Merge branch 'topic/ctxfi' into 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 #define RB_EVENT_HDR_SIZE 4
82
83 int alloc_cpu_buffers(void)
84 {
85         int i;
86
87         unsigned long buffer_size = oprofile_cpu_buffer_size;
88         unsigned long byte_size = buffer_size * (sizeof(struct op_sample) +
89                                                  RB_EVENT_HDR_SIZE);
90
91         op_ring_buffer_read = ring_buffer_alloc(byte_size, OP_BUFFER_FLAGS);
92         if (!op_ring_buffer_read)
93                 goto fail;
94         op_ring_buffer_write = ring_buffer_alloc(byte_size, OP_BUFFER_FLAGS);
95         if (!op_ring_buffer_write)
96                 goto fail;
97
98         for_each_possible_cpu(i) {
99                 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
100
101                 b->last_task = NULL;
102                 b->last_is_kernel = -1;
103                 b->tracing = 0;
104                 b->buffer_size = buffer_size;
105                 b->sample_received = 0;
106                 b->sample_lost_overflow = 0;
107                 b->backtrace_aborted = 0;
108                 b->sample_invalid_eip = 0;
109                 b->cpu = i;
110                 INIT_DELAYED_WORK(&b->work, wq_sync_buffer);
111         }
112         return 0;
113
114 fail:
115         free_cpu_buffers();
116         return -ENOMEM;
117 }
118
119 void start_cpu_work(void)
120 {
121         int i;
122
123         work_enabled = 1;
124
125         for_each_online_cpu(i) {
126                 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
127
128                 /*
129                  * Spread the work by 1 jiffy per cpu so they dont all
130                  * fire at once.
131                  */
132                 schedule_delayed_work_on(i, &b->work, DEFAULT_TIMER_EXPIRE + i);
133         }
134 }
135
136 void end_cpu_work(void)
137 {
138         int i;
139
140         work_enabled = 0;
141
142         for_each_online_cpu(i) {
143                 struct oprofile_cpu_buffer *b = &per_cpu(cpu_buffer, i);
144
145                 cancel_delayed_work(&b->work);
146         }
147
148         flush_scheduled_work();
149 }
150
151 /*
152  * This function prepares the cpu buffer to write a sample.
153  *
154  * Struct op_entry is used during operations on the ring buffer while
155  * struct op_sample contains the data that is stored in the ring
156  * buffer. Struct entry can be uninitialized. The function reserves a
157  * data array that is specified by size. Use
158  * op_cpu_buffer_write_commit() after preparing the sample. In case of
159  * errors a null pointer is returned, otherwise the pointer to the
160  * sample.
161  *
162  */
163 struct op_sample
164 *op_cpu_buffer_write_reserve(struct op_entry *entry, unsigned long size)
165 {
166         entry->event = ring_buffer_lock_reserve
167                 (op_ring_buffer_write, sizeof(struct op_sample) +
168                  size * sizeof(entry->sample->data[0]));
169         if (entry->event)
170                 entry->sample = ring_buffer_event_data(entry->event);
171         else
172                 entry->sample = NULL;
173
174         if (!entry->sample)
175                 return NULL;
176
177         entry->size = size;
178         entry->data = entry->sample->data;
179
180         return entry->sample;
181 }
182
183 int op_cpu_buffer_write_commit(struct op_entry *entry)
184 {
185         return ring_buffer_unlock_commit(op_ring_buffer_write, entry->event);
186 }
187
188 struct op_sample *op_cpu_buffer_read_entry(struct op_entry *entry, int cpu)
189 {
190         struct ring_buffer_event *e;
191         e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL);
192         if (e)
193                 goto event;
194         if (ring_buffer_swap_cpu(op_ring_buffer_read,
195                                  op_ring_buffer_write,
196                                  cpu))
197                 return NULL;
198         e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL);
199         if (e)
200                 goto event;
201         return NULL;
202
203 event:
204         entry->event = e;
205         entry->sample = ring_buffer_event_data(e);
206         entry->size = (ring_buffer_event_length(e) - sizeof(struct op_sample))
207                 / sizeof(entry->sample->data[0]);
208         entry->data = entry->sample->data;
209         return entry->sample;
210 }
211
212 unsigned long op_cpu_buffer_entries(int cpu)
213 {
214         return ring_buffer_entries_cpu(op_ring_buffer_read, cpu)
