]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Generic ring buffer | |
3 | * | |
4 | * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com> | |
5 | */ | |
6 | #include <linux/ring_buffer.h> | |
7 | #include <linux/trace_clock.h> | |
8 | #include <linux/ftrace_irq.h> | |
9 | #include <linux/spinlock.h> | |
10 | #include <linux/debugfs.h> | |
11 | #include <linux/uaccess.h> | |
12 | #include <linux/hardirq.h> | |
13 | #include <linux/kmemcheck.h> | |
14 | #include <linux/module.h> | |
15 | #include <linux/percpu.h> | |
16 | #include <linux/mutex.h> | |
17 | #include <linux/slab.h> | |
18 | #include <linux/init.h> | |
19 | #include <linux/hash.h> | |
20 | #include <linux/list.h> | |
21 | #include <linux/cpu.h> | |
22 | #include <linux/fs.h> | |
23 | ||
24 | #include <asm/local.h> | |
25 | #include "trace.h" | |
26 | ||
27 | /* | |
28 | * The ring buffer header is special. We must manually up keep it. | |
29 | */ | |
30 | int ring_buffer_print_entry_header(struct trace_seq *s) | |
31 | { | |
32 | int ret; | |
33 | ||
34 | ret = trace_seq_printf(s, "# compressed entry header\n"); | |
35 | ret = trace_seq_printf(s, "\ttype_len : 5 bits\n"); | |
36 | ret = trace_seq_printf(s, "\ttime_delta : 27 bits\n"); | |
37 | ret = trace_seq_printf(s, "\tarray : 32 bits\n"); | |
38 | ret = trace_seq_printf(s, "\n"); | |
39 | ret = trace_seq_printf(s, "\tpadding : type == %d\n", | |
40 | RINGBUF_TYPE_PADDING); | |
41 | ret = trace_seq_printf(s, "\ttime_extend : type == %d\n", | |
42 | RINGBUF_TYPE_TIME_EXTEND); | |
43 | ret = trace_seq_printf(s, "\tdata max type_len == %d\n", | |
44 | RINGBUF_TYPE_DATA_TYPE_LEN_MAX); | |
45 | ||
46 | return ret; | |
47 | } | |
48 | ||
49 | /* | |
50 | * The ring buffer is made up of a list of pages. A separate list of pages is | |
51 | * allocated for each CPU. A writer may only write to a buffer that is | |
52 | * associated with the CPU it is currently executing on. A reader may read | |
53 | * from any per cpu buffer. | |
54 | * | |
55 | * The reader is special. For each per cpu buffer, the reader has its own | |
56 | * reader page. When a reader has read the entire reader page, this reader | |
57 | * page is swapped with another page in the ring buffer. | |
58 | * | |
59 | * Now, as long as the writer is off the reader page, the reader can do what | |
60 | * ever it wants with that page. The writer will never write to that page | |
61 | * again (as long as it is out of the ring buffer). | |
62 | * | |
63 | * Here's some silly ASCII art. | |
64 | * | |
65 | * +------+ | |
66 | * |reader| RING BUFFER | |
67 | * |page | | |
68 | * +------+ +---+ +---+ +---+ | |
69 | * | |-->| |-->| | | |
70 | * +---+ +---+ +---+ | |
71 | * ^ | | |
72 | * | | | |
73 | * +---------------+ | |
74 | * | |
75 | * | |
76 | * +------+ | |
77 | * |reader| RING BUFFER | |
78 | * |page |------------------v | |
79 | * +------+ +---+ +---+ +---+ | |
80 | * | |-->| |-->| | | |
81 | * +---+ +---+ +---+ | |
82 | * ^ | | |
83 | * | | | |
84 | * +---------------+ | |
85 | * | |
86 | * | |
87 | * +------+ | |
88 | * |reader| RING BUFFER | |
89 | * |page |------------------v | |
90 | * +------+ +---+ +---+ +---+ | |
91 | * ^ | |-->| |-->| | | |
92 | * | +---+ +---+ +---+ | |
93 | * | | | |
94 | * | | | |
95 | * +------------------------------+ | |
96 | * | |
97 | * | |
98 | * +------+ | |
99 | * |buffer| RING BUFFER | |
100 | * |page |------------------v | |
101 | * +------+ +---+ +---+ +---+ | |
102 | * ^ | | | |-->| | | |
103 | * | New +---+ +---+ +---+ | |
104 | * | Reader------^ | | |
105 | * | page | | |
106 | * +------------------------------+ | |
107 | * | |
108 | * | |
109 | * After we make this swap, the reader can hand this page off to the splice | |
110 | * code and be done with it. It can even allocate a new page if it needs to | |
111 | * and swap that into the ring buffer. | |
112 | * | |
113 | * We will be using cmpxchg soon to make all this lockless. | |
114 | * | |
115 | */ | |
116 | ||
117 | /* | |
118 | * A fast way to enable or disable all ring buffers is to | |
119 | * call tracing_on or tracing_off. Turning off the ring buffers | |
120 | * prevents all ring buffers from being recorded to. | |
121 | * Turning this switch on, makes it OK to write to the | |
122 | * ring buffer, if the ring buffer is enabled itself. | |
123 | * | |
124 | * There's three layers that must be on in order to write | |
125 | * to the ring buffer. | |
126 | * | |
127 | * 1) This global flag must be set. | |
128 | * 2) The ring buffer must be enabled for recording. | |
129 | * 3) The per cpu buffer must be enabled for recording. | |
130 | * | |
131 | * In case of an anomaly, this global flag has a bit set that | |
132 | * will permantly disable all ring buffers. | |
133 | */ | |
134 | ||
135 | /* | |
136 | * Global flag to disable all recording to ring buffers | |
137 | * This has two bits: ON, DISABLED | |
138 | * | |
139 | * ON DISABLED | |
140 | * ---- ---------- | |
141 | * 0 0 : ring buffers are off | |
142 | * 1 0 : ring buffers are on | |
143 | * X 1 : ring buffers are permanently disabled | |
144 | */ | |
145 | ||
146 | enum { | |
147 | RB_BUFFERS_ON_BIT = 0, | |
148 | RB_BUFFERS_DISABLED_BIT = 1, | |
149 | }; | |
150 | ||
151 | enum { | |
152 | RB_BUFFERS_ON = 1 << RB_BUFFERS_ON_BIT, | |
153 | RB_BUFFERS_DISABLED = 1 << RB_BUFFERS_DISABLED_BIT, | |
154 | }; | |
155 | ||
156 | static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON; | |
157 | ||
158 | #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data) | |
159 | ||
160 | /** | |
161 | * tracing_on - enable all tracing buffers | |
162 | * | |
163 | * This function enables all tracing buffers that may have been | |
164 | * disabled with tracing_off. | |
165 | */ | |
166 | void tracing_on(void) | |
167 | { | |
168 | set_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags); | |
169 | } | |
170 | EXPORT_SYMBOL_GPL(tracing_on); | |
171 | ||
172 | /** | |
173 | * tracing_off - turn off all tracing buffers | |
174 | * | |
175 | * This function stops all tracing buffers from recording data. | |
176 | * It does not disable any overhead the tracers themselves may | |
177 | * be causing. This function simply causes all recording to | |
178 | * the ring buffers to fail. | |
179 | */ | |
180 | void tracing_off(void) | |
181 | { | |
182 | clear_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags); | |
183 | } | |
184 | EXPORT_SYMBOL_GPL(tracing_off); | |
185 | ||
186 | /** | |
187 | * tracing_off_permanent - permanently disable ring buffers | |
188 | * | |
189 | * This function, once called, will disable all ring buffers | |
190 | * permanently. | |
191 | */ | |
192 | void tracing_off_permanent(void) | |
193 | { | |
194 | set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags); | |
195 | } | |
196 | ||
197 | /** | |
198 | * tracing_is_on - show state of ring buffers enabled | |
199 | */ | |
200 | int tracing_is_on(void) | |
201 | { | |
202 | return ring_buffer_flags == RB_BUFFERS_ON; | |
203 | } | |
204 | EXPORT_SYMBOL_GPL(tracing_is_on); | |
205 | ||
206 | #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array)) | |
207 | #define RB_ALIGNMENT 4U | |
208 | #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX) | |
209 | #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */ | |
210 | ||
211 | #if !defined(CONFIG_64BIT) || defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) | |
212 | # define RB_FORCE_8BYTE_ALIGNMENT 0 | |
213 | # define RB_ARCH_ALIGNMENT RB_ALIGNMENT | |
214 | #else | |
215 | # define RB_FORCE_8BYTE_ALIGNMENT 1 | |
216 | # define RB_ARCH_ALIGNMENT 8U | |
217 | #endif | |
218 | ||
219 | /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */ | |
220 | #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX | |
221 | ||
222 | enum { | |
223 | RB_LEN_TIME_EXTEND = 8, | |
224 | RB_LEN_TIME_STAMP = 16, | |
225 | }; | |
226 | ||
227 | static inline int rb_null_event(struct ring_buffer_event *event) | |
228 | { | |
229 | return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta; | |
230 | } | |
231 | ||
232 | static void rb_event_set_padding(struct ring_buffer_event *event) | |
233 | { | |
234 | /* padding has a NULL time_delta */ | |
235 | event->type_len = RINGBUF_TYPE_PADDING; | |
236 | event->time_delta = 0; | |
237 | } | |
238 | ||
239 | static unsigned | |
240 | rb_event_data_length(struct ring_buffer_event *event) | |
241 | { | |
242 | unsigned length; | |
243 | ||
244 | if (event->type_len) | |
245 | length = event->type_len * RB_ALIGNMENT; | |
246 | else | |
247 | length = event->array[0]; | |
248 | return length + RB_EVNT_HDR_SIZE; | |
249 | } | |
250 | ||
251 | /* inline for ring buffer fast paths */ | |
252 | static unsigned | |
253 | rb_event_length(struct ring_buffer_event *event) | |
254 | { | |
255 | switch (event->type_len) { | |
256 | case RINGBUF_TYPE_PADDING: | |
257 | if (rb_null_event(event)) | |
258 | /* undefined */ | |
259 | return -1; | |
260 | return event->array[0] + RB_EVNT_HDR_SIZE; | |
261 | ||
262 | case RINGBUF_TYPE_TIME_EXTEND: | |
263 | return RB_LEN_TIME_EXTEND; | |
264 | ||
265 | case RINGBUF_TYPE_TIME_STAMP: | |
266 | return RB_LEN_TIME_STAMP; | |
267 | ||
268 | case RINGBUF_TYPE_DATA: | |
269 | return rb_event_data_length(event); | |
270 | default: | |
271 | BUG(); | |
272 | } | |
273 | /* not hit */ | |
274 | return 0; | |
275 | } | |
276 | ||
277 | /** | |
278 | * ring_buffer_event_length - return the length of the event | |
279 | * @event: the event to get the length of | |
280 | */ | |
281 | unsigned ring_buffer_event_length(struct ring_buffer_event *event) | |
282 | { | |
283 | unsigned length = rb_event_length(event); | |
284 | if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX) | |
285 | return length; | |
286 | length -= RB_EVNT_HDR_SIZE; | |
287 | if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0])) | |
288 | length -= sizeof(event->array[0]); | |
289 | return length; | |
290 | } | |
291 | EXPORT_SYMBOL_GPL(ring_buffer_event_length); | |
292 | ||
293 | /* inline for ring buffer fast paths */ | |
294 | static void * | |
295 | rb_event_data(struct ring_buffer_event *event) | |
296 | { | |
297 | BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX); | |
298 | /* If length is in len field, then array[0] has the data */ | |
299 | if (event->type_len) | |
300 | return (void *)&event->array[0]; | |
301 | /* Otherwise length is in array[0] and array[1] has the data */ | |
302 | return (void *)&event->array[1]; | |
303 | } | |
304 | ||
305 | /** | |
306 | * ring_buffer_event_data - return the data of the event | |
307 | * @event: the event to get the data from | |
308 | */ | |
309 | void *ring_buffer_event_data(struct ring_buffer_event *event) | |
310 | { | |
311 | return rb_event_data(event); | |
312 | } | |
313 | EXPORT_SYMBOL_GPL(ring_buffer_event_data); | |
314 | ||
315 | #define for_each_buffer_cpu(buffer, cpu) \ | |
316 | for_each_cpu(cpu, buffer->cpumask) | |
317 | ||
318 | #define TS_SHIFT 27 | |
319 | #define TS_MASK ((1ULL << TS_SHIFT) - 1) | |
320 | #define TS_DELTA_TEST (~TS_MASK) | |
321 | ||
322 | /* Flag when events were overwritten */ | |
323 | #define RB_MISSED_EVENTS (1 << 31) | |
324 | /* Missed count stored at end */ | |
325 | #define RB_MISSED_STORED (1 << 30) | |
326 | ||
327 | struct buffer_data_page { | |
328 | u64 time_stamp; /* page time stamp */ | |
329 | local_t commit; /* write committed index */ | |
330 | unsigned char data[]; /* data of buffer page */ | |
331 | }; | |
332 | ||
333 | /* | |
334 | * Note, the buffer_page list must be first. The buffer pages | |
335 | * are allocated in cache lines, which means that each buffer | |
336 | * page will be at the beginning of a cache line, and thus | |
337 | * the least significant bits will be zero. We use this to | |
338 | * add flags in the list struct pointers, to make the ring buffer | |
339 | * lockless. | |
340 | */ | |
341 | struct buffer_page { | |
342 | struct list_head list; /* list of buffer pages */ | |
343 | local_t write; /* index for next write */ | |
344 | unsigned read; /* index for next read */ | |
345 | local_t entries; /* entries on this page */ | |
346 | unsigned long real_end; /* real end of data */ | |
347 | struct buffer_data_page *page; /* Actual data page */ | |
348 | }; | |
349 | ||
350 | /* | |
351 | * The buffer page counters, write and entries, must be reset | |
352 | * atomically when crossing page boundaries. To synchronize this | |
353 | * update, two counters are inserted into the number. One is | |
354 | * the actual counter for the write position or count on the page. | |
355 | * | |
356 | * The other is a counter of updaters. Before an update happens | |
357 | * the update partition of the counter is incremented. This will | |
358 | * allow the updater to update the counter atomically. | |
359 | * | |
360 | * The counter is 20 bits, and the state data is 12. | |
361 | */ | |
362 | #define RB_WRITE_MASK 0xfffff | |
363 | #define RB_WRITE_INTCNT (1 << 20) | |
364 | ||
365 | static void rb_init_page(struct buffer_data_page *bpage) | |
366 | { | |
367 | local_set(&bpage->commit, 0); | |
368 | } | |
369 | ||
370 | /** | |
371 | * ring_buffer_page_len - the size of data on the page. | |
372 | * @page: The page to read | |
373 | * | |
374 | * Returns the amount of data on the page, including buffer page header. | |
375 | */ | |
376 | size_t ring_buffer_page_len(void *page) | |
377 | { | |
378 | return local_read(&((struct buffer_data_page *)page)->commit) | |
379 | + BUF_PAGE_HDR_SIZE; | |
380 | } | |
381 | ||
382 | /* | |
383 | * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing | |
384 | * this issue out. | |
385 | */ | |
386 | static void free_buffer_page(struct buffer_page *bpage) | |
387 | { | |
388 | free_page((unsigned long)bpage->page); | |
389 | kfree(bpage); | |
390 | } | |
391 | ||
392 | /* | |
393 | * We need to fit the time_stamp delta into 27 bits. | |
394 | */ | |
395 | static inline int test_time_stamp(u64 delta) | |
396 | { | |
397 | if (delta & TS_DELTA_TEST) | |
398 | return 1; | |
399 | return 0; | |
400 | } | |
401 | ||
402 | #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE) | |
403 | ||
404 | /* Max payload is BUF_PAGE_SIZE - header (8bytes) */ | |
405 | #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2)) | |
406 | ||
407 | /* Max number of timestamps that can fit on a page */ | |
408 | #define RB_TIMESTAMPS_PER_PAGE (BUF_PAGE_SIZE / RB_LEN_TIME_EXTEND) | |
409 | ||
410 | int ring_buffer_print_page_header(struct trace_seq *s) | |
411 | { | |
412 | struct buffer_data_page field; | |
413 | int ret; | |
414 | ||
415 | ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t" | |
416 | "offset:0;\tsize:%u;\tsigned:%u;\n", | |
417 | (unsigned int)sizeof(field.time_stamp), | |
418 | (unsigned int)is_signed_type(u64)); | |
419 | ||
420 | ret = trace_seq_printf(s, "\tfield: local_t commit;\t" | |
421 | "offset:%u;\tsize:%u;\tsigned:%u;\n", | |
422 | (unsigned int)offsetof(typeof(field), commit), | |
423 | (unsigned int)sizeof(field.commit), | |
424 | (unsigned int)is_signed_type(long)); | |
425 | ||
426 | ret = trace_seq_printf(s, "\tfield: int overwrite;\t" | |
427 | "offset:%u;\tsize:%u;\tsigned:%u;\n", | |
428 | (unsigned int)offsetof(typeof(field), commit), | |
429 | 1, | |
430 | (unsigned int)is_signed_type(long)); | |
431 | ||
432 | ret = trace_seq_printf(s, "\tfield: char data;\t" | |
433 | "offset:%u;\tsize:%u;\tsigned:%u;\n", | |
434 | (unsigned int)offsetof(typeof(field), data), | |
435 | (unsigned int)BUF_PAGE_SIZE, | |
436 | (unsigned int)is_signed_type(char)); | |
437 | ||
438 | return ret; | |
439 | } | |
440 | ||
441 | /* | |
442 | * head_page == tail_page && head == tail then buffer is empty. | |
443 | */ | |
444 | struct ring_buffer_per_cpu { | |
445 | int cpu; | |
446 | atomic_t record_disabled; | |
447 | struct ring_buffer *buffer; | |
448 | spinlock_t reader_lock; /* serialize readers */ | |
449 | arch_spinlock_t lock; | |
450 | struct lock_class_key lock_key; | |
451 | struct list_head *pages; | |
452 | struct buffer_page *head_page; /* read from head */ | |
453 | struct buffer_page *tail_page; /* write to tail */ | |
454 | struct buffer_page *commit_page; /* committed pages */ | |
455 | struct buffer_page *reader_page; | |
456 | unsigned long lost_events; | |
457 | unsigned long last_overrun; | |
458 | local_t commit_overrun; | |
459 | local_t overrun; | |
460 | local_t entries; | |
461 | local_t committing; | |
462 | local_t commits; | |
463 | unsigned long read; | |
464 | u64 write_stamp; | |
465 | u64 read_stamp; | |
466 | }; | |
467 | ||
468 | struct ring_buffer { | |
469 | unsigned pages; | |
470 | unsigned flags; | |
471 | int cpus; | |
472 | atomic_t record_disabled; | |
473 | cpumask_var_t cpumask; | |
474 | ||
475 | struct lock_class_key *reader_lock_key; | |
476 | ||
477 | struct mutex mutex; | |
478 | ||
479 | struct ring_buffer_per_cpu **buffers; | |
480 | ||
481 | #ifdef CONFIG_HOTPLUG_CPU | |
482 | struct notifier_block cpu_notify; | |
483 | #endif | |
484 | u64 (*clock)(void); | |
485 | }; | |
486 | ||
487 | struct ring_buffer_iter { | |
488 | struct ring_buffer_per_cpu *cpu_buffer; | |
489 | unsigned long head; | |
490 | struct buffer_page *head_page; | |
491 | struct buffer_page *cache_reader_page; | |
492 | unsigned long cache_read; | |
493 | u64 read_stamp; | |
494 | }; | |
495 | ||
