Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Generic ring buffer
4 *
5 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 */
7#include <linux/trace_events.h>
8#include <linux/ring_buffer.h>
9#include <linux/trace_clock.h>
10#include <linux/sched/clock.h>
11#include <linux/trace_seq.h>
12#include <linux/spinlock.h>
13#include <linux/irq_work.h>
14#include <linux/uaccess.h>
15#include <linux/hardirq.h>
16#include <linux/kthread.h> /* for self test */
17#include <linux/module.h>
18#include <linux/percpu.h>
19#include <linux/mutex.h>
20#include <linux/delay.h>
21#include <linux/slab.h>
22#include <linux/init.h>
23#include <linux/hash.h>
24#include <linux/list.h>
25#include <linux/cpu.h>
26#include <linux/oom.h>
27
28#include <asm/local.h>
29
30static void update_pages_handler(struct work_struct *work);
31
32/*
33 * The ring buffer header is special. We must manually up keep it.
34 */
35int ring_buffer_print_entry_header(struct trace_seq *s)
36{
37 trace_seq_puts(s, "# compressed entry header\n");
38 trace_seq_puts(s, "\ttype_len : 5 bits\n");
39 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
40 trace_seq_puts(s, "\tarray : 32 bits\n");
41 trace_seq_putc(s, '\n');
42 trace_seq_printf(s, "\tpadding : type == %d\n",
43 RINGBUF_TYPE_PADDING);
44 trace_seq_printf(s, "\ttime_extend : type == %d\n",
45 RINGBUF_TYPE_TIME_EXTEND);
46 trace_seq_printf(s, "\ttime_stamp : type == %d\n",
47 RINGBUF_TYPE_TIME_STAMP);
48 trace_seq_printf(s, "\tdata max type_len == %d\n",
49 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
50
51 return !trace_seq_has_overflowed(s);
52}
53
54/*
55 * The ring buffer is made up of a list of pages. A separate list of pages is
56 * allocated for each CPU. A writer may only write to a buffer that is
57 * associated with the CPU it is currently executing on. A reader may read
58 * from any per cpu buffer.
59 *
60 * The reader is special. For each per cpu buffer, the reader has its own
61 * reader page. When a reader has read the entire reader page, this reader
62 * page is swapped with another page in the ring buffer.
63 *
64 * Now, as long as the writer is off the reader page, the reader can do what
65 * ever it wants with that page. The writer will never write to that page
66 * again (as long as it is out of the ring buffer).
67 *
68 * Here's some silly ASCII art.
69 *
70 * +------+
71 * |reader| RING BUFFER
72 * |page |
73 * +------+ +---+ +---+ +---+
74 * | |-->| |-->| |
75 * +---+ +---+ +---+
76 * ^ |
77 * | |
78 * +---------------+
79 *
80 *
81 * +------+
82 * |reader| RING BUFFER
83 * |page |------------------v
84 * +------+ +---+ +---+ +---+
85 * | |-->| |-->| |
86 * +---+ +---+ +---+
87 * ^ |
88 * | |
89 * +---------------+
90 *
91 *
92 * +------+
93 * |reader| RING BUFFER
94 * |page |------------------v
95 * +------+ +---+ +---+ +---+
96 * ^ | |-->| |-->| |
97 * | +---+ +---+ +---+
98 * | |
99 * | |
100 * +------------------------------+
101 *
102 *
103 * +------+
104 * |buffer| RING BUFFER
105 * |page |------------------v
106 * +------+ +---+ +---+ +---+
107 * ^ | | | |-->| |
108 * | New +---+ +---+ +---+
109 * | Reader------^ |
110 * | page |
111 * +------------------------------+
112 *
113 *
114 * After we make this swap, the reader can hand this page off to the splice
115 * code and be done with it. It can even allocate a new page if it needs to
116 * and swap that into the ring buffer.
117 *
118 * We will be using cmpxchg soon to make all this lockless.
119 *
120 */
121
122/* Used for individual buffers (after the counter) */
123#define RB_BUFFER_OFF (1 << 20)
124
125#define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
126
127#define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
128#define RB_ALIGNMENT 4U
129#define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
130#define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
131#define RB_ALIGN_DATA __aligned(RB_ALIGNMENT)
132
133/* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
134#define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
135
136enum {
137 RB_LEN_TIME_EXTEND = 8,
138 RB_LEN_TIME_STAMP = 8,
139};
140
141#define skip_time_extend(event) \
142 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
143
144#define extended_time(event) \
145 (event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
146
147static inline int rb_null_event(struct ring_buffer_event *event)
148{
149 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
150}
151
152static void rb_event_set_padding(struct ring_buffer_event *event)
153{
154 /* padding has a NULL time_delta */
155 event->type_len = RINGBUF_TYPE_PADDING;
156 event->time_delta = 0;
157}
158
159static unsigned
160rb_event_data_length(struct ring_buffer_event *event)
161{
162 unsigned length;
163
164 if (event->type_len)
165 length = event->type_len * RB_ALIGNMENT;
166 else
167 length = event->array[0];
168 return length + RB_EVNT_HDR_SIZE;
169}
170
171/*
172 * Return the length of the given event. Will return
173 * the length of the time extend if the event is a
174 * time extend.
175 */
176static inline unsigned
177rb_event_length(struct ring_buffer_event *event)
178{
179 switch (event->type_len) {
180 case RINGBUF_TYPE_PADDING:
181 if (rb_null_event(event))
182 /* undefined */
183 return -1;
184 return event->array[0] + RB_EVNT_HDR_SIZE;
185
186 case RINGBUF_TYPE_TIME_EXTEND:
187 return RB_LEN_TIME_EXTEND;
188
189 case RINGBUF_TYPE_TIME_STAMP:
190 return RB_LEN_TIME_STAMP;
191
192 case RINGBUF_TYPE_DATA:
193 return rb_event_data_length(event);
194 default:
195 BUG();
196 }
197 /* not hit */
198 return 0;
199}
200
201/*
202 * Return total length of time extend and data,
203 * or just the event length for all other events.
204 */
205static inline unsigned
206rb_event_ts_length(struct ring_buffer_event *event)
207{
208 unsigned len = 0;
209
210 if (extended_time(event)) {
211 /* time extends include the data event after it */
212 len = RB_LEN_TIME_EXTEND;
213 event = skip_time_extend(event);
214 }
215 return len + rb_event_length(event);
216}
217
218/**
219 * ring_buffer_event_length - return the length of the event
220 * @event: the event to get the length of
221 *
222 * Returns the size of the data load of a data event.
223 * If the event is something other than a data event, it
224 * returns the size of the event itself. With the exception
225 * of a TIME EXTEND, where it still returns the size of the
226 * data load of the data event after it.
227 */
228unsigned ring_buffer_event_length(struct ring_buffer_event *event)
229{
230 unsigned length;
231
232 if (extended_time(event))
233 event = skip_time_extend(event);
234
235 length = rb_event_length(event);
236 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
237 return length;
238 length -= RB_EVNT_HDR_SIZE;
239 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
240 length -= sizeof(event->array[0]);
241 return length;
242}
243EXPORT_SYMBOL_GPL(ring_buffer_event_length);
244
245/* inline for ring buffer fast paths */
246static __always_inline void *
247rb_event_data(struct ring_buffer_event *event)
248{
249 if (extended_time(event))
250 event = skip_time_extend(event);
251 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
252 /* If length is in len field, then array[0] has the data */
253 if (event->type_len)
254 return (void *)&event->array[0];
255 /* Otherwise length is in array[0] and array[1] has the data */
256 return (void *)&event->array[1];
257}
258
259/**
260 * ring_buffer_event_data - return the data of the event
261 * @event: the event to get the data from
262 */
263void *ring_buffer_event_data(struct ring_buffer_event *event)
264{
265 return rb_event_data(event);
266}
267EXPORT_SYMBOL_GPL(ring_buffer_event_data);
268
269#define for_each_buffer_cpu(buffer, cpu) \
270 for_each_cpu(cpu, buffer->cpumask)
271
272#define TS_SHIFT 27
273#define TS_MASK ((1ULL << TS_SHIFT) - 1)
274#define TS_DELTA_TEST (~TS_MASK)
275
276/**
277 * ring_buffer_event_time_stamp - return the event's extended timestamp
278 * @event: the event to get the timestamp of
279 *
280 * Returns the extended timestamp associated with a data event.