215                 + ring_buffer_entries_cpu(op_ring_buffer_write, cpu);
216 }
217
218 static int
219 op_add_code(struct oprofile_cpu_buffer *cpu_buf, unsigned long backtrace,
220             int is_kernel, struct task_struct *task)
221 {
222         struct op_entry entry;
223         struct op_sample *sample;
224         unsigned long flags;
225         int size;
226
227         flags = 0;
228
229         if (backtrace)
230                 flags |= TRACE_BEGIN;
231
232         /* notice a switch from user->kernel or vice versa */
233         is_kernel = !!is_kernel;
234         if (cpu_buf->last_is_kernel != is_kernel) {
235                 cpu_buf->last_is_kernel = is_kernel;
236                 flags |= KERNEL_CTX_SWITCH;
237                 if (is_kernel)
238                         flags |= IS_KERNEL;
239         }
240
241         /* notice a task switch */
242         if (cpu_buf->last_task != task) {
243                 cpu_buf->last_task = task;
244                 flags |= USER_CTX_SWITCH;
245         }
246
247         if (!flags)
248                 /* nothing to do */
249                 return 0;
250
251         if (flags & USER_CTX_SWITCH)
252                 size = 1;
253         else
254                 size = 0;
255
256         sample = op_cpu_buffer_write_reserve(&entry, size);
257         if (!sample)
258                 return -ENOMEM;
259
260         sample->eip = ESCAPE_CODE;
261         sample->event = flags;
262
263         if (size)
264                 op_cpu_buffer_add_data(&entry, (unsigned long)task);
265
266         op_cpu_buffer_write_commit(&entry);
267
268         return 0;
269 }
270
271 static inline int
272 op_add_sample(struct oprofile_cpu_buffer *cpu_buf,
273               unsigned long pc, unsigned long event)
274 {
275         struct op_entry entry;
276         struct op_sample *sample;
277
278         sample = op_cpu_buffer_write_reserve(&entry, 0);
279         if (!sample)
280                 return -ENOMEM;
281
282         sample->eip = pc;
283         sample->event = event;
284
285         return op_cpu_buffer_write_commit(&entry);
286 }
287
288 /*
289  * This must be safe from any context.
290  *
291  * is_kernel is needed because on some architectures you cannot
292  * tell if you are in kernel or user space simply by looking at
293  * pc. We tag this in the buffer by generating kernel enter/exit
294  * events whenever is_kernel changes
295  */
296 static int
297 log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc,
298            unsigned long backtrace, int is_kernel, unsigned long event)
299 {
300         cpu_buf->sample_received++;
301
302         if (pc == ESCAPE_CODE) {
303                 cpu_buf->sample_invalid_eip++;
304                 return 0;
305         }
306
307         if (op_add_code(cpu_buf, backtrace, is_kernel, current))
308                 goto fail;
309
310         if (op_add_sample(cpu_buf, pc, event))
311                 goto fail;
312
313         return 1;
314
315 fail:
316         cpu_buf->sample_lost_overflow++;
317         return 0;
318 }
319
320 static inline void oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
321 {
322         cpu_buf->tracing = 1;
323 }
324
325 static inline void oprofile_end_trace(struct oprofile_cpu_buffer *cpu_buf)
326 {
327         cpu_buf->tracing = 0;
328 }
329
330 static inline void
331 __oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
332                           unsigned long event, int is_kernel)
333 {
334         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
335         unsigned long backtrace = oprofile_backtrace_depth;
336
337         /*
338          * if log_sample() fail we can't backtrace since we lost the
339          * source of this event
340          */
341         if (!log_sample(cpu_buf, pc, backtrace, is_kernel, event))
342                 /* failed */
343                 return;
344
345         if (!backtrace)
346                 return;
347
348         oprofile_begin_trace(cpu_buf);
349         oprofile_ops.backtrace(regs, backtrace);
350         oprofile_end_trace(cpu_buf);
351 }
352
353 void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
354                              unsigned long event, int is_kernel)
355 {
356         __oprofile_add_ext_sample(pc, regs, event, is_kernel);
357 }
358
359 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
360 {
361         int is_kernel = !user_mode(regs);
362         unsigned long pc = profile_pc(regs);
363
364         __oprofile_add_ext_sample(pc, regs, event, is_kernel);
365 }
366
367 /*
368  * Add samples with data to the ring buffer.
