oprofile: rework implementation of cpu buffer events
[linux-2.6.git] / 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]), &entry->irq_flags);
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                                          entry->irq_flags);
183 }
184
185 struct op_sample *op_cpu_buffer_read_entry(struct op_entry *entry, int cpu)
186 {
187         struct ring_buffer_event *e;
188         e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL);
189         if (e)
190                 goto event;
191         if (ring_buffer_swap_cpu(op_ring_buffer_read,
192                                  op_ring_buffer_write,
193                                  cpu))
194                 return NULL;
195         e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL);
196         if (e)
197                 goto event;
198         return NULL;
199
200 event:
201         entry->event = e;
202         entry->sample = ring_buffer_event_data(e);
203         entry->size = (ring_buffer_event_length(e) - sizeof(struct op_sample))
204                 / sizeof(entry->sample->data[0]);
205         entry->data = entry->sample->data;
206         return entry->sample;
207 }
208
209 unsigned long op_cpu_buffer_entries(int cpu)
210 {
211         return ring_buffer_entries_cpu(op_ring_buffer_read, cpu)
212                 + ring_buffer_entries_cpu(op_ring_buffer_write, cpu);
213 }
214
215 static int
216 op_add_code(struct oprofile_cpu_buffer *cpu_buf, unsigned long backtrace,
217             int is_kernel, struct task_struct *task)
218 {
219         struct op_entry entry;
220         struct op_sample *sample;
221         unsigned long flags;
222         int size;
223
224         flags = 0;
225
226         if (backtrace)
227                 flags |= TRACE_BEGIN;
228
229         /* notice a switch from user->kernel or vice versa */
230         is_kernel = !!is_kernel;
231         if (cpu_buf->last_is_kernel != is_kernel) {
232                 cpu_buf->last_is_kernel = is_kernel;
233                 flags |= KERNEL_CTX_SWITCH;
234                 if (is_kernel)
235                         flags |= IS_KERNEL;
236         }
237
238         /* notice a task switch */
239         if (cpu_buf->last_task != task) {
240                 cpu_buf->last_task = task;
241                 flags |= USER_CTX_SWITCH;
242         }
243
244         if (!flags)
245                 /* nothing to do */
246                 return 0;
247
248         if (flags & USER_CTX_SWITCH)
249                 size = 1;
250         else
251                 size = 0;
252
253         sample = op_cpu_buffer_write_reserve(&entry, size);
254         if (!sample)
255                 return -ENOMEM;
256
257         sample->eip = ESCAPE_CODE;
258         sample->event = flags;
259
260         if (size)
261                 sample->data[0] = (unsigned long)task;
262
263         op_cpu_buffer_write_commit(&entry);
264
265         return 0;
266 }
267
268 static inline int
269 op_add_sample(struct oprofile_cpu_buffer *cpu_buf,
270               unsigned long pc, unsigned long event)
271 {
272         struct op_entry entry;
273         struct op_sample *sample;
274
275         sample = op_cpu_buffer_write_reserve(&entry, 0);
276         if (!sample)
277                 return -ENOMEM;
278
279         sample->eip = pc;
280         sample->event = event;
281
282         return op_cpu_buffer_write_commit(&entry);
283 }
284
285 /*
286  * This must be safe from any context.