281 * An extended time_stamp is a 64-bit timestamp represented
282 * internally in a special way that makes the best use of space
283 * contained within a ring buffer event. This function decodes
284 * it and maps it to a straight u64 value.
285 */
286u64 ring_buffer_event_time_stamp(struct ring_buffer_event *event)
287{
288 u64 ts;
289
290 ts = event->array[0];
291 ts <<= TS_SHIFT;
292 ts += event->time_delta;
293
294 return ts;
295}
296
297/* Flag when events were overwritten */
298#define RB_MISSED_EVENTS (1 << 31)
299/* Missed count stored at end */
300#define RB_MISSED_STORED (1 << 30)
301
302#define RB_MISSED_FLAGS (RB_MISSED_EVENTS|RB_MISSED_STORED)
303
304struct buffer_data_page {
305 u64 time_stamp; /* page time stamp */
306 local_t commit; /* write committed index */
307 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
308};
309
310/*
311 * Note, the buffer_page list must be first. The buffer pages
312 * are allocated in cache lines, which means that each buffer
313 * page will be at the beginning of a cache line, and thus
314 * the least significant bits will be zero. We use this to
315 * add flags in the list struct pointers, to make the ring buffer
316 * lockless.
317 */
318struct buffer_page {
319 struct list_head list; /* list of buffer pages */
320 local_t write; /* index for next write */
321 unsigned read; /* index for next read */
322 local_t entries; /* entries on this page */
323 unsigned long real_end; /* real end of data */
324 struct buffer_data_page *page; /* Actual data page */
325};
326
327/*
328 * The buffer page counters, write and entries, must be reset
329 * atomically when crossing page boundaries. To synchronize this
330 * update, two counters are inserted into the number. One is
331 * the actual counter for the write position or count on the page.
332 *
333 * The other is a counter of updaters. Before an update happens
334 * the update partition of the counter is incremented. This will
335 * allow the updater to update the counter atomically.
336 *
337 * The counter is 20 bits, and the state data is 12.
338 */
339#define RB_WRITE_MASK 0xfffff
340#define RB_WRITE_INTCNT (1 << 20)
341
342static void rb_init_page(struct buffer_data_page *bpage)
343{
344 local_set(&bpage->commit, 0);
345}
346
347/*
348 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
349 * this issue out.
350 */
351static void free_buffer_page(struct buffer_page *bpage)
352{
353 free_page((unsigned long)bpage->page);
354 kfree(bpage);
355}
356
357/*
358 * We need to fit the time_stamp delta into 27 bits.
359 */
360static inline int test_time_stamp(u64 delta)
361{
362 if (delta & TS_DELTA_TEST)
363 return 1;
364 return 0;
365}
366
367#define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
368
369/* Max payload is BUF_PAGE_SIZE - header (8bytes) */
370#define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
371
372int ring_buffer_print_page_header(struct trace_seq *s)
373{
374 struct buffer_data_page field;
375
376 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
377 "offset:0;\tsize:%u;\tsigned:%u;\n",
378 (unsigned int)sizeof(field.time_stamp),
379 (unsigned int)is_signed_type(u64));
380
381 trace_seq_printf(s, "\tfield: local_t commit;\t"
382 "offset:%u;\tsize:%u;\tsigned:%u;\n",
383 (unsigned int)offsetof(typeof(field), commit),
384 (unsigned int)sizeof(field.commit),
385 (unsigned int)is_signed_type(long));
386
387 trace_seq_printf(s, "\tfield: int overwrite;\t"
388 "offset:%u;\tsize:%u;\tsigned:%u;\n",
389 (unsigned int)offsetof(typeof(field), commit),
390 1,
391 (unsigned int)is_signed_type(long));
392
393 trace_seq_printf(s, "\tfield: char data;\t"
394 "offset:%u;\tsize:%u;\tsigned:%u;\n",
395 (unsigned int)offsetof(typeof(field), data),
396 (unsigned int)BUF_PAGE_SIZE,
397 (unsigned int)is_signed_type(char));
398
399 return !trace_seq_has_overflowed(s);
400}
401
402struct rb_irq_work {
403 struct irq_work work;
404 wait_queue_head_t waiters;
405 wait_queue_head_t full_waiters;
406 bool waiters_pending;
407 bool full_waiters_pending;
408 bool wakeup_full;
409};
410
411/*
412 * Structure to hold event state and handle nested events.
413 */
414struct rb_event_info {
415 u64 ts;
416 u64 delta;
417 unsigned long length;
418 struct buffer_page *tail_page;
419 int add_timestamp;
420};
421
422/*
423 * Used for which event context the event is in.
424 * NMI = 0
425 * IRQ = 1
426 * SOFTIRQ = 2
427 * NORMAL = 3
428 *
429 * See trace_recursive_lock() comment below for more details.
430 */
431enum {
432 RB_CTX_NMI,
433 RB_CTX_IRQ,
434 RB_CTX_SOFTIRQ,
435 RB_CTX_NORMAL,
436 RB_CTX_MAX
437};
438
439/*
440 * head_page == tail_page && head == tail then buffer is empty.
441 */
442struct ring_buffer_per_cpu {
443 int cpu;
444 atomic_t record_disabled;
445 struct ring_buffer *buffer;
446 raw_spinlock_t reader_lock; /* serialize readers */
447 arch_spinlock_t lock;
448 struct lock_class_key lock_key;
449 struct buffer_data_page *free_page;
450 unsigned long nr_pages;
451 unsigned int current_context;
452 struct list_head *pages;
453 struct buffer_page *head_page; /* read from head */
454 struct buffer_page *tail_page; /* write to tail */
455 struct buffer_page *commit_page; /* committed pages */
456 struct buffer_page *reader_page;
457 unsigned long lost_events;
458 unsigned long last_overrun;
459 unsigned long nest;
460 local_t entries_bytes;
461 local_t entries;
462 local_t overrun;
463 local_t commit_overrun;
464 local_t dropped_events;
465 local_t committing;
466 local_t commits;
467 local_t pages_touched;
468 local_t pages_read;
469 long last_pages_touch;
470 size_t shortest_full;
471 unsigned long read;
472 unsigned long read_bytes;
473 u64 write_stamp;
474 u64 read_stamp;
475 /* ring buffer pages to update, > 0 to add, < 0 to remove */
476 long nr_pages_to_update;
477 struct list_head new_pages; /* new pages to add */
478 struct work_struct update_pages_work;
479 struct completion update_done;
480
481 struct rb_irq_work irq_work;
482};
483
484struct ring_buffer {
485 unsigned flags;
486 int cpus;
487 atomic_t record_disabled;
488 atomic_t resize_disabled;
489 cpumask_var_t cpumask;
490
491 struct lock_class_key *reader_lock_key;
492
493 struct mutex mutex;
494
495 struct ring_buffer_per_cpu **buffers;
496
497 struct hlist_node node;
498 u64 (*clock)(void);
499
500 struct rb_irq_work irq_work;
501 bool time_stamp_abs;
502};
503
504struct ring_buffer_iter {
505 struct ring_buffer_per_cpu *cpu_buffer;
506 unsigned long head;
507 struct buffer_page *head_page;
508 struct buffer_page *cache_reader_page;
509 unsigned long cache_read;
510 u64 read_stamp;
511};
512
513/**
514 * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
515 * @buffer: The ring_buffer to get the number of pages from
516 * @cpu: The cpu of the ring_buffer to get the number of pages from
517 *
518 * Returns the number of pages used by a per_cpu buffer of the ring buffer.
