Merge branch 'tracing/core-v2' into tracing-for-linus
[linux-2.6] / drivers / oprofile / cpu_buffer.c
1 /**
2  * @file cpu_buffer.c
3  *
4  * @remark Copyright 2002-2009 OProfile authors
5  * @remark Read the file COPYING
6  *
7  * @author John Levon <levon@movementarian.org>
8  * @author Barry Kasindorf <barry.kasindorf@amd.com>
9  * @author Robert Richter <robert.richter@amd.com>
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>
24 #include <linux/vmalloc.h>
25 #include <linux/errno.h>
26
27 #include "event_buffer.h"
28 #include "cpu_buffer.h"
29 #include "buffer_sync.h"
30 #include "oprof.h"
31
32 #define OP_BUFFER_FLAGS 0
33
34 /*
35  * Read and write access is using spin locking. Thus, writing to the
36  * buffer by NMI handler (x86) could occur also during critical
37  * sections when reading the buffer. To avoid this, there are 2
38  * buffers for independent read and write access. Read access is in
39  * process context only, write access only in the NMI handler. If the
40  * read buffer runs empty, both buffers are swapped atomically. There
41  * is potentially a small window during swapping where the buffers are
42  * disabled and samples could be lost.
43  *
44  * Using 2 buffers is a little bit overhead, but the solution is clear
45  * and does not require changes in the ring buffer implementation. It
46  * can be changed to a single buffer solution when the ring buffer
47  * access is implemented as non-locking atomic code.
48  */
49 static struct ring_buffer *op_ring_buffer_read;
50 static struct ring_buffer *op_ring_buffer_write;
51 DEFINE_PER_CPU(struct oprofile_cpu_buffer, cpu_buffer);
52
53 static void wq_sync_buffer(struct work_struct *work);
54
55 #define DEFAULT_TIMER_EXPIRE (HZ / 10)
56 static int work_enabled;
57
58 unsigned long oprofile_get_cpu_buffer_size(void)
59 {
60         return oprofile_cpu_buffer_size;
61 }
62
63 void oprofile_cpu_buffer_inc_smpl_lost(void)
64 {
65         struct oprofile_cpu_buffer *cpu_buf
66                 = &__get_cpu_var(cpu_buffer);
67
68         cpu_buf->sample_lost_overflow++;
69 }
70
71 void free_cpu_buffers(void)
72 {
73         if (op_ring_buffer_read)
74                 ring_buffer_free(op_ring_buffer_read);
75         op_ring_buffer_read = NULL;
76         if (op_ring_buffer_write)
77                 ring_buffer_free(op_ring_buffer_write);
78         op_ring_buffer_write = NULL;
79 }
80
81 int alloc_cpu_buffers(void)
82 {
83         int i;
84
85         unsigned long buffer_size = oprofile_cpu_buffer_size;
86
87         op_ring_buffer_read = ring_buffer_alloc(buffer_size, OP_BUFFER_FLAGS);
88         if (!op_ring_buffer_read)
89                 goto fail;
90         op_ring_buffer_write = ring_buffer_alloc(buffer_size, OP_BUFFER_FLAGS);
91         if (!op_ring_buffer_write)
92                 goto fail;
93
94         for_each_possible_cpu(i) {
95                 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
96
97                 b->last_task = NULL;
98                 b->last_is_kernel = -1;
99                 b->tracing = 0;
100                 b->buffer_size = buffer_size;
101                 b->sample_received = 0;
102                 b->sample_lost_overflow = 0;
103                 b->backtrace_aborted = 0;
104                 b->sample_invalid_eip = 0;
105                 b->cpu = i;
106                 INIT_DELAYED_WORK(&b->work, wq_sync_buffer);
107         }
108         return 0;
109
110 fail:
111         free_cpu_buffers();
112         return -ENOMEM;
113 }
114
115 void start_cpu_work(void)
116 {
117         int i;
118
119         work_enabled = 1;
120
121         for_each_online_cpu(i) {
122                 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
123
124                 /*
125                  * Spread the work by 1 jiffy per cpu so they dont all
126                  * fire at once.
127                  */
128                 schedule_delayed_work_on(i, &b->work, DEFAULT_TIMER_EXPIRE + i);
129         }
130 }
131
132 void end_cpu_work(void)
133 {
134         int i;
135
136         work_enabled = 0;
137
138         for_each_online_cpu(i) {
139                 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
140
141                 cancel_delayed_work(&b->work);
142         }
143
144         flush_scheduled_work();
145 }
146
147 /*
148  * This function prepares the cpu buffer to write a sample.
