4 * @remark Copyright 2002 OProfile authors
5 * @remark Read the file COPYING
7 * @author John Levon <levon@movementarian.org>
8 * @author Barry Kasindorf <barry.kasindorf@amd.com>
10 * Each CPU has a local buffer that stores PC value/event
11 * pairs. We also log context switches when we notice them.
12 * Eventually each CPU's buffer is processed into the global
13 * event buffer by sync_buffer().
15 * We use a local buffer for two reasons: an NMI or similar
16 * interrupt cannot synchronise, and high sampling rates
17 * would lead to catastrophic global synchronisation if
18 * a global buffer was used.
21 #include <linux/sched.h>
22 #include <linux/oprofile.h>
23 #include <linux/vmalloc.h>
24 #include <linux/errno.h>
26 #include "event_buffer.h"
27 #include "cpu_buffer.h"
28 #include "buffer_sync.h"
31 DEFINE_PER_CPU(struct oprofile_cpu_buffer, cpu_buffer);
33 static void wq_sync_buffer(struct work_struct *work);
35 #define DEFAULT_TIMER_EXPIRE (HZ / 10)
36 static int work_enabled;
38 void free_cpu_buffers(void)
42 for_each_online_cpu(i) {
43 vfree(per_cpu(cpu_buffer, i).buffer);
44 per_cpu(cpu_buffer, i).buffer = NULL;
48 int alloc_cpu_buffers(void)
52 unsigned long buffer_size = fs_cpu_buffer_size;
54 for_each_online_cpu(i) {
55 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
57 b->buffer = vmalloc_node(sizeof(struct op_sample) * buffer_size,
63 b->last_is_kernel = -1;
65 b->buffer_size = buffer_size;
68 b->sample_received = 0;
69 b->sample_lost_overflow = 0;
70 b->backtrace_aborted = 0;
71 b->sample_invalid_eip = 0;
73 INIT_DELAYED_WORK(&b->work, wq_sync_buffer);
82 void start_cpu_work(void)
88 for_each_online_cpu(i) {
89 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
92 * Spread the work by 1 jiffy per cpu so they dont all
95 schedule_delayed_work_on(i, &b->work, DEFAULT_TIMER_EXPIRE + i);
99 void end_cpu_work(void)
105 for_each_online_cpu(i) {
106 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
108 cancel_delayed_work(&b->work);
111 flush_scheduled_work();
114 /* Resets the cpu buffer to a sane state. */
115 void cpu_buffer_reset(struct oprofile_cpu_buffer *cpu_buf)
117 /* reset these to invalid values; the next sample
118 * collected will populate the buffer with proper
119 * values to initialize the buffer
121 cpu_buf->last_is_kernel = -1;
122 cpu_buf->last_task = NULL;
125 /* compute number of available slots in cpu_buffer queue */
126 static unsigned long nr_available_slots(struct oprofile_cpu_buffer const *b)
128 unsigned long head = b->head_pos;
129 unsigned long tail = b->tail_pos;
132 return (tail - head) - 1;
134 return tail + (b->buffer_size - head) - 1;
137 static void increment_head(struct oprofile_cpu_buffer *b)
139 unsigned long new_head = b->head_pos + 1;
141 /* Ensure anything written to the slot before we
142 * increment is visible */
145 if (new_head < b->buffer_size)
146 b->head_pos = new_head;
152 add_sample(struct oprofile_cpu_buffer *cpu_buf,
153 unsigned long pc, unsigned long event)
155 struct op_sample *entry = &cpu_buf->buffer[cpu_buf->head_pos];
157 entry->event = event;
158 increment_head(cpu_buf);
162 add_code(struct oprofile_cpu_buffer *buffer, unsigned long value)
164 add_sample(buffer, ESCAPE_CODE, value);
167 /* This must be safe from any context. It's safe writing here
168 * because of the head/tail separation of the writer and reader
171 * is_kernel is needed because on some architectures you cannot
172 * tell if you are in kernel or user space simply by looking at
173 * pc. We tag this in the buffer by generating kernel enter/exit
174 * events whenever is_kernel changes
176 static int log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc,
177 int is_kernel, unsigned long event)
179 struct task_struct *task;
181 cpu_buf->sample_received++;
183 if (pc == ESCAPE_CODE) {
184 cpu_buf->sample_invalid_eip++;
188 if (nr_available_slots(cpu_buf) < 3) {
189 cpu_buf->sample_lost_overflow++;