519 */
520size_t ring_buffer_nr_pages(struct ring_buffer *buffer, int cpu)
521{
522 return buffer->buffers[cpu]->nr_pages;
523}
524
525/**
526 * ring_buffer_nr_pages_dirty - get the number of used pages in the ring buffer
527 * @buffer: The ring_buffer to get the number of pages from
528 * @cpu: The cpu of the ring_buffer to get the number of pages from
529 *
530 * Returns the number of pages that have content in the ring buffer.
531 */
532size_t ring_buffer_nr_dirty_pages(struct ring_buffer *buffer, int cpu)
533{
534 size_t read;
535 size_t cnt;
536
537 read = local_read(&buffer->buffers[cpu]->pages_read);
538 cnt = local_read(&buffer->buffers[cpu]->pages_touched);
539 /* The reader can read an empty page, but not more than that */
540 if (cnt < read) {
541 WARN_ON_ONCE(read > cnt + 1);
542 return 0;
543 }
544
545 return cnt - read;
546}
547
548/*
549 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
550 *
551 * Schedules a delayed work to wake up any task that is blocked on the
552 * ring buffer waiters queue.
553 */
554static void rb_wake_up_waiters(struct irq_work *work)
555{
556 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
557
558 wake_up_all(&rbwork->waiters);
559 if (rbwork->wakeup_full) {
560 rbwork->wakeup_full = false;
561 wake_up_all(&rbwork->full_waiters);
562 }
563}
564
565/**
566 * ring_buffer_wait - wait for input to the ring buffer
567 * @buffer: buffer to wait on
568 * @cpu: the cpu buffer to wait on
569 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
570 *
571 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
572 * as data is added to any of the @buffer's cpu buffers. Otherwise
573 * it will wait for data to be added to a specific cpu buffer.
574 */
575int ring_buffer_wait(struct ring_buffer *buffer, int cpu, int full)
576{
577 struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
578 DEFINE_WAIT(wait);
579 struct rb_irq_work *work;
580 int ret = 0;
581
582 /*
583 * Depending on what the caller is waiting for, either any
584 * data in any cpu buffer, or a specific buffer, put the
585 * caller on the appropriate wait queue.
586 */
587 if (cpu == RING_BUFFER_ALL_CPUS) {
588 work = &buffer->irq_work;
589 /* Full only makes sense on per cpu reads */
590 full = 0;
591 } else {
592 if (!cpumask_test_cpu(cpu, buffer->cpumask))
593 return -ENODEV;
594 cpu_buffer = buffer->buffers[cpu];
595 work = &cpu_buffer->irq_work;
596 }
597
598
599 while (true) {
600 if (full)
601 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
602 else
603 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
604
605 /*
606 * The events can happen in critical sections where
607 * checking a work queue can cause deadlocks.
608 * After adding a task to the queue, this flag is set
609 * only to notify events to try to wake up the queue
610 * using irq_work.
611 *
612 * We don't clear it even if the buffer is no longer
613 * empty. The flag only causes the next event to run
614 * irq_work to do the work queue wake up. The worse
615 * that can happen if we race with !trace_empty() is that
616 * an event will cause an irq_work to try to wake up
617 * an empty queue.
618 *
619 * There's no reason to protect this flag either, as
620 * the work queue and irq_work logic will do the necessary
621 * synchronization for the wake ups. The only thing
622 * that is necessary is that the wake up happens after
623 * a task has been queued. It's OK for spurious wake ups.
624 */
625 if (full)
626 work->full_waiters_pending = true;
627 else
628 work->waiters_pending = true;
629
630 if (signal_pending(current)) {
631 ret = -EINTR;
632 break;
633 }
634
635 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
636 break;
637
638 if (cpu != RING_BUFFER_ALL_CPUS &&
639 !ring_buffer_empty_cpu(buffer, cpu)) {
640 unsigned long flags;
641 bool pagebusy;
642 size_t nr_pages;
643 size_t dirty;
644
645 if (!full)
646 break;
647
648 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
649 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
650 nr_pages = cpu_buffer->nr_pages;
651 dirty = ring_buffer_nr_dirty_pages(buffer, cpu);
652 if (!cpu_buffer->shortest_full ||
653 cpu_buffer->shortest_full < full)
654 cpu_buffer->shortest_full = full;
655 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
656 if (!pagebusy &&
657 (!nr_pages || (dirty * 100) > full * nr_pages))
658 break;
659 }
660
661 schedule();
662 }
663
664 if (full)
665 finish_wait(&work->full_waiters, &wait);
666 else
667 finish_wait(&work->waiters, &wait);
668
669 return ret;
670}
671
672/**
673 * ring_buffer_poll_wait - poll on buffer input
674 * @buffer: buffer to wait on
675 * @cpu: the cpu buffer to wait on
676 * @filp: the file descriptor
677 * @poll_table: The poll descriptor
678 *
679 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
680 * as data is added to any of the @buffer's cpu buffers. Otherwise
681 * it will wait for data to be added to a specific cpu buffer.
682 *
683 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
684 * zero otherwise.
685 */
686__poll_t ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
687 struct file *filp, poll_table *poll_table)
688{
689 struct ring_buffer_per_cpu *cpu_buffer;
690 struct rb_irq_work *work;
691
692 if (cpu == RING_BUFFER_ALL_CPUS)
693 work = &buffer->irq_work;
694 else {
695 if (!cpumask_test_cpu(cpu, buffer->cpumask))
696 return -EINVAL;
697
698 cpu_buffer = buffer->buffers[cpu];
699 work = &cpu_buffer->irq_work;
700 }
701
702 poll_wait(filp, &work->waiters, poll_table);
703 work->waiters_pending = true;
704 /*
705 * There's a tight race between setting the waiters_pending and
706 * checking if the ring buffer is empty. Once the waiters_pending bit
707 * is set, the next event will wake the task up, but we can get stuck
708 * if there's only a single event in.
709 *
710 * FIXME: Ideally, we need a memory barrier on the writer side as well,
711 * but adding a memory barrier to all events will cause too much of a
712 * performance hit in the fast path. We only need a memory barrier when
713 * the buffer goes from empty to having content. But as this race is
714 * extremely small, and it's not a problem if another event comes in, we
715 * will fix it later.
716 */
717 smp_mb();
718
719 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
720 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
721 return EPOLLIN | EPOLLRDNORM;
722 return 0;
723}
724
725/* buffer may be either ring_buffer or ring_buffer_per_cpu */
726#define RB_WARN_ON(b, cond) \
727 ({ \
728 int _____ret = unlikely(cond); \
729 if (_____ret) { \
730 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
731 struct ring_buffer_per_cpu *__b = \
732 (void *)b; \
733 atomic_inc(&__b->buffer->record_disabled); \
734 } else \
735 atomic_inc(&b->record_disabled); \
736 WARN_ON(1); \
737 } \
738 _____ret; \
739 })
740
741/* Up this if you want to test the TIME_EXTENTS and normalization */
742#define DEBUG_SHIFT 0
743
744static inline u64 rb_time_stamp(struct ring_buffer *buffer)
745{
746 /* shift to debug/test normalization and TIME_EXTENTS */
747 return buffer->clock() << DEBUG_SHIFT;
748}
749
750u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
751{
752 u64 time;
753
754 preempt_disable_notrace();
755 time = rb_time_stamp(buffer);
756 preempt_enable_notrace();
757
758 return time;
759}
760EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
761
762void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
763 int cpu, u64 *ts)
764{
765 /* Just stupid testing the normalize function and deltas */
766 *ts >>= DEBUG_SHIFT;
767}
768EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
769
770/*
771 * Making the ring buffer lockless makes things tricky.
