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 OP_BUFFER_FLAGS 0
34 * Read and write access is using spin locking. Thus, writing to the
35 * buffer by NMI handler (x86) could occur also during critical
36 * sections when reading the buffer. To avoid this, there are 2
37 * buffers for independent read and write access. Read access is in
38 * process context only, write access only in the NMI handler. If the
39 * read buffer runs empty, both buffers are swapped atomically. There
40 * is potentially a small window during swapping where the buffers are
41 * disabled and samples could be lost.
43 * Using 2 buffers is a little bit overhead, but the solution is clear
44 * and does not require changes in the ring buffer implementation. It
45 * can be changed to a single buffer solution when the ring buffer
46 * access is implemented as non-locking atomic code.
48 struct ring_buffer *op_ring_buffer_read;
49 struct ring_buffer *op_ring_buffer_write;
50 DEFINE_PER_CPU(struct oprofile_cpu_buffer, cpu_buffer);
52 static void wq_sync_buffer(struct work_struct *work);
54 #define DEFAULT_TIMER_EXPIRE (HZ / 10)
55 static int work_enabled;
57 void free_cpu_buffers(void)
59 if (op_ring_buffer_read)
60 ring_buffer_free(op_ring_buffer_read);
61 op_ring_buffer_read = NULL;
62 if (op_ring_buffer_write)
63 ring_buffer_free(op_ring_buffer_write);
64 op_ring_buffer_write = NULL;
67 unsigned long oprofile_get_cpu_buffer_size(void)
69 return fs_cpu_buffer_size;
72 void oprofile_cpu_buffer_inc_smpl_lost(void)
74 struct oprofile_cpu_buffer *cpu_buf
75 = &__get_cpu_var(cpu_buffer);
77 cpu_buf->sample_lost_overflow++;
80 int alloc_cpu_buffers(void)
84 unsigned long buffer_size = fs_cpu_buffer_size;
86 op_ring_buffer_read = ring_buffer_alloc(buffer_size, OP_BUFFER_FLAGS);
87 if (!op_ring_buffer_read)
89 op_ring_buffer_write = ring_buffer_alloc(buffer_size, OP_BUFFER_FLAGS);
90 if (!op_ring_buffer_write)
93 for_each_possible_cpu(i) {
94 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
97 b->last_is_kernel = -1;
99 b->buffer_size = buffer_size;
102 b->sample_received = 0;
103 b->sample_lost_overflow = 0;
104 b->backtrace_aborted = 0;
105 b->sample_invalid_eip = 0;
107 INIT_DELAYED_WORK(&b->work, wq_sync_buffer);
116 void start_cpu_work(void)
122 for_each_online_cpu(i) {
123 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
126 * Spread the work by 1 jiffy per cpu so they dont all
129 schedule_delayed_work_on(i, &b->work, DEFAULT_TIMER_EXPIRE + i);
133 void end_cpu_work(void)
139 for_each_online_cpu(i) {
140 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
142 cancel_delayed_work(&b->work);
145 flush_scheduled_work();
148 /* compute number of available slots in cpu_buffer queue */
149 static unsigned long nr_available_slots(struct oprofile_cpu_buffer const *b)
151 unsigned long head = b->head_pos;
152 unsigned long tail = b->tail_pos;
155 return (tail - head) - 1;
157 return tail + (b->buffer_size - head) - 1;
161 add_sample(struct oprofile_cpu_buffer *cpu_buf,
162 unsigned long pc, unsigned long event)
164 struct op_entry entry;
166 if (cpu_buffer_write_entry(&entry))
169 entry.sample->eip = pc;
170 entry.sample->event = event;
172 if (cpu_buffer_write_commit(&entry))
178 cpu_buf->sample_lost_overflow++;
183 add_code(struct oprofile_cpu_buffer *buffer, unsigned long value)
185 add_sample(buffer, ESCAPE_CODE, value);
188 /* This must be safe from any context. It's safe writing here
189 * because of the head/tail separation of the writer and reader
192 * is_kernel is needed because on some architectures you cannot
193 * tell if you are in kernel or user space simply by looking at
194 * pc. We tag this in the buffer by generating kernel enter/exit
195 * events whenever is_kernel changes
197 static int log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc,
198 int is_kernel, unsigned long event)
200 struct task_struct *task;
202 cpu_buf->sample_received++;
204 if (pc == ESCAPE_CODE) {
205 cpu_buf->sample_invalid_eip++;
209 if (nr_available_slots(cpu_buf) < 3) {
