Commit | Line | Data |
---|---|---|
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 | ||
33 | /* | |
34 | * Read and write access is using spin locking. Thus, writing to the | |
35 | * buffer by NMI handler (x86) could occur also during critical | |
36 | * sections when reading the buffer. To avoid this, there are 2 | |
37 | * buffers for independent read and write access. Read access is in | |
38 | * process context only, write access only in the NMI handler. If the | |
39 | * read buffer runs empty, both buffers are swapped atomically. There | |
40 | * is potentially a small window during swapping where the buffers are | |
41 | * disabled and samples could be lost. | |
42 | * | |
43 | * Using 2 buffers is a little bit overhead, but the solution is clear | |
44 | * and does not require changes in the ring buffer implementation. It | |
45 | * can be changed to a single buffer solution when the ring buffer | |
46 | * access is implemented as non-locking atomic code. | |
47 | */ | |
9966718d RR |
48 | static struct ring_buffer *op_ring_buffer_read; |
49 | static struct ring_buffer *op_ring_buffer_write; | |
b3e9f672 | 50 | DEFINE_PER_CPU(struct oprofile_cpu_buffer, op_cpu_buffer); |
1da177e4 | 51 | |
c4028958 | 52 | static void wq_sync_buffer(struct work_struct *work); |
1da177e4 LT |
53 | |
54 | #define DEFAULT_TIMER_EXPIRE (HZ / 10) | |
55 | static int work_enabled; | |
56 | ||
a5598ca0 CL |
57 | unsigned long oprofile_get_cpu_buffer_size(void) |
58 | { | |
bd2172f5 | 59 | return oprofile_cpu_buffer_size; |
a5598ca0 CL |
60 | } |
61 | ||
62 | void oprofile_cpu_buffer_inc_smpl_lost(void) | |
63 | { | |
b3e9f672 | 64 | struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(op_cpu_buffer); |
a5598ca0 CL |
65 | |
66 | cpu_buf->sample_lost_overflow++; | |
67 | } | |
68 | ||
30015776 RR |
69 | void free_cpu_buffers(void) |
70 | { | |
71 | if (op_ring_buffer_read) | |
72 | ring_buffer_free(op_ring_buffer_read); | |
73 | op_ring_buffer_read = NULL; | |
74 | if (op_ring_buffer_write) | |
75 | ring_buffer_free(op_ring_buffer_write); | |
76 | op_ring_buffer_write = NULL; | |
77 | } | |
78 | ||
54f2c841 RR |
79 | #define RB_EVENT_HDR_SIZE 4 |
80 | ||
1da177e4 LT |
81 | int alloc_cpu_buffers(void) |
82 | { | |
83 | int i; | |
6a18037d | 84 | |
bd2172f5 | 85 | unsigned long buffer_size = oprofile_cpu_buffer_size; |
54f2c841 RR |
86 | unsigned long byte_size = buffer_size * (sizeof(struct op_sample) + |
87 | RB_EVENT_HDR_SIZE); | |
6a18037d | 88 | |
54f2c841 | 89 | op_ring_buffer_read = ring_buffer_alloc(byte_size, OP_BUFFER_FLAGS); |
6dad828b RR |
90 | if (!op_ring_buffer_read) |
91 | goto fail; | |
54f2c841 | 92 | op_ring_buffer_write = ring_buffer_alloc(byte_size, OP_BUFFER_FLAGS); |
6dad828b RR |
93 | if (!op_ring_buffer_write) |
94 | goto fail; | |
95 | ||
4bd9b9dc | 96 | for_each_possible_cpu(i) { |
b3e9f672 | 97 | struct oprofile_cpu_buffer *b = &per_cpu(op_cpu_buffer, i); |
6a18037d | 98 | |
1da177e4 LT |
99 | b->last_task = NULL; |
100 | b->last_is_kernel = -1; | |
101 | b->tracing = 0; | |
102 | b->buffer_size = buffer_size; | |
1da177e4 LT |
103 | b->sample_received = 0; |
104 | b->sample_lost_overflow = 0; | |
df9d177a PE |
105 | b->backtrace_aborted = 0; |
106 | b->sample_invalid_eip = 0; | |
1da177e4 | 107 | b->cpu = i; |