772 * Although writes only happen on the CPU that they are on,
773 * and they only need to worry about interrupts. Reads can
774 * happen on any CPU.
775 *
776 * The reader page is always off the ring buffer, but when the
777 * reader finishes with a page, it needs to swap its page with
778 * a new one from the buffer. The reader needs to take from
779 * the head (writes go to the tail). But if a writer is in overwrite
780 * mode and wraps, it must push the head page forward.
781 *
782 * Here lies the problem.
783 *
784 * The reader must be careful to replace only the head page, and
785 * not another one. As described at the top of the file in the
786 * ASCII art, the reader sets its old page to point to the next
787 * page after head. It then sets the page after head to point to
788 * the old reader page. But if the writer moves the head page
789 * during this operation, the reader could end up with the tail.
790 *
791 * We use cmpxchg to help prevent this race. We also do something
792 * special with the page before head. We set the LSB to 1.
793 *
794 * When the writer must push the page forward, it will clear the
795 * bit that points to the head page, move the head, and then set
796 * the bit that points to the new head page.
797 *
798 * We also don't want an interrupt coming in and moving the head
799 * page on another writer. Thus we use the second LSB to catch
800 * that too. Thus:
801 *
802 * head->list->prev->next bit 1 bit 0
803 * ------- -------
804 * Normal page 0 0
805 * Points to head page 0 1
806 * New head page 1 0
807 *
808 * Note we can not trust the prev pointer of the head page, because:
809 *
810 * +----+ +-----+ +-----+
811 * | |------>| T |---X--->| N |
812 * | |<------| | | |
813 * +----+ +-----+ +-----+
814 * ^ ^ |
815 * | +-----+ | |
816 * +----------| R |----------+ |
817 * | |<-----------+
818 * +-----+
819 *
820 * Key: ---X--> HEAD flag set in pointer
821 * T Tail page
822 * R Reader page
823 * N Next page
824 *
825 * (see __rb_reserve_next() to see where this happens)
826 *
827 * What the above shows is that the reader just swapped out
828 * the reader page with a page in the buffer, but before it
829 * could make the new header point back to the new page added
830 * it was preempted by a writer. The writer moved forward onto
831 * the new page added by the reader and is about to move forward
832 * again.
833 *
834 * You can see, it is legitimate for the previous pointer of
835 * the head (or any page) not to point back to itself. But only
836 * temporarily.
837 */
838
839#define RB_PAGE_NORMAL 0UL
840#define RB_PAGE_HEAD 1UL
841#define RB_PAGE_UPDATE 2UL
842
843
844#define RB_FLAG_MASK 3UL
845
846/* PAGE_MOVED is not part of the mask */
847#define RB_PAGE_MOVED 4UL
848
849/*
850 * rb_list_head - remove any bit
851 */
852static struct list_head *rb_list_head(struct list_head *list)
853{
854 unsigned long val = (unsigned long)list;
855
856 return (struct list_head *)(val & ~RB_FLAG_MASK);
857}
858
859/*
860 * rb_is_head_page - test if the given page is the head page
861 *
862 * Because the reader may move the head_page pointer, we can
863 * not trust what the head page is (it may be pointing to
864 * the reader page). But if the next page is a header page,
865 * its flags will be non zero.
866 */
867static inline int
868rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
869 struct buffer_page *page, struct list_head *list)
870{
871 unsigned long val;
872
873 val = (unsigned long)list->next;
874
875 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
876 return RB_PAGE_MOVED;
877
878 return val & RB_FLAG_MASK;
879}
880
881/*
882 * rb_is_reader_page
883 *
884 * The unique thing about the reader page, is that, if the
885 * writer is ever on it, the previous pointer never points
886 * back to the reader page.
887 */
888static bool rb_is_reader_page(struct buffer_page *page)
889{
890 struct list_head *list = page->list.prev;
891
892 return rb_list_head(list->next) != &page->list;
893}
894
895/*
896 * rb_set_list_to_head - set a list_head to be pointing to head.
897 */
898static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
899 struct list_head *list)
900{
901 unsigned long *ptr;
902
903 ptr = (unsigned long *)&list->next;
904 *ptr |= RB_PAGE_HEAD;
905 *ptr &= ~RB_PAGE_UPDATE;
906}
907
908/*
909 * rb_head_page_activate - sets up head page
910 */
911static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
912{
913 struct buffer_page *head;
914
915 head = cpu_buffer->head_page;
916 if (!head)
917 return;
918
919 /*
920 * Set the previous list pointer to have the HEAD flag.
921 */
922 rb_set_list_to_head(cpu_buffer, head->list.prev);
923}
924
925static void rb_list_head_clear(struct list_head *list)
926{
927 unsigned long *ptr = (unsigned long *)&list->next;
928
929 *ptr &= ~RB_FLAG_MASK;
930}
931
932/*
933 * rb_head_page_deactivate - clears head page ptr (for free list)
934 */
935static void
936rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
937{
938 struct list_head *hd;
939
940 /* Go through the whole list and clear any pointers found. */
941 rb_list_head_clear(cpu_buffer->pages);
942
943 list_for_each(hd, cpu_buffer->pages)
944 rb_list_head_clear(hd);
945}
946
947static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
948 struct buffer_page *head,
949 struct buffer_page *prev,
950 int old_flag, int new_flag)
951{
952 struct list_head *list;
953 unsigned long val = (unsigned long)&head->list;
954 unsigned long ret;
955
956 list = &prev->list;
957
958 val &= ~RB_FLAG_MASK;
959
960 ret = cmpxchg((unsigned long *)&list->next,
961 val | old_flag, val | new_flag);
962
963 /* check if the reader took the page */
964 if ((ret & ~RB_FLAG_MASK) != val)
965 return RB_PAGE_MOVED;
966
967 return ret & RB_FLAG_MASK;
968}
969
970static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
971 struct buffer_page *head,
972 struct buffer_page *prev,
973 int old_flag)
974{
975 return rb_head_page_set(cpu_buffer, head, prev,
976 old_flag, RB_PAGE_UPDATE);
977}
978
979static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
980 struct buffer_page *head,
981 struct buffer_page *prev,
982 int old_flag)
983{
984 return rb_head_page_set(cpu_buffer, head, prev,
985 old_flag, RB_PAGE_HEAD);
986}
987
988static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
989 struct buffer_page *head,
990 struct buffer_page *prev,
991 int old_flag)
992{
993 return rb_head_page_set(cpu_buffer, head, prev,
994 old_flag, RB_PAGE_NORMAL);
995}
996
997static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
998 struct buffer_page **bpage)
999{
1000 struct list_head *p = rb_list_head((*bpage)->list.next);
1001
1002 *bpage = list_entry(p, struct buffer_page, list);
1003}
1004
1005static struct buffer_page *
1006rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1007{
1008 struct buffer_page *head;
1009 struct buffer_page *page;
1010 struct list_head *list;
1011 int i;
1012
1013 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1014 return NULL;
1015
1016 /* sanity check */
1017 list = cpu_buffer->pages;
1018 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1019 return NULL;
1020
1021 page = head = cpu_buffer->head_page;
1022 /*
1023 * It is possible that the writer moves the header behind
1024 * where we started, and we miss in one loop.
