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);
46 int alloc_cpu_buffers(void)
50 unsigned long buffer_size = fs_cpu_buffer_size;
52 for_each_online_cpu(i) {
53 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
55 b->buffer = vmalloc_node(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;
68 b->backtrace_aborted = 0;
69 b->sample_invalid_eip = 0;
71 INIT_DELAYED_WORK(&b->work, wq_sync_buffer);
80 void start_cpu_work(void)
86 for_each_online_cpu(i) {
87 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
90 * Spread the work by 1 jiffy per cpu so they dont all
93 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 = &per_cpu(cpu_buffer, i);
106 cancel_delayed_work(&b->work);
109 flush_scheduled_work();
112 /* Resets the cpu buffer to a sane state. */
113 void cpu_buffer_reset(struct oprofile_cpu_buffer * cpu_buf)
115 /* reset these to invalid values; the next sample
116 * collected will populate the buffer with proper
117 * values to initialize the buffer
119 cpu_buf->last_is_kernel = -1;
120 cpu_buf->last_task = NULL;
123 /* compute number of available slots in cpu_buffer queue */
124 static unsigned long nr_available_slots(struct oprofile_cpu_buffer const * b)
126 unsigned long head = b->head_pos;
127 unsigned long tail = b->tail_pos;
130 return (tail - head) - 1;
132 return tail + (b->buffer_size - head) - 1;
135 static void increment_head(struct oprofile_cpu_buffer * b)
137 unsigned long new_head = b->head_pos + 1;
139 /* Ensure anything written to the slot before we
140 * increment is visible */
143 if (new_head < b->buffer_size)
144 b->head_pos = new_head;
150 add_sample(struct oprofile_cpu_buffer * cpu_buf,
151 unsigned long pc, unsigned long event)
153 struct op_sample * entry = &cpu_buf->buffer[cpu_buf->head_pos];
155 entry->event = event;
156 increment_head(cpu_buf);
160 add_code(struct oprofile_cpu_buffer * buffer, unsigned long value)
162 add_sample(buffer, ESCAPE_CODE, value);
165 /* This must be safe from any context. It's safe writing here
166 * because of the head/tail separation of the writer and reader
169 * is_kernel is needed because on some architectures you cannot
170 * tell if you are in kernel or user space simply by looking at
171 * pc. We tag this in the buffer by generating kernel enter/exit
172 * events whenever is_kernel changes
174 static int log_sample(struct oprofile_cpu_buffer * cpu_buf, unsigned long pc,
175 int is_kernel, unsigned long event)
177 struct task_struct * task;
179 cpu_buf->sample_received++;
181 if (pc == ESCAPE_CODE) {
182 cpu_buf->sample_invalid_eip++;
186 if (nr_available_slots(cpu_buf) < 3) {
187 cpu_buf->sample_lost_overflow++;
191 is_kernel = !!is_kernel;
195 /* notice a switch from user->kernel or vice versa */
196 if (cpu_buf->last_is_kernel != is_kernel) {
197 cpu_buf->last_is_kernel = is_kernel;
198 add_code(cpu_buf, is_kernel);
201 /* notice a task switch */
202 if (cpu_buf->last_task != task) {
203 cpu_buf->last_task = task;
204 add_code(cpu_buf, (unsigned long)task);
207 add_sample(cpu_buf, pc, event);
211 static int oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
213 if (nr_available_slots(cpu_buf) < 4) {
214 cpu_buf->sample_lost_overflow++;
218 add_code(cpu_buf, CPU_TRACE_BEGIN);
219 cpu_buf->tracing = 1;
223 static void oprofile_end_trace(struct oprofile_cpu_buffer * cpu_buf)
225 cpu_buf->tracing = 0;
228 void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