149  *
150  * Struct op_entry is used during operations on the ring buffer while
151  * struct op_sample contains the data that is stored in the ring
152  * buffer. Struct entry can be uninitialized. The function reserves a
153  * data array that is specified by size. Use
154  * op_cpu_buffer_write_commit() after preparing the sample. In case of
155  * errors a null pointer is returned, otherwise the pointer to the
156  * sample.
157  *
158  */
159 struct op_sample
160 *op_cpu_buffer_write_reserve(struct op_entry *entry, unsigned long size)
161 {
162         entry->event = ring_buffer_lock_reserve
163                 (op_ring_buffer_write, sizeof(struct op_sample) +
164                  size * sizeof(entry->sample->data[0]));
165         if (entry->event)
166                 entry->sample = ring_buffer_event_data(entry->event);
167         else
168                 entry->sample = NULL;
169
170         if (!entry->sample)
171                 return NULL;
172
173         entry->size = size;
174         entry->data = entry->sample->data;
175
176         return entry->sample;
177 }
178
179 int op_cpu_buffer_write_commit(struct op_entry *entry)
180 {
181         return ring_buffer_unlock_commit(op_ring_buffer_write, entry->event);
182 }
183
184 struct op_sample *op_cpu_buffer_read_entry(struct op_entry *entry, int cpu)
185 {
186         struct ring_buffer_event *e;
187         e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL);
188         if (e)
189                 goto event;
190         if (ring_buffer_swap_cpu(op_ring_buffer_read,
191                                  op_ring_buffer_write,
192                                  cpu))
193                 return NULL;
194         e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL);
195         if (e)
196                 goto event;
197         return NULL;
198
199 event:
200         entry->event = e;
201         entry->sample = ring_buffer_event_data(e);
202         entry->size = (ring_buffer_event_length(e) - sizeof(struct op_sample))
203                 / sizeof(entry->sample->data[0]);
204         entry->data = entry->sample->data;
205         return entry->sample;
206 }
207
208 unsigned long op_cpu_buffer_entries(int cpu)
209 {
210         return ring_buffer_entries_cpu(op_ring_buffer_read, cpu)
211                 + ring_buffer_entries_cpu(op_ring_buffer_write, cpu);
212 }
213
214 static int
215 op_add_code(struct oprofile_cpu_buffer *cpu_buf, unsigned long backtrace,
216             int is_kernel, struct task_struct *task)
217 {
218         struct op_entry entry;
219         struct op_sample *sample;
220         unsigned long flags;
221         int size;
222
223         flags = 0;
224
225         if (backtrace)
226                 flags |= TRACE_BEGIN;
227
228         /* notice a switch from user->kernel or vice versa */
229         is_kernel = !!is_kernel;
230         if (cpu_buf->last_is_kernel != is_kernel) {
231                 cpu_buf->last_is_kernel = is_kernel;
232                 flags |= KERNEL_CTX_SWITCH;
233                 if (is_kernel)
234                         flags |= IS_KERNEL;
235         }
236
237         /* notice a task switch */
238         if (cpu_buf->last_task != task) {
239                 cpu_buf->last_task = task;
240                 flags |= USER_CTX_SWITCH;
241         }
242
243         if (!flags)
244                 /* nothing to do */
245                 return 0;
246
247         if (flags & USER_CTX_SWITCH)
248                 size = 1;
249         else
250                 size = 0;
251
252         sample = op_cpu_buffer_write_reserve(&entry, size);
253         if (!sample)
254                 return -ENOMEM;
255
256         sample->eip = ESCAPE_CODE;
257         sample->event = flags;
258
259         if (size)
260                 op_cpu_buffer_add_data(&entry, (unsigned long)task);
261
262         op_cpu_buffer_write_commit(&entry);
263
264         return 0;
265 }
266
267 static inline int
268 op_add_sample(struct oprofile_cpu_buffer *cpu_buf,
269               unsigned long pc, unsigned long event)
270 {
271         struct op_entry entry;
272         struct op_sample *sample;
273
274         sample = op_cpu_buffer_write_reserve(&entry, 0);
275         if (!sample)
276                 return -ENOMEM;
277
278         sample->eip = pc;
279         sample->event = event;
280
281         return op_cpu_buffer_write_commit(&entry);
282 }
283
284 /*
285  * This must be safe from any context.
