Loading...
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
28static void update_pages_handler(struct work_struct *work);
29
30/*
31 * The ring buffer header is special. We must manually up keep it.
32 */
33int 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
141enum {
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
149static inline int rb_null_event(struct ring_buffer_event *event)
150{
151 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
152}
153
154static 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
161static unsigned
162rb_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 */
178static inline unsigned
179rb_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 */
207static inline unsigned
208rb_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 */
230unsigned 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}
245EXPORT_SYMBOL_GPL(ring_buffer_event_length);
246
247/* inline for ring buffer fast paths */
248static void *
249rb_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 */
265void *ring_buffer_event_data(struct ring_buffer_event *event)
266{
267 return rb_event_data(event);
268}
269EXPORT_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
283struct 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 */
297struct 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
321static 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 */
332size_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 */
342static 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 */
351static 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
363int 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
393struct 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 */
405struct 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 */
422enum {
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 */
433struct 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 int 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 int 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
469struct 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
490struct 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 */
505static 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 */
526int 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 */
630int 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
688static 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
694u64 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}
704EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
705
706void 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}
712EXPORT_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 */
796static 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 */
811static inline int
812rb_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 */
832static 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 */
842static 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 */
855static 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
869static 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 */
879static void
880rb_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
891static 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
914static 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
923static 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
932static 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
941static 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
949static struct buffer_page *
950rb_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
987static 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 */
1005static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1006 struct buffer_page *tail_page,
1007 struct buffer_page *next_page)
1008{
1009 unsigned long old_entries;
1010 unsigned long old_write;
1011
1012 /*
1013 * The tail page now needs to be moved forward.
1014 *
1015 * We need to reset the tail page, but without messing
1016 * with possible erasing of data brought in by interrupts
1017 * that have moved the tail page and are currently on it.
1018 *
1019 * We add a counter to the write field to denote this.
1020 */
1021 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1022 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1023
1024 /*
1025 * Just make sure we have seen our old_write and synchronize
1026 * with any interrupts that come in.
1027 */
1028 barrier();
1029
1030 /*
1031 * If the tail page is still the same as what we think
1032 * it is, then it is up to us to update the tail
1033 * pointer.
1034 */
1035 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1036 /* Zero the write counter */
1037 unsigned long val = old_write & ~RB_WRITE_MASK;
1038 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1039
1040 /*
1041 * This will only succeed if an interrupt did
1042 * not come in and change it. In which case, we
1043 * do not want to modify it.
1044 *
1045 * We add (void) to let the compiler know that we do not care
1046 * about the return value of these functions. We use the
1047 * cmpxchg to only update if an interrupt did not already
1048 * do it for us. If the cmpxchg fails, we don't care.
1049 */
1050 (void)local_cmpxchg(&next_page->write, old_write, val);
1051 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1052
1053 /*
1054 * No need to worry about races with clearing out the commit.
1055 * it only can increment when a commit takes place. But that
1056 * only happens in the outer most nested commit.
1057 */
1058 local_set(&next_page->page->commit, 0);
1059
1060 /* Again, either we update tail_page or an interrupt does */
1061 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1062 }
1063}
1064
1065static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1066 struct buffer_page *bpage)
1067{
1068 unsigned long val = (unsigned long)bpage;
1069
1070 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1071 return 1;
1072
1073 return 0;
1074}
1075
1076/**
1077 * rb_check_list - make sure a pointer to a list has the last bits zero
1078 */
1079static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1080 struct list_head *list)
1081{
1082 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1083 return 1;
1084 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1085 return 1;
1086 return 0;
1087}
1088
1089/**
1090 * rb_check_pages - integrity check of buffer pages
1091 * @cpu_buffer: CPU buffer with pages to test
1092 *
1093 * As a safety measure we check to make sure the data pages have not
1094 * been corrupted.
1095 */
1096static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1097{
1098 struct list_head *head = cpu_buffer->pages;
1099 struct buffer_page *bpage, *tmp;
1100
1101 /* Reset the head page if it exists */
1102 if (cpu_buffer->head_page)
1103 rb_set_head_page(cpu_buffer);
1104
1105 rb_head_page_deactivate(cpu_buffer);
1106
1107 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1108 return -1;
1109 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1110 return -1;
1111
1112 if (rb_check_list(cpu_buffer, head))
1113 return -1;
1114
1115 list_for_each_entry_safe(bpage, tmp, head, list) {
1116 if (RB_WARN_ON(cpu_buffer,
1117 bpage->list.next->prev != &bpage->list))
1118 return -1;
1119 if (RB_WARN_ON(cpu_buffer,
1120 bpage->list.prev->next != &bpage->list))
1121 return -1;
1122 if (rb_check_list(cpu_buffer, &bpage->list))
1123 return -1;
1124 }
1125
1126 rb_head_page_activate(cpu_buffer);
1127
1128 return 0;
1129}
1130
1131static int __rb_allocate_pages(int nr_pages, struct list_head *pages, int cpu)
1132{
1133 int i;
1134 struct buffer_page *bpage, *tmp;
1135
1136 for (i = 0; i < nr_pages; i++) {
1137 struct page *page;
1138 /*
1139 * __GFP_NORETRY flag makes sure that the allocation fails
1140 * gracefully without invoking oom-killer and the system is
1141 * not destabilized.
1142 */
1143 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1144 GFP_KERNEL | __GFP_NORETRY,
1145 cpu_to_node(cpu));
1146 if (!bpage)
1147 goto free_pages;
1148
1149 list_add(&bpage->list, pages);
1150
1151 page = alloc_pages_node(cpu_to_node(cpu),
1152 GFP_KERNEL | __GFP_NORETRY, 0);
1153 if (!page)
1154 goto free_pages;
1155 bpage->page = page_address(page);
1156 rb_init_page(bpage->page);
1157 }
1158
1159 return 0;
1160
1161free_pages:
1162 list_for_each_entry_safe(bpage, tmp, pages, list) {
1163 list_del_init(&bpage->list);
1164 free_buffer_page(bpage);
1165 }
1166
1167 return -ENOMEM;
1168}
1169
1170static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1171 unsigned nr_pages)
1172{
1173 LIST_HEAD(pages);
1174
1175 WARN_ON(!nr_pages);
1176
1177 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1178 return -ENOMEM;
1179
1180 /*
1181 * The ring buffer page list is a circular list that does not
1182 * start and end with a list head. All page list items point to
1183 * other pages.
1184 */
1185 cpu_buffer->pages = pages.next;
1186 list_del(&pages);
1187
1188 cpu_buffer->nr_pages = nr_pages;
1189
1190 rb_check_pages(cpu_buffer);
1191
1192 return 0;
1193}
1194
1195static struct ring_buffer_per_cpu *
1196rb_allocate_cpu_buffer(struct ring_buffer *buffer, int nr_pages, int cpu)
1197{
1198 struct ring_buffer_per_cpu *cpu_buffer;
1199 struct buffer_page *bpage;
1200 struct page *page;
1201 int ret;
1202
1203 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1204 GFP_KERNEL, cpu_to_node(cpu));
1205 if (!cpu_buffer)
1206 return NULL;
1207
1208 cpu_buffer->cpu = cpu;
1209 cpu_buffer->buffer = buffer;
1210 raw_spin_lock_init(&cpu_buffer->reader_lock);
1211 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1212 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1213 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1214 init_completion(&cpu_buffer->update_done);
1215 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1216 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1217 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1218
1219 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1220 GFP_KERNEL, cpu_to_node(cpu));
1221 if (!bpage)
1222 goto fail_free_buffer;
1223
1224 rb_check_bpage(cpu_buffer, bpage);
1225
1226 cpu_buffer->reader_page = bpage;
1227 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1228 if (!page)
1229 goto fail_free_reader;
1230 bpage->page = page_address(page);
1231 rb_init_page(bpage->page);
1232
1233 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1234 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1235
1236 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1237 if (ret < 0)
1238 goto fail_free_reader;
1239
1240 cpu_buffer->head_page
1241 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1242 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1243
1244 rb_head_page_activate(cpu_buffer);
1245
1246 return cpu_buffer;
1247
1248 fail_free_reader:
1249 free_buffer_page(cpu_buffer->reader_page);
1250
1251 fail_free_buffer:
1252 kfree(cpu_buffer);
1253 return NULL;
1254}
1255
1256static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1257{
1258 struct list_head *head = cpu_buffer->pages;
1259 struct buffer_page *bpage, *tmp;
1260
1261 free_buffer_page(cpu_buffer->reader_page);
1262
1263 rb_head_page_deactivate(cpu_buffer);
1264
1265 if (head) {
1266 list_for_each_entry_safe(bpage, tmp, head, list) {
1267 list_del_init(&bpage->list);
1268 free_buffer_page(bpage);
1269 }
1270 bpage = list_entry(head, struct buffer_page, list);
1271 free_buffer_page(bpage);
1272 }
1273
1274 kfree(cpu_buffer);
1275}
1276
1277#ifdef CONFIG_HOTPLUG_CPU
1278static int rb_cpu_notify(struct notifier_block *self,
1279 unsigned long action, void *hcpu);
1280#endif
1281
1282/**
1283 * __ring_buffer_alloc - allocate a new ring_buffer
1284 * @size: the size in bytes per cpu that is needed.
1285 * @flags: attributes to set for the ring buffer.
1286 *
1287 * Currently the only flag that is available is the RB_FL_OVERWRITE
1288 * flag. This flag means that the buffer will overwrite old data
1289 * when the buffer wraps. If this flag is not set, the buffer will
1290 * drop data when the tail hits the head.
1291 */
1292struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1293 struct lock_class_key *key)
1294{
1295 struct ring_buffer *buffer;
1296 int bsize;
1297 int cpu, nr_pages;
1298
1299 /* keep it in its own cache line */
1300 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1301 GFP_KERNEL);
1302 if (!buffer)
1303 return NULL;
1304
1305 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1306 goto fail_free_buffer;
1307
1308 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1309 buffer->flags = flags;
1310 buffer->clock = trace_clock_local;
1311 buffer->reader_lock_key = key;
1312
1313 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1314 init_waitqueue_head(&buffer->irq_work.waiters);
1315
1316 /* need at least two pages */
1317 if (nr_pages < 2)
1318 nr_pages = 2;
1319
1320 /*
1321 * In case of non-hotplug cpu, if the ring-buffer is allocated
1322 * in early initcall, it will not be notified of secondary cpus.
1323 * In that off case, we need to allocate for all possible cpus.
1324 */
1325#ifdef CONFIG_HOTPLUG_CPU
1326 cpu_notifier_register_begin();
1327 cpumask_copy(buffer->cpumask, cpu_online_mask);
1328#else
1329 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1330#endif
1331 buffer->cpus = nr_cpu_ids;
1332
1333 bsize = sizeof(void *) * nr_cpu_ids;
1334 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1335 GFP_KERNEL);
1336 if (!buffer->buffers)
1337 goto fail_free_cpumask;
1338
1339 for_each_buffer_cpu(buffer, cpu) {
1340 buffer->buffers[cpu] =
1341 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1342 if (!buffer->buffers[cpu])
1343 goto fail_free_buffers;
1344 }
1345
1346#ifdef CONFIG_HOTPLUG_CPU
1347 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1348 buffer->cpu_notify.priority = 0;
1349 __register_cpu_notifier(&buffer->cpu_notify);
1350 cpu_notifier_register_done();
1351#endif
1352
1353 mutex_init(&buffer->mutex);
1354
1355 return buffer;
1356
1357 fail_free_buffers:
1358 for_each_buffer_cpu(buffer, cpu) {
1359 if (buffer->buffers[cpu])
1360 rb_free_cpu_buffer(buffer->buffers[cpu]);
1361 }
1362 kfree(buffer->buffers);
1363
1364 fail_free_cpumask:
1365 free_cpumask_var(buffer->cpumask);
1366#ifdef CONFIG_HOTPLUG_CPU
1367 cpu_notifier_register_done();
1368#endif
1369
1370 fail_free_buffer:
1371 kfree(buffer);
1372 return NULL;
1373}
1374EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1375
1376/**
1377 * ring_buffer_free - free a ring buffer.
1378 * @buffer: the buffer to free.
1379 */
1380void
1381ring_buffer_free(struct ring_buffer *buffer)
1382{
1383 int cpu;
1384
1385#ifdef CONFIG_HOTPLUG_CPU
1386 cpu_notifier_register_begin();
1387 __unregister_cpu_notifier(&buffer->cpu_notify);
1388#endif
1389
1390 for_each_buffer_cpu(buffer, cpu)
1391 rb_free_cpu_buffer(buffer->buffers[cpu]);
1392
1393#ifdef CONFIG_HOTPLUG_CPU
1394 cpu_notifier_register_done();
1395#endif
1396
1397 kfree(buffer->buffers);
1398 free_cpumask_var(buffer->cpumask);
1399
1400 kfree(buffer);
1401}
1402EXPORT_SYMBOL_GPL(ring_buffer_free);
1403
1404void ring_buffer_set_clock(struct ring_buffer *buffer,
1405 u64 (*clock)(void))
1406{
1407 buffer->clock = clock;
1408}
1409
1410static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1411
1412static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1413{
1414 return local_read(&bpage->entries) & RB_WRITE_MASK;
1415}
1416
1417static inline unsigned long rb_page_write(struct buffer_page *bpage)
1418{
1419 return local_read(&bpage->write) & RB_WRITE_MASK;
1420}
1421
1422static int
1423rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned int nr_pages)
1424{
1425 struct list_head *tail_page, *to_remove, *next_page;
1426 struct buffer_page *to_remove_page, *tmp_iter_page;
1427 struct buffer_page *last_page, *first_page;
1428 unsigned int nr_removed;
1429 unsigned long head_bit;
1430 int page_entries;
1431
1432 head_bit = 0;
1433
1434 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1435 atomic_inc(&cpu_buffer->record_disabled);
1436 /*
1437 * We don't race with the readers since we have acquired the reader
1438 * lock. We also don't race with writers after disabling recording.
1439 * This makes it easy to figure out the first and the last page to be
1440 * removed from the list. We unlink all the pages in between including
1441 * the first and last pages. This is done in a busy loop so that we
1442 * lose the least number of traces.
1443 * The pages are freed after we restart recording and unlock readers.
1444 */
1445 tail_page = &cpu_buffer->tail_page->list;
1446
1447 /*
1448 * tail page might be on reader page, we remove the next page
1449 * from the ring buffer
1450 */
1451 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1452 tail_page = rb_list_head(tail_page->next);
1453 to_remove = tail_page;
1454
1455 /* start of pages to remove */
1456 first_page = list_entry(rb_list_head(to_remove->next),
1457 struct buffer_page, list);
1458
1459 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1460 to_remove = rb_list_head(to_remove)->next;
1461 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1462 }
1463
1464 next_page = rb_list_head(to_remove)->next;
1465
1466 /*
1467 * Now we remove all pages between tail_page and next_page.
1468 * Make sure that we have head_bit value preserved for the
1469 * next page
1470 */
1471 tail_page->next = (struct list_head *)((unsigned long)next_page |
1472 head_bit);
1473 next_page = rb_list_head(next_page);
1474 next_page->prev = tail_page;
1475
1476 /* make sure pages points to a valid page in the ring buffer */
1477 cpu_buffer->pages = next_page;
1478
1479 /* update head page */
1480 if (head_bit)
1481 cpu_buffer->head_page = list_entry(next_page,
1482 struct buffer_page, list);
1483
1484 /*
1485 * change read pointer to make sure any read iterators reset
1486 * themselves
1487 */
1488 cpu_buffer->read = 0;
1489
1490 /* pages are removed, resume tracing and then free the pages */
1491 atomic_dec(&cpu_buffer->record_disabled);
1492 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1493
1494 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1495
1496 /* last buffer page to remove */
1497 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1498 list);
1499 tmp_iter_page = first_page;
1500
1501 do {
1502 to_remove_page = tmp_iter_page;
1503 rb_inc_page(cpu_buffer, &tmp_iter_page);
1504
1505 /* update the counters */
1506 page_entries = rb_page_entries(to_remove_page);
1507 if (page_entries) {
1508 /*
1509 * If something was added to this page, it was full
1510 * since it is not the tail page. So we deduct the
1511 * bytes consumed in ring buffer from here.
1512 * Increment overrun to account for the lost events.
1513 */
1514 local_add(page_entries, &cpu_buffer->overrun);
1515 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1516 }
1517
1518 /*
1519 * We have already removed references to this list item, just
1520 * free up the buffer_page and its page
1521 */
1522 free_buffer_page(to_remove_page);
1523 nr_removed--;
1524
1525 } while (to_remove_page != last_page);
1526
1527 RB_WARN_ON(cpu_buffer, nr_removed);
1528
1529 return nr_removed == 0;
1530}
1531
1532static int
1533rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1534{
1535 struct list_head *pages = &cpu_buffer->new_pages;
1536 int retries, success;
1537
1538 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1539 /*
1540 * We are holding the reader lock, so the reader page won't be swapped
1541 * in the ring buffer. Now we are racing with the writer trying to
1542 * move head page and the tail page.
1543 * We are going to adapt the reader page update process where:
1544 * 1. We first splice the start and end of list of new pages between
1545 * the head page and its previous page.
1546 * 2. We cmpxchg the prev_page->next to point from head page to the
1547 * start of new pages list.
1548 * 3. Finally, we update the head->prev to the end of new list.
1549 *
1550 * We will try this process 10 times, to make sure that we don't keep
1551 * spinning.
1552 */
1553 retries = 10;
1554 success = 0;
1555 while (retries--) {
1556 struct list_head *head_page, *prev_page, *r;
1557 struct list_head *last_page, *first_page;
1558 struct list_head *head_page_with_bit;
1559
1560 head_page = &rb_set_head_page(cpu_buffer)->list;
1561 if (!head_page)
1562 break;
1563 prev_page = head_page->prev;
1564
1565 first_page = pages->next;
1566 last_page = pages->prev;
1567
1568 head_page_with_bit = (struct list_head *)
1569 ((unsigned long)head_page | RB_PAGE_HEAD);
1570
1571 last_page->next = head_page_with_bit;
1572 first_page->prev = prev_page;
1573
1574 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1575
1576 if (r == head_page_with_bit) {
1577 /*
1578 * yay, we replaced the page pointer to our new list,
1579 * now, we just have to update to head page's prev
1580 * pointer to point to end of list
1581 */
1582 head_page->prev = last_page;
1583 success = 1;
1584 break;
1585 }
1586 }
1587
1588 if (success)
1589 INIT_LIST_HEAD(pages);
1590 /*
1591 * If we weren't successful in adding in new pages, warn and stop
1592 * tracing
1593 */
1594 RB_WARN_ON(cpu_buffer, !success);
1595 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1596
1597 /* free pages if they weren't inserted */
1598 if (!success) {
1599 struct buffer_page *bpage, *tmp;
1600 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1601 list) {
1602 list_del_init(&bpage->list);
1603 free_buffer_page(bpage);
1604 }
1605 }
1606 return success;
1607}
1608
1609static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1610{
1611 int success;
1612
1613 if (cpu_buffer->nr_pages_to_update > 0)
1614 success = rb_insert_pages(cpu_buffer);
1615 else
1616 success = rb_remove_pages(cpu_buffer,
1617 -cpu_buffer->nr_pages_to_update);
1618
1619 if (success)
1620 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1621}
1622
1623static void update_pages_handler(struct work_struct *work)
1624{
1625 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1626 struct ring_buffer_per_cpu, update_pages_work);
1627 rb_update_pages(cpu_buffer);
1628 complete(&cpu_buffer->update_done);
1629}
1630
1631/**
1632 * ring_buffer_resize - resize the ring buffer
1633 * @buffer: the buffer to resize.
1634 * @size: the new size.
1635 * @cpu_id: the cpu buffer to resize
1636 *
1637 * Minimum size is 2 * BUF_PAGE_SIZE.
1638 *
1639 * Returns 0 on success and < 0 on failure.
1640 */
1641int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1642 int cpu_id)
1643{
1644 struct ring_buffer_per_cpu *cpu_buffer;
1645 unsigned nr_pages;
1646 int cpu, err = 0;
1647
1648 /*
1649 * Always succeed at resizing a non-existent buffer:
1650 */
1651 if (!buffer)
1652 return size;
1653
1654 /* Make sure the requested buffer exists */
1655 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1656 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1657 return size;
1658
1659 size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1660 size *= BUF_PAGE_SIZE;
1661
1662 /* we need a minimum of two pages */
1663 if (size < BUF_PAGE_SIZE * 2)
1664 size = BUF_PAGE_SIZE * 2;
1665
1666 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1667
1668 /*
1669 * Don't succeed if resizing is disabled, as a reader might be
1670 * manipulating the ring buffer and is expecting a sane state while
1671 * this is true.
1672 */
1673 if (atomic_read(&buffer->resize_disabled))
1674 return -EBUSY;
1675
1676 /* prevent another thread from changing buffer sizes */
1677 mutex_lock(&buffer->mutex);
1678
1679 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1680 /* calculate the pages to update */
1681 for_each_buffer_cpu(buffer, cpu) {
1682 cpu_buffer = buffer->buffers[cpu];
1683
1684 cpu_buffer->nr_pages_to_update = nr_pages -
1685 cpu_buffer->nr_pages;
1686 /*
1687 * nothing more to do for removing pages or no update
1688 */
1689 if (cpu_buffer->nr_pages_to_update <= 0)
1690 continue;
1691 /*
1692 * to add pages, make sure all new pages can be
1693 * allocated without receiving ENOMEM
1694 */
1695 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1696 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1697 &cpu_buffer->new_pages, cpu)) {
1698 /* not enough memory for new pages */
1699 err = -ENOMEM;
1700 goto out_err;
1701 }
1702 }
1703
1704 get_online_cpus();
1705 /*
1706 * Fire off all the required work handlers
1707 * We can't schedule on offline CPUs, but it's not necessary
1708 * since we can change their buffer sizes without any race.
1709 */
1710 for_each_buffer_cpu(buffer, cpu) {
1711 cpu_buffer = buffer->buffers[cpu];
1712 if (!cpu_buffer->nr_pages_to_update)
1713 continue;
1714
1715 /* Can't run something on an offline CPU. */
1716 if (!cpu_online(cpu)) {
1717 rb_update_pages(cpu_buffer);
1718 cpu_buffer->nr_pages_to_update = 0;
1719 } else {
1720 schedule_work_on(cpu,
1721 &cpu_buffer->update_pages_work);
1722 }
1723 }
1724
1725 /* wait for all the updates to complete */
1726 for_each_buffer_cpu(buffer, cpu) {
1727 cpu_buffer = buffer->buffers[cpu];
1728 if (!cpu_buffer->nr_pages_to_update)
1729 continue;
1730
1731 if (cpu_online(cpu))
1732 wait_for_completion(&cpu_buffer->update_done);
1733 cpu_buffer->nr_pages_to_update = 0;
1734 }
1735
1736 put_online_cpus();
1737 } else {
1738 /* Make sure this CPU has been intitialized */
1739 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1740 goto out;
1741
1742 cpu_buffer = buffer->buffers[cpu_id];
1743
1744 if (nr_pages == cpu_buffer->nr_pages)
1745 goto out;
1746
1747 cpu_buffer->nr_pages_to_update = nr_pages -
1748 cpu_buffer->nr_pages;
1749
1750 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1751 if (cpu_buffer->nr_pages_to_update > 0 &&
1752 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1753 &cpu_buffer->new_pages, cpu_id)) {
1754 err = -ENOMEM;
1755 goto out_err;
1756 }
1757
1758 get_online_cpus();
1759
1760 /* Can't run something on an offline CPU. */
1761 if (!cpu_online(cpu_id))
1762 rb_update_pages(cpu_buffer);
1763 else {
1764 schedule_work_on(cpu_id,
1765 &cpu_buffer->update_pages_work);
1766 wait_for_completion(&cpu_buffer->update_done);
1767 }
1768
1769 cpu_buffer->nr_pages_to_update = 0;
1770 put_online_cpus();
1771 }
1772
1773 out:
1774 /*
1775 * The ring buffer resize can happen with the ring buffer
1776 * enabled, so that the update disturbs the tracing as little
1777 * as possible. But if the buffer is disabled, we do not need
1778 * to worry about that, and we can take the time to verify
1779 * that the buffer is not corrupt.
1780 */
1781 if (atomic_read(&buffer->record_disabled)) {
1782 atomic_inc(&buffer->record_disabled);
1783 /*
1784 * Even though the buffer was disabled, we must make sure
1785 * that it is truly disabled before calling rb_check_pages.
1786 * There could have been a race between checking
1787 * record_disable and incrementing it.
1788 */
1789 synchronize_sched();
1790 for_each_buffer_cpu(buffer, cpu) {
1791 cpu_buffer = buffer->buffers[cpu];
1792 rb_check_pages(cpu_buffer);
1793 }
1794 atomic_dec(&buffer->record_disabled);
1795 }
1796
1797 mutex_unlock(&buffer->mutex);
1798 return size;
1799
1800 out_err:
1801 for_each_buffer_cpu(buffer, cpu) {
1802 struct buffer_page *bpage, *tmp;
1803
1804 cpu_buffer = buffer->buffers[cpu];
1805 cpu_buffer->nr_pages_to_update = 0;
1806
1807 if (list_empty(&cpu_buffer->new_pages))
1808 continue;
1809
1810 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1811 list) {
1812 list_del_init(&bpage->list);
1813 free_buffer_page(bpage);
1814 }
1815 }
1816 mutex_unlock(&buffer->mutex);
1817 return err;
1818}
1819EXPORT_SYMBOL_GPL(ring_buffer_resize);
1820
1821void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1822{
1823 mutex_lock(&buffer->mutex);
1824 if (val)
1825 buffer->flags |= RB_FL_OVERWRITE;
1826 else
1827 buffer->flags &= ~RB_FL_OVERWRITE;
1828 mutex_unlock(&buffer->mutex);
1829}
1830EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1831
1832static inline void *
1833__rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1834{
1835 return bpage->data + index;
1836}
1837
1838static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1839{
1840 return bpage->page->data + index;
1841}
1842
1843static inline struct ring_buffer_event *
1844rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1845{
1846 return __rb_page_index(cpu_buffer->reader_page,
1847 cpu_buffer->reader_page->read);
1848}
1849
1850static inline struct ring_buffer_event *
1851rb_iter_head_event(struct ring_buffer_iter *iter)
1852{
1853 return __rb_page_index(iter->head_page, iter->head);
1854}
1855
1856static inline unsigned rb_page_commit(struct buffer_page *bpage)
1857{
1858 return local_read(&bpage->page->commit);
1859}
1860
1861/* Size is determined by what has been committed */
1862static inline unsigned rb_page_size(struct buffer_page *bpage)
1863{
1864 return rb_page_commit(bpage);
1865}
1866
1867static inline unsigned
1868rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1869{
1870 return rb_page_commit(cpu_buffer->commit_page);
1871}
1872
1873static inline unsigned
1874rb_event_index(struct ring_buffer_event *event)
1875{
1876 unsigned long addr = (unsigned long)event;
1877
1878 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1879}
1880
1881static void rb_inc_iter(struct ring_buffer_iter *iter)
1882{
1883 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1884
1885 /*
1886 * The iterator could be on the reader page (it starts there).
1887 * But the head could have moved, since the reader was
1888 * found. Check for this case and assign the iterator
1889 * to the head page instead of next.
1890 */
1891 if (iter->head_page == cpu_buffer->reader_page)
1892 iter->head_page = rb_set_head_page(cpu_buffer);
1893 else
1894 rb_inc_page(cpu_buffer, &iter->head_page);
1895
1896 iter->read_stamp = iter->head_page->page->time_stamp;
1897 iter->head = 0;
1898}
1899
1900/*
1901 * rb_handle_head_page - writer hit the head page
1902 *
1903 * Returns: +1 to retry page
1904 * 0 to continue
1905 * -1 on error
1906 */
1907static int
1908rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1909 struct buffer_page *tail_page,
1910 struct buffer_page *next_page)
1911{
1912 struct buffer_page *new_head;
1913 int entries;
1914 int type;
1915 int ret;
1916
1917 entries = rb_page_entries(next_page);
1918
1919 /*
1920 * The hard part is here. We need to move the head
1921 * forward, and protect against both readers on
1922 * other CPUs and writers coming in via interrupts.
1923 */
1924 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1925 RB_PAGE_HEAD);
1926
1927 /*
1928 * type can be one of four:
1929 * NORMAL - an interrupt already moved it for us
1930 * HEAD - we are the first to get here.
1931 * UPDATE - we are the interrupt interrupting
1932 * a current move.
1933 * MOVED - a reader on another CPU moved the next
1934 * pointer to its reader page. Give up
1935 * and try again.
1936 */
1937
1938 switch (type) {
1939 case RB_PAGE_HEAD:
1940 /*
1941 * We changed the head to UPDATE, thus
1942 * it is our responsibility to update
1943 * the counters.
1944 */
1945 local_add(entries, &cpu_buffer->overrun);
1946 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1947
1948 /*
1949 * The entries will be zeroed out when we move the
1950 * tail page.
1951 */
1952
1953 /* still more to do */
1954 break;
1955
1956 case RB_PAGE_UPDATE:
1957 /*
1958 * This is an interrupt that interrupt the
1959 * previous update. Still more to do.
1960 */
1961 break;
1962 case RB_PAGE_NORMAL:
1963 /*
1964 * An interrupt came in before the update
1965 * and processed this for us.
1966 * Nothing left to do.
1967 */
1968 return 1;
1969 case RB_PAGE_MOVED:
1970 /*
1971 * The reader is on another CPU and just did
1972 * a swap with our next_page.
1973 * Try again.
1974 */
1975 return 1;
1976 default:
1977 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1978 return -1;
1979 }
1980
1981 /*
1982 * Now that we are here, the old head pointer is
1983 * set to UPDATE. This will keep the reader from
1984 * swapping the head page with the reader page.