c4028958 | 108 | INIT_DELAYED_WORK(&b->work, wq_sync_buffer); |
1da177e4 LT |
109 | } |
110 | return 0; | |
111 | ||
112 | fail: | |
113 | free_cpu_buffers(); | |
114 | return -ENOMEM; | |
115 | } | |
1da177e4 LT |
116 | |
117 | void start_cpu_work(void) | |
118 | { | |
119 | int i; | |
120 | ||
121 | work_enabled = 1; | |
122 | ||
123 | for_each_online_cpu(i) { | |
b3e9f672 | 124 | struct oprofile_cpu_buffer *b = &per_cpu(op_cpu_buffer, i); |
1da177e4 LT |
125 | |
126 | /* | |
127 | * Spread the work by 1 jiffy per cpu so they dont all | |
128 | * fire at once. | |
129 | */ | |
130 | schedule_delayed_work_on(i, &b->work, DEFAULT_TIMER_EXPIRE + i); | |
131 | } | |
132 | } | |
133 | ||
1da177e4 LT |
134 | void end_cpu_work(void) |
135 | { | |
136 | int i; | |
137 | ||
138 | work_enabled = 0; | |
139 | ||
140 | for_each_online_cpu(i) { | |
b3e9f672 | 141 | struct oprofile_cpu_buffer *b = &per_cpu(op_cpu_buffer, i); |
1da177e4 LT |
142 | |
143 | cancel_delayed_work(&b->work); | |
144 | } | |
145 | ||
146 | flush_scheduled_work(); | |
147 | } | |
148 | ||
2cc28b9f RR |
149 | /* |
150 | * This function prepares the cpu buffer to write a sample. | |
151 | * | |
152 | * Struct op_entry is used during operations on the ring buffer while | |
153 | * struct op_sample contains the data that is stored in the ring | |
154 | * buffer. Struct entry can be uninitialized. The function reserves a | |
155 | * data array that is specified by size. Use | |
156 | * op_cpu_buffer_write_commit() after preparing the sample. In case of | |
157 | * errors a null pointer is returned, otherwise the pointer to the | |
158 | * sample. | |
159 | * | |
160 | */ | |
161 | struct op_sample | |
162 | *op_cpu_buffer_write_reserve(struct op_entry *entry, unsigned long size) | |
9966718d | 163 | { |
2cc28b9f RR |
164 | entry->event = ring_buffer_lock_reserve |
165 | (op_ring_buffer_write, sizeof(struct op_sample) + | |
304cc6ae | 166 | size * sizeof(entry->sample->data[0])); |
9966718d RR |
167 | if (entry->event) |
168 | entry->sample = ring_buffer_event_data(entry->event); | |
169 | else | |
170 | entry->sample = NULL; | |
171 | ||
172 | if (!entry->sample) | |
2cc28b9f | 173 | return NULL; |
9966718d | 174 | |
2cc28b9f RR |
175 | entry->size = size; |
176 | entry->data = entry->sample->data; | |
177 | ||
178 | return entry->sample; | |
9966718d RR |
179 | } |
180 | ||
181 | int op_cpu_buffer_write_commit(struct op_entry *entry) | |
182 | { | |
304cc6ae | 183 | return ring_buffer_unlock_commit(op_ring_buffer_write, entry->event); |
9966718d RR |
184 | } |
185 | ||
2d87b14c | 186 | struct op_sample *op_cpu_buffer_read_entry(struct op_entry *entry, int cpu) |
9966718d RR |
187 | { |
188 | struct ring_buffer_event *e; | |
189 | e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL); | |
190 | if (e) | |
2d87b14c | 191 | goto event; |
9966718d RR |
192 | if (ring_buffer_swap_cpu(op_ring_buffer_read, |
193 | op_ring_buffer_write, | |
194 | cpu)) | |
195 | return NULL; | |
196 | e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL); | |
197 | if (e) | |
2d87b14c | 198 | goto event; |
9966718d | 199 | return NULL; |
2d87b14c RR |
200 | |
201 | event: | |
202 | entry->event = e; | |
203 | entry->sample = ring_buffer_event_data(e); | |
204 | entry->size = (ring_buffer_event_length(e) - sizeof(struct op_sample)) | |
