4 * @remark Copyright 2004 Oprofile Authors
5 * @remark Copyright 2010 ARM Ltd.
6 * @remark Read the file COPYING
8 * @author Zwane Mwaikambo
9 * @author Will Deacon [move to perf]
12 #include <linux/cpumask.h>
13 #include <linux/err.h>
14 #include <linux/errno.h>
15 #include <linux/init.h>
16 #include <linux/mutex.h>
17 #include <linux/oprofile.h>
18 #include <linux/perf_event.h>
19 #include <linux/platform_device.h>
20 #include <linux/slab.h>
21 #include <asm/stacktrace.h>
22 #include <linux/uaccess.h>
24 #include <asm/perf_event.h>
25 #include <asm/ptrace.h>
27 #ifdef CONFIG_HW_PERF_EVENTS
29 * Per performance monitor configuration as set via oprofilefs.
31 struct op_counter_config {
33 unsigned long enabled;
35 unsigned long unit_mask;
38 struct perf_event_attr attr;
41 static int op_arm_enabled;
42 static DEFINE_MUTEX(op_arm_mutex);
44 static struct op_counter_config *counter_config;
45 static struct perf_event **perf_events[nr_cpumask_bits];
46 static int perf_num_counters;
49 * Overflow callback for oprofile.
51 static void op_overflow_handler(struct perf_event *event, int unused,
52 struct perf_sample_data *data, struct pt_regs *regs)
55 u32 cpu = smp_processor_id();
57 for (id = 0; id < perf_num_counters; ++id)
58 if (perf_events[cpu][id] == event)
61 if (id != perf_num_counters)
62 oprofile_add_sample(regs, id);
64 pr_warning("oprofile: ignoring spurious overflow "
69 * Called by op_arm_setup to create perf attributes to mirror the oprofile
70 * settings in counter_config. Attributes are created as `pinned' events and
71 * so are permanently scheduled on the PMU.
73 static void op_perf_setup(void)
76 u32 size = sizeof(struct perf_event_attr);
77 struct perf_event_attr *attr;
79 for (i = 0; i < perf_num_counters; ++i) {
80 attr = &counter_config[i].attr;
81 memset(attr, 0, size);
82 attr->type = PERF_TYPE_RAW;
84 attr->config = counter_config[i].event;
85 attr->sample_period = counter_config[i].count;
90 static int op_create_counter(int cpu, int event)
93 struct perf_event *pevent;
95 if (!counter_config[event].enabled || (perf_events[cpu][event] != NULL))
98 pevent = perf_event_create_kernel_counter(&counter_config[event].attr,
100 op_overflow_handler);
102 if (IS_ERR(pevent)) {
103 ret = PTR_ERR(pevent);
104 } else if (pevent->state != PERF_EVENT_STATE_ACTIVE) {
105 perf_event_release_kernel(pevent);
106 pr_warning("oprofile: failed to enable event %d "
107 "on CPU %d\n", event, cpu);
110 perf_events[cpu][event] = pevent;
116 static void op_destroy_counter(int cpu, int event)
118 struct perf_event *pevent = perf_events[cpu][event];
121 perf_event_release_kernel(pevent);
122 perf_events[cpu][event] = NULL;
127 * Called by op_arm_start to create active perf events based on the
128 * perviously configured attributes.
130 static int op_perf_start(void)
132 int cpu, event, ret = 0;
134 for_each_online_cpu(cpu) {
135 for (event = 0; event < perf_num_counters; ++event) {
136 ret = op_create_counter(cpu, event);
147 * Called by op_arm_stop at the end of a profiling run.
149 static void op_perf_stop(void)
153 for_each_online_cpu(cpu)
154 for (event = 0; event < perf_num_counters; ++event)
155 op_destroy_counter(cpu, event);
159 static char *op_name_from_perf_id(enum arm_perf_pmu_ids id)
162 case ARM_PERF_PMU_ID_XSCALE1:
163 return "arm/xscale1";
164 case ARM_PERF_PMU_ID_XSCALE2:
165 return "arm/xscale2";
166 case ARM_PERF_PMU_ID_V6:
168 case ARM_PERF_PMU_ID_V6MP:
170 case ARM_PERF_PMU_ID_CA8:
172 case ARM_PERF_PMU_ID_CA9:
173 return "arm/armv7-ca9";
179 static int op_arm_create_files(struct super_block *sb, struct dentry *root)
183 for (i = 0; i < perf_num_counters; i++) {
187 snprintf(buf, sizeof buf, "%d", i);
188 dir = oprofilefs_mkdir(sb, root, buf);
189 oprofilefs_create_ulong(sb, dir, "enabled", &counter_config[i].enabled);
190 oprofilefs_create_ulong(sb, dir, "event", &counter_config[i].event);
191 oprofilefs_create_ulong(sb, dir, "count", &counter_config[i].count);
192 oprofilefs_create_ulong(sb, dir, "unit_mask", &counter_config[i].unit_mask);
193 oprofilefs_create_ulong(sb, dir, "kernel", &counter_config[i].kernel);
194 oprofilefs_create_ulong(sb, dir, "user", &counter_config[i].user);
200 static int op_arm_setup(void)
202 spin_lock(&oprofilefs_lock);
204 spin_unlock(&oprofilefs_lock);