496 | /* buffer may be either ring_buffer or ring_buffer_per_cpu */ | |
497 | #define RB_WARN_ON(b, cond) \ | |
498 | ({ \ | |
499 | int _____ret = unlikely(cond); \ | |
500 | if (_____ret) { \ | |
501 | if (__same_type(*(b), struct ring_buffer_per_cpu)) { \ | |
502 | struct ring_buffer_per_cpu *__b = \ | |
503 | (void *)b; \ | |
504 | atomic_inc(&__b->buffer->record_disabled); \ | |
505 | } else \ | |
506 | atomic_inc(&b->record_disabled); \ | |
507 | WARN_ON(1); \ | |
508 | } \ | |
509 | _____ret; \ | |
510 | }) | |
511 | ||
512 | /* Up this if you want to test the TIME_EXTENTS and normalization */ | |
513 | #define DEBUG_SHIFT 0 | |
514 | ||
515 | static inline u64 rb_time_stamp(struct ring_buffer *buffer) | |
516 | { | |
517 | /* shift to debug/test normalization and TIME_EXTENTS */ | |
518 | return buffer->clock() << DEBUG_SHIFT; | |
519 | } | |
520 | ||
521 | u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu) | |
522 | { | |
523 | u64 time; | |
524 | ||
525 | preempt_disable_notrace(); | |
526 | time = rb_time_stamp(buffer); | |
527 | preempt_enable_no_resched_notrace(); | |
528 | ||
529 | return time; | |
530 | } | |
531 | EXPORT_SYMBOL_GPL(ring_buffer_time_stamp); | |
532 | ||
533 | void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer, | |
534 | int cpu, u64 *ts) | |
535 | { | |
536 | /* Just stupid testing the normalize function and deltas */ | |
537 | *ts >>= DEBUG_SHIFT; | |
538 | } | |
539 | EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp); | |
540 | ||
541 | /* | |
542 | * Making the ring buffer lockless makes things tricky. | |
543 | * Although writes only happen on the CPU that they are on, | |
544 | * and they only need to worry about interrupts. Reads can | |
545 | * happen on any CPU. | |
546 | * | |
547 | * The reader page is always off the ring buffer, but when the | |
548 | * reader finishes with a page, it needs to swap its page with | |
549 | * a new one from the buffer. The reader needs to take from | |
550 | * the head (writes go to the tail). But if a writer is in overwrite | |
551 | * mode and wraps, it must push the head page forward. | |
552 | * | |
553 | * Here lies the problem. | |
554 | * | |
555 | * The reader must be careful to replace only the head page, and | |
556 | * not another one. As described at the top of the file in the | |
557 | * ASCII art, the reader sets its old page to point to the next | |
558 | * page after head. It then sets the page after head to point to | |
559 | * the old reader page. But if the writer moves the head page | |
560 | * during this operation, the reader could end up with the tail. | |
561 | * | |
562 | * We use cmpxchg to help prevent this race. We also do something | |
563 | * special with the page before head. We set the LSB to 1. | |
564 | * | |
565 | * When the writer must push the page forward, it will clear the | |
566 | * bit that points to the head page, move the head, and then set | |
567 | * the bit that points to the new head page. | |
568 | * | |
569 | * We also don't want an interrupt coming in and moving the head | |
570 | * page on another writer. Thus we use the second LSB to catch | |
571 | * that too. Thus: | |
572 | * | |
573 | * head->list->prev->next bit 1 bit 0 | |
574 | * ------- ------- | |
575 | * Normal page 0 0 | |
576 | * Points to head page 0 1 | |
577 | * New head page 1 0 | |
578 | * | |
579 | * Note we can not trust the prev pointer of the head page, because: | |
580 | * | |
581 | * +----+ +-----+ +-----+ | |
582 | * | |------>| T |---X--->| N | | |
583 | * | |<------| | | | | |
584 | * +----+ +-----+ +-----+ | |
585 | * ^ ^ | | |
586 | * | +-----+ | | | |
587 | * +----------| R |----------+ | | |
588 | * | |<-----------+ | |
589 | * +-----+ | |
590 | * | |
591 | * Key: ---X--> HEAD flag set in pointer | |
592 | * T Tail page | |
593 | * R Reader page | |
594 | * N Next page | |
595 | * | |
596 | * (see __rb_reserve_next() to see where this happens) | |
597 | * | |
598 | * What the above shows is that the reader just swapped out | |
599 | * the reader page with a page in the buffer, but before it | |
600 | * could make the new header point back to the new page added | |
601 | * it was preempted by a writer. The writer moved forward onto | |
602 | * the new page added by the reader and is about to move forward | |
603 | * again. | |
604 | * | |
605 | * You can see, it is legitimate for the previous pointer of | |
606 | * the head (or any page) not to point back to itself. But only | |
607 | * temporarially. | |
608 | */ | |
609 | ||
610 | #define RB_PAGE_NORMAL 0UL | |
611 | #define RB_PAGE_HEAD 1UL | |
612 | #define RB_PAGE_UPDATE 2UL | |
613 | ||
614 | ||
615 | #define RB_FLAG_MASK 3UL | |
616 | ||
617 | /* PAGE_MOVED is not part of the mask */ | |
618 | #define RB_PAGE_MOVED 4UL | |
619 | ||
620 | /* | |
621 | * rb_list_head - remove any bit | |
622 | */ | |
623 | static struct list_head *rb_list_head(struct list_head *list) | |
624 | { | |
625 | unsigned long val = (unsigned long)list; | |
626 | ||
627 | return (struct list_head *)(val & ~RB_FLAG_MASK); | |
628 | } | |
629 | ||
630 | /* | |
631 | * rb_is_head_page - test if the given page is the head page | |
632 | * | |
633 | * Because the reader may move the head_page pointer, we can | |
634 | * not trust what the head page is (it may be pointing to | |
635 | * the reader page). But if the next page is a header page, | |
636 | * its flags will be non zero. | |
637 | */ | |
638 | static int inline | |
639 | rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer, | |
640 | struct buffer_page *page, struct list_head *list) | |
641 | { | |
642 | unsigned long val; | |
643 | ||
644 | val = (unsigned long)list->next; | |
645 | ||
646 | if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list) | |
647 | return RB_PAGE_MOVED; | |
648 | ||
649 | return val & RB_FLAG_MASK; | |
650 | } | |
651 | ||
652 | /* | |
653 | * rb_is_reader_page | |
654 | * | |
655 | * The unique thing about the reader page, is that, if the | |
656 | * writer is ever on it, the previous pointer never points | |
657 | * back to the reader page. | |
658 | */ | |
659 | static int rb_is_reader_page(struct buffer_page *page) | |
660 | { | |
661 | struct list_head *list = page->list.prev; | |
662 | ||
663 | return rb_list_head(list->next) != &page->list; | |
664 | } | |
665 | ||
666 | /* | |
667 | * rb_set_list_to_head - set a list_head to be pointing to head. | |
668 | */ | |
669 | static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer, | |
670 | struct list_head *list) | |
671 | { | |
672 | unsigned long *ptr; | |
673 | ||
674 | ptr = (unsigned long *)&list->next; | |
675 | *ptr |= RB_PAGE_HEAD; | |
676 | *ptr &= ~RB_PAGE_UPDATE; | |
677 | } | |
678 | ||
679 | /* | |
680 | * rb_head_page_activate - sets up head page | |
681 | */ | |
682 | static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer) | |
683 | { | |
684 | struct buffer_page *head; | |
685 | ||
686 | head = cpu_buffer->head_page; | |
687 | if (!head) | |
688 | return; | |
689 | ||
690 | /* | |
691 | * Set the previous list pointer to have the HEAD flag. | |
692 | */ | |
693 | rb_set_list_to_head(cpu_buffer, head->list.prev); | |
694 | } | |
695 | ||
696 | static void rb_list_head_clear(struct list_head *list) | |
697 | { | |
698 | unsigned long *ptr = (unsigned long *)&list->next; | |
699 | ||
700 | *ptr &= ~RB_FLAG_MASK; | |
701 | } | |
702 | ||
703 | /* | |
704 | * rb_head_page_dactivate - clears head page ptr (for free list) | |
705 | */ | |
706 | static void | |
707 | rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer) | |
708 | { | |
709 | struct list_head *hd; | |
710 | ||
711 | /* Go through the whole list and clear any pointers found. */ | |
712 | rb_list_head_clear(cpu_buffer->pages); | |
713 | ||
714 | list_for_each(hd, cpu_buffer->pages) | |
715 | rb_list_head_clear(hd); | |
716 | } | |
717 | ||
718 | static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer, | |
719 | struct buffer_page *head, | |
720 | struct buffer_page *prev, | |
721 | int old_flag, int new_flag) | |
722 | { | |
723 | struct list_head *list; | |
724 | unsigned long val = (unsigned long)&head->list; | |
725 | unsigned long ret; | |
726 | ||
727 | list = &prev->list; | |
728 | ||
729 | val &= ~RB_FLAG_MASK; | |
730 | ||
731 | ret = cmpxchg((unsigned long *)&list->next, | |
732 | val | old_flag, val | new_flag); | |
733 | ||
734 | /* check if the reader took the page */ | |
735 | if ((ret & ~RB_FLAG_MASK) != val) | |
736 | return RB_PAGE_MOVED; | |
737 | ||
738 | return ret & RB_FLAG_MASK; | |
739 | } | |
740 | ||
741 | static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer, | |
742 | struct buffer_page *head, | |
743 | struct buffer_page *prev, | |
744 | int old_flag) | |
745 | { | |
746 | return rb_head_page_set(cpu_buffer, head, prev, | |
747 | old_flag, RB_PAGE_UPDATE); | |
748 | } | |
749 | ||
750 | static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer, | |
751 | struct buffer_page *head, | |
752 | struct buffer_page *prev, | |
753 | int old_flag) | |
754 | { | |
755 | return rb_head_page_set(cpu_buffer, head, prev, | |
756 | old_flag, RB_PAGE_HEAD); | |
757 | } | |
758 | ||
759 | static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer, | |
760 | struct buffer_page *head, | |
761 | struct buffer_page *prev, | |
762 | int old_flag) | |
763 | { | |
764 | return rb_head_page_set(cpu_buffer, head, prev, | |
765 | old_flag, RB_PAGE_NORMAL); | |
766 | } | |
767 | ||
768 | static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer, | |
769 | struct buffer_page **bpage) | |
770 | { | |
771 | struct list_head *p = rb_list_head((*bpage)->list.next); | |
772 | ||
773 | *bpage = list_entry(p, struct buffer_page, list); | |
774 | } | |
775 | ||
776 | static struct buffer_page * | |
777 | rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer) | |
778 | { | |
779 | struct buffer_page *head; | |
780 | struct buffer_page *page; | |
781 | struct list_head *list; | |
782 | int i; | |
783 | ||
784 | if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page)) | |
785 | return NULL; | |
786 | ||
787 | /* sanity check */ | |
788 | list = cpu_buffer->pages; | |
789 | if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list)) | |
790 | return NULL; | |
791 | ||
792 | page = head = cpu_buffer->head_page; | |
793 | /* | |
794 | * It is possible that the writer moves the header behind | |
795 | * where we started, and we miss in one loop. | |
796 | * A second loop should grab the header, but we'll do | |
797 | * three loops just because I'm paranoid. | |
798 | */ | |
799 | for (i = 0; i < 3; i++) { | |
800 | do { | |
801 | if (rb_is_head_page(cpu_buffer, page, page->list.prev)) { | |
802 | cpu_buffer->head_page = page; | |
803 | return page; | |
804 | } | |
805 | rb_inc_page(cpu_buffer, &page); | |
806 | } while (page != head); | |
807 | } | |
808 | ||
809 | RB_WARN_ON(cpu_buffer, 1); | |
810 | ||
811 | return NULL; | |
812 | } | |
813 | ||
814 | static int rb_head_page_replace(struct buffer_page *old, | |
815 | struct buffer_page *new) | |
816 | { | |
817 | unsigned long *ptr = (unsigned long *)&old->list.prev->next; | |
818 | unsigned long val; | |
819 | unsigned long ret; | |
820 | ||
821 | val = *ptr & ~RB_FLAG_MASK; | |
822 | val |= RB_PAGE_HEAD; | |
823 | ||
824 | ret = cmpxchg(ptr, val, (unsigned long)&new->list); | |
825 | ||
826 | return ret == val; | |
827 | } | |
828 | ||
829 | /* | |
830 | * rb_tail_page_update - move the tail page forward | |
831 | * | |
832 | * Returns 1 if moved tail page, 0 if someone else did. | |
833 | */ | |
834 | static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer, | |
835 | struct buffer_page *tail_page, | |
836 | struct buffer_page *next_page) | |
837 | { | |
838 | struct buffer_page *old_tail; | |
839 | unsigned long old_entries; | |
840 | unsigned long old_write; | |
841 | int ret = 0; | |
842 | ||
843 | /* | |
844 | * The tail page now needs to be moved forward. | |
845 | * | |
846 | * We need to reset the tail page, but without messing | |
847 | * with possible erasing of data brought in by interrupts | |
848 | * that have moved the tail page and are currently on it. | |
849 | * | |
850 | * We add a counter to the write field to denote this. | |
851 | */ | |
852 | old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write); | |
853 | old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries); | |
854 | ||
855 | /* | |
856 | * Just make sure we have seen our old_write and synchronize | |
857 | * with any interrupts that come in. | |
858 | */ | |
859 | barrier(); | |
860 | ||
861 | /* | |
862 | * If the tail page is still the same as what we think | |
863 | * it is, then it is up to us to update the tail | |
864 | * pointer. | |
865 | */ | |
866 | if (tail_page == cpu_buffer->tail_page) { | |
867 | /* Zero the write counter */ | |
868 | unsigned long val = old_write & ~RB_WRITE_MASK; | |
869 | unsigned long eval = old_entries & ~RB_WRITE_MASK; | |
870 | ||
871 | /* | |
872 | * This will only succeed if an interrupt did | |
873 | * not come in and change it. In which case, we | |
874 | * do not want to modify it. | |
875 | * | |
876 | * We add (void) to let the compiler know that we do not care | |
877 | * about the return value of these functions. We use the | |
878 | * cmpxchg to only update if an interrupt did not already | |
879 | * do it for us. If the cmpxchg fails, we don't care. | |
880 | */ | |
881 | (void)local_cmpxchg(&next_page->write, old_write, val); | |
882 | (void)local_cmpxchg(&next_page->entries, old_entries, eval); | |
883 | ||
884 | /* | |
885 | * No need to worry about races with clearing out the commit. | |
886 | * it only can increment when a commit takes place. But that | |
887 | * only happens in the outer most nested commit. | |
888 | */ | |
889 | local_set(&next_page->page->commit, 0); | |
890 | ||
891 | old_tail = cmpxchg(&cpu_buffer->tail_page, | |
892 | tail_page, next_page); | |
893 | ||
894 | if (old_tail == tail_page) | |
895 | ret = 1; | |
896 | } | |
897 | ||
898 | return ret; | |
899 | } | |
900 | ||
901 | static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer, | |
902 | struct buffer_page *bpage) | |
903 | { | |
904 | unsigned long val = (unsigned long)bpage; | |
905 | ||
906 | if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK)) | |
907 | return 1; | |
908 | ||
909 | return 0; | |
910 | } | |
911 | ||
912 | /** | |
913 | * rb_check_list - make sure a pointer to a list has the last bits zero | |
914 | */ | |
915 | static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer, | |
916 | struct list_head *list) | |
917 | { | |
918 | if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev)) | |
919 | return 1; | |
920 | if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next)) | |
921 | return 1; | |
922 | return 0; | |
923 | } | |
924 | ||
925 | /** | |
926 | * check_pages - integrity check of buffer pages | |
927 | * @cpu_buffer: CPU buffer with pages to test | |
928 | * | |
929 | * As a safety measure we check to make sure the data pages have not | |
930 | * been corrupted. | |
931 | */ | |
932 | static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer) | |
933 | { | |
934 | struct list_head *head = cpu_buffer->pages; | |
935 | struct buffer_page *bpage, *tmp; | |
936 | ||
937 | rb_head_page_deactivate(cpu_buffer); | |
938 | ||
939 | if (RB_WARN_ON(cpu_buffer, head->next->prev != head)) | |
940 | return -1; | |
941 | if (RB_WARN_ON(cpu_buffer, head->prev->next != head)) | |
942 | return -1; | |
943 | ||
944 | if (rb_check_list(cpu_buffer, head)) | |
945 | return -1; | |
946 | ||
947 | list_for_each_entry_safe(bpage, tmp, head, list) { | |
948 | if (RB_WARN_ON(cpu_buffer, | |
949 | bpage->list.next->prev != &bpage->list)) | |
950 | return -1; | |
951 | if (RB_WARN_ON(cpu_buffer, | |
952 | bpage->list.prev->next != &bpage->list)) | |
953 | return -1; | |
954 | if (rb_check_list(cpu_buffer, &bpage->list)) | |
955 | return -1; | |
956 | } | |
957 | ||
958 | rb_head_page_activate(cpu_buffer); | |
959 | ||
960 | return 0; | |
961 | } | |
962 | ||
963 | static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer, | |
964 | unsigned nr_pages) | |
965 | { | |
966 | struct buffer_page *bpage, *tmp; | |
967 | unsigned long addr; | |
968 | LIST_HEAD(pages); | |
969 | unsigned i; | |
970 | ||
971 | WARN_ON(!nr_pages); | |
972 | ||
973 | for (i = 0; i < nr_pages; i++) { | |
974 | bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()), | |
975 | GFP_KERNEL, cpu_to_node(cpu_buffer->cpu)); | |
976 | if (!bpage) | |
977 | goto free_pages; | |
978 | ||
979 | rb_check_bpage(cpu_buffer, bpage); | |
980 | ||
981 | list_add(&bpage->list, &pages); | |
982 | ||
983 | addr = __get_free_page(GFP_KERNEL); | |
984 | if (!addr) | |
985 | goto free_pages; | |
986 | bpage->page = (void *)addr; | |
987 | rb_init_page(bpage->page); | |
988 | } | |
989 | ||
990 | /* | |
991 | * The ring buffer page list is a circular list that does not | |
992 | * start and end with a list head. All page list items point to | |
993 | * other pages. | |
994 | */ | |
995 | cpu_buffer->pages = pages.next; | |
996 | list_del(&pages); | |
997 | ||
998 | rb_check_pages(cpu_buffer); | |
999 | ||
1000 | return 0; | |
1001 | ||
1002 | free_pages: | |
1003 | list_for_each_entry_safe(bpage, tmp, &pages, list) { | |
1004 | list_del_init(&bpage->list); | |
1005 | free_buffer_page(bpage); | |
