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 static struct ring_buffer *op_ring_buffer_read;
49 static 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 unsigned long oprofile_get_cpu_buffer_size(void)
59 return oprofile_cpu_buffer_size;
62 void oprofile_cpu_buffer_inc_smpl_lost(void)
64 struct oprofile_cpu_buffer *cpu_buf
65 = &__get_cpu_var(cpu_buffer);
67 cpu_buf->sample_lost_overflow++;
70 void free_cpu_buffers(void)
72 if (op_ring_buffer_read)
73 ring_buffer_free(op_ring_buffer_read);
74 op_ring_buffer_read = NULL;
75 if (op_ring_buffer_write)
76 ring_buffer_free(op_ring_buffer_write);
77 op_ring_buffer_write = NULL;
80 int alloc_cpu_buffers(void)
84 unsigned long buffer_size = oprofile_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 int op_cpu_buffer_write_entry(struct op_entry *entry)
150 entry->event = ring_buffer_lock_reserve(op_ring_buffer_write,
151 sizeof(struct op_sample),
154 entry->sample = ring_buffer_event_data(entry->event);
156 entry->sample = NULL;
164 int op_cpu_buffer_write_commit(struct op_entry *entry)
166 return ring_buffer_unlock_commit(op_ring_buffer_write, entry->event,
170 struct op_sample *op_cpu_buffer_read_entry(int cpu)
172 struct ring_buffer_event *e;
173 e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL);
175 return ring_buffer_event_data(e);
176 if (ring_buffer_swap_cpu(op_ring_buffer_read,
177 op_ring_buffer_write,
180 e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL);
182 return ring_buffer_event_data(e);
186 unsigned long op_cpu_buffer_entries(int cpu)
188 return ring_buffer_entries_cpu(op_ring_buffer_read, cpu)
189 + ring_buffer_entries_cpu(op_ring_buffer_write, cpu);
193 add_sample(struct oprofile_cpu_buffer *cpu_buf,
194 unsigned long pc, unsigned long event)
196 struct op_entry entry;
199 ret = op_cpu_buffer_write_entry(&entry);
203 entry.sample->eip = pc;
204 entry.sample->event = event;
206 ret = op_cpu_buffer_write_commit(&entry);
214 add_code(struct oprofile_cpu_buffer *buffer, unsigned long value)
216 return add_sample(buffer, ESCAPE_CODE, value);
219 /* This must be safe from any context. It's safe writing here
220 * because of the head/tail separation of the writer and reader
223 * is_kernel is needed because on some architectures you cannot
224 * tell if you are in kernel or user space simply by looking at
225 * pc. We tag this in the buffer by generating kernel enter/exit
226 * events whenever is_kernel changes
228 static int log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc,
229 int is_kernel, unsigned long event)
231 struct task_struct *task;
233 cpu_buf->sample_received++;
235 if (pc == ESCAPE_CODE) {
236 cpu_buf->sample_invalid_eip++;
240 is_kernel = !!is_kernel;
244 /* notice a switch from user->kernel or vice versa */
245 if (cpu_buf->last_is_kernel != is_kernel) {
246 cpu_buf->last_is_kernel = is_kernel;
247 if (add_code(cpu_buf, is_kernel))
251 /* notice a task switch */
252 if (cpu_buf->last_task != task) {
253 cpu_buf->last_task = task;
254 if (add_code(cpu_buf, (unsigned long)task))
258 if (add_sample(cpu_buf, pc, event))
264 cpu_buf->sample_lost_overflow++;
268 static inline void oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
270 add_code(cpu_buf, CPU_TRACE_BEGIN);
271 cpu_buf->tracing = 1;
274 static inline void oprofile_end_trace(struct oprofile_cpu_buffer *cpu_buf)
276 cpu_buf->tracing = 0;
280 __oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
281 unsigned long event, int is_kernel)
283 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
285 if (!oprofile_backtrace_depth) {
286 log_sample(cpu_buf, pc, is_kernel, event);
290 oprofile_begin_trace(cpu_buf);
293 * if log_sample() fail we can't backtrace since we lost the
294 * source of this event
296 if (log_sample(cpu_buf, pc, is_kernel, event))
297 oprofile_ops.backtrace(regs, oprofile_backtrace_depth);
299 oprofile_end_trace(cpu_buf);
302 void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
303 unsigned long event, int is_kernel)
305 __oprofile_add_ext_sample(pc, regs, event, is_kernel);
308 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
310 int is_kernel = !user_mode(regs);
311 unsigned long pc = profile_pc(regs);
313 __oprofile_add_ext_sample(pc, regs, event, is_kernel);
316 #ifdef CONFIG_OPROFILE_IBS
318 #define MAX_IBS_SAMPLE_SIZE 14
320 void oprofile_add_ibs_sample(struct pt_regs * const regs,
321 unsigned int * const ibs_sample, int ibs_code)
323 int is_kernel = !user_mode(regs);
324 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
325 struct task_struct *task;
328 cpu_buf->sample_received++;
330 /* notice a switch from user->kernel or vice versa */
331 if (cpu_buf->last_is_kernel != is_kernel) {
332 if (add_code(cpu_buf, is_kernel))
334 cpu_buf->last_is_kernel = is_kernel;
337 /* notice a task switch */
340 if (cpu_buf->last_task != task) {
341 if (add_code(cpu_buf, (unsigned long)task))
343 cpu_buf->last_task = task;
347 fail = fail || add_code(cpu_buf, ibs_code);
348 fail = fail || add_sample(cpu_buf, ibs_sample[0], ibs_sample[1]);
349 fail = fail || add_sample(cpu_buf, ibs_sample[2], ibs_sample[3]);
350 fail = fail || add_sample(cpu_buf, ibs_sample[4], ibs_sample[5]);
352 if (ibs_code == IBS_OP_BEGIN) {
353 fail = fail || add_sample(cpu_buf, ibs_sample[6], ibs_sample[7]);
354 fail = fail || add_sample(cpu_buf, ibs_sample[8], ibs_sample[9]);
355 fail = fail || add_sample(cpu_buf, ibs_sample[10], ibs_sample[11]);
361 if (oprofile_backtrace_depth)
362 oprofile_ops.backtrace(regs, oprofile_backtrace_depth);
367 cpu_buf->sample_lost_overflow++;
373 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
375 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
376 log_sample(cpu_buf, pc, is_kernel, event);
379 void oprofile_add_trace(unsigned long pc)
381 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
383 if (!cpu_buf->tracing)
387 * broken frame can give an eip with the same value as an
388 * escape code, abort the trace if we get it
390 if (pc == ESCAPE_CODE)
393 if (add_sample(cpu_buf, pc, 0))
398 cpu_buf->tracing = 0;
399 cpu_buf->backtrace_aborted++;
404 * This serves to avoid cpu buffer overflow, and makes sure
405 * the task mortuary progresses
407 * By using schedule_delayed_work_on and then schedule_delayed_work
408 * we guarantee this will stay on the correct cpu
410 static void wq_sync_buffer(struct work_struct *work)
412 struct oprofile_cpu_buffer *b =
413 container_of(work, struct oprofile_cpu_buffer, work.work);
414 if (b->cpu != smp_processor_id()) {
415 printk(KERN_DEBUG "WQ on CPU%d, prefer CPU%d\n",
416 smp_processor_id(), b->cpu);
418 if (!cpu_online(b->cpu)) {
419 cancel_delayed_work(&b->work);
425 /* don't re-add the work if we're shutting down */
427 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);