))
 38			default_do_nmi(regs);
 39
 40		nmi_exit();
 41	}
 42
 43The do_nmi() function processes each NMI.  It first disables preemption
 44in the same way that a hardware irq would, then increments the per-CPU
 45count of NMIs.  It then invokes the NMI handler stored in the nmi_callback
 46function pointer.  If this handler returns zero, do_nmi() invokes the
 47default_do_nmi() function to handle a machine-specific NMI.  Finally,
 48preemption is restored.
 49
 50In theory, rcu_dereference_sched() is not needed, since this code runs
 51only on i386, which in theory does not need rcu_dereference_sched()
 52anyway.  However, in practice it is a good documentation aid, particularly
 53for anyone attempting to do something similar on Alpha or on systems
 54with aggressive optimizing compilers.
 55
 56Quick Quiz:  Why might the rcu_dereference_sched() be necessary on Alpha,
 57	     given that the code referenced by the pointer is read-only?
 58
 59
 60Back to the discussion of NMI and RCU...
 61
 62	void set_nmi_callback(nmi_callback_t callback)
 63	{
 64		rcu_assign_pointer(nmi_callback, callback);
 65	}
 66
 67The set_nmi_callback() function registers an NMI handler.  Note that any
 68data that is to be used by the callback must be initialized up -before-
 69the call to set_nmi_callback().  On architectures that do not order
 70writes, the rcu_assign_pointer() ensures that the NMI handler sees the
 71initialized values.
 72
 73	void unset_nmi_callback(void)
 74	{
 75		rcu_assign_pointer(nmi_callback, dummy_nmi_callback);
 76	}
 77
 78This function unregisters an NMI handler, restoring the original
 79dummy_nmi_handler().  However, there may well be an NMI handler
 80currently executing on some other CPU.  We therefore cannot free
 81up any data structures used by the old NMI handler until execution
 82of it completes on all other CPUs.
 83
 84One way to accomplish this is via synchronize_sched(), perhaps as
 85follows:
 86
 87	unset_nmi_callback();
 88	synchronize_sched();
 89	kfree(my_nmi_data);
 90
 91This works because synchronize_sched() blocks until all CPUs complete
 92any preemption-disabled segments of code that they were executing.
 93Since NMI handlers disable preemption, synchronize_sched() is guaranteed
 94not to return until all ongoing NMI handlers exit.  It is therefore safe
 95to free up the handler's data as soon as synchronize_sched() returns.
 96
 97Important note: for this to work, the architecture in question must
 98invoke nmi_enter() and nmi_exit() on NMI entry and exit, respectively.
 99
100
101Answer to Quick Quiz
102
103	Why might the rcu_dereference_sched() be necessary on Alpha, given
104	that the code referenced by the pointer is read-only?
105
106	Answer: The caller to set_nmi_callback() might well have
107		initialized some data that is to be used by the new NMI
108		handler.  In this case, the rcu_dereference_sched() would
109		be needed, because otherwise a CPU that received an NMI
110		just after the new handler was set might see the pointer
111		to the new NMI handler, but the old pre-initialized
112		version of the handler's data.
113
114		This same sad story can happen on other CPUs when using
115		a compiler with aggressive pointer-value speculation
116		optimizations.
117
118		More important, the rcu_dereference_sched() makes it
119		clear to someone reading the code that the pointer is
120		being protected by RCU-sched.