1985 * The reader (on another CPU) will spin till
1986 * we are finished.
1987 *
1988 * We just need to protect against interrupts
1989 * doing the job. We will set the next pointer
1990 * to HEAD. After that, we set the old pointer
1991 * to NORMAL, but only if it was HEAD before.
1992 * otherwise we are an interrupt, and only
1993 * want the outer most commit to reset it.
1994 */
1995 new_head = next_page;
1996 rb_inc_page(cpu_buffer, &new_head);
1997
1998 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1999 RB_PAGE_NORMAL);
2000
2001 /*
2002 * Valid returns are:
2003 * HEAD - an interrupt came in and already set it.
2004 * NORMAL - One of two things:
2005 * 1) We really set it.
2006 * 2) A bunch of interrupts came in and moved
2007 * the page forward again.
2008 */
2009 switch (ret) {
2010 case RB_PAGE_HEAD:
2011 case RB_PAGE_NORMAL:
2012 /* OK */
2013 break;
2014 default:
2015 RB_WARN_ON(cpu_buffer, 1);
2016 return -1;
2017 }
2018
2019 /*
2020 * It is possible that an interrupt came in,
2021 * set the head up, then more interrupts came in
2022 * and moved it again. When we get back here,
2023 * the page would have been set to NORMAL but we
2024 * just set it back to HEAD.
2025 *
2026 * How do you detect this? Well, if that happened
2027 * the tail page would have moved.
2028 */
2029 if (ret == RB_PAGE_NORMAL) {
2030 struct buffer_page *buffer_tail_page;
2031
2032 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2033 /*
2034 * If the tail had moved passed next, then we need
2035 * to reset the pointer.
2036 */
2037 if (buffer_tail_page != tail_page &&
2038 buffer_tail_page != next_page)
2039 rb_head_page_set_normal(cpu_buffer, new_head,
2040 next_page,
2041 RB_PAGE_HEAD);
2042 }
2043
2044 /*
2045 * If this was the outer most commit (the one that
2046 * changed the original pointer from HEAD to UPDATE),
2047 * then it is up to us to reset it to NORMAL.
2048 */
2049 if (type == RB_PAGE_HEAD) {
2050 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2051 tail_page,
2052 RB_PAGE_UPDATE);
2053 if (RB_WARN_ON(cpu_buffer,
2054 ret != RB_PAGE_UPDATE))
2055 return -1;
2056 }
2057
2058 return 0;
2059}
2060
2061static inline void
2062rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2063 unsigned long tail, struct rb_event_info *info)
2064{
2065 struct buffer_page *tail_page = info->tail_page;
2066 struct ring_buffer_event *event;
2067 unsigned long length = info->length;
2068
2069 /*
2070 * Only the event that crossed the page boundary
2071 * must fill the old tail_page with padding.
2072 */
2073 if (tail >= BUF_PAGE_SIZE) {
2074 /*
2075 * If the page was filled, then we still need
2076 * to update the real_end. Reset it to zero
2077 * and the reader will ignore it.
2078 */
2079 if (tail == BUF_PAGE_SIZE)
2080 tail_page->real_end = 0;
2081
2082 local_sub(length, &tail_page->write);
2083 return;
2084 }
2085
2086 event = __rb_page_index(tail_page, tail);
2087 kmemcheck_annotate_bitfield(event, bitfield);
2088
2089 /* account for padding bytes */
2090 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2091
2092 /*
2093 * Save the original length to the meta data.
2094 * This will be used by the reader to add lost event
2095 * counter.
2096 */
2097 tail_page->real_end = tail;
2098
2099 /*
2100 * If this event is bigger than the minimum size, then
2101 * we need to be careful that we don't subtract the
2102 * write counter enough to allow another writer to slip
2103 * in on this page.
2104 * We put in a discarded commit instead, to make sure
2105 * that this space is not used again.
2106 *
2107 * If we are less than the minimum size, we don't need to
2108 * worry about it.
2109 */
2110 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2111 /* No room for any events */
2112
2113 /* Mark the rest of the page with padding */
2114 rb_event_set_padding(event);
2115
2116 /* Set the write back to the previous setting */
2117 local_sub(length, &tail_page->write);
2118 return;
2119 }
2120
2121 /* Put in a discarded event */
2122 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2123 event->type_len = RINGBUF_TYPE_PADDING;
2124 /* time delta must be non zero */
2125 event->time_delta = 1;
2126
2127 /* Set write to end of buffer */
2128 length = (tail + length) - BUF_PAGE_SIZE;
2129 local_sub(length, &tail_page->write);
2130}
2131
2132static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2133
2134/*
2135 * This is the slow path, force gcc not to inline it.
2136 */
2137static noinline struct ring_buffer_event *
2138rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2139 unsigned long tail, struct rb_event_info *info)
2140{
2141 struct buffer_page *tail_page = info->tail_page;
2142 struct buffer_page *commit_page = cpu_buffer->commit_page;
2143 struct ring_buffer *buffer = cpu_buffer->buffer;
2144 struct buffer_page *next_page;
2145 int ret;
2146
2147 next_page = tail_page;
2148
2149 rb_inc_page(cpu_buffer, &next_page);
2150
2151 /*
2152 * If for some reason, we had an interrupt storm that made
2153 * it all the way around the buffer, bail, and warn
2154 * about it.
2155 */
2156 if (unlikely(next_page == commit_page)) {
2157 local_inc(&cpu_buffer->commit_overrun);
2158 goto out_reset;
2159 }
2160
2161 /*
2162 * This is where the fun begins!
2163 *
2164 * We are fighting against races between a reader that
2165 * could be on another CPU trying to swap its reader
2166 * page with the buffer head.
2167 *
2168 * We are also fighting against interrupts coming in and
2169 * moving the head or tail on us as well.
2170 *
2171 * If the next page is the head page then we have filled
2172 * the buffer, unless the commit page is still on the
2173 * reader page.
2174 */
2175 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2176
2177 /*
2178 * If the commit is not on the reader page, then
2179 * move the header page.
2180 */
2181 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2182 /*
2183 * If we are not in overwrite mode,
2184 * this is easy, just stop here.
2185 */
2186 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2187 local_inc(&cpu_buffer->dropped_events);
2188 goto out_reset;
2189 }
2190
2191 ret = rb_handle_head_page(cpu_buffer,
2192 tail_page,
2193 next_page);
2194 if (ret < 0)
2195 goto out_reset;
2196 if (ret)
2197 goto out_again;
2198 } else {
2199 /*
2200 * We need to be careful here too. The
2201 * commit page could still be on the reader
2202 * page. We could have a small buffer, and
2203 * have filled up the buffer with events
2204 * from interrupts and such, and wrapped.
2205 *
2206 * Note, if the tail page is also the on the
2207 * reader_page, we let it move out.
2208 */
2209 if (unlikely((cpu_buffer->commit_page !=
2210 cpu_buffer->tail_page) &&
2211 (cpu_buffer->commit_page ==
2212 cpu_buffer->reader_page))) {
2213 local_inc(&cpu_buffer->commit_overrun);
2214 goto out_reset;
2215 }
2216 }
2217 }
2218
2219 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2220
2221 out_again:
2222
2223 rb_reset_tail(cpu_buffer, tail, info);
2224
2225 /* Commit what we have for now. */
2226 rb_end_commit(cpu_buffer);
2227 /* rb_end_commit() decs committing */
2228 local_inc(&cpu_buffer->committing);
2229
2230 /* fail and let the caller try again */
2231 return ERR_PTR(-EAGAIN);
2232
2233 out_reset:
2234 /* reset write */
2235 rb_reset_tail(cpu_buffer, tail, info);
2236
2237 return NULL;
2238}
2239
2240/* Slow path, do not inline */
2241static noinline struct ring_buffer_event *
2242rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
2243{
2244 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2245
2246 /* Not the first event on the page? */
2247 if (rb_event_index(event)) {
2248 event->time_delta = delta & TS_MASK;
2249 event->array[0] = delta >> TS_SHIFT;
2250 } else {
2251 /* nope, just zero it */
2252 event->time_delta = 0;
2253 event->array[0] = 0;
2254 }
2255
2256 return skip_time_extend(event);
2257}
2258
2259static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2260 struct ring_buffer_event *event);
2261
2262/**
2263 * rb_update_event - update event type and data
2264 * @event: the event to update
2265 * @type: the type of event
2266 * @length: the size of the event field in the ring buffer
2267 *
2268 * Update the type and data fields of the event. The length
2269 * is the actual size that is written to the ring buffer,
2270 * and with this, we can determine what to place into the
2271 * data field.
2272 */
2273static void
2274rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2275 struct ring_buffer_event *event,
2276 struct rb_event_info *info)
2277{
2278 unsigned length = info->length;
2279 u64 delta = info->delta;
2280
2281 /* Only a commit updates the timestamp */
2282 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2283 delta = 0;
2284
2285 /*
2286 * If we need to add a timestamp, then we
2287 * add it to the start of the resevered space.
2288 */
2289 if (unlikely(info->add_timestamp)) {
2290 event = rb_add_time_stamp(event, delta);
2291 length -= RB_LEN_TIME_EXTEND;
2292 delta = 0;
2293 }
2294
2295 event->time_delta = delta;
2296 length -= RB_EVNT_HDR_SIZE;
2297 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2298 event->type_len = 0;
2299 event->array[0] = length;
2300 } else
2301 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2302}
2303
2304static unsigned rb_calculate_event_length(unsigned length)
2305{
2306 struct ring_buffer_event event; /* Used only for sizeof array */
2307
2308 /* zero length can cause confusions */
2309 if (!length)
2310 length++;
2311
2312 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2313 length += sizeof(event.array[0]);
2314
2315 length += RB_EVNT_HDR_SIZE;
2316 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2317
2318 /*
2319 * In case the time delta is larger than the 27 bits for it
2320 * in the header, we need to add a timestamp. If another
2321 * event comes in when trying to discard this one to increase
2322 * the length, then the timestamp will be added in the allocated
2323 * space of this event. If length is bigger than the size needed
2324 * for the TIME_EXTEND, then padding has to be used. The events
2325 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2326 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2327 * As length is a multiple of 4, we only need to worry if it
2328 * is 12 (RB_LEN_TIME_EXTEND + 4).
2329 */
2330 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2331 length += RB_ALIGNMENT;
2332
2333 return length;
2334}
2335
2336#ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2337static inline bool sched_clock_stable(void)
2338{
2339 return true;
2340}
2341#endif
2342
2343static inline int
2344rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2345 struct ring_buffer_event *event)
2346{
2347 unsigned long new_index, old_index;
2348 struct buffer_page *bpage;
2349 unsigned long index;
2350 unsigned long addr;
2351
2352 new_index = rb_event_index(event);
2353 old_index = new_index + rb_event_ts_length(event);
2354 addr = (unsigned long)event;
2355 addr &= PAGE_MASK;
2356
2357 bpage = READ_ONCE(cpu_buffer->tail_page);
2358
2359 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2360 unsigned long write_mask =
2361 local_read(&bpage->write) & ~RB_WRITE_MASK;
2362 unsigned long event_length = rb_event_length(event);
2363 /*
2364 * This is on the tail page. It is possible that
2365 * a write could come in and move the tail page
2366 * and write to the next page. That is fine
2367 * because we just shorten what is on this page.
2368 */
2369 old_index += write_mask;
2370 new_index += write_mask;
2371 index = local_cmpxchg(&bpage->write, old_index, new_index);
2372 if (index == old_index) {
2373 /* update counters */
2374 local_sub(event_length, &cpu_buffer->entries_bytes);
2375 return 1;
2376 }
2377 }
2378
2379 /* could not discard */
2380 return 0;
2381}
2382
2383static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2384{
2385 local_inc(&cpu_buffer->committing);
2386 local_inc(&cpu_buffer->commits);
2387}
2388
2389static void
2390rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2391{
2392 unsigned long max_count;
2393
2394 /*
2395 * We only race with interrupts and NMIs on this CPU.
2396 * If we own the commit event, then we can commit
2397 * all others that interrupted us, since the interruptions
2398 * are in stack format (they finish before they come
2399 * back to us). This allows us to do a simple loop to
2400 * assign the commit to the tail.
2401 */
2402 again:
2403 max_count = cpu_buffer->nr_pages * 100;
2404
2405 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2406 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2407 return;
2408 if (RB_WARN_ON(cpu_buffer,
2409 rb_is_reader_page(cpu_buffer->tail_page)))
2410 return;
2411 local_set(&cpu_buffer->commit_page->page->commit,
2412 rb_page_write(cpu_buffer->commit_page));
2413 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2414 /* Only update the write stamp if the page has an event */
2415 if (rb_page_write(cpu_buffer->commit_page))
2416 cpu_buffer->write_stamp =
2417 cpu_buffer->commit_page->page->time_stamp;
2418 /* add barrier to keep gcc from optimizing too much */
2419 barrier();
2420 }
2421 while (rb_commit_index(cpu_buffer) !=
2422 rb_page_write(cpu_buffer->commit_page)) {
2423
2424 local_set(&cpu_buffer->commit_page->page->commit,
2425 rb_page_write(cpu_buffer->commit_page));
2426 RB_WARN_ON(cpu_buffer,
2427 local_read(&cpu_buffer->commit_page->page->commit) &
2428 ~RB_WRITE_MASK);
2429 barrier();
2430 }
2431
2432 /* again, keep gcc from optimizing */
2433 barrier();
2434
2435 /*
2436 * If an interrupt came in just after the first while loop
2437 * and pushed the tail page forward, we will be left with
2438 * a dangling commit that will never go forward.
2439 */
2440 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2441 goto again;
2442}
2443
2444static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2445{
2446 unsigned long commits;
2447
2448 if (RB_WARN_ON(cpu_buffer,
2449 !local_read(&cpu_buffer->committing)))
2450 return;
2451
2452 again:
2453 commits = local_read(&cpu_buffer->commits);
2454 /* synchronize with interrupts */
2455 barrier();
2456 if (local_read(&cpu_buffer->committing) == 1)
2457 rb_set_commit_to_write(cpu_buffer);
2458
2459 local_dec(&cpu_buffer->committing);
2460
2461 /* synchronize with interrupts */
2462 barrier();
2463
2464 /*
2465 * Need to account for interrupts coming in between the
2466 * updating of the commit page and the clearing of the
2467 * committing counter.
2468 */
2469 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2470 !local_read(&cpu_buffer->committing)) {
2471 local_inc(&cpu_buffer->committing);
2472 goto again;
2473 }
2474}
2475
2476static inline void rb_event_discard(struct ring_buffer_event *event)
2477{
2478 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2479 event = skip_time_extend(event);
2480
2481 /* array[0] holds the actual length for the discarded event */
2482 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2483 event->type_len = RINGBUF_TYPE_PADDING;
2484 /* time delta must be non zero */
2485 if (!event->time_delta)
2486 event->time_delta = 1;
2487}
2488
2489static inline bool
2490rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2491 struct ring_buffer_event *event)
2492{
2493 unsigned long addr = (unsigned long)event;
2494 unsigned long index;
2495
2496 index = rb_event_index(event);
2497 addr &= PAGE_MASK;
2498
2499 return cpu_buffer->commit_page->page == (void *)addr &&
2500 rb_commit_index(cpu_buffer) == index;
2501}
2502
2503static void
2504rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2505 struct ring_buffer_event *event)
2506{
2507 u64 delta;
2508
2509 /*
2510 * The event first in the commit queue updates the
2511 * time stamp.
2512 */
2513 if (rb_event_is_commit(cpu_buffer, event)) {
2514 /*
2515 * A commit event that is first on a page
2516 * updates the write timestamp with the page stamp
2517 */
2518 if (!rb_event_index(event))
2519 cpu_buffer->write_stamp =
2520 cpu_buffer->commit_page->page->time_stamp;
2521 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2522 delta = event->array[0];
2523 delta <<= TS_SHIFT;
2524 delta += event->time_delta;
2525 cpu_buffer->write_stamp += delta;
2526 } else
2527 cpu_buffer->write_stamp += event->time_delta;
2528 }
2529}
2530
2531static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2532 struct ring_buffer_event *event)
2533{
2534 local_inc(&cpu_buffer->entries);
2535 rb_update_write_stamp(cpu_buffer, event);
2536 rb_end_commit(cpu_buffer);
2537}
2538
2539static __always_inline void
2540rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2541{
2542 bool pagebusy;
2543
2544 if (buffer->irq_work.waiters_pending) {
2545 buffer->irq_work.waiters_pending = false;
2546 /* irq_work_queue() supplies it's own memory barriers */
2547 irq_work_queue(&buffer->irq_work.work);
2548 }
2549
2550 if (cpu_buffer->irq_work.waiters_pending) {
2551 cpu_buffer->irq_work.waiters_pending = false;
2552 /* irq_work_queue() supplies it's own memory barriers */
2553 irq_work_queue(&cpu_buffer->irq_work.work);
2554 }
2555
2556 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
2557
2558 if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2559 cpu_buffer->irq_work.wakeup_full = true;
2560 cpu_buffer->irq_work.full_waiters_pending = false;
2561 /* irq_work_queue() supplies it's own memory barriers */
2562 irq_work_queue(&cpu_buffer->irq_work.work);
2563 }
2564}
2565
2566/*
2567 * The lock and unlock are done within a preempt disable section.
2568 * The current_context per_cpu variable can only be modified
2569 * by the current task between lock and unlock. But it can
2570 * be modified more than once via an interrupt. To pass this
2571 * information from the lock to the unlock without having to
2572 * access the 'in_interrupt()' functions again (which do show
2573 * a bit of overhead in something as critical as function tracing,
2574 * we use a bitmask trick.
2575 *
2576 * bit 0 = NMI context
2577 * bit 1 = IRQ context
2578 * bit 2 = SoftIRQ context
2579 * bit 3 = normal context.
2580 *
2581 * This works because this is the order of contexts that can
2582 * preempt other contexts. A SoftIRQ never preempts an IRQ
2583 * context.
2584 *
2585 * When the context is determined, the corresponding bit is
2586 * checked and set (if it was set, then a recursion of that context
2587 * happened).
2588 *
2589 * On unlock, we need to clear this bit. To do so, just subtract
2590 * 1 from the current_context and AND it to itself.
2591 *
2592 * (binary)
2593 * 101 - 1 = 100
2594 * 101 & 100 = 100 (clearing bit zero)
2595 *
2596 * 1010 - 1 = 1001
2597 * 1010 & 1001 = 1000 (clearing bit 1)
2598 *
2599 * The least significant bit can be cleared this way, and it
2600 * just so happens that it is the same bit corresponding to
2601 * the current context.
2602 */
2603
2604static __always_inline int
2605trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2606{
2607 unsigned int val = cpu_buffer->current_context;
2608 int bit;
2609
2610 if (in_interrupt()) {
2611 if (in_nmi())
2612 bit = RB_CTX_NMI;
2613 else if (in_irq())
2614 bit = RB_CTX_IRQ;
2615 else
2616 bit = RB_CTX_SOFTIRQ;
2617 } else
2618 bit = RB_CTX_NORMAL;
2619
2620 if (unlikely(val & (1 << bit)))
2621 return 1;
2622
2623 val |= (1 << bit);
2624 cpu_buffer->current_context = val;
2625
2626 return 0;
2627}
2628
2629static __always_inline void
2630trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2631{
2632 cpu_buffer->current_context &= cpu_buffer->current_context - 1;
2633}
2634
2635/**
2636 * ring_buffer_unlock_commit - commit a reserved
2637 * @buffer: The buffer to commit to
2638 * @event: The event pointer to commit.
2639 *
2640 * This commits the data to the ring buffer, and releases any locks held.
2641 *
2642 * Must be paired with ring_buffer_lock_reserve.
2643 */
2644int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2645 struct ring_buffer_event *event)
2646{
2647 struct ring_buffer_per_cpu *cpu_buffer;
2648 int cpu = raw_smp_processor_id();
2649
2650 cpu_buffer = buffer->buffers[cpu];
2651
2652 rb_commit(cpu_buffer, event);
2653
2654 rb_wakeups(buffer, cpu_buffer);
2655
2656 trace_recursive_unlock(cpu_buffer);
2657
2658 preempt_enable_notrace();
2659
2660 return 0;
2661}
2662EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2663
2664static noinline void
2665rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2666 struct rb_event_info *info)
2667{
2668 WARN_ONCE(info->delta > (1ULL << 59),
2669 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2670 (unsigned long long)info->delta,
2671 (unsigned long long)info->ts,
2672 (unsigned long long)cpu_buffer->write_stamp,
2673 sched_clock_stable() ? "" :
2674 "If you just came from a suspend/resume,\n"
2675 "please switch to the trace global clock:\n"
2676 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2677 info->add_timestamp = 1;
2678}
2679
2680static struct ring_buffer_event *
2681__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2682 struct rb_event_info *info)
2683{
2684 struct ring_buffer_event *event;
2685 struct buffer_page *tail_page;
2686 unsigned long tail, write;
2687
2688 /*
2689 * If the time delta since the last event is too big to
2690 * hold in the time field of the event, then we append a
2691 * TIME EXTEND event ahead of the data event.
2692 */
2693 if (unlikely(info->add_timestamp))
2694 info->length += RB_LEN_TIME_EXTEND;
2695
2696 /* Don't let the compiler play games with cpu_buffer->tail_page */
2697 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2698 write = local_add_return(info->length, &tail_page->write);
2699
2700 /* set write to only the index of the write */
2701 write &= RB_WRITE_MASK;
2702 tail = write - info->length;
2703
2704 /*
2705 * If this is the first commit on the page, then it has the same
2706 * timestamp as the page itself.
2707 */
2708 if (!tail)
2709 info->delta = 0;
2710
2711 /* See if we shot pass the end of this buffer page */
2712 if (unlikely(write > BUF_PAGE_SIZE))
2713 return rb_move_tail(cpu_buffer, tail, info);
2714
2715 /* We reserved something on the buffer */
2716
2717 event = __rb_page_index(tail_page, tail);
2718 kmemcheck_annotate_bitfield(event, bitfield);
2719 rb_update_event(cpu_buffer, event, info);
2720
2721 local_inc(&tail_page->entries);
2722
2723 /*
2724 * If this is the first commit on the page, then update
2725 * its timestamp.
2726 */
2727 if (!tail)
2728 tail_page->page->time_stamp = info->ts;
2729
2730 /* account for these added bytes */
2731 local_add(info->length, &cpu_buffer->entries_bytes);
2732
2733 return event;
2734}
2735
2736static struct ring_buffer_event *
2737rb_reserve_next_event(struct ring_buffer *buffer,
2738 struct ring_buffer_per_cpu *cpu_buffer,
2739 unsigned long length)
2740{
2741 struct ring_buffer_event *event;
2742 struct rb_event_info info;
2743 int nr_loops = 0;
2744 u64 diff;
2745
2746 rb_start_commit(cpu_buffer);
2747
2748#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2749 /*
2750 * Due to the ability to swap a cpu buffer from a buffer
2751 * it is possible it was swapped before we committed.
2752 * (committing stops a swap). We check for it here and
2753 * if it happened, we have to fail the write.
2754 */
2755 barrier();
2756 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2757 local_dec(&cpu_buffer->committing);
2758 local_dec(&cpu_buffer->commits);
2759 return NULL;
2760 }
2761#endif
2762
2763 info.length = rb_calculate_event_length(length);
2764 again:
2765 info.add_timestamp = 0;
2766 info.delta = 0;
2767
2768 /*
2769 * We allow for interrupts to reenter here and do a trace.
2770 * If one does, it will cause this original code to loop
2771 * back here. Even with heavy interrupts happening, this
2772 * should only happen a few times in a row. If this happens
2773 * 1000 times in a row, there must be either an interrupt
2774 * storm or we have something buggy.
2775 * Bail!
2776 */
2777 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2778 goto out_fail;
2779
2780 info.ts = rb_time_stamp(cpu_buffer->buffer);
2781 diff = info.ts - cpu_buffer->write_stamp;
2782
2783 /* make sure this diff is calculated here */
2784 barrier();
2785
2786 /* Did the write stamp get updated already? */
2787 if (likely(info.ts >= cpu_buffer->write_stamp)) {
2788 info.delta = diff;
2789 if (unlikely(test_time_stamp(info.delta)))
2790 rb_handle_timestamp(cpu_buffer, &info);
2791 }
2792
2793 event = __rb_reserve_next(cpu_buffer, &info);
2794
2795 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2796 if (info.add_timestamp)
2797 info.length -= RB_LEN_TIME_EXTEND;
2798 goto again;
2799 }
2800
2801 if (!event)
2802 goto out_fail;
2803
2804 return event;
2805
2806 out_fail:
2807 rb_end_commit(cpu_buffer);
2808 return NULL;
2809}
2810
2811/**
2812 * ring_buffer_lock_reserve - reserve a part of the buffer
2813 * @buffer: the ring buffer to reserve from
2814 * @length: the length of the data to reserve (excluding event header)
2815 *
2816 * Returns a reseverd event on the ring buffer to copy directly to.
2817 * The user of this interface will need to get the body to write into
2818 * and can use the ring_buffer_event_data() interface.
2819 *
2820 * The length is the length of the data needed, not the event length
2821 * which also includes the event header.
2822 *
2823 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2824 * If NULL is returned, then nothing has been allocated or locked.
2825 */
2826struct ring_buffer_event *
2827ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2828{
2829 struct ring_buffer_per_cpu *cpu_buffer;
2830 struct ring_buffer_event *event;
2831 int cpu;
2832
2833 /* If we are tracing schedule, we don't want to recurse */
2834 preempt_disable_notrace();
2835
2836 if (unlikely(atomic_read(&buffer->record_disabled)))
2837 goto out;
2838
2839 cpu = raw_smp_processor_id();
2840
2841 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2842 goto out;
2843
2844 cpu_buffer = buffer->buffers[cpu];
2845
2846 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2847 goto out;
2848
2849 if (unlikely(length > BUF_MAX_DATA_SIZE))
2850 goto out;
2851
2852 if (unlikely(trace_recursive_lock(cpu_buffer)))
2853 goto out;
2854
2855 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2856 if (!event)
2857 goto out_unlock;
2858
2859 return event;
2860
2861 out_unlock:
2862 trace_recursive_unlock(cpu_buffer);
2863 out:
2864 preempt_enable_notrace();
2865 return NULL;
2866}
2867EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2868
2869/*
2870 * Decrement the entries to the page that an event is on.
2871 * The event does not even need to exist, only the pointer
2872 * to the page it is on. This may only be called before the commit
2873 * takes place.
2874 */
2875static inline void
2876rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2877 struct ring_buffer_event *event)
2878{
2879 unsigned long addr = (unsigned long)event;
2880 struct buffer_page *bpage = cpu_buffer->commit_page;
2881 struct buffer_page *start;
2882
2883 addr &= PAGE_MASK;
2884
2885 /* Do the likely case first */
2886 if (likely(bpage->page == (void *)addr)) {
2887 local_dec(&bpage->entries);
2888 return;
2889 }
2890
2891 /*
2892 * Because the commit page may be on the reader page we
2893 * start with the next page and check the end loop there.
2894 */
2895 rb_inc_page(cpu_buffer, &bpage);
2896 start = bpage;
2897 do {
2898 if (bpage->page == (void *)addr) {
2899 local_dec(&bpage->entries);
2900 return;
2901 }
2902 rb_inc_page(cpu_buffer, &bpage);
2903 } while (bpage != start);
2904
2905 /* commit not part of this buffer?? */
2906 RB_WARN_ON(cpu_buffer, 1);
2907}
2908
2909/**
2910 * ring_buffer_commit_discard - discard an event that has not been committed
2911 * @buffer: the ring buffer
2912 * @event: non committed event to discard
2913 *
2914 * Sometimes an event that is in the ring buffer needs to be ignored.
2915 * This function lets the user discard an event in the ring buffer
2916 * and then that event will not be read later.
2917 *
2918 * This function only works if it is called before the the item has been
2919 * committed. It will try to free the event from the ring buffer
2920 * if another event has not been added behind it.
2921 *
2922 * If another event has been added behind it, it will set the event
2923 * up as discarded, and perform the commit.
2924 *
2925 * If this function is called, do not call ring_buffer_unlock_commit on
2926 * the event.
2927 */
2928void ring_buffer_discard_commit(struct ring_buffer *buffer,
2929 struct ring_buffer_event *event)
2930{
2931 struct ring_buffer_per_cpu *cpu_buffer;
2932 int cpu;
2933
2934 /* The event is discarded regardless */
2935 rb_event_discard(event);
2936
2937 cpu = smp_processor_id();
2938 cpu_buffer = buffer->buffers[cpu];
2939
2940 /*
2941 * This must only be called if the event has not been
2942 * committed yet. Thus we can assume that preemption
2943 * is still disabled.
2944 */
2945 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2946
2947 rb_decrement_entry(cpu_buffer, event);
2948 if (rb_try_to_discard(cpu_buffer, event))
2949 goto out;
2950
2951 /*
2952 * The commit is still visible by the reader, so we
2953 * must still update the timestamp.
2954 */
2955 rb_update_write_stamp(cpu_buffer, event);
2956 out:
2957 rb_end_commit(cpu_buffer);
2958
2959 trace_recursive_unlock(cpu_buffer);
2960
2961 preempt_enable_notrace();
2962
2963}
2964EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2965
2966/**
2967 * ring_buffer_write - write data to the buffer without reserving
2968 * @buffer: The ring buffer to write to.
2969 * @length: The length of the data being written (excluding the event header)
2970 * @data: The data to write to the buffer.
2971 *
2972 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2973 * one function. If you already have the data to write to the buffer, it
2974 * may be easier to simply call this function.
2975 *
2976 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2977 * and not the length of the event which would hold the header.