1025 * A second loop should grab the header, but we'll do
1026 * three loops just because I'm paranoid.
1027 */
1028 for (i = 0; i < 3; i++) {
1029 do {
1030 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
1031 cpu_buffer->head_page = page;
1032 return page;
1033 }
1034 rb_inc_page(cpu_buffer, &page);
1035 } while (page != head);
1036 }
1037
1038 RB_WARN_ON(cpu_buffer, 1);
1039
1040 return NULL;
1041}
1042
1043static int rb_head_page_replace(struct buffer_page *old,
1044 struct buffer_page *new)
1045{
1046 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1047 unsigned long val;
1048 unsigned long ret;
1049
1050 val = *ptr & ~RB_FLAG_MASK;
1051 val |= RB_PAGE_HEAD;
1052
1053 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1054
1055 return ret == val;
1056}
1057
1058/*
1059 * rb_tail_page_update - move the tail page forward
1060 */
1061static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1062 struct buffer_page *tail_page,
1063 struct buffer_page *next_page)
1064{
1065 unsigned long old_entries;
1066 unsigned long old_write;
1067
1068 /*
1069 * The tail page now needs to be moved forward.
1070 *
1071 * We need to reset the tail page, but without messing
1072 * with possible erasing of data brought in by interrupts
1073 * that have moved the tail page and are currently on it.
1074 *
1075 * We add a counter to the write field to denote this.
1076 */
1077 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1078 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1079
1080 local_inc(&cpu_buffer->pages_touched);
1081 /*
1082 * Just make sure we have seen our old_write and synchronize
1083 * with any interrupts that come in.
1084 */
1085 barrier();
1086
1087 /*
1088 * If the tail page is still the same as what we think
1089 * it is, then it is up to us to update the tail
1090 * pointer.
1091 */
1092 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1093 /* Zero the write counter */
1094 unsigned long val = old_write & ~RB_WRITE_MASK;
1095 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1096
1097 /*
1098 * This will only succeed if an interrupt did
1099 * not come in and change it. In which case, we
1100 * do not want to modify it.
1101 *
1102 * We add (void) to let the compiler know that we do not care
1103 * about the return value of these functions. We use the
1104 * cmpxchg to only update if an interrupt did not already
1105 * do it for us. If the cmpxchg fails, we don't care.
1106 */
1107 (void)local_cmpxchg(&next_page->write, old_write, val);
1108 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1109
1110 /*
1111 * No need to worry about races with clearing out the commit.
1112 * it only can increment when a commit takes place. But that
1113 * only happens in the outer most nested commit.
1114 */
1115 local_set(&next_page->page->commit, 0);
1116
1117 /* Again, either we update tail_page or an interrupt does */
1118 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1119 }
1120}
1121
1122static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1123 struct buffer_page *bpage)
1124{
1125 unsigned long val = (unsigned long)bpage;
1126
1127 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1128 return 1;
1129
1130 return 0;
1131}
1132
1133/**
1134 * rb_check_list - make sure a pointer to a list has the last bits zero
1135 */
1136static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1137 struct list_head *list)
1138{
1139 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1140 return 1;
1141 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1142 return 1;
1143 return 0;
1144}
1145
1146/**
1147 * rb_check_pages - integrity check of buffer pages
1148 * @cpu_buffer: CPU buffer with pages to test
1149 *
1150 * As a safety measure we check to make sure the data pages have not
1151 * been corrupted.
1152 */
1153static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1154{
1155 struct list_head *head = cpu_buffer->pages;
1156 struct buffer_page *bpage, *tmp;
1157
1158 /* Reset the head page if it exists */
1159 if (cpu_buffer->head_page)
1160 rb_set_head_page(cpu_buffer);
1161
1162 rb_head_page_deactivate(cpu_buffer);
1163
1164 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1165 return -1;
1166 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1167 return -1;
1168
1169 if (rb_check_list(cpu_buffer, head))
1170 return -1;
1171
1172 list_for_each_entry_safe(bpage, tmp, head, list) {
1173 if (RB_WARN_ON(cpu_buffer,
1174 bpage->list.next->prev != &bpage->list))
1175 return -1;
1176 if (RB_WARN_ON(cpu_buffer,
1177 bpage->list.prev->next != &bpage->list))
1178 return -1;
1179 if (rb_check_list(cpu_buffer, &bpage->list))
1180 return -1;
1181 }
1182
1183 rb_head_page_activate(cpu_buffer);
1184
1185 return 0;
1186}
1187
1188static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1189{
1190 struct buffer_page *bpage, *tmp;
1191 bool user_thread = current->mm != NULL;
1192 gfp_t mflags;
1193 long i;
1194
1195 /*
1196 * Check if the available memory is there first.
1197 * Note, si_mem_available() only gives us a rough estimate of available
1198 * memory. It may not be accurate. But we don't care, we just want
1199 * to prevent doing any allocation when it is obvious that it is
1200 * not going to succeed.
1201 */
1202 i = si_mem_available();
1203 if (i < nr_pages)
1204 return -ENOMEM;
1205
1206 /*
1207 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1208 * gracefully without invoking oom-killer and the system is not
1209 * destabilized.
1210 */
1211 mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1212
1213 /*
1214 * If a user thread allocates too much, and si_mem_available()
1215 * reports there's enough memory, even though there is not.
1216 * Make sure the OOM killer kills this thread. This can happen
1217 * even with RETRY_MAYFAIL because another task may be doing
1218 * an allocation after this task has taken all memory.
1219 * This is the task the OOM killer needs to take out during this
1220 * loop, even if it was triggered by an allocation somewhere else.
1221 */
1222 if (user_thread)
1223 set_current_oom_origin();
1224 for (i = 0; i < nr_pages; i++) {
1225 struct page *page;
1226
1227 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1228 mflags, cpu_to_node(cpu));
1229 if (!bpage)
1230 goto free_pages;
1231
1232 list_add(&bpage->list, pages);
1233
1234 page = alloc_pages_node(cpu_to_node(cpu), mflags, 0);
1235 if (!page)
1236 goto free_pages;
1237 bpage->page = page_address(page);
1238 rb_init_page(bpage->page);
1239
1240 if (user_thread && fatal_signal_pending(current))
1241 goto free_pages;
1242 }
1243 if (user_thread)
1244 clear_current_oom_origin();
1245
1246 return 0;
1247
1248free_pages:
1249 list_for_each_entry_safe(bpage, tmp, pages, list) {
1250 list_del_init(&bpage->list);
1251 free_buffer_page(bpage);
1252 }
1253 if (user_thread)
1254 clear_current_oom_origin();
1255
1256 return -ENOMEM;
1257}
1258
1259static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1260 unsigned long nr_pages)
1261{
1262 LIST_HEAD(pages);
1263
1264 WARN_ON(!nr_pages);
1265
1266 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1267 return -ENOMEM;
1268
1269 /*
1270 * The ring buffer page list is a circular list that does not
1271 * start and end with a list head. All page list items point to
1272 * other pages.
