path: root/Documentation/RCU/NMI-RCU.txt
diff options
Diffstat (limited to 'Documentation/RCU/NMI-RCU.txt')
1 files changed, 120 insertions, 0 deletions
diff --git a/Documentation/RCU/NMI-RCU.txt b/Documentation/RCU/NMI-RCU.txt
new file mode 100644
index 00000000..687777f8
--- /dev/null
+++ b/Documentation/RCU/NMI-RCU.txt
@@ -0,0 +1,120 @@
+Using RCU to Protect Dynamic NMI Handlers
+Although RCU is usually used to protect read-mostly data structures,
+it is possible to use RCU to provide dynamic non-maskable interrupt
+handlers, as well as dynamic irq handlers. This document describes
+how to do this, drawing loosely from Zwane Mwaikambo's NMI-timer
+work in "arch/x86/oprofile/nmi_timer_int.c" and in
+The relevant pieces of code are listed below, each followed by a
+brief explanation.
+ static int dummy_nmi_callback(struct pt_regs *regs, int cpu)
+ {
+ return 0;
+ }
+The dummy_nmi_callback() function is a "dummy" NMI handler that does
+nothing, but returns zero, thus saying that it did nothing, allowing
+the NMI handler to take the default machine-specific action.
+ static nmi_callback_t nmi_callback = dummy_nmi_callback;
+This nmi_callback variable is a global function pointer to the current
+NMI handler.
+ void do_nmi(struct pt_regs * regs, long error_code)
+ {
+ int cpu;
+ nmi_enter();
+ cpu = smp_processor_id();
+ ++nmi_count(cpu);
+ if (!rcu_dereference_sched(nmi_callback)(regs, cpu))
+ default_do_nmi(regs);
+ nmi_exit();
+ }
+The do_nmi() function processes each NMI. It first disables preemption
+in the same way that a hardware irq would, then increments the per-CPU
+count of NMIs. It then invokes the NMI handler stored in the nmi_callback
+function pointer. If this handler returns zero, do_nmi() invokes the
+default_do_nmi() function to handle a machine-specific NMI. Finally,
+preemption is restored.
+In theory, rcu_dereference_sched() is not needed, since this code runs
+only on i386, which in theory does not need rcu_dereference_sched()
+anyway. However, in practice it is a good documentation aid, particularly
+for anyone attempting to do something similar on Alpha or on systems
+with aggressive optimizing compilers.
+Quick Quiz: Why might the rcu_dereference_sched() be necessary on Alpha,
+ given that the code referenced by the pointer is read-only?
+Back to the discussion of NMI and RCU...
+ void set_nmi_callback(nmi_callback_t callback)
+ {
+ rcu_assign_pointer(nmi_callback, callback);
+ }
+The set_nmi_callback() function registers an NMI handler. Note that any
+data that is to be used by the callback must be initialized up -before-
+the call to set_nmi_callback(). On architectures that do not order
+writes, the rcu_assign_pointer() ensures that the NMI handler sees the
+initialized values.
+ void unset_nmi_callback(void)
+ {
+ rcu_assign_pointer(nmi_callback, dummy_nmi_callback);
+ }
+This function unregisters an NMI handler, restoring the original
+dummy_nmi_handler(). However, there may well be an NMI handler
+currently executing on some other CPU. We therefore cannot free
+up any data structures used by the old NMI handler until execution
+of it completes on all other CPUs.
+One way to accomplish this is via synchronize_sched(), perhaps as
+ unset_nmi_callback();
+ synchronize_sched();
+ kfree(my_nmi_data);
+This works because synchronize_sched() blocks until all CPUs complete
+any preemption-disabled segments of code that they were executing.
+Since NMI handlers disable preemption, synchronize_sched() is guaranteed
+not to return until all ongoing NMI handlers exit. It is therefore safe
+to free up the handler's data as soon as synchronize_sched() returns.
+Important note: for this to work, the architecture in question must
+invoke nmi_enter() and nmi_exit() on NMI entry and exit, respectively.
+Answer to Quick Quiz
+ Why might the rcu_dereference_sched() be necessary on Alpha, given
+ that the code referenced by the pointer is read-only?
+ Answer: The caller to set_nmi_callback() might well have
+ initialized some data that is to be used by the new NMI
+ handler. In this case, the rcu_dereference_sched() would
+ be needed, because otherwise a CPU that received an NMI
+ just after the new handler was set might see the pointer
+ to the new NMI handler, but the old pre-initialized
+ version of the handler's data.
+ This same sad story can happen on other CPUs when using
+ a compiler with aggressive pointer-value speculation
+ optimizations.
+ More important, the rcu_dereference_sched() makes it
+ clear to someone reading the code that the pointer is
+ being protected by RCU-sched.