Merge branch 'wimax-2.6.35.y' of git://git.kernel.org/pub/scm/linux/kernel/git/inaky...
[GitHub/mt8127/android_kernel_alcatel_ttab.git] / drivers / oprofile / cpu_buffer.c
CommitLineData
1da177e4
LT
1/**
2 * @file cpu_buffer.c
3 *
2cc28b9f 4 * @remark Copyright 2002-2009 OProfile authors
1da177e4
LT
5 * @remark Read the file COPYING
6 *
7 * @author John Levon <levon@movementarian.org>
345c2573 8 * @author Barry Kasindorf <barry.kasindorf@amd.com>
2cc28b9f 9 * @author Robert Richter <robert.richter@amd.com>
1da177e4
LT
10 *
11 * Each CPU has a local buffer that stores PC value/event
12 * pairs. We also log context switches when we notice them.
13 * Eventually each CPU's buffer is processed into the global
14 * event buffer by sync_buffer().
15 *
16 * We use a local buffer for two reasons: an NMI or similar
17 * interrupt cannot synchronise, and high sampling rates
18 * would lead to catastrophic global synchronisation if
19 * a global buffer was used.
20 */
21
22#include <linux/sched.h>
23#include <linux/oprofile.h>
1da177e4 24#include <linux/errno.h>
6a18037d 25
1da177e4
LT
26#include "event_buffer.h"
27#include "cpu_buffer.h"
28#include "buffer_sync.h"
29#include "oprof.h"
30
6dad828b
RR
31#define OP_BUFFER_FLAGS 0
32
cb6e943c 33static struct ring_buffer *op_ring_buffer;
b3e9f672 34DEFINE_PER_CPU(struct oprofile_cpu_buffer, op_cpu_buffer);
1da177e4 35
c4028958 36static void wq_sync_buffer(struct work_struct *work);
1da177e4
LT
37
38#define DEFAULT_TIMER_EXPIRE (HZ / 10)
39static int work_enabled;
40
a5598ca0
CL
41unsigned long oprofile_get_cpu_buffer_size(void)
42{
bd2172f5 43 return oprofile_cpu_buffer_size;
a5598ca0
CL
44}
45
46void oprofile_cpu_buffer_inc_smpl_lost(void)
47{
b3e9f672 48 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(op_cpu_buffer);
a5598ca0
CL
49
50 cpu_buf->sample_lost_overflow++;
51}
52
30015776
RR
53void free_cpu_buffers(void)
54{
cb6e943c
AK
55 if (op_ring_buffer)
56 ring_buffer_free(op_ring_buffer);
57 op_ring_buffer = NULL;
30015776
RR
58}
59
54f2c841
RR
60#define RB_EVENT_HDR_SIZE 4
61
1da177e4
LT
62int alloc_cpu_buffers(void)
63{
64 int i;
6a18037d 65
bd2172f5 66 unsigned long buffer_size = oprofile_cpu_buffer_size;
54f2c841
RR
67 unsigned long byte_size = buffer_size * (sizeof(struct op_sample) +
68 RB_EVENT_HDR_SIZE);
6a18037d 69
cb6e943c
AK
70 op_ring_buffer = ring_buffer_alloc(byte_size, OP_BUFFER_FLAGS);
71 if (!op_ring_buffer)
6dad828b
RR
72 goto fail;
73
4bd9b9dc 74 for_each_possible_cpu(i) {
b3e9f672 75 struct oprofile_cpu_buffer *b = &per_cpu(op_cpu_buffer, i);
6a18037d 76
1da177e4
LT
77 b->last_task = NULL;
78 b->last_is_kernel = -1;
79 b->tracing = 0;
80 b->buffer_size = buffer_size;
1da177e4
LT
81 b->sample_received = 0;
82 b->sample_lost_overflow = 0;
df9d177a
PE
83 b->backtrace_aborted = 0;
84 b->sample_invalid_eip = 0;
1da177e4 85 b->cpu = i;
c4028958 86 INIT_DELAYED_WORK(&b->work, wq_sync_buffer);
1da177e4
LT
87 }
88 return 0;
89
90fail:
91 free_cpu_buffers();
92 return -ENOMEM;
93}
1da177e4
LT
94
95void start_cpu_work(void)
96{
97 int i;
98
99 work_enabled = 1;
100
101 for_each_online_cpu(i) {
b3e9f672 102 struct oprofile_cpu_buffer *b = &per_cpu(op_cpu_buffer, i);
1da177e4
LT
103
104 /*
105 * Spread the work by 1 jiffy per cpu so they dont all
106 * fire at once.