287  *
288  * is_kernel is needed because on some architectures you cannot
289  * tell if you are in kernel or user space simply by looking at
290  * pc. We tag this in the buffer by generating kernel enter/exit
291  * events whenever is_kernel changes
292  */
293 static int
294 log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc,
295            unsigned long backtrace, int is_kernel, unsigned long event)
296 {
297         cpu_buf->sample_received++;
298
299         if (pc == ESCAPE_CODE) {
300                 cpu_buf->sample_invalid_eip++;
301                 return 0;
302         }
303
304         if (op_add_code(cpu_buf, backtrace, is_kernel, current))
305                 goto fail;
306
307         if (op_add_sample(cpu_buf, pc, event))
308                 goto fail;
309
310         return 1;
311
312 fail:
313         cpu_buf->sample_lost_overflow++;
314         return 0;
315 }
316
317 static inline void oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
318 {
319         cpu_buf->tracing = 1;
320 }
321
322 static inline void oprofile_end_trace(struct oprofile_cpu_buffer *cpu_buf)
323 {
324         cpu_buf->tracing = 0;
325 }
326
327 static inline void
328 __oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
329                           unsigned long event, int is_kernel)
330 {
331         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
332         unsigned long backtrace = oprofile_backtrace_depth;
333
334         /*
335          * if log_sample() fail we can't backtrace since we lost the
336          * source of this event
337          */
338         if (!log_sample(cpu_buf, pc, backtrace, is_kernel, event))
339                 /* failed */
340                 return;
341
342         if (!backtrace)
343                 return;
344
345         oprofile_begin_trace(cpu_buf);
346         oprofile_ops.backtrace(regs, backtrace);
347         oprofile_end_trace(cpu_buf);
348 }
349
350 void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
351                              unsigned long event, int is_kernel)
352 {
353         __oprofile_add_ext_sample(pc, regs, event, is_kernel);
354 }
355
356 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
357 {
358         int is_kernel = !user_mode(regs);
359         unsigned long pc = profile_pc(regs);
360
361         __oprofile_add_ext_sample(pc, regs, event, is_kernel);
362 }
363
364 #ifdef CONFIG_OPROFILE_IBS
365
366 void oprofile_add_ibs_sample(struct pt_regs * const regs,
367                              unsigned int * const ibs_sample, int ibs_code)
368 {
369         int is_kernel = !user_mode(regs);
370         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
371         int fail = 0;
372
373         cpu_buf->sample_received++;
374
375         /* backtraces disabled for ibs */
376         fail = fail || op_add_code(cpu_buf, 0, is_kernel, current);
377
378         fail = fail || op_add_sample(cpu_buf, ESCAPE_CODE,   ibs_code);
379         fail = fail || op_add_sample(cpu_buf, ibs_sample[0], ibs_sample[1]);
380         fail = fail || op_add_sample(cpu_buf, ibs_sample[2], ibs_sample[3]);
381         fail = fail || op_add_sample(cpu_buf, ibs_sample[4], ibs_sample[5]);
382
383         if (ibs_code == IBS_OP_BEGIN) {
384                 fail = fail || op_add_sample(cpu_buf, ibs_sample[6], ibs_sample[7]);
385                 fail = fail || op_add_sample(cpu_buf, ibs_sample[8], ibs_sample[9]);
386                 fail = fail || op_add_sample(cpu_buf, ibs_sample[10], ibs_sample[11]);
387         }
388
389         if (fail)
390                 cpu_buf->sample_lost_overflow++;
391 }
392
393 #endif
394
395 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
396 {
397         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
398         log_sample(cpu_buf, pc, 0, is_kernel, event);
399 }
400
401 void oprofile_add_trace(unsigned long pc)
402 {
403         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
404
405         if (!cpu_buf->tracing)
406                 return;
407
408         /*
409          * broken frame can give an eip with the same value as an
410          * escape code, abort the trace if we get it
411          */
412         if (pc == ESCAPE_CODE)
413                 goto fail;
414
415         if (op_add_sample(cpu_buf, pc, 0))
416                 goto fail;
417
418         return;
419 fail:
420         cpu_buf->tracing = 0;
421         cpu_buf->backtrace_aborted++;
422         return;
423 }
424
425 /*
426  * This serves to avoid cpu buffer overflow, and makes sure
427  * the task mortuary progresses
428  *
429  * By using schedule_delayed_work_on and then schedule_delayed_work
430  * we guarantee this will stay on the correct cpu
431  */
432 static void wq_sync_buffer(struct work_struct *work)
433 {
434         struct oprofile_cpu_buffer *b =
435                 container_of(work, struct oprofile_cpu_buffer, work.work);
436         if (b->cpu != smp_processor_id()) {
437                 printk(KERN_DEBUG "WQ on CPU%d, prefer CPU%d\n",
438                        smp_processor_id(), b->cpu);
439
440                 if (!cpu_online(b->cpu)) {
441                         cancel_delayed_work(&b->work);
442                         return;
443                 }
444         }
445         sync_buffer(b->cpu);
446
447         /* don't re-add the work if we're shutting down */
448         if (work_enabled)
449                 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);
450 }