1006 | } | |
1007 | return -ENOMEM; | |
1008 | } | |
1009 | ||
1010 | static struct ring_buffer_per_cpu * | |
1011 | rb_allocate_cpu_buffer(struct ring_buffer *buffer, int cpu) | |
1012 | { | |
1013 | struct ring_buffer_per_cpu *cpu_buffer; | |
1014 | struct buffer_page *bpage; | |
1015 | unsigned long addr; | |
1016 | int ret; | |
1017 | ||
1018 | cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()), | |
1019 | GFP_KERNEL, cpu_to_node(cpu)); | |
1020 | if (!cpu_buffer) | |
1021 | return NULL; | |
1022 | ||
1023 | cpu_buffer->cpu = cpu; | |
1024 | cpu_buffer->buffer = buffer; | |
1025 | spin_lock_init(&cpu_buffer->reader_lock); | |
1026 | lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key); | |
1027 | cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED; | |
1028 | ||
1029 | bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()), | |
1030 | GFP_KERNEL, cpu_to_node(cpu)); | |
1031 | if (!bpage) | |
1032 | goto fail_free_buffer; | |
1033 | ||
1034 | rb_check_bpage(cpu_buffer, bpage); | |
1035 | ||
1036 | cpu_buffer->reader_page = bpage; | |
1037 | addr = __get_free_page(GFP_KERNEL); | |
1038 | if (!addr) | |
1039 | goto fail_free_reader; | |
1040 | bpage->page = (void *)addr; | |
1041 | rb_init_page(bpage->page); | |
1042 | ||
1043 | INIT_LIST_HEAD(&cpu_buffer->reader_page->list); | |
1044 | ||
1045 | ret = rb_allocate_pages(cpu_buffer, buffer->pages); | |
1046 | if (ret < 0) | |
1047 | goto fail_free_reader; | |
1048 | ||
1049 | cpu_buffer->head_page | |
1050 | = list_entry(cpu_buffer->pages, struct buffer_page, list); | |
1051 | cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page; | |
1052 | ||
1053 | rb_head_page_activate(cpu_buffer); | |
1054 | ||
1055 | return cpu_buffer; | |
1056 | ||
1057 | fail_free_reader: | |
1058 | free_buffer_page(cpu_buffer->reader_page); | |
1059 | ||
1060 | fail_free_buffer: | |
1061 | kfree(cpu_buffer); | |
1062 | return NULL; | |
1063 | } | |
1064 | ||
1065 | static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer) | |
1066 | { | |
1067 | struct list_head *head = cpu_buffer->pages; | |
1068 | struct buffer_page *bpage, *tmp; | |
1069 | ||
1070 | free_buffer_page(cpu_buffer->reader_page); | |
1071 | ||
1072 | rb_head_page_deactivate(cpu_buffer); | |
1073 | ||
1074 | if (head) { | |
1075 | list_for_each_entry_safe(bpage, tmp, head, list) { | |
1076 | list_del_init(&bpage->list); | |
1077 | free_buffer_page(bpage); | |
1078 | } | |
1079 | bpage = list_entry(head, struct buffer_page, list); | |
1080 | free_buffer_page(bpage); | |
1081 | } | |
1082 | ||
1083 | kfree(cpu_buffer); | |
1084 | } | |
1085 | ||
1086 | #ifdef CONFIG_HOTPLUG_CPU | |
1087 | static int rb_cpu_notify(struct notifier_block *self, | |
1088 | unsigned long action, void *hcpu); | |
1089 | #endif | |
1090 | ||
1091 | /** | |
1092 | * ring_buffer_alloc - allocate a new ring_buffer | |
1093 | * @size: the size in bytes per cpu that is needed. | |
1094 | * @flags: attributes to set for the ring buffer. | |
1095 | * | |
1096 | * Currently the only flag that is available is the RB_FL_OVERWRITE | |
1097 | * flag. This flag means that the buffer will overwrite old data | |
1098 | * when the buffer wraps. If this flag is not set, the buffer will | |
1099 | * drop data when the tail hits the head. | |
1100 | */ | |
1101 | struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags, | |
1102 | struct lock_class_key *key) | |
1103 | { | |
1104 | struct ring_buffer *buffer; | |
1105 | int bsize; | |
1106 | int cpu; | |
1107 | ||
1108 | /* keep it in its own cache line */ | |
1109 | buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()), | |
1110 | GFP_KERNEL); | |
1111 | if (!buffer) | |
1112 | return NULL; | |
1113 | ||
1114 | if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL)) | |
1115 | goto fail_free_buffer; | |
1116 | ||
1117 | buffer->pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE); | |
1118 | buffer->flags = flags; | |
1119 | buffer->clock = trace_clock_local; | |
1120 | buffer->reader_lock_key = key; | |
1121 | ||
1122 | /* need at least two pages */ | |
1123 | if (buffer->pages < 2) | |
1124 | buffer->pages = 2; | |
1125 | ||
1126 | /* | |
1127 | * In case of non-hotplug cpu, if the ring-buffer is allocated | |
1128 | * in early initcall, it will not be notified of secondary cpus. | |
1129 | * In that off case, we need to allocate for all possible cpus. | |
1130 | */ | |
1131 | #ifdef CONFIG_HOTPLUG_CPU | |
1132 | get_online_cpus(); | |
1133 | cpumask_copy(buffer->cpumask, cpu_online_mask); | |
1134 | #else | |
1135 | cpumask_copy(buffer->cpumask, cpu_possible_mask); | |
1136 | #endif | |
1137 | buffer->cpus = nr_cpu_ids; | |
1138 | ||
1139 | bsize = sizeof(void *) * nr_cpu_ids; | |
1140 | buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()), | |
1141 | GFP_KERNEL); | |
1142 | if (!buffer->buffers) | |
1143 | goto fail_free_cpumask; | |
1144 | ||
1145 | for_each_buffer_cpu(buffer, cpu) { | |
1146 | buffer->buffers[cpu] = | |
1147 | rb_allocate_cpu_buffer(buffer, cpu); | |
1148 | if (!buffer->buffers[cpu]) | |
1149 | goto fail_free_buffers; | |
1150 | } | |
1151 | ||
1152 | #ifdef CONFIG_HOTPLUG_CPU | |
1153 | buffer->cpu_notify.notifier_call = rb_cpu_notify; | |
1154 | buffer->cpu_notify.priority = 0; | |
1155 | register_cpu_notifier(&buffer->cpu_notify); | |
1156 | #endif | |
1157 | ||
1158 | put_online_cpus(); | |
1159 | mutex_init(&buffer->mutex); | |
1160 | ||
1161 | return buffer; | |
1162 | ||
1163 | fail_free_buffers: | |
1164 | for_each_buffer_cpu(buffer, cpu) { | |
1165 | if (buffer->buffers[cpu]) | |
1166 | rb_free_cpu_buffer(buffer->buffers[cpu]); | |
1167 | } | |
1168 | kfree(buffer->buffers); | |
1169 | ||
1170 | fail_free_cpumask: | |
1171 | free_cpumask_var(buffer->cpumask); | |
1172 | put_online_cpus(); | |
1173 | ||
1174 | fail_free_buffer: | |
1175 | kfree(buffer); | |
1176 | return NULL; | |
1177 | } | |
1178 | EXPORT_SYMBOL_GPL(__ring_buffer_alloc); | |
1179 | ||
1180 | /** | |
1181 | * ring_buffer_free - free a ring buffer. | |
1182 | * @buffer: the buffer to free. | |
1183 | */ | |
1184 | void | |
1185 | ring_buffer_free(struct ring_buffer *buffer) | |
1186 | { | |
1187 | int cpu; | |
1188 | ||
1189 | get_online_cpus(); | |
1190 | ||
1191 | #ifdef CONFIG_HOTPLUG_CPU | |
1192 | unregister_cpu_notifier(&buffer->cpu_notify); | |
1193 | #endif | |
1194 | ||
1195 | for_each_buffer_cpu(buffer, cpu) | |
1196 | rb_free_cpu_buffer(buffer->buffers[cpu]); | |
1197 | ||
1198 | put_online_cpus(); | |
1199 | ||
1200 | kfree(buffer->buffers); | |
1201 | free_cpumask_var(buffer->cpumask); | |
1202 | ||
1203 | kfree(buffer); | |
1204 | } | |
1205 | EXPORT_SYMBOL_GPL(ring_buffer_free); | |
1206 | ||
1207 | void ring_buffer_set_clock(struct ring_buffer *buffer, | |
1208 | u64 (*clock)(void)) | |
1209 | { | |
1210 | buffer->clock = clock; | |
1211 | } | |
1212 | ||
1213 | static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer); | |
1214 | ||
1215 | static void | |
1216 | rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned nr_pages) | |
1217 | { | |
1218 | struct buffer_page *bpage; | |
1219 | struct list_head *p; | |
1220 | unsigned i; | |
1221 | ||
1222 | spin_lock_irq(&cpu_buffer->reader_lock); | |
1223 | rb_head_page_deactivate(cpu_buffer); | |
1224 | ||
1225 | for (i = 0; i < nr_pages; i++) { | |
1226 | if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages))) | |
1227 | goto out; | |
1228 | p = cpu_buffer->pages->next; | |
1229 | bpage = list_entry(p, struct buffer_page, list); | |
1230 | list_del_init(&bpage->list); | |
1231 | free_buffer_page(bpage); | |
1232 | } | |
1233 | if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages))) | |
1234 | goto out; | |
1235 | ||
1236 | rb_reset_cpu(cpu_buffer); | |
1237 | rb_check_pages(cpu_buffer); | |
1238 | ||
1239 | out: | |
1240 | spin_unlock_irq(&cpu_buffer->reader_lock); | |
1241 | } | |
1242 | ||
1243 | static void | |
1244 | rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer, | |
1245 | struct list_head *pages, unsigned nr_pages) | |
1246 | { | |
1247 | struct buffer_page *bpage; | |
1248 | struct list_head *p; | |
1249 | unsigned i; | |
1250 | ||
1251 | spin_lock_irq(&cpu_buffer->reader_lock); | |
1252 | rb_head_page_deactivate(cpu_buffer); | |
1253 | ||
1254 | for (i = 0; i < nr_pages; i++) { | |
1255 | if (RB_WARN_ON(cpu_buffer, list_empty(pages))) | |
1256 | goto out; | |
1257 | p = pages->next; | |
1258 | bpage = list_entry(p, struct buffer_page, list); | |
1259 | list_del_init(&bpage->list); | |
1260 | list_add_tail(&bpage->list, cpu_buffer->pages); | |
1261 | } | |
1262 | rb_reset_cpu(cpu_buffer); | |
1263 | rb_check_pages(cpu_buffer); | |
1264 | ||
1265 | out: | |
1266 | spin_unlock_irq(&cpu_buffer->reader_lock); | |
1267 | } | |
1268 | ||
1269 | /** | |
1270 | * ring_buffer_resize - resize the ring buffer | |
1271 | * @buffer: the buffer to resize. | |
1272 | * @size: the new size. | |
1273 | * | |
1274 | * Minimum size is 2 * BUF_PAGE_SIZE. | |
1275 | * | |
1276 | * Returns -1 on failure. | |
1277 | */ | |
1278 | int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size) | |
1279 | { | |
1280 | struct ring_buffer_per_cpu *cpu_buffer; | |
1281 | unsigned nr_pages, rm_pages, new_pages; | |
1282 | struct buffer_page *bpage, *tmp; | |
1283 | unsigned long buffer_size; | |
1284 | unsigned long addr; | |
1285 | LIST_HEAD(pages); | |
1286 | int i, cpu; | |
1287 | ||
1288 | /* | |
1289 | * Always succeed at resizing a non-existent buffer: | |
1290 | */ | |
1291 | if (!buffer) | |
1292 | return size; | |
1293 | ||
1294 | size = DIV_ROUND_UP(size, BUF_PAGE_SIZE); | |
1295 | size *= BUF_PAGE_SIZE; | |
1296 | buffer_size = buffer->pages * BUF_PAGE_SIZE; | |
1297 | ||
1298 | /* we need a minimum of two pages */ | |
1299 | if (size < BUF_PAGE_SIZE * 2) | |
1300 | size = BUF_PAGE_SIZE * 2; | |
1301 | ||
1302 | if (size == buffer_size) | |
1303 | return size; | |
1304 | ||
1305 | atomic_inc(&buffer->record_disabled); | |
1306 | ||
1307 | /* Make sure all writers are done with this buffer. */ | |
1308 | synchronize_sched(); | |
1309 | ||
1310 | mutex_lock(&buffer->mutex); | |
1311 | get_online_cpus(); | |
1312 | ||
1313 | nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE); | |
1314 | ||
1315 | if (size < buffer_size) { | |
1316 | ||
1317 | /* easy case, just free pages */ | |
1318 | if (RB_WARN_ON(buffer, nr_pages >= buffer->pages)) | |
1319 | goto out_fail; | |
1320 | ||
1321 | rm_pages = buffer->pages - nr_pages; | |
1322 | ||
1323 | for_each_buffer_cpu(buffer, cpu) { | |
1324 | cpu_buffer = buffer->buffers[cpu]; | |
1325 | rb_remove_pages(cpu_buffer, rm_pages); | |
1326 | } | |
1327 | goto out; | |
1328 | } | |
1329 | ||
1330 | /* | |
1331 | * This is a bit more difficult. We only want to add pages | |
1332 | * when we can allocate enough for all CPUs. We do this | |
1333 | * by allocating all the pages and storing them on a local | |
1334 | * link list. If we succeed in our allocation, then we | |
1335 | * add these pages to the cpu_buffers. Otherwise we just free | |
1336 | * them all and return -ENOMEM; | |
1337 | */ | |
1338 | if (RB_WARN_ON(buffer, nr_pages <= buffer->pages)) | |
1339 | goto out_fail; | |
1340 | ||
1341 | new_pages = nr_pages - buffer->pages; | |
1342 | ||
1343 | for_each_buffer_cpu(buffer, cpu) { | |
1344 | for (i = 0; i < new_pages; i++) { | |
1345 | bpage = kzalloc_node(ALIGN(sizeof(*bpage), | |
1346 | cache_line_size()), | |
1347 | GFP_KERNEL, cpu_to_node(cpu)); | |
1348 | if (!bpage) | |
1349 | goto free_pages; | |
1350 | list_add(&bpage->list, &pages); | |
1351 | addr = __get_free_page(GFP_KERNEL); | |
1352 | if (!addr) | |
1353 | goto free_pages; | |
1354 | bpage->page = (void *)addr; | |
1355 | rb_init_page(bpage->page); | |
1356 | } | |
1357 | } | |
1358 | ||
1359 | for_each_buffer_cpu(buffer, cpu) { | |
1360 | cpu_buffer = buffer->buffers[cpu]; | |
1361 | rb_insert_pages(cpu_buffer, &pages, new_pages); | |
1362 | } | |
1363 | ||
1364 | if (RB_WARN_ON(buffer, !list_empty(&pages))) | |
1365 | goto out_fail; | |
1366 | ||
1367 | out: | |
1368 | buffer->pages = nr_pages; | |
1369 | put_online_cpus(); | |
1370 | mutex_unlock(&buffer->mutex); | |
1371 | ||
1372 | atomic_dec(&buffer->record_disabled); | |
1373 | ||
1374 | return size; | |
1375 | ||
1376 | free_pages: | |
1377 | list_for_each_entry_safe(bpage, tmp, &pages, list) { | |
1378 | list_del_init(&bpage->list); | |
1379 | free_buffer_page(bpage); | |
1380 | } | |
1381 | put_online_cpus(); | |
1382 | mutex_unlock(&buffer->mutex); | |
1383 | atomic_dec(&buffer->record_disabled); | |
1384 | return -ENOMEM; | |
1385 | ||
1386 | /* | |
1387 | * Something went totally wrong, and we are too paranoid | |
1388 | * to even clean up the mess. | |
1389 | */ | |
1390 | out_fail: | |
1391 | put_online_cpus(); | |
1392 | mutex_unlock(&buffer->mutex); | |
1393 | atomic_dec(&buffer->record_disabled); | |
1394 | return -1; | |
1395 | } | |
1396 | EXPORT_SYMBOL_GPL(ring_buffer_resize); | |
1397 | ||
1398 | static inline void * | |
1399 | __rb_data_page_index(struct buffer_data_page *bpage, unsigned index) | |
1400 | { | |
1401 | return bpage->data + index; | |
1402 | } | |
1403 | ||
1404 | static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index) | |
1405 | { | |
1406 | return bpage->page->data + index; | |
1407 | } | |
1408 | ||
1409 | static inline struct ring_buffer_event * | |
1410 | rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer) | |
1411 | { | |
1412 | return __rb_page_index(cpu_buffer->reader_page, | |
1413 | cpu_buffer->reader_page->read); | |
1414 | } | |
1415 | ||
1416 | static inline struct ring_buffer_event * | |
1417 | rb_iter_head_event(struct ring_buffer_iter *iter) | |
1418 | { | |
1419 | return __rb_page_index(iter->head_page, iter->head); | |
1420 | } | |
1421 | ||
1422 | static inline unsigned long rb_page_write(struct buffer_page *bpage) | |
1423 | { | |
1424 | return local_read(&bpage->write) & RB_WRITE_MASK; | |
1425 | } | |
1426 | ||
1427 | static inline unsigned rb_page_commit(struct buffer_page *bpage) | |
1428 | { | |
1429 | return local_read(&bpage->page->commit); | |
1430 | } | |
1431 | ||
1432 | static inline unsigned long rb_page_entries(struct buffer_page *bpage) | |
1433 | { | |
1434 | return local_read(&bpage->entries) & RB_WRITE_MASK; | |
1435 | } | |
1436 | ||
1437 | /* Size is determined by what has been commited */ | |
1438 | static inline unsigned rb_page_size(struct buffer_page *bpage) | |
1439 | { | |
1440 | return rb_page_commit(bpage); | |
1441 | } | |
1442 | ||
1443 | static inline unsigned | |
1444 | rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer) | |
1445 | { | |
1446 | return rb_page_commit(cpu_buffer->commit_page); | |
1447 | } | |
1448 | ||
1449 | static inline unsigned | |
1450 | rb_event_index(struct ring_buffer_event *event) | |
1451 | { | |
1452 | unsigned long addr = (unsigned long)event; | |
1453 | ||
1454 | return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE; | |
1455 | } | |
1456 | ||
1457 | static inline int | |
1458 | rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer, | |
1459 | struct ring_buffer_event *event) | |
1460 | { | |
1461 | unsigned long addr = (unsigned long)event; | |
1462 | unsigned long index; | |
1463 | ||
1464 | index = rb_event_index(event); | |
1465 | addr &= PAGE_MASK; | |
1466 | ||
1467 | return cpu_buffer->commit_page->page == (void *)addr && | |
1468 | rb_commit_index(cpu_buffer) == index; | |
1469 | } | |
1470 | ||
1471 | static void | |
1472 | rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer) | |
1473 | { | |
1474 | unsigned long max_count; | |
1475 | ||
1476 | /* | |
1477 | * We only race with interrupts and NMIs on this CPU. | |
1478 | * If we own the commit event, then we can commit | |
1479 | * all others that interrupted us, since the interruptions | |
1480 | * are in stack format (they finish before they come | |
1481 | * back to us). This allows us to do a simple loop to | |
1482 | * assign the commit to the tail. | |
1483 | */ | |
1484 | again: | |
1485 | max_count = cpu_buffer->buffer->pages * 100; | |
1486 | ||
1487 | while (cpu_buffer->commit_page != cpu_buffer->tail_page) { | |
1488 | if (RB_WARN_ON(cpu_buffer, !(--max_count))) | |
1489 | return; | |
1490 | if (RB_WARN_ON(cpu_buffer, | |
1491 | rb_is_reader_page(cpu_buffer->tail_page))) | |
1492 | return; | |
1493 | local_set(&cpu_buffer->commit_page->page->commit, | |
1494 | rb_page_write(cpu_buffer->commit_page)); | |
1495 | rb_inc_page(cpu_buffer, &cpu_buffer->commit_page); | |
1496 | cpu_buffer->write_stamp = | |
1497 | cpu_buffer->commit_page->page->time_stamp; | |
1498 | /* add barrier to keep gcc from optimizing too much */ | |
1499 | barrier(); | |
1500 | } | |
1501 | while (rb_commit_index(cpu_buffer) != | |
1502 | rb_page_write(cpu_buffer->commit_page)) { | |
1503 | ||
1504 | local_set(&cpu_buffer->commit_page->page->commit, | |
1505 | rb_page_write(cpu_buffer->commit_page)); | |
1506 | RB_WARN_ON(cpu_buffer, | |
1507 | local_read(&cpu_buffer->commit_page->page->commit) & | |
1508 | ~RB_WRITE_MASK); | |
1509 | barrier(); | |
1510 | } | |
1511 | ||
1512 | /* again, keep gcc from optimizing */ | |
1513 | barrier(); | |
1514 | ||
1515 | /* | |
1516 | * If an interrupt came in just after the first while loop | |
1517 | * and pushed the tail page forward, we will be left with | |
1518 | * a dangling commit that will never go forward. | |
