1 /*
2 * Generic ring buffer
3 *
4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
5 */
6 #include <linux/trace_events.h>
7 #include <linux/ring_buffer.h>
8 #include <linux/trace_clock.h>
9 #include <linux/trace_seq.h>
10 #include <linux/spinlock.h>
11 #include <linux/irq_work.h>
12 #include <linux/uaccess.h>
13 #include <linux/hardirq.h>
14 #include <linux/kthread.h> /* for self test */
15 #include <linux/kmemcheck.h>
16 #include <linux/module.h>
17 #include <linux/percpu.h>
18 #include <linux/mutex.h>
19 #include <linux/delay.h>
20 #include <linux/slab.h>
21 #include <linux/init.h>
22 #include <linux/hash.h>
23 #include <linux/list.h>
24 #include <linux/cpu.h>
25
26 #include <asm/local.h>
27
28 static void update_pages_handler(struct work_struct *work);
29
30 /*
31 * The ring buffer header is special. We must manually up keep it.
32 */
ring_buffer_print_entry_header(struct trace_seq * s)33 int ring_buffer_print_entry_header(struct trace_seq *s)
34 {
35 trace_seq_puts(s, "# compressed entry header\n");
36 trace_seq_puts(s, "\ttype_len : 5 bits\n");
37 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
38 trace_seq_puts(s, "\tarray : 32 bits\n");
39 trace_seq_putc(s, '\n');
40 trace_seq_printf(s, "\tpadding : type == %d\n",
41 RINGBUF_TYPE_PADDING);
42 trace_seq_printf(s, "\ttime_extend : type == %d\n",
43 RINGBUF_TYPE_TIME_EXTEND);
44 trace_seq_printf(s, "\tdata max type_len == %d\n",
45 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
46
47 return !trace_seq_has_overflowed(s);
48 }
49
50 /*
51 * The ring buffer is made up of a list of pages. A separate list of pages is
52 * allocated for each CPU. A writer may only write to a buffer that is
53 * associated with the CPU it is currently executing on. A reader may read
54 * from any per cpu buffer.
55 *
56 * The reader is special. For each per cpu buffer, the reader has its own
57 * reader page. When a reader has read the entire reader page, this reader
58 * page is swapped with another page in the ring buffer.
59 *
60 * Now, as long as the writer is off the reader page, the reader can do what
61 * ever it wants with that page. The writer will never write to that page
62 * again (as long as it is out of the ring buffer).
63 *
64 * Here's some silly ASCII art.
65 *
66 * +------+
67 * |reader| RING BUFFER
68 * |page |
69 * +------+ +---+ +---+ +---+
70 * | |-->| |-->| |
71 * +---+ +---+ +---+
72 * ^ |
73 * | |
74 * +---------------+
75 *
76 *
77 * +------+
78 * |reader| RING BUFFER
79 * |page |------------------v
80 * +------+ +---+ +---+ +---+
81 * | |-->| |-->| |
82 * +---+ +---+ +---+
83 * ^ |
84 * | |
85 * +---------------+
86 *
87 *
88 * +------+
89 * |reader| RING BUFFER
90 * |page |------------------v
91 * +------+ +---+ +---+ +---+
92 * ^ | |-->| |-->| |
93 * | +---+ +---+ +---+
94 * | |
95 * | |
96 * +------------------------------+
97 *
98 *
99 * +------+
100 * |buffer| RING BUFFER
101 * |page |------------------v
102 * +------+ +---+ +---+ +---+
103 * ^ | | | |-->| |
104 * | New +---+ +---+ +---+
105 * | Reader------^ |
106 * | page |
107 * +------------------------------+
108 *
109 *
110 * After we make this swap, the reader can hand this page off to the splice
111 * code and be done with it. It can even allocate a new page if it needs to
112 * and swap that into the ring buffer.
113 *
114 * We will be using cmpxchg soon to make all this lockless.
115 *
116 */
117
118 /* Used for individual buffers (after the counter) */
119 #define RB_BUFFER_OFF (1 << 20)
120
121 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
122
123 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
124 #define RB_ALIGNMENT 4U
125 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
126 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
127
128 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
129 # define RB_FORCE_8BYTE_ALIGNMENT 0
130 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
131 #else
132 # define RB_FORCE_8BYTE_ALIGNMENT 1
133 # define RB_ARCH_ALIGNMENT 8U
134 #endif
135
136 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
137
138 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
139 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
140
141 enum {
142 RB_LEN_TIME_EXTEND = 8,
143 RB_LEN_TIME_STAMP = 16,
144 };
145
146 #define skip_time_extend(event) \
147 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
148
rb_null_event(struct ring_buffer_event * event)149 static inline int rb_null_event(struct ring_buffer_event *event)
150 {
151 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
152 }
153
rb_event_set_padding(struct ring_buffer_event * event)154 static void rb_event_set_padding(struct ring_buffer_event *event)
155 {
156 /* padding has a NULL time_delta */
157 event->type_len = RINGBUF_TYPE_PADDING;
158 event->time_delta = 0;
159 }
160
161 static unsigned
rb_event_data_length(struct ring_buffer_event * event)162 rb_event_data_length(struct ring_buffer_event *event)
163 {
164 unsigned length;
165
166 if (event->type_len)
167 length = event->type_len * RB_ALIGNMENT;
168 else
169 length = event->array[0];
170 return length + RB_EVNT_HDR_SIZE;
171 }
172
173 /*
174 * Return the length of the given event. Will return
175 * the length of the time extend if the event is a
176 * time extend.
177 */
178 static inline unsigned
rb_event_length(struct ring_buffer_event * event)179 rb_event_length(struct ring_buffer_event *event)
180 {
181 switch (event->type_len) {
182 case RINGBUF_TYPE_PADDING:
183 if (rb_null_event(event))
184 /* undefined */
185 return -1;
186 return event->array[0] + RB_EVNT_HDR_SIZE;
187
188 case RINGBUF_TYPE_TIME_EXTEND:
189 return RB_LEN_TIME_EXTEND;
190
191 case RINGBUF_TYPE_TIME_STAMP:
192 return RB_LEN_TIME_STAMP;
193
194 case RINGBUF_TYPE_DATA:
195 return rb_event_data_length(event);
196 default:
197 BUG();
198 }
199 /* not hit */
200 return 0;
201 }
202
203 /*
204 * Return total length of time extend and data,
205 * or just the event length for all other events.
206 */
207 static inline unsigned
rb_event_ts_length(struct ring_buffer_event * event)208 rb_event_ts_length(struct ring_buffer_event *event)
209 {
210 unsigned len = 0;
211
212 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
213 /* time extends include the data event after it */
214 len = RB_LEN_TIME_EXTEND;
215 event = skip_time_extend(event);
216 }
217 return len + rb_event_length(event);
218 }
219
220 /**
221 * ring_buffer_event_length - return the length of the event
222 * @event: the event to get the length of
223 *
224 * Returns the size of the data load of a data event.
225 * If the event is something other than a data event, it
226 * returns the size of the event itself. With the exception
227 * of a TIME EXTEND, where it still returns the size of the
228 * data load of the data event after it.
229 */
ring_buffer_event_length(struct ring_buffer_event * event)230 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
231 {
232 unsigned length;
233
234 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
235 event = skip_time_extend(event);
236
237 length = rb_event_length(event);
238 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
239 return length;
240 length -= RB_EVNT_HDR_SIZE;
241 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
242 length -= sizeof(event->array[0]);
243 return length;
244 }
245 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
246
247 /* inline for ring buffer fast paths */
248 static void *
rb_event_data(struct ring_buffer_event * event)249 rb_event_data(struct ring_buffer_event *event)
250 {
251 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
252 event = skip_time_extend(event);
253 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
254 /* If length is in len field, then array[0] has the data */
255 if (event->type_len)
256 return (void *)&event->array[0];
257 /* Otherwise length is in array[0] and array[1] has the data */
258 return (void *)&event->array[1];
259 }
260
261 /**
262 * ring_buffer_event_data - return the data of the event
263 * @event: the event to get the data from
264 */
ring_buffer_event_data(struct ring_buffer_event * event)265 void *ring_buffer_event_data(struct ring_buffer_event *event)
266 {
267 return rb_event_data(event);
268 }
269 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
270
271 #define for_each_buffer_cpu(buffer, cpu) \
272 for_each_cpu(cpu, buffer->cpumask)
273
274 #define TS_SHIFT 27
275 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
276 #define TS_DELTA_TEST (~TS_MASK)
277
278 /* Flag when events were overwritten */
279 #define RB_MISSED_EVENTS (1 << 31)
280 /* Missed count stored at end */
281 #define RB_MISSED_STORED (1 << 30)
282
283 struct buffer_data_page {
284 u64 time_stamp; /* page time stamp */
285 local_t commit; /* write committed index */
286 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
287 };
288
289 /*
290 * Note, the buffer_page list must be first. The buffer pages
291 * are allocated in cache lines, which means that each buffer
292 * page will be at the beginning of a cache line, and thus
293 * the least significant bits will be zero. We use this to
294 * add flags in the list struct pointers, to make the ring buffer
295 * lockless.
296 */
297 struct buffer_page {
298 struct list_head list; /* list of buffer pages */
299 local_t write; /* index for next write */
300 unsigned read; /* index for next read */
301 local_t entries; /* entries on this page */
302 unsigned long real_end; /* real end of data */
303 struct buffer_data_page *page; /* Actual data page */
304 };
305
306 /*
307 * The buffer page counters, write and entries, must be reset
308 * atomically when crossing page boundaries. To synchronize this
309 * update, two counters are inserted into the number. One is
310 * the actual counter for the write position or count on the page.
311 *
312 * The other is a counter of updaters. Before an update happens
313 * the update partition of the counter is incremented. This will
314 * allow the updater to update the counter atomically.
315 *
316 * The counter is 20 bits, and the state data is 12.
317 */
318 #define RB_WRITE_MASK 0xfffff
319 #define RB_WRITE_INTCNT (1 << 20)
320
rb_init_page(struct buffer_data_page * bpage)321 static void rb_init_page(struct buffer_data_page *bpage)
322 {
323 local_set(&bpage->commit, 0);
324 }
325
326 /**
327 * ring_buffer_page_len - the size of data on the page.
328 * @page: The page to read
329 *
330 * Returns the amount of data on the page, including buffer page header.
331 */
ring_buffer_page_len(void * page)332 size_t ring_buffer_page_len(void *page)
333 {
334 return local_read(&((struct buffer_data_page *)page)->commit)
335 + BUF_PAGE_HDR_SIZE;
336 }
337
338 /*
339 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
340 * this issue out.
341 */
free_buffer_page(struct buffer_page * bpage)342 static void free_buffer_page(struct buffer_page *bpage)
343 {
344 free_page((unsigned long)bpage->page);
345 kfree(bpage);
346 }
347
348 /*
349 * We need to fit the time_stamp delta into 27 bits.
350 */
test_time_stamp(u64 delta)351 static inline int test_time_stamp(u64 delta)
352 {
353 if (delta & TS_DELTA_TEST)
354 return 1;
355 return 0;
356 }
357
358 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
359
360 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
361 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
362
ring_buffer_print_page_header(struct trace_seq * s)363 int ring_buffer_print_page_header(struct trace_seq *s)
364 {
365 struct buffer_data_page field;
366
367 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
368 "offset:0;\tsize:%u;\tsigned:%u;\n",
369 (unsigned int)sizeof(field.time_stamp),
370 (unsigned int)is_signed_type(u64));
371
372 trace_seq_printf(s, "\tfield: local_t commit;\t"
373 "offset:%u;\tsize:%u;\tsigned:%u;\n",
374 (unsigned int)offsetof(typeof(field), commit),
375 (unsigned int)sizeof(field.commit),
376 (unsigned int)is_signed_type(long));
377
378 trace_seq_printf(s, "\tfield: int overwrite;\t"
379 "offset:%u;\tsize:%u;\tsigned:%u;\n",
380 (unsigned int)offsetof(typeof(field), commit),
381 1,
382 (unsigned int)is_signed_type(long));
383
384 trace_seq_printf(s, "\tfield: char data;\t"
385 "offset:%u;\tsize:%u;\tsigned:%u;\n",
386 (unsigned int)offsetof(typeof(field), data),
387 (unsigned int)BUF_PAGE_SIZE,
388 (unsigned int)is_signed_type(char));
389
390 return !trace_seq_has_overflowed(s);
391 }
392
393 struct rb_irq_work {
394 struct irq_work work;
395 wait_queue_head_t waiters;
396 wait_queue_head_t full_waiters;
397 bool waiters_pending;
398 bool full_waiters_pending;
399 bool wakeup_full;
400 };
401
402 /*
403 * Structure to hold event state and handle nested events.
404 */
405 struct rb_event_info {
406 u64 ts;
407 u64 delta;
408 unsigned long length;
409 struct buffer_page *tail_page;
410 int add_timestamp;
411 };
412
413 /*
414 * Used for which event context the event is in.
415 * NMI = 0
416 * IRQ = 1
417 * SOFTIRQ = 2
418 * NORMAL = 3
419 *
420 * See trace_recursive_lock() comment below for more details.
421 */
422 enum {
423 RB_CTX_NMI,
424 RB_CTX_IRQ,
425 RB_CTX_SOFTIRQ,
426 RB_CTX_NORMAL,
427 RB_CTX_MAX
428 };
429
430 /*
431 * head_page == tail_page && head == tail then buffer is empty.
432 */
433 struct ring_buffer_per_cpu {
434 int cpu;
435 atomic_t record_disabled;
436 struct ring_buffer *buffer;
437 raw_spinlock_t reader_lock; /* serialize readers */
438 arch_spinlock_t lock;
439 struct lock_class_key lock_key;
440 unsigned long nr_pages;
441 unsigned int current_context;
442 struct list_head *pages;
443 struct buffer_page *head_page; /* read from head */
444 struct buffer_page *tail_page; /* write to tail */
445 struct buffer_page *commit_page; /* committed pages */
446 struct buffer_page *reader_page;
447 unsigned long lost_events;
448 unsigned long last_overrun;
449 local_t entries_bytes;
450 local_t entries;
451 local_t overrun;
452 local_t commit_overrun;
453 local_t dropped_events;
454 local_t committing;
455 local_t commits;
456 unsigned long read;
457 unsigned long read_bytes;
458 u64 write_stamp;
459 u64 read_stamp;
460 /* ring buffer pages to update, > 0 to add, < 0 to remove */
461 long nr_pages_to_update;
462 struct list_head new_pages; /* new pages to add */
463 struct work_struct update_pages_work;
464 struct completion update_done;
465
466 struct rb_irq_work irq_work;
467 };
468
469 struct ring_buffer {
470 unsigned flags;
471 int cpus;
472 atomic_t record_disabled;
473 atomic_t resize_disabled;
474 cpumask_var_t cpumask;
475
476 struct lock_class_key *reader_lock_key;
477
478 struct mutex mutex;
479
480 struct ring_buffer_per_cpu **buffers;
481
482 #ifdef CONFIG_HOTPLUG_CPU
483 struct notifier_block cpu_notify;
484 #endif
485 u64 (*clock)(void);
486
487 struct rb_irq_work irq_work;
488 };
489
490 struct ring_buffer_iter {
491 struct ring_buffer_per_cpu *cpu_buffer;
492 unsigned long head;
493 struct buffer_page *head_page;
494 struct buffer_page *cache_reader_page;
495 unsigned long cache_read;
496 u64 read_stamp;
497 };
498
499 /*
500 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
501 *
502 * Schedules a delayed work to wake up any task that is blocked on the
503 * ring buffer waiters queue.
504 */
rb_wake_up_waiters(struct irq_work * work)505 static void rb_wake_up_waiters(struct irq_work *work)
506 {
507 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
508
509 wake_up_all(&rbwork->waiters);
510 if (rbwork->wakeup_full) {
511 rbwork->wakeup_full = false;
512 wake_up_all(&rbwork->full_waiters);
513 }
514 }
515
516 /**
517 * ring_buffer_wait - wait for input to the ring buffer
518 * @buffer: buffer to wait on
519 * @cpu: the cpu buffer to wait on
520 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
521 *
522 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
523 * as data is added to any of the @buffer's cpu buffers. Otherwise
524 * it will wait for data to be added to a specific cpu buffer.
525 */
ring_buffer_wait(struct ring_buffer * buffer,int cpu,bool full)526 int ring_buffer_wait(struct ring_buffer *buffer, int cpu, bool full)
527 {
528 struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
529 DEFINE_WAIT(wait);
530 struct rb_irq_work *work;
531 int ret = 0;
532
533 /*
534 * Depending on what the caller is waiting for, either any
535 * data in any cpu buffer, or a specific buffer, put the
536 * caller on the appropriate wait queue.
537 */
538 if (cpu == RING_BUFFER_ALL_CPUS) {
539 work = &buffer->irq_work;
540 /* Full only makes sense on per cpu reads */
541 full = false;
542 } else {
543 if (!cpumask_test_cpu(cpu, buffer->cpumask))
544 return -ENODEV;
545 cpu_buffer = buffer->buffers[cpu];
546 work = &cpu_buffer->irq_work;
547 }
548
549
550 while (true) {
551 if (full)
552 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
553 else
554 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
555
556 /*
557 * The events can happen in critical sections where
558 * checking a work queue can cause deadlocks.
559 * After adding a task to the queue, this flag is set
560 * only to notify events to try to wake up the queue
561 * using irq_work.
562 *
563 * We don't clear it even if the buffer is no longer
564 * empty. The flag only causes the next event to run
565 * irq_work to do the work queue wake up. The worse
566 * that can happen if we race with !trace_empty() is that
567 * an event will cause an irq_work to try to wake up
568 * an empty queue.
569 *
570 * There's no reason to protect this flag either, as
571 * the work queue and irq_work logic will do the necessary
572 * synchronization for the wake ups. The only thing
573 * that is necessary is that the wake up happens after
574 * a task has been queued. It's OK for spurious wake ups.
575 */
576 if (full)
577 work->full_waiters_pending = true;
578 else
579 work->waiters_pending = true;
580
581 if (signal_pending(current)) {
582 ret = -EINTR;
583 break;
584 }
585
586 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
587 break;
588
589 if (cpu != RING_BUFFER_ALL_CPUS &&
590 !ring_buffer_empty_cpu(buffer, cpu)) {
591 unsigned long flags;
592 bool pagebusy;
593
594 if (!full)
595 break;
596
597 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
598 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
599 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
600
601 if (!pagebusy)
602 break;
603 }
604
605 schedule();
606 }
607
608 if (full)
609 finish_wait(&work->full_waiters, &wait);
610 else
611 finish_wait(&work->waiters, &wait);
612
613 return ret;
614 }
615
616 /**
617 * ring_buffer_poll_wait - poll on buffer input
618 * @buffer: buffer to wait on
619 * @cpu: the cpu buffer to wait on
620 * @filp: the file descriptor
621 * @poll_table: The poll descriptor
622 *
623 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
624 * as data is added to any of the @buffer's cpu buffers. Otherwise
625 * it will wait for data to be added to a specific cpu buffer.
626 *
627 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
628 * zero otherwise.
629 */
ring_buffer_poll_wait(struct ring_buffer * buffer,int cpu,struct file * filp,poll_table * poll_table)630 int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
631 struct file *filp, poll_table *poll_table)
632 {
633 struct ring_buffer_per_cpu *cpu_buffer;
634 struct rb_irq_work *work;
635
636 if (cpu == RING_BUFFER_ALL_CPUS)
637 work = &buffer->irq_work;
638 else {
639 if (!cpumask_test_cpu(cpu, buffer->cpumask))
640 return -EINVAL;
641
642 cpu_buffer = buffer->buffers[cpu];
643 work = &cpu_buffer->irq_work;
644 }
645
646 poll_wait(filp, &work->waiters, poll_table);
647 work->waiters_pending = true;
648 /*
649 * There's a tight race between setting the waiters_pending and
650 * checking if the ring buffer is empty. Once the waiters_pending bit
651 * is set, the next event will wake the task up, but we can get stuck
652 * if there's only a single event in.