205 | / sizeof(entry->sample->data[0]); | |
206 | entry->data = entry->sample->data; | |
207 | return entry->sample; | |
9966718d RR |
208 | } |
209 | ||
210 | unsigned long op_cpu_buffer_entries(int cpu) | |
211 | { | |
212 | return ring_buffer_entries_cpu(op_ring_buffer_read, cpu) | |
213 | + ring_buffer_entries_cpu(op_ring_buffer_write, cpu); | |
214 | } | |
215 | ||
ae735e99 RR |
216 | static int |
217 | op_add_code(struct oprofile_cpu_buffer *cpu_buf, unsigned long backtrace, | |
218 | int is_kernel, struct task_struct *task) | |
219 | { | |
220 | struct op_entry entry; | |
221 | struct op_sample *sample; | |
222 | unsigned long flags; | |
223 | int size; | |
224 | ||
225 | flags = 0; | |
226 | ||
227 | if (backtrace) | |
228 | flags |= TRACE_BEGIN; | |
229 | ||
230 | /* notice a switch from user->kernel or vice versa */ | |
231 | is_kernel = !!is_kernel; | |
232 | if (cpu_buf->last_is_kernel != is_kernel) { | |
233 | cpu_buf->last_is_kernel = is_kernel; | |
234 | flags |= KERNEL_CTX_SWITCH; | |
235 | if (is_kernel) | |
236 | flags |= IS_KERNEL; | |
237 | } | |
238 | ||
239 | /* notice a task switch */ | |
240 | if (cpu_buf->last_task != task) { | |
241 | cpu_buf->last_task = task; | |
242 | flags |= USER_CTX_SWITCH; | |
243 | } | |
244 | ||
245 | if (!flags) | |
246 | /* nothing to do */ | |
247 | return 0; | |
248 | ||
249 | if (flags & USER_CTX_SWITCH) | |
250 | size = 1; | |
251 | else | |
252 | size = 0; | |
253 | ||
254 | sample = op_cpu_buffer_write_reserve(&entry, size); | |
255 | if (!sample) | |
256 | return -ENOMEM; | |
257 | ||
258 | sample->eip = ESCAPE_CODE; | |
259 | sample->event = flags; | |
260 | ||
261 | if (size) | |
d9928c25 | 262 | op_cpu_buffer_add_data(&entry, (unsigned long)task); |
ae735e99 RR |
263 | |
264 | op_cpu_buffer_write_commit(&entry); | |
265 | ||
266 | return 0; | |
267 | } | |
268 | ||
211117ff | 269 | static inline int |
d0e23384 RR |
270 | op_add_sample(struct oprofile_cpu_buffer *cpu_buf, |
271 | unsigned long pc, unsigned long event) | |
1da177e4 | 272 | { |
6dad828b | 273 | struct op_entry entry; |
2cc28b9f | 274 | struct op_sample *sample; |
6dad828b | 275 | |
2cc28b9f RR |
276 | sample = op_cpu_buffer_write_reserve(&entry, 0); |
277 | if (!sample) | |
278 | return -ENOMEM; | |
6dad828b | 279 | |
2cc28b9f RR |
280 | sample->eip = pc; |
281 | sample->event = event; | |
6dad828b | 282 | |
3967e93e | 283 | return op_cpu_buffer_write_commit(&entry); |
1da177e4 LT |
284 | } |
285 | ||
ae735e99 RR |
286 | /* |
287 | * This must be safe from any context. | |
1da177e4 LT |
288 | * |
289 | * is_kernel is needed because on some architectures you cannot | |
290 | * tell if you are in kernel or user space simply by looking at | |
291 | * pc. We tag this in the buffer by generating kernel enter/exit | |
292 | * events whenever is_kernel changes | |
293 | */ | |
ae735e99 RR |
294 | static int |
295 | log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc, | |
296 | unsigned long backtrace, int is_kernel, unsigned long event) | |
1da177e4 | 297 | { |
1da177e4 LT |
298 | cpu_buf->sample_received++; |
299 | ||
df9d177a PE |
300 | if (pc == ESCAPE_CODE) { |
301 | cpu_buf->sample_invalid_eip++; | |
302 | return 0; | |
303 | } | |
304 | ||
ae735e99 RR |