210 cpu_buf->sample_lost_overflow++;
214 is_kernel = !!is_kernel;
218 /* notice a switch from user->kernel or vice versa */
219 if (cpu_buf->last_is_kernel != is_kernel) {
220 cpu_buf->last_is_kernel = is_kernel;
221 add_code(cpu_buf, is_kernel);
224 /* notice a task switch */
225 if (cpu_buf->last_task != task) {
226 cpu_buf->last_task = task;
227 add_code(cpu_buf, (unsigned long)task);
230 add_sample(cpu_buf, pc, event);
234 static int oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
236 if (nr_available_slots(cpu_buf) < 4) {
237 cpu_buf->sample_lost_overflow++;
241 add_code(cpu_buf, CPU_TRACE_BEGIN);
242 cpu_buf->tracing = 1;
246 static void oprofile_end_trace(struct oprofile_cpu_buffer *cpu_buf)
248 cpu_buf->tracing = 0;
251 void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
252 unsigned long event, int is_kernel)
254 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
256 if (!backtrace_depth) {
257 log_sample(cpu_buf, pc, is_kernel, event);
261 if (!oprofile_begin_trace(cpu_buf))
265 * if log_sample() fail we can't backtrace since we lost the
266 * source of this event
268 if (log_sample(cpu_buf, pc, is_kernel, event))
269 oprofile_ops.backtrace(regs, backtrace_depth);
270 oprofile_end_trace(cpu_buf);
273 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
275 int is_kernel = !user_mode(regs);
276 unsigned long pc = profile_pc(regs);
278 oprofile_add_ext_sample(pc, regs, event, is_kernel);
281 #ifdef CONFIG_OPROFILE_IBS
283 #define MAX_IBS_SAMPLE_SIZE 14
285 void oprofile_add_ibs_sample(struct pt_regs * const regs,
286 unsigned int * const ibs_sample, int ibs_code)
288 int is_kernel = !user_mode(regs);
289 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
290 struct task_struct *task;
292 cpu_buf->sample_received++;
294 if (nr_available_slots(cpu_buf) < MAX_IBS_SAMPLE_SIZE) {
295 /* we can't backtrace since we lost the source of this event */
296 cpu_buf->sample_lost_overflow++;
300 /* notice a switch from user->kernel or vice versa */
301 if (cpu_buf->last_is_kernel != is_kernel) {
302 cpu_buf->last_is_kernel = is_kernel;
303 add_code(cpu_buf, is_kernel);
306 /* notice a task switch */
309 if (cpu_buf->last_task != task) {
310 cpu_buf->last_task = task;
311 add_code(cpu_buf, (unsigned long)task);
315 add_code(cpu_buf, ibs_code);
316 add_sample(cpu_buf, ibs_sample[0], ibs_sample[1]);
317 add_sample(cpu_buf, ibs_sample[2], ibs_sample[3]);
318 add_sample(cpu_buf, ibs_sample[4], ibs_sample[5]);
320 if (ibs_code == IBS_OP_BEGIN) {
321 add_sample(cpu_buf, ibs_sample[6], ibs_sample[7]);
322 add_sample(cpu_buf, ibs_sample[8], ibs_sample[9]);
323 add_sample(cpu_buf, ibs_sample[10], ibs_sample[11]);
327 oprofile_ops.backtrace(regs, backtrace_depth);
332 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
334 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
335 log_sample(cpu_buf, pc, is_kernel, event);
338 void oprofile_add_trace(unsigned long pc)
340 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
342 if (!cpu_buf->tracing)
345 if (nr_available_slots(cpu_buf) < 1) {
346 cpu_buf->tracing = 0;
347 cpu_buf->sample_lost_overflow++;
352 * broken frame can give an eip with the same value as an
353 * escape code, abort the trace if we get it
355 if (pc == ESCAPE_CODE) {
356 cpu_buf->tracing = 0;
357 cpu_buf->backtrace_aborted++;
361 add_sample(cpu_buf, pc, 0);
365 * This serves to avoid cpu buffer overflow, and makes sure
366 * the task mortuary progresses
368 * By using schedule_delayed_work_on and then schedule_delayed_work
369 * we guarantee this will stay on the correct cpu
371 static void wq_sync_buffer(struct work_struct *work)
373 struct oprofile_cpu_buffer *b =
374 container_of(work, struct oprofile_cpu_buffer, work.work);
375 if (b->cpu != smp_processor_id()) {
376 printk(KERN_DEBUG "WQ on CPU%d, prefer CPU%d\n",
377 smp_processor_id(), b->cpu);
379 if (!cpu_online(b->cpu)) {
380 cancel_delayed_work(&b->work);
386 /* don't re-add the work if we're shutting down */
388 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);