653 *
654 * FIXME: Ideally, we need a memory barrier on the writer side as well,
655 * but adding a memory barrier to all events will cause too much of a
656 * performance hit in the fast path. We only need a memory barrier when
657 * the buffer goes from empty to having content. But as this race is
658 * extremely small, and it's not a problem if another event comes in, we
659 * will fix it later.
660 */
661 smp_mb();
662
663 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
664 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
665 return POLLIN | POLLRDNORM;
666 return 0;
667 }
668
669 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
670 #define RB_WARN_ON(b, cond) \
671 ({ \
672 int _____ret = unlikely(cond); \
673 if (_____ret) { \
674 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
675 struct ring_buffer_per_cpu *__b = \
676 (void *)b; \
677 atomic_inc(&__b->buffer->record_disabled); \
678 } else \
679 atomic_inc(&b->record_disabled); \
680 WARN_ON(1); \
681 } \
682 _____ret; \
683 })
684
685 /* Up this if you want to test the TIME_EXTENTS and normalization */
686 #define DEBUG_SHIFT 0
687
rb_time_stamp(struct ring_buffer * buffer)688 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
689 {
690 /* shift to debug/test normalization and TIME_EXTENTS */
691 return buffer->clock() << DEBUG_SHIFT;
692 }
693
ring_buffer_time_stamp(struct ring_buffer * buffer,int cpu)694 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
695 {
696 u64 time;
697
698 preempt_disable_notrace();
699 time = rb_time_stamp(buffer);
700 preempt_enable_no_resched_notrace();
701
702 return time;
703 }
704 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
705
ring_buffer_normalize_time_stamp(struct ring_buffer * buffer,int cpu,u64 * ts)706 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
707 int cpu, u64 *ts)
708 {
709 /* Just stupid testing the normalize function and deltas */
710 *ts >>= DEBUG_SHIFT;
711 }
712 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
713
714 /*
715 * Making the ring buffer lockless makes things tricky.
716 * Although writes only happen on the CPU that they are on,
717 * and they only need to worry about interrupts. Reads can
718 * happen on any CPU.
719 *
720 * The reader page is always off the ring buffer, but when the
721 * reader finishes with a page, it needs to swap its page with
722 * a new one from the buffer. The reader needs to take from
723 * the head (writes go to the tail). But if a writer is in overwrite
724 * mode and wraps, it must push the head page forward.
725 *
726 * Here lies the problem.
727 *
728 * The reader must be careful to replace only the head page, and
729 * not another one. As described at the top of the file in the
730 * ASCII art, the reader sets its old page to point to the next
731 * page after head. It then sets the page after head to point to
732 * the old reader page. But if the writer moves the head page
733 * during this operation, the reader could end up with the tail.
734 *
735 * We use cmpxchg to help prevent this race. We also do something
736 * special with the page before head. We set the LSB to 1.
737 *
738 * When the writer must push the page forward, it will clear the
739 * bit that points to the head page, move the head, and then set
740 * the bit that points to the new head page.
741 *
742 * We also don't want an interrupt coming in and moving the head
743 * page on another writer. Thus we use the second LSB to catch
744 * that too. Thus:
745 *
746 * head->list->prev->next bit 1 bit 0
747 * ------- -------
748 * Normal page 0 0
749 * Points to head page 0 1
750 * New head page 1 0
751 *
752 * Note we can not trust the prev pointer of the head page, because:
753 *
754 * +----+ +-----+ +-----+
755 * | |------>| T |---X--->| N |
756 * | |<------| | | |
757 * +----+ +-----+ +-----+
758 * ^ ^ |
759 * | +-----+ | |
760 * +----------| R |----------+ |
761 * | |<-----------+
762 * +-----+
763 *
764 * Key: ---X--> HEAD flag set in pointer
765 * T Tail page
766 * R Reader page
767 * N Next page
768 *
769 * (see __rb_reserve_next() to see where this happens)
770 *
771 * What the above shows is that the reader just swapped out
772 * the reader page with a page in the buffer, but before it
773 * could make the new header point back to the new page added
774 * it was preempted by a writer. The writer moved forward onto
775 * the new page added by the reader and is about to move forward
776 * again.
777 *
778 * You can see, it is legitimate for the previous pointer of
779 * the head (or any page) not to point back to itself. But only
780 * temporarially.
781 */
782
783 #define RB_PAGE_NORMAL 0UL
784 #define RB_PAGE_HEAD 1UL
785 #define RB_PAGE_UPDATE 2UL
786
787
788 #define RB_FLAG_MASK 3UL
789
790 /* PAGE_MOVED is not part of the mask */
791 #define RB_PAGE_MOVED 4UL
792
793 /*
794 * rb_list_head - remove any bit
795 */
rb_list_head(struct list_head * list)796 static struct list_head *rb_list_head(struct list_head *list)
797 {
798 unsigned long val = (unsigned long)list;
799
800 return (struct list_head *)(val & ~RB_FLAG_MASK);
801 }
802
803 /*
804 * rb_is_head_page - test if the given page is the head page
805 *
806 * Because the reader may move the head_page pointer, we can
807 * not trust what the head page is (it may be pointing to
808 * the reader page). But if the next page is a header page,
809 * its flags will be non zero.
810 */
811 static inline int
rb_is_head_page(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * page,struct list_head * list)812 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
813 struct buffer_page *page, struct list_head *list)
814 {
815 unsigned long val;
816
817 val = (unsigned long)list->next;
818
819 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
820 return RB_PAGE_MOVED;
821
822 return val & RB_FLAG_MASK;
823 }
824
825 /*
826 * rb_is_reader_page
827 *
828 * The unique thing about the reader page, is that, if the
829 * writer is ever on it, the previous pointer never points
830 * back to the reader page.
831 */
rb_is_reader_page(struct buffer_page * page)832 static bool rb_is_reader_page(struct buffer_page *page)
833 {
834 struct list_head *list = page->list.prev;
835
836 return rb_list_head(list->next) != &page->list;
837 }
838
839 /*
840 * rb_set_list_to_head - set a list_head to be pointing to head.
841 */
rb_set_list_to_head(struct ring_buffer_per_cpu * cpu_buffer,struct list_head * list)842 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
843 struct list_head *list)
844 {
845 unsigned long *ptr;
846
847 ptr = (unsigned long *)&list->next;
848 *ptr |= RB_PAGE_HEAD;
849 *ptr &= ~RB_PAGE_UPDATE;
850 }
851
852 /*
853 * rb_head_page_activate - sets up head page
854 */
rb_head_page_activate(struct ring_buffer_per_cpu * cpu_buffer)855 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
856 {
857 struct buffer_page *head;
858
859 head = cpu_buffer->head_page;
860 if (!head)
861 return;
862
863 /*
864 * Set the previous list pointer to have the HEAD flag.
865 */
866 rb_set_list_to_head(cpu_buffer, head->list.prev);
867 }
868
rb_list_head_clear(struct list_head * list)869 static void rb_list_head_clear(struct list_head *list)
870 {
871 unsigned long *ptr = (unsigned long *)&list->next;
872
873 *ptr &= ~RB_FLAG_MASK;
874 }
875
876 /*
877 * rb_head_page_dactivate - clears head page ptr (for free list)
878 */
879 static void
rb_head_page_deactivate(struct ring_buffer_per_cpu * cpu_buffer)880 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
881 {
882 struct list_head *hd;
883
884 /* Go through the whole list and clear any pointers found. */
885 rb_list_head_clear(cpu_buffer->pages);
886
887 list_for_each(hd, cpu_buffer->pages)
888 rb_list_head_clear(hd);
889 }
890
rb_head_page_set(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * head,struct buffer_page * prev,int old_flag,int new_flag)891 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
892 struct buffer_page *head,
893 struct buffer_page *prev,
894 int old_flag, int new_flag)
895 {
896 struct list_head *list;
897 unsigned long val = (unsigned long)&head->list;
898 unsigned long ret;
899
900 list = &prev->list;
901
902 val &= ~RB_FLAG_MASK;
903
904 ret = cmpxchg((unsigned long *)&list->next,
905 val | old_flag, val | new_flag);
906
907 /* check if the reader took the page */
908 if ((ret & ~RB_FLAG_MASK) != val)
909 return RB_PAGE_MOVED;
910
911 return ret & RB_FLAG_MASK;
912 }
913
rb_head_page_set_update(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * head,struct buffer_page * prev,int old_flag)914 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
915 struct buffer_page *head,
916 struct buffer_page *prev,
917 int old_flag)
918 {
919 return rb_head_page_set(cpu_buffer, head, prev,
920 old_flag, RB_PAGE_UPDATE);
921 }
922
rb_head_page_set_head(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * head,struct buffer_page * prev,int old_flag)923 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
924 struct buffer_page *head,
925 struct buffer_page *prev,
926 int old_flag)
927 {
928 return rb_head_page_set(cpu_buffer, head, prev,
929 old_flag, RB_PAGE_HEAD);
930 }
931
rb_head_page_set_normal(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * head,struct buffer_page * prev,int old_flag)932 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
933 struct buffer_page *head,
934 struct buffer_page *prev,
935 int old_flag)
936 {
937 return rb_head_page_set(cpu_buffer, head, prev,
938 old_flag, RB_PAGE_NORMAL);
939 }
940
rb_inc_page(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page ** bpage)941 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
942 struct buffer_page **bpage)
943 {
944 struct list_head *p = rb_list_head((*bpage)->list.next);
945
946 *bpage = list_entry(p, struct buffer_page, list);
947 }
948
949 static struct buffer_page *
rb_set_head_page(struct ring_buffer_per_cpu * cpu_buffer)950 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
951 {
952 struct buffer_page *head;
953 struct buffer_page *page;
954 struct list_head *list;
955 int i;
956
957 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
958 return NULL;
959
960 /* sanity check */
961 list = cpu_buffer->pages;
962 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
963 return NULL;
964
965 page = head = cpu_buffer->head_page;
966 /*
967 * It is possible that the writer moves the header behind
968 * where we started, and we miss in one loop.
969 * A second loop should grab the header, but we'll do
970 * three loops just because I'm paranoid.
971 */
972 for (i = 0; i < 3; i++) {
973 do {
974 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
975 cpu_buffer->head_page = page;
976 return page;
977 }
978 rb_inc_page(cpu_buffer, &page);
979 } while (page != head);
980 }
981
982 RB_WARN_ON(cpu_buffer, 1);
983
984 return NULL;
985 }
986
rb_head_page_replace(struct buffer_page * old,struct buffer_page * new)987 static int rb_head_page_replace(struct buffer_page *old,
988 struct buffer_page *new)
989 {
990 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
991 unsigned long val;
992 unsigned long ret;
993
994 val = *ptr & ~RB_FLAG_MASK;
995 val |= RB_PAGE_HEAD;
996
997 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
998
999 return ret == val;
1000 }
1001
1002 /*
1003 * rb_tail_page_update - move the tail page forward
1004 *
1005 * Returns 1 if moved tail page, 0 if someone else did.
1006 */
rb_tail_page_update(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * tail_page,struct buffer_page * next_page)1007 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1008 struct buffer_page *tail_page,
1009 struct buffer_page *next_page)
1010 {
1011 struct buffer_page *old_tail;
1012 unsigned long old_entries;
1013 unsigned long old_write;
1014 int ret = 0;
1015
1016 /*
1017 * The tail page now needs to be moved forward.
1018 *
1019 * We need to reset the tail page, but without messing
1020 * with possible erasing of data brought in by interrupts
1021 * that have moved the tail page and are currently on it.
1022 *
1023 * We add a counter to the write field to denote this.
1024 */
1025 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1026 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1027
1028 /*
1029 * Just make sure we have seen our old_write and synchronize
1030 * with any interrupts that come in.
1031 */
1032 barrier();
1033
1034 /*
1035 * If the tail page is still the same as what we think
1036 * it is, then it is up to us to update the tail
1037 * pointer.
1038 */
1039 if (tail_page == cpu_buffer->tail_page) {
1040 /* Zero the write counter */
1041 unsigned long val = old_write & ~RB_WRITE_MASK;
1042 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1043
1044 /*
1045 * This will only succeed if an interrupt did
1046 * not come in and change it. In which case, we
1047 * do not want to modify it.
1048 *
1049 * We add (void) to let the compiler know that we do not care
1050 * about the return value of these functions. We use the
1051 * cmpxchg to only update if an interrupt did not already
1052 * do it for us. If the cmpxchg fails, we don't care.
1053 */
1054 (void)local_cmpxchg(&next_page->write, old_write, val);
1055 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1056
1057 /*
1058 * No need to worry about races with clearing out the commit.
1059 * it only can increment when a commit takes place. But that
1060 * only happens in the outer most nested commit.
1061 */
1062 local_set(&next_page->page->commit, 0);
1063
1064 old_tail = cmpxchg(&cpu_buffer->tail_page,
1065 tail_page, next_page);
1066
1067 if (old_tail == tail_page)
1068 ret = 1;
1069 }
1070
1071 return ret;
1072 }
1073
rb_check_bpage(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * bpage)1074 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1075 struct buffer_page *bpage)
1076 {
1077 unsigned long val = (unsigned long)bpage;
1078
1079 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1080 return 1;
1081
1082 return 0;
1083 }
1084
1085 /**
1086 * rb_check_list - make sure a pointer to a list has the last bits zero
1087 */
rb_check_list(struct ring_buffer_per_cpu * cpu_buffer,struct list_head * list)1088 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1089 struct list_head *list)
1090 {
1091 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1092 return 1;
1093 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1094 return 1;
1095 return 0;
1096 }
1097
1098 /**
1099 * rb_check_pages - integrity check of buffer pages
1100 * @cpu_buffer: CPU buffer with pages to test
1101 *
1102 * As a safety measure we check to make sure the data pages have not
1103 * been corrupted.
1104 */
rb_check_pages(struct ring_buffer_per_cpu * cpu_buffer)1105 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1106 {
1107 struct list_head *head = cpu_buffer->pages;
1108 struct buffer_page *bpage, *tmp;
1109
1110 /* Reset the head page if it exists */
1111 if (cpu_buffer->head_page)
1112 rb_set_head_page(cpu_buffer);
1113
1114 rb_head_page_deactivate(cpu_buffer);
1115
1116 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1117 return -1;
1118 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1119 return -1;
1120
1121 if (rb_check_list(cpu_buffer, head))
1122 return -1;
1123
1124 list_for_each_entry_safe(bpage, tmp, head, list) {
1125 if (RB_WARN_ON(cpu_buffer,
1126 bpage->list.next->prev != &bpage->list))
1127 return -1;
1128 if (RB_WARN_ON(cpu_buffer,
1129 bpage->list.prev->next != &bpage->list))
1130 return -1;
1131 if (rb_check_list(cpu_buffer, &bpage->list))
1132 return -1;
1133 }
1134
1135 rb_head_page_activate(cpu_buffer);
1136
1137 return 0;
1138 }
1139
__rb_allocate_pages(long nr_pages,struct list_head * pages,int cpu)1140 static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1141 {
1142 struct buffer_page *bpage, *tmp;
1143 long i;
1144
1145 for (i = 0; i < nr_pages; i++) {
1146 struct page *page;
1147 /*
1148 * __GFP_NORETRY flag makes sure that the allocation fails
1149 * gracefully without invoking oom-killer and the system is
1150 * not destabilized.
1151 */
1152 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1153 GFP_KERNEL | __GFP_NORETRY,
1154 cpu_to_node(cpu));
1155 if (!bpage)
1156 goto free_pages;
1157
1158 list_add(&bpage->list, pages);
1159
1160 page = alloc_pages_node(cpu_to_node(cpu),
1161 GFP_KERNEL | __GFP_NORETRY, 0);
1162 if (!page)
1163 goto free_pages;
1164 bpage->page = page_address(page);
1165 rb_init_page(bpage->page);
1166 }
1167
1168 return 0;
1169
1170 free_pages:
1171 list_for_each_entry_safe(bpage, tmp, pages, list) {
1172 list_del_init(&bpage->list);
1173 free_buffer_page(bpage);
1174 }
1175
1176 return -ENOMEM;
1177 }
1178
rb_allocate_pages(struct ring_buffer_per_cpu * cpu_buffer,unsigned long nr_pages)1179 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1180 unsigned long nr_pages)
1181 {
1182 LIST_HEAD(pages);
1183
1184 WARN_ON(!nr_pages);
1185
1186 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1187 return -ENOMEM;
1188
1189 /*
1190 * The ring buffer page list is a circular list that does not
1191 * start and end with a list head. All page list items point to
1192 * other pages.
1193 */
1194 cpu_buffer->pages = pages.next;
1195 list_del(&pages);
1196
1197 cpu_buffer->nr_pages = nr_pages;
1198
1199 rb_check_pages(cpu_buffer);
1200
1201 return 0;
1202 }
1203
1204 static struct ring_buffer_per_cpu *
rb_allocate_cpu_buffer(struct ring_buffer * buffer,long nr_pages,int cpu)1205 rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1206 {
1207 struct ring_buffer_per_cpu *cpu_buffer;
1208 struct buffer_page *bpage;
1209 struct page *page;
1210 int ret;
1211
1212 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1213 GFP_KERNEL, cpu_to_node(cpu));
1214 if (!cpu_buffer)
1215 return NULL;
1216
1217 cpu_buffer->cpu = cpu;
1218 cpu_buffer->buffer = buffer;
1219 raw_spin_lock_init(&cpu_buffer->reader_lock);
1220 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1221 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1222 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1223 init_completion(&cpu_buffer->update_done);
1224 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1225 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1226 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1227
1228 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1229 GFP_KERNEL, cpu_to_node(cpu));
1230 if (!bpage)
1231 goto fail_free_buffer;
1232
1233 rb_check_bpage(cpu_buffer, bpage);
1234
1235 cpu_buffer->reader_page = bpage;
1236 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1237 if (!page)
1238 goto fail_free_reader;
1239 bpage->page = page_address(page);
1240 rb_init_page(bpage->page);
1241
1242 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1243 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1244
1245 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1246 if (ret < 0)
1247 goto fail_free_reader;
1248
1249 cpu_buffer->head_page
1250 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1251 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1252
1253 rb_head_page_activate(cpu_buffer);
1254
1255 return cpu_buffer;
1256
1257 fail_free_reader:
1258 free_buffer_page(cpu_buffer->reader_page);
1259
1260 fail_free_buffer:
1261 kfree(cpu_buffer);
1262 return NULL;
1263 }
1264
rb_free_cpu_buffer(struct ring_buffer_per_cpu * cpu_buffer)1265 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1266 {
1267 struct list_head *head = cpu_buffer->pages;
1268 struct buffer_page *bpage, *tmp;
1269
1270 free_buffer_page(cpu_buffer->reader_page);
1271
1272 rb_head_page_deactivate(cpu_buffer);
1273
1274 if (head) {
1275 list_for_each_entry_safe(bpage, tmp, head, list) {
1276 list_del_init(&bpage->list);
1277 free_buffer_page(bpage);
1278 }
1279 bpage = list_entry(head, struct buffer_page, list);
1280 free_buffer_page(bpage);
1281 }
1282
1283 kfree(cpu_buffer);
1284 }
1285
1286 #ifdef CONFIG_HOTPLUG_CPU
1287 static int rb_cpu_notify(struct notifier_block *self,
1288 unsigned long action, void *hcpu);
1289 #endif
1290
1291 /**
1292 * __ring_buffer_alloc - allocate a new ring_buffer
1293 * @size: the size in bytes per cpu that is needed.