1273 */
1274 cpu_buffer->pages = pages.next;
1275 list_del(&pages);
1276
1277 cpu_buffer->nr_pages = nr_pages;
1278
1279 rb_check_pages(cpu_buffer);
1280
1281 return 0;
1282}
1283
1284static struct ring_buffer_per_cpu *
1285rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1286{
1287 struct ring_buffer_per_cpu *cpu_buffer;
1288 struct buffer_page *bpage;
1289 struct page *page;
1290 int ret;
1291
1292 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1293 GFP_KERNEL, cpu_to_node(cpu));
1294 if (!cpu_buffer)
1295 return NULL;
1296
1297 cpu_buffer->cpu = cpu;
1298 cpu_buffer->buffer = buffer;
1299 raw_spin_lock_init(&cpu_buffer->reader_lock);
1300 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1301 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1302 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1303 init_completion(&cpu_buffer->update_done);
1304 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1305 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1306 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1307
1308 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1309 GFP_KERNEL, cpu_to_node(cpu));
1310 if (!bpage)
1311 goto fail_free_buffer;
1312
1313 rb_check_bpage(cpu_buffer, bpage);
1314
1315 cpu_buffer->reader_page = bpage;
1316 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1317 if (!page)
1318 goto fail_free_reader;
1319 bpage->page = page_address(page);
1320 rb_init_page(bpage->page);
1321
1322 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1323 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1324
1325 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1326 if (ret < 0)
1327 goto fail_free_reader;
1328
1329 cpu_buffer->head_page
1330 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1331 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1332
1333 rb_head_page_activate(cpu_buffer);
1334
1335 return cpu_buffer;
1336
1337 fail_free_reader:
1338 free_buffer_page(cpu_buffer->reader_page);
1339
1340 fail_free_buffer:
1341 kfree(cpu_buffer);
1342 return NULL;
1343}
1344
1345static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1346{
1347 struct list_head *head = cpu_buffer->pages;
1348 struct buffer_page *bpage, *tmp;
1349
1350 free_buffer_page(cpu_buffer->reader_page);
1351
1352 rb_head_page_deactivate(cpu_buffer);
1353
1354 if (head) {
1355 list_for_each_entry_safe(bpage, tmp, head, list) {
1356 list_del_init(&bpage->list);
1357 free_buffer_page(bpage);
1358 }
1359 bpage = list_entry(head, struct buffer_page, list);
1360 free_buffer_page(bpage);
1361 }
1362
1363 kfree(cpu_buffer);
1364}
1365
1366/**
1367 * __ring_buffer_alloc - allocate a new ring_buffer
1368 * @size: the size in bytes per cpu that is needed.
1369 * @flags: attributes to set for the ring buffer.
1370 *
1371 * Currently the only flag that is available is the RB_FL_OVERWRITE
1372 * flag. This flag means that the buffer will overwrite old data
1373 * when the buffer wraps. If this flag is not set, the buffer will
1374 * drop data when the tail hits the head.
1375 */
1376struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1377 struct lock_class_key *key)
1378{
1379 struct ring_buffer *buffer;
1380 long nr_pages;
1381 int bsize;
1382 int cpu;
1383 int ret;
1384
1385 /* keep it in its own cache line */
1386 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1387 GFP_KERNEL);
1388 if (!buffer)
1389 return NULL;
1390
1391 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1392 goto fail_free_buffer;
1393
1394 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1395 buffer->flags = flags;
1396 buffer->clock = trace_clock_local;
1397 buffer->reader_lock_key = key;
1398
1399 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1400 init_waitqueue_head(&buffer->irq_work.waiters);
1401
1402 /* need at least two pages */
1403 if (nr_pages < 2)
1404 nr_pages = 2;
1405
1406 buffer->cpus = nr_cpu_ids;
1407
1408 bsize = sizeof(void *) * nr_cpu_ids;
1409 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1410 GFP_KERNEL);
1411 if (!buffer->buffers)
1412 goto fail_free_cpumask;
1413
1414 cpu = raw_smp_processor_id();
1415 cpumask_set_cpu(cpu, buffer->cpumask);
1416 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1417 if (!buffer->buffers[cpu])
1418 goto fail_free_buffers;
1419
1420 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1421 if (ret < 0)
1422 goto fail_free_buffers;
1423
1424 mutex_init(&buffer->mutex);
1425
1426 return buffer;
1427
1428 fail_free_buffers:
1429 for_each_buffer_cpu(buffer, cpu) {
1430 if (buffer->buffers[cpu])
1431 rb_free_cpu_buffer(buffer->buffers[cpu]);
1432 }
1433 kfree(buffer->buffers);
1434
1435 fail_free_cpumask:
1436 free_cpumask_var(buffer->cpumask);
1437
1438 fail_free_buffer:
1439 kfree(buffer);
1440 return NULL;
1441}
1442EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1443
1444/**
1445 * ring_buffer_free - free a ring buffer.
1446 * @buffer: the buffer to free.
1447 */
1448void
1449ring_buffer_free(struct ring_buffer *buffer)
1450{
1451 int cpu;
1452
1453 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1454
1455 for_each_buffer_cpu(buffer, cpu)
1456 rb_free_cpu_buffer(buffer->buffers[cpu]);
1457
1458 kfree(buffer->buffers);
1459 free_cpumask_var(buffer->cpumask);
1460
1461 kfree(buffer);
1462}
1463EXPORT_SYMBOL_GPL(ring_buffer_free);
1464
1465void ring_buffer_set_clock(struct ring_buffer *buffer,
1466 u64 (*clock)(void))
1467{
1468 buffer->clock = clock;
1469}
1470
1471void ring_buffer_set_time_stamp_abs(struct ring_buffer *buffer, bool abs)
1472{
1473 buffer->time_stamp_abs = abs;
1474}
1475
1476bool ring_buffer_time_stamp_abs(struct ring_buffer *buffer)
1477{
1478 return buffer->time_stamp_abs;
1479}
1480
1481static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1482
1483static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1484{
1485 return local_read(&bpage->entries) & RB_WRITE_MASK;
1486}
1487
1488static inline unsigned long rb_page_write(struct buffer_page *bpage)
1489{
1490 return local_read(&bpage->write) & RB_WRITE_MASK;
1491}
1492
1493static int
1494rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1495{
1496 struct list_head *tail_page, *to_remove, *next_page;
1497 struct buffer_page *to_remove_page, *tmp_iter_page;
1498 struct buffer_page *last_page, *first_page;
1499 unsigned long nr_removed;
1500 unsigned long head_bit;
1501 int page_entries;
1502
1503 head_bit = 0;
1504
1505 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1506 atomic_inc(&cpu_buffer->record_disabled);
1507 /*
1508 * We don't race with the readers since we have acquired the reader
1509 * lock. We also don't race with writers after disabling recording.
1510 * This makes it easy to figure out the first and the last page to be
1511 * removed from the list. We unlink all the pages in between including
1512 * the first and last pages. This is done in a busy loop so that we
1513 * lose the least number of traces.
1514 * The pages are freed after we restart recording and unlock readers.
1515 */
1516 tail_page = &cpu_buffer->tail_page->list;
1517
1518 /*
1519 * tail page might be on reader page, we remove the next page
1520 * from the ring buffer
1521 */
1522 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1523 tail_page = rb_list_head(tail_page->next);
1524 to_remove = tail_page;
1525
1526 /* start of pages to remove */
1527 first_page = list_entry(rb_list_head(to_remove->next),
1528 struct buffer_page, list);
1529
1530 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1531 to_remove = rb_list_head(to_remove)->next;
1532 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1533 }
1534
1535 next_page = rb_list_head(to_remove)->next;
1536
1537 /*
1538 * Now we remove all pages between tail_page and next_page.