305 | if (op_add_code(cpu_buf, backtrace, is_kernel, current)) |
306 | goto fail; | |
6a18037d | 307 | |
d0e23384 | 308 | if (op_add_sample(cpu_buf, pc, event)) |
211117ff RR |
309 | goto fail; |
310 | ||
1da177e4 | 311 | return 1; |
211117ff RR |
312 | |
313 | fail: | |
314 | cpu_buf->sample_lost_overflow++; | |
315 | return 0; | |
1da177e4 LT |
316 | } |
317 | ||
6352d92d | 318 | static inline void oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf) |
1da177e4 | 319 | { |
1da177e4 | 320 | cpu_buf->tracing = 1; |
1da177e4 LT |
321 | } |
322 | ||
6352d92d | 323 | static inline void oprofile_end_trace(struct oprofile_cpu_buffer *cpu_buf) |
1da177e4 LT |
324 | { |
325 | cpu_buf->tracing = 0; | |
326 | } | |
327 | ||
d45d23be RR |
328 | static inline void |
329 | __oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs, | |
330 | unsigned long event, int is_kernel) | |
1da177e4 | 331 | { |
b3e9f672 | 332 | struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(op_cpu_buffer); |
ae735e99 | 333 | unsigned long backtrace = oprofile_backtrace_depth; |
1da177e4 | 334 | |
fd13f6c8 RR |
335 | /* |
336 | * if log_sample() fail we can't backtrace since we lost the | |
337 | * source of this event | |
338 | */ | |
ae735e99 RR |
339 | if (!log_sample(cpu_buf, pc, backtrace, is_kernel, event)) |
340 | /* failed */ | |
341 | return; | |
342 | ||
343 | if (!backtrace) | |
344 | return; | |
6352d92d | 345 | |
ae735e99 RR |
346 | oprofile_begin_trace(cpu_buf); |
347 | oprofile_ops.backtrace(regs, backtrace); | |
1da177e4 LT |
348 | oprofile_end_trace(cpu_buf); |
349 | } | |
350 | ||
d45d23be RR |
351 | void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs, |
352 | unsigned long event, int is_kernel) | |
353 | { | |
354 | __oprofile_add_ext_sample(pc, regs, event, is_kernel); | |
355 | } | |
356 | ||
27357716 BR |
357 | void oprofile_add_sample(struct pt_regs * const regs, unsigned long event) |
358 | { | |
359 | int is_kernel = !user_mode(regs); | |
360 | unsigned long pc = profile_pc(regs); | |
361 | ||
d45d23be | 362 | __oprofile_add_ext_sample(pc, regs, event, is_kernel); |
27357716 BR |
363 | } |
364 | ||
1acda878 RR |
365 | /* |
366 | * Add samples with data to the ring buffer. | |
367 | * | |
14f0ca8e RR |
368 | * Use oprofile_add_data(&entry, val) to add data and |
369 | * oprofile_write_commit(&entry) to commit the sample. | |
1acda878 | 370 | */ |
14f0ca8e RR |
371 | void |
372 | oprofile_write_reserve(struct op_entry *entry, struct pt_regs * const regs, | |
1acda878 | 373 | unsigned long pc, int code, int size) |
345c2573 | 374 | { |
1acda878 | 375 | struct op_sample *sample; |
e2fee276 | 376 | int is_kernel = !user_mode(regs); |
b3e9f672 | 377 | struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(op_cpu_buffer); |
345c2573 BK |
378 | |
379 | cpu_buf->sample_received++; | |
380 | ||
1acda878 RR |
381 | /* no backtraces for samples with data */ |
382 | if (op_add_code(cpu_buf, 0, is_kernel, current)) | |
383 | goto fail; | |
384 | ||
385 | sample = op_cpu_buffer_write_reserve(entry, size + 2); | |
386 | if (!sample) | |
387 | goto fail; | |
388 | sample->eip = ESCAPE_CODE; | |
389 | sample->event = 0; /* no flags */ | |
345c2573 | 390 | |
1acda878 RR |
391 | op_cpu_buffer_add_data(entry, code); |
392 | op_cpu_buffer_add_data(entry, pc); | |