1294 * @flags: attributes to set for the ring buffer.
1295 *
1296 * Currently the only flag that is available is the RB_FL_OVERWRITE
1297 * flag. This flag means that the buffer will overwrite old data
1298 * when the buffer wraps. If this flag is not set, the buffer will
1299 * drop data when the tail hits the head.
1300 */
__ring_buffer_alloc(unsigned long size,unsigned flags,struct lock_class_key * key)1301 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1302 struct lock_class_key *key)
1303 {
1304 struct ring_buffer *buffer;
1305 long nr_pages;
1306 int bsize;
1307 int cpu;
1308
1309 /* keep it in its own cache line */
1310 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1311 GFP_KERNEL);
1312 if (!buffer)
1313 return NULL;
1314
1315 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1316 goto fail_free_buffer;
1317
1318 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1319 buffer->flags = flags;
1320 buffer->clock = trace_clock_local;
1321 buffer->reader_lock_key = key;
1322
1323 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1324 init_waitqueue_head(&buffer->irq_work.waiters);
1325
1326 /* need at least two pages */
1327 if (nr_pages < 2)
1328 nr_pages = 2;
1329
1330 /*
1331 * In case of non-hotplug cpu, if the ring-buffer is allocated
1332 * in early initcall, it will not be notified of secondary cpus.
1333 * In that off case, we need to allocate for all possible cpus.
1334 */
1335 #ifdef CONFIG_HOTPLUG_CPU
1336 cpu_notifier_register_begin();
1337 cpumask_copy(buffer->cpumask, cpu_online_mask);
1338 #else
1339 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1340 #endif
1341 buffer->cpus = nr_cpu_ids;
1342
1343 bsize = sizeof(void *) * nr_cpu_ids;
1344 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1345 GFP_KERNEL);
1346 if (!buffer->buffers)
1347 goto fail_free_cpumask;
1348
1349 for_each_buffer_cpu(buffer, cpu) {
1350 buffer->buffers[cpu] =
1351 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1352 if (!buffer->buffers[cpu])
1353 goto fail_free_buffers;
1354 }
1355
1356 #ifdef CONFIG_HOTPLUG_CPU
1357 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1358 buffer->cpu_notify.priority = 0;
1359 __register_cpu_notifier(&buffer->cpu_notify);
1360 cpu_notifier_register_done();
1361 #endif
1362
1363 mutex_init(&buffer->mutex);
1364
1365 return buffer;
1366
1367 fail_free_buffers:
1368 for_each_buffer_cpu(buffer, cpu) {
1369 if (buffer->buffers[cpu])
1370 rb_free_cpu_buffer(buffer->buffers[cpu]);
1371 }
1372 kfree(buffer->buffers);
1373
1374 fail_free_cpumask:
1375 free_cpumask_var(buffer->cpumask);
1376 #ifdef CONFIG_HOTPLUG_CPU
1377 cpu_notifier_register_done();
1378 #endif
1379
1380 fail_free_buffer:
1381 kfree(buffer);
1382 return NULL;
1383 }
1384 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1385
1386 /**
1387 * ring_buffer_free - free a ring buffer.
1388 * @buffer: the buffer to free.
1389 */
1390 void
ring_buffer_free(struct ring_buffer * buffer)1391 ring_buffer_free(struct ring_buffer *buffer)
1392 {
1393 int cpu;
1394
1395 #ifdef CONFIG_HOTPLUG_CPU
1396 cpu_notifier_register_begin();
1397 __unregister_cpu_notifier(&buffer->cpu_notify);
1398 #endif
1399
1400 for_each_buffer_cpu(buffer, cpu)
1401 rb_free_cpu_buffer(buffer->buffers[cpu]);
1402
1403 #ifdef CONFIG_HOTPLUG_CPU
1404 cpu_notifier_register_done();
1405 #endif
1406
1407 kfree(buffer->buffers);
1408 free_cpumask_var(buffer->cpumask);
1409
1410 kfree(buffer);
1411 }
1412 EXPORT_SYMBOL_GPL(ring_buffer_free);
1413
ring_buffer_set_clock(struct ring_buffer * buffer,u64 (* clock)(void))1414 void ring_buffer_set_clock(struct ring_buffer *buffer,
1415 u64 (*clock)(void))
1416 {
1417 buffer->clock = clock;
1418 }
1419
1420 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1421
rb_page_entries(struct buffer_page * bpage)1422 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1423 {
1424 return local_read(&bpage->entries) & RB_WRITE_MASK;
1425 }
1426
rb_page_write(struct buffer_page * bpage)1427 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1428 {
1429 return local_read(&bpage->write) & RB_WRITE_MASK;
1430 }
1431
1432 static int
rb_remove_pages(struct ring_buffer_per_cpu * cpu_buffer,unsigned long nr_pages)1433 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1434 {
1435 struct list_head *tail_page, *to_remove, *next_page;
1436 struct buffer_page *to_remove_page, *tmp_iter_page;
1437 struct buffer_page *last_page, *first_page;
1438 unsigned long nr_removed;
1439 unsigned long head_bit;
1440 int page_entries;
1441
1442 head_bit = 0;
1443
1444 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1445 atomic_inc(&cpu_buffer->record_disabled);
1446 /*
1447 * We don't race with the readers since we have acquired the reader
1448 * lock. We also don't race with writers after disabling recording.
1449 * This makes it easy to figure out the first and the last page to be
1450 * removed from the list. We unlink all the pages in between including
1451 * the first and last pages. This is done in a busy loop so that we
1452 * lose the least number of traces.
1453 * The pages are freed after we restart recording and unlock readers.
1454 */
1455 tail_page = &cpu_buffer->tail_page->list;
1456
1457 /*
1458 * tail page might be on reader page, we remove the next page
1459 * from the ring buffer
1460 */
1461 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1462 tail_page = rb_list_head(tail_page->next);
1463 to_remove = tail_page;
1464
1465 /* start of pages to remove */
1466 first_page = list_entry(rb_list_head(to_remove->next),
1467 struct buffer_page, list);
1468
1469 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1470 to_remove = rb_list_head(to_remove)->next;
1471 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1472 }
1473
1474 next_page = rb_list_head(to_remove)->next;
1475
1476 /*
1477 * Now we remove all pages between tail_page and next_page.
1478 * Make sure that we have head_bit value preserved for the
1479 * next page
1480 */
1481 tail_page->next = (struct list_head *)((unsigned long)next_page |
1482 head_bit);
1483 next_page = rb_list_head(next_page);
1484 next_page->prev = tail_page;
1485
1486 /* make sure pages points to a valid page in the ring buffer */
1487 cpu_buffer->pages = next_page;
1488
1489 /* update head page */
1490 if (head_bit)
1491 cpu_buffer->head_page = list_entry(next_page,
1492 struct buffer_page, list);
1493
1494 /*
1495 * change read pointer to make sure any read iterators reset
1496 * themselves
1497 */
1498 cpu_buffer->read = 0;
1499
1500 /* pages are removed, resume tracing and then free the pages */
1501 atomic_dec(&cpu_buffer->record_disabled);
1502 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1503
1504 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1505
1506 /* last buffer page to remove */
1507 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1508 list);
1509 tmp_iter_page = first_page;
1510
1511 do {
1512 to_remove_page = tmp_iter_page;
1513 rb_inc_page(cpu_buffer, &tmp_iter_page);
1514
1515 /* update the counters */
1516 page_entries = rb_page_entries(to_remove_page);
1517 if (page_entries) {
1518 /*
1519 * If something was added to this page, it was full
1520 * since it is not the tail page. So we deduct the
1521 * bytes consumed in ring buffer from here.
1522 * Increment overrun to account for the lost events.
1523 */
1524 local_add(page_entries, &cpu_buffer->overrun);
1525 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1526 }
1527
1528 /*
1529 * We have already removed references to this list item, just
1530 * free up the buffer_page and its page
1531 */
1532 free_buffer_page(to_remove_page);
1533 nr_removed--;
1534
1535 } while (to_remove_page != last_page);
1536
1537 RB_WARN_ON(cpu_buffer, nr_removed);
1538
1539 return nr_removed == 0;
1540 }
1541
1542 static int
rb_insert_pages(struct ring_buffer_per_cpu * cpu_buffer)1543 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1544 {
1545 struct list_head *pages = &cpu_buffer->new_pages;
1546 int retries, success;
1547
1548 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1549 /*
1550 * We are holding the reader lock, so the reader page won't be swapped
1551 * in the ring buffer. Now we are racing with the writer trying to
1552 * move head page and the tail page.
1553 * We are going to adapt the reader page update process where:
1554 * 1. We first splice the start and end of list of new pages between
1555 * the head page and its previous page.
1556 * 2. We cmpxchg the prev_page->next to point from head page to the
1557 * start of new pages list.
1558 * 3. Finally, we update the head->prev to the end of new list.
1559 *
1560 * We will try this process 10 times, to make sure that we don't keep
1561 * spinning.
1562 */
1563 retries = 10;
1564 success = 0;
1565 while (retries--) {
1566 struct list_head *head_page, *prev_page, *r;
1567 struct list_head *last_page, *first_page;
1568 struct list_head *head_page_with_bit;
1569
1570 head_page = &rb_set_head_page(cpu_buffer)->list;
1571 if (!head_page)
1572 break;
1573 prev_page = head_page->prev;
1574
1575 first_page = pages->next;
1576 last_page = pages->prev;
1577
1578 head_page_with_bit = (struct list_head *)
1579 ((unsigned long)head_page | RB_PAGE_HEAD);
1580
1581 last_page->next = head_page_with_bit;
1582 first_page->prev = prev_page;
1583
1584 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1585
1586 if (r == head_page_with_bit) {
1587 /*
1588 * yay, we replaced the page pointer to our new list,
1589 * now, we just have to update to head page's prev
1590 * pointer to point to end of list
1591 */
1592 head_page->prev = last_page;
1593 success = 1;
1594 break;
1595 }
1596 }
1597
1598 if (success)
1599 INIT_LIST_HEAD(pages);
1600 /*
1601 * If we weren't successful in adding in new pages, warn and stop
1602 * tracing
1603 */
1604 RB_WARN_ON(cpu_buffer, !success);
1605 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1606
1607 /* free pages if they weren't inserted */
1608 if (!success) {
1609 struct buffer_page *bpage, *tmp;
1610 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1611 list) {
1612 list_del_init(&bpage->list);
1613 free_buffer_page(bpage);
1614 }
1615 }
1616 return success;
1617 }
1618
rb_update_pages(struct ring_buffer_per_cpu * cpu_buffer)1619 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1620 {
1621 int success;
1622
1623 if (cpu_buffer->nr_pages_to_update > 0)
1624 success = rb_insert_pages(cpu_buffer);
1625 else
1626 success = rb_remove_pages(cpu_buffer,
1627 -cpu_buffer->nr_pages_to_update);
1628
1629 if (success)
1630 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1631 }
1632
update_pages_handler(struct work_struct * work)1633 static void update_pages_handler(struct work_struct *work)
1634 {
1635 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1636 struct ring_buffer_per_cpu, update_pages_work);
1637 rb_update_pages(cpu_buffer);
1638 complete(&cpu_buffer->update_done);
1639 }
1640
1641 /**
1642 * ring_buffer_resize - resize the ring buffer
1643 * @buffer: the buffer to resize.
1644 * @size: the new size.
1645 * @cpu_id: the cpu buffer to resize
1646 *
1647 * Minimum size is 2 * BUF_PAGE_SIZE.
1648 *
1649 * Returns 0 on success and < 0 on failure.
1650 */
ring_buffer_resize(struct ring_buffer * buffer,unsigned long size,int cpu_id)1651 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1652 int cpu_id)
1653 {
1654 struct ring_buffer_per_cpu *cpu_buffer;
1655 unsigned long nr_pages;
1656 int cpu, err = 0;
1657
1658 /*
1659 * Always succeed at resizing a non-existent buffer:
1660 */
1661 if (!buffer)
1662 return size;
1663
1664 /* Make sure the requested buffer exists */
1665 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1666 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1667 return size;
1668
1669 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1670
1671 /* we need a minimum of two pages */
1672 if (nr_pages < 2)
1673 nr_pages = 2;
1674
1675 size = nr_pages * BUF_PAGE_SIZE;
1676
1677 /*
1678 * Don't succeed if resizing is disabled, as a reader might be
1679 * manipulating the ring buffer and is expecting a sane state while
1680 * this is true.
1681 */
1682 if (atomic_read(&buffer->resize_disabled))
1683 return -EBUSY;
1684
1685 /* prevent another thread from changing buffer sizes */
1686 mutex_lock(&buffer->mutex);
1687
1688 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1689 /* calculate the pages to update */
1690 for_each_buffer_cpu(buffer, cpu) {
1691 cpu_buffer = buffer->buffers[cpu];
1692
1693 cpu_buffer->nr_pages_to_update = nr_pages -
1694 cpu_buffer->nr_pages;
1695 /*
1696 * nothing more to do for removing pages or no update
1697 */
1698 if (cpu_buffer->nr_pages_to_update <= 0)
1699 continue;
1700 /*
1701 * to add pages, make sure all new pages can be
1702 * allocated without receiving ENOMEM
1703 */
1704 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1705 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1706 &cpu_buffer->new_pages, cpu)) {
1707 /* not enough memory for new pages */
1708 err = -ENOMEM;
1709 goto out_err;
1710 }
1711 }
1712
1713 get_online_cpus();
1714 /*
1715 * Fire off all the required work handlers
1716 * We can't schedule on offline CPUs, but it's not necessary
1717 * since we can change their buffer sizes without any race.
1718 */
1719 for_each_buffer_cpu(buffer, cpu) {
1720 cpu_buffer = buffer->buffers[cpu];
1721 if (!cpu_buffer->nr_pages_to_update)
1722 continue;
1723
1724 /* Can't run something on an offline CPU. */
1725 if (!cpu_online(cpu)) {
1726 rb_update_pages(cpu_buffer);
1727 cpu_buffer->nr_pages_to_update = 0;
1728 } else {
1729 schedule_work_on(cpu,
1730 &cpu_buffer->update_pages_work);
1731 }
1732 }
1733
1734 /* wait for all the updates to complete */
1735 for_each_buffer_cpu(buffer, cpu) {
1736 cpu_buffer = buffer->buffers[cpu];
1737 if (!cpu_buffer->nr_pages_to_update)
1738 continue;
1739
1740 if (cpu_online(cpu))
1741 wait_for_completion(&cpu_buffer->update_done);
1742 cpu_buffer->nr_pages_to_update = 0;
1743 }
1744
1745 put_online_cpus();
1746 } else {
1747 /* Make sure this CPU has been intitialized */
1748 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1749 goto out;
1750
1751 cpu_buffer = buffer->buffers[cpu_id];
1752
1753 if (nr_pages == cpu_buffer->nr_pages)
1754 goto out;
1755
1756 cpu_buffer->nr_pages_to_update = nr_pages -
1757 cpu_buffer->nr_pages;
1758
1759 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1760 if (cpu_buffer->nr_pages_to_update > 0 &&
1761 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1762 &cpu_buffer->new_pages, cpu_id)) {
1763 err = -ENOMEM;
1764 goto out_err;
1765 }
1766
1767 get_online_cpus();
1768
1769 /* Can't run something on an offline CPU. */
1770 if (!cpu_online(cpu_id))
1771 rb_update_pages(cpu_buffer);
1772 else {
1773 schedule_work_on(cpu_id,
1774 &cpu_buffer->update_pages_work);
1775 wait_for_completion(&cpu_buffer->update_done);
1776 }
1777
1778 cpu_buffer->nr_pages_to_update = 0;
1779 put_online_cpus();
1780 }
1781
1782 out:
1783 /*
1784 * The ring buffer resize can happen with the ring buffer
1785 * enabled, so that the update disturbs the tracing as little
1786 * as possible. But if the buffer is disabled, we do not need
1787 * to worry about that, and we can take the time to verify
1788 * that the buffer is not corrupt.
1789 */
1790 if (atomic_read(&buffer->record_disabled)) {
1791 atomic_inc(&buffer->record_disabled);
1792 /*
1793 * Even though the buffer was disabled, we must make sure
1794 * that it is truly disabled before calling rb_check_pages.
1795 * There could have been a race between checking
1796 * record_disable and incrementing it.
1797 */
1798 synchronize_sched();
1799 for_each_buffer_cpu(buffer, cpu) {
1800 cpu_buffer = buffer->buffers[cpu];
1801 rb_check_pages(cpu_buffer);
1802 }
1803 atomic_dec(&buffer->record_disabled);
1804 }
1805
1806 mutex_unlock(&buffer->mutex);
1807 return size;
1808
1809 out_err:
1810 for_each_buffer_cpu(buffer, cpu) {
1811 struct buffer_page *bpage, *tmp;
1812
1813 cpu_buffer = buffer->buffers[cpu];
1814 cpu_buffer->nr_pages_to_update = 0;
1815
1816 if (list_empty(&cpu_buffer->new_pages))
1817 continue;
1818
1819 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1820 list) {
1821 list_del_init(&bpage->list);
1822 free_buffer_page(bpage);
1823 }
1824 }
1825 mutex_unlock(&buffer->mutex);
1826 return err;
1827 }
1828 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1829
ring_buffer_change_overwrite(struct ring_buffer * buffer,int val)1830 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1831 {
1832 mutex_lock(&buffer->mutex);
1833 if (val)
1834 buffer->flags |= RB_FL_OVERWRITE;
1835 else
1836 buffer->flags &= ~RB_FL_OVERWRITE;
1837 mutex_unlock(&buffer->mutex);
1838 }
1839 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1840
1841 static inline void *
__rb_data_page_index(struct buffer_data_page * bpage,unsigned index)1842 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1843 {
1844 return bpage->data + index;
1845 }
1846
__rb_page_index(struct buffer_page * bpage,unsigned index)1847 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1848 {
1849 return bpage->page->data + index;
1850 }
1851
1852 static inline struct ring_buffer_event *
rb_reader_event(struct ring_buffer_per_cpu * cpu_buffer)1853 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1854 {
1855 return __rb_page_index(cpu_buffer->reader_page,
1856 cpu_buffer->reader_page->read);
1857 }
1858
1859 static inline struct ring_buffer_event *
rb_iter_head_event(struct ring_buffer_iter * iter)1860 rb_iter_head_event(struct ring_buffer_iter *iter)
1861 {
1862 return __rb_page_index(iter->head_page, iter->head);
1863 }
1864
rb_page_commit(struct buffer_page * bpage)1865 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1866 {
1867 return local_read(&bpage->page->commit);
1868 }
1869
1870 /* Size is determined by what has been committed */
rb_page_size(struct buffer_page * bpage)1871 static inline unsigned rb_page_size(struct buffer_page *bpage)
1872 {
1873 return rb_page_commit(bpage);
1874 }
1875
1876 static inline unsigned
rb_commit_index(struct ring_buffer_per_cpu * cpu_buffer)1877 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1878 {
1879 return rb_page_commit(cpu_buffer->commit_page);
1880 }
1881
1882 static inline unsigned
rb_event_index(struct ring_buffer_event * event)1883 rb_event_index(struct ring_buffer_event *event)
1884 {
1885 unsigned long addr = (unsigned long)event;
1886
1887 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1888 }
1889
rb_inc_iter(struct ring_buffer_iter * iter)1890 static void rb_inc_iter(struct ring_buffer_iter *iter)
1891 {
1892 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1893
1894 /*
1895 * The iterator could be on the reader page (it starts there).
1896 * But the head could have moved, since the reader was
1897 * found. Check for this case and assign the iterator
1898 * to the head page instead of next.