369  *
370  * Use oprofile_add_data(&entry, val) to add data and
371  * oprofile_write_commit(&entry) to commit the sample.
372  */
373 void
374 oprofile_write_reserve(struct op_entry *entry, struct pt_regs * const regs,
375                        unsigned long pc, int code, int size)
376 {
377         struct op_sample *sample;
378         int is_kernel = !user_mode(regs);
379         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
380
381         cpu_buf->sample_received++;
382
383         /* no backtraces for samples with data */
384         if (op_add_code(cpu_buf, 0, is_kernel, current))
385                 goto fail;
386
387         sample = op_cpu_buffer_write_reserve(entry, size + 2);
388         if (!sample)
389                 goto fail;
390         sample->eip = ESCAPE_CODE;
391         sample->event = 0;              /* no flags */
392
393         op_cpu_buffer_add_data(entry, code);
394         op_cpu_buffer_add_data(entry, pc);
395
396         return;
397
398 fail:
399         entry->event = NULL;
400         cpu_buf->sample_lost_overflow++;
401 }
402
403 int oprofile_add_data(struct op_entry *entry, unsigned long val)
404 {
405         if (!entry->event)
406                 return 0;
407         return op_cpu_buffer_add_data(entry, val);
408 }
409
410 int oprofile_write_commit(struct op_entry *entry)
411 {
412         if (!entry->event)
413                 return -EINVAL;
414         return op_cpu_buffer_write_commit(entry);
415 }
416
417 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
418 {
419         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
420         log_sample(cpu_buf, pc, 0, is_kernel, event);
421 }
422
423 void oprofile_add_trace(unsigned long pc)
424 {
425         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
426
427         if (!cpu_buf->tracing)
428                 return;
429
430         /*
431          * broken frame can give an eip with the same value as an
432          * escape code, abort the trace if we get it
433          */
434         if (pc == ESCAPE_CODE)
435                 goto fail;
436
437         if (op_add_sample(cpu_buf, pc, 0))
438                 goto fail;
439
440         return;
441 fail:
442         cpu_buf->tracing = 0;
443         cpu_buf->backtrace_aborted++;
444         return;
445 }
446
447 /*
448  * This serves to avoid cpu buffer overflow, and makes sure
449  * the task mortuary progresses
450  *
451  * By using schedule_delayed_work_on and then schedule_delayed_work
452  * we guarantee this will stay on the correct cpu
453  */
454 static void wq_sync_buffer(struct work_struct *work)
455 {
456         struct oprofile_cpu_buffer *b =
457                 container_of(work, struct oprofile_cpu_buffer, work.work);
458         if (b->cpu != smp_processor_id()) {
459                 printk(KERN_DEBUG "WQ on CPU%d, prefer CPU%d\n",
460                        smp_processor_id(), b->cpu);
461
462                 if (!cpu_online(b->cpu)) {
463                         cancel_delayed_work(&b->work);
464                         return;
465                 }
466         }
467         sync_buffer(b->cpu);
468
469         /* don't re-add the work if we're shutting down */
470         if (work_enabled)
471                 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);
472 }