1539 * Make sure that we have head_bit value preserved for the
1540 * next page
1541 */
1542 tail_page->next = (struct list_head *)((unsigned long)next_page |
1543 head_bit);
1544 next_page = rb_list_head(next_page);
1545 next_page->prev = tail_page;
1546
1547 /* make sure pages points to a valid page in the ring buffer */
1548 cpu_buffer->pages = next_page;
1549
1550 /* update head page */
1551 if (head_bit)
1552 cpu_buffer->head_page = list_entry(next_page,
1553 struct buffer_page, list);
1554
1555 /*
1556 * change read pointer to make sure any read iterators reset
1557 * themselves
1558 */
1559 cpu_buffer->read = 0;
1560
1561 /* pages are removed, resume tracing and then free the pages */
1562 atomic_dec(&cpu_buffer->record_disabled);
1563 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1564
1565 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1566
1567 /* last buffer page to remove */
1568 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1569 list);
1570 tmp_iter_page = first_page;
1571
1572 do {
1573 cond_resched();
1574
1575 to_remove_page = tmp_iter_page;
1576 rb_inc_page(cpu_buffer, &tmp_iter_page);
1577
1578 /* update the counters */
1579 page_entries = rb_page_entries(to_remove_page);
1580 if (page_entries) {
1581 /*
1582 * If something was added to this page, it was full
1583 * since it is not the tail page. So we deduct the
1584 * bytes consumed in ring buffer from here.
1585 * Increment overrun to account for the lost events.
1586 */
1587 local_add(page_entries, &cpu_buffer->overrun);
1588 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1589 }
1590
1591 /*
1592 * We have already removed references to this list item, just
1593 * free up the buffer_page and its page
1594 */
1595 free_buffer_page(to_remove_page);
1596 nr_removed--;
1597
1598 } while (to_remove_page != last_page);
1599
1600 RB_WARN_ON(cpu_buffer, nr_removed);
1601
1602 return nr_removed == 0;
1603}
1604
1605static int
1606rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1607{
1608 struct list_head *pages = &cpu_buffer->new_pages;
1609 int retries, success;
1610
1611 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1612 /*
1613 * We are holding the reader lock, so the reader page won't be swapped
1614 * in the ring buffer. Now we are racing with the writer trying to
1615 * move head page and the tail page.
1616 * We are going to adapt the reader page update process where:
1617 * 1. We first splice the start and end of list of new pages between
1618 * the head page and its previous page.
1619 * 2. We cmpxchg the prev_page->next to point from head page to the
1620 * start of new pages list.
1621 * 3. Finally, we update the head->prev to the end of new list.
1622 *
1623 * We will try this process 10 times, to make sure that we don't keep
1624 * spinning.
1625 */
1626 retries = 10;
1627 success = 0;
1628 while (retries--) {
1629 struct list_head *head_page, *prev_page, *r;
1630 struct list_head *last_page, *first_page;
1631 struct list_head *head_page_with_bit;
1632
1633 head_page = &rb_set_head_page(cpu_buffer)->list;
1634 if (!head_page)
1635 break;
1636 prev_page = head_page->prev;
1637
1638 first_page = pages->next;
1639 last_page = pages->prev;
1640
1641 head_page_with_bit = (struct list_head *)
1642 ((unsigned long)head_page | RB_PAGE_HEAD);
1643
1644 last_page->next = head_page_with_bit;
1645 first_page->prev = prev_page;
1646
1647 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1648
1649 if (r == head_page_with_bit) {
1650 /*
1651 * yay, we replaced the page pointer to our new list,
1652 * now, we just have to update to head page's prev
1653 * pointer to point to end of list
1654 */
1655 head_page->prev = last_page;
1656 success = 1;
1657 break;
1658 }
1659 }
1660
1661 if (success)
1662 INIT_LIST_HEAD(pages);
1663 /*
1664 * If we weren't successful in adding in new pages, warn and stop
1665 * tracing
1666 */
1667 RB_WARN_ON(cpu_buffer, !success);
1668 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1669
1670 /* free pages if they weren't inserted */
1671 if (!success) {
1672 struct buffer_page *bpage, *tmp;
1673 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1674 list) {
1675 list_del_init(&bpage->list);
1676 free_buffer_page(bpage);
1677 }
1678 }
1679 return success;
1680}
1681
1682static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1683{
1684 int success;
1685
1686 if (cpu_buffer->nr_pages_to_update > 0)
1687 success = rb_insert_pages(cpu_buffer);
1688 else
1689 success = rb_remove_pages(cpu_buffer,
1690 -cpu_buffer->nr_pages_to_update);
1691
1692 if (success)
1693 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1694}
1695
1696static void update_pages_handler(struct work_struct *work)
1697{
1698 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1699 struct ring_buffer_per_cpu, update_pages_work);
1700 rb_update_pages(cpu_buffer);
1701 complete(&cpu_buffer->update_done);
1702}
1703
1704/**
1705 * ring_buffer_resize - resize the ring buffer
1706 * @buffer: the buffer to resize.
1707 * @size: the new size.
1708 * @cpu_id: the cpu buffer to resize
1709 *
1710 * Minimum size is 2 * BUF_PAGE_SIZE.
1711 *
1712 * Returns 0 on success and < 0 on failure.
1713 */
1714int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1715 int cpu_id)
1716{
1717 struct ring_buffer_per_cpu *cpu_buffer;
1718 unsigned long nr_pages;
1719 int cpu, err = 0;
1720
1721 /*
1722 * Always succeed at resizing a non-existent buffer:
1723 */
1724 if (!buffer)
1725 return size;
1726
1727 /* Make sure the requested buffer exists */
1728 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1729 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1730 return size;
1731
1732 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1733
1734 /* we need a minimum of two pages */
1735 if (nr_pages < 2)
1736 nr_pages = 2;
1737
1738 size = nr_pages * BUF_PAGE_SIZE;
1739
1740 /*
1741 * Don't succeed if resizing is disabled, as a reader might be
1742 * manipulating the ring buffer and is expecting a sane state while
1743 * this is true.
1744 */
1745 if (atomic_read(&buffer->resize_disabled))
1746 return -EBUSY;
1747
1748 /* prevent another thread from changing buffer sizes */
1749 mutex_lock(&buffer->mutex);
1750
1751 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1752 /* calculate the pages to update */
1753 for_each_buffer_cpu(buffer, cpu) {
1754 cpu_buffer = buffer->buffers[cpu];
1755
1756 cpu_buffer->nr_pages_to_update = nr_pages -
1757 cpu_buffer->nr_pages;
1758 /*
1759 * nothing more to do for removing pages or no update
1760 */
1761 if (cpu_buffer->nr_pages_to_update <= 0)
1762 continue;
1763 /*
1764 * to add pages, make sure all new pages can be
1765 * allocated without receiving ENOMEM
1766 */
1767 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1768 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1769 &cpu_buffer->new_pages, cpu)) {
1770 /* not enough memory for new pages */
1771 err = -ENOMEM;
1772 goto out_err;
1773 }
1774 }
1775
1776 get_online_cpus();
1777 /*
1778 * Fire off all the required work handlers
1779 * We can't schedule on offline CPUs, but it's not necessary
1780 * since we can change their buffer sizes without any race.