2978 */
2979int ring_buffer_write(struct ring_buffer *buffer,
2980 unsigned long length,
2981 void *data)
2982{
2983 struct ring_buffer_per_cpu *cpu_buffer;
2984 struct ring_buffer_event *event;
2985 void *body;
2986 int ret = -EBUSY;
2987 int cpu;
2988
2989 preempt_disable_notrace();
2990
2991 if (atomic_read(&buffer->record_disabled))
2992 goto out;
2993
2994 cpu = raw_smp_processor_id();
2995
2996 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2997 goto out;
2998
2999 cpu_buffer = buffer->buffers[cpu];
3000
3001 if (atomic_read(&cpu_buffer->record_disabled))
3002 goto out;
3003
3004 if (length > BUF_MAX_DATA_SIZE)
3005 goto out;
3006
3007 if (unlikely(trace_recursive_lock(cpu_buffer)))
3008 goto out;
3009
3010 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3011 if (!event)
3012 goto out_unlock;
3013
3014 body = rb_event_data(event);
3015
3016 memcpy(body, data, length);
3017
3018 rb_commit(cpu_buffer, event);
3019
3020 rb_wakeups(buffer, cpu_buffer);
3021
3022 ret = 0;
3023
3024 out_unlock:
3025 trace_recursive_unlock(cpu_buffer);
3026
3027 out:
3028 preempt_enable_notrace();
3029
3030 return ret;
3031}
3032EXPORT_SYMBOL_GPL(ring_buffer_write);
3033
3034static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3035{
3036 struct buffer_page *reader = cpu_buffer->reader_page;
3037 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3038 struct buffer_page *commit = cpu_buffer->commit_page;
3039
3040 /* In case of error, head will be NULL */
3041 if (unlikely(!head))
3042 return true;
3043
3044 return reader->read == rb_page_commit(reader) &&
3045 (commit == reader ||
3046 (commit == head &&
3047 head->read == rb_page_commit(commit)));
3048}
3049
3050/**
3051 * ring_buffer_record_disable - stop all writes into the buffer
3052 * @buffer: The ring buffer to stop writes to.
3053 *
3054 * This prevents all writes to the buffer. Any attempt to write
3055 * to the buffer after this will fail and return NULL.
3056 *
3057 * The caller should call synchronize_sched() after this.
3058 */
3059void ring_buffer_record_disable(struct ring_buffer *buffer)
3060{
3061 atomic_inc(&buffer->record_disabled);
3062}
3063EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3064
3065/**
3066 * ring_buffer_record_enable - enable writes to the buffer
3067 * @buffer: The ring buffer to enable writes
3068 *
3069 * Note, multiple disables will need the same number of enables
3070 * to truly enable the writing (much like preempt_disable).
3071 */
3072void ring_buffer_record_enable(struct ring_buffer *buffer)
3073{
3074 atomic_dec(&buffer->record_disabled);
3075}
3076EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3077
3078/**
3079 * ring_buffer_record_off - stop all writes into the buffer
3080 * @buffer: The ring buffer to stop writes to.
3081 *
3082 * This prevents all writes to the buffer. Any attempt to write
3083 * to the buffer after this will fail and return NULL.
3084 *
3085 * This is different than ring_buffer_record_disable() as
3086 * it works like an on/off switch, where as the disable() version
3087 * must be paired with a enable().
3088 */
3089void ring_buffer_record_off(struct ring_buffer *buffer)
3090{
3091 unsigned int rd;
3092 unsigned int new_rd;
3093
3094 do {
3095 rd = atomic_read(&buffer->record_disabled);
3096 new_rd = rd | RB_BUFFER_OFF;
3097 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3098}
3099EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3100
3101/**
3102 * ring_buffer_record_on - restart writes into the buffer
3103 * @buffer: The ring buffer to start writes to.
3104 *
3105 * This enables all writes to the buffer that was disabled by
3106 * ring_buffer_record_off().
3107 *
3108 * This is different than ring_buffer_record_enable() as
3109 * it works like an on/off switch, where as the enable() version
3110 * must be paired with a disable().
3111 */
3112void ring_buffer_record_on(struct ring_buffer *buffer)
3113{
3114 unsigned int rd;
3115 unsigned int new_rd;
3116
3117 do {
3118 rd = atomic_read(&buffer->record_disabled);
3119 new_rd = rd & ~RB_BUFFER_OFF;
3120 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3121}
3122EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3123
3124/**
3125 * ring_buffer_record_is_on - return true if the ring buffer can write
3126 * @buffer: The ring buffer to see if write is enabled
3127 *
3128 * Returns true if the ring buffer is in a state that it accepts writes.
3129 */
3130int ring_buffer_record_is_on(struct ring_buffer *buffer)
3131{
3132 return !atomic_read(&buffer->record_disabled);
3133}
3134
3135/**
3136 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3137 * @buffer: The ring buffer to stop writes to.
3138 * @cpu: The CPU buffer to stop
3139 *
3140 * This prevents all writes to the buffer. Any attempt to write
3141 * to the buffer after this will fail and return NULL.
3142 *
3143 * The caller should call synchronize_sched() after this.
3144 */
3145void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3146{
3147 struct ring_buffer_per_cpu *cpu_buffer;
3148
3149 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3150 return;
3151
3152 cpu_buffer = buffer->buffers[cpu];
3153 atomic_inc(&cpu_buffer->record_disabled);
3154}
3155EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3156
3157/**
3158 * ring_buffer_record_enable_cpu - enable writes to the buffer
3159 * @buffer: The ring buffer to enable writes
3160 * @cpu: The CPU to enable.
3161 *
3162 * Note, multiple disables will need the same number of enables
3163 * to truly enable the writing (much like preempt_disable).
3164 */
3165void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3166{
3167 struct ring_buffer_per_cpu *cpu_buffer;
3168
3169 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3170 return;
3171
3172 cpu_buffer = buffer->buffers[cpu];
3173 atomic_dec(&cpu_buffer->record_disabled);
3174}
3175EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3176
3177/*
3178 * The total entries in the ring buffer is the running counter
3179 * of entries entered into the ring buffer, minus the sum of
3180 * the entries read from the ring buffer and the number of
3181 * entries that were overwritten.
3182 */
3183static inline unsigned long
3184rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3185{
3186 return local_read(&cpu_buffer->entries) -
3187 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3188}
3189
3190/**
3191 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3192 * @buffer: The ring buffer
3193 * @cpu: The per CPU buffer to read from.
3194 */
3195u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3196{
3197 unsigned long flags;
3198 struct ring_buffer_per_cpu *cpu_buffer;
3199 struct buffer_page *bpage;
3200 u64 ret = 0;
3201
3202 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3203 return 0;
3204
3205 cpu_buffer = buffer->buffers[cpu];
3206 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3207 /*
3208 * if the tail is on reader_page, oldest time stamp is on the reader
3209 * page
3210 */
3211 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3212 bpage = cpu_buffer->reader_page;
3213 else
3214 bpage = rb_set_head_page(cpu_buffer);
3215 if (bpage)
3216 ret = bpage->page->time_stamp;
3217 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3218
3219 return ret;
3220}
3221EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3222
3223/**
3224 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3225 * @buffer: The ring buffer
3226 * @cpu: The per CPU buffer to read from.
3227 */
3228unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3229{
3230 struct ring_buffer_per_cpu *cpu_buffer;
3231 unsigned long ret;
3232
3233 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3234 return 0;
3235
3236 cpu_buffer = buffer->buffers[cpu];
3237 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3238
3239 return ret;
3240}
3241EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3242
3243/**
3244 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3245 * @buffer: The ring buffer
3246 * @cpu: The per CPU buffer to get the entries from.
3247 */
3248unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3249{
3250 struct ring_buffer_per_cpu *cpu_buffer;
3251
3252 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3253 return 0;
3254
3255 cpu_buffer = buffer->buffers[cpu];
3256
3257 return rb_num_of_entries(cpu_buffer);
3258}
3259EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3260
3261/**
3262 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3263 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3264 * @buffer: The ring buffer
3265 * @cpu: The per CPU buffer to get the number of overruns from
3266 */
3267unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3268{
3269 struct ring_buffer_per_cpu *cpu_buffer;
3270 unsigned long ret;
3271
3272 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3273 return 0;
3274
3275 cpu_buffer = buffer->buffers[cpu];
3276 ret = local_read(&cpu_buffer->overrun);
3277
3278 return ret;
3279}
3280EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3281
3282/**
3283 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3284 * commits failing due to the buffer wrapping around while there are uncommitted
3285 * events, such as during an interrupt storm.
3286 * @buffer: The ring buffer
3287 * @cpu: The per CPU buffer to get the number of overruns from
3288 */
3289unsigned long
3290ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3291{
3292 struct ring_buffer_per_cpu *cpu_buffer;
3293 unsigned long ret;
3294
3295 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3296 return 0;
3297
3298 cpu_buffer = buffer->buffers[cpu];
3299 ret = local_read(&cpu_buffer->commit_overrun);
3300
3301 return ret;
3302}
3303EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3304
3305/**
3306 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3307 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3308 * @buffer: The ring buffer
3309 * @cpu: The per CPU buffer to get the number of overruns from
3310 */
3311unsigned long
3312ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3313{
3314 struct ring_buffer_per_cpu *cpu_buffer;
3315 unsigned long ret;
3316
3317 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3318 return 0;
3319
3320 cpu_buffer = buffer->buffers[cpu];
3321 ret = local_read(&cpu_buffer->dropped_events);
3322
3323 return ret;
3324}
3325EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3326
3327/**
3328 * ring_buffer_read_events_cpu - get the number of events successfully read
3329 * @buffer: The ring buffer
3330 * @cpu: The per CPU buffer to get the number of events read
3331 */
3332unsigned long
3333ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3334{
3335 struct ring_buffer_per_cpu *cpu_buffer;
3336
3337 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3338 return 0;
3339
3340 cpu_buffer = buffer->buffers[cpu];
3341 return cpu_buffer->read;
3342}
3343EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3344
3345/**
3346 * ring_buffer_entries - get the number of entries in a buffer
3347 * @buffer: The ring buffer
3348 *
3349 * Returns the total number of entries in the ring buffer
3350 * (all CPU entries)
3351 */
3352unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3353{
3354 struct ring_buffer_per_cpu *cpu_buffer;
3355 unsigned long entries = 0;
3356 int cpu;
3357
3358 /* if you care about this being correct, lock the buffer */
3359 for_each_buffer_cpu(buffer, cpu) {
3360 cpu_buffer = buffer->buffers[cpu];
3361 entries += rb_num_of_entries(cpu_buffer);
3362 }
3363
3364 return entries;
3365}
3366EXPORT_SYMBOL_GPL(ring_buffer_entries);
3367
3368/**
3369 * ring_buffer_overruns - get the number of overruns in buffer
3370 * @buffer: The ring buffer
3371 *
3372 * Returns the total number of overruns in the ring buffer
3373 * (all CPU entries)
3374 */
3375unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3376{
3377 struct ring_buffer_per_cpu *cpu_buffer;
3378 unsigned long overruns = 0;
3379 int cpu;
3380
3381 /* if you care about this being correct, lock the buffer */
3382 for_each_buffer_cpu(buffer, cpu) {
3383 cpu_buffer = buffer->buffers[cpu];
3384 overruns += local_read(&cpu_buffer->overrun);
3385 }
3386
3387 return overruns;
3388}
3389EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3390
3391static void rb_iter_reset(struct ring_buffer_iter *iter)
3392{
3393 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3394
3395 /* Iterator usage is expected to have record disabled */
3396 iter->head_page = cpu_buffer->reader_page;
3397 iter->head = cpu_buffer->reader_page->read;
3398
3399 iter->cache_reader_page = iter->head_page;
3400 iter->cache_read = cpu_buffer->read;
3401
3402 if (iter->head)
3403 iter->read_stamp = cpu_buffer->read_stamp;
3404 else
3405 iter->read_stamp = iter->head_page->page->time_stamp;
3406}
3407
3408/**
3409 * ring_buffer_iter_reset - reset an iterator
3410 * @iter: The iterator to reset
3411 *
3412 * Resets the iterator, so that it will start from the beginning
3413 * again.
3414 */
3415void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3416{
3417 struct ring_buffer_per_cpu *cpu_buffer;
3418 unsigned long flags;
3419
3420 if (!iter)
3421 return;
3422
3423 cpu_buffer = iter->cpu_buffer;
3424
3425 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3426 rb_iter_reset(iter);
3427 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3428}
3429EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3430
3431/**
3432 * ring_buffer_iter_empty - check if an iterator has no more to read
3433 * @iter: The iterator to check
3434 */
3435int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3436{
3437 struct ring_buffer_per_cpu *cpu_buffer;
3438
3439 cpu_buffer = iter->cpu_buffer;
3440
3441 return iter->head_page == cpu_buffer->commit_page &&
3442 iter->head == rb_commit_index(cpu_buffer);
3443}
3444EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3445
3446static void
3447rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3448 struct ring_buffer_event *event)
3449{
3450 u64 delta;
3451
3452 switch (event->type_len) {
3453 case RINGBUF_TYPE_PADDING:
3454 return;
3455
3456 case RINGBUF_TYPE_TIME_EXTEND:
3457 delta = event->array[0];
3458 delta <<= TS_SHIFT;
3459 delta += event->time_delta;
3460 cpu_buffer->read_stamp += delta;
3461 return;
3462
3463 case RINGBUF_TYPE_TIME_STAMP:
3464 /* FIXME: not implemented */
3465 return;
3466
3467 case RINGBUF_TYPE_DATA:
3468 cpu_buffer->read_stamp += event->time_delta;
3469 return;
3470
3471 default:
3472 BUG();
3473 }
3474 return;
3475}
3476
3477static void
3478rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3479 struct ring_buffer_event *event)
3480{
3481 u64 delta;
3482
3483 switch (event->type_len) {
3484 case RINGBUF_TYPE_PADDING:
3485 return;
3486
3487 case RINGBUF_TYPE_TIME_EXTEND:
3488 delta = event->array[0];
3489 delta <<= TS_SHIFT;
3490 delta += event->time_delta;
3491 iter->read_stamp += delta;
3492 return;
3493
3494 case RINGBUF_TYPE_TIME_STAMP:
3495 /* FIXME: not implemented */
3496 return;
3497
3498 case RINGBUF_TYPE_DATA:
3499 iter->read_stamp += event->time_delta;
3500 return;
3501
3502 default:
3503 BUG();
3504 }
3505 return;
3506}
3507
3508static struct buffer_page *
3509rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3510{
3511 struct buffer_page *reader = NULL;
3512 unsigned long overwrite;
3513 unsigned long flags;
3514 int nr_loops = 0;
3515 int ret;
3516
3517 local_irq_save(flags);
3518 arch_spin_lock(&cpu_buffer->lock);
3519
3520 again:
3521 /*
3522 * This should normally only loop twice. But because the
3523 * start of the reader inserts an empty page, it causes
3524 * a case where we will loop three times. There should be no
3525 * reason to loop four times (that I know of).
3526 */
3527 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3528 reader = NULL;
3529 goto out;
3530 }
3531
3532 reader = cpu_buffer->reader_page;
3533
3534 /* If there's more to read, return this page */
3535 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3536 goto out;
3537
3538 /* Never should we have an index greater than the size */
3539 if (RB_WARN_ON(cpu_buffer,
3540 cpu_buffer->reader_page->read > rb_page_size(reader)))
3541 goto out;
3542
3543 /* check if we caught up to the tail */
3544 reader = NULL;
3545 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3546 goto out;
3547
3548 /* Don't bother swapping if the ring buffer is empty */
3549 if (rb_num_of_entries(cpu_buffer) == 0)
3550 goto out;
3551
3552 /*
3553 * Reset the reader page to size zero.
3554 */
3555 local_set(&cpu_buffer->reader_page->write, 0);
3556 local_set(&cpu_buffer->reader_page->entries, 0);
3557 local_set(&cpu_buffer->reader_page->page->commit, 0);
3558 cpu_buffer->reader_page->real_end = 0;
3559
3560 spin:
3561 /*
3562 * Splice the empty reader page into the list around the head.
3563 */
3564 reader = rb_set_head_page(cpu_buffer);
3565 if (!reader)
3566 goto out;
3567 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3568 cpu_buffer->reader_page->list.prev = reader->list.prev;
3569
3570 /*
3571 * cpu_buffer->pages just needs to point to the buffer, it
3572 * has no specific buffer page to point to. Lets move it out
3573 * of our way so we don't accidentally swap it.
3574 */
3575 cpu_buffer->pages = reader->list.prev;
3576
3577 /* The reader page will be pointing to the new head */
3578 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3579
3580 /*
3581 * We want to make sure we read the overruns after we set up our
3582 * pointers to the next object. The writer side does a
3583 * cmpxchg to cross pages which acts as the mb on the writer
3584 * side. Note, the reader will constantly fail the swap
3585 * while the writer is updating the pointers, so this
3586 * guarantees that the overwrite recorded here is the one we
3587 * want to compare with the last_overrun.
3588 */
3589 smp_mb();
3590 overwrite = local_read(&(cpu_buffer->overrun));
3591
3592 /*
3593 * Here's the tricky part.
3594 *
3595 * We need to move the pointer past the header page.
3596 * But we can only do that if a writer is not currently
3597 * moving it. The page before the header page has the
3598 * flag bit '1' set if it is pointing to the page we want.
3599 * but if the writer is in the process of moving it
3600 * than it will be '2' or already moved '0'.
3601 */
3602
3603 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3604
3605 /*
3606 * If we did not convert it, then we must try again.
3607 */
3608 if (!ret)
3609 goto spin;
3610
3611 /*
3612 * Yeah! We succeeded in replacing the page.
3613 *
3614 * Now make the new head point back to the reader page.
3615 */
3616 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3617 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3618
3619 /* Finally update the reader page to the new head */
3620 cpu_buffer->reader_page = reader;
3621 cpu_buffer->reader_page->read = 0;
3622
3623 if (overwrite != cpu_buffer->last_overrun) {
3624 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3625 cpu_buffer->last_overrun = overwrite;
3626 }
3627
3628 goto again;
3629
3630 out:
3631 /* Update the read_stamp on the first event */
3632 if (reader && reader->read == 0)
3633 cpu_buffer->read_stamp = reader->page->time_stamp;
3634
3635 arch_spin_unlock(&cpu_buffer->lock);
3636 local_irq_restore(flags);
3637
3638 return reader;
3639}
3640
3641static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3642{
3643 struct ring_buffer_event *event;
3644 struct buffer_page *reader;
3645 unsigned length;
3646
3647 reader = rb_get_reader_page(cpu_buffer);
3648
3649 /* This function should not be called when buffer is empty */
3650 if (RB_WARN_ON(cpu_buffer, !reader))
3651 return;
3652
3653 event = rb_reader_event(cpu_buffer);
3654
3655 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3656 cpu_buffer->read++;
3657
3658 rb_update_read_stamp(cpu_buffer, event);
3659
3660 length = rb_event_length(event);
3661 cpu_buffer->reader_page->read += length;
3662}
3663
3664static void rb_advance_iter(struct ring_buffer_iter *iter)
3665{
3666 struct ring_buffer_per_cpu *cpu_buffer;
3667 struct ring_buffer_event *event;
3668 unsigned length;
3669
3670 cpu_buffer = iter->cpu_buffer;
3671
3672 /*
3673 * Check if we are at the end of the buffer.
3674 */
3675 if (iter->head >= rb_page_size(iter->head_page)) {
3676 /* discarded commits can make the page empty */
3677 if (iter->head_page == cpu_buffer->commit_page)
3678 return;
3679 rb_inc_iter(iter);
3680 return;
3681 }
3682
3683 event = rb_iter_head_event(iter);
3684
3685 length = rb_event_length(event);
3686
3687 /*
3688 * This should not be called to advance the header if we are
3689 * at the tail of the buffer.
3690 */
3691 if (RB_WARN_ON(cpu_buffer,
3692 (iter->head_page == cpu_buffer->commit_page) &&
3693 (iter->head + length > rb_commit_index(cpu_buffer))))
3694 return;
3695
3696 rb_update_iter_read_stamp(iter, event);
3697
3698 iter->head += length;
3699
3700 /* check for end of page padding */
3701 if ((iter->head >= rb_page_size(iter->head_page)) &&
3702 (iter->head_page != cpu_buffer->commit_page))
3703 rb_inc_iter(iter);
3704}
3705
3706static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3707{
3708 return cpu_buffer->lost_events;
3709}
3710
3711static struct ring_buffer_event *
3712rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3713 unsigned long *lost_events)
3714{
3715 struct ring_buffer_event *event;
3716 struct buffer_page *reader;
3717 int nr_loops = 0;
3718
3719 again:
3720 /*
3721 * We repeat when a time extend is encountered.
3722 * Since the time extend is always attached to a data event,
3723 * we should never loop more than once.
3724 * (We never hit the following condition more than twice).
3725 */
3726 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3727 return NULL;
3728
3729 reader = rb_get_reader_page(cpu_buffer);
3730 if (!reader)
3731 return NULL;
3732
3733 event = rb_reader_event(cpu_buffer);
3734
3735 switch (event->type_len) {
3736 case RINGBUF_TYPE_PADDING:
3737 if (rb_null_event(event))
3738 RB_WARN_ON(cpu_buffer, 1);
3739 /*
3740 * Because the writer could be discarding every
3741 * event it creates (which would probably be bad)
3742 * if we were to go back to "again" then we may never
3743 * catch up, and will trigger the warn on, or lock
3744 * the box. Return the padding, and we will release
3745 * the current locks, and try again.
3746 */
3747 return event;
3748
3749 case RINGBUF_TYPE_TIME_EXTEND:
3750 /* Internal data, OK to advance */
3751 rb_advance_reader(cpu_buffer);
3752 goto again;
3753
3754 case RINGBUF_TYPE_TIME_STAMP:
3755 /* FIXME: not implemented */
3756 rb_advance_reader(cpu_buffer);
3757 goto again;
3758
3759 case RINGBUF_TYPE_DATA:
3760 if (ts) {
3761 *ts = cpu_buffer->read_stamp + event->time_delta;
3762 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3763 cpu_buffer->cpu, ts);
3764 }
3765 if (lost_events)
3766 *lost_events = rb_lost_events(cpu_buffer);
3767 return event;
3768
3769 default:
3770 BUG();
3771 }
3772
3773 return NULL;
3774}
3775EXPORT_SYMBOL_GPL(ring_buffer_peek);
3776
3777static struct ring_buffer_event *
3778rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3779{
3780 struct ring_buffer *buffer;
3781 struct ring_buffer_per_cpu *cpu_buffer;
3782 struct ring_buffer_event *event;
3783 int nr_loops = 0;
3784
3785 cpu_buffer = iter->cpu_buffer;
3786 buffer = cpu_buffer->buffer;
3787
3788 /*
3789 * Check if someone performed a consuming read to
3790 * the buffer. A consuming read invalidates the iterator
3791 * and we need to reset the iterator in this case.
3792 */
3793 if (unlikely(iter->cache_read != cpu_buffer->read ||
3794 iter->cache_reader_page != cpu_buffer->reader_page))
3795 rb_iter_reset(iter);
3796
3797 again:
3798 if (ring_buffer_iter_empty(iter))
3799 return NULL;
3800
3801 /*
3802 * We repeat when a time extend is encountered or we hit
3803 * the end of the page. Since the time extend is always attached
3804 * to a data event, we should never loop more than three times.
3805 * Once for going to next page, once on time extend, and
3806 * finally once to get the event.
3807 * (We never hit the following condition more than thrice).
3808 */
3809 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3810 return NULL;
3811
3812 if (rb_per_cpu_empty(cpu_buffer))
3813 return NULL;
3814
3815 if (iter->head >= rb_page_size(iter->head_page)) {
3816 rb_inc_iter(iter);
3817 goto again;
3818 }
3819
3820 event = rb_iter_head_event(iter);
3821
3822 switch (event->type_len) {
3823 case RINGBUF_TYPE_PADDING:
3824 if (rb_null_event(event)) {
3825 rb_inc_iter(iter);
3826 goto again;
3827 }
3828 rb_advance_iter(iter);
3829 return event;
3830
3831 case RINGBUF_TYPE_TIME_EXTEND:
3832 /* Internal data, OK to advance */
3833 rb_advance_iter(iter);
3834 goto again;
3835
3836 case RINGBUF_TYPE_TIME_STAMP:
3837 /* FIXME: not implemented */
3838 rb_advance_iter(iter);
3839 goto again;
3840
3841 case RINGBUF_TYPE_DATA:
3842 if (ts) {
3843 *ts = iter->read_stamp + event->time_delta;
3844 ring_buffer_normalize_time_stamp(buffer,
3845 cpu_buffer->cpu, ts);
3846 }
3847 return event;
3848
3849 default:
3850 BUG();
3851 }
3852
3853 return NULL;
3854}
3855EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3856
3857static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
3858{
3859 if (likely(!in_nmi())) {
3860 raw_spin_lock(&cpu_buffer->reader_lock);
3861 return true;
3862 }
3863
3864 /*
3865 * If an NMI die dumps out the content of the ring buffer
3866 * trylock must be used to prevent a deadlock if the NMI
3867 * preempted a task that holds the ring buffer locks. If
3868 * we get the lock then all is fine, if not, then continue
3869 * to do the read, but this can corrupt the ring buffer,
3870 * so it must be permanently disabled from future writes.
3871 * Reading from NMI is a oneshot deal.
3872 */
3873 if (raw_spin_trylock(&cpu_buffer->reader_lock))
3874 return true;
3875
3876 /* Continue without locking, but disable the ring buffer */
3877 atomic_inc(&cpu_buffer->record_disabled);
3878 return false;
3879}
3880
3881static inline void
3882rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
3883{
3884 if (likely(locked))
3885 raw_spin_unlock(&cpu_buffer->reader_lock);
3886 return;
3887}
3888
3889/**
3890 * ring_buffer_peek - peek at the next event to be read
3891 * @buffer: The ring buffer to read
3892 * @cpu: The cpu to peak at
3893 * @ts: The timestamp counter of this event.
3894 * @lost_events: a variable to store if events were lost (may be NULL)
3895 *
3896 * This will return the event that will be read next, but does
3897 * not consume the data.
3898 */
3899struct ring_buffer_event *
3900ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3901 unsigned long *lost_events)
3902{
3903 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3904 struct ring_buffer_event *event;
3905 unsigned long flags;
3906 bool dolock;
3907
3908 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3909 return NULL;
3910
3911 again:
3912 local_irq_save(flags);
3913 dolock = rb_reader_lock(cpu_buffer);
3914 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3915 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3916 rb_advance_reader(cpu_buffer);
3917 rb_reader_unlock(cpu_buffer, dolock);
3918 local_irq_restore(flags);
3919
3920 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3921 goto again;
3922
3923 return event;
3924}
3925
3926/**
3927 * ring_buffer_iter_peek - peek at the next event to be read
3928 * @iter: The ring buffer iterator
3929 * @ts: The timestamp counter of this event.
3930 *
3931 * This will return the event that will be read next, but does
3932 * not increment the iterator.
3933 */
3934struct ring_buffer_event *
3935ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3936{
3937 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3938 struct ring_buffer_event *event;
3939 unsigned long flags;
3940
3941 again:
3942 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3943 event = rb_iter_peek(iter, ts);
3944 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3945
3946 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3947 goto again;
3948
3949 return event;
3950}
3951
3952/**
3953 * ring_buffer_consume - return an event and consume it
3954 * @buffer: The ring buffer to get the next event from
3955 * @cpu: the cpu to read the buffer from
3956 * @ts: a variable to store the timestamp (may be NULL)
3957 * @lost_events: a variable to store if events were lost (may be NULL)
3958 *
3959 * Returns the next event in the ring buffer, and that event is consumed.
3960 * Meaning, that sequential reads will keep returning a different event,
3961 * and eventually empty the ring buffer if the producer is slower.
3962 */
3963struct ring_buffer_event *
3964ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3965 unsigned long *lost_events)
3966{
3967 struct ring_buffer_per_cpu *cpu_buffer;
3968 struct ring_buffer_event *event = NULL;
3969 unsigned long flags;
3970 bool dolock;
3971
3972 again:
3973 /* might be called in atomic */
3974 preempt_disable();
3975
3976 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3977 goto out;
3978
3979 cpu_buffer = buffer->buffers[cpu];
3980 local_irq_save(flags);
3981 dolock = rb_reader_lock(cpu_buffer);
3982
3983 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3984 if (event) {
3985 cpu_buffer->lost_events = 0;
3986 rb_advance_reader(cpu_buffer);
3987 }
3988
3989 rb_reader_unlock(cpu_buffer, dolock);
3990 local_irq_restore(flags);
3991
3992 out:
3993 preempt_enable();
3994
3995 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3996 goto again;
3997
3998 return event;
3999}
4000EXPORT_SYMBOL_GPL(ring_buffer_consume);
4001
4002/**
4003 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4004 * @buffer: The ring buffer to read from
4005 * @cpu: The cpu buffer to iterate over
4006 *
4007 * This performs the initial preparations necessary to iterate
4008 * through the buffer. Memory is allocated, buffer recording
4009 * is disabled, and the iterator pointer is returned to the caller.
4010 *
4011 * Disabling buffer recordng prevents the reading from being
4012 * corrupted. This is not a consuming read, so a producer is not
4013 * expected.
4014 *
4015 * After a sequence of ring_buffer_read_prepare calls, the user is
4016 * expected to make at least one call to ring_buffer_read_prepare_sync.
4017 * Afterwards, ring_buffer_read_start is invoked to get things going
4018 * for real.
4019 *
4020 * This overall must be paired with ring_buffer_read_finish.
4021 */
4022struct ring_buffer_iter *
4023ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
4024{
4025 struct ring_buffer_per_cpu *cpu_buffer;
4026 struct ring_buffer_iter *iter;
4027
4028 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4029 return NULL;
4030
4031 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
4032 if (!iter)
4033 return NULL;
4034
4035 cpu_buffer = buffer->buffers[cpu];
4036
4037 iter->cpu_buffer = cpu_buffer;
4038
4039 atomic_inc(&buffer->resize_disabled);
4040 atomic_inc(&cpu_buffer->record_disabled);
4041
4042 return iter;
4043}
4044EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4045
4046/**
4047 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4048 *
4049 * All previously invoked ring_buffer_read_prepare calls to prepare
4050 * iterators will be synchronized. Afterwards, read_buffer_read_start
4051 * calls on those iterators are allowed.
