4 * @remark Copyright 2002 OProfile authors
5 * @remark Read the file COPYING
7 * @author John Levon <levon@movementarian.org>
9 * Each CPU has a local buffer that stores PC value/event
10 * pairs. We also log context switches when we notice them.
11 * Eventually each CPU's buffer is processed into the global
12 * event buffer by sync_buffer().
14 * We use a local buffer for two reasons: an NMI or similar
15 * interrupt cannot synchronise, and high sampling rates
16 * would lead to catastrophic global synchronisation if
17 * a global buffer was used.
20 #include <linux/sched.h>
21 #include <linux/oprofile.h>
22 #include <linux/vmalloc.h>
23 #include <linux/errno.h>
25 #include "event_buffer.h"
26 #include "cpu_buffer.h"
27 #include "buffer_sync.h"
30 struct oprofile_cpu_buffer cpu_buffer[NR_CPUS] __cacheline_aligned;
32 static void wq_sync_buffer(void *);
34 #define DEFAULT_TIMER_EXPIRE (HZ / 10)
35 static int work_enabled;
37 void free_cpu_buffers(void)
41 for_each_online_cpu(i) {
42 vfree(cpu_buffer[i].buffer);
47 int alloc_cpu_buffers(void)
51 unsigned long buffer_size = fs_cpu_buffer_size;
53 for_each_online_cpu(i) {
54 struct oprofile_cpu_buffer * b = &cpu_buffer[i];
56 b->buffer = vmalloc(sizeof(struct op_sample) * buffer_size);
61 b->last_is_kernel = -1;
63 b->buffer_size = buffer_size;
66 b->sample_received = 0;
67 b->sample_lost_overflow = 0;
69 INIT_WORK(&b->work, wq_sync_buffer, b);
79 void start_cpu_work(void)
85 for_each_online_cpu(i) {
86 struct oprofile_cpu_buffer * b = &cpu_buffer[i];
89 * Spread the work by 1 jiffy per cpu so they dont all
92 schedule_delayed_work_on(i, &b->work, DEFAULT_TIMER_EXPIRE + i);
97 void end_cpu_work(void)
103 for_each_online_cpu(i) {
104 struct oprofile_cpu_buffer * b = &cpu_buffer[i];
106 cancel_delayed_work(&b->work);
109 flush_scheduled_work();
113 /* Resets the cpu buffer to a sane state. */
114 void cpu_buffer_reset(struct oprofile_cpu_buffer * cpu_buf)
116 /* reset these to invalid values; the next sample
117 * collected will populate the buffer with proper
118 * values to initialize the buffer
120 cpu_buf->last_is_kernel = -1;
121 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;
138 static void increment_head(struct oprofile_cpu_buffer * b)
140 unsigned long new_head = b->head_pos + 1;
142 /* Ensure anything written to the slot before we
143 * increment is visible */
146 if (new_head < b->buffer_size)
147 b->head_pos = new_head;
156 add_sample(struct oprofile_cpu_buffer * cpu_buf,
157 unsigned long pc, unsigned long event)
159 struct op_sample * entry = &cpu_buf->buffer[cpu_buf->head_pos];
161 entry->event = event;
162 increment_head(cpu_buf);
167 add_code(struct oprofile_cpu_buffer * buffer, unsigned long value)
169 add_sample(buffer, ESCAPE_CODE, value);
173 /* This must be safe from any context. It's safe writing here
174 * because of the head/tail separation of the writer and reader
177 * is_kernel is needed because on some architectures you cannot
178 * tell if you are in kernel or user space simply by looking at
179 * pc. We tag this in the buffer by generating kernel enter/exit
180 * events whenever is_kernel changes
182 static int log_sample(struct oprofile_cpu_buffer * cpu_buf, unsigned long pc,
183 int is_kernel, unsigned long event)
185 struct task_struct * task;
187 cpu_buf->sample_received++;
189 if (nr_available_slots(cpu_buf) < 3) {
190 cpu_buf->sample_lost_overflow++;
194 is_kernel = !!is_kernel;
198 /* notice a switch from user->kernel or vice versa */
199 if (cpu_buf->last_is_kernel != is_kernel) {
200 cpu_buf->last_is_kernel = is_kernel;
201 add_code(cpu_buf, is_kernel);
204 /* notice a task switch */
205 if (cpu_buf->last_task != task) {
206 cpu_buf->last_task = task;
207 add_code(cpu_buf, (unsigned long)task);
210 add_sample(cpu_buf, pc, event);
214 static int oprofile_begin_trace(struct oprofile_cpu_buffer * cpu_buf)
216 if (nr_available_slots(cpu_buf) < 4) {
217 cpu_buf->sample_lost_overflow++;
221 add_code(cpu_buf, CPU_TRACE_BEGIN);
222 cpu_buf->tracing = 1;
227 static void oprofile_end_trace(struct oprofile_cpu_buffer * cpu_buf)
229 cpu_buf->tracing = 0;
233 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
235 struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[smp_processor_id()];
236 unsigned long pc = profile_pc(regs);
237 int is_kernel = !user_mode(regs);
239 if (!backtrace_depth) {
240 log_sample(cpu_buf, pc, is_kernel, event);
244 if (!oprofile_begin_trace(cpu_buf))
247 /* if log_sample() fail we can't backtrace since we lost the source
249 if (log_sample(cpu_buf, pc, is_kernel, event))
250 oprofile_ops.backtrace(regs, backtrace_depth);
251 oprofile_end_trace(cpu_buf);
255 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
257 struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[smp_processor_id()];
258 log_sample(cpu_buf, pc, is_kernel, event);
262 void oprofile_add_trace(unsigned long pc)
264 struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[smp_processor_id()];
266 if (!cpu_buf->tracing)
269 if (nr_available_slots(cpu_buf) < 1) {
270 cpu_buf->tracing = 0;
271 cpu_buf->sample_lost_overflow++;
275 /* broken frame can give an eip with the same value as an escape code,
276 * abort the trace if we get it */
277 if (pc == ESCAPE_CODE) {
278 cpu_buf->tracing = 0;
279 cpu_buf->backtrace_aborted++;
283 add_sample(cpu_buf, pc, 0);
289 * This serves to avoid cpu buffer overflow, and makes sure
290 * the task mortuary progresses
292 * By using schedule_delayed_work_on and then schedule_delayed_work
293 * we guarantee this will stay on the correct cpu
295 static void wq_sync_buffer(void * data)
297 struct oprofile_cpu_buffer * b = data;
298 if (b->cpu != smp_processor_id()) {
299 printk("WQ on CPU%d, prefer CPU%d\n",
300 smp_processor_id(), b->cpu);
304 /* don't re-add the work if we're shutting down */
306 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);