1519 | */ | |
1520 | if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page)) | |
1521 | goto again; | |
1522 | } | |
1523 | ||
1524 | static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer) | |
1525 | { | |
1526 | cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp; | |
1527 | cpu_buffer->reader_page->read = 0; | |
1528 | } | |
1529 | ||
1530 | static void rb_inc_iter(struct ring_buffer_iter *iter) | |
1531 | { | |
1532 | struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer; | |
1533 | ||
1534 | /* | |
1535 | * The iterator could be on the reader page (it starts there). | |
1536 | * But the head could have moved, since the reader was | |
1537 | * found. Check for this case and assign the iterator | |
1538 | * to the head page instead of next. | |
1539 | */ | |
1540 | if (iter->head_page == cpu_buffer->reader_page) | |
1541 | iter->head_page = rb_set_head_page(cpu_buffer); | |
1542 | else | |
1543 | rb_inc_page(cpu_buffer, &iter->head_page); | |
1544 | ||
1545 | iter->read_stamp = iter->head_page->page->time_stamp; | |
1546 | iter->head = 0; | |
1547 | } | |
1548 | ||
1549 | /** | |
1550 | * ring_buffer_update_event - update event type and data | |
1551 | * @event: the even to update | |
1552 | * @type: the type of event | |
1553 | * @length: the size of the event field in the ring buffer | |
1554 | * | |
1555 | * Update the type and data fields of the event. The length | |
1556 | * is the actual size that is written to the ring buffer, | |
1557 | * and with this, we can determine what to place into the | |
1558 | * data field. | |
1559 | */ | |
1560 | static void | |
1561 | rb_update_event(struct ring_buffer_event *event, | |
1562 | unsigned type, unsigned length) | |
1563 | { | |
1564 | event->type_len = type; | |
1565 | ||
1566 | switch (type) { | |
1567 | ||
1568 | case RINGBUF_TYPE_PADDING: | |
1569 | case RINGBUF_TYPE_TIME_EXTEND: | |
1570 | case RINGBUF_TYPE_TIME_STAMP: | |
1571 | break; | |
1572 | ||
1573 | case 0: | |
1574 | length -= RB_EVNT_HDR_SIZE; | |
1575 | if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) | |
1576 | event->array[0] = length; | |
1577 | else | |
1578 | event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT); | |
1579 | break; | |
1580 | default: | |
1581 | BUG(); | |
1582 | } | |
1583 | } | |
1584 | ||
1585 | /* | |
1586 | * rb_handle_head_page - writer hit the head page | |
1587 | * | |
1588 | * Returns: +1 to retry page | |
1589 | * 0 to continue | |
1590 | * -1 on error | |
1591 | */ | |
1592 | static int | |
1593 | rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer, | |
1594 | struct buffer_page *tail_page, | |
1595 | struct buffer_page *next_page) | |
1596 | { | |
1597 | struct buffer_page *new_head; | |
1598 | int entries; | |
1599 | int type; | |
1600 | int ret; | |
1601 | ||
1602 | entries = rb_page_entries(next_page); | |
1603 | ||
1604 | /* | |
1605 | * The hard part is here. We need to move the head | |
1606 | * forward, and protect against both readers on | |
1607 | * other CPUs and writers coming in via interrupts. | |
1608 | */ | |
1609 | type = rb_head_page_set_update(cpu_buffer, next_page, tail_page, | |
1610 | RB_PAGE_HEAD); | |
1611 | ||
1612 | /* | |
1613 | * type can be one of four: | |
1614 | * NORMAL - an interrupt already moved it for us | |
1615 | * HEAD - we are the first to get here. | |
1616 | * UPDATE - we are the interrupt interrupting | |
1617 | * a current move. | |
1618 | * MOVED - a reader on another CPU moved the next | |
1619 | * pointer to its reader page. Give up | |
1620 | * and try again. | |
1621 | */ | |
1622 | ||
1623 | switch (type) { | |
1624 | case RB_PAGE_HEAD: | |
1625 | /* | |
1626 | * We changed the head to UPDATE, thus | |
1627 | * it is our responsibility to update | |
1628 | * the counters. | |
1629 | */ | |
1630 | local_add(entries, &cpu_buffer->overrun); | |
1631 | ||
1632 | /* | |
1633 | * The entries will be zeroed out when we move the | |
1634 | * tail page. | |
1635 | */ | |
1636 | ||
1637 | /* still more to do */ | |
1638 | break; | |
1639 | ||
1640 | case RB_PAGE_UPDATE: | |
1641 | /* | |
1642 | * This is an interrupt that interrupt the | |
1643 | * previous update. Still more to do. | |
1644 | */ | |
1645 | break; | |
1646 | case RB_PAGE_NORMAL: | |
1647 | /* | |
1648 | * An interrupt came in before the update | |
1649 | * and processed this for us. | |
1650 | * Nothing left to do. | |
1651 | */ | |
1652 | return 1; | |
1653 | case RB_PAGE_MOVED: | |
1654 | /* | |
1655 | * The reader is on another CPU and just did | |
1656 | * a swap with our next_page. | |
1657 | * Try again. | |
1658 | */ | |
1659 | return 1; | |
1660 | default: | |
1661 | RB_WARN_ON(cpu_buffer, 1); /* WTF??? */ | |
1662 | return -1; | |
1663 | } | |
1664 | ||
1665 | /* | |
1666 | * Now that we are here, the old head pointer is | |
1667 | * set to UPDATE. This will keep the reader from | |
1668 | * swapping the head page with the reader page. | |
1669 | * The reader (on another CPU) will spin till | |
1670 | * we are finished. | |
1671 | * | |
1672 | * We just need to protect against interrupts | |
1673 | * doing the job. We will set the next pointer | |
1674 | * to HEAD. After that, we set the old pointer | |
1675 | * to NORMAL, but only if it was HEAD before. | |
1676 | * otherwise we are an interrupt, and only | |
1677 | * want the outer most commit to reset it. | |
1678 | */ | |
1679 | new_head = next_page; | |
1680 | rb_inc_page(cpu_buffer, &new_head); | |
1681 | ||
1682 | ret = rb_head_page_set_head(cpu_buffer, new_head, next_page, | |
1683 | RB_PAGE_NORMAL); | |
1684 | ||
1685 | /* | |
1686 | * Valid returns are: | |
1687 | * HEAD - an interrupt came in and already set it. | |
1688 | * NORMAL - One of two things: | |
1689 | * 1) We really set it. | |
1690 | * 2) A bunch of interrupts came in and moved | |
1691 | * the page forward again. | |
1692 | */ | |
1693 | switch (ret) { | |
1694 | case RB_PAGE_HEAD: | |
1695 | case RB_PAGE_NORMAL: | |
1696 | /* OK */ | |
1697 | break; | |
1698 | default: | |
1699 | RB_WARN_ON(cpu_buffer, 1); | |
1700 | return -1; | |
1701 | } | |
1702 | ||
1703 | /* | |
1704 | * It is possible that an interrupt came in, | |
1705 | * set the head up, then more interrupts came in | |
1706 | * and moved it again. When we get back here, | |
1707 | * the page would have been set to NORMAL but we | |
1708 | * just set it back to HEAD. | |
1709 | * | |
1710 | * How do you detect this? Well, if that happened | |
1711 | * the tail page would have moved. | |
1712 | */ | |
1713 | if (ret == RB_PAGE_NORMAL) { | |
1714 | /* | |
1715 | * If the tail had moved passed next, then we need | |
1716 | * to reset the pointer. | |
1717 | */ | |
1718 | if (cpu_buffer->tail_page != tail_page && | |
1719 | cpu_buffer->tail_page != next_page) | |
1720 | rb_head_page_set_normal(cpu_buffer, new_head, | |
1721 | next_page, | |
1722 | RB_PAGE_HEAD); | |
1723 | } | |
1724 | ||
1725 | /* | |
1726 | * If this was the outer most commit (the one that | |
1727 | * changed the original pointer from HEAD to UPDATE), | |
1728 | * then it is up to us to reset it to NORMAL. | |
1729 | */ | |
1730 | if (type == RB_PAGE_HEAD) { | |
1731 | ret = rb_head_page_set_normal(cpu_buffer, next_page, | |
1732 | tail_page, | |
1733 | RB_PAGE_UPDATE); | |
1734 | if (RB_WARN_ON(cpu_buffer, | |
1735 | ret != RB_PAGE_UPDATE)) | |
1736 | return -1; | |
1737 | } | |
1738 | ||
1739 | return 0; | |
1740 | } | |
1741 | ||
1742 | static unsigned rb_calculate_event_length(unsigned length) | |
1743 | { | |
1744 | struct ring_buffer_event event; /* Used only for sizeof array */ | |
1745 | ||
1746 | /* zero length can cause confusions */ | |
1747 | if (!length) | |
1748 | length = 1; | |
1749 | ||
1750 | if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) | |
1751 | length += sizeof(event.array[0]); | |
1752 | ||
1753 | length += RB_EVNT_HDR_SIZE; | |
1754 | length = ALIGN(length, RB_ARCH_ALIGNMENT); | |
1755 | ||
1756 | return length; | |
1757 | } | |
1758 | ||
1759 | static inline void | |
1760 | rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer, | |
1761 | struct buffer_page *tail_page, | |
1762 | unsigned long tail, unsigned long length) | |
1763 | { | |
1764 | struct ring_buffer_event *event; | |
1765 | ||
1766 | /* | |
1767 | * Only the event that crossed the page boundary | |
1768 | * must fill the old tail_page with padding. | |
1769 | */ | |
1770 | if (tail >= BUF_PAGE_SIZE) { | |
1771 | /* | |
1772 | * If the page was filled, then we still need | |
1773 | * to update the real_end. Reset it to zero | |
1774 | * and the reader will ignore it. | |
1775 | */ | |
1776 | if (tail == BUF_PAGE_SIZE) | |
1777 | tail_page->real_end = 0; | |
1778 | ||
1779 | local_sub(length, &tail_page->write); | |
1780 | return; | |
1781 | } | |
1782 | ||
1783 | event = __rb_page_index(tail_page, tail); | |
1784 | kmemcheck_annotate_bitfield(event, bitfield); | |
1785 | ||
1786 | /* | |
1787 | * Save the original length to the meta data. | |
1788 | * This will be used by the reader to add lost event | |
1789 | * counter. | |
1790 | */ | |
1791 | tail_page->real_end = tail; | |
1792 | ||
1793 | /* | |
1794 | * If this event is bigger than the minimum size, then | |
1795 | * we need to be careful that we don't subtract the | |
1796 | * write counter enough to allow another writer to slip | |
1797 | * in on this page. | |
1798 | * We put in a discarded commit instead, to make sure | |
1799 | * that this space is not used again. | |
1800 | * | |
1801 | * If we are less than the minimum size, we don't need to | |
1802 | * worry about it. | |
1803 | */ | |
1804 | if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) { | |
1805 | /* No room for any events */ | |
1806 | ||
1807 | /* Mark the rest of the page with padding */ | |
1808 | rb_event_set_padding(event); | |
1809 | ||
1810 | /* Set the write back to the previous setting */ | |
1811 | local_sub(length, &tail_page->write); | |
1812 | return; | |
1813 | } | |
1814 | ||
1815 | /* Put in a discarded event */ | |
1816 | event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE; | |
1817 | event->type_len = RINGBUF_TYPE_PADDING; | |
1818 | /* time delta must be non zero */ | |
1819 | event->time_delta = 1; | |
1820 | ||
1821 | /* Set write to end of buffer */ | |
1822 | length = (tail + length) - BUF_PAGE_SIZE; | |
1823 | local_sub(length, &tail_page->write); | |
1824 | } | |
1825 | ||
1826 | static struct ring_buffer_event * | |
1827 | rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer, | |
1828 | unsigned long length, unsigned long tail, | |
1829 | struct buffer_page *tail_page, u64 *ts) | |
1830 | { | |
1831 | struct buffer_page *commit_page = cpu_buffer->commit_page; | |
1832 | struct ring_buffer *buffer = cpu_buffer->buffer; | |
1833 | struct buffer_page *next_page; | |
1834 | int ret; | |
1835 | ||
1836 | next_page = tail_page; | |
1837 | ||
1838 | rb_inc_page(cpu_buffer, &next_page); | |
1839 | ||
1840 | /* | |
1841 | * If for some reason, we had an interrupt storm that made | |
1842 | * it all the way around the buffer, bail, and warn | |
1843 | * about it. | |
1844 | */ | |
1845 | if (unlikely(next_page == commit_page)) { | |
1846 | local_inc(&cpu_buffer->commit_overrun); | |
1847 | goto out_reset; | |
1848 | } | |
1849 | ||
1850 | /* | |
1851 | * This is where the fun begins! | |
1852 | * | |
1853 | * We are fighting against races between a reader that | |
1854 | * could be on another CPU trying to swap its reader | |
1855 | * page with the buffer head. | |
1856 | * | |
1857 | * We are also fighting against interrupts coming in and | |
1858 | * moving the head or tail on us as well. | |
1859 | * | |
1860 | * If the next page is the head page then we have filled | |
1861 | * the buffer, unless the commit page is still on the | |
1862 | * reader page. | |
1863 | */ | |
1864 | if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) { | |
1865 | ||
1866 | /* | |
1867 | * If the commit is not on the reader page, then | |
1868 | * move the header page. | |
1869 | */ | |
1870 | if (!rb_is_reader_page(cpu_buffer->commit_page)) { | |
1871 | /* | |
1872 | * If we are not in overwrite mode, | |
1873 | * this is easy, just stop here. | |
1874 | */ | |
1875 | if (!(buffer->flags & RB_FL_OVERWRITE)) | |
1876 | goto out_reset; | |
1877 | ||
1878 | ret = rb_handle_head_page(cpu_buffer, | |
1879 | tail_page, | |
1880 | next_page); | |
1881 | if (ret < 0) | |
1882 | goto out_reset; | |
1883 | if (ret) | |
1884 | goto out_again; | |
1885 | } else { | |
1886 | /* | |
1887 | * We need to be careful here too. The | |
1888 | * commit page could still be on the reader | |
1889 | * page. We could have a small buffer, and | |
1890 | * have filled up the buffer with events | |
1891 | * from interrupts and such, and wrapped. | |
1892 | * | |
1893 | * Note, if the tail page is also the on the | |
1894 | * reader_page, we let it move out. | |
1895 | */ | |
1896 | if (unlikely((cpu_buffer->commit_page != | |
1897 | cpu_buffer->tail_page) && | |
1898 | (cpu_buffer->commit_page == | |
1899 | cpu_buffer->reader_page))) { | |
1900 | local_inc(&cpu_buffer->commit_overrun); | |
1901 | goto out_reset; | |
1902 | } | |
1903 | } | |
1904 | } | |
1905 | ||
1906 | ret = rb_tail_page_update(cpu_buffer, tail_page, next_page); | |
1907 | if (ret) { | |
1908 | /* | |
1909 | * Nested commits always have zero deltas, so | |
1910 | * just reread the time stamp | |
1911 | */ | |
1912 | *ts = rb_time_stamp(buffer); | |
1913 | next_page->page->time_stamp = *ts; | |
1914 | } | |
1915 | ||
1916 | out_again: | |
1917 | ||
1918 | rb_reset_tail(cpu_buffer, tail_page, tail, length); | |
1919 | ||
1920 | /* fail and let the caller try again */ | |
1921 | return ERR_PTR(-EAGAIN); | |
1922 | ||
1923 | out_reset: | |
1924 | /* reset write */ | |
1925 | rb_reset_tail(cpu_buffer, tail_page, tail, length); | |
1926 | ||
1927 | return NULL; | |
1928 | } | |
1929 | ||
1930 | static struct ring_buffer_event * | |
1931 | __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer, | |
1932 | unsigned type, unsigned long length, u64 *ts) | |
1933 | { | |
1934 | struct buffer_page *tail_page; | |
1935 | struct ring_buffer_event *event; | |
1936 | unsigned long tail, write; | |
1937 | ||
1938 | tail_page = cpu_buffer->tail_page; | |
1939 | write = local_add_return(length, &tail_page->write); | |
1940 | ||
1941 | /* set write to only the index of the write */ | |
1942 | write &= RB_WRITE_MASK; | |
1943 | tail = write - length; | |
1944 | ||
1945 | /* See if we shot pass the end of this buffer page */ | |
1946 | if (write > BUF_PAGE_SIZE) | |
1947 | return rb_move_tail(cpu_buffer, length, tail, | |
1948 | tail_page, ts); | |
1949 | ||
1950 | /* We reserved something on the buffer */ | |
1951 | ||
1952 | event = __rb_page_index(tail_page, tail); | |
1953 | kmemcheck_annotate_bitfield(event, bitfield); | |
1954 | rb_update_event(event, type, length); | |
1955 | ||
1956 | /* The passed in type is zero for DATA */ | |
1957 | if (likely(!type)) | |
1958 | local_inc(&tail_page->entries); | |
1959 | ||
1960 | /* | |
1961 | * If this is the first commit on the page, then update | |
1962 | * its timestamp. | |
1963 | */ | |
1964 | if (!tail) | |
1965 | tail_page->page->time_stamp = *ts; | |
1966 | ||
1967 | return event; | |
1968 | } | |
1969 | ||
1970 | static inline int | |
1971 | rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer, | |
1972 | struct ring_buffer_event *event) | |
1973 | { | |
1974 | unsigned long new_index, old_index; | |
1975 | struct buffer_page *bpage; | |
1976 | unsigned long index; | |
1977 | unsigned long addr; | |
1978 | ||
1979 | new_index = rb_event_index(event); | |
1980 | old_index = new_index + rb_event_length(event); | |
1981 | addr = (unsigned long)event; | |
1982 | addr &= PAGE_MASK; | |
1983 | ||
1984 | bpage = cpu_buffer->tail_page; | |
1985 | ||
1986 | if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) { | |
1987 | unsigned long write_mask = | |
1988 | local_read(&bpage->write) & ~RB_WRITE_MASK; | |
1989 | /* | |
1990 | * This is on the tail page. It is possible that | |
1991 | * a write could come in and move the tail page | |
1992 | * and write to the next page. That is fine | |
1993 | * because we just shorten what is on this page. | |
1994 | */ | |
1995 | old_index += write_mask; | |
1996 | new_index += write_mask; | |
1997 | index = local_cmpxchg(&bpage->write, old_index, new_index); | |
1998 | if (index == old_index) | |
1999 | return 1; | |
2000 | } | |
2001 | ||
2002 | /* could not discard */ | |
2003 | return 0; | |
2004 | } | |
2005 | ||
2006 | static int | |
2007 | rb_add_time_stamp(struct ring_buffer_per_cpu *cpu_buffer, | |
2008 | u64 *ts, u64 *delta) | |
2009 | { | |
2010 | struct ring_buffer_event *event; | |
2011 | int ret; | |
2012 | ||
2013 | WARN_ONCE(*delta > (1ULL << 59), | |
2014 | KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n", | |
2015 | (unsigned long long)*delta, | |
2016 | (unsigned long long)*ts, | |
2017 | (unsigned long long)cpu_buffer->write_stamp); | |
2018 | ||
2019 | /* | |
2020 | * The delta is too big, we to add a | |
2021 | * new timestamp. | |
2022 | */ | |
2023 | event = __rb_reserve_next(cpu_buffer, | |
2024 | RINGBUF_TYPE_TIME_EXTEND, | |
2025 | RB_LEN_TIME_EXTEND, | |
2026 | ts); | |
2027 | if (!event) | |
2028 | return -EBUSY; | |
2029 | ||
