1	     Semantics and Behavior of Local Atomic Operations
  2
  3			    Mathieu Desnoyers
  4
  5
  6	This document explains the purpose of the local atomic operations, how
  7to implement them for any given architecture and shows how they can be used
  8properly. It also stresses on the precautions that must be taken when reading
  9those local variables across CPUs when the order of memory writes matters.
 10
 11
 12
 13* Purpose of local atomic operations
 14
 15Local atomic operations are meant to provide fast and highly reentrant per CPU
 16counters. They minimize the performance cost of standard atomic operations by
 17removing the LOCK prefix and memory barriers normally required to synchronize
 18across CPUs.
 19
 20Having fast per CPU atomic counters is interesting in many cases : it does not
 21require disabling interrupts to protect from interrupt handlers and it permits
 22coherent counters in NMI handlers. It is especially useful for tracing purposes
 23and for various performance monitoring counters.
 24
 25Local atomic operations only guarantee variable modification atomicity wrt the
 26CPU which owns the data. Therefore, care must taken to make sure that only one
 27CPU writes to the local_t data. This is done by using per cpu data and making
 28sure that we modify it from within a preemption safe context. It is however
 29permitted to read local_t data from any CPU : it will then appear to be written
 30out of order wrt other memory writes by the owner CPU.
 31
 32
 33* Implementation for a given architecture
 34
 35It can be done by slightly modifying the standard atomic operations : only
 36their UP variant must be kept. It typically means removing LOCK prefix (on
 37i386 and x86_64) and any SMP synchronization barrier. If the architecture does
 38not have a different behavior between SMP and UP, including asm-generic/local.h
 39in your architecture's local.h is sufficient.
 40
 41The local_t type is defined as an opaque signed long by embedding an
 42atomic_long_t inside a structure. This is made so a cast from this type to a
 43long fails. The definition looks like :
 44
 45typedef struct { atomic_long_t a; } local_t;
 46
 47
 48* Rules to follow when using local atomic operations
 49
 50- Variables touched by local ops must be per cpu variables.
 51- _Only_ the CPU owner of these variables must write to them.
 52- This CPU can use local ops from any context (process, irq, softirq, nmi, ...)
 53  to update its local_t variables.
 54- Preemption (or interrupts) must be disabled when using local ops in
 55  process context to   make sure the process won't be migrated to a
 56  different CPU between getting the per-cpu variable and doing the
 57  actual local op.
 58- When using local ops in interrupt context, no special care must be
 59  taken on a mainline kernel, since they will run on the local CPU with
 60  preemption already disabled. I suggest, however, to explicitly
 61  disable preemption anyway to make sure it will still work correctly on
 62  -rt kernels.
 63- Reading the local cpu variable will provide the current copy of the
 64  variable.
 65- Reads of these variables can be done from any CPU, because updates to
 66  "long", aligned, variables are always atomic. Since no memory
 67  synchronization is done by the writer CPU, an outdated copy of the
 68  variable can be read when reading some _other_ cpu's variables.
 69
 70
 71* How to use local atomic operations
 72
 73#include <linux/percpu.h>
 74#include <asm/local.h>
 75
 76static DEFINE_PER_CPU(local_t, counters) = LOCAL_INIT(0);
 77
 78
 79* Counting
 80
 81Counting is done on all the bits of a signed long.
 82
 83In preemptible context, use get_cpu_var() and put_cpu_var() around local atomic
 84operations : it makes sure that preemption is disabled around write access to
 85the per cpu variable. For instance :
 86
 87	local_inc(&get_cpu_var(counters));
 88	put_cpu_var(counters);
 89
 90If you are already in a preemption-safe context, you can directly use
 91__get_cpu_var() instead.
 92
 93	local_inc(&__get_cpu_var(counters));
 94
 95
 96
 97* Reading the counters
 98
 99Those local counters can be read from foreign CPUs to sum the count. Note that
100the data seen by local_read across CPUs must be considered to be out of order
101relatively to other memory writes happening on the CPU that owns the data.
102
103	long sum = 0;
104	for_each_online_cpu(cpu)
105		sum += local_read(&per_cpu(counters, cpu));
106
107If you want to use a remote local_read to synchronize access to a resource
108between CPUs, explicit smp_wmb() and smp_rmb() memory barriers must be used
109respectively on the writer and the reader CPUs. It would be the case if you use
110the local_t variable as a counter of bytes written in a buffer : there should
111be a smp_wmb() between the buffer write and the counter increment and also a
112smp_rmb() between the counter read and the buffer read.
113
114
115Here is a sample module which implements a basic per cpu counter using local.h.
116
117--- BEGIN ---
118/* test-local.c
119 *
120 * Sample module for local.h usage.
121 */
122
123
124#include <asm/local.h>
125#include <linux/module.h>
126#include <linux/timer.h>
127
128static DEFINE_PER_CPU(local_t, counters) = LOCAL_INIT(0);
129
130static struct timer_list test_timer;
131
132/* IPI called on each CPU. */
133static void test_each(void *info)
134{
135	/* Increment the counter from a non preemptible context */
136	printk("Increment on cpu %d\n", smp_processor_id());
137	local_inc(&__get_cpu_var(counters));
138
139	/* This is what incrementing the variable would look like within a
140	 * preemptible context (it disables preemption) :
141	 *
142	 * local_inc(&get_cpu_var(counters));
143	 * put_cpu_var(counters);
144	 */
145}
146
147static void do_test_timer(unsigned long data)
148{
149	int cpu;
150
151	/* Increment the counters */
152	on_each_cpu(test_each, NULL, 1);
153	/* Read all the counters */
154	printk("Counters read from CPU %d\n", smp_processor_id());
155	for_each_online_cpu(cpu) {
156		printk("Read : CPU %d, count %ld\n", cpu,
157			local_read(&per_cpu(counters, cpu)));
158	}
159	del_timer(&test_timer);
160	test_timer.expires = jiffies + 1000;
161	add_timer(&test_timer);
162}
163
164static int __init test_init(void)
165{
166	/* initialize the timer that will increment the counter */
167	init_timer(&test_timer);
168	test_timer.function = do_test_timer;
169	test_timer.expires = jiffies + 1;
170	add_timer(&test_timer);
171
172	return 0;
173}
174
175static void __exit test_exit(void)
176{
177	del_timer_sync(&test_timer);
178}
179
180module_init(test_init);
181module_exit(test_exit);
182
183MODULE_LICENSE("GPL");
184MODULE_AUTHOR("Mathieu Desnoyers");
185MODULE_DESCRIPTION("Local Atomic Ops");
186--- END ---