4052 */
4053void
4054ring_buffer_read_prepare_sync(void)
4055{
4056 synchronize_sched();
4057}
4058EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4059
4060/**
4061 * ring_buffer_read_start - start a non consuming read of the buffer
4062 * @iter: The iterator returned by ring_buffer_read_prepare
4063 *
4064 * This finalizes the startup of an iteration through the buffer.
4065 * The iterator comes from a call to ring_buffer_read_prepare and
4066 * an intervening ring_buffer_read_prepare_sync must have been
4067 * performed.
4068 *
4069 * Must be paired with ring_buffer_read_finish.
4070 */
4071void
4072ring_buffer_read_start(struct ring_buffer_iter *iter)
4073{
4074 struct ring_buffer_per_cpu *cpu_buffer;
4075 unsigned long flags;
4076
4077 if (!iter)
4078 return;
4079
4080 cpu_buffer = iter->cpu_buffer;
4081
4082 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4083 arch_spin_lock(&cpu_buffer->lock);
4084 rb_iter_reset(iter);
4085 arch_spin_unlock(&cpu_buffer->lock);
4086 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4087}
4088EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4089
4090/**
4091 * ring_buffer_read_finish - finish reading the iterator of the buffer
4092 * @iter: The iterator retrieved by ring_buffer_start
4093 *
4094 * This re-enables the recording to the buffer, and frees the
4095 * iterator.
4096 */
4097void
4098ring_buffer_read_finish(struct ring_buffer_iter *iter)
4099{
4100 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4101 unsigned long flags;
4102
4103 /*
4104 * Ring buffer is disabled from recording, here's a good place
4105 * to check the integrity of the ring buffer.
4106 * Must prevent readers from trying to read, as the check
4107 * clears the HEAD page and readers require it.
4108 */
4109 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4110 rb_check_pages(cpu_buffer);
4111 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4112
4113 atomic_dec(&cpu_buffer->record_disabled);
4114 atomic_dec(&cpu_buffer->buffer->resize_disabled);
4115 kfree(iter);
4116}
4117EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4118
4119/**
4120 * ring_buffer_read - read the next item in the ring buffer by the iterator
4121 * @iter: The ring buffer iterator
4122 * @ts: The time stamp of the event read.
4123 *
4124 * This reads the next event in the ring buffer and increments the iterator.
4125 */
4126struct ring_buffer_event *
4127ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4128{
4129 struct ring_buffer_event *event;
4130 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4131 unsigned long flags;
4132
4133 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4134 again:
4135 event = rb_iter_peek(iter, ts);
4136 if (!event)
4137 goto out;
4138
4139 if (event->type_len == RINGBUF_TYPE_PADDING)
4140 goto again;
4141
4142 rb_advance_iter(iter);
4143 out:
4144 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4145
4146 return event;
4147}
4148EXPORT_SYMBOL_GPL(ring_buffer_read);
4149
4150/**
4151 * ring_buffer_size - return the size of the ring buffer (in bytes)
4152 * @buffer: The ring buffer.
4153 */
4154unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4155{
4156 /*
4157 * Earlier, this method returned
4158 * BUF_PAGE_SIZE * buffer->nr_pages
4159 * Since the nr_pages field is now removed, we have converted this to
4160 * return the per cpu buffer value.
4161 */
4162 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4163 return 0;
4164
4165 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4166}
4167EXPORT_SYMBOL_GPL(ring_buffer_size);
4168
4169static void
4170rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4171{
4172 rb_head_page_deactivate(cpu_buffer);
4173
4174 cpu_buffer->head_page
4175 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4176 local_set(&cpu_buffer->head_page->write, 0);
4177 local_set(&cpu_buffer->head_page->entries, 0);
4178 local_set(&cpu_buffer->head_page->page->commit, 0);
4179
4180 cpu_buffer->head_page->read = 0;
4181
4182 cpu_buffer->tail_page = cpu_buffer->head_page;
4183 cpu_buffer->commit_page = cpu_buffer->head_page;
4184
4185 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4186 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4187 local_set(&cpu_buffer->reader_page->write, 0);
4188 local_set(&cpu_buffer->reader_page->entries, 0);
4189 local_set(&cpu_buffer->reader_page->page->commit, 0);
4190 cpu_buffer->reader_page->read = 0;
4191
4192 local_set(&cpu_buffer->entries_bytes, 0);
4193 local_set(&cpu_buffer->overrun, 0);
4194 local_set(&cpu_buffer->commit_overrun, 0);
4195 local_set(&cpu_buffer->dropped_events, 0);
4196 local_set(&cpu_buffer->entries, 0);
4197 local_set(&cpu_buffer->committing, 0);
4198 local_set(&cpu_buffer->commits, 0);
4199 cpu_buffer->read = 0;
4200 cpu_buffer->read_bytes = 0;
4201
4202 cpu_buffer->write_stamp = 0;
4203 cpu_buffer->read_stamp = 0;
4204
4205 cpu_buffer->lost_events = 0;
4206 cpu_buffer->last_overrun = 0;
4207
4208 rb_head_page_activate(cpu_buffer);
4209}
4210
4211/**
4212 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4213 * @buffer: The ring buffer to reset a per cpu buffer of
4214 * @cpu: The CPU buffer to be reset
4215 */
4216void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4217{
4218 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4219 unsigned long flags;
4220
4221 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4222 return;
4223
4224 atomic_inc(&buffer->resize_disabled);
4225 atomic_inc(&cpu_buffer->record_disabled);
4226
4227 /* Make sure all commits have finished */
4228 synchronize_sched();
4229
4230 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4231
4232 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4233 goto out;
4234
4235 arch_spin_lock(&cpu_buffer->lock);
4236
4237 rb_reset_cpu(cpu_buffer);
4238
4239 arch_spin_unlock(&cpu_buffer->lock);
4240
4241 out:
4242 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4243
4244 atomic_dec(&cpu_buffer->record_disabled);
4245 atomic_dec(&buffer->resize_disabled);
4246}
4247EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4248
4249/**
4250 * ring_buffer_reset - reset a ring buffer
4251 * @buffer: The ring buffer to reset all cpu buffers
4252 */
4253void ring_buffer_reset(struct ring_buffer *buffer)
4254{
4255 int cpu;
4256
4257 for_each_buffer_cpu(buffer, cpu)
4258 ring_buffer_reset_cpu(buffer, cpu);
4259}
4260EXPORT_SYMBOL_GPL(ring_buffer_reset);
4261
4262/**
4263 * rind_buffer_empty - is the ring buffer empty?
4264 * @buffer: The ring buffer to test
4265 */
4266bool ring_buffer_empty(struct ring_buffer *buffer)
4267{
4268 struct ring_buffer_per_cpu *cpu_buffer;
4269 unsigned long flags;
4270 bool dolock;
4271 int cpu;
4272 int ret;
4273
4274 /* yes this is racy, but if you don't like the race, lock the buffer */
4275 for_each_buffer_cpu(buffer, cpu) {
4276 cpu_buffer = buffer->buffers[cpu];
4277 local_irq_save(flags);
4278 dolock = rb_reader_lock(cpu_buffer);
4279 ret = rb_per_cpu_empty(cpu_buffer);
4280 rb_reader_unlock(cpu_buffer, dolock);
4281 local_irq_restore(flags);
4282
4283 if (!ret)
4284 return false;
4285 }
4286
4287 return true;
4288}
4289EXPORT_SYMBOL_GPL(ring_buffer_empty);
4290
4291/**
4292 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4293 * @buffer: The ring buffer
4294 * @cpu: The CPU buffer to test
4295 */
4296bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4297{
4298 struct ring_buffer_per_cpu *cpu_buffer;
4299 unsigned long flags;
4300 bool dolock;
4301 int ret;
4302
4303 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4304 return true;
4305
4306 cpu_buffer = buffer->buffers[cpu];
4307 local_irq_save(flags);
4308 dolock = rb_reader_lock(cpu_buffer);
4309 ret = rb_per_cpu_empty(cpu_buffer);
4310 rb_reader_unlock(cpu_buffer, dolock);
4311 local_irq_restore(flags);
4312
4313 return ret;
4314}
4315EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4316
4317#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4318/**
4319 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4320 * @buffer_a: One buffer to swap with
4321 * @buffer_b: The other buffer to swap with
4322 *
4323 * This function is useful for tracers that want to take a "snapshot"
4324 * of a CPU buffer and has another back up buffer lying around.
4325 * it is expected that the tracer handles the cpu buffer not being
4326 * used at the moment.
4327 */
4328int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4329 struct ring_buffer *buffer_b, int cpu)
4330{
4331 struct ring_buffer_per_cpu *cpu_buffer_a;
4332 struct ring_buffer_per_cpu *cpu_buffer_b;
4333 int ret = -EINVAL;
4334
4335 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4336 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4337 goto out;
4338
4339 cpu_buffer_a = buffer_a->buffers[cpu];
4340 cpu_buffer_b = buffer_b->buffers[cpu];
4341
4342 /* At least make sure the two buffers are somewhat the same */
4343 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4344 goto out;
4345
4346 ret = -EAGAIN;
4347
4348 if (atomic_read(&buffer_a->record_disabled))
4349 goto out;
4350
4351 if (atomic_read(&buffer_b->record_disabled))
4352 goto out;
4353
4354 if (atomic_read(&cpu_buffer_a->record_disabled))
4355 goto out;
4356
4357 if (atomic_read(&cpu_buffer_b->record_disabled))
4358 goto out;
4359
4360 /*
4361 * We can't do a synchronize_sched here because this
4362 * function can be called in atomic context.
4363 * Normally this will be called from the same CPU as cpu.
4364 * If not it's up to the caller to protect this.
4365 */
4366 atomic_inc(&cpu_buffer_a->record_disabled);
4367 atomic_inc(&cpu_buffer_b->record_disabled);
4368
4369 ret = -EBUSY;
4370 if (local_read(&cpu_buffer_a->committing))
4371 goto out_dec;
4372 if (local_read(&cpu_buffer_b->committing))
4373 goto out_dec;
4374
4375 buffer_a->buffers[cpu] = cpu_buffer_b;
4376 buffer_b->buffers[cpu] = cpu_buffer_a;
4377
4378 cpu_buffer_b->buffer = buffer_a;
4379 cpu_buffer_a->buffer = buffer_b;
4380
4381 ret = 0;
4382
4383out_dec:
4384 atomic_dec(&cpu_buffer_a->record_disabled);
4385 atomic_dec(&cpu_buffer_b->record_disabled);
4386out:
4387 return ret;
4388}
4389EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4390#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4391
4392/**
4393 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4394 * @buffer: the buffer to allocate for.
4395 * @cpu: the cpu buffer to allocate.
4396 *
4397 * This function is used in conjunction with ring_buffer_read_page.
4398 * When reading a full page from the ring buffer, these functions
4399 * can be used to speed up the process. The calling function should
4400 * allocate a few pages first with this function. Then when it
4401 * needs to get pages from the ring buffer, it passes the result
4402 * of this function into ring_buffer_read_page, which will swap
4403 * the page that was allocated, with the read page of the buffer.
4404 *
4405 * Returns:
4406 * The page allocated, or NULL on error.
4407 */
4408void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4409{
4410 struct buffer_data_page *bpage;
4411 struct page *page;
4412
4413 page = alloc_pages_node(cpu_to_node(cpu),
4414 GFP_KERNEL | __GFP_NORETRY, 0);
4415 if (!page)
4416 return NULL;
4417
4418 bpage = page_address(page);
4419
4420 rb_init_page(bpage);
4421
4422 return bpage;
4423}
4424EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4425
4426/**
4427 * ring_buffer_free_read_page - free an allocated read page
4428 * @buffer: the buffer the page was allocate for
4429 * @data: the page to free
4430 *
4431 * Free a page allocated from ring_buffer_alloc_read_page.
4432 */
4433void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4434{
4435 free_page((unsigned long)data);
4436}
4437EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4438
4439/**
4440 * ring_buffer_read_page - extract a page from the ring buffer
4441 * @buffer: buffer to extract from
4442 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4443 * @len: amount to extract
4444 * @cpu: the cpu of the buffer to extract
4445 * @full: should the extraction only happen when the page is full.
4446 *
4447 * This function will pull out a page from the ring buffer and consume it.
4448 * @data_page must be the address of the variable that was returned
4449 * from ring_buffer_alloc_read_page. This is because the page might be used
4450 * to swap with a page in the ring buffer.
4451 *
4452 * for example:
4453 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4454 * if (!rpage)
4455 * return error;
4456 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4457 * if (ret >= 0)
4458 * process_page(rpage, ret);
4459 *
4460 * When @full is set, the function will not return true unless
4461 * the writer is off the reader page.
4462 *
4463 * Note: it is up to the calling functions to handle sleeps and wakeups.
4464 * The ring buffer can be used anywhere in the kernel and can not
4465 * blindly call wake_up. The layer that uses the ring buffer must be
4466 * responsible for that.
4467 *
4468 * Returns:
4469 * >=0 if data has been transferred, returns the offset of consumed data.
4470 * <0 if no data has been transferred.
4471 */
4472int ring_buffer_read_page(struct ring_buffer *buffer,
4473 void **data_page, size_t len, int cpu, int full)
4474{
4475 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4476 struct ring_buffer_event *event;
4477 struct buffer_data_page *bpage;
4478 struct buffer_page *reader;
4479 unsigned long missed_events;
4480 unsigned long flags;
4481 unsigned int commit;
4482 unsigned int read;
4483 u64 save_timestamp;
4484 int ret = -1;
4485
4486 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4487 goto out;
4488
4489 /*
4490 * If len is not big enough to hold the page header, then
4491 * we can not copy anything.
4492 */
4493 if (len <= BUF_PAGE_HDR_SIZE)
4494 goto out;
4495
4496 len -= BUF_PAGE_HDR_SIZE;
4497
4498 if (!data_page)
4499 goto out;
4500
4501 bpage = *data_page;
4502 if (!bpage)
4503 goto out;
4504
4505 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4506
4507 reader = rb_get_reader_page(cpu_buffer);
4508 if (!reader)
4509 goto out_unlock;
4510
4511 event = rb_reader_event(cpu_buffer);
4512
4513 read = reader->read;
4514 commit = rb_page_commit(reader);
4515
4516 /* Check if any events were dropped */
4517 missed_events = cpu_buffer->lost_events;
4518
4519 /*
4520 * If this page has been partially read or
4521 * if len is not big enough to read the rest of the page or
4522 * a writer is still on the page, then
4523 * we must copy the data from the page to the buffer.
4524 * Otherwise, we can simply swap the page with the one passed in.
4525 */
4526 if (read || (len < (commit - read)) ||
4527 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4528 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4529 unsigned int rpos = read;
4530 unsigned int pos = 0;
4531 unsigned int size;
4532
4533 if (full)
4534 goto out_unlock;
4535
4536 if (len > (commit - read))
4537 len = (commit - read);
4538
4539 /* Always keep the time extend and data together */
4540 size = rb_event_ts_length(event);
4541
4542 if (len < size)
4543 goto out_unlock;
4544
4545 /* save the current timestamp, since the user will need it */
4546 save_timestamp = cpu_buffer->read_stamp;
4547
4548 /* Need to copy one event at a time */
4549 do {
4550 /* We need the size of one event, because
4551 * rb_advance_reader only advances by one event,
4552 * whereas rb_event_ts_length may include the size of
4553 * one or two events.
4554 * We have already ensured there's enough space if this
4555 * is a time extend. */
4556 size = rb_event_length(event);
4557 memcpy(bpage->data + pos, rpage->data + rpos, size);
4558
4559 len -= size;
4560
4561 rb_advance_reader(cpu_buffer);
4562 rpos = reader->read;
4563 pos += size;
4564
4565 if (rpos >= commit)
4566 break;
4567
4568 event = rb_reader_event(cpu_buffer);
4569 /* Always keep the time extend and data together */
4570 size = rb_event_ts_length(event);
4571 } while (len >= size);
4572
4573 /* update bpage */
4574 local_set(&bpage->commit, pos);
4575 bpage->time_stamp = save_timestamp;
4576
4577 /* we copied everything to the beginning */
4578 read = 0;
4579 } else {
4580 /* update the entry counter */
4581 cpu_buffer->read += rb_page_entries(reader);
4582 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4583
4584 /* swap the pages */
4585 rb_init_page(bpage);
4586 bpage = reader->page;
4587 reader->page = *data_page;
4588 local_set(&reader->write, 0);
4589 local_set(&reader->entries, 0);
4590 reader->read = 0;
4591 *data_page = bpage;
4592
4593 /*
4594 * Use the real_end for the data size,
4595 * This gives us a chance to store the lost events
4596 * on the page.
4597 */
4598 if (reader->real_end)
4599 local_set(&bpage->commit, reader->real_end);
4600 }
4601 ret = read;
4602
4603 cpu_buffer->lost_events = 0;
4604
4605 commit = local_read(&bpage->commit);
4606 /*
4607 * Set a flag in the commit field if we lost events
4608 */
4609 if (missed_events) {
4610 /* If there is room at the end of the page to save the
4611 * missed events, then record it there.
4612 */
4613 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4614 memcpy(&bpage->data[commit], &missed_events,
4615 sizeof(missed_events));
4616 local_add(RB_MISSED_STORED, &bpage->commit);
4617 commit += sizeof(missed_events);
4618 }
4619 local_add(RB_MISSED_EVENTS, &bpage->commit);
4620 }
4621
4622 /*
4623 * This page may be off to user land. Zero it out here.
4624 */
4625 if (commit < BUF_PAGE_SIZE)
4626 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4627
4628 out_unlock:
4629 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4630
4631 out:
4632 return ret;
4633}
4634EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4635
4636#ifdef CONFIG_HOTPLUG_CPU
4637static int rb_cpu_notify(struct notifier_block *self,
4638 unsigned long action, void *hcpu)
4639{
4640 struct ring_buffer *buffer =
4641 container_of(self, struct ring_buffer, cpu_notify);
4642 long cpu = (long)hcpu;
4643 int cpu_i, nr_pages_same;
4644 unsigned int nr_pages;
4645
4646 switch (action) {
4647 case CPU_UP_PREPARE:
4648 case CPU_UP_PREPARE_FROZEN:
4649 if (cpumask_test_cpu(cpu, buffer->cpumask))
4650 return NOTIFY_OK;
4651
4652 nr_pages = 0;
4653 nr_pages_same = 1;
4654 /* check if all cpu sizes are same */
4655 for_each_buffer_cpu(buffer, cpu_i) {
4656 /* fill in the size from first enabled cpu */
4657 if (nr_pages == 0)
4658 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4659 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4660 nr_pages_same = 0;
4661 break;
4662 }
4663 }
4664 /* allocate minimum pages, user can later expand it */
4665 if (!nr_pages_same)
4666 nr_pages = 2;
4667 buffer->buffers[cpu] =
4668 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4669 if (!buffer->buffers[cpu]) {
4670 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4671 cpu);
4672 return NOTIFY_OK;
4673 }
4674 smp_wmb();
4675 cpumask_set_cpu(cpu, buffer->cpumask);
4676 break;
4677 case CPU_DOWN_PREPARE:
4678 case CPU_DOWN_PREPARE_FROZEN:
4679 /*
4680 * Do nothing.
4681 * If we were to free the buffer, then the user would
4682 * lose any trace that was in the buffer.
4683 */
4684 break;
4685 default:
4686 break;
4687 }
4688 return NOTIFY_OK;
4689}
4690#endif
4691
4692#ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4693/*
4694 * This is a basic integrity check of the ring buffer.
4695 * Late in the boot cycle this test will run when configured in.
4696 * It will kick off a thread per CPU that will go into a loop
4697 * writing to the per cpu ring buffer various sizes of data.
4698 * Some of the data will be large items, some small.
4699 *
4700 * Another thread is created that goes into a spin, sending out
4701 * IPIs to the other CPUs to also write into the ring buffer.
4702 * this is to test the nesting ability of the buffer.
4703 *
4704 * Basic stats are recorded and reported. If something in the
4705 * ring buffer should happen that's not expected, a big warning
4706 * is displayed and all ring buffers are disabled.
4707 */
4708static struct task_struct *rb_threads[NR_CPUS] __initdata;
4709
4710struct rb_test_data {
4711 struct ring_buffer *buffer;
4712 unsigned long events;
4713 unsigned long bytes_written;
4714 unsigned long bytes_alloc;
4715 unsigned long bytes_dropped;
4716 unsigned long events_nested;
4717 unsigned long bytes_written_nested;
4718 unsigned long bytes_alloc_nested;
4719 unsigned long bytes_dropped_nested;
4720 int min_size_nested;
4721 int max_size_nested;
4722 int max_size;
4723 int min_size;
4724 int cpu;
4725 int cnt;
4726};
4727
4728static struct rb_test_data rb_data[NR_CPUS] __initdata;
4729
4730/* 1 meg per cpu */
4731#define RB_TEST_BUFFER_SIZE 1048576
4732
4733static char rb_string[] __initdata =
4734 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4735 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4736 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4737
4738static bool rb_test_started __initdata;
4739
4740struct rb_item {
4741 int size;
4742 char str[];
4743};
4744
4745static __init int rb_write_something(struct rb_test_data *data, bool nested)
4746{
4747 struct ring_buffer_event *event;
4748 struct rb_item *item;
4749 bool started;
4750 int event_len;
4751 int size;
4752 int len;
4753 int cnt;
4754
4755 /* Have nested writes different that what is written */
4756 cnt = data->cnt + (nested ? 27 : 0);
4757
4758 /* Multiply cnt by ~e, to make some unique increment */
4759 size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4760
4761 len = size + sizeof(struct rb_item);
4762
4763 started = rb_test_started;
4764 /* read rb_test_started before checking buffer enabled */
4765 smp_rmb();
4766
4767 event = ring_buffer_lock_reserve(data->buffer, len);
4768 if (!event) {
4769 /* Ignore dropped events before test starts. */
4770 if (started) {
4771 if (nested)
4772 data->bytes_dropped += len;
4773 else
4774 data->bytes_dropped_nested += len;
4775 }
4776 return len;
4777 }
4778
4779 event_len = ring_buffer_event_length(event);
4780
4781 if (RB_WARN_ON(data->buffer, event_len < len))
4782 goto out;
4783
4784 item = ring_buffer_event_data(event);
4785 item->size = size;
4786 memcpy(item->str, rb_string, size);
4787
4788 if (nested) {
4789 data->bytes_alloc_nested += event_len;
4790 data->bytes_written_nested += len;
4791 data->events_nested++;
4792 if (!data->min_size_nested || len < data->min_size_nested)
4793 data->min_size_nested = len;
4794 if (len > data->max_size_nested)
4795 data->max_size_nested = len;
4796 } else {
4797 data->bytes_alloc += event_len;
4798 data->bytes_written += len;
4799 data->events++;
4800 if (!data->min_size || len < data->min_size)
4801 data->max_size = len;
4802 if (len > data->max_size)
4803 data->max_size = len;
4804 }
4805
4806 out:
4807 ring_buffer_unlock_commit(data->buffer, event);
4808
4809 return 0;
4810}
4811
4812static __init int rb_test(void *arg)
4813{
4814 struct rb_test_data *data = arg;
4815
4816 while (!kthread_should_stop()) {
4817 rb_write_something(data, false);
4818 data->cnt++;
4819
4820 set_current_state(TASK_INTERRUPTIBLE);
4821 /* Now sleep between a min of 100-300us and a max of 1ms */
4822 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4823 }
4824
4825 return 0;
4826}
4827
4828static __init void rb_ipi(void *ignore)
4829{
4830 struct rb_test_data *data;
4831 int cpu = smp_processor_id();
4832
4833 data = &rb_data[cpu];
4834 rb_write_something(data, true);
4835}
4836
4837static __init int rb_hammer_test(void *arg)
4838{
4839 while (!kthread_should_stop()) {
4840
4841 /* Send an IPI to all cpus to write data! */
4842 smp_call_function(rb_ipi, NULL, 1);
4843 /* No sleep, but for non preempt, let others run */
4844 schedule();
4845 }
4846
4847 return 0;
4848}
4849
4850static __init int test_ringbuffer(void)
4851{
4852 struct task_struct *rb_hammer;
4853 struct ring_buffer *buffer;
4854 int cpu;
4855 int ret = 0;
4856
4857 pr_info("Running ring buffer tests...\n");
4858
4859 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4860 if (WARN_ON(!buffer))
4861 return 0;
4862
4863 /* Disable buffer so that threads can't write to it yet */
4864 ring_buffer_record_off(buffer);
4865
4866 for_each_online_cpu(cpu) {
4867 rb_data[cpu].buffer = buffer;
4868 rb_data[cpu].cpu = cpu;
4869 rb_data[cpu].cnt = cpu;
4870 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4871 "rbtester/%d", cpu);
4872 if (WARN_ON(!rb_threads[cpu])) {
4873 pr_cont("FAILED\n");
4874 ret = -1;
4875 goto out_free;
4876 }
4877
4878 kthread_bind(rb_threads[cpu], cpu);
4879 wake_up_process(rb_threads[cpu]);
4880 }
4881
4882 /* Now create the rb hammer! */
4883 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4884 if (WARN_ON(!rb_hammer)) {
4885 pr_cont("FAILED\n");
4886 ret = -1;
4887 goto out_free;
4888 }
4889
4890 ring_buffer_record_on(buffer);
4891 /*
4892 * Show buffer is enabled before setting rb_test_started.
4893 * Yes there's a small race window where events could be
4894 * dropped and the thread wont catch it. But when a ring
4895 * buffer gets enabled, there will always be some kind of
4896 * delay before other CPUs see it. Thus, we don't care about
4897 * those dropped events. We care about events dropped after
4898 * the threads see that the buffer is active.
4899 */
4900 smp_wmb();
4901 rb_test_started = true;
4902
4903 set_current_state(TASK_INTERRUPTIBLE);
4904 /* Just run for 10 seconds */;
4905 schedule_timeout(10 * HZ);
4906
4907 kthread_stop(rb_hammer);
4908
4909 out_free:
4910 for_each_online_cpu(cpu) {
4911 if (!rb_threads[cpu])
4912 break;
4913 kthread_stop(rb_threads[cpu]);
4914 }
4915 if (ret) {
4916 ring_buffer_free(buffer);
4917 return ret;
4918 }
4919
4920 /* Report! */
4921 pr_info("finished\n");
4922 for_each_online_cpu(cpu) {
4923 struct ring_buffer_event *event;
4924 struct rb_test_data *data = &rb_data[cpu];
4925 struct rb_item *item;
4926 unsigned long total_events;
4927 unsigned long total_dropped;
4928 unsigned long total_written;
4929 unsigned long total_alloc;
4930 unsigned long total_read = 0;
4931 unsigned long total_size = 0;
4932 unsigned long total_len = 0;
4933 unsigned long total_lost = 0;
4934 unsigned long lost;
4935 int big_event_size;
4936 int small_event_size;
4937
4938 ret = -1;
4939
4940 total_events = data->events + data->events_nested;
4941 total_written = data->bytes_written + data->bytes_written_nested;
4942 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
4943 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
4944
4945 big_event_size = data->max_size + data->max_size_nested;
4946 small_event_size = data->min_size + data->min_size_nested;
4947
4948 pr_info("CPU %d:\n", cpu);
4949 pr_info(" events: %ld\n", total_events);
4950 pr_info(" dropped bytes: %ld\n", total_dropped);
4951 pr_info(" alloced bytes: %ld\n", total_alloc);
4952 pr_info(" written bytes: %ld\n", total_written);
4953 pr_info(" biggest event: %d\n", big_event_size);
4954 pr_info(" smallest event: %d\n", small_event_size);
4955
4956 if (RB_WARN_ON(buffer, total_dropped))
4957 break;
4958
4959 ret = 0;
4960
4961 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
4962 total_lost += lost;
4963 item = ring_buffer_event_data(event);
4964 total_len += ring_buffer_event_length(event);
4965 total_size += item->size + sizeof(struct rb_item);
4966 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
4967 pr_info("FAILED!\n");
4968 pr_info("buffer had: %.*s\n", item->size, item->str);
4969 pr_info("expected: %.*s\n", item->size, rb_string);
4970 RB_WARN_ON(buffer, 1);
4971 ret = -1;
4972 break;
4973 }
4974 total_read++;
4975 }
4976 if (ret)
4977 break;
4978
4979 ret = -1;
4980
4981 pr_info(" read events: %ld\n", total_read);
4982 pr_info(" lost events: %ld\n", total_lost);
4983 pr_info(" total events: %ld\n", total_lost + total_read);
4984 pr_info(" recorded len bytes: %ld\n", total_len);
4985 pr_info(" recorded size bytes: %ld\n", total_size);
4986 if (total_lost)
4987 pr_info(" With dropped events, record len and size may not match\n"
4988 " alloced and written from above\n");
4989 if (!total_lost) {
4990 if (RB_WARN_ON(buffer, total_len != total_alloc ||
4991 total_size != total_written))
4992 break;
4993 }
4994 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
4995 break;
4996
4997 ret = 0;
4998 }
4999 if (!ret)
5000 pr_info("Ring buffer PASSED!\n");
5001
5002 ring_buffer_free(buffer);
5003 return 0;
5004}
5005
5006late_initcall(test_ringbuffer);
5007#endif /* CONFIG_RING_BUFFER_STARTUP_TEST */
1/*
2 * Generic ring buffer
3 *
4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
5 */
6#include <linux/ring_buffer.h>
7#include <linux/trace_clock.h>
8#include <linux/spinlock.h>
9#include <linux/debugfs.h>
10#include <linux/uaccess.h>
11#include <linux/hardirq.h>
12#include <linux/kmemcheck.h>
13#include <linux/module.h>
14#include <linux/percpu.h>
15#include <linux/mutex.h>
16#include <linux/slab.h>
17#include <linux/init.h>
18#include <linux/hash.h>
19#include <linux/list.h>
20#include <linux/cpu.h>
21#include <linux/fs.h>
22
23#include <asm/local.h>
24#include "trace.h"
25
26/*
27 * The ring buffer header is special. We must manually up keep it.