208 static int op_arm_start(void)
212 mutex_lock(&op_arm_mutex);
213 if (!op_arm_enabled) {
218 mutex_unlock(&op_arm_mutex);
222 static void op_arm_stop(void)
224 mutex_lock(&op_arm_mutex);
228 mutex_unlock(&op_arm_mutex);
232 static int op_arm_suspend(struct platform_device *dev, pm_message_t state)
234 mutex_lock(&op_arm_mutex);
237 mutex_unlock(&op_arm_mutex);
241 static int op_arm_resume(struct platform_device *dev)
243 mutex_lock(&op_arm_mutex);
244 if (op_arm_enabled && op_perf_start())
246 mutex_unlock(&op_arm_mutex);
250 static struct platform_driver oprofile_driver = {
252 .name = "arm-oprofile",
254 .resume = op_arm_resume,
255 .suspend = op_arm_suspend,
258 static struct platform_device *oprofile_pdev;
260 static int __init init_driverfs(void)
264 ret = platform_driver_register(&oprofile_driver);
268 oprofile_pdev = platform_device_register_simple(
269 oprofile_driver.driver.name, 0, NULL, 0);
270 if (IS_ERR(oprofile_pdev)) {
271 ret = PTR_ERR(oprofile_pdev);
272 platform_driver_unregister(&oprofile_driver);
279 static void __exit exit_driverfs(void)
281 platform_device_unregister(oprofile_pdev);
282 platform_driver_unregister(&oprofile_driver);
285 static int __init init_driverfs(void) { return 0; }
286 #define exit_driverfs() do { } while (0)
287 #endif /* CONFIG_PM */
289 static int report_trace(struct stackframe *frame, void *d)
291 unsigned int *depth = d;
294 oprofile_add_trace(frame->pc);
302 * The registers we're interested in are at the end of the variable
303 * length saved register structure. The fp points at the end of this
304 * structure so the address of this struct is:
305 * (struct frame_tail *)(xxx->fp)-1
308 struct frame_tail *fp;
311 } __attribute__((packed));
313 static struct frame_tail* user_backtrace(struct frame_tail *tail)
315 struct frame_tail buftail[2];
317 /* Also check accessibility of one struct frame_tail beyond */
318 if (!access_ok(VERIFY_READ, tail, sizeof(buftail)))
320 if (__copy_from_user_inatomic(buftail, tail, sizeof(buftail)))
323 oprofile_add_trace(buftail[0].lr);
325 /* frame pointers should strictly progress back up the stack
326 * (towards higher addresses) */
327 if (tail >= buftail[0].fp)
330 return buftail[0].fp-1;
333 static void arm_backtrace(struct pt_regs * const regs, unsigned int depth)
335 struct frame_tail *tail = ((struct frame_tail *) regs->ARM_fp) - 1;
337 if (!user_mode(regs)) {
338 struct stackframe frame;
339 frame.fp = regs->ARM_fp;
340 frame.sp = regs->ARM_sp;
341 frame.lr = regs->ARM_lr;
342 frame.pc = regs->ARM_pc;
343 walk_stackframe(&frame, report_trace, &depth);
347 while (depth-- && tail && !((unsigned long) tail & 3))
348 tail = user_backtrace(tail);
351 int __init oprofile_arch_init(struct oprofile_operations *ops)
355 memset(&perf_events, 0, sizeof(perf_events));
357 perf_num_counters = armpmu_get_max_events();
359 counter_config = kcalloc(perf_num_counters,
360 sizeof(struct op_counter_config), GFP_KERNEL);
362 if (!counter_config) {
363 pr_info("oprofile: failed to allocate %d "
364 "counters\n", perf_num_counters);
369 ret = init_driverfs();
373 for_each_possible_cpu(cpu) {
374 perf_events[cpu] = kcalloc(perf_num_counters,
375 sizeof(struct perf_event *), GFP_KERNEL);
376 if (!perf_events[cpu]) {
377 pr_info("oprofile: failed to allocate %d perf events "
378 "for cpu %d\n", perf_num_counters, cpu);
384 ops->backtrace = arm_backtrace;
385 ops->create_files = op_arm_create_files;
386 ops->setup = op_arm_setup;
387 ops->start = op_arm_start;
388 ops->stop = op_arm_stop;
389 ops->shutdown = op_arm_stop;
390 ops->cpu_type = op_name_from_perf_id(armpmu_get_pmu_id());
395 pr_info("oprofile: using %s\n", ops->cpu_type);
399 for_each_possible_cpu(cpu)
400 kfree(perf_events[cpu]);
401 kfree(counter_config);
407 void __exit oprofile_arch_exit(void)
410 struct perf_event *event;
412 for_each_possible_cpu(cpu) {
413 for (id = 0; id < perf_num_counters; ++id) {
414 event = perf_events[cpu][id];
416 perf_event_release_kernel(event);
419 kfree(perf_events[cpu]);
422 kfree(counter_config);
426 int __init oprofile_arch_init(struct oprofile_operations *ops)
428 pr_info("oprofile: hardware counters not available\n");
431 void __exit oprofile_arch_exit(void) {}
432 #endif /* CONFIG_HW_PERF_EVENTS */