1899 */
1900 if (iter->head_page == cpu_buffer->reader_page)
1901 iter->head_page = rb_set_head_page(cpu_buffer);
1902 else
1903 rb_inc_page(cpu_buffer, &iter->head_page);
1904
1905 iter->read_stamp = iter->head_page->page->time_stamp;
1906 iter->head = 0;
1907 }
1908
1909 /*
1910 * rb_handle_head_page - writer hit the head page
1911 *
1912 * Returns: +1 to retry page
1913 * 0 to continue
1914 * -1 on error
1915 */
1916 static int
rb_handle_head_page(struct ring_buffer_per_cpu * cpu_buffer,struct buffer_page * tail_page,struct buffer_page * next_page)1917 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1918 struct buffer_page *tail_page,
1919 struct buffer_page *next_page)
1920 {
1921 struct buffer_page *new_head;
1922 int entries;
1923 int type;
1924 int ret;
1925
1926 entries = rb_page_entries(next_page);
1927
1928 /*
1929 * The hard part is here. We need to move the head
1930 * forward, and protect against both readers on
1931 * other CPUs and writers coming in via interrupts.
1932 */
1933 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1934 RB_PAGE_HEAD);
1935
1936 /*
1937 * type can be one of four:
1938 * NORMAL - an interrupt already moved it for us
1939 * HEAD - we are the first to get here.
1940 * UPDATE - we are the interrupt interrupting
1941 * a current move.
1942 * MOVED - a reader on another CPU moved the next
1943 * pointer to its reader page. Give up
1944 * and try again.
1945 */
1946
1947 switch (type) {
1948 case RB_PAGE_HEAD:
1949 /*
1950 * We changed the head to UPDATE, thus
1951 * it is our responsibility to update
1952 * the counters.
1953 */
1954 local_add(entries, &cpu_buffer->overrun);
1955 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1956
1957 /*
1958 * The entries will be zeroed out when we move the
1959 * tail page.
1960 */
1961
1962 /* still more to do */
1963 break;
1964
1965 case RB_PAGE_UPDATE:
1966 /*
1967 * This is an interrupt that interrupt the
1968 * previous update. Still more to do.
1969 */
1970 break;
1971 case RB_PAGE_NORMAL:
1972 /*
1973 * An interrupt came in before the update
1974 * and processed this for us.
1975 * Nothing left to do.
1976 */
1977 return 1;
1978 case RB_PAGE_MOVED:
1979 /*
1980 * The reader is on another CPU and just did
1981 * a swap with our next_page.
1982 * Try again.
1983 */
1984 return 1;
1985 default:
1986 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1987 return -1;
1988 }
1989
1990 /*
1991 * Now that we are here, the old head pointer is
1992 * set to UPDATE. This will keep the reader from
1993 * swapping the head page with the reader page.
1994 * The reader (on another CPU) will spin till
1995 * we are finished.
1996 *
1997 * We just need to protect against interrupts
1998 * doing the job. We will set the next pointer
1999 * to HEAD. After that, we set the old pointer
2000 * to NORMAL, but only if it was HEAD before.
2001 * otherwise we are an interrupt, and only
2002 * want the outer most commit to reset it.
2003 */
2004 new_head = next_page;
2005 rb_inc_page(cpu_buffer, &new_head);
2006
2007 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2008 RB_PAGE_NORMAL);
2009
2010 /*
2011 * Valid returns are:
2012 * HEAD - an interrupt came in and already set it.
2013 * NORMAL - One of two things:
2014 * 1) We really set it.
2015 * 2) A bunch of interrupts came in and moved
2016 * the page forward again.
2017 */
2018 switch (ret) {
2019 case RB_PAGE_HEAD:
2020 case RB_PAGE_NORMAL:
2021 /* OK */
2022 break;
2023 default:
2024 RB_WARN_ON(cpu_buffer, 1);
2025 return -1;
2026 }
2027
2028 /*
2029 * It is possible that an interrupt came in,
2030 * set the head up, then more interrupts came in
2031 * and moved it again. When we get back here,
2032 * the page would have been set to NORMAL but we
2033 * just set it back to HEAD.
2034 *
2035 * How do you detect this? Well, if that happened
2036 * the tail page would have moved.
2037 */
2038 if (ret == RB_PAGE_NORMAL) {
2039 /*
2040 * If the tail had moved passed next, then we need
2041 * to reset the pointer.
2042 */
2043 if (cpu_buffer->tail_page != tail_page &&
2044 cpu_buffer->tail_page != next_page)
2045 rb_head_page_set_normal(cpu_buffer, new_head,
2046 next_page,
2047 RB_PAGE_HEAD);
2048 }
2049
2050 /*
2051 * If this was the outer most commit (the one that
2052 * changed the original pointer from HEAD to UPDATE),
2053 * then it is up to us to reset it to NORMAL.
2054 */
2055 if (type == RB_PAGE_HEAD) {
2056 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2057 tail_page,
2058 RB_PAGE_UPDATE);
2059 if (RB_WARN_ON(cpu_buffer,
2060 ret != RB_PAGE_UPDATE))
2061 return -1;
2062 }
2063
2064 return 0;
2065 }
2066
2067 static inline void
rb_reset_tail(struct ring_buffer_per_cpu * cpu_buffer,unsigned long tail,struct rb_event_info * info)2068 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2069 unsigned long tail, struct rb_event_info *info)
2070 {
2071 struct buffer_page *tail_page = info->tail_page;
2072 struct ring_buffer_event *event;
2073 unsigned long length = info->length;
2074
2075 /*
2076 * Only the event that crossed the page boundary
2077 * must fill the old tail_page with padding.
2078 */
2079 if (tail >= BUF_PAGE_SIZE) {
2080 /*
2081 * If the page was filled, then we still need
2082 * to update the real_end. Reset it to zero
2083 * and the reader will ignore it.
2084 */
2085 if (tail == BUF_PAGE_SIZE)
2086 tail_page->real_end = 0;
2087
2088 local_sub(length, &tail_page->write);
2089 return;
2090 }
2091
2092 event = __rb_page_index(tail_page, tail);
2093 kmemcheck_annotate_bitfield(event, bitfield);
2094
2095 /* account for padding bytes */
2096 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2097
2098 /*
2099 * Save the original length to the meta data.
2100 * This will be used by the reader to add lost event
2101 * counter.
2102 */
2103 tail_page->real_end = tail;
2104
2105 /*
2106 * If this event is bigger than the minimum size, then
2107 * we need to be careful that we don't subtract the
2108 * write counter enough to allow another writer to slip
2109 * in on this page.
2110 * We put in a discarded commit instead, to make sure
2111 * that this space is not used again.
2112 *
2113 * If we are less than the minimum size, we don't need to
2114 * worry about it.
2115 */
2116 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2117 /* No room for any events */
2118
2119 /* Mark the rest of the page with padding */
2120 rb_event_set_padding(event);
2121
2122 /* Set the write back to the previous setting */
2123 local_sub(length, &tail_page->write);
2124 return;
2125 }
2126
2127 /* Put in a discarded event */
2128 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2129 event->type_len = RINGBUF_TYPE_PADDING;
2130 /* time delta must be non zero */
2131 event->time_delta = 1;
2132
2133 /* Set write to end of buffer */
2134 length = (tail + length) - BUF_PAGE_SIZE;
2135 local_sub(length, &tail_page->write);
2136 }
2137
2138 /*
2139 * This is the slow path, force gcc not to inline it.
2140 */
2141 static noinline struct ring_buffer_event *
rb_move_tail(struct ring_buffer_per_cpu * cpu_buffer,unsigned long tail,struct rb_event_info * info)2142 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2143 unsigned long tail, struct rb_event_info *info)
2144 {
2145 struct buffer_page *tail_page = info->tail_page;
2146 struct buffer_page *commit_page = cpu_buffer->commit_page;
2147 struct ring_buffer *buffer = cpu_buffer->buffer;
2148 struct buffer_page *next_page;
2149 int ret;
2150 u64 ts;
2151
2152 next_page = tail_page;
2153
2154 rb_inc_page(cpu_buffer, &next_page);
2155
2156 /*
2157 * If for some reason, we had an interrupt storm that made
2158 * it all the way around the buffer, bail, and warn
2159 * about it.
2160 */
2161 if (unlikely(next_page == commit_page)) {
2162 local_inc(&cpu_buffer->commit_overrun);
2163 goto out_reset;
2164 }
2165
2166 /*
2167 * This is where the fun begins!
2168 *
2169 * We are fighting against races between a reader that
2170 * could be on another CPU trying to swap its reader
2171 * page with the buffer head.
2172 *
2173 * We are also fighting against interrupts coming in and
2174 * moving the head or tail on us as well.
2175 *
2176 * If the next page is the head page then we have filled
2177 * the buffer, unless the commit page is still on the
2178 * reader page.
2179 */
2180 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2181
2182 /*
2183 * If the commit is not on the reader page, then
2184 * move the header page.
2185 */
2186 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2187 /*
2188 * If we are not in overwrite mode,
2189 * this is easy, just stop here.
2190 */
2191 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2192 local_inc(&cpu_buffer->dropped_events);
2193 goto out_reset;
2194 }
2195
2196 ret = rb_handle_head_page(cpu_buffer,
2197 tail_page,
2198 next_page);
2199 if (ret < 0)
2200 goto out_reset;
2201 if (ret)
2202 goto out_again;
2203 } else {
2204 /*
2205 * We need to be careful here too. The
2206 * commit page could still be on the reader
2207 * page. We could have a small buffer, and
2208 * have filled up the buffer with events
2209 * from interrupts and such, and wrapped.
2210 *
2211 * Note, if the tail page is also the on the
2212 * reader_page, we let it move out.
2213 */
2214 if (unlikely((cpu_buffer->commit_page !=
2215 cpu_buffer->tail_page) &&
2216 (cpu_buffer->commit_page ==
2217 cpu_buffer->reader_page))) {
2218 local_inc(&cpu_buffer->commit_overrun);
2219 goto out_reset;
2220 }
2221 }
2222 }
2223
2224 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
2225 if (ret) {
2226 /*
2227 * Nested commits always have zero deltas, so
2228 * just reread the time stamp
2229 */
2230 ts = rb_time_stamp(buffer);
2231 next_page->page->time_stamp = ts;
2232 }
2233
2234 out_again:
2235
2236 rb_reset_tail(cpu_buffer, tail, info);
2237
2238 /* fail and let the caller try again */
2239 return ERR_PTR(-EAGAIN);
2240
2241 out_reset:
2242 /* reset write */
2243 rb_reset_tail(cpu_buffer, tail, info);
2244
2245 return NULL;
2246 }
2247
2248 /* Slow path, do not inline */
2249 static noinline struct ring_buffer_event *
rb_add_time_stamp(struct ring_buffer_event * event,u64 delta)2250 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
2251 {
2252 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2253
2254 /* Not the first event on the page? */
2255 if (rb_event_index(event)) {
2256 event->time_delta = delta & TS_MASK;
2257 event->array[0] = delta >> TS_SHIFT;
2258 } else {
2259 /* nope, just zero it */
2260 event->time_delta = 0;
2261 event->array[0] = 0;
2262 }
2263
2264 return skip_time_extend(event);
2265 }
2266
2267 static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2268 struct ring_buffer_event *event);
2269
2270 /**
2271 * rb_update_event - update event type and data
2272 * @event: the event to update
2273 * @type: the type of event
2274 * @length: the size of the event field in the ring buffer
2275 *
2276 * Update the type and data fields of the event. The length
2277 * is the actual size that is written to the ring buffer,
2278 * and with this, we can determine what to place into the
2279 * data field.
2280 */
2281 static void
rb_update_event(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event,struct rb_event_info * info)2282 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2283 struct ring_buffer_event *event,
2284 struct rb_event_info *info)
2285 {
2286 unsigned length = info->length;
2287 u64 delta = info->delta;
2288
2289 /* Only a commit updates the timestamp */
2290 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2291 delta = 0;
2292
2293 /*
2294 * If we need to add a timestamp, then we
2295 * add it to the start of the resevered space.
2296 */
2297 if (unlikely(info->add_timestamp)) {
2298 event = rb_add_time_stamp(event, delta);
2299 length -= RB_LEN_TIME_EXTEND;
2300 delta = 0;
2301 }
2302
2303 event->time_delta = delta;
2304 length -= RB_EVNT_HDR_SIZE;
2305 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2306 event->type_len = 0;
2307 event->array[0] = length;
2308 } else
2309 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2310 }
2311
rb_calculate_event_length(unsigned length)2312 static unsigned rb_calculate_event_length(unsigned length)
2313 {
2314 struct ring_buffer_event event; /* Used only for sizeof array */
2315
2316 /* zero length can cause confusions */
2317 if (!length)
2318 length++;
2319
2320 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2321 length += sizeof(event.array[0]);
2322
2323 length += RB_EVNT_HDR_SIZE;
2324 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2325
2326 /*
2327 * In case the time delta is larger than the 27 bits for it
2328 * in the header, we need to add a timestamp. If another
2329 * event comes in when trying to discard this one to increase
2330 * the length, then the timestamp will be added in the allocated
2331 * space of this event. If length is bigger than the size needed
2332 * for the TIME_EXTEND, then padding has to be used. The events
2333 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2334 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2335 * As length is a multiple of 4, we only need to worry if it
2336 * is 12 (RB_LEN_TIME_EXTEND + 4).
2337 */
2338 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2339 length += RB_ALIGNMENT;
2340
2341 return length;
2342 }
2343
2344 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
sched_clock_stable(void)2345 static inline bool sched_clock_stable(void)
2346 {
2347 return true;
2348 }
2349 #endif
2350
2351 static inline int
rb_try_to_discard(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event)2352 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2353 struct ring_buffer_event *event)
2354 {
2355 unsigned long new_index, old_index;
2356 struct buffer_page *bpage;
2357 unsigned long index;
2358 unsigned long addr;
2359
2360 new_index = rb_event_index(event);
2361 old_index = new_index + rb_event_ts_length(event);
2362 addr = (unsigned long)event;
2363 addr &= PAGE_MASK;
2364
2365 bpage = cpu_buffer->tail_page;
2366
2367 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2368 unsigned long write_mask =
2369 local_read(&bpage->write) & ~RB_WRITE_MASK;
2370 unsigned long event_length = rb_event_length(event);
2371 /*
2372 * This is on the tail page. It is possible that
2373 * a write could come in and move the tail page
2374 * and write to the next page. That is fine
2375 * because we just shorten what is on this page.
2376 */
2377 old_index += write_mask;
2378 new_index += write_mask;
2379 index = local_cmpxchg(&bpage->write, old_index, new_index);
2380 if (index == old_index) {
2381 /* update counters */
2382 local_sub(event_length, &cpu_buffer->entries_bytes);
2383 return 1;
2384 }
2385 }
2386
2387 /* could not discard */
2388 return 0;
2389 }
2390
rb_start_commit(struct ring_buffer_per_cpu * cpu_buffer)2391 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2392 {
2393 local_inc(&cpu_buffer->committing);
2394 local_inc(&cpu_buffer->commits);
2395 }
2396
2397 static void
rb_set_commit_to_write(struct ring_buffer_per_cpu * cpu_buffer)2398 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2399 {
2400 unsigned long max_count;
2401
2402 /*
2403 * We only race with interrupts and NMIs on this CPU.
2404 * If we own the commit event, then we can commit
2405 * all others that interrupted us, since the interruptions
2406 * are in stack format (they finish before they come
2407 * back to us). This allows us to do a simple loop to
2408 * assign the commit to the tail.
2409 */
2410 again:
2411 max_count = cpu_buffer->nr_pages * 100;
2412
2413 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
2414 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2415 return;
2416 if (RB_WARN_ON(cpu_buffer,
2417 rb_is_reader_page(cpu_buffer->tail_page)))
2418 return;
2419 local_set(&cpu_buffer->commit_page->page->commit,
2420 rb_page_write(cpu_buffer->commit_page));
2421 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2422 cpu_buffer->write_stamp =
2423 cpu_buffer->commit_page->page->time_stamp;
2424 /* add barrier to keep gcc from optimizing too much */
2425 barrier();
2426 }
2427 while (rb_commit_index(cpu_buffer) !=
2428 rb_page_write(cpu_buffer->commit_page)) {
2429
2430 local_set(&cpu_buffer->commit_page->page->commit,
2431 rb_page_write(cpu_buffer->commit_page));
2432 RB_WARN_ON(cpu_buffer,
2433 local_read(&cpu_buffer->commit_page->page->commit) &
2434 ~RB_WRITE_MASK);
2435 barrier();
2436 }
2437
2438 /* again, keep gcc from optimizing */
2439 barrier();
2440
2441 /*
2442 * If an interrupt came in just after the first while loop
2443 * and pushed the tail page forward, we will be left with
2444 * a dangling commit that will never go forward.
2445 */
2446 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
2447 goto again;
2448 }
2449
rb_end_commit(struct ring_buffer_per_cpu * cpu_buffer)2450 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2451 {
2452 unsigned long commits;
2453
2454 if (RB_WARN_ON(cpu_buffer,
2455 !local_read(&cpu_buffer->committing)))
2456 return;
2457
2458 again:
2459 commits = local_read(&cpu_buffer->commits);
2460 /* synchronize with interrupts */
2461 barrier();
2462 if (local_read(&cpu_buffer->committing) == 1)
2463 rb_set_commit_to_write(cpu_buffer);
2464
2465 local_dec(&cpu_buffer->committing);
2466
2467 /* synchronize with interrupts */
2468 barrier();
2469
2470 /*
2471 * Need to account for interrupts coming in between the
2472 * updating of the commit page and the clearing of the
2473 * committing counter.
2474 */
2475 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2476 !local_read(&cpu_buffer->committing)) {
2477 local_inc(&cpu_buffer->committing);
2478 goto again;
2479 }
2480 }
2481
rb_event_discard(struct ring_buffer_event * event)2482 static inline void rb_event_discard(struct ring_buffer_event *event)
2483 {
2484 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2485 event = skip_time_extend(event);
2486
2487 /* array[0] holds the actual length for the discarded event */
2488 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2489 event->type_len = RINGBUF_TYPE_PADDING;
2490 /* time delta must be non zero */
2491 if (!event->time_delta)
2492 event->time_delta = 1;
2493 }
2494
2495 static inline bool
rb_event_is_commit(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event)2496 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2497 struct ring_buffer_event *event)
2498 {
2499 unsigned long addr = (unsigned long)event;
2500 unsigned long index;
2501
2502 index = rb_event_index(event);
2503 addr &= PAGE_MASK;
2504
2505 return cpu_buffer->commit_page->page == (void *)addr &&
2506 rb_commit_index(cpu_buffer) == index;
2507 }
2508
2509 static void
rb_update_write_stamp(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event)2510 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2511 struct ring_buffer_event *event)
2512 {
2513 u64 delta;
2514
2515 /*
2516 * The event first in the commit queue updates the
2517 * time stamp.