193 is_kernel = !!is_kernel;
197 /* notice a switch from user->kernel or vice versa */
198 if (cpu_buf->last_is_kernel != is_kernel) {
199 cpu_buf->last_is_kernel = is_kernel;
200 add_code(cpu_buf, is_kernel);
203 /* notice a task switch */
204 if (cpu_buf->last_task != task) {
205 cpu_buf->last_task = task;
206 add_code(cpu_buf, (unsigned long)task);
209 add_sample(cpu_buf, pc, event);
213 static int oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
215 if (nr_available_slots(cpu_buf) < 4) {
216 cpu_buf->sample_lost_overflow++;
220 add_code(cpu_buf, CPU_TRACE_BEGIN);
221 cpu_buf->tracing = 1;
225 static void oprofile_end_trace(struct oprofile_cpu_buffer *cpu_buf)
227 cpu_buf->tracing = 0;
230 void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
231 unsigned long event, int is_kernel)
233 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
235 if (!backtrace_depth) {
236 log_sample(cpu_buf, pc, is_kernel, event);
240 if (!oprofile_begin_trace(cpu_buf))
243 /* if log_sample() fail we can't backtrace since we lost the source
245 if (log_sample(cpu_buf, pc, is_kernel, event))
246 oprofile_ops.backtrace(regs, backtrace_depth);
247 oprofile_end_trace(cpu_buf);
250 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
252 int is_kernel = !user_mode(regs);
253 unsigned long pc = profile_pc(regs);
255 oprofile_add_ext_sample(pc, regs, event, is_kernel);
258 #ifdef CONFIG_OPROFILE_IBS
260 #define MAX_IBS_SAMPLE_SIZE 14
262 void oprofile_add_ibs_sample(struct pt_regs *const regs,
263 unsigned int *const ibs_sample, int ibs_code)
265 int is_kernel = !user_mode(regs);
266 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
267 struct task_struct *task;
269 cpu_buf->sample_received++;
271 if (nr_available_slots(cpu_buf) < MAX_IBS_SAMPLE_SIZE) {
272 /* we can't backtrace since we lost the source of this event */
273 cpu_buf->sample_lost_overflow++;
277 /* notice a switch from user->kernel or vice versa */
278 if (cpu_buf->last_is_kernel != is_kernel) {
279 cpu_buf->last_is_kernel = is_kernel;
280 add_code(cpu_buf, is_kernel);
283 /* notice a task switch */
286 if (cpu_buf->last_task != task) {
287 cpu_buf->last_task = task;
288 add_code(cpu_buf, (unsigned long)task);
292 add_code(cpu_buf, ibs_code);
293 add_sample(cpu_buf, ibs_sample[0], ibs_sample[1]);
294 add_sample(cpu_buf, ibs_sample[2], ibs_sample[3]);
295 add_sample(cpu_buf, ibs_sample[4], ibs_sample[5]);
297 if (ibs_code == IBS_OP_BEGIN) {
298 add_sample(cpu_buf, ibs_sample[6], ibs_sample[7]);
299 add_sample(cpu_buf, ibs_sample[8], ibs_sample[9]);
300 add_sample(cpu_buf, ibs_sample[10], ibs_sample[11]);
304 oprofile_ops.backtrace(regs, backtrace_depth);
309 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
311 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
312 log_sample(cpu_buf, pc, is_kernel, event);
315 void oprofile_add_trace(unsigned long pc)
317 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
319 if (!cpu_buf->tracing)
322 if (nr_available_slots(cpu_buf) < 1) {
323 cpu_buf->tracing = 0;
324 cpu_buf->sample_lost_overflow++;
328 /* broken frame can give an eip with the same value as an escape code,
329 * abort the trace if we get it */
330 if (pc == ESCAPE_CODE) {
331 cpu_buf->tracing = 0;
332 cpu_buf->backtrace_aborted++;
336 add_sample(cpu_buf, pc, 0);
340 * This serves to avoid cpu buffer overflow, and makes sure
341 * the task mortuary progresses
343 * By using schedule_delayed_work_on and then schedule_delayed_work
344 * we guarantee this will stay on the correct cpu
346 static void wq_sync_buffer(struct work_struct *work)
348 struct oprofile_cpu_buffer *b =
349 container_of(work, struct oprofile_cpu_buffer, work.work);
350 if (b->cpu != smp_processor_id()) {
351 printk(KERN_DEBUG "WQ on CPU%d, prefer CPU%d\n",
352 smp_processor_id(), b->cpu);
356 /* don't re-add the work if we're shutting down */
358 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);