28 */
29int ring_buffer_print_entry_header(struct trace_seq *s)
30{
31 int ret;
32
33 ret = trace_seq_printf(s, "# compressed entry header\n");
34 ret = trace_seq_printf(s, "\ttype_len : 5 bits\n");
35 ret = trace_seq_printf(s, "\ttime_delta : 27 bits\n");
36 ret = trace_seq_printf(s, "\tarray : 32 bits\n");
37 ret = trace_seq_printf(s, "\n");
38 ret = trace_seq_printf(s, "\tpadding : type == %d\n",
39 RINGBUF_TYPE_PADDING);
40 ret = trace_seq_printf(s, "\ttime_extend : type == %d\n",
41 RINGBUF_TYPE_TIME_EXTEND);
42 ret = trace_seq_printf(s, "\tdata max type_len == %d\n",
43 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
44
45 return ret;
46}
47
48/*
49 * The ring buffer is made up of a list of pages. A separate list of pages is
50 * allocated for each CPU. A writer may only write to a buffer that is
51 * associated with the CPU it is currently executing on. A reader may read
52 * from any per cpu buffer.
53 *
54 * The reader is special. For each per cpu buffer, the reader has its own
55 * reader page. When a reader has read the entire reader page, this reader
56 * page is swapped with another page in the ring buffer.
57 *
58 * Now, as long as the writer is off the reader page, the reader can do what
59 * ever it wants with that page. The writer will never write to that page
60 * again (as long as it is out of the ring buffer).
61 *
62 * Here's some silly ASCII art.
63 *
64 * +------+
65 * |reader| RING BUFFER
66 * |page |
67 * +------+ +---+ +---+ +---+
68 * | |-->| |-->| |
69 * +---+ +---+ +---+
70 * ^ |
71 * | |
72 * +---------------+
73 *
74 *
75 * +------+
76 * |reader| RING BUFFER
77 * |page |------------------v
78 * +------+ +---+ +---+ +---+
79 * | |-->| |-->| |
80 * +---+ +---+ +---+
81 * ^ |
82 * | |
83 * +---------------+
84 *
85 *
86 * +------+
87 * |reader| RING BUFFER
88 * |page |------------------v
89 * +------+ +---+ +---+ +---+
90 * ^ | |-->| |-->| |
91 * | +---+ +---+ +---+
92 * | |
93 * | |
94 * +------------------------------+
95 *
96 *
97 * +------+
98 * |buffer| RING BUFFER
99 * |page |------------------v
100 * +------+ +---+ +---+ +---+
101 * ^ | | | |-->| |
102 * | New +---+ +---+ +---+
103 * | Reader------^ |
104 * | page |
105 * +------------------------------+
106 *
107 *
108 * After we make this swap, the reader can hand this page off to the splice
109 * code and be done with it. It can even allocate a new page if it needs to
110 * and swap that into the ring buffer.
111 *
112 * We will be using cmpxchg soon to make all this lockless.
113 *
114 */
115
116/*
117 * A fast way to enable or disable all ring buffers is to
118 * call tracing_on or tracing_off. Turning off the ring buffers
119 * prevents all ring buffers from being recorded to.
120 * Turning this switch on, makes it OK to write to the
121 * ring buffer, if the ring buffer is enabled itself.
122 *
123 * There's three layers that must be on in order to write
124 * to the ring buffer.
125 *
126 * 1) This global flag must be set.
127 * 2) The ring buffer must be enabled for recording.
128 * 3) The per cpu buffer must be enabled for recording.
129 *
130 * In case of an anomaly, this global flag has a bit set that
131 * will permantly disable all ring buffers.
132 */
133
134/*
135 * Global flag to disable all recording to ring buffers
136 * This has two bits: ON, DISABLED
137 *
138 * ON DISABLED
139 * ---- ----------
140 * 0 0 : ring buffers are off
141 * 1 0 : ring buffers are on
142 * X 1 : ring buffers are permanently disabled
143 */
144
145enum {
146 RB_BUFFERS_ON_BIT = 0,
147 RB_BUFFERS_DISABLED_BIT = 1,
148};
149
150enum {
151 RB_BUFFERS_ON = 1 << RB_BUFFERS_ON_BIT,
152 RB_BUFFERS_DISABLED = 1 << RB_BUFFERS_DISABLED_BIT,
153};
154
155static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON;
156
157#define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
158
159/**
160 * tracing_on - enable all tracing buffers
161 *
162 * This function enables all tracing buffers that may have been
163 * disabled with tracing_off.
164 */
165void tracing_on(void)
166{
167 set_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
168}
169EXPORT_SYMBOL_GPL(tracing_on);
170
171/**
172 * tracing_off - turn off all tracing buffers
173 *
174 * This function stops all tracing buffers from recording data.
175 * It does not disable any overhead the tracers themselves may
176 * be causing. This function simply causes all recording to
177 * the ring buffers to fail.
178 */
179void tracing_off(void)
180{
181 clear_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
182}
183EXPORT_SYMBOL_GPL(tracing_off);
184
185/**
186 * tracing_off_permanent - permanently disable ring buffers
187 *
188 * This function, once called, will disable all ring buffers
189 * permanently.
190 */
191void tracing_off_permanent(void)
192{
193 set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags);
194}
195
196/**
197 * tracing_is_on - show state of ring buffers enabled
198 */
199int tracing_is_on(void)
200{
201 return ring_buffer_flags == RB_BUFFERS_ON;
202}
203EXPORT_SYMBOL_GPL(tracing_is_on);
204
205#define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
206#define RB_ALIGNMENT 4U
207#define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
208#define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
209
210#if !defined(CONFIG_64BIT) || defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
211# define RB_FORCE_8BYTE_ALIGNMENT 0
212# define RB_ARCH_ALIGNMENT RB_ALIGNMENT
213#else
214# define RB_FORCE_8BYTE_ALIGNMENT 1
215# define RB_ARCH_ALIGNMENT 8U
216#endif
217
218/* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
219#define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
220
221enum {
222 RB_LEN_TIME_EXTEND = 8,
223 RB_LEN_TIME_STAMP = 16,
224};
225
226#define skip_time_extend(event) \
227 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
228
229static inline int rb_null_event(struct ring_buffer_event *event)
230{
231 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
232}
233
234static void rb_event_set_padding(struct ring_buffer_event *event)
235{
236 /* padding has a NULL time_delta */
237 event->type_len = RINGBUF_TYPE_PADDING;
238 event->time_delta = 0;
239}
240
241static unsigned
242rb_event_data_length(struct ring_buffer_event *event)
243{
244 unsigned length;
245
246 if (event->type_len)
247 length = event->type_len * RB_ALIGNMENT;
248 else
249 length = event->array[0];
250 return length + RB_EVNT_HDR_SIZE;
251}
252
253/*
254 * Return the length of the given event. Will return
255 * the length of the time extend if the event is a
256 * time extend.
257 */
258static inline unsigned
259rb_event_length(struct ring_buffer_event *event)
260{
261 switch (event->type_len) {
262 case RINGBUF_TYPE_PADDING:
263 if (rb_null_event(event))
264 /* undefined */
265 return -1;
266 return event->array[0] + RB_EVNT_HDR_SIZE;
267
268 case RINGBUF_TYPE_TIME_EXTEND:
269 return RB_LEN_TIME_EXTEND;
270
271 case RINGBUF_TYPE_TIME_STAMP:
272 return RB_LEN_TIME_STAMP;
273
274 case RINGBUF_TYPE_DATA:
275 return rb_event_data_length(event);
276 default:
277 BUG();
278 }
279 /* not hit */
280 return 0;
281}
282
283/*
284 * Return total length of time extend and data,
285 * or just the event length for all other events.
286 */
287static inline unsigned
288rb_event_ts_length(struct ring_buffer_event *event)
289{
290 unsigned len = 0;
291
292 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
293 /* time extends include the data event after it */
294 len = RB_LEN_TIME_EXTEND;
295 event = skip_time_extend(event);
296 }
297 return len + rb_event_length(event);
298}
299
300/**
301 * ring_buffer_event_length - return the length of the event
302 * @event: the event to get the length of
303 *
304 * Returns the size of the data load of a data event.
305 * If the event is something other than a data event, it
306 * returns the size of the event itself. With the exception
307 * of a TIME EXTEND, where it still returns the size of the
308 * data load of the data event after it.
309 */
310unsigned ring_buffer_event_length(struct ring_buffer_event *event)
311{
312 unsigned length;
313
314 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
315 event = skip_time_extend(event);
316
317 length = rb_event_length(event);
318 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
319 return length;
320 length -= RB_EVNT_HDR_SIZE;
321 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
322 length -= sizeof(event->array[0]);
323 return length;
324}
325EXPORT_SYMBOL_GPL(ring_buffer_event_length);
326
327/* inline for ring buffer fast paths */
328static void *
329rb_event_data(struct ring_buffer_event *event)
330{
331 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
332 event = skip_time_extend(event);
333 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
334 /* If length is in len field, then array[0] has the data */
335 if (event->type_len)
336 return (void *)&event->array[0];
337 /* Otherwise length is in array[0] and array[1] has the data */
338 return (void *)&event->array[1];
339}
340
341/**
342 * ring_buffer_event_data - return the data of the event
343 * @event: the event to get the data from
344 */
345void *ring_buffer_event_data(struct ring_buffer_event *event)
346{
347 return rb_event_data(event);
348}
349EXPORT_SYMBOL_GPL(ring_buffer_event_data);
350
351#define for_each_buffer_cpu(buffer, cpu) \
352 for_each_cpu(cpu, buffer->cpumask)
353
354#define TS_SHIFT 27
355#define TS_MASK ((1ULL << TS_SHIFT) - 1)
356#define TS_DELTA_TEST (~TS_MASK)
357
358/* Flag when events were overwritten */
359#define RB_MISSED_EVENTS (1 << 31)
360/* Missed count stored at end */
361#define RB_MISSED_STORED (1 << 30)
362
363struct buffer_data_page {
364 u64 time_stamp; /* page time stamp */
365 local_t commit; /* write committed index */
366 unsigned char data[]; /* data of buffer page */
367};
368
369/*
370 * Note, the buffer_page list must be first. The buffer pages
371 * are allocated in cache lines, which means that each buffer
372 * page will be at the beginning of a cache line, and thus
373 * the least significant bits will be zero. We use this to
374 * add flags in the list struct pointers, to make the ring buffer
375 * lockless.
376 */
377struct buffer_page {
378 struct list_head list; /* list of buffer pages */
379 local_t write; /* index for next write */
380 unsigned read; /* index for next read */
381 local_t entries; /* entries on this page */
382 unsigned long real_end; /* real end of data */
383 struct buffer_data_page *page; /* Actual data page */
384};
385
386/*
387 * The buffer page counters, write and entries, must be reset
388 * atomically when crossing page boundaries. To synchronize this
389 * update, two counters are inserted into the number. One is
390 * the actual counter for the write position or count on the page.
391 *
392 * The other is a counter of updaters. Before an update happens
393 * the update partition of the counter is incremented. This will
394 * allow the updater to update the counter atomically.
395 *
396 * The counter is 20 bits, and the state data is 12.
397 */
398#define RB_WRITE_MASK 0xfffff
399#define RB_WRITE_INTCNT (1 << 20)
400
401static void rb_init_page(struct buffer_data_page *bpage)
402{
403 local_set(&bpage->commit, 0);
404}
405
406/**
407 * ring_buffer_page_len - the size of data on the page.
408 * @page: The page to read
409 *
410 * Returns the amount of data on the page, including buffer page header.
411 */
412size_t ring_buffer_page_len(void *page)
413{
414 return local_read(&((struct buffer_data_page *)page)->commit)
415 + BUF_PAGE_HDR_SIZE;
416}
417
418/*
419 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
420 * this issue out.
421 */
422static void free_buffer_page(struct buffer_page *bpage)
423{
424 free_page((unsigned long)bpage->page);
425 kfree(bpage);
426}
427
428/*
429 * We need to fit the time_stamp delta into 27 bits.
430 */
431static inline int test_time_stamp(u64 delta)
432{
433 if (delta & TS_DELTA_TEST)
434 return 1;
435 return 0;
436}
437
438#define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
439
440/* Max payload is BUF_PAGE_SIZE - header (8bytes) */
441#define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
442
443int ring_buffer_print_page_header(struct trace_seq *s)
444{
445 struct buffer_data_page field;
446 int ret;
447
448 ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t"
449 "offset:0;\tsize:%u;\tsigned:%u;\n",
450 (unsigned int)sizeof(field.time_stamp),
451 (unsigned int)is_signed_type(u64));
452
453 ret = trace_seq_printf(s, "\tfield: local_t commit;\t"
454 "offset:%u;\tsize:%u;\tsigned:%u;\n",
455 (unsigned int)offsetof(typeof(field), commit),
456 (unsigned int)sizeof(field.commit),
457 (unsigned int)is_signed_type(long));
458
459 ret = trace_seq_printf(s, "\tfield: int overwrite;\t"
460 "offset:%u;\tsize:%u;\tsigned:%u;\n",
461 (unsigned int)offsetof(typeof(field), commit),
462 1,
463 (unsigned int)is_signed_type(long));
464
465 ret = trace_seq_printf(s, "\tfield: char data;\t"
466 "offset:%u;\tsize:%u;\tsigned:%u;\n",
467 (unsigned int)offsetof(typeof(field), data),
468 (unsigned int)BUF_PAGE_SIZE,
469 (unsigned int)is_signed_type(char));
470
471 return ret;
472}
473
474/*
475 * head_page == tail_page && head == tail then buffer is empty.
476 */
477struct ring_buffer_per_cpu {
478 int cpu;
479 atomic_t record_disabled;
480 struct ring_buffer *buffer;
481 spinlock_t reader_lock; /* serialize readers */
482 arch_spinlock_t lock;
483 struct lock_class_key lock_key;
484 struct list_head *pages;
485 struct buffer_page *head_page; /* read from head */
486 struct buffer_page *tail_page; /* write to tail */
487 struct buffer_page *commit_page; /* committed pages */
488 struct buffer_page *reader_page;
489 unsigned long lost_events;
490 unsigned long last_overrun;
491 local_t commit_overrun;
492 local_t overrun;
493 local_t entries;
494 local_t committing;
495 local_t commits;
496 unsigned long read;
497 u64 write_stamp;
498 u64 read_stamp;
499};
500
501struct ring_buffer {
502 unsigned pages;
503 unsigned flags;
504 int cpus;
505 atomic_t record_disabled;
506 cpumask_var_t cpumask;
507
508 struct lock_class_key *reader_lock_key;
509
510 struct mutex mutex;
511
512 struct ring_buffer_per_cpu **buffers;
513
514#ifdef CONFIG_HOTPLUG_CPU
515 struct notifier_block cpu_notify;
516#endif
517 u64 (*clock)(void);
518};
519
520struct ring_buffer_iter {
521 struct ring_buffer_per_cpu *cpu_buffer;
522 unsigned long head;
523 struct buffer_page *head_page;
524 struct buffer_page *cache_reader_page;
525 unsigned long cache_read;
526 u64 read_stamp;
527};
528
529/* buffer may be either ring_buffer or ring_buffer_per_cpu */
530#define RB_WARN_ON(b, cond) \
531 ({ \
532 int _____ret = unlikely(cond); \
533 if (_____ret) { \
534 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
535 struct ring_buffer_per_cpu *__b = \
536 (void *)b; \
537 atomic_inc(&__b->buffer->record_disabled); \
538 } else \
539 atomic_inc(&b->record_disabled); \
540 WARN_ON(1); \
541 } \
542 _____ret; \
543 })
544
545/* Up this if you want to test the TIME_EXTENTS and normalization */
546#define DEBUG_SHIFT 0
547
548static inline u64 rb_time_stamp(struct ring_buffer *buffer)
549{
550 /* shift to debug/test normalization and TIME_EXTENTS */
551 return buffer->clock() << DEBUG_SHIFT;
552}
553
554u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
555{
556 u64 time;
557
558 preempt_disable_notrace();
559 time = rb_time_stamp(buffer);
560 preempt_enable_no_resched_notrace();
561
562 return time;
563}
564EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
565
566void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
567 int cpu, u64 *ts)
568{
569 /* Just stupid testing the normalize function and deltas */
570 *ts >>= DEBUG_SHIFT;
571}
572EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
573
574/*
575 * Making the ring buffer lockless makes things tricky.
576 * Although writes only happen on the CPU that they are on,
577 * and they only need to worry about interrupts. Reads can
578 * happen on any CPU.
579 *
580 * The reader page is always off the ring buffer, but when the
581 * reader finishes with a page, it needs to swap its page with
582 * a new one from the buffer. The reader needs to take from
583 * the head (writes go to the tail). But if a writer is in overwrite
584 * mode and wraps, it must push the head page forward.
585 *
586 * Here lies the problem.
587 *
588 * The reader must be careful to replace only the head page, and
589 * not another one. As described at the top of the file in the
590 * ASCII art, the reader sets its old page to point to the next
591 * page after head. It then sets the page after head to point to
592 * the old reader page. But if the writer moves the head page
593 * during this operation, the reader could end up with the tail.
594 *
595 * We use cmpxchg to help prevent this race. We also do something
596 * special with the page before head. We set the LSB to 1.
597 *
598 * When the writer must push the page forward, it will clear the
599 * bit that points to the head page, move the head, and then set
600 * the bit that points to the new head page.
601 *
602 * We also don't want an interrupt coming in and moving the head
603 * page on another writer. Thus we use the second LSB to catch
604 * that too. Thus:
605 *
606 * head->list->prev->next bit 1 bit 0
607 * ------- -------
608 * Normal page 0 0
609 * Points to head page 0 1
610 * New head page 1 0
611 *
612 * Note we can not trust the prev pointer of the head page, because:
613 *
614 * +----+ +-----+ +-----+
615 * | |------>| T |---X--->| N |
616 * | |<------| | | |
617 * +----+ +-----+ +-----+
618 * ^ ^ |
619 * | +-----+ | |
620 * +----------| R |----------+ |
621 * | |<-----------+
622 * +-----+
623 *
624 * Key: ---X--> HEAD flag set in pointer
625 * T Tail page
626 * R Reader page
627 * N Next page
628 *
629 * (see __rb_reserve_next() to see where this happens)
630 *
631 * What the above shows is that the reader just swapped out
632 * the reader page with a page in the buffer, but before it
633 * could make the new header point back to the new page added
634 * it was preempted by a writer. The writer moved forward onto
635 * the new page added by the reader and is about to move forward
636 * again.
637 *
638 * You can see, it is legitimate for the previous pointer of
639 * the head (or any page) not to point back to itself. But only
640 * temporarially.
641 */
642
643#define RB_PAGE_NORMAL 0UL
644#define RB_PAGE_HEAD 1UL
645#define RB_PAGE_UPDATE 2UL
646
647
648#define RB_FLAG_MASK 3UL
649
650/* PAGE_MOVED is not part of the mask */
651#define RB_PAGE_MOVED 4UL
652
653/*
654 * rb_list_head - remove any bit
655 */
656static struct list_head *rb_list_head(struct list_head *list)
657{
658 unsigned long val = (unsigned long)list;
659
660 return (struct list_head *)(val & ~RB_FLAG_MASK);
661}
662
663/*
664 * rb_is_head_page - test if the given page is the head page
665 *
666 * Because the reader may move the head_page pointer, we can
667 * not trust what the head page is (it may be pointing to
668 * the reader page). But if the next page is a header page,
669 * its flags will be non zero.
670 */
671static inline int
672rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
673 struct buffer_page *page, struct list_head *list)
674{
675 unsigned long val;
676
677 val = (unsigned long)list->next;
678
679 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
680 return RB_PAGE_MOVED;
681
682 return val & RB_FLAG_MASK;
683}
684
685/*
686 * rb_is_reader_page
687 *
688 * The unique thing about the reader page, is that, if the
689 * writer is ever on it, the previous pointer never points
690 * back to the reader page.
691 */
692static int rb_is_reader_page(struct buffer_page *page)
693{
694 struct list_head *list = page->list.prev;
695
696 return rb_list_head(list->next) != &page->list;
697}
698
699/*
700 * rb_set_list_to_head - set a list_head to be pointing to head.
701 */
702static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
703 struct list_head *list)
704{
705 unsigned long *ptr;
706
707 ptr = (unsigned long *)&list->next;
708 *ptr |= RB_PAGE_HEAD;
709 *ptr &= ~RB_PAGE_UPDATE;
710}
711
712/*
713 * rb_head_page_activate - sets up head page
714 */
715static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
716{
717 struct buffer_page *head;
718
719 head = cpu_buffer->head_page;
720 if (!head)
721 return;
722
723 /*
724 * Set the previous list pointer to have the HEAD flag.
725 */
726 rb_set_list_to_head(cpu_buffer, head->list.prev);
727}
728
729static void rb_list_head_clear(struct list_head *list)
730{
731 unsigned long *ptr = (unsigned long *)&list->next;
732
733 *ptr &= ~RB_FLAG_MASK;
734}
735
736/*
737 * rb_head_page_dactivate - clears head page ptr (for free list)
738 */
739static void
740rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
741{
742 struct list_head *hd;
743
744 /* Go through the whole list and clear any pointers found. */
745 rb_list_head_clear(cpu_buffer->pages);
746
747 list_for_each(hd, cpu_buffer->pages)
748 rb_list_head_clear(hd);
749}
750
751static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
752 struct buffer_page *head,
753 struct buffer_page *prev,
754 int old_flag, int new_flag)
755{
756 struct list_head *list;
757 unsigned long val = (unsigned long)&head->list;
758 unsigned long ret;
759
760 list = &prev->list;
761
762 val &= ~RB_FLAG_MASK;
763
764 ret = cmpxchg((unsigned long *)&list->next,
765 val | old_flag, val | new_flag);
766
767 /* check if the reader took the page */
768 if ((ret & ~RB_FLAG_MASK) != val)
769 return RB_PAGE_MOVED;
770
771 return ret & RB_FLAG_MASK;
772}
773
774static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
775 struct buffer_page *head,
776 struct buffer_page *prev,
777 int old_flag)
778{
779 return rb_head_page_set(cpu_buffer, head, prev,
780 old_flag, RB_PAGE_UPDATE);
781}
782
783static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
784 struct buffer_page *head,
785 struct buffer_page *prev,
786 int old_flag)
787{
788 return rb_head_page_set(cpu_buffer, head, prev,
789 old_flag, RB_PAGE_HEAD);
790}
791
792static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
793 struct buffer_page *head,
794 struct buffer_page *prev,
795 int old_flag)
796{
797 return rb_head_page_set(cpu_buffer, head, prev,
798 old_flag, RB_PAGE_NORMAL);
799}
800
801static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
802 struct buffer_page **bpage)
803{
804 struct list_head *p = rb_list_head((*bpage)->list.next);
805
806 *bpage = list_entry(p, struct buffer_page, list);
807}
808
809static struct buffer_page *
810rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
811{
812 struct buffer_page *head;
813 struct buffer_page *page;
814 struct list_head *list;
815 int i;
816
817 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
818 return NULL;
819
820 /* sanity check */
821 list = cpu_buffer->pages;
822 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
823 return NULL;
824
825 page = head = cpu_buffer->head_page;
826 /*
827 * It is possible that the writer moves the header behind
828 * where we started, and we miss in one loop.
829 * A second loop should grab the header, but we'll do
830 * three loops just because I'm paranoid.
831 */
832 for (i = 0; i < 3; i++) {
833 do {
834 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
835 cpu_buffer->head_page = page;
836 return page;
837 }
838 rb_inc_page(cpu_buffer, &page);
839 } while (page != head);
840 }
841
842 RB_WARN_ON(cpu_buffer, 1);
843
844 return NULL;
845}
846
847static int rb_head_page_replace(struct buffer_page *old,
848 struct buffer_page *new)
849{
850 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
851 unsigned long val;
852 unsigned long ret;
853
854 val = *ptr & ~RB_FLAG_MASK;
855 val |= RB_PAGE_HEAD;
856
857 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
858
859 return ret == val;
860}
861
862/*
863 * rb_tail_page_update - move the tail page forward
864 *
865 * Returns 1 if moved tail page, 0 if someone else did.
866 */
867static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
868 struct buffer_page *tail_page,
869 struct buffer_page *next_page)
870{
871 struct buffer_page *old_tail;
872 unsigned long old_entries;
873 unsigned long old_write;
874 int ret = 0;
875
876 /*
877 * The tail page now needs to be moved forward.
878 *
879 * We need to reset the tail page, but without messing
880 * with possible erasing of data brought in by interrupts
881 * that have moved the tail page and are currently on it.
882 *
883 * We add a counter to the write field to denote this.
884 */
885 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
886 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
887
888 /*
889 * Just make sure we have seen our old_write and synchronize
890 * with any interrupts that come in.
891 */
892 barrier();
893
894 /*
895 * If the tail page is still the same as what we think
896 * it is, then it is up to us to update the tail
897 * pointer.
898 */
899 if (tail_page == cpu_buffer->tail_page) {
900 /* Zero the write counter */
901 unsigned long val = old_write & ~RB_WRITE_MASK;
902 unsigned long eval = old_entries & ~RB_WRITE_MASK;
903
904 /*
905 * This will only succeed if an interrupt did
906 * not come in and change it. In which case, we
907 * do not want to modify it.
908 *
909 * We add (void) to let the compiler know that we do not care
910 * about the return value of these functions. We use the
911 * cmpxchg to only update if an interrupt did not already
912 * do it for us. If the cmpxchg fails, we don't care.
913 */
914 (void)local_cmpxchg(&next_page->write, old_write, val);
915 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
916
917 /*
918 * No need to worry about races with clearing out the commit.
919 * it only can increment when a commit takes place. But that
920 * only happens in the outer most nested commit.
921 */
922 local_set(&next_page->page->commit, 0);
923
924 old_tail = cmpxchg(&cpu_buffer->tail_page,
925 tail_page, next_page);
926
927 if (old_tail == tail_page)
928 ret = 1;
929 }
930
931 return ret;
932}
933
934static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
935 struct buffer_page *bpage)
936{
937 unsigned long val = (unsigned long)bpage;
938
939 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
940 return 1;
941
942 return 0;
943}
944
945/**
946 * rb_check_list - make sure a pointer to a list has the last bits zero
947 */
948static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
949 struct list_head *list)
950{
951 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
952 return 1;
953 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
954 return 1;
955 return 0;
956}
957
958/**
959 * check_pages - integrity check of buffer pages
960 * @cpu_buffer: CPU buffer with pages to test
961 *
962 * As a safety measure we check to make sure the data pages have not
963 * been corrupted.
964 */
965static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
966{
967 struct list_head *head = cpu_buffer->pages;
968 struct buffer_page *bpage, *tmp;
969
970 rb_head_page_deactivate(cpu_buffer);
971
972 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
973 return -1;
974 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
975 return -1;
976
977 if (rb_check_list(cpu_buffer, head))
978 return -1;
979
980 list_for_each_entry_safe(bpage, tmp, head, list) {
981 if (RB_WARN_ON(cpu_buffer,
982 bpage->list.next->prev != &bpage->list))
983 return -1;
984 if (RB_WARN_ON(cpu_buffer,
985 bpage->list.prev->next != &bpage->list))
986 return -1;
987 if (rb_check_list(cpu_buffer, &bpage->list))
988 return -1;
989 }
990
991 rb_head_page_activate(cpu_buffer);
992
993 return 0;
994}
995
996static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
997 unsigned nr_pages)
998{
999 struct buffer_page *bpage, *tmp;
1000 LIST_HEAD(pages);
1001 unsigned i;
1002
1003 WARN_ON(!nr_pages);
1004
1005 for (i = 0; i < nr_pages; i++) {
1006 struct page *page;
1007 /*
1008 * __GFP_NORETRY flag makes sure that the allocation fails
1009 * gracefully without invoking oom-killer and the system is
1010 * not destabilized.
1011 */
1012 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1013 GFP_KERNEL | __GFP_NORETRY,
1014 cpu_to_node(cpu_buffer->cpu));
1015 if (!bpage)
1016 goto free_pages;
1017
1018 rb_check_bpage(cpu_buffer, bpage);
1019
1020 list_add(&bpage->list, &pages);
1021
1022 page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu),
1023 GFP_KERNEL | __GFP_NORETRY, 0);
1024 if (!page)
1025 goto free_pages;
1026 bpage->page = page_address(page);
1027 rb_init_page(bpage->page);
1028 }
1029
1030 /*
1031 * The ring buffer page list is a circular list that does not
1032 * start and end with a list head. All page list items point to
1033 * other pages.