107 */
108 schedule_delayed_work_on(i, &b->work, DEFAULT_TIMER_EXPIRE + i);
109 }
110}
111
1da177e4
LT
112void end_cpu_work(void)
113{
114 int i;
115
116 work_enabled = 0;
117
118 for_each_online_cpu(i) {
b3e9f672 119 struct oprofile_cpu_buffer *b = &per_cpu(op_cpu_buffer, i);
1da177e4
LT
120
121 cancel_delayed_work(&b->work);
122 }
123
124 flush_scheduled_work();
125}
126
2cc28b9f
RR
127/*
128 * This function prepares the cpu buffer to write a sample.
129 *
130 * Struct op_entry is used during operations on the ring buffer while
131 * struct op_sample contains the data that is stored in the ring
132 * buffer. Struct entry can be uninitialized. The function reserves a
133 * data array that is specified by size. Use
134 * op_cpu_buffer_write_commit() after preparing the sample. In case of
135 * errors a null pointer is returned, otherwise the pointer to the
136 * sample.
137 *
138 */
139struct op_sample
140*op_cpu_buffer_write_reserve(struct op_entry *entry, unsigned long size)
9966718d 141{
2cc28b9f 142 entry->event = ring_buffer_lock_reserve
cb6e943c 143 (op_ring_buffer, sizeof(struct op_sample) +
304cc6ae 144 size * sizeof(entry->sample->data[0]));
cb6e943c 145 if (!entry->event)
2cc28b9f 146 return NULL;
cb6e943c 147 entry->sample = ring_buffer_event_data(entry->event);
2cc28b9f
RR
148 entry->size = size;
149 entry->data = entry->sample->data;
150
151 return entry->sample;
9966718d
RR
152}
153
154int op_cpu_buffer_write_commit(struct op_entry *entry)
155{
cb6e943c 156 return ring_buffer_unlock_commit(op_ring_buffer, entry->event);
9966718d
RR
157}
158
2d87b14c 159struct op_sample *op_cpu_buffer_read_entry(struct op_entry *entry, int cpu)
9966718d
RR
160{
161 struct ring_buffer_event *e;
b971f061 162 e = ring_buffer_consume(op_ring_buffer, cpu, NULL, NULL);
cb6e943c 163 if (!e)
9966718d 164 return NULL;
2d87b14c 165
2d87b14c
RR
166 entry->event = e;
167 entry->sample = ring_buffer_event_data(e);
168 entry->size = (ring_buffer_event_length(e) - sizeof(struct op_sample))
169 / sizeof(entry->sample->data[0]);
170 entry->data = entry->sample->data;
171 return entry->sample;
9966718d
RR
172}
173
174unsigned long op_cpu_buffer_entries(int cpu)
175{
cb6e943c 176 return ring_buffer_entries_cpu(op_ring_buffer, cpu);
9966718d
RR
177}
178
ae735e99
RR
179static int
180op_add_code(struct oprofile_cpu_buffer *cpu_buf, unsigned long backtrace,
181 int is_kernel, struct task_struct *task)
182{
183 struct op_entry entry;
184 struct op_sample *sample;
185 unsigned long flags;
186 int size;
187
188 flags = 0;
189
190 if (backtrace)
191 flags |= TRACE_BEGIN;
192
193 /* notice a switch from user->kernel or vice versa */
194 is_kernel = !!is_kernel;
195 if (cpu_buf->last_is_kernel != is_kernel) {
196 cpu_buf->last_is_kernel = is_kernel;
197 flags |= KERNEL_CTX_SWITCH;
198 if (is_kernel)
199 flags |= IS_KERNEL;
200 }
201
202 /* notice a task switch */
203 if (cpu_buf->last_task != task) {
204 cpu_buf->last_task = task;
205 flags |= USER_CTX_SWITCH;
206 }
207
208 if (!flags)
209 /* nothing to do */
210 return 0;
211
212 if (flags & USER_CTX_SWITCH)
213 size = 1;
214 else
215 size = 0;
216
217 sample = op_cpu_buffer_write_reserve(&entry, size);
218 if (!sample)
219 return -ENOMEM;
220
221 sample->eip = ESCAPE_CODE;
222 sample->event = flags;
223
224 if (size)
d9928c25 225 op_cpu_buffer_add_data(&entry, (unsigned long)task);
ae735e99
RR
226
227 op_cpu_buffer_write_commit(&entry);
228
229 return 0;
230}
231
211117ff 232static inline int
d0e23384
RR
233op_add_sample(struct oprofile_cpu_buffer *cpu_buf,
234 unsigned long pc, unsigned long event)
1da177e4 235{
6dad828b 236 struct op_entry entry;
2cc28b9f 237 struct op_sample *sample;
6dad828b 238
2cc28b9f
RR
239 sample = op_cpu_buffer_write_reserve(&entry, 0);
240 if (!sample)
241 return -ENOMEM;
6dad828b 242
2cc28b9f
RR
243 sample->eip = pc;
244 sample->event = event;
6dad828b 245
3967e93e 246 return op_cpu_buffer_write_commit(&entry);
1da177e4
LT
247}
248
ae735e99
RR
249/*
250 * This must be safe from any context.