229 unsigned long event, int is_kernel)
231 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
233 if (!backtrace_depth) {
234 log_sample(cpu_buf, pc, is_kernel, event);
238 if (!oprofile_begin_trace(cpu_buf))
241 /* if log_sample() fail we can't backtrace since we lost the source
243 if (log_sample(cpu_buf, pc, is_kernel, event))
244 oprofile_ops.backtrace(regs, backtrace_depth);
245 oprofile_end_trace(cpu_buf);
248 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
250 int is_kernel = !user_mode(regs);
251 unsigned long pc = profile_pc(regs);
253 oprofile_add_ext_sample(pc, regs, event, is_kernel);
256 #define MAX_IBS_SAMPLE_SIZE 14
257 static int log_ibs_sample(struct oprofile_cpu_buffer *cpu_buf,
258 unsigned long pc, int is_kernel, unsigned int *ibs, int ibs_code)
260 struct task_struct *task;
262 cpu_buf->sample_received++;
264 if (nr_available_slots(cpu_buf) < MAX_IBS_SAMPLE_SIZE) {
265 cpu_buf->sample_lost_overflow++;
269 is_kernel = !!is_kernel;
271 /* notice a switch from user->kernel or vice versa */
272 if (cpu_buf->last_is_kernel != is_kernel) {
273 cpu_buf->last_is_kernel = is_kernel;
274 add_code(cpu_buf, is_kernel);
277 /* notice a task switch */
281 if (cpu_buf->last_task != task) {
282 cpu_buf->last_task = task;
283 add_code(cpu_buf, (unsigned long)task);
287 add_code(cpu_buf, ibs_code);
288 add_sample(cpu_buf, ibs[0], ibs[1]);
289 add_sample(cpu_buf, ibs[2], ibs[3]);
290 add_sample(cpu_buf, ibs[4], ibs[5]);
292 if (ibs_code == IBS_OP_BEGIN) {
293 add_sample(cpu_buf, ibs[6], ibs[7]);
294 add_sample(cpu_buf, ibs[8], ibs[9]);
295 add_sample(cpu_buf, ibs[10], ibs[11]);
301 void oprofile_add_ibs_sample(struct pt_regs *const regs,
302 unsigned int * const ibs_sample, u8 code)
304 int is_kernel = !user_mode(regs);
305 unsigned long pc = profile_pc(regs);
307 struct oprofile_cpu_buffer *cpu_buf =
308 &per_cpu(cpu_buffer, smp_processor_id());
310 if (!backtrace_depth) {
311 log_ibs_sample(cpu_buf, pc, is_kernel, ibs_sample, code);
315 /* if log_sample() fails we can't backtrace since we lost the source
317 if (log_ibs_sample(cpu_buf, pc, is_kernel, ibs_sample, code))
318 oprofile_ops.backtrace(regs, backtrace_depth);
321 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
323 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
324 log_sample(cpu_buf, pc, is_kernel, event);
327 void oprofile_add_trace(unsigned long pc)
329 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
331 if (!cpu_buf->tracing)
334 if (nr_available_slots(cpu_buf) < 1) {
335 cpu_buf->tracing = 0;
336 cpu_buf->sample_lost_overflow++;
340 /* broken frame can give an eip with the same value as an escape code,
341 * abort the trace if we get it */
342 if (pc == ESCAPE_CODE) {
343 cpu_buf->tracing = 0;
344 cpu_buf->backtrace_aborted++;
348 add_sample(cpu_buf, pc, 0);
352 * This serves to avoid cpu buffer overflow, and makes sure
353 * the task mortuary progresses
355 * By using schedule_delayed_work_on and then schedule_delayed_work
356 * we guarantee this will stay on the correct cpu
358 static void wq_sync_buffer(struct work_struct *work)
360 struct oprofile_cpu_buffer * b =
361 container_of(work, struct oprofile_cpu_buffer, work.work);
362 if (b->cpu != smp_processor_id()) {
363 printk("WQ on CPU%d, prefer CPU%d\n",
364 smp_processor_id(), b->cpu);
368 /* don't re-add the work if we're shutting down */
370 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);