1034 */
1035 cpu_buffer->pages = pages.next;
1036 list_del(&pages);
1037
1038 rb_check_pages(cpu_buffer);
1039
1040 return 0;
1041
1042 free_pages:
1043 list_for_each_entry_safe(bpage, tmp, &pages, list) {
1044 list_del_init(&bpage->list);
1045 free_buffer_page(bpage);
1046 }
1047 return -ENOMEM;
1048}
1049
1050static struct ring_buffer_per_cpu *
1051rb_allocate_cpu_buffer(struct ring_buffer *buffer, int cpu)
1052{
1053 struct ring_buffer_per_cpu *cpu_buffer;
1054 struct buffer_page *bpage;
1055 struct page *page;
1056 int ret;
1057
1058 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1059 GFP_KERNEL, cpu_to_node(cpu));
1060 if (!cpu_buffer)
1061 return NULL;
1062
1063 cpu_buffer->cpu = cpu;
1064 cpu_buffer->buffer = buffer;
1065 spin_lock_init(&cpu_buffer->reader_lock);
1066 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1067 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1068
1069 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1070 GFP_KERNEL, cpu_to_node(cpu));
1071 if (!bpage)
1072 goto fail_free_buffer;
1073
1074 rb_check_bpage(cpu_buffer, bpage);
1075
1076 cpu_buffer->reader_page = bpage;
1077 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1078 if (!page)
1079 goto fail_free_reader;
1080 bpage->page = page_address(page);
1081 rb_init_page(bpage->page);
1082
1083 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1084
1085 ret = rb_allocate_pages(cpu_buffer, buffer->pages);
1086 if (ret < 0)
1087 goto fail_free_reader;
1088
1089 cpu_buffer->head_page
1090 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1091 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1092
1093 rb_head_page_activate(cpu_buffer);
1094
1095 return cpu_buffer;
1096
1097 fail_free_reader:
1098 free_buffer_page(cpu_buffer->reader_page);
1099
1100 fail_free_buffer:
1101 kfree(cpu_buffer);
1102 return NULL;
1103}
1104
1105static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1106{
1107 struct list_head *head = cpu_buffer->pages;
1108 struct buffer_page *bpage, *tmp;
1109
1110 free_buffer_page(cpu_buffer->reader_page);
1111
1112 rb_head_page_deactivate(cpu_buffer);
1113
1114 if (head) {
1115 list_for_each_entry_safe(bpage, tmp, head, list) {
1116 list_del_init(&bpage->list);
1117 free_buffer_page(bpage);
1118 }
1119 bpage = list_entry(head, struct buffer_page, list);
1120 free_buffer_page(bpage);
1121 }
1122
1123 kfree(cpu_buffer);
1124}
1125
1126#ifdef CONFIG_HOTPLUG_CPU
1127static int rb_cpu_notify(struct notifier_block *self,
1128 unsigned long action, void *hcpu);
1129#endif
1130
1131/**
1132 * ring_buffer_alloc - allocate a new ring_buffer
1133 * @size: the size in bytes per cpu that is needed.
1134 * @flags: attributes to set for the ring buffer.
1135 *
1136 * Currently the only flag that is available is the RB_FL_OVERWRITE
1137 * flag. This flag means that the buffer will overwrite old data
1138 * when the buffer wraps. If this flag is not set, the buffer will
1139 * drop data when the tail hits the head.
1140 */
1141struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1142 struct lock_class_key *key)
1143{
1144 struct ring_buffer *buffer;
1145 int bsize;
1146 int cpu;
1147
1148 /* keep it in its own cache line */
1149 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1150 GFP_KERNEL);
1151 if (!buffer)
1152 return NULL;
1153
1154 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1155 goto fail_free_buffer;
1156
1157 buffer->pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1158 buffer->flags = flags;
1159 buffer->clock = trace_clock_local;
1160 buffer->reader_lock_key = key;
1161
1162 /* need at least two pages */
1163 if (buffer->pages < 2)
1164 buffer->pages = 2;
1165
1166 /*
1167 * In case of non-hotplug cpu, if the ring-buffer is allocated
1168 * in early initcall, it will not be notified of secondary cpus.
1169 * In that off case, we need to allocate for all possible cpus.
1170 */
1171#ifdef CONFIG_HOTPLUG_CPU
1172 get_online_cpus();
1173 cpumask_copy(buffer->cpumask, cpu_online_mask);
1174#else
1175 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1176#endif
1177 buffer->cpus = nr_cpu_ids;
1178
1179 bsize = sizeof(void *) * nr_cpu_ids;
1180 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1181 GFP_KERNEL);
1182 if (!buffer->buffers)
1183 goto fail_free_cpumask;
1184
1185 for_each_buffer_cpu(buffer, cpu) {
1186 buffer->buffers[cpu] =
1187 rb_allocate_cpu_buffer(buffer, cpu);
1188 if (!buffer->buffers[cpu])
1189 goto fail_free_buffers;
1190 }
1191
1192#ifdef CONFIG_HOTPLUG_CPU
1193 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1194 buffer->cpu_notify.priority = 0;
1195 register_cpu_notifier(&buffer->cpu_notify);
1196#endif
1197
1198 put_online_cpus();
1199 mutex_init(&buffer->mutex);
1200
1201 return buffer;
1202
1203 fail_free_buffers:
1204 for_each_buffer_cpu(buffer, cpu) {
1205 if (buffer->buffers[cpu])
1206 rb_free_cpu_buffer(buffer->buffers[cpu]);
1207 }
1208 kfree(buffer->buffers);
1209
1210 fail_free_cpumask:
1211 free_cpumask_var(buffer->cpumask);
1212 put_online_cpus();
1213
1214 fail_free_buffer:
1215 kfree(buffer);
1216 return NULL;
1217}
1218EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1219
1220/**
1221 * ring_buffer_free - free a ring buffer.
1222 * @buffer: the buffer to free.
1223 */
1224void
1225ring_buffer_free(struct ring_buffer *buffer)
1226{
1227 int cpu;
1228
1229 get_online_cpus();
1230
1231#ifdef CONFIG_HOTPLUG_CPU
1232 unregister_cpu_notifier(&buffer->cpu_notify);
1233#endif
1234
1235 for_each_buffer_cpu(buffer, cpu)
1236 rb_free_cpu_buffer(buffer->buffers[cpu]);
1237
1238 put_online_cpus();
1239
1240 kfree(buffer->buffers);
1241 free_cpumask_var(buffer->cpumask);
1242
1243 kfree(buffer);
1244}
1245EXPORT_SYMBOL_GPL(ring_buffer_free);
1246
1247void ring_buffer_set_clock(struct ring_buffer *buffer,
1248 u64 (*clock)(void))
1249{
1250 buffer->clock = clock;
1251}
1252
1253static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1254
1255static void
1256rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned nr_pages)
1257{
1258 struct buffer_page *bpage;
1259 struct list_head *p;
1260 unsigned i;
1261
1262 spin_lock_irq(&cpu_buffer->reader_lock);
1263 rb_head_page_deactivate(cpu_buffer);
1264
1265 for (i = 0; i < nr_pages; i++) {
1266 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1267 goto out;
1268 p = cpu_buffer->pages->next;
1269 bpage = list_entry(p, struct buffer_page, list);
1270 list_del_init(&bpage->list);
1271 free_buffer_page(bpage);
1272 }
1273 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1274 goto out;
1275
1276 rb_reset_cpu(cpu_buffer);
1277 rb_check_pages(cpu_buffer);
1278
1279out:
1280 spin_unlock_irq(&cpu_buffer->reader_lock);
1281}
1282
1283static void
1284rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer,
1285 struct list_head *pages, unsigned nr_pages)
1286{
1287 struct buffer_page *bpage;
1288 struct list_head *p;
1289 unsigned i;
1290
1291 spin_lock_irq(&cpu_buffer->reader_lock);
1292 rb_head_page_deactivate(cpu_buffer);
1293
1294 for (i = 0; i < nr_pages; i++) {
1295 if (RB_WARN_ON(cpu_buffer, list_empty(pages)))
1296 goto out;
1297 p = pages->next;
1298 bpage = list_entry(p, struct buffer_page, list);
1299 list_del_init(&bpage->list);
1300 list_add_tail(&bpage->list, cpu_buffer->pages);
1301 }
1302 rb_reset_cpu(cpu_buffer);
1303 rb_check_pages(cpu_buffer);
1304
1305out:
1306 spin_unlock_irq(&cpu_buffer->reader_lock);
1307}
1308
1309/**
1310 * ring_buffer_resize - resize the ring buffer
1311 * @buffer: the buffer to resize.
1312 * @size: the new size.
1313 *
1314 * Minimum size is 2 * BUF_PAGE_SIZE.
1315 *
1316 * Returns -1 on failure.
1317 */
1318int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size)
1319{
1320 struct ring_buffer_per_cpu *cpu_buffer;
1321 unsigned nr_pages, rm_pages, new_pages;
1322 struct buffer_page *bpage, *tmp;
1323 unsigned long buffer_size;
1324 LIST_HEAD(pages);
1325 int i, cpu;
1326
1327 /*
1328 * Always succeed at resizing a non-existent buffer:
1329 */
1330 if (!buffer)
1331 return size;
1332
1333 size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1334 size *= BUF_PAGE_SIZE;
1335 buffer_size = buffer->pages * BUF_PAGE_SIZE;
1336
1337 /* we need a minimum of two pages */
1338 if (size < BUF_PAGE_SIZE * 2)
1339 size = BUF_PAGE_SIZE * 2;
1340
1341 if (size == buffer_size)
1342 return size;
1343
1344 atomic_inc(&buffer->record_disabled);
1345
1346 /* Make sure all writers are done with this buffer. */
1347 synchronize_sched();
1348
1349 mutex_lock(&buffer->mutex);
1350 get_online_cpus();
1351
1352 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1353
1354 if (size < buffer_size) {
1355
1356 /* easy case, just free pages */
1357 if (RB_WARN_ON(buffer, nr_pages >= buffer->pages))
1358 goto out_fail;
1359
1360 rm_pages = buffer->pages - nr_pages;
1361
1362 for_each_buffer_cpu(buffer, cpu) {
1363 cpu_buffer = buffer->buffers[cpu];
1364 rb_remove_pages(cpu_buffer, rm_pages);
1365 }
1366 goto out;
1367 }
1368
1369 /*
1370 * This is a bit more difficult. We only want to add pages
1371 * when we can allocate enough for all CPUs. We do this
1372 * by allocating all the pages and storing them on a local
1373 * link list. If we succeed in our allocation, then we
1374 * add these pages to the cpu_buffers. Otherwise we just free
1375 * them all and return -ENOMEM;
1376 */
1377 if (RB_WARN_ON(buffer, nr_pages <= buffer->pages))
1378 goto out_fail;
1379
1380 new_pages = nr_pages - buffer->pages;
1381
1382 for_each_buffer_cpu(buffer, cpu) {
1383 for (i = 0; i < new_pages; i++) {
1384 struct page *page;
1385 /*
1386 * __GFP_NORETRY flag makes sure that the allocation
1387 * fails gracefully without invoking oom-killer and
1388 * the system is not destabilized.
1389 */
1390 bpage = kzalloc_node(ALIGN(sizeof(*bpage),
1391 cache_line_size()),
1392 GFP_KERNEL | __GFP_NORETRY,
1393 cpu_to_node(cpu));
1394 if (!bpage)
1395 goto free_pages;
1396 list_add(&bpage->list, &pages);
1397 page = alloc_pages_node(cpu_to_node(cpu),
1398 GFP_KERNEL | __GFP_NORETRY, 0);
1399 if (!page)
1400 goto free_pages;
1401 bpage->page = page_address(page);
1402 rb_init_page(bpage->page);
1403 }
1404 }
1405
1406 for_each_buffer_cpu(buffer, cpu) {
1407 cpu_buffer = buffer->buffers[cpu];
1408 rb_insert_pages(cpu_buffer, &pages, new_pages);
1409 }
1410
1411 if (RB_WARN_ON(buffer, !list_empty(&pages)))
1412 goto out_fail;
1413
1414 out:
1415 buffer->pages = nr_pages;
1416 put_online_cpus();
1417 mutex_unlock(&buffer->mutex);
1418
1419 atomic_dec(&buffer->record_disabled);
1420
1421 return size;
1422
1423 free_pages:
1424 list_for_each_entry_safe(bpage, tmp, &pages, list) {
1425 list_del_init(&bpage->list);
1426 free_buffer_page(bpage);
1427 }
1428 put_online_cpus();
1429 mutex_unlock(&buffer->mutex);
1430 atomic_dec(&buffer->record_disabled);
1431 return -ENOMEM;
1432
1433 /*
1434 * Something went totally wrong, and we are too paranoid
1435 * to even clean up the mess.
1436 */
1437 out_fail:
1438 put_online_cpus();
1439 mutex_unlock(&buffer->mutex);
1440 atomic_dec(&buffer->record_disabled);
1441 return -1;
1442}
1443EXPORT_SYMBOL_GPL(ring_buffer_resize);
1444
1445void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1446{
1447 mutex_lock(&buffer->mutex);
1448 if (val)
1449 buffer->flags |= RB_FL_OVERWRITE;
1450 else
1451 buffer->flags &= ~RB_FL_OVERWRITE;
1452 mutex_unlock(&buffer->mutex);
1453}
1454EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1455
1456static inline void *
1457__rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1458{
1459 return bpage->data + index;
1460}
1461
1462static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1463{
1464 return bpage->page->data + index;
1465}
1466
1467static inline struct ring_buffer_event *
1468rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1469{
1470 return __rb_page_index(cpu_buffer->reader_page,
1471 cpu_buffer->reader_page->read);
1472}
1473
1474static inline struct ring_buffer_event *
1475rb_iter_head_event(struct ring_buffer_iter *iter)
1476{
1477 return __rb_page_index(iter->head_page, iter->head);
1478}
1479
1480static inline unsigned long rb_page_write(struct buffer_page *bpage)
1481{
1482 return local_read(&bpage->write) & RB_WRITE_MASK;
1483}
1484
1485static inline unsigned rb_page_commit(struct buffer_page *bpage)
1486{
1487 return local_read(&bpage->page->commit);
1488}
1489
1490static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1491{
1492 return local_read(&bpage->entries) & RB_WRITE_MASK;
1493}
1494
1495/* Size is determined by what has been committed */
1496static inline unsigned rb_page_size(struct buffer_page *bpage)
1497{
1498 return rb_page_commit(bpage);
1499}
1500
1501static inline unsigned
1502rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1503{
1504 return rb_page_commit(cpu_buffer->commit_page);
1505}
1506
1507static inline unsigned
1508rb_event_index(struct ring_buffer_event *event)
1509{
1510 unsigned long addr = (unsigned long)event;
1511
1512 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1513}
1514
1515static inline int
1516rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1517 struct ring_buffer_event *event)
1518{
1519 unsigned long addr = (unsigned long)event;
1520 unsigned long index;
1521
1522 index = rb_event_index(event);
1523 addr &= PAGE_MASK;
1524
1525 return cpu_buffer->commit_page->page == (void *)addr &&
1526 rb_commit_index(cpu_buffer) == index;
1527}
1528
1529static void
1530rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1531{
1532 unsigned long max_count;
1533
1534 /*
1535 * We only race with interrupts and NMIs on this CPU.
1536 * If we own the commit event, then we can commit
1537 * all others that interrupted us, since the interruptions
1538 * are in stack format (they finish before they come
1539 * back to us). This allows us to do a simple loop to
1540 * assign the commit to the tail.
1541 */
1542 again:
1543 max_count = cpu_buffer->buffer->pages * 100;
1544
1545 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1546 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1547 return;
1548 if (RB_WARN_ON(cpu_buffer,
1549 rb_is_reader_page(cpu_buffer->tail_page)))
1550 return;
1551 local_set(&cpu_buffer->commit_page->page->commit,
1552 rb_page_write(cpu_buffer->commit_page));
1553 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1554 cpu_buffer->write_stamp =
1555 cpu_buffer->commit_page->page->time_stamp;
1556 /* add barrier to keep gcc from optimizing too much */
1557 barrier();
1558 }
1559 while (rb_commit_index(cpu_buffer) !=
1560 rb_page_write(cpu_buffer->commit_page)) {
1561
1562 local_set(&cpu_buffer->commit_page->page->commit,
1563 rb_page_write(cpu_buffer->commit_page));
1564 RB_WARN_ON(cpu_buffer,
1565 local_read(&cpu_buffer->commit_page->page->commit) &
1566 ~RB_WRITE_MASK);
1567 barrier();
1568 }
1569
1570 /* again, keep gcc from optimizing */
1571 barrier();
1572
1573 /*
1574 * If an interrupt came in just after the first while loop
1575 * and pushed the tail page forward, we will be left with
1576 * a dangling commit that will never go forward.
1577 */
1578 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1579 goto again;
1580}
1581
1582static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1583{
1584 cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1585 cpu_buffer->reader_page->read = 0;
1586}
1587
1588static void rb_inc_iter(struct ring_buffer_iter *iter)
1589{
1590 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1591
1592 /*
1593 * The iterator could be on the reader page (it starts there).
1594 * But the head could have moved, since the reader was
1595 * found. Check for this case and assign the iterator
1596 * to the head page instead of next.
1597 */
1598 if (iter->head_page == cpu_buffer->reader_page)
1599 iter->head_page = rb_set_head_page(cpu_buffer);
1600 else
1601 rb_inc_page(cpu_buffer, &iter->head_page);
1602
1603 iter->read_stamp = iter->head_page->page->time_stamp;
1604 iter->head = 0;
1605}
1606
1607/* Slow path, do not inline */
1608static noinline struct ring_buffer_event *
1609rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
1610{
1611 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
1612
1613 /* Not the first event on the page? */
1614 if (rb_event_index(event)) {
1615 event->time_delta = delta & TS_MASK;
1616 event->array[0] = delta >> TS_SHIFT;
1617 } else {
1618 /* nope, just zero it */
1619 event->time_delta = 0;
1620 event->array[0] = 0;
1621 }
1622
1623 return skip_time_extend(event);
1624}
1625
1626/**
1627 * ring_buffer_update_event - update event type and data
1628 * @event: the even to update
1629 * @type: the type of event
1630 * @length: the size of the event field in the ring buffer
1631 *
1632 * Update the type and data fields of the event. The length
1633 * is the actual size that is written to the ring buffer,
1634 * and with this, we can determine what to place into the
1635 * data field.
1636 */
1637static void
1638rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
1639 struct ring_buffer_event *event, unsigned length,
1640 int add_timestamp, u64 delta)
1641{
1642 /* Only a commit updates the timestamp */
1643 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
1644 delta = 0;
1645
1646 /*
1647 * If we need to add a timestamp, then we
1648 * add it to the start of the resevered space.
1649 */
1650 if (unlikely(add_timestamp)) {
1651 event = rb_add_time_stamp(event, delta);
1652 length -= RB_LEN_TIME_EXTEND;
1653 delta = 0;
1654 }
1655
1656 event->time_delta = delta;
1657 length -= RB_EVNT_HDR_SIZE;
1658 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
1659 event->type_len = 0;
1660 event->array[0] = length;
1661 } else
1662 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
1663}
1664
1665/*
1666 * rb_handle_head_page - writer hit the head page
1667 *
1668 * Returns: +1 to retry page
1669 * 0 to continue
1670 * -1 on error
1671 */
1672static int
1673rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1674 struct buffer_page *tail_page,
1675 struct buffer_page *next_page)
1676{
1677 struct buffer_page *new_head;
1678 int entries;
1679 int type;
1680 int ret;
1681
1682 entries = rb_page_entries(next_page);
1683
1684 /*
1685 * The hard part is here. We need to move the head
1686 * forward, and protect against both readers on
1687 * other CPUs and writers coming in via interrupts.
1688 */
1689 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1690 RB_PAGE_HEAD);
1691
1692 /*
1693 * type can be one of four:
1694 * NORMAL - an interrupt already moved it for us
1695 * HEAD - we are the first to get here.
1696 * UPDATE - we are the interrupt interrupting
1697 * a current move.
1698 * MOVED - a reader on another CPU moved the next
1699 * pointer to its reader page. Give up
1700 * and try again.
1701 */
1702
1703 switch (type) {
1704 case RB_PAGE_HEAD:
1705 /*
1706 * We changed the head to UPDATE, thus
1707 * it is our responsibility to update
1708 * the counters.
1709 */
1710 local_add(entries, &cpu_buffer->overrun);
1711
1712 /*
1713 * The entries will be zeroed out when we move the
1714 * tail page.
1715 */
1716
1717 /* still more to do */
1718 break;
1719
1720 case RB_PAGE_UPDATE:
1721 /*
1722 * This is an interrupt that interrupt the
1723 * previous update. Still more to do.
1724 */
1725 break;
1726 case RB_PAGE_NORMAL:
1727 /*
1728 * An interrupt came in before the update
1729 * and processed this for us.
1730 * Nothing left to do.
1731 */
1732 return 1;
1733 case RB_PAGE_MOVED:
1734 /*
1735 * The reader is on another CPU and just did
1736 * a swap with our next_page.
1737 * Try again.
1738 */
1739 return 1;
1740 default:
1741 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1742 return -1;
1743 }
1744
1745 /*
1746 * Now that we are here, the old head pointer is
1747 * set to UPDATE. This will keep the reader from
1748 * swapping the head page with the reader page.
1749 * The reader (on another CPU) will spin till
1750 * we are finished.
1751 *
1752 * We just need to protect against interrupts
1753 * doing the job. We will set the next pointer
1754 * to HEAD. After that, we set the old pointer
1755 * to NORMAL, but only if it was HEAD before.
1756 * otherwise we are an interrupt, and only
1757 * want the outer most commit to reset it.
1758 */
1759 new_head = next_page;
1760 rb_inc_page(cpu_buffer, &new_head);
1761
1762 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1763 RB_PAGE_NORMAL);
1764
1765 /*
1766 * Valid returns are:
1767 * HEAD - an interrupt came in and already set it.
1768 * NORMAL - One of two things:
1769 * 1) We really set it.
1770 * 2) A bunch of interrupts came in and moved
1771 * the page forward again.
1772 */
1773 switch (ret) {
1774 case RB_PAGE_HEAD:
1775 case RB_PAGE_NORMAL:
1776 /* OK */
1777 break;
1778 default:
1779 RB_WARN_ON(cpu_buffer, 1);
1780 return -1;
1781 }
1782
1783 /*
1784 * It is possible that an interrupt came in,
1785 * set the head up, then more interrupts came in
1786 * and moved it again. When we get back here,
1787 * the page would have been set to NORMAL but we
1788 * just set it back to HEAD.
1789 *
1790 * How do you detect this? Well, if that happened
1791 * the tail page would have moved.
1792 */
1793 if (ret == RB_PAGE_NORMAL) {
1794 /*
1795 * If the tail had moved passed next, then we need
1796 * to reset the pointer.
1797 */
1798 if (cpu_buffer->tail_page != tail_page &&
1799 cpu_buffer->tail_page != next_page)
1800 rb_head_page_set_normal(cpu_buffer, new_head,
1801 next_page,
1802 RB_PAGE_HEAD);
1803 }
1804
1805 /*
1806 * If this was the outer most commit (the one that
1807 * changed the original pointer from HEAD to UPDATE),
1808 * then it is up to us to reset it to NORMAL.
1809 */
1810 if (type == RB_PAGE_HEAD) {
1811 ret = rb_head_page_set_normal(cpu_buffer, next_page,
1812 tail_page,
1813 RB_PAGE_UPDATE);
1814 if (RB_WARN_ON(cpu_buffer,
1815 ret != RB_PAGE_UPDATE))
1816 return -1;
1817 }
1818
1819 return 0;
1820}
1821
1822static unsigned rb_calculate_event_length(unsigned length)
1823{
1824 struct ring_buffer_event event; /* Used only for sizeof array */
1825
1826 /* zero length can cause confusions */
1827 if (!length)
1828 length = 1;
1829
1830 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
1831 length += sizeof(event.array[0]);
1832
1833 length += RB_EVNT_HDR_SIZE;
1834 length = ALIGN(length, RB_ARCH_ALIGNMENT);
1835
1836 return length;
1837}
1838
1839static inline void
1840rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
1841 struct buffer_page *tail_page,
1842 unsigned long tail, unsigned long length)
1843{
1844 struct ring_buffer_event *event;
1845
1846 /*
1847 * Only the event that crossed the page boundary
1848 * must fill the old tail_page with padding.
1849 */
1850 if (tail >= BUF_PAGE_SIZE) {
1851 /*
1852 * If the page was filled, then we still need
1853 * to update the real_end. Reset it to zero
1854 * and the reader will ignore it.
1855 */
1856 if (tail == BUF_PAGE_SIZE)
1857 tail_page->real_end = 0;
1858
1859 local_sub(length, &tail_page->write);
1860 return;
1861 }
1862
1863 event = __rb_page_index(tail_page, tail);
1864 kmemcheck_annotate_bitfield(event, bitfield);
1865
1866 /*
1867 * Save the original length to the meta data.
1868 * This will be used by the reader to add lost event
1869 * counter.
1870 */
1871 tail_page->real_end = tail;
1872
1873 /*
1874 * If this event is bigger than the minimum size, then
1875 * we need to be careful that we don't subtract the
1876 * write counter enough to allow another writer to slip
1877 * in on this page.
1878 * We put in a discarded commit instead, to make sure
1879 * that this space is not used again.
1880 *
1881 * If we are less than the minimum size, we don't need to
1882 * worry about it.
1883 */
1884 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
1885 /* No room for any events */
1886
1887 /* Mark the rest of the page with padding */
1888 rb_event_set_padding(event);
1889
1890 /* Set the write back to the previous setting */
1891 local_sub(length, &tail_page->write);
1892 return;
1893 }
1894
1895 /* Put in a discarded event */
1896 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
1897 event->type_len = RINGBUF_TYPE_PADDING;
1898 /* time delta must be non zero */
1899 event->time_delta = 1;
1900
1901 /* Set write to end of buffer */
1902 length = (tail + length) - BUF_PAGE_SIZE;
1903 local_sub(length, &tail_page->write);
1904}
1905
1906/*
1907 * This is the slow path, force gcc not to inline it.
1908 */
1909static noinline struct ring_buffer_event *
1910rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
1911 unsigned long length, unsigned long tail,
1912 struct buffer_page *tail_page, u64 ts)
1913{
1914 struct buffer_page *commit_page = cpu_buffer->commit_page;
1915 struct ring_buffer *buffer = cpu_buffer->buffer;
1916 struct buffer_page *next_page;
1917 int ret;
1918
1919 next_page = tail_page;
1920
1921 rb_inc_page(cpu_buffer, &next_page);
1922
1923 /*
1924 * If for some reason, we had an interrupt storm that made
1925 * it all the way around the buffer, bail, and warn
1926 * about it.
1927 */
1928 if (unlikely(next_page == commit_page)) {
1929 local_inc(&cpu_buffer->commit_overrun);
1930 goto out_reset;
1931 }
1932
1933 /*
1934 * This is where the fun begins!
1935 *
1936 * We are fighting against races between a reader that
1937 * could be on another CPU trying to swap its reader
1938 * page with the buffer head.
1939 *
1940 * We are also fighting against interrupts coming in and
1941 * moving the head or tail on us as well.
1942 *
1943 * If the next page is the head page then we have filled
1944 * the buffer, unless the commit page is still on the
1945 * reader page.
1946 */
1947 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
1948
1949 /*
1950 * If the commit is not on the reader page, then
1951 * move the header page.
1952 */
1953 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
1954 /*
1955 * If we are not in overwrite mode,
1956 * this is easy, just stop here.
1957 */
1958 if (!(buffer->flags & RB_FL_OVERWRITE))
1959 goto out_reset;
1960
1961 ret = rb_handle_head_page(cpu_buffer,
1962 tail_page,
1963 next_page);
1964 if (ret < 0)
1965 goto out_reset;
1966 if (ret)
1967 goto out_again;
1968 } else {
1969 /*
1970 * We need to be careful here too. The
1971 * commit page could still be on the reader
1972 * page. We could have a small buffer, and
1973 * have filled up the buffer with events
1974 * from interrupts and such, and wrapped.
1975 *
1976 * Note, if the tail page is also the on the
1977 * reader_page, we let it move out.
1978 */
1979 if (unlikely((cpu_buffer->commit_page !=
1980 cpu_buffer->tail_page) &&
1981 (cpu_buffer->commit_page ==
1982 cpu_buffer->reader_page))) {
1983 local_inc(&cpu_buffer->commit_overrun);
1984 goto out_reset;
1985 }
1986 }
1987 }
1988
1989 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
1990 if (ret) {
1991 /*
1992 * Nested commits always have zero deltas, so
1993 * just reread the time stamp
1994 */
1995 ts = rb_time_stamp(buffer);
1996 next_page->page->time_stamp = ts;
1997 }
1998
1999 out_again:
2000
2001 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2002
2003 /* fail and let the caller try again */
2004 return ERR_PTR(-EAGAIN);
2005
2006 out_reset:
2007 /* reset write */
2008 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2009
2010 return NULL;
2011}
2012
2013static struct ring_buffer_event *
2014__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2015 unsigned long length, u64 ts,
2016 u64 delta, int add_timestamp)
2017{
2018 struct buffer_page *tail_page;
2019 struct ring_buffer_event *event;
2020 unsigned long tail, write;
2021
2022 /*
2023 * If the time delta since the last event is too big to
2024 * hold in the time field of the event, then we append a
2025 * TIME EXTEND event ahead of the data event.
2026 */
2027 if (unlikely(add_timestamp))
2028 length += RB_LEN_TIME_EXTEND;
2029
2030 tail_page = cpu_buffer->tail_page;
2031 write = local_add_return(length, &tail_page->write);
2032
2033 /* set write to only the index of the write */
2034 write &= RB_WRITE_MASK;
2035 tail = write - length;
2036
2037 /* See if we shot pass the end of this buffer page */
2038 if (unlikely(write > BUF_PAGE_SIZE))
2039 return rb_move_tail(cpu_buffer, length, tail,
2040 tail_page, ts);
2041
2042 /* We reserved something on the buffer */
2043
2044 event = __rb_page_index(tail_page, tail);
2045 kmemcheck_annotate_bitfield(event, bitfield);
2046 rb_update_event(cpu_buffer, event, length, add_timestamp, delta);
2047
2048 local_inc(&tail_page->entries);
2049
2050 /*
2051 * If this is the first commit on the page, then update
2052 * its timestamp.