2030 | if (PTR_ERR(event) == -EAGAIN) | |
2031 | return -EAGAIN; | |
2032 | ||
2033 | /* Only a commited time event can update the write stamp */ | |
2034 | if (rb_event_is_commit(cpu_buffer, event)) { | |
2035 | /* | |
2036 | * If this is the first on the page, then it was | |
2037 | * updated with the page itself. Try to discard it | |
2038 | * and if we can't just make it zero. | |
2039 | */ | |
2040 | if (rb_event_index(event)) { | |
2041 | event->time_delta = *delta & TS_MASK; | |
2042 | event->array[0] = *delta >> TS_SHIFT; | |
2043 | } else { | |
2044 | /* try to discard, since we do not need this */ | |
2045 | if (!rb_try_to_discard(cpu_buffer, event)) { | |
2046 | /* nope, just zero it */ | |
2047 | event->time_delta = 0; | |
2048 | event->array[0] = 0; | |
2049 | } | |
2050 | } | |
2051 | cpu_buffer->write_stamp = *ts; | |
2052 | /* let the caller know this was the commit */ | |
2053 | ret = 1; | |
2054 | } else { | |
2055 | /* Try to discard the event */ | |
2056 | if (!rb_try_to_discard(cpu_buffer, event)) { | |
2057 | /* Darn, this is just wasted space */ | |
2058 | event->time_delta = 0; | |
2059 | event->array[0] = 0; | |
2060 | } | |
2061 | ret = 0; | |
2062 | } | |
2063 | ||
2064 | *delta = 0; | |
2065 | ||
2066 | return ret; | |
2067 | } | |
2068 | ||
2069 | static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer) | |
2070 | { | |
2071 | local_inc(&cpu_buffer->committing); | |
2072 | local_inc(&cpu_buffer->commits); | |
2073 | } | |
2074 | ||
2075 | static void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer) | |
2076 | { | |
2077 | unsigned long commits; | |
2078 | ||
2079 | if (RB_WARN_ON(cpu_buffer, | |
2080 | !local_read(&cpu_buffer->committing))) | |
2081 | return; | |
2082 | ||
2083 | again: | |
2084 | commits = local_read(&cpu_buffer->commits); | |
2085 | /* synchronize with interrupts */ | |
2086 | barrier(); | |
2087 | if (local_read(&cpu_buffer->committing) == 1) | |
2088 | rb_set_commit_to_write(cpu_buffer); | |
2089 | ||
2090 | local_dec(&cpu_buffer->committing); | |
2091 | ||
2092 | /* synchronize with interrupts */ | |
2093 | barrier(); | |
2094 | ||
2095 | /* | |
2096 | * Need to account for interrupts coming in between the | |
2097 | * updating of the commit page and the clearing of the | |
2098 | * committing counter. | |
2099 | */ | |
2100 | if (unlikely(local_read(&cpu_buffer->commits) != commits) && | |
2101 | !local_read(&cpu_buffer->committing)) { | |
2102 | local_inc(&cpu_buffer->committing); | |
2103 | goto again; | |
2104 | } | |
2105 | } | |
2106 | ||
2107 | static struct ring_buffer_event * | |
2108 | rb_reserve_next_event(struct ring_buffer *buffer, | |
2109 | struct ring_buffer_per_cpu *cpu_buffer, | |
2110 | unsigned long length) | |
2111 | { | |
2112 | struct ring_buffer_event *event; | |
2113 | u64 ts, delta = 0; | |
2114 | int commit = 0; | |
2115 | int nr_loops = 0; | |
2116 | ||
2117 | rb_start_commit(cpu_buffer); | |
2118 | ||
2119 | #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP | |
2120 | /* | |
2121 | * Due to the ability to swap a cpu buffer from a buffer | |
2122 | * it is possible it was swapped before we committed. | |
2123 | * (committing stops a swap). We check for it here and | |
2124 | * if it happened, we have to fail the write. | |
2125 | */ | |
2126 | barrier(); | |
2127 | if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) { | |
2128 | local_dec(&cpu_buffer->committing); | |
2129 | local_dec(&cpu_buffer->commits); | |
2130 | return NULL; | |
2131 | } | |
2132 | #endif | |
2133 | ||
2134 | length = rb_calculate_event_length(length); | |
2135 | again: | |
2136 | /* | |
2137 | * We allow for interrupts to reenter here and do a trace. | |
2138 | * If one does, it will cause this original code to loop | |
2139 | * back here. Even with heavy interrupts happening, this | |
2140 | * should only happen a few times in a row. If this happens | |
2141 | * 1000 times in a row, there must be either an interrupt | |
2142 | * storm or we have something buggy. | |
2143 | * Bail! | |
2144 | */ | |
2145 | if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000)) | |
2146 | goto out_fail; | |
2147 | ||
2148 | ts = rb_time_stamp(cpu_buffer->buffer); | |
2149 | ||
2150 | /* | |
2151 | * Only the first commit can update the timestamp. | |
2152 | * Yes there is a race here. If an interrupt comes in | |
2153 | * just after the conditional and it traces too, then it | |
2154 | * will also check the deltas. More than one timestamp may | |
2155 | * also be made. But only the entry that did the actual | |
2156 | * commit will be something other than zero. | |
2157 | */ | |
2158 | if (likely(cpu_buffer->tail_page == cpu_buffer->commit_page && | |
2159 | rb_page_write(cpu_buffer->tail_page) == | |
2160 | rb_commit_index(cpu_buffer))) { | |
2161 | u64 diff; | |
2162 | ||
2163 | diff = ts - cpu_buffer->write_stamp; | |
2164 | ||
2165 | /* make sure this diff is calculated here */ | |
2166 | barrier(); | |
2167 | ||
2168 | /* Did the write stamp get updated already? */ | |
2169 | if (unlikely(ts < cpu_buffer->write_stamp)) | |
2170 | goto get_event; | |
2171 | ||
2172 | delta = diff; | |
2173 | if (unlikely(test_time_stamp(delta))) { | |
2174 | ||
2175 | commit = rb_add_time_stamp(cpu_buffer, &ts, &delta); | |
2176 | if (commit == -EBUSY) | |
2177 | goto out_fail; | |
2178 | ||
2179 | if (commit == -EAGAIN) | |
2180 | goto again; | |
2181 | ||
2182 | RB_WARN_ON(cpu_buffer, commit < 0); | |
2183 | } | |
2184 | } | |
2185 | ||
2186 | get_event: | |
2187 | event = __rb_reserve_next(cpu_buffer, 0, length, &ts); | |
2188 | if (unlikely(PTR_ERR(event) == -EAGAIN)) | |
2189 | goto again; | |
2190 | ||
2191 | if (!event) | |
2192 | goto out_fail; | |
2193 | ||
2194 | if (!rb_event_is_commit(cpu_buffer, event)) | |
2195 | delta = 0; | |
2196 | ||
2197 | event->time_delta = delta; | |
2198 | ||
2199 | return event; | |
2200 | ||
2201 | out_fail: | |
2202 | rb_end_commit(cpu_buffer); | |
2203 | return NULL; | |
2204 | } | |
2205 | ||
2206 | #ifdef CONFIG_TRACING | |
2207 | ||
2208 | #define TRACE_RECURSIVE_DEPTH 16 | |
2209 | ||
2210 | static int trace_recursive_lock(void) | |
2211 | { | |
2212 | current->trace_recursion++; | |
2213 | ||
2214 | if (likely(current->trace_recursion < TRACE_RECURSIVE_DEPTH)) | |
2215 | return 0; | |
2216 | ||
2217 | /* Disable all tracing before we do anything else */ | |
2218 | tracing_off_permanent(); | |
2219 | ||
2220 | printk_once(KERN_WARNING "Tracing recursion: depth[%ld]:" | |
2221 | "HC[%lu]:SC[%lu]:NMI[%lu]\n", | |
2222 | current->trace_recursion, | |
2223 | hardirq_count() >> HARDIRQ_SHIFT, | |
2224 | softirq_count() >> SOFTIRQ_SHIFT, | |
2225 | in_nmi()); | |
2226 | ||
2227 | WARN_ON_ONCE(1); | |
2228 | return -1; | |
2229 | } | |
2230 | ||
2231 | static void trace_recursive_unlock(void) | |
2232 | { | |
2233 | WARN_ON_ONCE(!current->trace_recursion); | |
2234 | ||
2235 | current->trace_recursion--; | |
2236 | } | |
2237 | ||
2238 | #else | |
2239 | ||
2240 | #define trace_recursive_lock() (0) | |
2241 | #define trace_recursive_unlock() do { } while (0) | |
2242 | ||
2243 | #endif | |
2244 | ||
2245 | /** | |
2246 | * ring_buffer_lock_reserve - reserve a part of the buffer | |
2247 | * @buffer: the ring buffer to reserve from | |
2248 | * @length: the length of the data to reserve (excluding event header) | |
2249 | * | |
2250 | * Returns a reseverd event on the ring buffer to copy directly to. | |
2251 | * The user of this interface will need to get the body to write into | |
2252 | * and can use the ring_buffer_event_data() interface. | |
2253 | * | |
2254 | * The length is the length of the data needed, not the event length | |
2255 | * which also includes the event header. | |
2256 | * | |
2257 | * Must be paired with ring_buffer_unlock_commit, unless NULL is returned. | |
2258 | * If NULL is returned, then nothing has been allocated or locked. | |
2259 | */ | |
2260 | struct ring_buffer_event * | |
2261 | ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length) | |
2262 | { | |
2263 | struct ring_buffer_per_cpu *cpu_buffer; | |
2264 | struct ring_buffer_event *event; | |
2265 | int cpu; | |
2266 | ||
2267 | if (ring_buffer_flags != RB_BUFFERS_ON) | |
2268 | return NULL; | |
2269 | ||
2270 | /* If we are tracing schedule, we don't want to recurse */ | |
2271 | preempt_disable_notrace(); | |
2272 | ||
2273 | if (atomic_read(&buffer->record_disabled)) | |
2274 | goto out_nocheck; | |
2275 | ||
2276 | if (trace_recursive_lock()) | |
2277 | goto out_nocheck; | |
2278 | ||
2279 | cpu = raw_smp_processor_id(); | |
2280 | ||
2281 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) | |
2282 | goto out; | |
2283 | ||
2284 | cpu_buffer = buffer->buffers[cpu]; | |
2285 | ||
2286 | if (atomic_read(&cpu_buffer->record_disabled)) | |
2287 | goto out; | |
2288 | ||
2289 | if (length > BUF_MAX_DATA_SIZE) | |
2290 | goto out; | |
2291 | ||
2292 | event = rb_reserve_next_event(buffer, cpu_buffer, length); | |
2293 | if (!event) | |
2294 | goto out; | |
2295 | ||
2296 | return event; | |
2297 | ||
2298 | out: | |
2299 | trace_recursive_unlock(); | |
2300 | ||
2301 | out_nocheck: | |
2302 | preempt_enable_notrace(); | |
2303 | return NULL; | |
2304 | } | |
2305 | EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve); | |
2306 | ||
2307 | static void | |
2308 | rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer, | |
2309 | struct ring_buffer_event *event) | |
2310 | { | |
2311 | /* | |
2312 | * The event first in the commit queue updates the | |
2313 | * time stamp. | |
2314 | */ | |
2315 | if (rb_event_is_commit(cpu_buffer, event)) | |
2316 | cpu_buffer->write_stamp += event->time_delta; | |
2317 | } | |
2318 | ||
2319 | static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer, | |
2320 | struct ring_buffer_event *event) | |
2321 | { | |
2322 | local_inc(&cpu_buffer->entries); | |
2323 | rb_update_write_stamp(cpu_buffer, event); | |
2324 | rb_end_commit(cpu_buffer); | |
2325 | } | |
2326 | ||
2327 | /** | |
2328 | * ring_buffer_unlock_commit - commit a reserved | |
2329 | * @buffer: The buffer to commit to | |
2330 | * @event: The event pointer to commit. | |
2331 | * | |
2332 | * This commits the data to the ring buffer, and releases any locks held. | |
2333 | * | |
2334 | * Must be paired with ring_buffer_lock_reserve. | |
2335 | */ | |
2336 | int ring_buffer_unlock_commit(struct ring_buffer *buffer, | |
2337 | struct ring_buffer_event *event) | |
2338 | { | |
2339 | struct ring_buffer_per_cpu *cpu_buffer; | |
2340 | int cpu = raw_smp_processor_id(); | |
2341 | ||
2342 | cpu_buffer = buffer->buffers[cpu]; | |
2343 | ||
2344 | rb_commit(cpu_buffer, event); | |
2345 | ||
2346 | trace_recursive_unlock(); | |
2347 | ||
2348 | preempt_enable_notrace(); | |
2349 | ||
2350 | return 0; | |
2351 | } | |
2352 | EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit); | |
2353 | ||
2354 | static inline void rb_event_discard(struct ring_buffer_event *event) | |
2355 | { | |
2356 | /* array[0] holds the actual length for the discarded event */ | |
2357 | event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE; | |
2358 | event->type_len = RINGBUF_TYPE_PADDING; | |
2359 | /* time delta must be non zero */ | |
2360 | if (!event->time_delta) | |
2361 | event->time_delta = 1; | |
2362 | } | |
2363 | ||
2364 | /* | |
2365 | * Decrement the entries to the page that an event is on. | |
2366 | * The event does not even need to exist, only the pointer | |
2367 | * to the page it is on. This may only be called before the commit | |
2368 | * takes place. | |
2369 | */ | |
2370 | static inline void | |
2371 | rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer, | |
2372 | struct ring_buffer_event *event) | |
2373 | { | |
2374 | unsigned long addr = (unsigned long)event; | |
2375 | struct buffer_page *bpage = cpu_buffer->commit_page; | |
2376 | struct buffer_page *start; | |
2377 | ||
2378 | addr &= PAGE_MASK; | |
2379 | ||
2380 | /* Do the likely case first */ | |
2381 | if (likely(bpage->page == (void *)addr)) { | |
2382 | local_dec(&bpage->entries); | |
2383 | return; | |
2384 | } | |
2385 | ||
2386 | /* | |
2387 | * Because the commit page may be on the reader page we | |
2388 | * start with the next page and check the end loop there. | |
2389 | */ | |
2390 | rb_inc_page(cpu_buffer, &bpage); | |
2391 | start = bpage; | |
2392 | do { | |
2393 | if (bpage->page == (void *)addr) { | |
2394 | local_dec(&bpage->entries); | |
2395 | return; | |
2396 | } | |
2397 | rb_inc_page(cpu_buffer, &bpage); | |
2398 | } while (bpage != start); | |
2399 | ||
2400 | /* commit not part of this buffer?? */ | |
2401 | RB_WARN_ON(cpu_buffer, 1); | |
2402 | } | |
2403 | ||
2404 | /** | |
2405 | * ring_buffer_commit_discard - discard an event that has not been committed | |
2406 | * @buffer: the ring buffer | |
2407 | * @event: non committed event to discard | |
2408 | * | |
2409 | * Sometimes an event that is in the ring buffer needs to be ignored. | |
2410 | * This function lets the user discard an event in the ring buffer | |
2411 | * and then that event will not be read later. | |
2412 | * | |
2413 | * This function only works if it is called before the the item has been | |
2414 | * committed. It will try to free the event from the ring buffer | |
2415 | * if another event has not been added behind it. | |
2416 | * | |
2417 | * If another event has been added behind it, it will set the event | |
2418 | * up as discarded, and perform the commit. | |
2419 | * | |
2420 | * If this function is called, do not call ring_buffer_unlock_commit on | |
2421 | * the event. | |
2422 | */ | |
2423 | void ring_buffer_discard_commit(struct ring_buffer *buffer, | |
2424 | struct ring_buffer_event *event) | |
2425 | { | |
2426 | struct ring_buffer_per_cpu *cpu_buffer; | |
2427 | int cpu; | |
2428 | ||
2429 | /* The event is discarded regardless */ | |
2430 | rb_event_discard(event); | |
2431 | ||
2432 | cpu = smp_processor_id(); | |
2433 | cpu_buffer = buffer->buffers[cpu]; | |
2434 | ||
2435 | /* | |
2436 | * This must only be called if the event has not been | |
2437 | * committed yet. Thus we can assume that preemption | |
2438 | * is still disabled. | |
2439 | */ | |
2440 | RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing)); | |
2441 | ||
2442 | rb_decrement_entry(cpu_buffer, event); | |
2443 | if (rb_try_to_discard(cpu_buffer, event)) | |
2444 | goto out; | |
2445 | ||
2446 | /* | |
2447 | * The commit is still visible by the reader, so we | |
2448 | * must still update the timestamp. | |
2449 | */ | |
2450 | rb_update_write_stamp(cpu_buffer, event); | |
2451 | out: | |
2452 | rb_end_commit(cpu_buffer); | |
2453 | ||
2454 | trace_recursive_unlock(); | |
2455 | ||
2456 | preempt_enable_notrace(); | |
2457 | ||
2458 | } | |
2459 | EXPORT_SYMBOL_GPL(ring_buffer_discard_commit); | |
2460 | ||
2461 | /** | |
2462 | * ring_buffer_write - write data to the buffer without reserving | |
2463 | * @buffer: The ring buffer to write to. | |
2464 | * @length: The length of the data being written (excluding the event header) | |
2465 | * @data: The data to write to the buffer. | |
2466 | * | |
2467 | * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as | |
2468 | * one function. If you already have the data to write to the buffer, it | |
2469 | * may be easier to simply call this function. | |
2470 | * | |
2471 | * Note, like ring_buffer_lock_reserve, the length is the length of the data | |
2472 | * and not the length of the event which would hold the header. | |
2473 | */ | |
2474 | int ring_buffer_write(struct ring_buffer *buffer, | |
2475 | unsigned long length, | |
2476 | void *data) | |
2477 | { | |
2478 | struct ring_buffer_per_cpu *cpu_buffer; | |
2479 | struct ring_buffer_event *event; | |
2480 | void *body; | |
2481 | int ret = -EBUSY; | |
2482 | int cpu; | |
2483 | ||
2484 | if (ring_buffer_flags != RB_BUFFERS_ON) | |
2485 | return -EBUSY; | |
2486 | ||
2487 | preempt_disable_notrace(); | |
2488 | ||
2489 | if (atomic_read(&buffer->record_disabled)) | |
2490 | goto out; | |
2491 | ||
2492 | cpu = raw_smp_processor_id(); | |
2493 | ||
2494 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) | |
2495 | goto out; | |
2496 | ||
2497 | cpu_buffer = buffer->buffers[cpu]; | |
2498 | ||
2499 | if (atomic_read(&cpu_buffer->record_disabled)) | |
2500 | goto out; | |
2501 | ||
2502 | if (length > BUF_MAX_DATA_SIZE) | |
2503 | goto out; | |
2504 | ||
2505 | event = rb_reserve_next_event(buffer, cpu_buffer, length); | |
2506 | if (!event) | |
2507 | goto out; | |
2508 | ||
2509 | body = rb_event_data(event); | |
2510 | ||
2511 | memcpy(body, data, length); | |
2512 | ||
2513 | rb_commit(cpu_buffer, event); | |
2514 | ||
2515 | ret = 0; | |
2516 | out: | |
2517 | preempt_enable_notrace(); | |
2518 | ||
2519 | return ret; | |
2520 | } | |
2521 | EXPORT_SYMBOL_GPL(ring_buffer_write); | |
2522 | ||
2523 | static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer) | |
2524 | { | |
2525 | struct buffer_page *reader = cpu_buffer->reader_page; | |
2526 | struct buffer_page *head = rb_set_head_page(cpu_buffer); | |
2527 | struct buffer_page *commit = cpu_buffer->commit_page; | |
2528 | ||
2529 | /* In case of error, head will be NULL */ | |
2530 | if (unlikely(!head)) | |
2531 | return 1; | |
2532 | ||
2533 | return reader->read == rb_page_commit(reader) && | |
2534 | (commit == reader || | |
