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);
45 int alloc_cpu_buffers(void)
49 unsigned long buffer_size = fs_cpu_buffer_size;
51 for_each_online_cpu(i) {
52 struct oprofile_cpu_buffer * b = &cpu_buffer[i];
54 b->buffer = vmalloc_node(sizeof(struct op_sample) * buffer_size,
60 b->last_is_kernel = -1;
62 b->buffer_size = buffer_size;
65 b->sample_received = 0;
66 b->sample_lost_overflow = 0;
68 INIT_WORK(&b->work, wq_sync_buffer, b);
77 void start_cpu_work(void)
83 for_each_online_cpu(i) {
84 struct oprofile_cpu_buffer * b = &cpu_buffer[i];
87 * Spread the work by 1 jiffy per cpu so they dont all
90 schedule_delayed_work_on(i, &b->work, DEFAULT_TIMER_EXPIRE + i);
94 void end_cpu_work(void)
100 for_each_online_cpu(i) {
101 struct oprofile_cpu_buffer * b = &cpu_buffer[i];
103 cancel_delayed_work(&b->work);
106 flush_scheduled_work();
109 /* Resets the cpu buffer to a sane state. */
110 void cpu_buffer_reset(struct oprofile_cpu_buffer * cpu_buf)
112 /* reset these to invalid values; the next sample
113 * collected will populate the buffer with proper
114 * values to initialize the buffer
116 cpu_buf->last_is_kernel = -1;
117 cpu_buf->last_task = NULL;
120 /* compute number of available slots in cpu_buffer queue */
121 static unsigned long nr_available_slots(struct oprofile_cpu_buffer const * b)
123 unsigned long head = b->head_pos;
124 unsigned long tail = b->tail_pos;
127 return (tail - head) - 1;
129 return tail + (b->buffer_size - head) - 1;
132 static void increment_head(struct oprofile_cpu_buffer * b)
134 unsigned long new_head = b->head_pos + 1;
136 /* Ensure anything written to the slot before we
137 * increment is visible */
140 if (new_head < b->buffer_size)
141 b->head_pos = new_head;
147 add_sample(struct oprofile_cpu_buffer * cpu_buf,
148 unsigned long pc, unsigned long event)
150 struct op_sample * entry = &cpu_buf->buffer[cpu_buf->head_pos];
152 entry->event = event;
153 increment_head(cpu_buf);
157 add_code(struct oprofile_cpu_buffer * buffer, unsigned long value)
159 add_sample(buffer, ESCAPE_CODE, value);
162 /* This must be safe from any context. It's safe writing here
163 * because of the head/tail separation of the writer and reader
166 * is_kernel is needed because on some architectures you cannot
167 * tell if you are in kernel or user space simply by looking at
168 * pc. We tag this in the buffer by generating kernel enter/exit
169 * events whenever is_kernel changes
171 static int log_sample(struct oprofile_cpu_buffer * cpu_buf, unsigned long pc,
172 int is_kernel, unsigned long event)
174 struct task_struct * task;
176 cpu_buf->sample_received++;
178 if (nr_available_slots(cpu_buf) < 3) {
179 cpu_buf->sample_lost_overflow++;
183 is_kernel = !!is_kernel;
187 /* notice a switch from user->kernel or vice versa */
188 if (cpu_buf->last_is_kernel != is_kernel) {
189 cpu_buf->last_is_kernel = is_kernel;
190 add_code(cpu_buf, is_kernel);
193 /* notice a task switch */
194 if (cpu_buf->last_task != task) {
195 cpu_buf->last_task = task;
196 add_code(cpu_buf, (unsigned long)task);
199 add_sample(cpu_buf, pc, event);
203 static int oprofile_begin_trace(struct oprofile_cpu_buffer * cpu_buf)
205 if (nr_available_slots(cpu_buf) < 4) {
206 cpu_buf->sample_lost_overflow++;
210 add_code(cpu_buf, CPU_TRACE_BEGIN);
211 cpu_buf->tracing = 1;
215 static void oprofile_end_trace(struct oprofile_cpu_buffer * cpu_buf)
217 cpu_buf->tracing = 0;
220 void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
221 unsigned long event, int is_kernel)
223 struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[smp_processor_id()];
225 if (!backtrace_depth) {
226 log_sample(cpu_buf, pc, is_kernel, event);
230 if (!oprofile_begin_trace(cpu_buf))
233 /* if log_sample() fail we can't backtrace since we lost the source
235 if (log_sample(cpu_buf, pc, is_kernel, event))
236 oprofile_ops.backtrace(regs, backtrace_depth);
237 oprofile_end_trace(cpu_buf);
240 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
242 int is_kernel = !user_mode(regs);
243 unsigned long pc = profile_pc(regs);
245 oprofile_add_ext_sample(pc, regs, event, is_kernel);
248 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
250 struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[smp_processor_id()];
251 log_sample(cpu_buf, pc, is_kernel, event);
254 void oprofile_add_trace(unsigned long pc)
256 struct oprofile_cpu_buffer * cpu_buf = &cpu_buffer[smp_processor_id()];
258 if (!cpu_buf->tracing)
261 if (nr_available_slots(cpu_buf) < 1) {
262 cpu_buf->tracing = 0;
263 cpu_buf->sample_lost_overflow++;
267 /* broken frame can give an eip with the same value as an escape code,
268 * abort the trace if we get it */
269 if (pc == ESCAPE_CODE) {
270 cpu_buf->tracing = 0;
271 cpu_buf->backtrace_aborted++;
275 add_sample(cpu_buf, pc, 0);
279 * This serves to avoid cpu buffer overflow, and makes sure
280 * the task mortuary progresses
282 * By using schedule_delayed_work_on and then schedule_delayed_work
283 * we guarantee this will stay on the correct cpu
285 static void wq_sync_buffer(void * data)
287 struct oprofile_cpu_buffer * b = data;
288 if (b->cpu != smp_processor_id()) {
289 printk("WQ on CPU%d, prefer CPU%d\n",
290 smp_processor_id(), b->cpu);
294 /* don't re-add the work if we're shutting down */
296 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);