2053 */
2054 if (!tail)
2055 tail_page->page->time_stamp = ts;
2056
2057 return event;
2058}
2059
2060static inline int
2061rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2062 struct ring_buffer_event *event)
2063{
2064 unsigned long new_index, old_index;
2065 struct buffer_page *bpage;
2066 unsigned long index;
2067 unsigned long addr;
2068
2069 new_index = rb_event_index(event);
2070 old_index = new_index + rb_event_ts_length(event);
2071 addr = (unsigned long)event;
2072 addr &= PAGE_MASK;
2073
2074 bpage = cpu_buffer->tail_page;
2075
2076 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2077 unsigned long write_mask =
2078 local_read(&bpage->write) & ~RB_WRITE_MASK;
2079 /*
2080 * This is on the tail page. It is possible that
2081 * a write could come in and move the tail page
2082 * and write to the next page. That is fine
2083 * because we just shorten what is on this page.
2084 */
2085 old_index += write_mask;
2086 new_index += write_mask;
2087 index = local_cmpxchg(&bpage->write, old_index, new_index);
2088 if (index == old_index)
2089 return 1;
2090 }
2091
2092 /* could not discard */
2093 return 0;
2094}
2095
2096static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2097{
2098 local_inc(&cpu_buffer->committing);
2099 local_inc(&cpu_buffer->commits);
2100}
2101
2102static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2103{
2104 unsigned long commits;
2105
2106 if (RB_WARN_ON(cpu_buffer,
2107 !local_read(&cpu_buffer->committing)))
2108 return;
2109
2110 again:
2111 commits = local_read(&cpu_buffer->commits);
2112 /* synchronize with interrupts */
2113 barrier();
2114 if (local_read(&cpu_buffer->committing) == 1)
2115 rb_set_commit_to_write(cpu_buffer);
2116
2117 local_dec(&cpu_buffer->committing);
2118
2119 /* synchronize with interrupts */
2120 barrier();
2121
2122 /*
2123 * Need to account for interrupts coming in between the
2124 * updating of the commit page and the clearing of the
2125 * committing counter.
2126 */
2127 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2128 !local_read(&cpu_buffer->committing)) {
2129 local_inc(&cpu_buffer->committing);
2130 goto again;
2131 }
2132}
2133
2134static struct ring_buffer_event *
2135rb_reserve_next_event(struct ring_buffer *buffer,
2136 struct ring_buffer_per_cpu *cpu_buffer,
2137 unsigned long length)
2138{
2139 struct ring_buffer_event *event;
2140 u64 ts, delta;
2141 int nr_loops = 0;
2142 int add_timestamp;
2143 u64 diff;
2144
2145 rb_start_commit(cpu_buffer);
2146
2147#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2148 /*
2149 * Due to the ability to swap a cpu buffer from a buffer
2150 * it is possible it was swapped before we committed.
2151 * (committing stops a swap). We check for it here and
2152 * if it happened, we have to fail the write.
2153 */
2154 barrier();
2155 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2156 local_dec(&cpu_buffer->committing);
2157 local_dec(&cpu_buffer->commits);
2158 return NULL;
2159 }
2160#endif
2161
2162 length = rb_calculate_event_length(length);
2163 again:
2164 add_timestamp = 0;
2165 delta = 0;
2166
2167 /*
2168 * We allow for interrupts to reenter here and do a trace.
2169 * If one does, it will cause this original code to loop
2170 * back here. Even with heavy interrupts happening, this
2171 * should only happen a few times in a row. If this happens
2172 * 1000 times in a row, there must be either an interrupt
2173 * storm or we have something buggy.
2174 * Bail!
2175 */
2176 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2177 goto out_fail;
2178
2179 ts = rb_time_stamp(cpu_buffer->buffer);
2180 diff = ts - cpu_buffer->write_stamp;
2181
2182 /* make sure this diff is calculated here */
2183 barrier();
2184
2185 /* Did the write stamp get updated already? */
2186 if (likely(ts >= cpu_buffer->write_stamp)) {
2187 delta = diff;
2188 if (unlikely(test_time_stamp(delta))) {
2189 int local_clock_stable = 1;
2190#ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2191 local_clock_stable = sched_clock_stable;
2192#endif
2193 WARN_ONCE(delta > (1ULL << 59),
2194 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2195 (unsigned long long)delta,
2196 (unsigned long long)ts,
2197 (unsigned long long)cpu_buffer->write_stamp,
2198 local_clock_stable ? "" :
2199 "If you just came from a suspend/resume,\n"
2200 "please switch to the trace global clock:\n"
2201 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2202 add_timestamp = 1;
2203 }
2204 }
2205
2206 event = __rb_reserve_next(cpu_buffer, length, ts,
2207 delta, add_timestamp);
2208 if (unlikely(PTR_ERR(event) == -EAGAIN))
2209 goto again;
2210
2211 if (!event)
2212 goto out_fail;
2213
2214 return event;
2215
2216 out_fail:
2217 rb_end_commit(cpu_buffer);
2218 return NULL;
2219}
2220
2221#ifdef CONFIG_TRACING
2222
2223#define TRACE_RECURSIVE_DEPTH 16
2224
2225/* Keep this code out of the fast path cache */
2226static noinline void trace_recursive_fail(void)
2227{
2228 /* Disable all tracing before we do anything else */
2229 tracing_off_permanent();
2230
2231 printk_once(KERN_WARNING "Tracing recursion: depth[%ld]:"
2232 "HC[%lu]:SC[%lu]:NMI[%lu]\n",
2233 trace_recursion_buffer(),
2234 hardirq_count() >> HARDIRQ_SHIFT,
2235 softirq_count() >> SOFTIRQ_SHIFT,
2236 in_nmi());
2237
2238 WARN_ON_ONCE(1);
2239}
2240
2241static inline int trace_recursive_lock(void)
2242{
2243 trace_recursion_inc();
2244
2245 if (likely(trace_recursion_buffer() < TRACE_RECURSIVE_DEPTH))
2246 return 0;
2247
2248 trace_recursive_fail();
2249
2250 return -1;
2251}
2252
2253static inline void trace_recursive_unlock(void)
2254{
2255 WARN_ON_ONCE(!trace_recursion_buffer());
2256
2257 trace_recursion_dec();
2258}
2259
2260#else
2261
2262#define trace_recursive_lock() (0)
2263#define trace_recursive_unlock() do { } while (0)
2264
2265#endif
2266
2267/**
2268 * ring_buffer_lock_reserve - reserve a part of the buffer
2269 * @buffer: the ring buffer to reserve from
2270 * @length: the length of the data to reserve (excluding event header)
2271 *
2272 * Returns a reseverd event on the ring buffer to copy directly to.
2273 * The user of this interface will need to get the body to write into
2274 * and can use the ring_buffer_event_data() interface.
2275 *
2276 * The length is the length of the data needed, not the event length
2277 * which also includes the event header.
2278 *
2279 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2280 * If NULL is returned, then nothing has been allocated or locked.
2281 */
2282struct ring_buffer_event *
2283ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2284{
2285 struct ring_buffer_per_cpu *cpu_buffer;
2286 struct ring_buffer_event *event;
2287 int cpu;
2288
2289 if (ring_buffer_flags != RB_BUFFERS_ON)
2290 return NULL;
2291
2292 /* If we are tracing schedule, we don't want to recurse */
2293 preempt_disable_notrace();
2294
2295 if (atomic_read(&buffer->record_disabled))
2296 goto out_nocheck;
2297
2298 if (trace_recursive_lock())
2299 goto out_nocheck;
2300
2301 cpu = raw_smp_processor_id();
2302
2303 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2304 goto out;
2305
2306 cpu_buffer = buffer->buffers[cpu];
2307
2308 if (atomic_read(&cpu_buffer->record_disabled))
2309 goto out;
2310
2311 if (length > BUF_MAX_DATA_SIZE)
2312 goto out;
2313
2314 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2315 if (!event)
2316 goto out;
2317
2318 return event;
2319
2320 out:
2321 trace_recursive_unlock();
2322
2323 out_nocheck:
2324 preempt_enable_notrace();
2325 return NULL;
2326}
2327EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2328
2329static void
2330rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2331 struct ring_buffer_event *event)
2332{
2333 u64 delta;
2334
2335 /*
2336 * The event first in the commit queue updates the
2337 * time stamp.
2338 */
2339 if (rb_event_is_commit(cpu_buffer, event)) {
2340 /*
2341 * A commit event that is first on a page
2342 * updates the write timestamp with the page stamp
2343 */
2344 if (!rb_event_index(event))
2345 cpu_buffer->write_stamp =
2346 cpu_buffer->commit_page->page->time_stamp;
2347 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2348 delta = event->array[0];
2349 delta <<= TS_SHIFT;
2350 delta += event->time_delta;
2351 cpu_buffer->write_stamp += delta;
2352 } else
2353 cpu_buffer->write_stamp += event->time_delta;
2354 }
2355}
2356
2357static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2358 struct ring_buffer_event *event)
2359{
2360 local_inc(&cpu_buffer->entries);
2361 rb_update_write_stamp(cpu_buffer, event);
2362 rb_end_commit(cpu_buffer);
2363}
2364
2365/**
2366 * ring_buffer_unlock_commit - commit a reserved
2367 * @buffer: The buffer to commit to
2368 * @event: The event pointer to commit.
2369 *
2370 * This commits the data to the ring buffer, and releases any locks held.
2371 *
2372 * Must be paired with ring_buffer_lock_reserve.
2373 */
2374int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2375 struct ring_buffer_event *event)
2376{
2377 struct ring_buffer_per_cpu *cpu_buffer;
2378 int cpu = raw_smp_processor_id();
2379
2380 cpu_buffer = buffer->buffers[cpu];
2381
2382 rb_commit(cpu_buffer, event);
2383
2384 trace_recursive_unlock();
2385
2386 preempt_enable_notrace();
2387
2388 return 0;
2389}
2390EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2391
2392static inline void rb_event_discard(struct ring_buffer_event *event)
2393{
2394 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2395 event = skip_time_extend(event);
2396
2397 /* array[0] holds the actual length for the discarded event */
2398 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2399 event->type_len = RINGBUF_TYPE_PADDING;
2400 /* time delta must be non zero */
2401 if (!event->time_delta)
2402 event->time_delta = 1;
2403}
2404
2405/*
2406 * Decrement the entries to the page that an event is on.
2407 * The event does not even need to exist, only the pointer
2408 * to the page it is on. This may only be called before the commit
2409 * takes place.
2410 */
2411static inline void
2412rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2413 struct ring_buffer_event *event)
2414{
2415 unsigned long addr = (unsigned long)event;
2416 struct buffer_page *bpage = cpu_buffer->commit_page;
2417 struct buffer_page *start;
2418
2419 addr &= PAGE_MASK;
2420
2421 /* Do the likely case first */
2422 if (likely(bpage->page == (void *)addr)) {
2423 local_dec(&bpage->entries);
2424 return;
2425 }
2426
2427 /*
2428 * Because the commit page may be on the reader page we
2429 * start with the next page and check the end loop there.
2430 */
2431 rb_inc_page(cpu_buffer, &bpage);
2432 start = bpage;
2433 do {
2434 if (bpage->page == (void *)addr) {
2435 local_dec(&bpage->entries);
2436 return;
2437 }
2438 rb_inc_page(cpu_buffer, &bpage);
2439 } while (bpage != start);
2440
2441 /* commit not part of this buffer?? */
2442 RB_WARN_ON(cpu_buffer, 1);
2443}
2444
2445/**
2446 * ring_buffer_commit_discard - discard an event that has not been committed
2447 * @buffer: the ring buffer
2448 * @event: non committed event to discard
2449 *
2450 * Sometimes an event that is in the ring buffer needs to be ignored.
2451 * This function lets the user discard an event in the ring buffer
2452 * and then that event will not be read later.
2453 *
2454 * This function only works if it is called before the the item has been
2455 * committed. It will try to free the event from the ring buffer
2456 * if another event has not been added behind it.
2457 *
2458 * If another event has been added behind it, it will set the event
2459 * up as discarded, and perform the commit.
2460 *
2461 * If this function is called, do not call ring_buffer_unlock_commit on
2462 * the event.
2463 */
2464void ring_buffer_discard_commit(struct ring_buffer *buffer,
2465 struct ring_buffer_event *event)
2466{
2467 struct ring_buffer_per_cpu *cpu_buffer;
2468 int cpu;
2469
2470 /* The event is discarded regardless */
2471 rb_event_discard(event);
2472
2473 cpu = smp_processor_id();
2474 cpu_buffer = buffer->buffers[cpu];
2475
2476 /*
2477 * This must only be called if the event has not been
2478 * committed yet. Thus we can assume that preemption
2479 * is still disabled.
2480 */
2481 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2482
2483 rb_decrement_entry(cpu_buffer, event);
2484 if (rb_try_to_discard(cpu_buffer, event))
2485 goto out;
2486
2487 /*
2488 * The commit is still visible by the reader, so we
2489 * must still update the timestamp.
2490 */
2491 rb_update_write_stamp(cpu_buffer, event);
2492 out:
2493 rb_end_commit(cpu_buffer);
2494
2495 trace_recursive_unlock();
2496
2497 preempt_enable_notrace();
2498
2499}
2500EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2501
2502/**
2503 * ring_buffer_write - write data to the buffer without reserving
2504 * @buffer: The ring buffer to write to.
2505 * @length: The length of the data being written (excluding the event header)
2506 * @data: The data to write to the buffer.
2507 *
2508 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2509 * one function. If you already have the data to write to the buffer, it
2510 * may be easier to simply call this function.
2511 *
2512 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2513 * and not the length of the event which would hold the header.
2514 */
2515int ring_buffer_write(struct ring_buffer *buffer,
2516 unsigned long length,
2517 void *data)
2518{
2519 struct ring_buffer_per_cpu *cpu_buffer;
2520 struct ring_buffer_event *event;
2521 void *body;
2522 int ret = -EBUSY;
2523 int cpu;
2524
2525 if (ring_buffer_flags != RB_BUFFERS_ON)
2526 return -EBUSY;
2527
2528 preempt_disable_notrace();
2529
2530 if (atomic_read(&buffer->record_disabled))
2531 goto out;
2532
2533 cpu = raw_smp_processor_id();
2534
2535 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2536 goto out;
2537
2538 cpu_buffer = buffer->buffers[cpu];
2539
2540 if (atomic_read(&cpu_buffer->record_disabled))
2541 goto out;
2542
2543 if (length > BUF_MAX_DATA_SIZE)
2544 goto out;
2545
2546 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2547 if (!event)
2548 goto out;
2549
2550 body = rb_event_data(event);
2551
2552 memcpy(body, data, length);
2553
2554 rb_commit(cpu_buffer, event);
2555
2556 ret = 0;
2557 out:
2558 preempt_enable_notrace();
2559
2560 return ret;
2561}
2562EXPORT_SYMBOL_GPL(ring_buffer_write);
2563
2564static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
2565{
2566 struct buffer_page *reader = cpu_buffer->reader_page;
2567 struct buffer_page *head = rb_set_head_page(cpu_buffer);
2568 struct buffer_page *commit = cpu_buffer->commit_page;
2569
2570 /* In case of error, head will be NULL */
2571 if (unlikely(!head))
2572 return 1;
2573
2574 return reader->read == rb_page_commit(reader) &&
2575 (commit == reader ||
2576 (commit == head &&
2577 head->read == rb_page_commit(commit)));
2578}
2579
2580/**
2581 * ring_buffer_record_disable - stop all writes into the buffer
2582 * @buffer: The ring buffer to stop writes to.
2583 *
2584 * This prevents all writes to the buffer. Any attempt to write
2585 * to the buffer after this will fail and return NULL.
2586 *
2587 * The caller should call synchronize_sched() after this.
2588 */
2589void ring_buffer_record_disable(struct ring_buffer *buffer)
2590{
2591 atomic_inc(&buffer->record_disabled);
2592}
2593EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
2594
2595/**
2596 * ring_buffer_record_enable - enable writes to the buffer
2597 * @buffer: The ring buffer to enable writes
2598 *
2599 * Note, multiple disables will need the same number of enables
2600 * to truly enable the writing (much like preempt_disable).
2601 */
2602void ring_buffer_record_enable(struct ring_buffer *buffer)
2603{
2604 atomic_dec(&buffer->record_disabled);
2605}
2606EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
2607
2608/**
2609 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
2610 * @buffer: The ring buffer to stop writes to.
2611 * @cpu: The CPU buffer to stop
2612 *
2613 * This prevents all writes to the buffer. Any attempt to write
2614 * to the buffer after this will fail and return NULL.
2615 *
2616 * The caller should call synchronize_sched() after this.
2617 */
2618void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
2619{
2620 struct ring_buffer_per_cpu *cpu_buffer;
2621
2622 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2623 return;
2624
2625 cpu_buffer = buffer->buffers[cpu];
2626 atomic_inc(&cpu_buffer->record_disabled);
2627}
2628EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
2629
2630/**
2631 * ring_buffer_record_enable_cpu - enable writes to the buffer
2632 * @buffer: The ring buffer to enable writes
2633 * @cpu: The CPU to enable.
2634 *
2635 * Note, multiple disables will need the same number of enables
2636 * to truly enable the writing (much like preempt_disable).
2637 */
2638void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
2639{
2640 struct ring_buffer_per_cpu *cpu_buffer;
2641
2642 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2643 return;
2644
2645 cpu_buffer = buffer->buffers[cpu];
2646 atomic_dec(&cpu_buffer->record_disabled);
2647}
2648EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
2649
2650/*
2651 * The total entries in the ring buffer is the running counter
2652 * of entries entered into the ring buffer, minus the sum of
2653 * the entries read from the ring buffer and the number of
2654 * entries that were overwritten.
2655 */
2656static inline unsigned long
2657rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
2658{
2659 return local_read(&cpu_buffer->entries) -
2660 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
2661}
2662
2663/**
2664 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
2665 * @buffer: The ring buffer
2666 * @cpu: The per CPU buffer to get the entries from.
2667 */
2668unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
2669{
2670 struct ring_buffer_per_cpu *cpu_buffer;
2671
2672 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2673 return 0;
2674
2675 cpu_buffer = buffer->buffers[cpu];
2676
2677 return rb_num_of_entries(cpu_buffer);
2678}
2679EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
2680
2681/**
2682 * ring_buffer_overrun_cpu - get the number of overruns in a cpu_buffer
2683 * @buffer: The ring buffer
2684 * @cpu: The per CPU buffer to get the number of overruns from
2685 */
2686unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
2687{
2688 struct ring_buffer_per_cpu *cpu_buffer;
2689 unsigned long ret;
2690
2691 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2692 return 0;
2693
2694 cpu_buffer = buffer->buffers[cpu];
2695 ret = local_read(&cpu_buffer->overrun);
2696
2697 return ret;
2698}
2699EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
2700
2701/**
2702 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by commits
2703 * @buffer: The ring buffer
2704 * @cpu: The per CPU buffer to get the number of overruns from
2705 */
2706unsigned long
2707ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
2708{
2709 struct ring_buffer_per_cpu *cpu_buffer;
2710 unsigned long ret;
2711
2712 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2713 return 0;
2714
2715 cpu_buffer = buffer->buffers[cpu];
2716 ret = local_read(&cpu_buffer->commit_overrun);
2717
2718 return ret;
2719}
2720EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
2721
2722/**
2723 * ring_buffer_entries - get the number of entries in a buffer
2724 * @buffer: The ring buffer
2725 *
2726 * Returns the total number of entries in the ring buffer
2727 * (all CPU entries)
2728 */
2729unsigned long ring_buffer_entries(struct ring_buffer *buffer)
2730{
2731 struct ring_buffer_per_cpu *cpu_buffer;
2732 unsigned long entries = 0;
2733 int cpu;
2734
2735 /* if you care about this being correct, lock the buffer */
2736 for_each_buffer_cpu(buffer, cpu) {
2737 cpu_buffer = buffer->buffers[cpu];
2738 entries += rb_num_of_entries(cpu_buffer);
2739 }
2740
2741 return entries;
2742}
2743EXPORT_SYMBOL_GPL(ring_buffer_entries);
2744
2745/**
2746 * ring_buffer_overruns - get the number of overruns in buffer
2747 * @buffer: The ring buffer
2748 *
2749 * Returns the total number of overruns in the ring buffer
2750 * (all CPU entries)
2751 */
2752unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
2753{
2754 struct ring_buffer_per_cpu *cpu_buffer;
2755 unsigned long overruns = 0;
2756 int cpu;
2757
2758 /* if you care about this being correct, lock the buffer */
2759 for_each_buffer_cpu(buffer, cpu) {
2760 cpu_buffer = buffer->buffers[cpu];
2761 overruns += local_read(&cpu_buffer->overrun);
2762 }
2763
2764 return overruns;
2765}
2766EXPORT_SYMBOL_GPL(ring_buffer_overruns);
2767
2768static void rb_iter_reset(struct ring_buffer_iter *iter)
2769{
2770 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2771
2772 /* Iterator usage is expected to have record disabled */
2773 if (list_empty(&cpu_buffer->reader_page->list)) {
2774 iter->head_page = rb_set_head_page(cpu_buffer);
2775 if (unlikely(!iter->head_page))
2776 return;
2777 iter->head = iter->head_page->read;
2778 } else {
2779 iter->head_page = cpu_buffer->reader_page;
2780 iter->head = cpu_buffer->reader_page->read;
2781 }
2782 if (iter->head)
2783 iter->read_stamp = cpu_buffer->read_stamp;
2784 else
2785 iter->read_stamp = iter->head_page->page->time_stamp;
2786 iter->cache_reader_page = cpu_buffer->reader_page;
2787 iter->cache_read = cpu_buffer->read;
2788}
2789
2790/**
2791 * ring_buffer_iter_reset - reset an iterator
2792 * @iter: The iterator to reset
2793 *
2794 * Resets the iterator, so that it will start from the beginning
2795 * again.
2796 */
2797void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
2798{
2799 struct ring_buffer_per_cpu *cpu_buffer;
2800 unsigned long flags;
2801
2802 if (!iter)
2803 return;
2804
2805 cpu_buffer = iter->cpu_buffer;
2806
2807 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2808 rb_iter_reset(iter);
2809 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2810}
2811EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
2812
2813/**
2814 * ring_buffer_iter_empty - check if an iterator has no more to read
2815 * @iter: The iterator to check
2816 */
2817int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
2818{
2819 struct ring_buffer_per_cpu *cpu_buffer;
2820
2821 cpu_buffer = iter->cpu_buffer;
2822
2823 return iter->head_page == cpu_buffer->commit_page &&
2824 iter->head == rb_commit_index(cpu_buffer);
2825}
2826EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
2827
2828static void
2829rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2830 struct ring_buffer_event *event)
2831{
2832 u64 delta;
2833
2834 switch (event->type_len) {
2835 case RINGBUF_TYPE_PADDING:
2836 return;
2837
2838 case RINGBUF_TYPE_TIME_EXTEND:
2839 delta = event->array[0];
2840 delta <<= TS_SHIFT;
2841 delta += event->time_delta;
2842 cpu_buffer->read_stamp += delta;
2843 return;
2844
2845 case RINGBUF_TYPE_TIME_STAMP:
2846 /* FIXME: not implemented */
2847 return;
2848
2849 case RINGBUF_TYPE_DATA:
2850 cpu_buffer->read_stamp += event->time_delta;
2851 return;
2852
2853 default:
2854 BUG();
2855 }
2856 return;
2857}
2858
2859static void
2860rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
2861 struct ring_buffer_event *event)
2862{
2863 u64 delta;
2864
2865 switch (event->type_len) {
2866 case RINGBUF_TYPE_PADDING:
2867 return;
2868
2869 case RINGBUF_TYPE_TIME_EXTEND:
2870 delta = event->array[0];
2871 delta <<= TS_SHIFT;
2872 delta += event->time_delta;
2873 iter->read_stamp += delta;
2874 return;
2875
2876 case RINGBUF_TYPE_TIME_STAMP:
2877 /* FIXME: not implemented */
2878 return;
2879
2880 case RINGBUF_TYPE_DATA:
2881 iter->read_stamp += event->time_delta;
2882 return;
2883
2884 default:
2885 BUG();
2886 }
2887 return;
2888}
2889
2890static struct buffer_page *
2891rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
2892{
2893 struct buffer_page *reader = NULL;
2894 unsigned long overwrite;
2895 unsigned long flags;
2896 int nr_loops = 0;
2897 int ret;
2898
2899 local_irq_save(flags);
2900 arch_spin_lock(&cpu_buffer->lock);
2901
2902 again:
2903 /*
2904 * This should normally only loop twice. But because the
2905 * start of the reader inserts an empty page, it causes
2906 * a case where we will loop three times. There should be no
2907 * reason to loop four times (that I know of).
2908 */
2909 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
2910 reader = NULL;
2911 goto out;
2912 }
2913
2914 reader = cpu_buffer->reader_page;
2915
2916 /* If there's more to read, return this page */
2917 if (cpu_buffer->reader_page->read < rb_page_size(reader))
2918 goto out;
2919
2920 /* Never should we have an index greater than the size */
2921 if (RB_WARN_ON(cpu_buffer,
2922 cpu_buffer->reader_page->read > rb_page_size(reader)))
2923 goto out;
2924
2925 /* check if we caught up to the tail */
2926 reader = NULL;
2927 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
2928 goto out;
2929
2930 /*
2931 * Reset the reader page to size zero.
2932 */
2933 local_set(&cpu_buffer->reader_page->write, 0);
2934 local_set(&cpu_buffer->reader_page->entries, 0);
2935 local_set(&cpu_buffer->reader_page->page->commit, 0);
2936 cpu_buffer->reader_page->real_end = 0;
2937
2938 spin:
2939 /*
2940 * Splice the empty reader page into the list around the head.
2941 */
2942 reader = rb_set_head_page(cpu_buffer);
2943 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
2944 cpu_buffer->reader_page->list.prev = reader->list.prev;
2945
2946 /*
2947 * cpu_buffer->pages just needs to point to the buffer, it
2948 * has no specific buffer page to point to. Lets move it out
2949 * of our way so we don't accidentally swap it.
2950 */
2951 cpu_buffer->pages = reader->list.prev;
2952
2953 /* The reader page will be pointing to the new head */
2954 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
2955
2956 /*
2957 * We want to make sure we read the overruns after we set up our
2958 * pointers to the next object. The writer side does a
2959 * cmpxchg to cross pages which acts as the mb on the writer
2960 * side. Note, the reader will constantly fail the swap
2961 * while the writer is updating the pointers, so this
2962 * guarantees that the overwrite recorded here is the one we
2963 * want to compare with the last_overrun.
2964 */
2965 smp_mb();
2966 overwrite = local_read(&(cpu_buffer->overrun));
2967
2968 /*
2969 * Here's the tricky part.
2970 *
2971 * We need to move the pointer past the header page.
2972 * But we can only do that if a writer is not currently
2973 * moving it. The page before the header page has the
2974 * flag bit '1' set if it is pointing to the page we want.
2975 * but if the writer is in the process of moving it
2976 * than it will be '2' or already moved '0'.
2977 */
2978
2979 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
2980
2981 /*
2982 * If we did not convert it, then we must try again.
2983 */
2984 if (!ret)
2985 goto spin;
2986
2987 /*
2988 * Yeah! We succeeded in replacing the page.
2989 *
2990 * Now make the new head point back to the reader page.
2991 */
2992 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
2993 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
2994
2995 /* Finally update the reader page to the new head */
2996 cpu_buffer->reader_page = reader;
2997 rb_reset_reader_page(cpu_buffer);
2998
2999 if (overwrite != cpu_buffer->last_overrun) {
3000 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3001 cpu_buffer->last_overrun = overwrite;
3002 }
3003
3004 goto again;
3005
3006 out:
3007 arch_spin_unlock(&cpu_buffer->lock);
3008 local_irq_restore(flags);
3009
3010 return reader;
3011}
3012
3013static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3014{
3015 struct ring_buffer_event *event;
3016 struct buffer_page *reader;
3017 unsigned length;
3018
3019 reader = rb_get_reader_page(cpu_buffer);
3020
3021 /* This function should not be called when buffer is empty */
3022 if (RB_WARN_ON(cpu_buffer, !reader))
3023 return;
3024
3025 event = rb_reader_event(cpu_buffer);
3026
3027 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3028 cpu_buffer->read++;
3029
3030 rb_update_read_stamp(cpu_buffer, event);
3031
3032 length = rb_event_length(event);
3033 cpu_buffer->reader_page->read += length;
3034}
3035
3036static void rb_advance_iter(struct ring_buffer_iter *iter)
3037{
3038 struct ring_buffer_per_cpu *cpu_buffer;
3039 struct ring_buffer_event *event;
3040 unsigned length;
3041
3042 cpu_buffer = iter->cpu_buffer;
3043
3044 /*
3045 * Check if we are at the end of the buffer.
3046 */
3047 if (iter->head >= rb_page_size(iter->head_page)) {
3048 /* discarded commits can make the page empty */
3049 if (iter->head_page == cpu_buffer->commit_page)
3050 return;
3051 rb_inc_iter(iter);
3052 return;
3053 }
3054
3055 event = rb_iter_head_event(iter);
3056
3057 length = rb_event_length(event);
3058
3059 /*
3060 * This should not be called to advance the header if we are
3061 * at the tail of the buffer.
3062 */
3063 if (RB_WARN_ON(cpu_buffer,
3064 (iter->head_page == cpu_buffer->commit_page) &&
3065 (iter->head + length > rb_commit_index(cpu_buffer))))
3066 return;
3067
3068 rb_update_iter_read_stamp(iter, event);
3069
3070 iter->head += length;
3071
3072 /* check for end of page padding */
3073 if ((iter->head >= rb_page_size(iter->head_page)) &&
3074 (iter->head_page != cpu_buffer->commit_page))
3075 rb_advance_iter(iter);
3076}
3077
3078static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3079{
3080 return cpu_buffer->lost_events;
3081}
3082
3083static struct ring_buffer_event *
3084rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3085 unsigned long *lost_events)
3086{
3087 struct ring_buffer_event *event;
3088 struct buffer_page *reader;
3089 int nr_loops = 0;
3090
3091 again:
3092 /*
3093 * We repeat when a time extend is encountered.
