oprofile: add op_cpu_buffer_write_reserve()
[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(int cpu)
186 {
187         struct ring_buffer_event *e;
188         e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL);
189         if (e)
190                 return ring_buffer_event_data(e);
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                 return ring_buffer_event_data(e);
198         return NULL;
199 }
200
201 unsigned long op_cpu_buffer_entries(int cpu)
202 {
203         return ring_buffer_entries_cpu(op_ring_buffer_read, cpu)
204                 + ring_buffer_entries_cpu(op_ring_buffer_write, cpu);
205 }
206
207 static inline int
208 op_add_sample(struct oprofile_cpu_buffer *cpu_buf,
209               unsigned long pc, unsigned long event)
210 {
211         struct op_entry entry;
212         struct op_sample *sample;
213
214         sample = op_cpu_buffer_write_reserve(&entry, 0);
215         if (!sample)
216                 return -ENOMEM;
217
218         sample->eip = pc;
219         sample->event = event;
220
221         return op_cpu_buffer_write_commit(&entry);
222 }
223
224 static inline int
225 add_code(struct oprofile_cpu_buffer *buffer, unsigned long value)
226 {
227         return op_add_sample(buffer, ESCAPE_CODE, value);
228 }
229
230 /* This must be safe from any context. It's safe writing here
231  * because of the head/tail separation of the writer and reader
232  * of the CPU buffer.
233  *
234  * is_kernel is needed because on some architectures you cannot
235  * tell if you are in kernel or user space simply by looking at
236  * pc. We tag this in the buffer by generating kernel enter/exit
237  * events whenever is_kernel changes
238  */
239 static int log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc,
240                       int is_kernel, unsigned long event)
241 {
242         struct task_struct *task;
243
244         cpu_buf->sample_received++;
245
246         if (pc == ESCAPE_CODE) {
247                 cpu_buf->sample_invalid_eip++;
248                 return 0;
249         }
250
251         is_kernel = !!is_kernel;
252
253         task = current;
254
255         /* notice a switch from user->kernel or vice versa */
256         if (cpu_buf->last_is_kernel != is_kernel) {
257                 cpu_buf->last_is_kernel = is_kernel;
258                 if (add_code(cpu_buf, is_kernel))
259                         goto fail;
260         }
261
262         /* notice a task switch */
263         if (cpu_buf->last_task != task) {
264                 cpu_buf->last_task = task;
265                 if (add_code(cpu_buf, (unsigned long)task))
266                         goto fail;
267         }
268
269         if (op_add_sample(cpu_buf, pc, event))
270                 goto fail;
271
272         return 1;
273
274 fail:
275         cpu_buf->sample_lost_overflow++;
276         return 0;
277 }
278
279 static inline void oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
280 {
281         add_code(cpu_buf, CPU_TRACE_BEGIN);
282         cpu_buf->tracing = 1;
283 }
284
285 static inline void oprofile_end_trace(struct oprofile_cpu_buffer *cpu_buf)
286 {
287         cpu_buf->tracing = 0;
288 }
289
290 static inline void
291 __oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
292                           unsigned long event, int is_kernel)
293 {
294         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
295
296         if (!oprofile_backtrace_depth) {
297                 log_sample(cpu_buf, pc, is_kernel, event);
298                 return;
299         }
300
301         oprofile_begin_trace(cpu_buf);
302
303         /*
304          * if log_sample() fail we can't backtrace since we lost the
305          * source of this event
306          */
307         if (log_sample(cpu_buf, pc, is_kernel, event))
308                 oprofile_ops.backtrace(regs, oprofile_backtrace_depth);
309
310         oprofile_end_trace(cpu_buf);
311 }
312
313 void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
314                              unsigned long event, int is_kernel)
315 {
316         __oprofile_add_ext_sample(pc, regs, event, is_kernel);
317 }
318
319 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
320 {
321         int is_kernel = !user_mode(regs);
322         unsigned long pc = profile_pc(regs);
323
324         __oprofile_add_ext_sample(pc, regs, event, is_kernel);
325 }
326
327 #ifdef CONFIG_OPROFILE_IBS
328
329 void oprofile_add_ibs_sample(struct pt_regs * const regs,
330                              unsigned int * const ibs_sample, int ibs_code)
331 {
332         int is_kernel = !user_mode(regs);
333         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
334         struct task_struct *task;
335         int fail = 0;
336
337         cpu_buf->sample_received++;
338
339         /* notice a switch from user->kernel or vice versa */
340         if (cpu_buf->last_is_kernel != is_kernel) {
341                 if (add_code(cpu_buf, is_kernel))
342                         goto fail;
343                 cpu_buf->last_is_kernel = is_kernel;
344         }
345
346         /* notice a task switch */
347         if (!is_kernel) {
348                 task = current;
349                 if (cpu_buf->last_task != task) {
350                         if (add_code(cpu_buf, (unsigned long)task))
351                                 goto fail;
352                         cpu_buf->last_task = task;
353                 }
354         }
355
356         fail = fail || add_code(cpu_buf, ibs_code);
357         fail = fail || op_add_sample(cpu_buf, ibs_sample[0], ibs_sample[1]);
358         fail = fail || op_add_sample(cpu_buf, ibs_sample[2], ibs_sample[3]);
359         fail = fail || op_add_sample(cpu_buf, ibs_sample[4], ibs_sample[5]);
360
361         if (ibs_code == IBS_OP_BEGIN) {
362                 fail = fail || op_add_sample(cpu_buf, ibs_sample[6], ibs_sample[7]);
363                 fail = fail || op_add_sample(cpu_buf, ibs_sample[8], ibs_sample[9]);
364                 fail = fail || op_add_sample(cpu_buf, ibs_sample[10], ibs_sample[11]);
365         }
366
367         if (!fail)
368                 return;
369
370 fail:
371         cpu_buf->sample_lost_overflow++;
372 }
373
374 #endif
375
376 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
377 {
378         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
379         log_sample(cpu_buf, pc, is_kernel, event);
380 }
381
382 void oprofile_add_trace(unsigned long pc)
383 {
384         struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
385
386         if (!cpu_buf->tracing)
387                 return;
388
389         /*
390          * broken frame can give an eip with the same value as an
391          * escape code, abort the trace if we get it
392          */
393         if (pc == ESCAPE_CODE)
394                 goto fail;
395
396         if (op_add_sample(cpu_buf, pc, 0))
397                 goto fail;
398
399         return;
400 fail:
401         cpu_buf->tracing = 0;
402         cpu_buf->backtrace_aborted++;
403         return;
404 }
405
406 /*
407  * This serves to avoid cpu buffer overflow, and makes sure
408  * the task mortuary progresses
409  *
410  * By using schedule_delayed_work_on and then schedule_delayed_work
411  * we guarantee this will stay on the correct cpu
412  */
413 static void wq_sync_buffer(struct work_struct *work)
414 {
415         struct oprofile_cpu_buffer *b =
416                 container_of(work, struct oprofile_cpu_buffer, work.work);
417         if (b->cpu != smp_processor_id()) {
418                 printk(KERN_DEBUG "WQ on CPU%d, prefer CPU%d\n",
419                        smp_processor_id(), b->cpu);
420
421                 if (!cpu_online(b->cpu)) {
422                         cancel_delayed_work(&b->work);
423                         return;
424                 }
425         }
426         sync_buffer(b->cpu);
427
428         /* don't re-add the work if we're shutting down */
429         if (work_enabled)
430                 schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE);
431 }