2518 */
2519 if (rb_event_is_commit(cpu_buffer, event)) {
2520 /*
2521 * A commit event that is first on a page
2522 * updates the write timestamp with the page stamp
2523 */
2524 if (!rb_event_index(event))
2525 cpu_buffer->write_stamp =
2526 cpu_buffer->commit_page->page->time_stamp;
2527 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2528 delta = event->array[0];
2529 delta <<= TS_SHIFT;
2530 delta += event->time_delta;
2531 cpu_buffer->write_stamp += delta;
2532 } else
2533 cpu_buffer->write_stamp += event->time_delta;
2534 }
2535 }
2536
rb_commit(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event)2537 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2538 struct ring_buffer_event *event)
2539 {
2540 local_inc(&cpu_buffer->entries);
2541 rb_update_write_stamp(cpu_buffer, event);
2542 rb_end_commit(cpu_buffer);
2543 }
2544
2545 static __always_inline void
rb_wakeups(struct ring_buffer * buffer,struct ring_buffer_per_cpu * cpu_buffer)2546 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2547 {
2548 bool pagebusy;
2549
2550 if (buffer->irq_work.waiters_pending) {
2551 buffer->irq_work.waiters_pending = false;
2552 /* irq_work_queue() supplies it's own memory barriers */
2553 irq_work_queue(&buffer->irq_work.work);
2554 }
2555
2556 if (cpu_buffer->irq_work.waiters_pending) {
2557 cpu_buffer->irq_work.waiters_pending = false;
2558 /* irq_work_queue() supplies it's own memory barriers */
2559 irq_work_queue(&cpu_buffer->irq_work.work);
2560 }
2561
2562 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
2563
2564 if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2565 cpu_buffer->irq_work.wakeup_full = true;
2566 cpu_buffer->irq_work.full_waiters_pending = false;
2567 /* irq_work_queue() supplies it's own memory barriers */
2568 irq_work_queue(&cpu_buffer->irq_work.work);
2569 }
2570 }
2571
2572 /*
2573 * The lock and unlock are done within a preempt disable section.
2574 * The current_context per_cpu variable can only be modified
2575 * by the current task between lock and unlock. But it can
2576 * be modified more than once via an interrupt. To pass this
2577 * information from the lock to the unlock without having to
2578 * access the 'in_interrupt()' functions again (which do show
2579 * a bit of overhead in something as critical as function tracing,
2580 * we use a bitmask trick.
2581 *
2582 * bit 0 = NMI context
2583 * bit 1 = IRQ context
2584 * bit 2 = SoftIRQ context
2585 * bit 3 = normal context.
2586 *
2587 * This works because this is the order of contexts that can
2588 * preempt other contexts. A SoftIRQ never preempts an IRQ
2589 * context.
2590 *
2591 * When the context is determined, the corresponding bit is
2592 * checked and set (if it was set, then a recursion of that context
2593 * happened).
2594 *
2595 * On unlock, we need to clear this bit. To do so, just subtract
2596 * 1 from the current_context and AND it to itself.
2597 *
2598 * (binary)
2599 * 101 - 1 = 100
2600 * 101 & 100 = 100 (clearing bit zero)
2601 *
2602 * 1010 - 1 = 1001
2603 * 1010 & 1001 = 1000 (clearing bit 1)
2604 *
2605 * The least significant bit can be cleared this way, and it
2606 * just so happens that it is the same bit corresponding to
2607 * the current context.
2608 */
2609
2610 static __always_inline int
trace_recursive_lock(struct ring_buffer_per_cpu * cpu_buffer)2611 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2612 {
2613 unsigned int val = cpu_buffer->current_context;
2614 int bit;
2615
2616 if (in_interrupt()) {
2617 if (in_nmi())
2618 bit = RB_CTX_NMI;
2619 else if (in_irq())
2620 bit = RB_CTX_IRQ;
2621 else
2622 bit = RB_CTX_SOFTIRQ;
2623 } else
2624 bit = RB_CTX_NORMAL;
2625
2626 if (unlikely(val & (1 << bit)))
2627 return 1;
2628
2629 val |= (1 << bit);
2630 cpu_buffer->current_context = val;
2631
2632 return 0;
2633 }
2634
2635 static __always_inline void
trace_recursive_unlock(struct ring_buffer_per_cpu * cpu_buffer)2636 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2637 {
2638 cpu_buffer->current_context &= cpu_buffer->current_context - 1;
2639 }
2640
2641 /**
2642 * ring_buffer_unlock_commit - commit a reserved
2643 * @buffer: The buffer to commit to
2644 * @event: The event pointer to commit.
2645 *
2646 * This commits the data to the ring buffer, and releases any locks held.
2647 *
2648 * Must be paired with ring_buffer_lock_reserve.
2649 */
ring_buffer_unlock_commit(struct ring_buffer * buffer,struct ring_buffer_event * event)2650 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2651 struct ring_buffer_event *event)
2652 {
2653 struct ring_buffer_per_cpu *cpu_buffer;
2654 int cpu = raw_smp_processor_id();
2655
2656 cpu_buffer = buffer->buffers[cpu];
2657
2658 rb_commit(cpu_buffer, event);
2659
2660 rb_wakeups(buffer, cpu_buffer);
2661
2662 trace_recursive_unlock(cpu_buffer);
2663
2664 preempt_enable_notrace();
2665
2666 return 0;
2667 }
2668 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2669
2670 static noinline void
rb_handle_timestamp(struct ring_buffer_per_cpu * cpu_buffer,struct rb_event_info * info)2671 rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2672 struct rb_event_info *info)
2673 {
2674 WARN_ONCE(info->delta > (1ULL << 59),
2675 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2676 (unsigned long long)info->delta,
2677 (unsigned long long)info->ts,
2678 (unsigned long long)cpu_buffer->write_stamp,
2679 sched_clock_stable() ? "" :
2680 "If you just came from a suspend/resume,\n"
2681 "please switch to the trace global clock:\n"
2682 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2683 info->add_timestamp = 1;
2684 }
2685
2686 static struct ring_buffer_event *
__rb_reserve_next(struct ring_buffer_per_cpu * cpu_buffer,struct rb_event_info * info)2687 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2688 struct rb_event_info *info)
2689 {
2690 struct ring_buffer_event *event;
2691 struct buffer_page *tail_page;
2692 unsigned long tail, write;
2693
2694 /*
2695 * If the time delta since the last event is too big to
2696 * hold in the time field of the event, then we append a
2697 * TIME EXTEND event ahead of the data event.
2698 */
2699 if (unlikely(info->add_timestamp))
2700 info->length += RB_LEN_TIME_EXTEND;
2701
2702 tail_page = info->tail_page = cpu_buffer->tail_page;
2703 write = local_add_return(info->length, &tail_page->write);
2704
2705 /* set write to only the index of the write */
2706 write &= RB_WRITE_MASK;
2707 tail = write - info->length;
2708
2709 /*
2710 * If this is the first commit on the page, then it has the same
2711 * timestamp as the page itself.
2712 */
2713 if (!tail)
2714 info->delta = 0;
2715
2716 /* See if we shot pass the end of this buffer page */
2717 if (unlikely(write > BUF_PAGE_SIZE))
2718 return rb_move_tail(cpu_buffer, tail, info);
2719
2720 /* We reserved something on the buffer */
2721
2722 event = __rb_page_index(tail_page, tail);
2723 kmemcheck_annotate_bitfield(event, bitfield);
2724 rb_update_event(cpu_buffer, event, info);
2725
2726 local_inc(&tail_page->entries);
2727
2728 /*
2729 * If this is the first commit on the page, then update
2730 * its timestamp.
2731 */
2732 if (!tail)
2733 tail_page->page->time_stamp = info->ts;
2734
2735 /* account for these added bytes */
2736 local_add(info->length, &cpu_buffer->entries_bytes);
2737
2738 return event;
2739 }
2740
2741 static struct ring_buffer_event *
rb_reserve_next_event(struct ring_buffer * buffer,struct ring_buffer_per_cpu * cpu_buffer,unsigned long length)2742 rb_reserve_next_event(struct ring_buffer *buffer,
2743 struct ring_buffer_per_cpu *cpu_buffer,
2744 unsigned long length)
2745 {
2746 struct ring_buffer_event *event;
2747 struct rb_event_info info;
2748 int nr_loops = 0;
2749 u64 diff;
2750
2751 rb_start_commit(cpu_buffer);
2752
2753 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2754 /*
2755 * Due to the ability to swap a cpu buffer from a buffer
2756 * it is possible it was swapped before we committed.
2757 * (committing stops a swap). We check for it here and
2758 * if it happened, we have to fail the write.
2759 */
2760 barrier();
2761 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2762 local_dec(&cpu_buffer->committing);
2763 local_dec(&cpu_buffer->commits);
2764 return NULL;
2765 }
2766 #endif
2767
2768 info.length = rb_calculate_event_length(length);
2769 again:
2770 info.add_timestamp = 0;
2771 info.delta = 0;
2772
2773 /*
2774 * We allow for interrupts to reenter here and do a trace.
2775 * If one does, it will cause this original code to loop
2776 * back here. Even with heavy interrupts happening, this
2777 * should only happen a few times in a row. If this happens
2778 * 1000 times in a row, there must be either an interrupt
2779 * storm or we have something buggy.
2780 * Bail!
2781 */
2782 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2783 goto out_fail;
2784
2785 info.ts = rb_time_stamp(cpu_buffer->buffer);
2786 diff = info.ts - cpu_buffer->write_stamp;
2787
2788 /* make sure this diff is calculated here */
2789 barrier();
2790
2791 /* Did the write stamp get updated already? */
2792 if (likely(info.ts >= cpu_buffer->write_stamp)) {
2793 info.delta = diff;
2794 if (unlikely(test_time_stamp(info.delta)))
2795 rb_handle_timestamp(cpu_buffer, &info);
2796 }
2797
2798 event = __rb_reserve_next(cpu_buffer, &info);
2799
2800 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2801 if (info.add_timestamp)
2802 info.length -= RB_LEN_TIME_EXTEND;
2803 goto again;
2804 }
2805
2806 if (!event)
2807 goto out_fail;
2808
2809 return event;
2810
2811 out_fail:
2812 rb_end_commit(cpu_buffer);
2813 return NULL;
2814 }
2815
2816 /**
2817 * ring_buffer_lock_reserve - reserve a part of the buffer
2818 * @buffer: the ring buffer to reserve from
2819 * @length: the length of the data to reserve (excluding event header)
2820 *
2821 * Returns a reseverd event on the ring buffer to copy directly to.
2822 * The user of this interface will need to get the body to write into
2823 * and can use the ring_buffer_event_data() interface.
2824 *
2825 * The length is the length of the data needed, not the event length
2826 * which also includes the event header.
2827 *
2828 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2829 * If NULL is returned, then nothing has been allocated or locked.
2830 */
2831 struct ring_buffer_event *
ring_buffer_lock_reserve(struct ring_buffer * buffer,unsigned long length)2832 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2833 {
2834 struct ring_buffer_per_cpu *cpu_buffer;
2835 struct ring_buffer_event *event;
2836 int cpu;
2837
2838 /* If we are tracing schedule, we don't want to recurse */
2839 preempt_disable_notrace();
2840
2841 if (unlikely(atomic_read(&buffer->record_disabled)))
2842 goto out;
2843
2844 cpu = raw_smp_processor_id();
2845
2846 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2847 goto out;
2848
2849 cpu_buffer = buffer->buffers[cpu];
2850
2851 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2852 goto out;
2853
2854 if (unlikely(length > BUF_MAX_DATA_SIZE))
2855 goto out;
2856
2857 if (unlikely(trace_recursive_lock(cpu_buffer)))
2858 goto out;
2859
2860 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2861 if (!event)
2862 goto out_unlock;
2863
2864 return event;
2865
2866 out_unlock:
2867 trace_recursive_unlock(cpu_buffer);
2868 out:
2869 preempt_enable_notrace();
2870 return NULL;
2871 }
2872 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2873
2874 /*
2875 * Decrement the entries to the page that an event is on.
2876 * The event does not even need to exist, only the pointer
2877 * to the page it is on. This may only be called before the commit
2878 * takes place.
2879 */
2880 static inline void
rb_decrement_entry(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event)2881 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2882 struct ring_buffer_event *event)
2883 {
2884 unsigned long addr = (unsigned long)event;
2885 struct buffer_page *bpage = cpu_buffer->commit_page;
2886 struct buffer_page *start;
2887
2888 addr &= PAGE_MASK;
2889
2890 /* Do the likely case first */
2891 if (likely(bpage->page == (void *)addr)) {
2892 local_dec(&bpage->entries);
2893 return;
2894 }
2895
2896 /*
2897 * Because the commit page may be on the reader page we
2898 * start with the next page and check the end loop there.
2899 */
2900 rb_inc_page(cpu_buffer, &bpage);
2901 start = bpage;
2902 do {
2903 if (bpage->page == (void *)addr) {
2904 local_dec(&bpage->entries);
2905 return;
2906 }
2907 rb_inc_page(cpu_buffer, &bpage);
2908 } while (bpage != start);
2909
2910 /* commit not part of this buffer?? */
2911 RB_WARN_ON(cpu_buffer, 1);
2912 }
2913
2914 /**
2915 * ring_buffer_commit_discard - discard an event that has not been committed
2916 * @buffer: the ring buffer
2917 * @event: non committed event to discard
2918 *
2919 * Sometimes an event that is in the ring buffer needs to be ignored.
2920 * This function lets the user discard an event in the ring buffer
2921 * and then that event will not be read later.
2922 *
2923 * This function only works if it is called before the the item has been
2924 * committed. It will try to free the event from the ring buffer
2925 * if another event has not been added behind it.
2926 *
2927 * If another event has been added behind it, it will set the event
2928 * up as discarded, and perform the commit.
2929 *
2930 * If this function is called, do not call ring_buffer_unlock_commit on
2931 * the event.
2932 */
ring_buffer_discard_commit(struct ring_buffer * buffer,struct ring_buffer_event * event)2933 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2934 struct ring_buffer_event *event)
2935 {
2936 struct ring_buffer_per_cpu *cpu_buffer;
2937 int cpu;
2938
2939 /* The event is discarded regardless */
2940 rb_event_discard(event);
2941
2942 cpu = smp_processor_id();
2943 cpu_buffer = buffer->buffers[cpu];
2944
2945 /*
2946 * This must only be called if the event has not been
2947 * committed yet. Thus we can assume that preemption
2948 * is still disabled.
2949 */
2950 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2951
2952 rb_decrement_entry(cpu_buffer, event);
2953 if (rb_try_to_discard(cpu_buffer, event))
2954 goto out;
2955
2956 /*
2957 * The commit is still visible by the reader, so we
2958 * must still update the timestamp.
2959 */
2960 rb_update_write_stamp(cpu_buffer, event);
2961 out:
2962 rb_end_commit(cpu_buffer);
2963
2964 trace_recursive_unlock(cpu_buffer);
2965
2966 preempt_enable_notrace();
2967
2968 }
2969 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2970
2971 /**
2972 * ring_buffer_write - write data to the buffer without reserving
2973 * @buffer: The ring buffer to write to.
2974 * @length: The length of the data being written (excluding the event header)
2975 * @data: The data to write to the buffer.
2976 *
2977 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2978 * one function. If you already have the data to write to the buffer, it
2979 * may be easier to simply call this function.
2980 *
2981 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2982 * and not the length of the event which would hold the header.
2983 */
ring_buffer_write(struct ring_buffer * buffer,unsigned long length,void * data)2984 int ring_buffer_write(struct ring_buffer *buffer,
2985 unsigned long length,
2986 void *data)
2987 {
2988 struct ring_buffer_per_cpu *cpu_buffer;
2989 struct ring_buffer_event *event;
2990 void *body;
2991 int ret = -EBUSY;
2992 int cpu;
2993
2994 preempt_disable_notrace();
2995
2996 if (atomic_read(&buffer->record_disabled))
2997 goto out;
2998
2999 cpu = raw_smp_processor_id();
3000
3001 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3002 goto out;
3003
3004 cpu_buffer = buffer->buffers[cpu];
3005
3006 if (atomic_read(&cpu_buffer->record_disabled))
3007 goto out;
3008
3009 if (length > BUF_MAX_DATA_SIZE)
3010 goto out;
3011
3012 if (unlikely(trace_recursive_lock(cpu_buffer)))
3013 goto out;
3014
3015 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3016 if (!event)
3017 goto out_unlock;
3018
3019 body = rb_event_data(event);
3020
3021 memcpy(body, data, length);
3022
3023 rb_commit(cpu_buffer, event);
3024
3025 rb_wakeups(buffer, cpu_buffer);
3026
3027 ret = 0;
3028
3029 out_unlock:
3030 trace_recursive_unlock(cpu_buffer);
3031
3032 out:
3033 preempt_enable_notrace();
3034
3035 return ret;
3036 }
3037 EXPORT_SYMBOL_GPL(ring_buffer_write);
3038
rb_per_cpu_empty(struct ring_buffer_per_cpu * cpu_buffer)3039 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3040 {
3041 struct buffer_page *reader = cpu_buffer->reader_page;
3042 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3043 struct buffer_page *commit = cpu_buffer->commit_page;
3044
3045 /* In case of error, head will be NULL */
3046 if (unlikely(!head))
3047 return true;
3048
3049 return reader->read == rb_page_commit(reader) &&
3050 (commit == reader ||
3051 (commit == head &&
3052 head->read == rb_page_commit(commit)));
3053 }
3054
3055 /**
3056 * ring_buffer_record_disable - stop all writes into the buffer
3057 * @buffer: The ring buffer to stop writes to.
3058 *
3059 * This prevents all writes to the buffer. Any attempt to write
3060 * to the buffer after this will fail and return NULL.
3061 *
3062 * The caller should call synchronize_sched() after this.
3063 */
ring_buffer_record_disable(struct ring_buffer * buffer)3064 void ring_buffer_record_disable(struct ring_buffer *buffer)
3065 {
3066 atomic_inc(&buffer->record_disabled);
3067 }
3068 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3069
3070 /**
3071 * ring_buffer_record_enable - enable writes to the buffer
3072 * @buffer: The ring buffer to enable writes
3073 *
3074 * Note, multiple disables will need the same number of enables
3075 * to truly enable the writing (much like preempt_disable).
3076 */
ring_buffer_record_enable(struct ring_buffer * buffer)3077 void ring_buffer_record_enable(struct ring_buffer *buffer)
3078 {
3079 atomic_dec(&buffer->record_disabled);
3080 }
3081 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3082
3083 /**
3084 * ring_buffer_record_off - stop all writes into the buffer
3085 * @buffer: The ring buffer to stop writes to.
3086 *
3087 * This prevents all writes to the buffer. Any attempt to write
3088 * to the buffer after this will fail and return NULL.
3089 *
3090 * This is different than ring_buffer_record_disable() as
3091 * it works like an on/off switch, where as the disable() version
3092 * must be paired with a enable().
3093 */
ring_buffer_record_off(struct ring_buffer * buffer)3094 void ring_buffer_record_off(struct ring_buffer *buffer)
3095 {
3096 unsigned int rd;
3097 unsigned int new_rd;
3098
3099 do {
3100 rd = atomic_read(&buffer->record_disabled);
3101 new_rd = rd | RB_BUFFER_OFF;
3102 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3103 }
3104 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3105
3106 /**
3107 * ring_buffer_record_on - restart writes into the buffer
3108 * @buffer: The ring buffer to start writes to.
3109 *
3110 * This enables all writes to the buffer that was disabled by
3111 * ring_buffer_record_off().
3112 *
3113 * This is different than ring_buffer_record_enable() as
3114 * it works like an on/off switch, where as the enable() version
3115 * must be paired with a disable().
3116 */
ring_buffer_record_on(struct ring_buffer * buffer)3117 void ring_buffer_record_on(struct ring_buffer *buffer)
3118 {
3119 unsigned int rd;
3120 unsigned int new_rd;
3121
3122 do {
3123 rd = atomic_read(&buffer->record_disabled);
3124 new_rd = rd & ~RB_BUFFER_OFF;
3125 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3126 }
3127 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3128
3129 /**
3130 * ring_buffer_record_is_on - return true if the ring buffer can write
3131 * @buffer: The ring buffer to see if write is enabled
3132 *
3133 * Returns true if the ring buffer is in a state that it accepts writes.