3094 * Since the time extend is always attached to a data event,
3095 * we should never loop more than once.
3096 * (We never hit the following condition more than twice).
3097 */
3098 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3099 return NULL;
3100
3101 reader = rb_get_reader_page(cpu_buffer);
3102 if (!reader)
3103 return NULL;
3104
3105 event = rb_reader_event(cpu_buffer);
3106
3107 switch (event->type_len) {
3108 case RINGBUF_TYPE_PADDING:
3109 if (rb_null_event(event))
3110 RB_WARN_ON(cpu_buffer, 1);
3111 /*
3112 * Because the writer could be discarding every
3113 * event it creates (which would probably be bad)
3114 * if we were to go back to "again" then we may never
3115 * catch up, and will trigger the warn on, or lock
3116 * the box. Return the padding, and we will release
3117 * the current locks, and try again.
3118 */
3119 return event;
3120
3121 case RINGBUF_TYPE_TIME_EXTEND:
3122 /* Internal data, OK to advance */
3123 rb_advance_reader(cpu_buffer);
3124 goto again;
3125
3126 case RINGBUF_TYPE_TIME_STAMP:
3127 /* FIXME: not implemented */
3128 rb_advance_reader(cpu_buffer);
3129 goto again;
3130
3131 case RINGBUF_TYPE_DATA:
3132 if (ts) {
3133 *ts = cpu_buffer->read_stamp + event->time_delta;
3134 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3135 cpu_buffer->cpu, ts);
3136 }
3137 if (lost_events)
3138 *lost_events = rb_lost_events(cpu_buffer);
3139 return event;
3140
3141 default:
3142 BUG();
3143 }
3144
3145 return NULL;
3146}
3147EXPORT_SYMBOL_GPL(ring_buffer_peek);
3148
3149static struct ring_buffer_event *
3150rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3151{
3152 struct ring_buffer *buffer;
3153 struct ring_buffer_per_cpu *cpu_buffer;
3154 struct ring_buffer_event *event;
3155 int nr_loops = 0;
3156
3157 cpu_buffer = iter->cpu_buffer;
3158 buffer = cpu_buffer->buffer;
3159
3160 /*
3161 * Check if someone performed a consuming read to
3162 * the buffer. A consuming read invalidates the iterator
3163 * and we need to reset the iterator in this case.
3164 */
3165 if (unlikely(iter->cache_read != cpu_buffer->read ||
3166 iter->cache_reader_page != cpu_buffer->reader_page))
3167 rb_iter_reset(iter);
3168
3169 again:
3170 if (ring_buffer_iter_empty(iter))
3171 return NULL;
3172
3173 /*
3174 * We repeat when a time extend is encountered.
3175 * Since the time extend is always attached to a data event,
3176 * we should never loop more than once.
3177 * (We never hit the following condition more than twice).
3178 */
3179 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3180 return NULL;
3181
3182 if (rb_per_cpu_empty(cpu_buffer))
3183 return NULL;
3184
3185 if (iter->head >= local_read(&iter->head_page->page->commit)) {
3186 rb_inc_iter(iter);
3187 goto again;
3188 }
3189
3190 event = rb_iter_head_event(iter);
3191
3192 switch (event->type_len) {
3193 case RINGBUF_TYPE_PADDING:
3194 if (rb_null_event(event)) {
3195 rb_inc_iter(iter);
3196 goto again;
3197 }
3198 rb_advance_iter(iter);
3199 return event;
3200
3201 case RINGBUF_TYPE_TIME_EXTEND:
3202 /* Internal data, OK to advance */
3203 rb_advance_iter(iter);
3204 goto again;
3205
3206 case RINGBUF_TYPE_TIME_STAMP:
3207 /* FIXME: not implemented */
3208 rb_advance_iter(iter);
3209 goto again;
3210
3211 case RINGBUF_TYPE_DATA:
3212 if (ts) {
3213 *ts = iter->read_stamp + event->time_delta;
3214 ring_buffer_normalize_time_stamp(buffer,
3215 cpu_buffer->cpu, ts);
3216 }
3217 return event;
3218
3219 default:
3220 BUG();
3221 }
3222
3223 return NULL;
3224}
3225EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3226
3227static inline int rb_ok_to_lock(void)
3228{
3229 /*
3230 * If an NMI die dumps out the content of the ring buffer
3231 * do not grab locks. We also permanently disable the ring
3232 * buffer too. A one time deal is all you get from reading
3233 * the ring buffer from an NMI.
3234 */
3235 if (likely(!in_nmi()))
3236 return 1;
3237
3238 tracing_off_permanent();
3239 return 0;
3240}
3241
3242/**
3243 * ring_buffer_peek - peek at the next event to be read
3244 * @buffer: The ring buffer to read
3245 * @cpu: The cpu to peak at
3246 * @ts: The timestamp counter of this event.
3247 * @lost_events: a variable to store if events were lost (may be NULL)
3248 *
3249 * This will return the event that will be read next, but does
3250 * not consume the data.
3251 */
3252struct ring_buffer_event *
3253ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3254 unsigned long *lost_events)
3255{
3256 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3257 struct ring_buffer_event *event;
3258 unsigned long flags;
3259 int dolock;
3260
3261 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3262 return NULL;
3263
3264 dolock = rb_ok_to_lock();
3265 again:
3266 local_irq_save(flags);
3267 if (dolock)
3268 spin_lock(&cpu_buffer->reader_lock);
3269 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3270 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3271 rb_advance_reader(cpu_buffer);
3272 if (dolock)
3273 spin_unlock(&cpu_buffer->reader_lock);
3274 local_irq_restore(flags);
3275
3276 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3277 goto again;
3278
3279 return event;
3280}
3281
3282/**
3283 * ring_buffer_iter_peek - peek at the next event to be read
3284 * @iter: The ring buffer iterator
3285 * @ts: The timestamp counter of this event.
3286 *
3287 * This will return the event that will be read next, but does
3288 * not increment the iterator.
3289 */
3290struct ring_buffer_event *
3291ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3292{
3293 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3294 struct ring_buffer_event *event;
3295 unsigned long flags;
3296
3297 again:
3298 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3299 event = rb_iter_peek(iter, ts);
3300 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3301
3302 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3303 goto again;
3304
3305 return event;
3306}
3307
3308/**
3309 * ring_buffer_consume - return an event and consume it
3310 * @buffer: The ring buffer to get the next event from
3311 * @cpu: the cpu to read the buffer from
3312 * @ts: a variable to store the timestamp (may be NULL)
3313 * @lost_events: a variable to store if events were lost (may be NULL)
3314 *
3315 * Returns the next event in the ring buffer, and that event is consumed.
3316 * Meaning, that sequential reads will keep returning a different event,
3317 * and eventually empty the ring buffer if the producer is slower.
3318 */
3319struct ring_buffer_event *
3320ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3321 unsigned long *lost_events)
3322{
3323 struct ring_buffer_per_cpu *cpu_buffer;
3324 struct ring_buffer_event *event = NULL;
3325 unsigned long flags;
3326 int dolock;
3327
3328 dolock = rb_ok_to_lock();
3329
3330 again:
3331 /* might be called in atomic */
3332 preempt_disable();
3333
3334 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3335 goto out;
3336
3337 cpu_buffer = buffer->buffers[cpu];
3338 local_irq_save(flags);
3339 if (dolock)
3340 spin_lock(&cpu_buffer->reader_lock);
3341
3342 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3343 if (event) {
3344 cpu_buffer->lost_events = 0;
3345 rb_advance_reader(cpu_buffer);
3346 }
3347
3348 if (dolock)
3349 spin_unlock(&cpu_buffer->reader_lock);
3350 local_irq_restore(flags);
3351
3352 out:
3353 preempt_enable();
3354
3355 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3356 goto again;
3357
3358 return event;
3359}
3360EXPORT_SYMBOL_GPL(ring_buffer_consume);
3361
3362/**
3363 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3364 * @buffer: The ring buffer to read from
3365 * @cpu: The cpu buffer to iterate over
3366 *
3367 * This performs the initial preparations necessary to iterate
3368 * through the buffer. Memory is allocated, buffer recording
3369 * is disabled, and the iterator pointer is returned to the caller.
3370 *
3371 * Disabling buffer recordng prevents the reading from being
3372 * corrupted. This is not a consuming read, so a producer is not
3373 * expected.
3374 *
3375 * After a sequence of ring_buffer_read_prepare calls, the user is
3376 * expected to make at least one call to ring_buffer_prepare_sync.
3377 * Afterwards, ring_buffer_read_start is invoked to get things going
3378 * for real.
3379 *
3380 * This overall must be paired with ring_buffer_finish.
3381 */
3382struct ring_buffer_iter *
3383ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3384{
3385 struct ring_buffer_per_cpu *cpu_buffer;
3386 struct ring_buffer_iter *iter;
3387
3388 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3389 return NULL;
3390
3391 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3392 if (!iter)
3393 return NULL;
3394
3395 cpu_buffer = buffer->buffers[cpu];
3396
3397 iter->cpu_buffer = cpu_buffer;
3398
3399 atomic_inc(&cpu_buffer->record_disabled);
3400
3401 return iter;
3402}
3403EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
3404
3405/**
3406 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
3407 *
3408 * All previously invoked ring_buffer_read_prepare calls to prepare
3409 * iterators will be synchronized. Afterwards, read_buffer_read_start
3410 * calls on those iterators are allowed.
3411 */
3412void
3413ring_buffer_read_prepare_sync(void)
3414{
3415 synchronize_sched();
3416}
3417EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
3418
3419/**
3420 * ring_buffer_read_start - start a non consuming read of the buffer
3421 * @iter: The iterator returned by ring_buffer_read_prepare
3422 *
3423 * This finalizes the startup of an iteration through the buffer.
3424 * The iterator comes from a call to ring_buffer_read_prepare and
3425 * an intervening ring_buffer_read_prepare_sync must have been
3426 * performed.
3427 *
3428 * Must be paired with ring_buffer_finish.
3429 */
3430void
3431ring_buffer_read_start(struct ring_buffer_iter *iter)
3432{
3433 struct ring_buffer_per_cpu *cpu_buffer;
3434 unsigned long flags;
3435
3436 if (!iter)
3437 return;
3438
3439 cpu_buffer = iter->cpu_buffer;
3440
3441 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3442 arch_spin_lock(&cpu_buffer->lock);
3443 rb_iter_reset(iter);
3444 arch_spin_unlock(&cpu_buffer->lock);
3445 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3446}
3447EXPORT_SYMBOL_GPL(ring_buffer_read_start);
3448
3449/**
3450 * ring_buffer_finish - finish reading the iterator of the buffer
3451 * @iter: The iterator retrieved by ring_buffer_start
3452 *
3453 * This re-enables the recording to the buffer, and frees the
3454 * iterator.
3455 */
3456void
3457ring_buffer_read_finish(struct ring_buffer_iter *iter)
3458{
3459 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3460
3461 atomic_dec(&cpu_buffer->record_disabled);
3462 kfree(iter);
3463}
3464EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
3465
3466/**
3467 * ring_buffer_read - read the next item in the ring buffer by the iterator
3468 * @iter: The ring buffer iterator
3469 * @ts: The time stamp of the event read.
3470 *
3471 * This reads the next event in the ring buffer and increments the iterator.
3472 */
3473struct ring_buffer_event *
3474ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
3475{
3476 struct ring_buffer_event *event;
3477 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3478 unsigned long flags;
3479
3480 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3481 again:
3482 event = rb_iter_peek(iter, ts);
3483 if (!event)
3484 goto out;
3485
3486 if (event->type_len == RINGBUF_TYPE_PADDING)
3487 goto again;
3488
3489 rb_advance_iter(iter);
3490 out:
3491 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3492
3493 return event;
3494}
3495EXPORT_SYMBOL_GPL(ring_buffer_read);
3496
3497/**
3498 * ring_buffer_size - return the size of the ring buffer (in bytes)
3499 * @buffer: The ring buffer.
3500 */
3501unsigned long ring_buffer_size(struct ring_buffer *buffer)
3502{
3503 return BUF_PAGE_SIZE * buffer->pages;
3504}
3505EXPORT_SYMBOL_GPL(ring_buffer_size);
3506
3507static void
3508rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
3509{
3510 rb_head_page_deactivate(cpu_buffer);
3511
3512 cpu_buffer->head_page
3513 = list_entry(cpu_buffer->pages, struct buffer_page, list);
3514 local_set(&cpu_buffer->head_page->write, 0);
3515 local_set(&cpu_buffer->head_page->entries, 0);
3516 local_set(&cpu_buffer->head_page->page->commit, 0);
3517
3518 cpu_buffer->head_page->read = 0;
3519
3520 cpu_buffer->tail_page = cpu_buffer->head_page;
3521 cpu_buffer->commit_page = cpu_buffer->head_page;
3522
3523 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
3524 local_set(&cpu_buffer->reader_page->write, 0);
3525 local_set(&cpu_buffer->reader_page->entries, 0);
3526 local_set(&cpu_buffer->reader_page->page->commit, 0);
3527 cpu_buffer->reader_page->read = 0;
3528
3529 local_set(&cpu_buffer->commit_overrun, 0);
3530 local_set(&cpu_buffer->overrun, 0);
3531 local_set(&cpu_buffer->entries, 0);
3532 local_set(&cpu_buffer->committing, 0);
3533 local_set(&cpu_buffer->commits, 0);
3534 cpu_buffer->read = 0;
3535
3536 cpu_buffer->write_stamp = 0;
3537 cpu_buffer->read_stamp = 0;
3538
3539 cpu_buffer->lost_events = 0;
3540 cpu_buffer->last_overrun = 0;
3541
3542 rb_head_page_activate(cpu_buffer);
3543}
3544
3545/**
3546 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
3547 * @buffer: The ring buffer to reset a per cpu buffer of
3548 * @cpu: The CPU buffer to be reset
3549 */
3550void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
3551{
3552 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3553 unsigned long flags;
3554
3555 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3556 return;
3557
3558 atomic_inc(&cpu_buffer->record_disabled);
3559
3560 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3561
3562 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
3563 goto out;
3564
3565 arch_spin_lock(&cpu_buffer->lock);
3566
3567 rb_reset_cpu(cpu_buffer);
3568
3569 arch_spin_unlock(&cpu_buffer->lock);
3570
3571 out:
3572 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3573
3574 atomic_dec(&cpu_buffer->record_disabled);
3575}
3576EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
3577
3578/**
3579 * ring_buffer_reset - reset a ring buffer
3580 * @buffer: The ring buffer to reset all cpu buffers
3581 */
3582void ring_buffer_reset(struct ring_buffer *buffer)
3583{
3584 int cpu;
3585
3586 for_each_buffer_cpu(buffer, cpu)
3587 ring_buffer_reset_cpu(buffer, cpu);
3588}
3589EXPORT_SYMBOL_GPL(ring_buffer_reset);
3590
3591/**
3592 * rind_buffer_empty - is the ring buffer empty?
3593 * @buffer: The ring buffer to test
3594 */
3595int ring_buffer_empty(struct ring_buffer *buffer)
3596{
3597 struct ring_buffer_per_cpu *cpu_buffer;
3598 unsigned long flags;
3599 int dolock;
3600 int cpu;
3601 int ret;
3602
3603 dolock = rb_ok_to_lock();
3604
3605 /* yes this is racy, but if you don't like the race, lock the buffer */
3606 for_each_buffer_cpu(buffer, cpu) {
3607 cpu_buffer = buffer->buffers[cpu];
3608 local_irq_save(flags);
3609 if (dolock)
3610 spin_lock(&cpu_buffer->reader_lock);
3611 ret = rb_per_cpu_empty(cpu_buffer);
3612 if (dolock)
3613 spin_unlock(&cpu_buffer->reader_lock);
3614 local_irq_restore(flags);
3615
3616 if (!ret)
3617 return 0;
3618 }
3619
3620 return 1;
3621}
3622EXPORT_SYMBOL_GPL(ring_buffer_empty);
3623
3624/**
3625 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
3626 * @buffer: The ring buffer
3627 * @cpu: The CPU buffer to test
3628 */
3629int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
3630{
3631 struct ring_buffer_per_cpu *cpu_buffer;
3632 unsigned long flags;
3633 int dolock;
3634 int ret;
3635
3636 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3637 return 1;
3638
3639 dolock = rb_ok_to_lock();
3640
3641 cpu_buffer = buffer->buffers[cpu];
3642 local_irq_save(flags);
3643 if (dolock)
3644 spin_lock(&cpu_buffer->reader_lock);
3645 ret = rb_per_cpu_empty(cpu_buffer);
3646 if (dolock)
3647 spin_unlock(&cpu_buffer->reader_lock);
3648 local_irq_restore(flags);
3649
3650 return ret;
3651}
3652EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
3653
3654#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3655/**
3656 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
3657 * @buffer_a: One buffer to swap with
3658 * @buffer_b: The other buffer to swap with
3659 *
3660 * This function is useful for tracers that want to take a "snapshot"
3661 * of a CPU buffer and has another back up buffer lying around.
3662 * it is expected that the tracer handles the cpu buffer not being
3663 * used at the moment.
3664 */
3665int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
3666 struct ring_buffer *buffer_b, int cpu)
3667{
3668 struct ring_buffer_per_cpu *cpu_buffer_a;
3669 struct ring_buffer_per_cpu *cpu_buffer_b;
3670 int ret = -EINVAL;
3671
3672 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
3673 !cpumask_test_cpu(cpu, buffer_b->cpumask))
3674 goto out;
3675
3676 /* At least make sure the two buffers are somewhat the same */
3677 if (buffer_a->pages != buffer_b->pages)
3678 goto out;
3679
3680 ret = -EAGAIN;
3681
3682 if (ring_buffer_flags != RB_BUFFERS_ON)
3683 goto out;
3684
3685 if (atomic_read(&buffer_a->record_disabled))
3686 goto out;
3687
3688 if (atomic_read(&buffer_b->record_disabled))
3689 goto out;
3690
3691 cpu_buffer_a = buffer_a->buffers[cpu];
3692 cpu_buffer_b = buffer_b->buffers[cpu];
3693
3694 if (atomic_read(&cpu_buffer_a->record_disabled))
3695 goto out;
3696
3697 if (atomic_read(&cpu_buffer_b->record_disabled))
3698 goto out;
3699
3700 /*
3701 * We can't do a synchronize_sched here because this
3702 * function can be called in atomic context.
3703 * Normally this will be called from the same CPU as cpu.
3704 * If not it's up to the caller to protect this.
3705 */
3706 atomic_inc(&cpu_buffer_a->record_disabled);
3707 atomic_inc(&cpu_buffer_b->record_disabled);
3708
3709 ret = -EBUSY;
3710 if (local_read(&cpu_buffer_a->committing))
3711 goto out_dec;
3712 if (local_read(&cpu_buffer_b->committing))
3713 goto out_dec;
3714
3715 buffer_a->buffers[cpu] = cpu_buffer_b;
3716 buffer_b->buffers[cpu] = cpu_buffer_a;
3717
3718 cpu_buffer_b->buffer = buffer_a;
3719 cpu_buffer_a->buffer = buffer_b;
3720
3721 ret = 0;
3722
3723out_dec:
3724 atomic_dec(&cpu_buffer_a->record_disabled);
3725 atomic_dec(&cpu_buffer_b->record_disabled);
3726out:
3727 return ret;
3728}
3729EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
3730#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
3731
3732/**
3733 * ring_buffer_alloc_read_page - allocate a page to read from buffer
3734 * @buffer: the buffer to allocate for.
3735 *
3736 * This function is used in conjunction with ring_buffer_read_page.
3737 * When reading a full page from the ring buffer, these functions
3738 * can be used to speed up the process. The calling function should
3739 * allocate a few pages first with this function. Then when it
3740 * needs to get pages from the ring buffer, it passes the result
3741 * of this function into ring_buffer_read_page, which will swap
3742 * the page that was allocated, with the read page of the buffer.
3743 *
3744 * Returns:
3745 * The page allocated, or NULL on error.
3746 */
3747void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
3748{
3749 struct buffer_data_page *bpage;
3750 struct page *page;
3751
3752 page = alloc_pages_node(cpu_to_node(cpu),
3753 GFP_KERNEL | __GFP_NORETRY, 0);
3754 if (!page)
3755 return NULL;
3756
3757 bpage = page_address(page);
3758
3759 rb_init_page(bpage);
3760
3761 return bpage;
3762}
3763EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
3764
3765/**
3766 * ring_buffer_free_read_page - free an allocated read page
3767 * @buffer: the buffer the page was allocate for
3768 * @data: the page to free
3769 *
3770 * Free a page allocated from ring_buffer_alloc_read_page.
3771 */
3772void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
3773{
3774 free_page((unsigned long)data);
3775}
3776EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
3777
3778/**
3779 * ring_buffer_read_page - extract a page from the ring buffer
3780 * @buffer: buffer to extract from
3781 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
3782 * @len: amount to extract
3783 * @cpu: the cpu of the buffer to extract
3784 * @full: should the extraction only happen when the page is full.
3785 *
3786 * This function will pull out a page from the ring buffer and consume it.
3787 * @data_page must be the address of the variable that was returned
3788 * from ring_buffer_alloc_read_page. This is because the page might be used
3789 * to swap with a page in the ring buffer.
3790 *
3791 * for example:
3792 * rpage = ring_buffer_alloc_read_page(buffer);
3793 * if (!rpage)
3794 * return error;
3795 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
3796 * if (ret >= 0)
3797 * process_page(rpage, ret);
3798 *
3799 * When @full is set, the function will not return true unless
3800 * the writer is off the reader page.
3801 *
3802 * Note: it is up to the calling functions to handle sleeps and wakeups.
3803 * The ring buffer can be used anywhere in the kernel and can not
3804 * blindly call wake_up. The layer that uses the ring buffer must be
3805 * responsible for that.
3806 *
3807 * Returns:
3808 * >=0 if data has been transferred, returns the offset of consumed data.
3809 * <0 if no data has been transferred.
3810 */
3811int ring_buffer_read_page(struct ring_buffer *buffer,
3812 void **data_page, size_t len, int cpu, int full)
3813{
3814 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3815 struct ring_buffer_event *event;
3816 struct buffer_data_page *bpage;
3817 struct buffer_page *reader;
3818 unsigned long missed_events;
3819 unsigned long flags;
3820 unsigned int commit;
3821 unsigned int read;
3822 u64 save_timestamp;
3823 int ret = -1;
3824
3825 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3826 goto out;
3827
3828 /*
3829 * If len is not big enough to hold the page header, then
3830 * we can not copy anything.
3831 */
3832 if (len <= BUF_PAGE_HDR_SIZE)
3833 goto out;
3834
3835 len -= BUF_PAGE_HDR_SIZE;
3836
3837 if (!data_page)
3838 goto out;
3839
3840 bpage = *data_page;
3841 if (!bpage)
3842 goto out;
3843
3844 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3845
3846 reader = rb_get_reader_page(cpu_buffer);
3847 if (!reader)
3848 goto out_unlock;
3849
3850 event = rb_reader_event(cpu_buffer);
3851
3852 read = reader->read;
3853 commit = rb_page_commit(reader);
3854
3855 /* Check if any events were dropped */
3856 missed_events = cpu_buffer->lost_events;
3857
3858 /*
3859 * If this page has been partially read or
3860 * if len is not big enough to read the rest of the page or
3861 * a writer is still on the page, then
3862 * we must copy the data from the page to the buffer.
3863 * Otherwise, we can simply swap the page with the one passed in.
3864 */
3865 if (read || (len < (commit - read)) ||
3866 cpu_buffer->reader_page == cpu_buffer->commit_page) {
3867 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
3868 unsigned int rpos = read;
3869 unsigned int pos = 0;
3870 unsigned int size;
3871
3872 if (full)
3873 goto out_unlock;
3874
3875 if (len > (commit - read))
3876 len = (commit - read);
3877
3878 /* Always keep the time extend and data together */
3879 size = rb_event_ts_length(event);
3880
3881 if (len < size)
3882 goto out_unlock;
3883
3884 /* save the current timestamp, since the user will need it */
3885 save_timestamp = cpu_buffer->read_stamp;
3886
3887 /* Need to copy one event at a time */
3888 do {
3889 /* We need the size of one event, because
3890 * rb_advance_reader only advances by one event,
3891 * whereas rb_event_ts_length may include the size of
3892 * one or two events.
3893 * We have already ensured there's enough space if this
3894 * is a time extend. */
3895 size = rb_event_length(event);
3896 memcpy(bpage->data + pos, rpage->data + rpos, size);
3897
3898 len -= size;
3899
3900 rb_advance_reader(cpu_buffer);
3901 rpos = reader->read;
3902 pos += size;
3903
3904 if (rpos >= commit)
3905 break;
3906
3907 event = rb_reader_event(cpu_buffer);
3908 /* Always keep the time extend and data together */
3909 size = rb_event_ts_length(event);
3910 } while (len >= size);
3911
3912 /* update bpage */
3913 local_set(&bpage->commit, pos);
3914 bpage->time_stamp = save_timestamp;
3915
3916 /* we copied everything to the beginning */
3917 read = 0;
3918 } else {
3919 /* update the entry counter */
3920 cpu_buffer->read += rb_page_entries(reader);
3921
3922 /* swap the pages */
3923 rb_init_page(bpage);
3924 bpage = reader->page;
3925 reader->page = *data_page;
3926 local_set(&reader->write, 0);
3927 local_set(&reader->entries, 0);
3928 reader->read = 0;
3929 *data_page = bpage;
3930
3931 /*
3932 * Use the real_end for the data size,
3933 * This gives us a chance to store the lost events
3934 * on the page.
3935 */
3936 if (reader->real_end)
3937 local_set(&bpage->commit, reader->real_end);
3938 }
3939 ret = read;
3940
3941 cpu_buffer->lost_events = 0;
3942
3943 commit = local_read(&bpage->commit);
3944 /*
3945 * Set a flag in the commit field if we lost events
3946 */
3947 if (missed_events) {
3948 /* If there is room at the end of the page to save the
3949 * missed events, then record it there.
3950 */
3951 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
3952 memcpy(&bpage->data[commit], &missed_events,
3953 sizeof(missed_events));
3954 local_add(RB_MISSED_STORED, &bpage->commit);
3955 commit += sizeof(missed_events);
3956 }
3957 local_add(RB_MISSED_EVENTS, &bpage->commit);
3958 }
3959
3960 /*
3961 * This page may be off to user land. Zero it out here.
3962 */
3963 if (commit < BUF_PAGE_SIZE)
3964 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
3965
3966 out_unlock:
3967 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3968
3969 out:
3970 return ret;
3971}
3972EXPORT_SYMBOL_GPL(ring_buffer_read_page);
3973
3974#ifdef CONFIG_TRACING
3975static ssize_t
3976rb_simple_read(struct file *filp, char __user *ubuf,
3977 size_t cnt, loff_t *ppos)
3978{
3979 unsigned long *p = filp->private_data;
3980 char buf[64];
3981 int r;
3982
3983 if (test_bit(RB_BUFFERS_DISABLED_BIT, p))
3984 r = sprintf(buf, "permanently disabled\n");
3985 else
3986 r = sprintf(buf, "%d\n", test_bit(RB_BUFFERS_ON_BIT, p));
3987
3988 return simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
3989}
3990
3991static ssize_t
3992rb_simple_write(struct file *filp, const char __user *ubuf,
3993 size_t cnt, loff_t *ppos)
3994{
3995 unsigned long *p = filp->private_data;
3996 unsigned long val;
3997 int ret;
3998
3999 ret = kstrtoul_from_user(ubuf, cnt, 10, &val);
4000 if (ret)
4001 return ret;
4002
4003 if (val)
4004 set_bit(RB_BUFFERS_ON_BIT, p);
4005 else
4006 clear_bit(RB_BUFFERS_ON_BIT, p);
4007
4008 (*ppos)++;
4009
4010 return cnt;
4011}
4012
4013static const struct file_operations rb_simple_fops = {
4014 .open = tracing_open_generic,
4015 .read = rb_simple_read,
4016 .write = rb_simple_write,
4017 .llseek = default_llseek,
4018};
4019
4020
4021static __init int rb_init_debugfs(void)
4022{
4023 struct dentry *d_tracer;
4024
4025 d_tracer = tracing_init_dentry();
4026
4027 trace_create_file("tracing_on", 0644, d_tracer,
4028 &ring_buffer_flags, &rb_simple_fops);
4029
4030 return 0;
4031}
4032
4033fs_initcall(rb_init_debugfs);
4034#endif
4035
4036#ifdef CONFIG_HOTPLUG_CPU
4037static int rb_cpu_notify(struct notifier_block *self,
4038 unsigned long action, void *hcpu)
4039{
4040 struct ring_buffer *buffer =
4041 container_of(self, struct ring_buffer, cpu_notify);
4042 long cpu = (long)hcpu;
4043
4044 switch (action) {
4045 case CPU_UP_PREPARE:
4046 case CPU_UP_PREPARE_FROZEN:
4047 if (cpumask_test_cpu(cpu, buffer->cpumask))
4048 return NOTIFY_OK;
4049
4050 buffer->buffers[cpu] =
4051 rb_allocate_cpu_buffer(buffer, cpu);
4052 if (!buffer->buffers[cpu]) {
4053 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4054 cpu);
4055 return NOTIFY_OK;
4056 }
4057 smp_wmb();
4058 cpumask_set_cpu(cpu, buffer->cpumask);
4059 break;
4060 case CPU_DOWN_PREPARE:
4061 case CPU_DOWN_PREPARE_FROZEN:
4062 /*
4063 * Do nothing.
4064 * If we were to free the buffer, then the user would
4065 * lose any trace that was in the buffer.
4066 */
4067 break;
4068 default:
4069 break;
4070 }
4071 return NOTIFY_OK;
4072}
4073#endif