2535 | (commit == head && | |
2536 | head->read == rb_page_commit(commit))); | |
2537 | } | |
2538 | ||
2539 | /** | |
2540 | * ring_buffer_record_disable - stop all writes into the buffer | |
2541 | * @buffer: The ring buffer to stop writes to. | |
2542 | * | |
2543 | * This prevents all writes to the buffer. Any attempt to write | |
2544 | * to the buffer after this will fail and return NULL. | |
2545 | * | |
2546 | * The caller should call synchronize_sched() after this. | |
2547 | */ | |
2548 | void ring_buffer_record_disable(struct ring_buffer *buffer) | |
2549 | { | |
2550 | atomic_inc(&buffer->record_disabled); | |
2551 | } | |
2552 | EXPORT_SYMBOL_GPL(ring_buffer_record_disable); | |
2553 | ||
2554 | /** | |
2555 | * ring_buffer_record_enable - enable writes to the buffer | |
2556 | * @buffer: The ring buffer to enable writes | |
2557 | * | |
2558 | * Note, multiple disables will need the same number of enables | |
2559 | * to truly enable the writing (much like preempt_disable). | |
2560 | */ | |
2561 | void ring_buffer_record_enable(struct ring_buffer *buffer) | |
2562 | { | |
2563 | atomic_dec(&buffer->record_disabled); | |
2564 | } | |
2565 | EXPORT_SYMBOL_GPL(ring_buffer_record_enable); | |
2566 | ||
2567 | /** | |
2568 | * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer | |
2569 | * @buffer: The ring buffer to stop writes to. | |
2570 | * @cpu: The CPU buffer to stop | |
2571 | * | |
2572 | * This prevents all writes to the buffer. Any attempt to write | |
2573 | * to the buffer after this will fail and return NULL. | |
2574 | * | |
2575 | * The caller should call synchronize_sched() after this. | |
2576 | */ | |
2577 | void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu) | |
2578 | { | |
2579 | struct ring_buffer_per_cpu *cpu_buffer; | |
2580 | ||
2581 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) | |
2582 | return; | |
2583 | ||
2584 | cpu_buffer = buffer->buffers[cpu]; | |
2585 | atomic_inc(&cpu_buffer->record_disabled); | |
2586 | } | |
2587 | EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu); | |
2588 | ||
2589 | /** | |
2590 | * ring_buffer_record_enable_cpu - enable writes to the buffer | |
2591 | * @buffer: The ring buffer to enable writes | |
2592 | * @cpu: The CPU to enable. | |
2593 | * | |
2594 | * Note, multiple disables will need the same number of enables | |
2595 | * to truly enable the writing (much like preempt_disable). | |
2596 | */ | |
2597 | void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu) | |
2598 | { | |
2599 | struct ring_buffer_per_cpu *cpu_buffer; | |
2600 | ||
2601 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) | |
2602 | return; | |
2603 | ||
2604 | cpu_buffer = buffer->buffers[cpu]; | |
2605 | atomic_dec(&cpu_buffer->record_disabled); | |
2606 | } | |
2607 | EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu); | |
2608 | ||
2609 | /** | |
2610 | * ring_buffer_entries_cpu - get the number of entries in a cpu buffer | |
2611 | * @buffer: The ring buffer | |
2612 | * @cpu: The per CPU buffer to get the entries from. | |
2613 | */ | |
2614 | unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu) | |
2615 | { | |
2616 | struct ring_buffer_per_cpu *cpu_buffer; | |
2617 | unsigned long ret; | |
2618 | ||
2619 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) | |
2620 | return 0; | |
2621 | ||
2622 | cpu_buffer = buffer->buffers[cpu]; | |
2623 | ret = (local_read(&cpu_buffer->entries) - local_read(&cpu_buffer->overrun)) | |
2624 | - cpu_buffer->read; | |
2625 | ||
2626 | return ret; | |
2627 | } | |
2628 | EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu); | |
2629 | ||
2630 | /** | |
2631 | * ring_buffer_overrun_cpu - get the number of overruns in a cpu_buffer | |
2632 | * @buffer: The ring buffer | |
2633 | * @cpu: The per CPU buffer to get the number of overruns from | |
2634 | */ | |
2635 | unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu) | |
2636 | { | |
2637 | struct ring_buffer_per_cpu *cpu_buffer; | |
2638 | unsigned long ret; | |
2639 | ||
2640 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) | |
2641 | return 0; | |
2642 | ||
2643 | cpu_buffer = buffer->buffers[cpu]; | |
2644 | ret = local_read(&cpu_buffer->overrun); | |
2645 | ||
2646 | return ret; | |
2647 | } | |
2648 | EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu); | |
2649 | ||
2650 | /** | |
2651 | * ring_buffer_commit_overrun_cpu - get the number of overruns caused by commits | |
2652 | * @buffer: The ring buffer | |
2653 | * @cpu: The per CPU buffer to get the number of overruns from | |
2654 | */ | |
2655 | unsigned long | |
2656 | ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu) | |
2657 | { | |
2658 | struct ring_buffer_per_cpu *cpu_buffer; | |
2659 | unsigned long ret; | |
2660 | ||
2661 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) | |
2662 | return 0; | |
2663 | ||
2664 | cpu_buffer = buffer->buffers[cpu]; | |
2665 | ret = local_read(&cpu_buffer->commit_overrun); | |
2666 | ||
2667 | return ret; | |
2668 | } | |
2669 | EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu); | |
2670 | ||
2671 | /** | |
2672 | * ring_buffer_entries - get the number of entries in a buffer | |
2673 | * @buffer: The ring buffer | |
2674 | * | |
2675 | * Returns the total number of entries in the ring buffer | |
2676 | * (all CPU entries) | |
2677 | */ | |
2678 | unsigned long ring_buffer_entries(struct ring_buffer *buffer) | |
2679 | { | |
2680 | struct ring_buffer_per_cpu *cpu_buffer; | |
2681 | unsigned long entries = 0; | |
2682 | int cpu; | |
2683 | ||
2684 | /* if you care about this being correct, lock the buffer */ | |
2685 | for_each_buffer_cpu(buffer, cpu) { | |
2686 | cpu_buffer = buffer->buffers[cpu]; | |
2687 | entries += (local_read(&cpu_buffer->entries) - | |
2688 | local_read(&cpu_buffer->overrun)) - cpu_buffer->read; | |
2689 | } | |
2690 | ||
2691 | return entries; | |
2692 | } | |
2693 | EXPORT_SYMBOL_GPL(ring_buffer_entries); | |
2694 | ||
2695 | /** | |
2696 | * ring_buffer_overruns - get the number of overruns in buffer | |
2697 | * @buffer: The ring buffer | |
2698 | * | |
2699 | * Returns the total number of overruns in the ring buffer | |
2700 | * (all CPU entries) | |
2701 | */ | |
2702 | unsigned long ring_buffer_overruns(struct ring_buffer *buffer) | |
2703 | { | |
2704 | struct ring_buffer_per_cpu *cpu_buffer; | |
2705 | unsigned long overruns = 0; | |
2706 | int cpu; | |
2707 | ||
2708 | /* if you care about this being correct, lock the buffer */ | |
2709 | for_each_buffer_cpu(buffer, cpu) { | |
2710 | cpu_buffer = buffer->buffers[cpu]; | |
2711 | overruns += local_read(&cpu_buffer->overrun); | |
2712 | } | |
2713 | ||
2714 | return overruns; | |
2715 | } | |
2716 | EXPORT_SYMBOL_GPL(ring_buffer_overruns); | |
2717 | ||
2718 | static void rb_iter_reset(struct ring_buffer_iter *iter) | |
2719 | { | |
2720 | struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer; | |
2721 | ||
2722 | /* Iterator usage is expected to have record disabled */ | |
2723 | if (list_empty(&cpu_buffer->reader_page->list)) { | |
2724 | iter->head_page = rb_set_head_page(cpu_buffer); | |
2725 | if (unlikely(!iter->head_page)) | |
2726 | return; | |
2727 | iter->head = iter->head_page->read; | |
2728 | } else { | |
2729 | iter->head_page = cpu_buffer->reader_page; | |
2730 | iter->head = cpu_buffer->reader_page->read; | |
2731 | } | |
2732 | if (iter->head) | |
2733 | iter->read_stamp = cpu_buffer->read_stamp; | |
2734 | else | |
2735 | iter->read_stamp = iter->head_page->page->time_stamp; | |
2736 | iter->cache_reader_page = cpu_buffer->reader_page; | |
2737 | iter->cache_read = cpu_buffer->read; | |
2738 | } | |
2739 | ||
2740 | /** | |
2741 | * ring_buffer_iter_reset - reset an iterator | |
2742 | * @iter: The iterator to reset | |
2743 | * | |
2744 | * Resets the iterator, so that it will start from the beginning | |
2745 | * again. | |
2746 | */ | |
2747 | void ring_buffer_iter_reset(struct ring_buffer_iter *iter) | |
2748 | { | |
2749 | struct ring_buffer_per_cpu *cpu_buffer; | |
2750 | unsigned long flags; | |
2751 | ||
2752 | if (!iter) | |
2753 | return; | |
2754 | ||
2755 | cpu_buffer = iter->cpu_buffer; | |
2756 | ||
2757 | spin_lock_irqsave(&cpu_buffer->reader_lock, flags); | |
2758 | rb_iter_reset(iter); | |
2759 | spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags); | |
2760 | } | |
2761 | EXPORT_SYMBOL_GPL(ring_buffer_iter_reset); | |
2762 | ||
2763 | /** | |
2764 | * ring_buffer_iter_empty - check if an iterator has no more to read | |
2765 | * @iter: The iterator to check | |
2766 | */ | |
2767 | int ring_buffer_iter_empty(struct ring_buffer_iter *iter) | |
2768 | { | |
2769 | struct ring_buffer_per_cpu *cpu_buffer; | |
2770 | ||
2771 | cpu_buffer = iter->cpu_buffer; | |
2772 | ||
2773 | return iter->head_page == cpu_buffer->commit_page && | |
2774 | iter->head == rb_commit_index(cpu_buffer); | |
2775 | } | |
2776 | EXPORT_SYMBOL_GPL(ring_buffer_iter_empty); | |
2777 | ||
2778 | static void | |
2779 | rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer, | |
2780 | struct ring_buffer_event *event) | |
2781 | { | |
2782 | u64 delta; | |
2783 | ||
2784 | switch (event->type_len) { | |
2785 | case RINGBUF_TYPE_PADDING: | |
2786 | return; | |
2787 | ||
2788 | case RINGBUF_TYPE_TIME_EXTEND: | |
2789 | delta = event->array[0]; | |
2790 | delta <<= TS_SHIFT; | |
2791 | delta += event->time_delta; | |
2792 | cpu_buffer->read_stamp += delta; | |
2793 | return; | |
2794 | ||
2795 | case RINGBUF_TYPE_TIME_STAMP: | |
2796 | /* FIXME: not implemented */ | |
2797 | return; | |
2798 | ||
2799 | case RINGBUF_TYPE_DATA: | |
2800 | cpu_buffer->read_stamp += event->time_delta; | |
2801 | return; | |
2802 | ||
2803 | default: | |
2804 | BUG(); | |
2805 | } | |
2806 | return; | |
2807 | } | |
2808 | ||
2809 | static void | |
2810 | rb_update_iter_read_stamp(struct ring_buffer_iter *iter, | |
2811 | struct ring_buffer_event *event) | |
2812 | { | |
2813 | u64 delta; | |
2814 | ||
2815 | switch (event->type_len) { | |
2816 | case RINGBUF_TYPE_PADDING: | |
2817 | return; | |
2818 | ||
2819 | case RINGBUF_TYPE_TIME_EXTEND: | |
2820 | delta = event->array[0]; | |
2821 | delta <<= TS_SHIFT; | |
2822 | delta += event->time_delta; | |
2823 | iter->read_stamp += delta; | |
2824 | return; | |
2825 | ||
2826 | case RINGBUF_TYPE_TIME_STAMP: | |
2827 | /* FIXME: not implemented */ | |
2828 | return; | |
2829 | ||
2830 | case RINGBUF_TYPE_DATA: | |
2831 | iter->read_stamp += event->time_delta; | |
2832 | return; | |
2833 | ||
2834 | default: | |
2835 | BUG(); | |
2836 | } | |
2837 | return; | |
2838 | } | |
2839 | ||
2840 | static struct buffer_page * | |
2841 | rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer) | |
2842 | { | |
2843 | struct buffer_page *reader = NULL; | |
2844 | unsigned long overwrite; | |
2845 | unsigned long flags; | |
2846 | int nr_loops = 0; | |
2847 | int ret; | |
2848 | ||
2849 | local_irq_save(flags); | |
2850 | arch_spin_lock(&cpu_buffer->lock); | |
2851 | ||
2852 | again: | |
2853 | /* | |
2854 | * This should normally only loop twice. But because the | |
2855 | * start of the reader inserts an empty page, it causes | |
2856 | * a case where we will loop three times. There should be no | |
2857 | * reason to loop four times (that I know of). | |
2858 | */ | |
2859 | if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) { | |
2860 | reader = NULL; | |
2861 | goto out; | |
2862 | } | |
2863 | ||
2864 | reader = cpu_buffer->reader_page; | |
2865 | ||
2866 | /* If there's more to read, return this page */ | |
2867 | if (cpu_buffer->reader_page->read < rb_page_size(reader)) | |
2868 | goto out; | |
2869 | ||
2870 | /* Never should we have an index greater than the size */ | |
2871 | if (RB_WARN_ON(cpu_buffer, | |
2872 | cpu_buffer->reader_page->read > rb_page_size(reader))) | |
2873 | goto out; | |
2874 | ||
2875 | /* check if we caught up to the tail */ | |
2876 | reader = NULL; | |
2877 | if (cpu_buffer->commit_page == cpu_buffer->reader_page) | |
2878 | goto out; | |
2879 | ||
2880 | /* | |
2881 | * Reset the reader page to size zero. | |
2882 | */ | |
2883 | local_set(&cpu_buffer->reader_page->write, 0); | |
2884 | local_set(&cpu_buffer->reader_page->entries, 0); | |
2885 | local_set(&cpu_buffer->reader_page->page->commit, 0); | |
2886 | cpu_buffer->reader_page->real_end = 0; | |
2887 | ||
2888 | spin: | |
2889 | /* | |
2890 | * Splice the empty reader page into the list around the head. | |
2891 | */ | |
2892 | reader = rb_set_head_page(cpu_buffer); | |
2893 | cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next); | |
2894 | cpu_buffer->reader_page->list.prev = reader->list.prev; | |
2895 | ||
2896 | /* | |
2897 | * cpu_buffer->pages just needs to point to the buffer, it | |
2898 | * has no specific buffer page to point to. Lets move it out | |
2899 | * of our way so we don't accidently swap it. | |
2900 | */ | |
2901 | cpu_buffer->pages = reader->list.prev; | |
2902 | ||
2903 | /* The reader page will be pointing to the new head */ | |
2904 | rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list); | |
2905 | ||
2906 | /* | |
2907 | * We want to make sure we read the overruns after we set up our | |
2908 | * pointers to the next object. The writer side does a | |
2909 | * cmpxchg to cross pages which acts as the mb on the writer | |
2910 | * side. Note, the reader will constantly fail the swap | |
2911 | * while the writer is updating the pointers, so this | |
2912 | * guarantees that the overwrite recorded here is the one we | |
2913 | * want to compare with the last_overrun. | |
2914 | */ | |
2915 | smp_mb(); | |
2916 | overwrite = local_read(&(cpu_buffer->overrun)); | |
2917 | ||
2918 | /* | |
2919 | * Here's the tricky part. | |
2920 | * | |
2921 | * We need to move the pointer past the header page. | |
2922 | * But we can only do that if a writer is not currently | |
2923 | * moving it. The page before the header page has the | |
2924 | * flag bit '1' set if it is pointing to the page we want. | |
2925 | * but if the writer is in the process of moving it | |
2926 | * than it will be '2' or already moved '0'. | |
2927 | */ | |
2928 | ||
2929 | ret = rb_head_page_replace(reader, cpu_buffer->reader_page); | |
2930 | ||
2931 | /* | |
2932 | * If we did not convert it, then we must try again. | |
2933 | */ | |
2934 | if (!ret) | |
2935 | goto spin; | |
2936 | ||
2937 | /* | |
2938 | * Yeah! We succeeded in replacing the page. | |
2939 | * | |
2940 | * Now make the new head point back to the reader page. | |
2941 | */ | |
2942 | rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list; | |
2943 | rb_inc_page(cpu_buffer, &cpu_buffer->head_page); | |
2944 | ||
2945 | /* Finally update the reader page to the new head */ | |
2946 | cpu_buffer->reader_page = reader; | |
2947 | rb_reset_reader_page(cpu_buffer); | |
2948 | ||
2949 | if (overwrite != cpu_buffer->last_overrun) { | |
2950 | cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun; | |
2951 | cpu_buffer->last_overrun = overwrite; | |
2952 | } | |
2953 | ||
2954 | goto again; | |
2955 | ||
2956 | out: | |
2957 | arch_spin_unlock(&cpu_buffer->lock); | |
2958 | local_irq_restore(flags); | |
2959 | ||
2960 | return reader; | |
2961 | } | |
2962 | ||
2963 | static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer) | |
2964 | { | |
2965 | struct ring_buffer_event *event; | |
2966 | struct buffer_page *reader; | |
2967 | unsigned length; | |
2968 | ||
2969 | reader = rb_get_reader_page(cpu_buffer); | |
2970 | ||
2971 | /* This function should not be called when buffer is empty */ | |
2972 | if (RB_WARN_ON(cpu_buffer, !reader)) | |
2973 | return; | |
2974 | ||
2975 | event = rb_reader_event(cpu_buffer); | |
2976 | ||
2977 | if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX) | |
2978 | cpu_buffer->read++; | |
2979 | ||
2980 | rb_update_read_stamp(cpu_buffer, event); | |
2981 | ||
2982 | length = rb_event_length(event); | |
2983 | cpu_buffer->reader_page->read += length; | |
2984 | } | |
2985 | ||
2986 | static void rb_advance_iter(struct ring_buffer_iter *iter) | |
2987 | { | |
2988 | struct ring_buffer_per_cpu *cpu_buffer; | |
2989 | struct ring_buffer_event *event; | |
2990 | unsigned length; | |
2991 | ||
2992 | cpu_buffer = iter->cpu_buffer; | |
2993 | ||
2994 | /* | |
2995 | * Check if we are at the end of the buffer. | |
2996 | */ | |
2997 | if (iter->head >= rb_page_size(iter->head_page)) { | |
2998 | /* discarded commits can make the page empty */ | |
2999 | if (iter->head_page == cpu_buffer->commit_page) | |
3000 | return; | |
3001 | rb_inc_iter(iter); | |
3002 | return; | |
3003 | } | |
3004 | ||
3005 | event = rb_iter_head_event(iter); | |
3006 | ||
3007 | length = rb_event_length(event); | |
3008 | ||
3009 | /* | |
3010 | * This should not be called to advance the header if we are | |
3011 | * at the tail of the buffer. | |
3012 | */ | |
3013 | if (RB_WARN_ON(cpu_buffer, | |
3014 | (iter->head_page == cpu_buffer->commit_page) && | |
3015 | (iter->head + length > rb_commit_index(cpu_buffer)))) | |
3016 | return; | |
3017 | ||
3018 | rb_update_iter_read_stamp(iter, event); | |
3019 | ||
3020 | iter->head += length; | |
3021 | ||
3022 | /* check for end of page padding */ | |
3023 | if ((iter->head >= rb_page_size(iter->head_page)) && | |
3024 | (iter->head_page != cpu_buffer->commit_page)) | |