286  *
287  * is_kernel is needed because on some architectures you cannot
288  * tell if you are in kernel or user space simply by looking at
289  * pc. We tag this in the buffer by generating kernel enter/exit
290  * events whenever is_kernel changes
291  */
292 static int
293 log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc,
294            unsigned long backtrace, int is_kernel, unsigned long event)
295 {
296         cpu_buf->sample_received++;
297
298         if (pc == ESCAPE_CODE) {
299                 cpu_buf->sample_invalid_eip++;
300                 return 0;
301         }
302
303         if (op_add_code(cpu_buf, backtrace, is_kernel, current))
304                 goto fail;
305
306         if (op_add_sample(cpu_buf, pc, event))
307                 goto fail;
308
309         return 1;
310
311 fail:
312         cpu_buf->sample_lost_overflow++;
313         return 0;
314 }
315
316 static inline void oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
317 {
318         cpu_buf->tracing = 1;
319 }
320
321 static inline void oprofile_end_trace(struct oprofile_cpu_buffer *cpu_buf)
322 {
323         cpu_buf->tracing = 0;
324 }
325
326 static inline void
327 __oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
328                           unsigned long event, int is_kernel)
329 {
330         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
331         unsigned long backtrace = oprofile_backtrace_depth;
332
333         /*
334          * if log_sample() fail we can't backtrace since we lost the
335          * source of this event
336          */
337         if (!log_sample(cpu_buf, pc, backtrace, is_kernel, event))
338                 /* failed */
339                 return;
340
341         if (!backtrace)
342                 return;
343
344         oprofile_begin_trace(cpu_buf);
345         oprofile_ops.backtrace(regs, backtrace);
346         oprofile_end_trace(cpu_buf);
347 }
348
349 void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
350                              unsigned long event, int is_kernel)
351 {
352         __oprofile_add_ext_sample(pc, regs, event, is_kernel);
353 }
354
355 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
356 {
357         int is_kernel = !user_mode(regs);
358         unsigned long pc = profile_pc(regs);
359
360         __oprofile_add_ext_sample(pc, regs, event, is_kernel);
361 }
362
363 /*
364  * Add samples with data to the ring buffer.
365  *
366  * Use oprofile_add_data(&entry, val) to add data and
367  * oprofile_write_commit(&entry) to commit the sample.
368  */
369 void
370 oprofile_write_reserve(struct op_entry *entry, struct pt_regs * const regs,
371                        unsigned long pc, int code, int size)
372 {
373         struct op_sample *sample;
374         int is_kernel = !user_mode(regs);
375         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
376
377         cpu_buf->sample_received++;
378
379         /* no backtraces for samples with data */
380         if (op_add_code(cpu_buf, 0, is_kernel, current))
381                 goto fail;
382
383         sample = op_cpu_buffer_write_reserve(entry, size + 2);
384         if (!sample)
385                 goto fail;
386         sample->eip = ESCAPE_CODE;
387         sample->event = 0;              /* no flags */
388
389         op_cpu_buffer_add_data(entry, code);
390         op_cpu_buffer_add_data(entry, pc);
391
392         return;
393
394 fail:
395         entry->event = NULL;
396         cpu_buf->sample_lost_overflow++;
397 }
398
399 int oprofile_add_data(struct op_entry *entry, unsigned long val)
400 {
401         if (!entry->event)
402                 return 0;
403         return op_cpu_buffer_add_data(entry, val);
404 }
405
406 int oprofile_write_commit(struct op_entry *entry)
407 {
408         if (!entry->event)
409                 return -EINVAL;
410         return op_cpu_buffer_write_commit(entry);
411 }
412
413 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
414 {
415         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
416         log_sample(cpu_buf, pc, 0, is_kernel, event);
417 }
418
419 void oprofile_add_trace(unsigned long pc)
420 {
421         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
422
423         if (!cpu_buf->tracing)
424                 return;
425
426         /*
427          * broken frame can give an eip with the same value as an
428          * escape code, abort the trace if we get it
429          */
430         if (pc == ESCAPE_CODE)
431                 goto fail;
432
433         if (op_add_sample(cpu_buf, pc, 0))
434                 goto fail;
435
436         return;
437 fail:
438         cpu_buf->tracing = 0;
439         cpu_buf->backtrace_aborted++;
440         return;
441 }
442
443 /*
444  * This serves to avoid cpu buffer overflow, and makes sure
445  * the task mortuary progresses
446  *
447  * By using schedule_delayed_work_on and then schedule_delayed_work
448  * we guarantee this will stay on the correct cpu
449  */
450 static void wq_sync_buffer(struct work_struct *work)
451 {
452         struct oprofile_cpu_buffer *b =
453                 container_of(work, struct oprofile_cpu_buffer, work.work);
454         if (b->cpu != smp_processor_id()) {
455                 printk(KERN_DEBUG "WQ on CPU%d, prefer CPU%d\n",
456                        smp_processor_id(), b->cpu);
457
458                 if (!cpu_online(b->cpu)) {
459                         cancel_delayed_work(&b->work);
460                         return;
461                 }
462         }
463         sync_buffer(b->cpu);
464
465         /* don't re-add the work if we're shutting down */
466         if (work_enabled)
467                 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);
468 }