3134 */
ring_buffer_record_is_on(struct ring_buffer * buffer)3135 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3136 {
3137 return !atomic_read(&buffer->record_disabled);
3138 }
3139
3140 /**
3141 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3142 * @buffer: The ring buffer to stop writes to.
3143 * @cpu: The CPU buffer to stop
3144 *
3145 * This prevents all writes to the buffer. Any attempt to write
3146 * to the buffer after this will fail and return NULL.
3147 *
3148 * The caller should call synchronize_sched() after this.
3149 */
ring_buffer_record_disable_cpu(struct ring_buffer * buffer,int cpu)3150 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3151 {
3152 struct ring_buffer_per_cpu *cpu_buffer;
3153
3154 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3155 return;
3156
3157 cpu_buffer = buffer->buffers[cpu];
3158 atomic_inc(&cpu_buffer->record_disabled);
3159 }
3160 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3161
3162 /**
3163 * ring_buffer_record_enable_cpu - enable writes to the buffer
3164 * @buffer: The ring buffer to enable writes
3165 * @cpu: The CPU to enable.
3166 *
3167 * Note, multiple disables will need the same number of enables
3168 * to truly enable the writing (much like preempt_disable).
3169 */
ring_buffer_record_enable_cpu(struct ring_buffer * buffer,int cpu)3170 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3171 {
3172 struct ring_buffer_per_cpu *cpu_buffer;
3173
3174 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3175 return;
3176
3177 cpu_buffer = buffer->buffers[cpu];
3178 atomic_dec(&cpu_buffer->record_disabled);
3179 }
3180 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3181
3182 /*
3183 * The total entries in the ring buffer is the running counter
3184 * of entries entered into the ring buffer, minus the sum of
3185 * the entries read from the ring buffer and the number of
3186 * entries that were overwritten.
3187 */
3188 static inline unsigned long
rb_num_of_entries(struct ring_buffer_per_cpu * cpu_buffer)3189 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3190 {
3191 return local_read(&cpu_buffer->entries) -
3192 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3193 }
3194
3195 /**
3196 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3197 * @buffer: The ring buffer
3198 * @cpu: The per CPU buffer to read from.
3199 */
ring_buffer_oldest_event_ts(struct ring_buffer * buffer,int cpu)3200 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3201 {
3202 unsigned long flags;
3203 struct ring_buffer_per_cpu *cpu_buffer;
3204 struct buffer_page *bpage;
3205 u64 ret = 0;
3206
3207 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3208 return 0;
3209
3210 cpu_buffer = buffer->buffers[cpu];
3211 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3212 /*
3213 * if the tail is on reader_page, oldest time stamp is on the reader
3214 * page
3215 */
3216 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3217 bpage = cpu_buffer->reader_page;
3218 else
3219 bpage = rb_set_head_page(cpu_buffer);
3220 if (bpage)
3221 ret = bpage->page->time_stamp;
3222 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3223
3224 return ret;
3225 }
3226 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3227
3228 /**
3229 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3230 * @buffer: The ring buffer
3231 * @cpu: The per CPU buffer to read from.
3232 */
ring_buffer_bytes_cpu(struct ring_buffer * buffer,int cpu)3233 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3234 {
3235 struct ring_buffer_per_cpu *cpu_buffer;
3236 unsigned long ret;
3237
3238 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3239 return 0;
3240
3241 cpu_buffer = buffer->buffers[cpu];
3242 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3243
3244 return ret;
3245 }
3246 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3247
3248 /**
3249 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3250 * @buffer: The ring buffer
3251 * @cpu: The per CPU buffer to get the entries from.
3252 */
ring_buffer_entries_cpu(struct ring_buffer * buffer,int cpu)3253 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3254 {
3255 struct ring_buffer_per_cpu *cpu_buffer;
3256
3257 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3258 return 0;
3259
3260 cpu_buffer = buffer->buffers[cpu];
3261
3262 return rb_num_of_entries(cpu_buffer);
3263 }
3264 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3265
3266 /**
3267 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3268 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3269 * @buffer: The ring buffer
3270 * @cpu: The per CPU buffer to get the number of overruns from
3271 */
ring_buffer_overrun_cpu(struct ring_buffer * buffer,int cpu)3272 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3273 {
3274 struct ring_buffer_per_cpu *cpu_buffer;
3275 unsigned long ret;
3276
3277 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3278 return 0;
3279
3280 cpu_buffer = buffer->buffers[cpu];
3281 ret = local_read(&cpu_buffer->overrun);
3282
3283 return ret;
3284 }
3285 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3286
3287 /**
3288 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3289 * commits failing due to the buffer wrapping around while there are uncommitted
3290 * events, such as during an interrupt storm.
3291 * @buffer: The ring buffer
3292 * @cpu: The per CPU buffer to get the number of overruns from
3293 */
3294 unsigned long
ring_buffer_commit_overrun_cpu(struct ring_buffer * buffer,int cpu)3295 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3296 {
3297 struct ring_buffer_per_cpu *cpu_buffer;
3298 unsigned long ret;
3299
3300 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3301 return 0;
3302
3303 cpu_buffer = buffer->buffers[cpu];
3304 ret = local_read(&cpu_buffer->commit_overrun);
3305
3306 return ret;
3307 }
3308 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3309
3310 /**
3311 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3312 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3313 * @buffer: The ring buffer
3314 * @cpu: The per CPU buffer to get the number of overruns from
3315 */
3316 unsigned long
ring_buffer_dropped_events_cpu(struct ring_buffer * buffer,int cpu)3317 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3318 {
3319 struct ring_buffer_per_cpu *cpu_buffer;
3320 unsigned long ret;
3321
3322 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3323 return 0;
3324
3325 cpu_buffer = buffer->buffers[cpu];
3326 ret = local_read(&cpu_buffer->dropped_events);
3327
3328 return ret;
3329 }
3330 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3331
3332 /**
3333 * ring_buffer_read_events_cpu - get the number of events successfully read
3334 * @buffer: The ring buffer
3335 * @cpu: The per CPU buffer to get the number of events read
3336 */
3337 unsigned long
ring_buffer_read_events_cpu(struct ring_buffer * buffer,int cpu)3338 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3339 {
3340 struct ring_buffer_per_cpu *cpu_buffer;
3341
3342 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3343 return 0;
3344
3345 cpu_buffer = buffer->buffers[cpu];
3346 return cpu_buffer->read;
3347 }
3348 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3349
3350 /**
3351 * ring_buffer_entries - get the number of entries in a buffer
3352 * @buffer: The ring buffer
3353 *
3354 * Returns the total number of entries in the ring buffer
3355 * (all CPU entries)
3356 */
ring_buffer_entries(struct ring_buffer * buffer)3357 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3358 {
3359 struct ring_buffer_per_cpu *cpu_buffer;
3360 unsigned long entries = 0;
3361 int cpu;
3362
3363 /* if you care about this being correct, lock the buffer */
3364 for_each_buffer_cpu(buffer, cpu) {
3365 cpu_buffer = buffer->buffers[cpu];
3366 entries += rb_num_of_entries(cpu_buffer);
3367 }
3368
3369 return entries;
3370 }
3371 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3372
3373 /**
3374 * ring_buffer_overruns - get the number of overruns in buffer
3375 * @buffer: The ring buffer
3376 *
3377 * Returns the total number of overruns in the ring buffer
3378 * (all CPU entries)
3379 */
ring_buffer_overruns(struct ring_buffer * buffer)3380 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3381 {
3382 struct ring_buffer_per_cpu *cpu_buffer;
3383 unsigned long overruns = 0;
3384 int cpu;
3385
3386 /* if you care about this being correct, lock the buffer */
3387 for_each_buffer_cpu(buffer, cpu) {
3388 cpu_buffer = buffer->buffers[cpu];
3389 overruns += local_read(&cpu_buffer->overrun);
3390 }
3391
3392 return overruns;
3393 }
3394 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3395
rb_iter_reset(struct ring_buffer_iter * iter)3396 static void rb_iter_reset(struct ring_buffer_iter *iter)
3397 {
3398 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3399
3400 /* Iterator usage is expected to have record disabled */
3401 iter->head_page = cpu_buffer->reader_page;
3402 iter->head = cpu_buffer->reader_page->read;
3403
3404 iter->cache_reader_page = iter->head_page;
3405 iter->cache_read = cpu_buffer->read;
3406
3407 if (iter->head)
3408 iter->read_stamp = cpu_buffer->read_stamp;
3409 else
3410 iter->read_stamp = iter->head_page->page->time_stamp;
3411 }
3412
3413 /**
3414 * ring_buffer_iter_reset - reset an iterator
3415 * @iter: The iterator to reset
3416 *
3417 * Resets the iterator, so that it will start from the beginning
3418 * again.
3419 */
ring_buffer_iter_reset(struct ring_buffer_iter * iter)3420 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3421 {
3422 struct ring_buffer_per_cpu *cpu_buffer;
3423 unsigned long flags;
3424
3425 if (!iter)
3426 return;
3427
3428 cpu_buffer = iter->cpu_buffer;
3429
3430 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3431 rb_iter_reset(iter);
3432 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3433 }
3434 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3435
3436 /**
3437 * ring_buffer_iter_empty - check if an iterator has no more to read
3438 * @iter: The iterator to check
3439 */
ring_buffer_iter_empty(struct ring_buffer_iter * iter)3440 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3441 {
3442 struct ring_buffer_per_cpu *cpu_buffer;
3443
3444 cpu_buffer = iter->cpu_buffer;
3445
3446 return iter->head_page == cpu_buffer->commit_page &&
3447 iter->head == rb_commit_index(cpu_buffer);
3448 }
3449 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3450
3451 static void
rb_update_read_stamp(struct ring_buffer_per_cpu * cpu_buffer,struct ring_buffer_event * event)3452 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3453 struct ring_buffer_event *event)
3454 {
3455 u64 delta;
3456
3457 switch (event->type_len) {
3458 case RINGBUF_TYPE_PADDING:
3459 return;
3460
3461 case RINGBUF_TYPE_TIME_EXTEND:
3462 delta = event->array[0];
3463 delta <<= TS_SHIFT;
3464 delta += event->time_delta;
3465 cpu_buffer->read_stamp += delta;
3466 return;
3467
3468 case RINGBUF_TYPE_TIME_STAMP:
3469 /* FIXME: not implemented */
3470 return;
3471
3472 case RINGBUF_TYPE_DATA:
3473 cpu_buffer->read_stamp += event->time_delta;
3474 return;
3475
3476 default:
3477 BUG();
3478 }
3479 return;
3480 }
3481
3482 static void
rb_update_iter_read_stamp(struct ring_buffer_iter * iter,struct ring_buffer_event * event)3483 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3484 struct ring_buffer_event *event)
3485 {
3486 u64 delta;
3487
3488 switch (event->type_len) {
3489 case RINGBUF_TYPE_PADDING:
3490 return;
3491
3492 case RINGBUF_TYPE_TIME_EXTEND:
3493 delta = event->array[0];
3494 delta <<= TS_SHIFT;
3495 delta += event->time_delta;
3496 iter->read_stamp += delta;
3497 return;
3498
3499 case RINGBUF_TYPE_TIME_STAMP:
3500 /* FIXME: not implemented */
3501 return;
3502
3503 case RINGBUF_TYPE_DATA:
3504 iter->read_stamp += event->time_delta;
3505 return;
3506
3507 default:
3508 BUG();
3509 }
3510 return;
3511 }
3512
3513 static struct buffer_page *
rb_get_reader_page(struct ring_buffer_per_cpu * cpu_buffer)3514 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3515 {
3516 struct buffer_page *reader = NULL;
3517 unsigned long overwrite;
3518 unsigned long flags;
3519 int nr_loops = 0;
3520 int ret;
3521
3522 local_irq_save(flags);
3523 arch_spin_lock(&cpu_buffer->lock);
3524
3525 again:
3526 /*
3527 * This should normally only loop twice. But because the
3528 * start of the reader inserts an empty page, it causes
3529 * a case where we will loop three times. There should be no
3530 * reason to loop four times (that I know of).
3531 */
3532 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3533 reader = NULL;
3534 goto out;
3535 }
3536
3537 reader = cpu_buffer->reader_page;
3538
3539 /* If there's more to read, return this page */
3540 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3541 goto out;
3542
3543 /* Never should we have an index greater than the size */
3544 if (RB_WARN_ON(cpu_buffer,
3545 cpu_buffer->reader_page->read > rb_page_size(reader)))
3546 goto out;
3547
3548 /* check if we caught up to the tail */
3549 reader = NULL;
3550 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3551 goto out;
3552
3553 /* Don't bother swapping if the ring buffer is empty */
3554 if (rb_num_of_entries(cpu_buffer) == 0)
3555 goto out;
3556
3557 /*
3558 * Reset the reader page to size zero.
3559 */
3560 local_set(&cpu_buffer->reader_page->write, 0);
3561 local_set(&cpu_buffer->reader_page->entries, 0);
3562 local_set(&cpu_buffer->reader_page->page->commit, 0);
3563 cpu_buffer->reader_page->real_end = 0;
3564
3565 spin:
3566 /*
3567 * Splice the empty reader page into the list around the head.
3568 */
3569 reader = rb_set_head_page(cpu_buffer);
3570 if (!reader)
3571 goto out;
3572 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3573 cpu_buffer->reader_page->list.prev = reader->list.prev;
3574
3575 /*
3576 * cpu_buffer->pages just needs to point to the buffer, it
3577 * has no specific buffer page to point to. Lets move it out
3578 * of our way so we don't accidentally swap it.
3579 */
3580 cpu_buffer->pages = reader->list.prev;
3581
3582 /* The reader page will be pointing to the new head */
3583 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3584
3585 /*
3586 * We want to make sure we read the overruns after we set up our
3587 * pointers to the next object. The writer side does a
3588 * cmpxchg to cross pages which acts as the mb on the writer
3589 * side. Note, the reader will constantly fail the swap
3590 * while the writer is updating the pointers, so this
3591 * guarantees that the overwrite recorded here is the one we
3592 * want to compare with the last_overrun.
3593 */
3594 smp_mb();
3595 overwrite = local_read(&(cpu_buffer->overrun));
3596
3597 /*
3598 * Here's the tricky part.
3599 *
3600 * We need to move the pointer past the header page.
3601 * But we can only do that if a writer is not currently
3602 * moving it. The page before the header page has the
3603 * flag bit '1' set if it is pointing to the page we want.
3604 * but if the writer is in the process of moving it
3605 * than it will be '2' or already moved '0'.
3606 */
3607
3608 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3609
3610 /*
3611 * If we did not convert it, then we must try again.
3612 */
3613 if (!ret)
3614 goto spin;
3615
3616 /*
3617 * Yeah! We succeeded in replacing the page.
3618 *
3619 * Now make the new head point back to the reader page.
3620 */
3621 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3622 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3623
3624 /* Finally update the reader page to the new head */
3625 cpu_buffer->reader_page = reader;
3626 cpu_buffer->reader_page->read = 0;
3627
3628 if (overwrite != cpu_buffer->last_overrun) {
3629 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3630 cpu_buffer->last_overrun = overwrite;
3631 }
3632
3633 goto again;
3634
3635 out:
3636 /* Update the read_stamp on the first event */
3637 if (reader && reader->read == 0)
3638 cpu_buffer->read_stamp = reader->page->time_stamp;
3639
3640 arch_spin_unlock(&cpu_buffer->lock);
3641 local_irq_restore(flags);
3642
3643 return reader;
3644 }
3645
rb_advance_reader(struct ring_buffer_per_cpu * cpu_buffer)3646 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3647 {
3648 struct ring_buffer_event *event;
3649 struct buffer_page *reader;
3650 unsigned length;
3651
3652 reader = rb_get_reader_page(cpu_buffer);
3653
3654 /* This function should not be called when buffer is empty */
3655 if (RB_WARN_ON(cpu_buffer, !reader))
3656 return;
3657
3658 event = rb_reader_event(cpu_buffer);
3659
3660 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3661 cpu_buffer->read++;
3662
3663 rb_update_read_stamp(cpu_buffer, event);
3664
3665 length = rb_event_length(event);
3666 cpu_buffer->reader_page->read += length;
3667 }
3668
rb_advance_iter(struct ring_buffer_iter * iter)3669 static void rb_advance_iter(struct ring_buffer_iter *iter)
3670 {
3671 struct ring_buffer_per_cpu *cpu_buffer;
3672 struct ring_buffer_event *event;
3673 unsigned length;
3674
3675 cpu_buffer = iter->cpu_buffer;
3676
3677 /*
3678 * Check if we are at the end of the buffer.
3679 */
3680 if (iter->head >= rb_page_size(iter->head_page)) {
3681 /* discarded commits can make the page empty */
3682 if (iter->head_page == cpu_buffer->commit_page)
3683 return;
3684 rb_inc_iter(iter);
3685 return;
3686 }
3687
3688 event = rb_iter_head_event(iter);
3689
3690 length = rb_event_length(event);
3691
3692 /*
3693 * This should not be called to advance the header if we are
3694 * at the tail of the buffer.
3695 */
3696 if (RB_WARN_ON(cpu_buffer,
3697 (iter->head_page == cpu_buffer->commit_page) &&
3698 (iter->head + length > rb_commit_index(cpu_buffer))))
3699 return;
3700
3701 rb_update_iter_read_stamp(iter, event);
3702
3703 iter->head += length;
3704
3705 /* check for end of page padding */
3706 if ((iter->head >= rb_page_size(iter->head_page)) &&
3707 (iter->head_page != cpu_buffer->commit_page))
3708 rb_inc_iter(iter);
3709 }
3710
rb_lost_events(struct ring_buffer_per_cpu * cpu_buffer)3711 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3712 {
3713 return cpu_buffer->lost_events;
3714 }
3715
3716 static struct ring_buffer_event *
rb_buffer_peek(struct ring_buffer_per_cpu * cpu_buffer,u64 * ts,unsigned long * lost_events)3717 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3718 unsigned long *lost_events)
3719 {
3720 struct ring_buffer_event *event;
3721 struct buffer_page *reader;
3722 int nr_loops = 0;
3723
3724 again:
3725 /*
3726 * We repeat when a time extend is encountered.
3727 * Since the time extend is always attached to a data event,
3728 * we should never loop more than once.
3729 * (We never hit the following condition more than twice).
3730 */
3731 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3732 return NULL;
3733
3734 reader = rb_get_reader_page(cpu_buffer);
3735 if (!reader)
3736 return NULL;
3737
3738 event = rb_reader_event(cpu_buffer);
3739
3740 switch (event->type_len) {
3741 case RINGBUF_TYPE_PADDING:
3742 if (rb_null_event(event))
3743 RB_WARN_ON(cpu_buffer, 1);
3744 /*
3745 * Because the writer could be discarding every
3746 * event it creates (which would probably be bad)
3747 * if we were to go back to "again" then we may never
3748 * catch up, and will trigger the warn on, or lock
3749 * the box. Return the padding, and we will release
3750 * the current locks, and try again.
3751 */
3752 return event;
3753
3754 case RINGBUF_TYPE_TIME_EXTEND:
3755 /* Internal data, OK to advance */
3756 rb_advance_reader(cpu_buffer);
3757 goto again;
3758
3759 case RINGBUF_TYPE_TIME_STAMP:
3760 /* FIXME: not implemented */
3761 rb_advance_reader(cpu_buffer);
3762 goto again;
3763
3764 case RINGBUF_TYPE_DATA:
3765 if (ts) {
3766 *ts = cpu_buffer->read_stamp + event->time_delta;
3767 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3768 cpu_buffer->cpu, ts);
3769 }
3770 if (lost_events)
3771 *lost_events = rb_lost_events(cpu_buffer);
3772 return event;
3773
3774 default:
3775 BUG();
3776 }
3777
3778 return NULL;
3779 }
3780 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3781
3782 static struct ring_buffer_event *
rb_iter_peek(struct ring_buffer_iter * iter,u64 * ts)3783 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3784 {
3785 struct ring_buffer *buffer;
3786 struct ring_buffer_per_cpu *cpu_buffer;
3787 struct ring_buffer_event *event;
3788 int nr_loops = 0;
3789
3790 cpu_buffer = iter->cpu_buffer;
3791 buffer = cpu_buffer->buffer;
3792
3793 /*
3794 * Check if someone performed a consuming read to
3795 * the buffer. A consuming read invalidates the iterator
3796 * and we need to reset the iterator in this case.