3025 | rb_advance_iter(iter); | |
3026 | } | |
3027 | ||
3028 | static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer) | |
3029 | { | |
3030 | return cpu_buffer->lost_events; | |
3031 | } | |
3032 | ||
3033 | static struct ring_buffer_event * | |
3034 | rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts, | |
3035 | unsigned long *lost_events) | |
3036 | { | |
3037 | struct ring_buffer_event *event; | |
3038 | struct buffer_page *reader; | |
3039 | int nr_loops = 0; | |
3040 | ||
3041 | again: | |
3042 | /* | |
3043 | * We repeat when a timestamp is encountered. It is possible | |
3044 | * to get multiple timestamps from an interrupt entering just | |
3045 | * as one timestamp is about to be written, or from discarded | |
3046 | * commits. The most that we can have is the number on a single page. | |
3047 | */ | |
3048 | if (RB_WARN_ON(cpu_buffer, ++nr_loops > RB_TIMESTAMPS_PER_PAGE)) | |
3049 | return NULL; | |
3050 | ||
3051 | reader = rb_get_reader_page(cpu_buffer); | |
3052 | if (!reader) | |
3053 | return NULL; | |
3054 | ||
3055 | event = rb_reader_event(cpu_buffer); | |
3056 | ||
3057 | switch (event->type_len) { | |
3058 | case RINGBUF_TYPE_PADDING: | |
3059 | if (rb_null_event(event)) | |
3060 | RB_WARN_ON(cpu_buffer, 1); | |
3061 | /* | |
3062 | * Because the writer could be discarding every | |
3063 | * event it creates (which would probably be bad) | |
3064 | * if we were to go back to "again" then we may never | |
3065 | * catch up, and will trigger the warn on, or lock | |
3066 | * the box. Return the padding, and we will release | |
3067 | * the current locks, and try again. | |
3068 | */ | |
3069 | return event; | |
3070 | ||
3071 | case RINGBUF_TYPE_TIME_EXTEND: | |
3072 | /* Internal data, OK to advance */ | |
3073 | rb_advance_reader(cpu_buffer); | |
3074 | goto again; | |
3075 | ||
3076 | case RINGBUF_TYPE_TIME_STAMP: | |
3077 | /* FIXME: not implemented */ | |
3078 | rb_advance_reader(cpu_buffer); | |
3079 | goto again; | |
3080 | ||
3081 | case RINGBUF_TYPE_DATA: | |
3082 | if (ts) { | |
3083 | *ts = cpu_buffer->read_stamp + event->time_delta; | |
3084 | ring_buffer_normalize_time_stamp(cpu_buffer->buffer, | |
3085 | cpu_buffer->cpu, ts); | |
3086 | } | |
3087 | if (lost_events) | |
3088 | *lost_events = rb_lost_events(cpu_buffer); | |
3089 | return event; | |
3090 | ||
3091 | default: | |
3092 | BUG(); | |
3093 | } | |
3094 | ||
3095 | return NULL; | |
3096 | } | |
3097 | EXPORT_SYMBOL_GPL(ring_buffer_peek); | |
3098 | ||
3099 | static struct ring_buffer_event * | |
3100 | rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts) | |
3101 | { | |
3102 | struct ring_buffer *buffer; | |
3103 | struct ring_buffer_per_cpu *cpu_buffer; | |
3104 | struct ring_buffer_event *event; | |
3105 | int nr_loops = 0; | |
3106 | ||
3107 | cpu_buffer = iter->cpu_buffer; | |
3108 | buffer = cpu_buffer->buffer; | |
3109 | ||
3110 | /* | |
3111 | * Check if someone performed a consuming read to | |
3112 | * the buffer. A consuming read invalidates the iterator | |
3113 | * and we need to reset the iterator in this case. | |
3114 | */ | |
3115 | if (unlikely(iter->cache_read != cpu_buffer->read || | |
3116 | iter->cache_reader_page != cpu_buffer->reader_page)) | |
3117 | rb_iter_reset(iter); | |
3118 | ||
3119 | again: | |
3120 | if (ring_buffer_iter_empty(iter)) | |
3121 | return NULL; | |
3122 | ||
3123 | /* | |
3124 | * We repeat when a timestamp is encountered. | |
3125 | * We can get multiple timestamps by nested interrupts or also | |
3126 | * if filtering is on (discarding commits). Since discarding | |
3127 | * commits can be frequent we can get a lot of timestamps. | |
3128 | * But we limit them by not adding timestamps if they begin | |
3129 | * at the start of a page. | |
3130 | */ | |
3131 | if (RB_WARN_ON(cpu_buffer, ++nr_loops > RB_TIMESTAMPS_PER_PAGE)) | |
3132 | return NULL; | |
3133 | ||
3134 | if (rb_per_cpu_empty(cpu_buffer)) | |
3135 | return NULL; | |
3136 | ||
3137 | if (iter->head >= local_read(&iter->head_page->page->commit)) { | |
3138 | rb_inc_iter(iter); | |
3139 | goto again; | |
3140 | } | |
3141 | ||
3142 | event = rb_iter_head_event(iter); | |
3143 | ||
3144 | switch (event->type_len) { | |
3145 | case RINGBUF_TYPE_PADDING: | |
3146 | if (rb_null_event(event)) { | |
3147 | rb_inc_iter(iter); | |
3148 | goto again; | |
3149 | } | |
3150 | rb_advance_iter(iter); | |
3151 | return event; | |
3152 | ||
3153 | case RINGBUF_TYPE_TIME_EXTEND: | |
3154 | /* Internal data, OK to advance */ | |
3155 | rb_advance_iter(iter); | |
3156 | goto again; | |
3157 | ||
3158 | case RINGBUF_TYPE_TIME_STAMP: | |
3159 | /* FIXME: not implemented */ | |
3160 | rb_advance_iter(iter); | |
3161 | goto again; | |
3162 | ||
3163 | case RINGBUF_TYPE_DATA: | |
3164 | if (ts) { | |
3165 | *ts = iter->read_stamp + event->time_delta; | |
3166 | ring_buffer_normalize_time_stamp(buffer, | |
3167 | cpu_buffer->cpu, ts); | |
3168 | } | |
3169 | return event; | |
3170 | ||
3171 | default: | |
3172 | BUG(); | |
3173 | } | |
3174 | ||
3175 | return NULL; | |
3176 | } | |
3177 | EXPORT_SYMBOL_GPL(ring_buffer_iter_peek); | |
3178 | ||
3179 | static inline int rb_ok_to_lock(void) | |
3180 | { | |
3181 | /* | |
3182 | * If an NMI die dumps out the content of the ring buffer | |
3183 | * do not grab locks. We also permanently disable the ring | |
3184 | * buffer too. A one time deal is all you get from reading | |
3185 | * the ring buffer from an NMI. | |
3186 | */ | |
3187 | if (likely(!in_nmi())) | |
3188 | return 1; | |
3189 | ||
3190 | tracing_off_permanent(); | |
3191 | return 0; | |
3192 | } | |
3193 | ||
3194 | /** | |
3195 | * ring_buffer_peek - peek at the next event to be read | |
3196 | * @buffer: The ring buffer to read | |
3197 | * @cpu: The cpu to peak at | |
3198 | * @ts: The timestamp counter of this event. | |
3199 | * @lost_events: a variable to store if events were lost (may be NULL) | |
3200 | * | |
3201 | * This will return the event that will be read next, but does | |
3202 | * not consume the data. | |
3203 | */ | |
3204 | struct ring_buffer_event * | |
3205 | ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts, | |
3206 | unsigned long *lost_events) | |
3207 | { | |
3208 | struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu]; | |
3209 | struct ring_buffer_event *event; | |
3210 | unsigned long flags; | |
3211 | int dolock; | |
3212 | ||
3213 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) | |
3214 | return NULL; | |
3215 | ||
3216 | dolock = rb_ok_to_lock(); | |
3217 | again: | |
3218 | local_irq_save(flags); | |
3219 | if (dolock) | |
3220 | spin_lock(&cpu_buffer->reader_lock); | |
3221 | event = rb_buffer_peek(cpu_buffer, ts, lost_events); | |
3222 | if (event && event->type_len == RINGBUF_TYPE_PADDING) | |
3223 | rb_advance_reader(cpu_buffer); | |
3224 | if (dolock) | |
3225 | spin_unlock(&cpu_buffer->reader_lock); | |
3226 | local_irq_restore(flags); | |
3227 | ||
3228 | if (event && event->type_len == RINGBUF_TYPE_PADDING) | |
3229 | goto again; | |
3230 | ||
3231 | return event; | |
3232 | } | |
3233 | ||
3234 | /** | |
3235 | * ring_buffer_iter_peek - peek at the next event to be read | |
3236 | * @iter: The ring buffer iterator | |
3237 | * @ts: The timestamp counter of this event. | |
3238 | * | |
3239 | * This will return the event that will be read next, but does | |
3240 | * not increment the iterator. | |
3241 | */ | |
3242 | struct ring_buffer_event * | |
3243 | ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts) | |
3244 | { | |
3245 | struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer; | |
3246 | struct ring_buffer_event *event; | |
3247 | unsigned long flags; | |
3248 | ||
3249 | again: | |
3250 | spin_lock_irqsave(&cpu_buffer->reader_lock, flags); | |
3251 | event = rb_iter_peek(iter, ts); | |
3252 | spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags); | |
3253 | ||
3254 | if (event && event->type_len == RINGBUF_TYPE_PADDING) | |
3255 | goto again; | |
3256 | ||
3257 | return event; | |
3258 | } | |
3259 | ||
3260 | /** | |
3261 | * ring_buffer_consume - return an event and consume it | |
3262 | * @buffer: The ring buffer to get the next event from | |
3263 | * @cpu: the cpu to read the buffer from | |
3264 | * @ts: a variable to store the timestamp (may be NULL) | |
3265 | * @lost_events: a variable to store if events were lost (may be NULL) | |
3266 | * | |
3267 | * Returns the next event in the ring buffer, and that event is consumed. | |
3268 | * Meaning, that sequential reads will keep returning a different event, | |
3269 | * and eventually empty the ring buffer if the producer is slower. | |
3270 | */ | |
3271 | struct ring_buffer_event * | |
3272 | ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts, | |
3273 | unsigned long *lost_events) | |
3274 | { | |
3275 | struct ring_buffer_per_cpu *cpu_buffer; | |
3276 | struct ring_buffer_event *event = NULL; | |
3277 | unsigned long flags; | |
3278 | int dolock; | |
3279 | ||
3280 | dolock = rb_ok_to_lock(); | |
3281 | ||
3282 | again: | |
3283 | /* might be called in atomic */ | |
3284 | preempt_disable(); | |
3285 | ||
3286 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) | |
3287 | goto out; | |
3288 | ||
3289 | cpu_buffer = buffer->buffers[cpu]; | |
3290 | local_irq_save(flags); | |
3291 | if (dolock) | |
3292 | spin_lock(&cpu_buffer->reader_lock); | |
3293 | ||
3294 | event = rb_buffer_peek(cpu_buffer, ts, lost_events); | |
3295 | if (event) { | |
3296 | cpu_buffer->lost_events = 0; | |
3297 | rb_advance_reader(cpu_buffer); | |
3298 | } | |
3299 | ||
3300 | if (dolock) | |
3301 | spin_unlock(&cpu_buffer->reader_lock); | |
3302 | local_irq_restore(flags); | |
3303 | ||
3304 | out: | |
3305 | preempt_enable(); | |
3306 | ||
3307 | if (event && event->type_len == RINGBUF_TYPE_PADDING) | |
3308 | goto again; | |
3309 | ||
3310 | return event; | |
3311 | } | |
3312 | EXPORT_SYMBOL_GPL(ring_buffer_consume); | |
3313 | ||
3314 | /** | |
3315 | * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer | |
3316 | * @buffer: The ring buffer to read from | |
3317 | * @cpu: The cpu buffer to iterate over | |
3318 | * | |
3319 | * This performs the initial preparations necessary to iterate | |
3320 | * through the buffer. Memory is allocated, buffer recording | |
3321 | * is disabled, and the iterator pointer is returned to the caller. | |
3322 | * | |
3323 | * Disabling buffer recordng prevents the reading from being | |
3324 | * corrupted. This is not a consuming read, so a producer is not | |
3325 | * expected. | |
3326 | * | |
3327 | * After a sequence of ring_buffer_read_prepare calls, the user is | |
3328 | * expected to make at least one call to ring_buffer_prepare_sync. | |
3329 | * Afterwards, ring_buffer_read_start is invoked to get things going | |
3330 | * for real. | |
3331 | * | |
3332 | * This overall must be paired with ring_buffer_finish. | |
3333 | */ | |
3334 | struct ring_buffer_iter * | |
3335 | ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu) | |
3336 | { | |
3337 | struct ring_buffer_per_cpu *cpu_buffer; | |
3338 | struct ring_buffer_iter *iter; | |
3339 | ||
3340 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) | |
3341 | return NULL; | |
3342 | ||
3343 | iter = kmalloc(sizeof(*iter), GFP_KERNEL); | |
3344 | if (!iter) | |
3345 | return NULL; | |
3346 | ||
3347 | cpu_buffer = buffer->buffers[cpu]; | |
3348 | ||
3349 | iter->cpu_buffer = cpu_buffer; | |
3350 | ||
3351 | atomic_inc(&cpu_buffer->record_disabled); | |
3352 | ||
3353 | return iter; | |
3354 | } | |
3355 | EXPORT_SYMBOL_GPL(ring_buffer_read_prepare); | |
3356 | ||
3357 | /** | |
3358 | * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls | |
3359 | * | |
3360 | * All previously invoked ring_buffer_read_prepare calls to prepare | |
3361 | * iterators will be synchronized. Afterwards, read_buffer_read_start | |
3362 | * calls on those iterators are allowed. | |
3363 | */ | |
3364 | void | |
3365 | ring_buffer_read_prepare_sync(void) | |
3366 | { | |
3367 | synchronize_sched(); | |
3368 | } | |
3369 | EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync); | |
3370 | ||
3371 | /** | |
3372 | * ring_buffer_read_start - start a non consuming read of the buffer | |
3373 | * @iter: The iterator returned by ring_buffer_read_prepare | |
3374 | * | |
3375 | * This finalizes the startup of an iteration through the buffer. | |
3376 | * The iterator comes from a call to ring_buffer_read_prepare and | |
3377 | * an intervening ring_buffer_read_prepare_sync must have been | |
3378 | * performed. | |
3379 | * | |
3380 | * Must be paired with ring_buffer_finish. | |
3381 | */ | |
3382 | void | |
3383 | ring_buffer_read_start(struct ring_buffer_iter *iter) | |
3384 | { | |
3385 | struct ring_buffer_per_cpu *cpu_buffer; | |
3386 | unsigned long flags; | |
3387 | ||
3388 | if (!iter) | |
3389 | return; | |
3390 | ||
3391 | cpu_buffer = iter->cpu_buffer; | |
3392 | ||
3393 | spin_lock_irqsave(&cpu_buffer->reader_lock, flags); | |
3394 | arch_spin_lock(&cpu_buffer->lock); | |
3395 | rb_iter_reset(iter); | |
3396 | arch_spin_unlock(&cpu_buffer->lock); | |
3397 | spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags); | |
3398 | } | |
3399 | EXPORT_SYMBOL_GPL(ring_buffer_read_start); | |
3400 | ||
3401 | /** | |
3402 | * ring_buffer_finish - finish reading the iterator of the buffer | |
3403 | * @iter: The iterator retrieved by ring_buffer_start | |
3404 | * | |
3405 | * This re-enables the recording to the buffer, and frees the | |
3406 | * iterator. | |
3407 | */ | |
3408 | void | |
3409 | ring_buffer_read_finish(struct ring_buffer_iter *iter) | |
3410 | { | |
3411 | struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer; | |
3412 | ||
3413 | atomic_dec(&cpu_buffer->record_disabled); | |
3414 | kfree(iter); | |
3415 | } | |
3416 | EXPORT_SYMBOL_GPL(ring_buffer_read_finish); | |
3417 | ||
3418 | /** | |
3419 | * ring_buffer_read - read the next item in the ring buffer by the iterator | |
3420 | * @iter: The ring buffer iterator | |
3421 | * @ts: The time stamp of the event read. | |
3422 | * | |
3423 | * This reads the next event in the ring buffer and increments the iterator. | |
3424 | */ | |
3425 | struct ring_buffer_event * | |
3426 | ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts) | |
3427 | { | |
3428 | struct ring_buffer_event *event; | |
3429 | struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer; | |
3430 | unsigned long flags; | |
3431 | ||
3432 | spin_lock_irqsave(&cpu_buffer->reader_lock, flags); | |
3433 | again: | |
3434 | event = rb_iter_peek(iter, ts); | |
3435 | if (!event) | |
3436 | goto out; | |
3437 | ||
3438 | if (event->type_len == RINGBUF_TYPE_PADDING) | |
3439 | goto again; | |
3440 | ||
3441 | rb_advance_iter(iter); | |
3442 | out: | |
3443 | spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags); | |
3444 | ||
3445 | return event; | |
3446 | } | |
3447 | EXPORT_SYMBOL_GPL(ring_buffer_read); | |
3448 | ||
3449 | /** | |
3450 | * ring_buffer_size - return the size of the ring buffer (in bytes) | |
3451 | * @buffer: The ring buffer. | |
3452 | */ | |
3453 | unsigned long ring_buffer_size(struct ring_buffer *buffer) | |
3454 | { | |
3455 | return BUF_PAGE_SIZE * buffer->pages; | |
3456 | } | |
3457 | EXPORT_SYMBOL_GPL(ring_buffer_size); | |
3458 | ||
3459 | static void | |
3460 | rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer) | |
3461 | { | |
3462 | rb_head_page_deactivate(cpu_buffer); | |
3463 | ||
3464 | cpu_buffer->head_page | |
3465 | = list_entry(cpu_buffer->pages, struct buffer_page, list); | |
3466 | local_set(&cpu_buffer->head_page->write, 0); | |
3467 | local_set(&cpu_buffer->head_page->entries, 0); | |
3468 | local_set(&cpu_buffer->head_page->page->commit, 0); | |
3469 | ||
3470 | cpu_buffer->head_page->read = 0; | |
3471 | ||
3472 | cpu_buffer->tail_page = cpu_buffer->head_page; | |
3473 | cpu_buffer->commit_page = cpu_buffer->head_page; | |
3474 | ||
3475 | INIT_LIST_HEAD(&cpu_buffer->reader_page->list); | |
3476 | local_set(&cpu_buffer->reader_page->write, 0); | |
3477 | local_set(&cpu_buffer->reader_page->entries, 0); | |
3478 | local_set(&cpu_buffer->reader_page->page->commit, 0); | |
3479 | cpu_buffer->reader_page->read = 0; | |
3480 | ||
3481 | local_set(&cpu_buffer->commit_overrun, 0); | |
3482 | local_set(&cpu_buffer->overrun, 0); | |
3483 | local_set(&cpu_buffer->entries, 0); | |
3484 | local_set(&cpu_buffer->committing, 0); | |
3485 | local_set(&cpu_buffer->commits, 0); | |
3486 | cpu_buffer->read = 0; | |
3487 | ||
3488 | cpu_buffer->write_stamp = 0; | |
3489 | cpu_buffer->read_stamp = 0; | |
3490 | ||
3491 | cpu_buffer->lost_events = 0; | |
3492 | cpu_buffer->last_overrun = 0; | |
3493 | ||
3494 | rb_head_page_activate(cpu_buffer); | |
3495 | } | |
3496 | ||
3497 | /** | |
3498 | * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer | |
3499 | * @buffer: The ring buffer to reset a per cpu buffer of | |
3500 | * @cpu: The CPU buffer to be reset | |
3501 | */ | |
3502 | void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu) | |
3503 | { | |
3504 | struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu]; | |
3505 | unsigned long flags; | |
3506 | ||
3507 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) | |
3508 | return; | |
3509 | ||
3510 | atomic_inc(&cpu_buffer->record_disabled); | |
3511 | ||