1da177e4
LT
251 *
252 * is_kernel is needed because on some architectures you cannot
253 * tell if you are in kernel or user space simply by looking at
254 * pc. We tag this in the buffer by generating kernel enter/exit
255 * events whenever is_kernel changes
256 */
ae735e99
RR
257static int
258log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc,
259 unsigned long backtrace, int is_kernel, unsigned long event)
1da177e4 260{
1da177e4
LT
261 cpu_buf->sample_received++;
262
df9d177a
PE
263 if (pc == ESCAPE_CODE) {
264 cpu_buf->sample_invalid_eip++;
265 return 0;
266 }
267
ae735e99
RR
268 if (op_add_code(cpu_buf, backtrace, is_kernel, current))
269 goto fail;
6a18037d 270
d0e23384 271 if (op_add_sample(cpu_buf, pc, event))
211117ff
RR
272 goto fail;
273
1da177e4 274 return 1;
211117ff
RR
275
276fail:
277 cpu_buf->sample_lost_overflow++;
278 return 0;
1da177e4
LT
279}
280
6352d92d 281static inline void oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
1da177e4 282{
1da177e4 283 cpu_buf->tracing = 1;
1da177e4
LT
284}
285
6352d92d 286static inline void oprofile_end_trace(struct oprofile_cpu_buffer *cpu_buf)
1da177e4
LT
287{
288 cpu_buf->tracing = 0;
289}
290
d45d23be
RR
291static inline void
292__oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
293 unsigned long event, int is_kernel)
1da177e4 294{
b3e9f672 295 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(op_cpu_buffer);
ae735e99 296 unsigned long backtrace = oprofile_backtrace_depth;
1da177e4 297
fd13f6c8
RR
298 /*
299 * if log_sample() fail we can't backtrace since we lost the
300 * source of this event
301 */
ae735e99
RR
302 if (!log_sample(cpu_buf, pc, backtrace, is_kernel, event))
303 /* failed */
304 return;
305
306 if (!backtrace)
307 return;
6352d92d 308
ae735e99
RR
309 oprofile_begin_trace(cpu_buf);
310 oprofile_ops.backtrace(regs, backtrace);
1da177e4
LT
311 oprofile_end_trace(cpu_buf);
312}
313
d45d23be
RR
314void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
315 unsigned long event, int is_kernel)
316{
317 __oprofile_add_ext_sample(pc, regs, event, is_kernel);
318}
319
27357716
BR
320void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
321{
9414e996
PC
322 int is_kernel;
323 unsigned long pc;
324
325 if (likely(regs)) {
326 is_kernel = !user_mode(regs);
327 pc = profile_pc(regs);
328 } else {
329 is_kernel = 0; /* This value will not be used */
330 pc = ESCAPE_CODE; /* as this causes an early return. */
331 }
27357716 332
d45d23be 333 __oprofile_add_ext_sample(pc, regs, event, is_kernel);
27357716
BR
334}
335
1acda878
RR
336/*
337 * Add samples with data to the ring buffer.
338 *
14f0ca8e
RR
339 * Use oprofile_add_data(&entry, val) to add data and
340 * oprofile_write_commit(&entry) to commit the sample.