345c2573 | 393 | |
1acda878 | 394 | return; |
345c2573 | 395 | |
1acda878 | 396 | fail: |
fdb6a8f4 | 397 | entry->event = NULL; |
1acda878 | 398 | cpu_buf->sample_lost_overflow++; |
345c2573 BK |
399 | } |
400 | ||
14f0ca8e RR |
401 | int oprofile_add_data(struct op_entry *entry, unsigned long val) |
402 | { | |
fdb6a8f4 RR |
403 | if (!entry->event) |
404 | return 0; | |
14f0ca8e RR |
405 | return op_cpu_buffer_add_data(entry, val); |
406 | } | |
407 | ||
51563a0e RR |
408 | int oprofile_add_data64(struct op_entry *entry, u64 val) |
409 | { | |
410 | if (!entry->event) | |
411 | return 0; | |
412 | if (op_cpu_buffer_get_size(entry) < 2) | |
413 | /* | |
414 | * the function returns 0 to indicate a too small | |
415 | * buffer, even if there is some space left | |
416 | */ | |
417 | return 0; | |
418 | if (!op_cpu_buffer_add_data(entry, (u32)val)) | |
419 | return 0; | |
420 | return op_cpu_buffer_add_data(entry, (u32)(val >> 32)); | |
421 | } | |
422 | ||
14f0ca8e RR |
423 | int oprofile_write_commit(struct op_entry *entry) |
424 | { | |
fdb6a8f4 RR |
425 | if (!entry->event) |
426 | return -EINVAL; | |
14f0ca8e RR |
427 | return op_cpu_buffer_write_commit(entry); |
428 | } | |
429 | ||
1da177e4 LT |
430 | void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event) |
431 | { | |
b3e9f672 | 432 | struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(op_cpu_buffer); |
ae735e99 | 433 | log_sample(cpu_buf, pc, 0, is_kernel, event); |
1da177e4 LT |
434 | } |
435 | ||
1da177e4 LT |
436 | void oprofile_add_trace(unsigned long pc) |
437 | { | |
b3e9f672 | 438 | struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(op_cpu_buffer); |
1da177e4 LT |
439 | |
440 | if (!cpu_buf->tracing) | |
441 | return; | |
442 | ||
fd13f6c8 RR |
443 | /* |
444 | * broken frame can give an eip with the same value as an | |
445 | * escape code, abort the trace if we get it | |
446 | */ | |
211117ff RR |
447 | if (pc == ESCAPE_CODE) |
448 | goto fail; | |
449 | ||
d0e23384 | 450 | if (op_add_sample(cpu_buf, pc, 0)) |
211117ff | 451 | goto fail; |
1da177e4 | 452 | |
211117ff RR |
453 | return; |
454 | fail: | |
455 | cpu_buf->tracing = 0; | |
456 | cpu_buf->backtrace_aborted++; | |
457 | return; | |
1da177e4 LT |
458 | } |
459 | ||
1da177e4 LT |
460 | /* |
461 | * This serves to avoid cpu buffer overflow, and makes sure | |
462 | * the task mortuary progresses | |
463 | * | |
464 | * By using schedule_delayed_work_on and then schedule_delayed_work | |
465 | * we guarantee this will stay on the correct cpu | |
466 | */ | |
c4028958 | 467 | static void wq_sync_buffer(struct work_struct *work) |
1da177e4 | 468 | { |
25ad2913 | 469 | struct oprofile_cpu_buffer *b = |
c4028958 | 470 | container_of(work, struct oprofile_cpu_buffer, work.work); |
1da177e4 | 471 | if (b->cpu != smp_processor_id()) { |
bd17b625 | 472 | printk(KERN_DEBUG "WQ on CPU%d, prefer CPU%d\n", |
1da177e4 | 473 | smp_processor_id(), b->cpu); |
4bd9b9dc CA |
474 | |
475 | if (!cpu_online(b->cpu)) { | |
476 | cancel_delayed_work(&b->work); | |
477 | return; | |
478 | } | |
1da177e4 LT |
479 | } |
480 | sync_buffer(b->cpu); | |
481 | ||
482 | /* don't re-add the work if we're shutting down */ | |
483 | if (work_enabled) | |
484 | schedule_delayed_work(&b->work, DEFAULT_TIMER_EXPIRE); | |
485 | } |