3512 | spin_lock_irqsave(&cpu_buffer->reader_lock, flags); | |
3513 | ||
3514 | if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing))) | |
3515 | goto out; | |
3516 | ||
3517 | arch_spin_lock(&cpu_buffer->lock); | |
3518 | ||
3519 | rb_reset_cpu(cpu_buffer); | |
3520 | ||
3521 | arch_spin_unlock(&cpu_buffer->lock); | |
3522 | ||
3523 | out: | |
3524 | spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags); | |
3525 | ||
3526 | atomic_dec(&cpu_buffer->record_disabled); | |
3527 | } | |
3528 | EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu); | |
3529 | ||
3530 | /** | |
3531 | * ring_buffer_reset - reset a ring buffer | |
3532 | * @buffer: The ring buffer to reset all cpu buffers | |
3533 | */ | |
3534 | void ring_buffer_reset(struct ring_buffer *buffer) | |
3535 | { | |
3536 | int cpu; | |
3537 | ||
3538 | for_each_buffer_cpu(buffer, cpu) | |
3539 | ring_buffer_reset_cpu(buffer, cpu); | |
3540 | } | |
3541 | EXPORT_SYMBOL_GPL(ring_buffer_reset); | |
3542 | ||
3543 | /** | |
3544 | * rind_buffer_empty - is the ring buffer empty? | |
3545 | * @buffer: The ring buffer to test | |
3546 | */ | |
3547 | int ring_buffer_empty(struct ring_buffer *buffer) | |
3548 | { | |
3549 | struct ring_buffer_per_cpu *cpu_buffer; | |
3550 | unsigned long flags; | |
3551 | int dolock; | |
3552 | int cpu; | |
3553 | int ret; | |
3554 | ||
3555 | dolock = rb_ok_to_lock(); | |
3556 | ||
3557 | /* yes this is racy, but if you don't like the race, lock the buffer */ | |
3558 | for_each_buffer_cpu(buffer, cpu) { | |
3559 | cpu_buffer = buffer->buffers[cpu]; | |
3560 | local_irq_save(flags); | |
3561 | if (dolock) | |
3562 | spin_lock(&cpu_buffer->reader_lock); | |
3563 | ret = rb_per_cpu_empty(cpu_buffer); | |
3564 | if (dolock) | |
3565 | spin_unlock(&cpu_buffer->reader_lock); | |
3566 | local_irq_restore(flags); | |
3567 | ||
3568 | if (!ret) | |
3569 | return 0; | |
3570 | } | |
3571 | ||
3572 | return 1; | |
3573 | } | |
3574 | EXPORT_SYMBOL_GPL(ring_buffer_empty); | |
3575 | ||
3576 | /** | |
3577 | * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty? | |
3578 | * @buffer: The ring buffer | |
3579 | * @cpu: The CPU buffer to test | |
3580 | */ | |
3581 | int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu) | |
3582 | { | |
3583 | struct ring_buffer_per_cpu *cpu_buffer; | |
3584 | unsigned long flags; | |
3585 | int dolock; | |
3586 | int ret; | |
3587 | ||
3588 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) | |
3589 | return 1; | |
3590 | ||
3591 | dolock = rb_ok_to_lock(); | |
3592 | ||
3593 | cpu_buffer = buffer->buffers[cpu]; | |
3594 | local_irq_save(flags); | |
3595 | if (dolock) | |
3596 | spin_lock(&cpu_buffer->reader_lock); | |
3597 | ret = rb_per_cpu_empty(cpu_buffer); | |
3598 | if (dolock) | |
3599 | spin_unlock(&cpu_buffer->reader_lock); | |
3600 | local_irq_restore(flags); | |
3601 | ||
3602 | return ret; | |
3603 | } | |
3604 | EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu); | |
3605 | ||
3606 | #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP | |
3607 | /** | |
3608 | * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers | |
3609 | * @buffer_a: One buffer to swap with | |
3610 | * @buffer_b: The other buffer to swap with | |
3611 | * | |
3612 | * This function is useful for tracers that want to take a "snapshot" | |
3613 | * of a CPU buffer and has another back up buffer lying around. | |
3614 | * it is expected that the tracer handles the cpu buffer not being | |
3615 | * used at the moment. | |
3616 | */ | |
3617 | int ring_buffer_swap_cpu(struct ring_buffer *buffer_a, | |
3618 | struct ring_buffer *buffer_b, int cpu) | |
3619 | { | |
3620 | struct ring_buffer_per_cpu *cpu_buffer_a; | |
3621 | struct ring_buffer_per_cpu *cpu_buffer_b; | |
3622 | int ret = -EINVAL; | |
3623 | ||
3624 | if (!cpumask_test_cpu(cpu, buffer_a->cpumask) || | |
3625 | !cpumask_test_cpu(cpu, buffer_b->cpumask)) | |
3626 | goto out; | |
3627 | ||
3628 | /* At least make sure the two buffers are somewhat the same */ | |
3629 | if (buffer_a->pages != buffer_b->pages) | |
3630 | goto out; | |
3631 | ||
3632 | ret = -EAGAIN; | |
3633 | ||
3634 | if (ring_buffer_flags != RB_BUFFERS_ON) | |
3635 | goto out; | |
3636 | ||
3637 | if (atomic_read(&buffer_a->record_disabled)) | |
3638 | goto out; | |
3639 | ||
3640 | if (atomic_read(&buffer_b->record_disabled)) | |
3641 | goto out; | |
3642 | ||
3643 | cpu_buffer_a = buffer_a->buffers[cpu]; | |
3644 | cpu_buffer_b = buffer_b->buffers[cpu]; | |
3645 | ||
3646 | if (atomic_read(&cpu_buffer_a->record_disabled)) | |
3647 | goto out; | |
3648 | ||
3649 | if (atomic_read(&cpu_buffer_b->record_disabled)) | |
3650 | goto out; | |
3651 | ||
3652 | /* | |
3653 | * We can't do a synchronize_sched here because this | |
3654 | * function can be called in atomic context. | |
3655 | * Normally this will be called from the same CPU as cpu. | |
3656 | * If not it's up to the caller to protect this. | |
3657 | */ | |
3658 | atomic_inc(&cpu_buffer_a->record_disabled); | |
3659 | atomic_inc(&cpu_buffer_b->record_disabled); | |
3660 | ||
3661 | ret = -EBUSY; | |
3662 | if (local_read(&cpu_buffer_a->committing)) | |
3663 | goto out_dec; | |
3664 | if (local_read(&cpu_buffer_b->committing)) | |
3665 | goto out_dec; | |
3666 | ||
3667 | buffer_a->buffers[cpu] = cpu_buffer_b; | |
3668 | buffer_b->buffers[cpu] = cpu_buffer_a; | |
3669 | ||
3670 | cpu_buffer_b->buffer = buffer_a; | |
3671 | cpu_buffer_a->buffer = buffer_b; | |
3672 | ||
3673 | ret = 0; | |
3674 | ||
3675 | out_dec: | |
3676 | atomic_dec(&cpu_buffer_a->record_disabled); | |
3677 | atomic_dec(&cpu_buffer_b->record_disabled); | |
3678 | out: | |
3679 | return ret; | |
3680 | } | |
3681 | EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu); | |
3682 | #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */ | |
3683 | ||
3684 | /** | |
3685 | * ring_buffer_alloc_read_page - allocate a page to read from buffer | |
3686 | * @buffer: the buffer to allocate for. | |
3687 | * | |
3688 | * This function is used in conjunction with ring_buffer_read_page. | |
3689 | * When reading a full page from the ring buffer, these functions | |
3690 | * can be used to speed up the process. The calling function should | |
3691 | * allocate a few pages first with this function. Then when it | |
3692 | * needs to get pages from the ring buffer, it passes the result | |
3693 | * of this function into ring_buffer_read_page, which will swap | |
3694 | * the page that was allocated, with the read page of the buffer. | |
3695 | * | |
3696 | * Returns: | |
3697 | * The page allocated, or NULL on error. | |
3698 | */ | |
3699 | void *ring_buffer_alloc_read_page(struct ring_buffer *buffer) | |
3700 | { | |
3701 | struct buffer_data_page *bpage; | |
3702 | unsigned long addr; | |
3703 | ||
3704 | addr = __get_free_page(GFP_KERNEL); | |
3705 | if (!addr) | |
3706 | return NULL; | |
3707 | ||
3708 | bpage = (void *)addr; | |
3709 | ||
3710 | rb_init_page(bpage); | |
3711 | ||
3712 | return bpage; | |
3713 | } | |
3714 | EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page); | |
3715 | ||
3716 | /** | |
3717 | * ring_buffer_free_read_page - free an allocated read page | |
3718 | * @buffer: the buffer the page was allocate for | |
3719 | * @data: the page to free | |
3720 | * | |
3721 | * Free a page allocated from ring_buffer_alloc_read_page. | |
3722 | */ | |
3723 | void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data) | |
3724 | { | |
3725 | free_page((unsigned long)data); | |
3726 | } | |
3727 | EXPORT_SYMBOL_GPL(ring_buffer_free_read_page); | |
3728 | ||
3729 | /** | |
3730 | * ring_buffer_read_page - extract a page from the ring buffer | |
3731 | * @buffer: buffer to extract from | |
3732 | * @data_page: the page to use allocated from ring_buffer_alloc_read_page | |
3733 | * @len: amount to extract | |
3734 | * @cpu: the cpu of the buffer to extract | |
3735 | * @full: should the extraction only happen when the page is full. | |
3736 | * | |
3737 | * This function will pull out a page from the ring buffer and consume it. | |
3738 | * @data_page must be the address of the variable that was returned | |
3739 | * from ring_buffer_alloc_read_page. This is because the page might be used | |
3740 | * to swap with a page in the ring buffer. | |
3741 | * | |
3742 | * for example: | |
3743 | * rpage = ring_buffer_alloc_read_page(buffer); | |
3744 | * if (!rpage) | |
3745 | * return error; | |
3746 | * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0); | |
3747 | * if (ret >= 0) | |
3748 | * process_page(rpage, ret); | |
3749 | * | |
3750 | * When @full is set, the function will not return true unless | |
3751 | * the writer is off the reader page. | |
3752 | * | |
3753 | * Note: it is up to the calling functions to handle sleeps and wakeups. | |
3754 | * The ring buffer can be used anywhere in the kernel and can not | |
3755 | * blindly call wake_up. The layer that uses the ring buffer must be | |
3756 | * responsible for that. | |
3757 | * | |
3758 | * Returns: | |
3759 | * >=0 if data has been transferred, returns the offset of consumed data. | |
3760 | * <0 if no data has been transferred. | |
3761 | */ | |
3762 | int ring_buffer_read_page(struct ring_buffer *buffer, | |
3763 | void **data_page, size_t len, int cpu, int full) | |
3764 | { | |
3765 | struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu]; | |
3766 | struct ring_buffer_event *event; | |
3767 | struct buffer_data_page *bpage; | |
3768 | struct buffer_page *reader; | |
3769 | unsigned long missed_events; | |
3770 | unsigned long flags; | |
3771 | unsigned int commit; | |
3772 | unsigned int read; | |
3773 | u64 save_timestamp; | |
3774 | int ret = -1; | |
3775 | ||
3776 | if (!cpumask_test_cpu(cpu, buffer->cpumask)) | |
3777 | goto out; | |
3778 | ||
3779 | /* | |
3780 | * If len is not big enough to hold the page header, then | |
3781 | * we can not copy anything. | |
3782 | */ | |
3783 | if (len <= BUF_PAGE_HDR_SIZE) | |
3784 | goto out; | |
3785 | ||
3786 | len -= BUF_PAGE_HDR_SIZE; | |
3787 | ||
3788 | if (!data_page) | |
3789 | goto out; | |
3790 | ||
3791 | bpage = *data_page; | |
3792 | if (!bpage) | |
3793 | goto out; | |
3794 | ||
3795 | spin_lock_irqsave(&cpu_buffer->reader_lock, flags); | |
3796 | ||
3797 | reader = rb_get_reader_page(cpu_buffer); | |
3798 | if (!reader) | |
3799 | goto out_unlock; | |
3800 | ||
3801 | event = rb_reader_event(cpu_buffer); | |
3802 | ||
3803 | read = reader->read; | |
3804 | commit = rb_page_commit(reader); | |
3805 | ||
3806 | /* Check if any events were dropped */ | |
3807 | missed_events = cpu_buffer->lost_events; | |
3808 | ||
3809 | /* | |
3810 | * If this page has been partially read or | |
3811 | * if len is not big enough to read the rest of the page or | |
3812 | * a writer is still on the page, then | |
3813 | * we must copy the data from the page to the buffer. | |
3814 | * Otherwise, we can simply swap the page with the one passed in. | |
3815 | */ | |
3816 | if (read || (len < (commit - read)) || | |
3817 | cpu_buffer->reader_page == cpu_buffer->commit_page) { | |
3818 | struct buffer_data_page *rpage = cpu_buffer->reader_page->page; | |
3819 | unsigned int rpos = read; | |
3820 | unsigned int pos = 0; | |
3821 | unsigned int size; | |
3822 | ||
3823 | if (full) | |
3824 | goto out_unlock; | |
3825 | ||
3826 | if (len > (commit - read)) | |
3827 | len = (commit - read); | |
3828 | ||
3829 | size = rb_event_length(event); | |
3830 | ||
3831 | if (len < size) | |
3832 | goto out_unlock; | |
3833 | ||
3834 | /* save the current timestamp, since the user will need it */ | |
3835 | save_timestamp = cpu_buffer->read_stamp; | |
3836 | ||
3837 | /* Need to copy one event at a time */ | |
3838 | do { | |
3839 | memcpy(bpage->data + pos, rpage->data + rpos, size); | |
3840 | ||
3841 | len -= size; | |
3842 | ||
3843 | rb_advance_reader(cpu_buffer); | |
3844 | rpos = reader->read; | |
3845 | pos += size; | |
3846 | ||
3847 | if (rpos >= commit) | |
3848 | break; | |
3849 | ||
3850 | event = rb_reader_event(cpu_buffer); | |
3851 | size = rb_event_length(event); | |
3852 | } while (len > size); | |
3853 | ||
3854 | /* update bpage */ | |
3855 | local_set(&bpage->commit, pos); | |
3856 | bpage->time_stamp = save_timestamp; | |
3857 | ||
3858 | /* we copied everything to the beginning */ | |
3859 | read = 0; | |
3860 | } else { | |
3861 | /* update the entry counter */ | |
3862 | cpu_buffer->read += rb_page_entries(reader); | |
3863 | ||
3864 | /* swap the pages */ | |
3865 | rb_init_page(bpage); | |
3866 | bpage = reader->page; | |
3867 | reader->page = *data_page; | |
3868 | local_set(&reader->write, 0); | |
3869 | local_set(&reader->entries, 0); | |
3870 | reader->read = 0; | |
3871 | *data_page = bpage; | |
3872 | ||
3873 | /* | |
3874 | * Use the real_end for the data size, | |
3875 | * This gives us a chance to store the lost events | |
3876 | * on the page. | |
3877 | */ | |
3878 | if (reader->real_end) | |
3879 | local_set(&bpage->commit, reader->real_end); | |
3880 | } | |
3881 | ret = read; | |
3882 | ||
3883 | cpu_buffer->lost_events = 0; | |
3884 | ||
3885 | commit = local_read(&bpage->commit); | |
3886 | /* | |
3887 | * Set a flag in the commit field if we lost events | |
3888 | */ | |
3889 | if (missed_events) { | |
3890 | /* If there is room at the end of the page to save the | |
3891 | * missed events, then record it there. | |
3892 | */ | |
3893 | if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) { | |
3894 | memcpy(&bpage->data[commit], &missed_events, | |
3895 | sizeof(missed_events)); | |
3896 | local_add(RB_MISSED_STORED, &bpage->commit); | |
3897 | commit += sizeof(missed_events); | |
3898 | } | |
3899 | local_add(RB_MISSED_EVENTS, &bpage->commit); | |
3900 | } | |
3901 | ||
3902 | /* | |
3903 | * This page may be off to user land. Zero it out here. | |
3904 | */ | |
3905 | if (commit < BUF_PAGE_SIZE) | |
3906 | memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit); | |
3907 | ||
3908 | out_unlock: | |
3909 | spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags); | |
3910 | ||
3911 | out: | |
3912 | return ret; | |
3913 | } | |
3914 | EXPORT_SYMBOL_GPL(ring_buffer_read_page); | |
3915 | ||
3916 | #ifdef CONFIG_TRACING | |
3917 | static ssize_t | |
3918 | rb_simple_read(struct file *filp, char __user *ubuf, | |
3919 | size_t cnt, loff_t *ppos) | |
3920 | { | |
3921 | unsigned long *p = filp->private_data; | |
3922 | char buf[64]; | |
3923 | int r; | |
3924 | ||
3925 | if (test_bit(RB_BUFFERS_DISABLED_BIT, p)) | |
3926 | r = sprintf(buf, "permanently disabled\n"); | |
3927 | else | |
3928 | r = sprintf(buf, "%d\n", test_bit(RB_BUFFERS_ON_BIT, p)); | |
3929 | ||
3930 | return simple_read_from_buffer(ubuf, cnt, ppos, buf, r); | |
3931 | } | |
3932 | ||
3933 | static ssize_t | |
3934 | rb_simple_write(struct file *filp, const char __user *ubuf, | |
3935 | size_t cnt, loff_t *ppos) | |
3936 | { | |
3937 | unsigned long *p = filp->private_data; | |
3938 | char buf[64]; | |
3939 | unsigned long val; | |
3940 | int ret; | |
3941 | ||
3942 | if (cnt >= sizeof(buf)) | |
3943 | return -EINVAL; | |
3944 | ||
3945 | if (copy_from_user(&buf, ubuf, cnt)) | |
3946 | return -EFAULT; | |
3947 | ||
3948 | buf[cnt] = 0; | |
3949 | ||
3950 | ret = strict_strtoul(buf, 10, &val); | |
3951 | if (ret < 0) | |
3952 | return ret; | |
3953 | ||
3954 | if (val) | |
3955 | set_bit(RB_BUFFERS_ON_BIT, p); | |
3956 | else | |
3957 | clear_bit(RB_BUFFERS_ON_BIT, p); | |
3958 | ||
3959 | (*ppos)++; | |
3960 | ||
3961 | return cnt; | |
3962 | } | |
3963 | ||
3964 | static const struct file_operations rb_simple_fops = { | |
3965 | .open = tracing_open_generic, | |
3966 | .read = rb_simple_read, | |
3967 | .write = rb_simple_write, | |
3968 | }; | |
3969 | ||
3970 | ||
3971 | static __init int rb_init_debugfs(void) | |
3972 | { | |
3973 | struct dentry *d_tracer; | |
3974 | ||
3975 | d_tracer = tracing_init_dentry(); | |
3976 | ||
3977 | trace_create_file("tracing_on", 0644, d_tracer, | |
3978 | &ring_buffer_flags, &rb_simple_fops); | |
3979 | ||
3980 | return 0; | |
3981 | } | |
3982 | ||
3983 | fs_initcall(rb_init_debugfs); | |
3984 | #endif | |
3985 | ||
3986 | #ifdef CONFIG_HOTPLUG_CPU | |
3987 | static int rb_cpu_notify(struct notifier_block *self, | |
3988 | unsigned long action, void *hcpu) | |
3989 | { | |
3990 | struct ring_buffer *buffer = | |
3991 | container_of(self, struct ring_buffer, cpu_notify); | |
3992 | long cpu = (long)hcpu; | |
3993 | ||
3994 | switch (action) { | |
3995 | case CPU_UP_PREPARE: | |
3996 | case CPU_UP_PREPARE_FROZEN: | |
3997 | if (cpumask_test_cpu(cpu, buffer->cpumask)) | |
3998 | return NOTIFY_OK; | |
3999 | ||
4000 | buffer->buffers[cpu] = | |
4001 | rb_allocate_cpu_buffer(buffer, cpu); | |
4002 | if (!buffer->buffers[cpu]) { | |
4003 | WARN(1, "failed to allocate ring buffer on CPU %ld\n", | |
4004 | cpu); | |
4005 | return NOTIFY_OK; | |
4006 | } | |
4007 | smp_wmb(); | |
4008 | cpumask_set_cpu(cpu, buffer->cpumask); | |
4009 | break; | |
4010 | case CPU_DOWN_PREPARE: | |
4011 | case CPU_DOWN_PREPARE_FROZEN: | |
4012 | /* | |
4013 | * Do nothing. | |
4014 | * If we were to free the buffer, then the user would | |
4015 | * lose any trace that was in the buffer. | |
4016 | */ | |
4017 | break; | |
4018 | default: | |
4019 | break; | |
4020 | } | |
4021 | return NOTIFY_OK; | |
4022 | } | |
4023 | #endif |