1acda878 341 */
14f0ca8e
RR
342void
343oprofile_write_reserve(struct op_entry *entry, struct pt_regs * const regs,
1acda878 344 unsigned long pc, int code, int size)
345c2573 345{
1acda878 346 struct op_sample *sample;
e2fee276 347 int is_kernel = !user_mode(regs);
b3e9f672 348 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(op_cpu_buffer);
345c2573
BK
349
350 cpu_buf->sample_received++;
351
1acda878
RR
352 /* no backtraces for samples with data */
353 if (op_add_code(cpu_buf, 0, is_kernel, current))
354 goto fail;
355
356 sample = op_cpu_buffer_write_reserve(entry, size + 2);
357 if (!sample)
358 goto fail;
359 sample->eip = ESCAPE_CODE;
360 sample->event = 0; /* no flags */
345c2573 361
1acda878
RR
362 op_cpu_buffer_add_data(entry, code);
363 op_cpu_buffer_add_data(entry, pc);
345c2573 364
1acda878 365 return;
345c2573 366
1acda878 367fail:
fdb6a8f4 368 entry->event = NULL;
1acda878 369 cpu_buf->sample_lost_overflow++;
345c2573
BK
370}
371
14f0ca8e
RR
372int oprofile_add_data(struct op_entry *entry, unsigned long val)
373{
fdb6a8f4
RR
374 if (!entry->event)
375 return 0;
14f0ca8e
RR
376 return op_cpu_buffer_add_data(entry, val);
377}
378
51563a0e
RR
379int oprofile_add_data64(struct op_entry *entry, u64 val)
380{
381 if (!entry->event)
382 return 0;
383 if (op_cpu_buffer_get_size(entry) < 2)
384 /*
385 * the function returns 0 to indicate a too small
386 * buffer, even if there is some space left
387 */
388 return 0;
389 if (!op_cpu_buffer_add_data(entry, (u32)val))
390 return 0;
391 return op_cpu_buffer_add_data(entry, (u32)(val >> 32));
392}
393
14f0ca8e
RR
394int oprofile_write_commit(struct op_entry *entry)
395{
fdb6a8f4
RR
396 if (!entry->event)
397 return -EINVAL;
14f0ca8e
RR
398 return op_cpu_buffer_write_commit(entry);
399}
400
1da177e4
LT
401void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
402{
b3e9f672 403 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(op_cpu_buffer);
ae735e99 404 log_sample(cpu_buf, pc, 0, is_kernel, event);
1da177e4
LT
405}
406
1da177e4
LT
407void oprofile_add_trace(unsigned long pc)
408{
b3e9f672 409 struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(op_cpu_buffer);
1da177e4
LT
410
411 if (!cpu_buf->tracing)
412 return;
413
fd13f6c8
RR
414 /*
415 * broken frame can give an eip with the same value as an
416 * escape code, abort the trace if we get it
417 */
211117ff
RR
418 if (pc == ESCAPE_CODE)
419 goto fail;
420
d0e23384 421 if (op_add_sample(cpu_buf, pc, 0))
211117ff 422 goto fail;
1da177e4 423
211117ff
RR
424 return;
425fail:
426 cpu_buf->tracing = 0;
427 cpu_buf->backtrace_aborted++;
428 return;
1da177e4
LT
429}
430
1da177e4
LT
431/*
432 * This serves to avoid cpu buffer overflow, and makes sure
433 * the task mortuary progresses
434 *
435 * By using schedule_delayed_work_on and then schedule_delayed_work
436 * we guarantee this will stay on the correct cpu
437 */
c4028958 438static void wq_sync_buffer(struct work_struct *work)
1da177e4 439{
25ad2913 440 struct oprofile_cpu_buffer *b =
c4028958 441 container_of(work, struct oprofile_cpu_buffer, work.work);
1da177e4 442 if (b->cpu != smp_processor_id()) {
bd17b625 443 printk(KERN_DEBUG "WQ on CPU%d, prefer CPU%d\n",
1da177e4 444 smp_processor_id(), b->cpu);
4bd9b9dc
CA
445
446 if (!cpu_online(b->cpu)) {
447 cancel_delayed_work(&b->work);
448 return;
449 }
1da177e4
LT
450 }
451 sync_buffer(b->cpu);
452
453 /* don't re-add the work if we're shutting down */
454 if (work_enabled)
455 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);
456}