3797 */
3798 if (unlikely(iter->cache_read != cpu_buffer->read ||
3799 iter->cache_reader_page != cpu_buffer->reader_page))
3800 rb_iter_reset(iter);
3801
3802 again:
3803 if (ring_buffer_iter_empty(iter))
3804 return NULL;
3805
3806 /*
3807 * We repeat when a time extend is encountered or we hit
3808 * the end of the page. Since the time extend is always attached
3809 * to a data event, we should never loop more than three times.
3810 * Once for going to next page, once on time extend, and
3811 * finally once to get the event.
3812 * (We never hit the following condition more than thrice).
3813 */
3814 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3815 return NULL;
3816
3817 if (rb_per_cpu_empty(cpu_buffer))
3818 return NULL;
3819
3820 if (iter->head >= rb_page_size(iter->head_page)) {
3821 rb_inc_iter(iter);
3822 goto again;
3823 }
3824
3825 event = rb_iter_head_event(iter);
3826
3827 switch (event->type_len) {
3828 case RINGBUF_TYPE_PADDING:
3829 if (rb_null_event(event)) {
3830 rb_inc_iter(iter);
3831 goto again;
3832 }
3833 rb_advance_iter(iter);
3834 return event;
3835
3836 case RINGBUF_TYPE_TIME_EXTEND:
3837 /* Internal data, OK to advance */
3838 rb_advance_iter(iter);
3839 goto again;
3840
3841 case RINGBUF_TYPE_TIME_STAMP:
3842 /* FIXME: not implemented */
3843 rb_advance_iter(iter);
3844 goto again;
3845
3846 case RINGBUF_TYPE_DATA:
3847 if (ts) {
3848 *ts = iter->read_stamp + event->time_delta;
3849 ring_buffer_normalize_time_stamp(buffer,
3850 cpu_buffer->cpu, ts);
3851 }
3852 return event;
3853
3854 default:
3855 BUG();
3856 }
3857
3858 return NULL;
3859 }
3860 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3861
rb_reader_lock(struct ring_buffer_per_cpu * cpu_buffer)3862 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
3863 {
3864 if (likely(!in_nmi())) {
3865 raw_spin_lock(&cpu_buffer->reader_lock);
3866 return true;
3867 }
3868
3869 /*
3870 * If an NMI die dumps out the content of the ring buffer
3871 * trylock must be used to prevent a deadlock if the NMI
3872 * preempted a task that holds the ring buffer locks. If
3873 * we get the lock then all is fine, if not, then continue
3874 * to do the read, but this can corrupt the ring buffer,
3875 * so it must be permanently disabled from future writes.
3876 * Reading from NMI is a oneshot deal.
3877 */
3878 if (raw_spin_trylock(&cpu_buffer->reader_lock))
3879 return true;
3880
3881 /* Continue without locking, but disable the ring buffer */
3882 atomic_inc(&cpu_buffer->record_disabled);
3883 return false;
3884 }
3885
3886 static inline void
rb_reader_unlock(struct ring_buffer_per_cpu * cpu_buffer,bool locked)3887 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
3888 {
3889 if (likely(locked))
3890 raw_spin_unlock(&cpu_buffer->reader_lock);
3891 return;
3892 }
3893
3894 /**
3895 * ring_buffer_peek - peek at the next event to be read
3896 * @buffer: The ring buffer to read
3897 * @cpu: The cpu to peak at
3898 * @ts: The timestamp counter of this event.
3899 * @lost_events: a variable to store if events were lost (may be NULL)
3900 *
3901 * This will return the event that will be read next, but does
3902 * not consume the data.
3903 */
3904 struct ring_buffer_event *
ring_buffer_peek(struct ring_buffer * buffer,int cpu,u64 * ts,unsigned long * lost_events)3905 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3906 unsigned long *lost_events)
3907 {
3908 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3909 struct ring_buffer_event *event;
3910 unsigned long flags;
3911 bool dolock;
3912
3913 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3914 return NULL;
3915
3916 again:
3917 local_irq_save(flags);
3918 dolock = rb_reader_lock(cpu_buffer);
3919 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3920 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3921 rb_advance_reader(cpu_buffer);
3922 rb_reader_unlock(cpu_buffer, dolock);
3923 local_irq_restore(flags);
3924
3925 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3926 goto again;
3927
3928 return event;
3929 }
3930
3931 /**
3932 * ring_buffer_iter_peek - peek at the next event to be read
3933 * @iter: The ring buffer iterator
3934 * @ts: The timestamp counter of this event.
3935 *
3936 * This will return the event that will be read next, but does
3937 * not increment the iterator.
3938 */
3939 struct ring_buffer_event *
ring_buffer_iter_peek(struct ring_buffer_iter * iter,u64 * ts)3940 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3941 {
3942 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3943 struct ring_buffer_event *event;
3944 unsigned long flags;
3945
3946 again:
3947 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3948 event = rb_iter_peek(iter, ts);
3949 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3950
3951 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3952 goto again;
3953
3954 return event;
3955 }
3956
3957 /**
3958 * ring_buffer_consume - return an event and consume it
3959 * @buffer: The ring buffer to get the next event from
3960 * @cpu: the cpu to read the buffer from
3961 * @ts: a variable to store the timestamp (may be NULL)
3962 * @lost_events: a variable to store if events were lost (may be NULL)
3963 *
3964 * Returns the next event in the ring buffer, and that event is consumed.
3965 * Meaning, that sequential reads will keep returning a different event,
3966 * and eventually empty the ring buffer if the producer is slower.
3967 */
3968 struct ring_buffer_event *
ring_buffer_consume(struct ring_buffer * buffer,int cpu,u64 * ts,unsigned long * lost_events)3969 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3970 unsigned long *lost_events)
3971 {
3972 struct ring_buffer_per_cpu *cpu_buffer;
3973 struct ring_buffer_event *event = NULL;
3974 unsigned long flags;
3975 bool dolock;
3976
3977 again:
3978 /* might be called in atomic */
3979 preempt_disable();
3980
3981 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3982 goto out;
3983
3984 cpu_buffer = buffer->buffers[cpu];
3985 local_irq_save(flags);
3986 dolock = rb_reader_lock(cpu_buffer);
3987
3988 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3989 if (event) {
3990 cpu_buffer->lost_events = 0;
3991 rb_advance_reader(cpu_buffer);
3992 }
3993
3994 rb_reader_unlock(cpu_buffer, dolock);
3995 local_irq_restore(flags);
3996
3997 out:
3998 preempt_enable();
3999
4000 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4001 goto again;
4002
4003 return event;
4004 }
4005 EXPORT_SYMBOL_GPL(ring_buffer_consume);
4006
4007 /**
4008 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4009 * @buffer: The ring buffer to read from
4010 * @cpu: The cpu buffer to iterate over
4011 *
4012 * This performs the initial preparations necessary to iterate
4013 * through the buffer. Memory is allocated, buffer recording
4014 * is disabled, and the iterator pointer is returned to the caller.
4015 *
4016 * Disabling buffer recordng prevents the reading from being
4017 * corrupted. This is not a consuming read, so a producer is not
4018 * expected.
4019 *
4020 * After a sequence of ring_buffer_read_prepare calls, the user is
4021 * expected to make at least one call to ring_buffer_read_prepare_sync.
4022 * Afterwards, ring_buffer_read_start is invoked to get things going
4023 * for real.
4024 *
4025 * This overall must be paired with ring_buffer_read_finish.
4026 */
4027 struct ring_buffer_iter *
ring_buffer_read_prepare(struct ring_buffer * buffer,int cpu)4028 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
4029 {
4030 struct ring_buffer_per_cpu *cpu_buffer;
4031 struct ring_buffer_iter *iter;
4032
4033 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4034 return NULL;
4035
4036 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
4037 if (!iter)
4038 return NULL;
4039
4040 cpu_buffer = buffer->buffers[cpu];
4041
4042 iter->cpu_buffer = cpu_buffer;
4043
4044 atomic_inc(&buffer->resize_disabled);
4045 atomic_inc(&cpu_buffer->record_disabled);
4046
4047 return iter;
4048 }
4049 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4050
4051 /**
4052 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4053 *
4054 * All previously invoked ring_buffer_read_prepare calls to prepare
4055 * iterators will be synchronized. Afterwards, read_buffer_read_start
4056 * calls on those iterators are allowed.
4057 */
4058 void
ring_buffer_read_prepare_sync(void)4059 ring_buffer_read_prepare_sync(void)
4060 {
4061 synchronize_sched();
4062 }
4063 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4064
4065 /**
4066 * ring_buffer_read_start - start a non consuming read of the buffer
4067 * @iter: The iterator returned by ring_buffer_read_prepare
4068 *
4069 * This finalizes the startup of an iteration through the buffer.
4070 * The iterator comes from a call to ring_buffer_read_prepare and
4071 * an intervening ring_buffer_read_prepare_sync must have been
4072 * performed.
4073 *
4074 * Must be paired with ring_buffer_read_finish.
4075 */
4076 void
ring_buffer_read_start(struct ring_buffer_iter * iter)4077 ring_buffer_read_start(struct ring_buffer_iter *iter)
4078 {
4079 struct ring_buffer_per_cpu *cpu_buffer;
4080 unsigned long flags;
4081
4082 if (!iter)
4083 return;
4084
4085 cpu_buffer = iter->cpu_buffer;
4086
4087 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4088 arch_spin_lock(&cpu_buffer->lock);
4089 rb_iter_reset(iter);
4090 arch_spin_unlock(&cpu_buffer->lock);
4091 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4092 }
4093 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4094
4095 /**
4096 * ring_buffer_read_finish - finish reading the iterator of the buffer
4097 * @iter: The iterator retrieved by ring_buffer_start
4098 *
4099 * This re-enables the recording to the buffer, and frees the
4100 * iterator.
4101 */
4102 void
ring_buffer_read_finish(struct ring_buffer_iter * iter)4103 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4104 {
4105 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4106 unsigned long flags;
4107
4108 /*
4109 * Ring buffer is disabled from recording, here's a good place
4110 * to check the integrity of the ring buffer.
4111 * Must prevent readers from trying to read, as the check
4112 * clears the HEAD page and readers require it.
4113 */
4114 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4115 rb_check_pages(cpu_buffer);
4116 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4117
4118 atomic_dec(&cpu_buffer->record_disabled);
4119 atomic_dec(&cpu_buffer->buffer->resize_disabled);
4120 kfree(iter);
4121 }
4122 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4123
4124 /**
4125 * ring_buffer_read - read the next item in the ring buffer by the iterator
4126 * @iter: The ring buffer iterator
4127 * @ts: The time stamp of the event read.
4128 *
4129 * This reads the next event in the ring buffer and increments the iterator.
4130 */
4131 struct ring_buffer_event *
ring_buffer_read(struct ring_buffer_iter * iter,u64 * ts)4132 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4133 {
4134 struct ring_buffer_event *event;
4135 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4136 unsigned long flags;
4137
4138 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4139 again:
4140 event = rb_iter_peek(iter, ts);
4141 if (!event)
4142 goto out;
4143
4144 if (event->type_len == RINGBUF_TYPE_PADDING)
4145 goto again;
4146
4147 rb_advance_iter(iter);
4148 out:
4149 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4150
4151 return event;
4152 }
4153 EXPORT_SYMBOL_GPL(ring_buffer_read);
4154
4155 /**
4156 * ring_buffer_size - return the size of the ring buffer (in bytes)
4157 * @buffer: The ring buffer.
4158 */
ring_buffer_size(struct ring_buffer * buffer,int cpu)4159 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4160 {
4161 /*
4162 * Earlier, this method returned
4163 * BUF_PAGE_SIZE * buffer->nr_pages
4164 * Since the nr_pages field is now removed, we have converted this to
4165 * return the per cpu buffer value.
4166 */
4167 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4168 return 0;
4169
4170 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4171 }
4172 EXPORT_SYMBOL_GPL(ring_buffer_size);
4173
4174 static void
rb_reset_cpu(struct ring_buffer_per_cpu * cpu_buffer)4175 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4176 {
4177 rb_head_page_deactivate(cpu_buffer);
4178
4179 cpu_buffer->head_page
4180 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4181 local_set(&cpu_buffer->head_page->write, 0);
4182 local_set(&cpu_buffer->head_page->entries, 0);
4183 local_set(&cpu_buffer->head_page->page->commit, 0);
4184
4185 cpu_buffer->head_page->read = 0;
4186
4187 cpu_buffer->tail_page = cpu_buffer->head_page;
4188 cpu_buffer->commit_page = cpu_buffer->head_page;
4189
4190 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4191 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4192 local_set(&cpu_buffer->reader_page->write, 0);
4193 local_set(&cpu_buffer->reader_page->entries, 0);
4194 local_set(&cpu_buffer->reader_page->page->commit, 0);
4195 cpu_buffer->reader_page->read = 0;
4196
4197 local_set(&cpu_buffer->entries_bytes, 0);
4198 local_set(&cpu_buffer->overrun, 0);
4199 local_set(&cpu_buffer->commit_overrun, 0);
4200 local_set(&cpu_buffer->dropped_events, 0);
4201 local_set(&cpu_buffer->entries, 0);
4202 local_set(&cpu_buffer->committing, 0);
4203 local_set(&cpu_buffer->commits, 0);
4204 cpu_buffer->read = 0;
4205 cpu_buffer->read_bytes = 0;
4206
4207 cpu_buffer->write_stamp = 0;
4208 cpu_buffer->read_stamp = 0;
4209
4210 cpu_buffer->lost_events = 0;
4211 cpu_buffer->last_overrun = 0;
4212
4213 rb_head_page_activate(cpu_buffer);
4214 }
4215
4216 /**
4217 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4218 * @buffer: The ring buffer to reset a per cpu buffer of
4219 * @cpu: The CPU buffer to be reset
4220 */
ring_buffer_reset_cpu(struct ring_buffer * buffer,int cpu)4221 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4222 {
4223 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4224 unsigned long flags;
4225
4226 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4227 return;
4228
4229 atomic_inc(&buffer->resize_disabled);
4230 atomic_inc(&cpu_buffer->record_disabled);
4231
4232 /* Make sure all commits have finished */
4233 synchronize_sched();
4234
4235 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4236
4237 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4238 goto out;
4239
4240 arch_spin_lock(&cpu_buffer->lock);
4241
4242 rb_reset_cpu(cpu_buffer);
4243
4244 arch_spin_unlock(&cpu_buffer->lock);
4245
4246 out:
4247 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4248
4249 atomic_dec(&cpu_buffer->record_disabled);
4250 atomic_dec(&buffer->resize_disabled);
4251 }
4252 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4253
4254 /**
4255 * ring_buffer_reset - reset a ring buffer
4256 * @buffer: The ring buffer to reset all cpu buffers
4257 */
ring_buffer_reset(struct ring_buffer * buffer)4258 void ring_buffer_reset(struct ring_buffer *buffer)
4259 {
4260 int cpu;
4261
4262 for_each_buffer_cpu(buffer, cpu)
4263 ring_buffer_reset_cpu(buffer, cpu);
4264 }
4265 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4266
4267 /**
4268 * rind_buffer_empty - is the ring buffer empty?
4269 * @buffer: The ring buffer to test
4270 */
ring_buffer_empty(struct ring_buffer * buffer)4271 bool ring_buffer_empty(struct ring_buffer *buffer)
4272 {
4273 struct ring_buffer_per_cpu *cpu_buffer;
4274 unsigned long flags;
4275 bool dolock;
4276 int cpu;
4277 int ret;
4278
4279 /* yes this is racy, but if you don't like the race, lock the buffer */
4280 for_each_buffer_cpu(buffer, cpu) {
4281 cpu_buffer = buffer->buffers[cpu];
4282 local_irq_save(flags);
4283 dolock = rb_reader_lock(cpu_buffer);
4284 ret = rb_per_cpu_empty(cpu_buffer);
4285 rb_reader_unlock(cpu_buffer, dolock);
4286 local_irq_restore(flags);
4287
4288 if (!ret)
4289 return false;
4290 }
4291
4292 return true;
4293 }
4294 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4295
4296 /**
4297 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4298 * @buffer: The ring buffer
4299 * @cpu: The CPU buffer to test
4300 */
ring_buffer_empty_cpu(struct ring_buffer * buffer,int cpu)4301 bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4302 {
4303 struct ring_buffer_per_cpu *cpu_buffer;
4304 unsigned long flags;
4305 bool dolock;
4306 int ret;
4307
4308 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4309 return true;
4310
4311 cpu_buffer = buffer->buffers[cpu];
4312 local_irq_save(flags);
4313 dolock = rb_reader_lock(cpu_buffer);
4314 ret = rb_per_cpu_empty(cpu_buffer);
4315 rb_reader_unlock(cpu_buffer, dolock);
4316 local_irq_restore(flags);
4317
4318 return ret;
4319 }
4320 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4321
4322 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4323 /**
4324 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4325 * @buffer_a: One buffer to swap with
4326 * @buffer_b: The other buffer to swap with
4327 *
4328 * This function is useful for tracers that want to take a "snapshot"
4329 * of a CPU buffer and has another back up buffer lying around.
4330 * it is expected that the tracer handles the cpu buffer not being
4331 * used at the moment.
4332 */
ring_buffer_swap_cpu(struct ring_buffer * buffer_a,struct ring_buffer * buffer_b,int cpu)4333 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4334 struct ring_buffer *buffer_b, int cpu)
4335 {
4336 struct ring_buffer_per_cpu *cpu_buffer_a;
4337 struct ring_buffer_per_cpu *cpu_buffer_b;
4338 int ret = -EINVAL;
4339
4340 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4341 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4342 goto out;
4343
4344 cpu_buffer_a = buffer_a->buffers[cpu];
4345 cpu_buffer_b = buffer_b->buffers[cpu];
4346
4347 /* At least make sure the two buffers are somewhat the same */
4348 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4349 goto out;
4350
4351 ret = -EAGAIN;
4352
4353 if (atomic_read(&buffer_a->record_disabled))
4354 goto out;
4355
4356 if (atomic_read(&buffer_b->record_disabled))
4357 goto out;
4358
4359 if (atomic_read(&cpu_buffer_a->record_disabled))
4360 goto out;
4361
4362 if (atomic_read(&cpu_buffer_b->record_disabled))
4363 goto out;
4364
4365 /*
4366 * We can't do a synchronize_sched here because this
4367 * function can be called in atomic context.
4368 * Normally this will be called from the same CPU as cpu.
4369 * If not it's up to the caller to protect this.
4370 */
4371 atomic_inc(&cpu_buffer_a->record_disabled);
4372 atomic_inc(&cpu_buffer_b->record_disabled);
4373
4374 ret = -EBUSY;
4375 if (local_read(&cpu_buffer_a->committing))
4376 goto out_dec;
4377 if (local_read(&cpu_buffer_b->committing))
4378 goto out_dec;
4379
4380 buffer_a->buffers[cpu] = cpu_buffer_b;
4381 buffer_b->buffers[cpu] = cpu_buffer_a;
4382
4383 cpu_buffer_b->buffer = buffer_a;
4384 cpu_buffer_a->buffer = buffer_b;
4385
4386 ret = 0;
4387
4388 out_dec:
4389 atomic_dec(&cpu_buffer_a->record_disabled);
4390 atomic_dec(&cpu_buffer_b->record_disabled);
4391 out:
4392 return ret;
4393 }
4394 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4395 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4396
4397 /**
4398 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4399 * @buffer: the buffer to allocate for.
4400 * @cpu: the cpu buffer to allocate.
4401 *
4402 * This function is used in conjunction with ring_buffer_read_page.
4403 * When reading a full page from the ring buffer, these functions
4404 * can be used to speed up the process. The calling function should
4405 * allocate a few pages first with this function. Then when it
4406 * needs to get pages from the ring buffer, it passes the result
4407 * of this function into ring_buffer_read_page, which will swap
4408 * the page that was allocated, with the read page of the buffer.
4409 *
4410 * Returns:
4411 * The page allocated, or NULL on error.
4412 */
ring_buffer_alloc_read_page(struct ring_buffer * buffer,int cpu)4413 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4414 {
4415 struct buffer_data_page *bpage;
4416 struct page *page;
4417
4418 page = alloc_pages_node(cpu_to_node(cpu),
4419 GFP_KERNEL | __GFP_NORETRY, 0);
4420 if (!page)
4421 return NULL;
4422
4423 bpage = page_address(page);
4424
4425 rb_init_page(bpage);
4426
4427 return bpage;
4428 }
4429 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4430
4431 /**
4432 * ring_buffer_free_read_page - free an allocated read page
4433 * @buffer: the buffer the page was allocate for
4434 * @data: the page to free
4435 *
4436 * Free a page allocated from ring_buffer_alloc_read_page.
4437 */
ring_buffer_free_read_page(struct ring_buffer * buffer,void * data)4438 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4439 {
4440 free_page((unsigned long)data);
4441 }
4442 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4443
4444 /**
4445 * ring_buffer_read_page - extract a page from the ring buffer
4446 * @buffer: buffer to extract from
4447 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4448 * @len: amount to extract
4449 * @cpu: the cpu of the buffer to extract
4450 * @full: should the extraction only happen when the page is full.
4451 *
4452 * This function will pull out a page from the ring buffer and consume it.
4453 * @data_page must be the address of the variable that was returned
4454 * from ring_buffer_alloc_read_page. This is because the page might be used
4455 * to swap with a page in the ring buffer.
4456 *
4457 * for example:
4458 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4459 * if (!rpage)
4460 * return error;
4461 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4462 * if (ret >= 0)
4463 * process_page(rpage, ret);
4464 *
4465 * When @full is set, the function will not return true unless
4466 * the writer is off the reader page.
4467 *
4468 * Note: it is up to the calling functions to handle sleeps and wakeups.
4469 * The ring buffer can be used anywhere in the kernel and can not
4470 * blindly call wake_up. The layer that uses the ring buffer must be
4471 * responsible for that.
4472 *
4473 * Returns:
4474 * >=0 if data has been transferred, returns the offset of consumed data.
4475 * <0 if no data has been transferred.
4476 */
ring_buffer_read_page(struct ring_buffer * buffer,void ** data_page,size_t len,int cpu,int full)4477 int ring_buffer_read_page(struct ring_buffer *buffer,
4478 void **data_page, size_t len, int cpu, int full)
4479 {
4480 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4481 struct ring_buffer_event *event;
4482 struct buffer_data_page *bpage;
4483 struct buffer_page *reader;
4484 unsigned long missed_events;
4485 unsigned long flags;
4486 unsigned int commit;
4487 unsigned int read;
4488 u64 save_timestamp;
4489 int ret = -1;
4490
4491 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4492 goto out;
4493
4494 /*
4495 * If len is not big enough to hold the page header, then
4496 * we can not copy anything.
4497 */
4498 if (len <= BUF_PAGE_HDR_SIZE)
4499 goto out;
4500
4501 len -= BUF_PAGE_HDR_SIZE;
4502
4503 if (!data_page)
4504 goto out;
4505
4506 bpage = *data_page;
4507 if (!bpage)
4508 goto out;
4509
4510 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4511
4512 reader = rb_get_reader_page(cpu_buffer);
4513 if (!reader)
4514 goto out_unlock;
4515
4516 event = rb_reader_event(cpu_buffer);
4517
4518 read = reader->read;
4519 commit = rb_page_commit(reader);
4520
4521 /* Check if any events were dropped */
4522 missed_events = cpu_buffer->lost_events;
4523
4524 /*
4525 * If this page has been partially read or
4526 * if len is not big enough to read the rest of the page or
4527 * a writer is still on the page, then
4528 * we must copy the data from the page to the buffer.
4529 * Otherwise, we can simply swap the page with the one passed in.
4530 */
4531 if (read || (len < (commit - read)) ||
4532 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4533 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4534 unsigned int rpos = read;
4535 unsigned int pos = 0;
4536 unsigned int size;
4537
4538 if (full)
4539 goto out_unlock;
4540
4541 if (len > (commit - read))
4542 len = (commit - read);
4543
4544 /* Always keep the time extend and data together */
4545 size = rb_event_ts_length(event);
4546
4547 if (len < size)
4548 goto out_unlock;
4549
4550 /* save the current timestamp, since the user will need it */
4551 save_timestamp = cpu_buffer->read_stamp;
4552
4553 /* Need to copy one event at a time */
4554 do {
4555 /* We need the size of one event, because
4556 * rb_advance_reader only advances by one event,
4557 * whereas rb_event_ts_length may include the size of
4558 * one or two events.
4559 * We have already ensured there's enough space if this
4560 * is a time extend. */
4561 size = rb_event_length(event);
4562 memcpy(bpage->data + pos, rpage->data + rpos, size);
4563
4564 len -= size;
4565
4566 rb_advance_reader(cpu_buffer);
4567 rpos = reader->read;
4568 pos += size;
4569
4570 if (rpos >= commit)
4571 break;
4572
4573 event = rb_reader_event(cpu_buffer);
4574 /* Always keep the time extend and data together */
4575 size = rb_event_ts_length(event);
4576 } while (len >= size);
4577
4578 /* update bpage */
4579 local_set(&bpage->commit, pos);
4580 bpage->time_stamp = save_timestamp;
4581
4582 /* we copied everything to the beginning */
4583 read = 0;
4584 } else {
4585 /* update the entry counter */
4586 cpu_buffer->read += rb_page_entries(reader);
4587 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4588
4589 /* swap the pages */
4590 rb_init_page(bpage);
4591 bpage = reader->page;
4592 reader->page = *data_page;
4593 local_set(&reader->write, 0);
4594 local_set(&reader->entries, 0);
4595 reader->read = 0;
4596 *data_page = bpage;
4597
4598 /*
4599 * Use the real_end for the data size,
4600 * This gives us a chance to store the lost events
4601 * on the page.
4602 */
4603 if (reader->real_end)
4604 local_set(&bpage->commit, reader->real_end);
4605 }
4606 ret = read;
4607
4608 cpu_buffer->lost_events = 0;
4609
4610 commit = local_read(&bpage->commit);
4611 /*
4612 * Set a flag in the commit field if we lost events
4613 */
4614 if (missed_events) {
4615 /* If there is room at the end of the page to save the
4616 * missed events, then record it there.
4617 */
4618 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4619 memcpy(&bpage->data[commit], &missed_events,
4620 sizeof(missed_events));
4621 local_add(RB_MISSED_STORED, &bpage->commit);
4622 commit += sizeof(missed_events);
4623 }
4624 local_add(RB_MISSED_EVENTS, &bpage->commit);
4625 }
4626
4627 /*
4628 * This page may be off to user land. Zero it out here.
4629 */
4630 if (commit < BUF_PAGE_SIZE)
4631 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4632
4633 out_unlock:
4634 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4635
4636 out:
4637 return ret;
4638 }
4639 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4640
4641 #ifdef CONFIG_HOTPLUG_CPU
rb_cpu_notify(struct notifier_block * self,unsigned long action,void * hcpu)4642 static int rb_cpu_notify(struct notifier_block *self,
4643 unsigned long action, void *hcpu)
4644 {
4645 struct ring_buffer *buffer =
4646 container_of(self, struct ring_buffer, cpu_notify);
4647 long cpu = (long)hcpu;
4648 long nr_pages_same;
4649 int cpu_i;
4650 unsigned long nr_pages;
4651
4652 switch (action) {
4653 case CPU_UP_PREPARE:
4654 case CPU_UP_PREPARE_FROZEN:
4655 if (cpumask_test_cpu(cpu, buffer->cpumask))
4656 return NOTIFY_OK;
4657
4658 nr_pages = 0;
4659 nr_pages_same = 1;
4660 /* check if all cpu sizes are same */
4661 for_each_buffer_cpu(buffer, cpu_i) {
4662 /* fill in the size from first enabled cpu */
4663 if (nr_pages == 0)
4664 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4665 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4666 nr_pages_same = 0;
4667 break;
4668 }
4669 }
4670 /* allocate minimum pages, user can later expand it */
4671 if (!nr_pages_same)
4672 nr_pages = 2;
4673 buffer->buffers[cpu] =
4674 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4675 if (!buffer->buffers[cpu]) {
4676 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4677 cpu);
4678 return NOTIFY_OK;
4679 }
4680 smp_wmb();
4681 cpumask_set_cpu(cpu, buffer->cpumask);
4682 break;
4683 case CPU_DOWN_PREPARE:
4684 case CPU_DOWN_PREPARE_FROZEN:
4685 /*
4686 * Do nothing.
4687 * If we were to free the buffer, then the user would
4688 * lose any trace that was in the buffer.
4689 */
4690 break;
4691 default:
4692 break;
4693 }
4694 return NOTIFY_OK;
4695 }
4696 #endif
4697
4698 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4699 /*
4700 * This is a basic integrity check of the ring buffer.
4701 * Late in the boot cycle this test will run when configured in.
4702 * It will kick off a thread per CPU that will go into a loop
4703 * writing to the per cpu ring buffer various sizes of data.
4704 * Some of the data will be large items, some small.
4705 *
4706 * Another thread is created that goes into a spin, sending out
4707 * IPIs to the other CPUs to also write into the ring buffer.
4708 * this is to test the nesting ability of the buffer.
4709 *
4710 * Basic stats are recorded and reported. If something in the
4711 * ring buffer should happen that's not expected, a big warning
4712 * is displayed and all ring buffers are disabled.
4713 */
4714 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4715
4716 struct rb_test_data {
4717 struct ring_buffer *buffer;
4718 unsigned long events;
4719 unsigned long bytes_written;
4720 unsigned long bytes_alloc;
4721 unsigned long bytes_dropped;
4722 unsigned long events_nested;
4723 unsigned long bytes_written_nested;
4724 unsigned long bytes_alloc_nested;
4725 unsigned long bytes_dropped_nested;
4726 int min_size_nested;
4727 int max_size_nested;
4728 int max_size;
4729 int min_size;
4730 int cpu;
4731 int cnt;
4732 };
4733
4734 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4735
4736 /* 1 meg per cpu */
4737 #define RB_TEST_BUFFER_SIZE 1048576
4738
4739 static char rb_string[] __initdata =
4740 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4741 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4742 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4743
4744 static bool rb_test_started __initdata;
4745
4746 struct rb_item {
4747 int size;
4748 char str[];
4749 };
4750
rb_write_something(struct rb_test_data * data,bool nested)4751 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4752 {
4753 struct ring_buffer_event *event;
4754 struct rb_item *item;
4755 bool started;
4756 int event_len;
4757 int size;
4758 int len;
4759 int cnt;
4760
4761 /* Have nested writes different that what is written */
4762 cnt = data->cnt + (nested ? 27 : 0);
4763
4764 /* Multiply cnt by ~e, to make some unique increment */
4765 size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4766
4767 len = size + sizeof(struct rb_item);
4768
4769 started = rb_test_started;
4770 /* read rb_test_started before checking buffer enabled */
4771 smp_rmb();
4772
4773 event = ring_buffer_lock_reserve(data->buffer, len);
4774 if (!event) {
4775 /* Ignore dropped events before test starts. */
4776 if (started) {
4777 if (nested)
4778 data->bytes_dropped += len;
4779 else
4780 data->bytes_dropped_nested += len;
4781 }
4782 return len;
4783 }
4784
4785 event_len = ring_buffer_event_length(event);
4786
4787 if (RB_WARN_ON(data->buffer, event_len < len))
4788 goto out;
4789
4790 item = ring_buffer_event_data(event);
4791 item->size = size;
4792 memcpy(item->str, rb_string, size);
4793
4794 if (nested) {
4795 data->bytes_alloc_nested += event_len;
4796 data->bytes_written_nested += len;
4797 data->events_nested++;
4798 if (!data->min_size_nested || len < data->min_size_nested)
4799 data->min_size_nested = len;
4800 if (len > data->max_size_nested)
4801 data->max_size_nested = len;
4802 } else {
4803 data->bytes_alloc += event_len;
4804 data->bytes_written += len;
4805 data->events++;
4806 if (!data->min_size || len < data->min_size)
4807 data->max_size = len;
4808 if (len > data->max_size)
4809 data->max_size = len;
4810 }
4811
4812 out:
4813 ring_buffer_unlock_commit(data->buffer, event);
4814
4815 return 0;
4816 }
4817
rb_test(void * arg)4818 static __init int rb_test(void *arg)
4819 {
4820 struct rb_test_data *data = arg;
4821
4822 while (!kthread_should_stop()) {
4823 rb_write_something(data, false);
4824 data->cnt++;
4825
4826 set_current_state(TASK_INTERRUPTIBLE);
4827 /* Now sleep between a min of 100-300us and a max of 1ms */
4828 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4829 }
4830
4831 return 0;
4832 }
4833
rb_ipi(void * ignore)4834 static __init void rb_ipi(void *ignore)
4835 {
4836 struct rb_test_data *data;
4837 int cpu = smp_processor_id();
4838
4839 data = &rb_data[cpu];
4840 rb_write_something(data, true);
4841 }
4842
rb_hammer_test(void * arg)4843 static __init int rb_hammer_test(void *arg)
4844 {
4845 while (!kthread_should_stop()) {
4846
4847 /* Send an IPI to all cpus to write data! */
4848 smp_call_function(rb_ipi, NULL, 1);
4849 /* No sleep, but for non preempt, let others run */
4850 schedule();
4851 }
4852
4853 return 0;
4854 }
4855
test_ringbuffer(void)4856 static __init int test_ringbuffer(void)
4857 {
4858 struct task_struct *rb_hammer;
4859 struct ring_buffer *buffer;
4860 int cpu;
4861 int ret = 0;
4862
4863 pr_info("Running ring buffer tests...\n");
4864
4865 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4866 if (WARN_ON(!buffer))
4867 return 0;
4868
4869 /* Disable buffer so that threads can't write to it yet */
4870 ring_buffer_record_off(buffer);
4871
4872 for_each_online_cpu(cpu) {
4873 rb_data[cpu].buffer = buffer;
4874 rb_data[cpu].cpu = cpu;
4875 rb_data[cpu].cnt = cpu;
4876 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4877 "rbtester/%d", cpu);
4878 if (WARN_ON(!rb_threads[cpu])) {
4879 pr_cont("FAILED\n");
4880 ret = -1;
4881 goto out_free;
4882 }
4883
4884 kthread_bind(rb_threads[cpu], cpu);
4885 wake_up_process(rb_threads[cpu]);
4886 }
4887
4888 /* Now create the rb hammer! */
4889 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4890 if (WARN_ON(!rb_hammer)) {
4891 pr_cont("FAILED\n");
4892 ret = -1;
4893 goto out_free;
4894 }
4895
4896 ring_buffer_record_on(buffer);
4897 /*
4898 * Show buffer is enabled before setting rb_test_started.
4899 * Yes there's a small race window where events could be
4900 * dropped and the thread wont catch it. But when a ring
4901 * buffer gets enabled, there will always be some kind of
4902 * delay before other CPUs see it. Thus, we don't care about
4903 * those dropped events. We care about events dropped after
4904 * the threads see that the buffer is active.
4905 */
4906 smp_wmb();
4907 rb_test_started = true;
4908
4909 set_current_state(TASK_INTERRUPTIBLE);
4910 /* Just run for 10 seconds */;
4911 schedule_timeout(10 * HZ);
4912
4913 kthread_stop(rb_hammer);
4914
4915 out_free:
4916 for_each_online_cpu(cpu) {
4917 if (!rb_threads[cpu])
4918 break;
4919 kthread_stop(rb_threads[cpu]);
4920 }
4921 if (ret) {
4922 ring_buffer_free(buffer);
4923 return ret;
4924 }
4925
4926 /* Report! */
4927 pr_info("finished\n");
4928 for_each_online_cpu(cpu) {
4929 struct ring_buffer_event *event;
4930 struct rb_test_data *data = &rb_data[cpu];
4931 struct rb_item *item;
4932 unsigned long total_events;
4933 unsigned long total_dropped;
4934 unsigned long total_written;
4935 unsigned long total_alloc;
4936 unsigned long total_read = 0;
4937 unsigned long total_size = 0;
4938 unsigned long total_len = 0;
4939 unsigned long total_lost = 0;
4940 unsigned long lost;
4941 int big_event_size;
4942 int small_event_size;
4943
4944 ret = -1;
4945
4946 total_events = data->events + data->events_nested;
4947 total_written = data->bytes_written + data->bytes_written_nested;
4948 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
4949 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
4950
4951 big_event_size = data->max_size + data->max_size_nested;
4952 small_event_size = data->min_size + data->min_size_nested;
4953
4954 pr_info("CPU %d:\n", cpu);
4955 pr_info(" events: %ld\n", total_events);
4956 pr_info(" dropped bytes: %ld\n", total_dropped);
4957 pr_info(" alloced bytes: %ld\n", total_alloc);
4958 pr_info(" written bytes: %ld\n", total_written);
4959 pr_info(" biggest event: %d\n", big_event_size);
4960 pr_info(" smallest event: %d\n", small_event_size);
4961
4962 if (RB_WARN_ON(buffer, total_dropped))
4963 break;
4964
4965 ret = 0;
4966
4967 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
4968 total_lost += lost;
4969 item = ring_buffer_event_data(event);
4970 total_len += ring_buffer_event_length(event);
4971 total_size += item->size + sizeof(struct rb_item);
4972 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
4973 pr_info("FAILED!\n");
4974 pr_info("buffer had: %.*s\n", item->size, item->str);
4975 pr_info("expected: %.*s\n", item->size, rb_string);
4976 RB_WARN_ON(buffer, 1);
4977 ret = -1;
4978 break;
4979 }
4980 total_read++;
4981 }
4982 if (ret)
4983 break;
4984
4985 ret = -1;
4986
4987 pr_info(" read events: %ld\n", total_read);
4988 pr_info(" lost events: %ld\n", total_lost);
4989 pr_info(" total events: %ld\n", total_lost + total_read);
4990 pr_info(" recorded len bytes: %ld\n", total_len);
4991 pr_info(" recorded size bytes: %ld\n", total_size);
4992 if (total_lost)
4993 pr_info(" With dropped events, record len and size may not match\n"
4994 " alloced and written from above\n");
4995 if (!total_lost) {
4996 if (RB_WARN_ON(buffer, total_len != total_alloc ||
4997 total_size != total_written))
4998 break;
4999 }
5000 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5001 break;
5002
5003 ret = 0;
5004 }
5005 if (!ret)
5006 pr_info("Ring buffer PASSED!\n");
5007
5008 ring_buffer_free(buffer);
5009 return 0;
5010 }
5011
5012 late_initcall(test_ringbuffer);
5013 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */
5014