1781 */
1782 for_each_buffer_cpu(buffer, cpu) {
1783 cpu_buffer = buffer->buffers[cpu];
1784 if (!cpu_buffer->nr_pages_to_update)
1785 continue;
1786
1787 /* Can't run something on an offline CPU. */
1788 if (!cpu_online(cpu)) {
1789 rb_update_pages(cpu_buffer);
1790 cpu_buffer->nr_pages_to_update = 0;
1791 } else {
1792 schedule_work_on(cpu,
1793 &cpu_buffer->update_pages_work);
1794 }
1795 }
1796
1797 /* wait for all the updates to complete */
1798 for_each_buffer_cpu(buffer, cpu) {
1799 cpu_buffer = buffer->buffers[cpu];
1800 if (!cpu_buffer->nr_pages_to_update)
1801 continue;
1802
1803 if (cpu_online(cpu))
1804 wait_for_completion(&cpu_buffer->update_done);
1805 cpu_buffer->nr_pages_to_update = 0;
1806 }
1807
1808 put_online_cpus();
1809 } else {
1810 /* Make sure this CPU has been initialized */
1811 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1812 goto out;
1813
1814 cpu_buffer = buffer->buffers[cpu_id];
1815
1816 if (nr_pages == cpu_buffer->nr_pages)
1817 goto out;
1818
1819 cpu_buffer->nr_pages_to_update = nr_pages -
1820 cpu_buffer->nr_pages;
1821
1822 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1823 if (cpu_buffer->nr_pages_to_update > 0 &&
1824 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1825 &cpu_buffer->new_pages, cpu_id)) {
1826 err = -ENOMEM;
1827 goto out_err;
1828 }
1829
1830 get_online_cpus();
1831
1832 /* Can't run something on an offline CPU. */
1833 if (!cpu_online(cpu_id))
1834 rb_update_pages(cpu_buffer);
1835 else {
1836 schedule_work_on(cpu_id,
1837 &cpu_buffer->update_pages_work);
1838 wait_for_completion(&cpu_buffer->update_done);
1839 }
1840
1841 cpu_buffer->nr_pages_to_update = 0;
1842 put_online_cpus();
1843 }
1844
1845 out:
1846 /*
1847 * The ring buffer resize can happen with the ring buffer
1848 * enabled, so that the update disturbs the tracing as little
1849 * as possible. But if the buffer is disabled, we do not need
1850 * to worry about that, and we can take the time to verify
1851 * that the buffer is not corrupt.
1852 */
1853 if (atomic_read(&buffer->record_disabled)) {
1854 atomic_inc(&buffer->record_disabled);
1855 /*
1856 * Even though the buffer was disabled, we must make sure
1857 * that it is truly disabled before calling rb_check_pages.
1858 * There could have been a race between checking
1859 * record_disable and incrementing it.
1860 */
1861 synchronize_rcu();
1862 for_each_buffer_cpu(buffer, cpu) {
1863 cpu_buffer = buffer->buffers[cpu];
1864 rb_check_pages(cpu_buffer);
1865 }
1866 atomic_dec(&buffer->record_disabled);
1867 }
1868
1869 mutex_unlock(&buffer->mutex);
1870 return size;
1871
1872 out_err:
1873 for_each_buffer_cpu(buffer, cpu) {
1874 struct buffer_page *bpage, *tmp;
1875
1876 cpu_buffer = buffer->buffers[cpu];
1877 cpu_buffer->nr_pages_to_update = 0;
1878
1879 if (list_empty(&cpu_buffer->new_pages))
1880 continue;
1881
1882 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1883 list) {
1884 list_del_init(&bpage->list);
1885 free_buffer_page(bpage);
1886 }
1887 }
1888 mutex_unlock(&buffer->mutex);
1889 return err;
1890}
1891EXPORT_SYMBOL_GPL(ring_buffer_resize);
1892
1893void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1894{
1895 mutex_lock(&buffer->mutex);
1896 if (val)
1897 buffer->flags |= RB_FL_OVERWRITE;
1898 else
1899 buffer->flags &= ~RB_FL_OVERWRITE;
1900 mutex_unlock(&buffer->mutex);
1901}
1902EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1903
1904static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1905{
1906 return bpage->page->data + index;
1907}
1908
1909static __always_inline struct ring_buffer_event *
1910rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1911{
1912 return __rb_page_index(cpu_buffer->reader_page,
1913 cpu_buffer->reader_page->read);
1914}
1915
1916static __always_inline struct ring_buffer_event *
1917rb_iter_head_event(struct ring_buffer_iter *iter)
1918{
1919 return __rb_page_index(iter->head_page, iter->head);
1920}
1921
1922static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
1923{
1924 return local_read(&bpage->page->commit);
1925}
1926
1927/* Size is determined by what has been committed */
1928static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
1929{
1930 return rb_page_commit(bpage);
1931}
1932
1933static __always_inline unsigned
1934rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1935{
1936 return rb_page_commit(cpu_buffer->commit_page);
1937}
1938
1939static __always_inline unsigned
1940rb_event_index(struct ring_buffer_event *event)
1941{
1942 unsigned long addr = (unsigned long)event;
1943
1944 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1945}
1946
1947static void rb_inc_iter(struct ring_buffer_iter *iter)
1948{
1949 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1950
1951 /*
1952 * The iterator could be on the reader page (it starts there).
1953 * But the head could have moved, since the reader was
1954 * found. Check for this case and assign the iterator
1955 * to the head page instead of next.
1956 */
1957 if (iter->head_page == cpu_buffer->reader_page)
1958 iter->head_page = rb_set_head_page(cpu_buffer);
1959 else
1960 rb_inc_page(cpu_buffer, &iter->head_page);
1961
1962 iter->read_stamp = iter->head_page->page->time_stamp;
1963 iter->head = 0;
1964}
1965
1966/*
1967 * rb_handle_head_page - writer hit the head page
1968 *
1969 * Returns: +1 to retry page
1970 * 0 to continue
1971 * -1 on error
1972 */
1973static int
1974rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1975 struct buffer_page *tail_page,
1976 struct buffer_page *next_page)
1977{
1978 struct buffer_page *new_head;
1979 int entries;
1980 int type;
1981 int ret;
1982
1983 entries = rb_page_entries(next_page);
1984
1985 /*
1986 * The hard part is here. We need to move the head
1987 * forward, and protect against both readers on
1988 * other CPUs and writers coming in via interrupts.
1989 */
1990 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1991 RB_PAGE_HEAD);
1992
1993 /*
1994 * type can be one of four:
1995 * NORMAL - an interrupt already moved it for us
1996 * HEAD - we are the first to get here.
1997 * UPDATE - we are the interrupt interrupting
1998 * a current move.
1999 * MOVED - a reader on another CPU moved the next
2000 * pointer to its reader page. Give up
2001 * and try again.
2002 */
2003
2004 switch (type) {
2005 case RB_PAGE_HEAD:
2006 /*
2007 * We changed the head to UPDATE, thus
2008 * it is our responsibility to update
2009 * the counters.
2010 */
2011 local_add(entries, &cpu_buffer->overrun);
2012 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2013
2014 /*
2015 * The entries will be zeroed out when we move the
2016 * tail page.
2017 */
2018
2019 /* still more to do */
2020 break;
2021
2022 case RB_PAGE_UPDATE:
2023 /*
2024 * This is an interrupt that interrupt the
2025 * previous update. Still more to do.
2026 */
2027 break;
2028 case RB_PAGE_NORMAL:
2029 /*
2030 * An interrupt came in before the update
2031 * and processed this for us.
2032 * Nothing left to do.
2033 */
2034 return 1;
2035 case RB_PAGE_MOVED:
2036 /*
2037 * The reader is on another CPU and just did
2038 * a swap with our next_page.
2039 * Try again.
2040 */
2041 return 1;
2042 default:
2043 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2044 return -1;
2045 }
2046
2047 /*
2048 * Now that we are here, the old head pointer is
2049 * set to UPDATE. This will keep the reader from
2050 * swapping the head page with the reader page.
2051 * The reader (on another CPU) will spin till
2052 * we are finished.
2053 *
2054 * We just need to protect against interrupts
2055 * doing the job. We will set the next pointer
2056 * to HEAD. After that, we set the old pointer
2057 * to NORMAL, but only if it was HEAD before.
2058 * otherwise we are an interrupt, and only
2059 * want the outer most commit to reset it.
2060 */
2061 new_head = next_page;
2062 rb_inc_page(cpu_buffer, &new_head);
2063
2064 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2065 RB_PAGE_NORMAL);
2066
2067 /*
2068 * Valid returns are:
2069 * HEAD - an interrupt came in and already set it.
2070 * NORMAL - One of two things:
2071 * 1) We really set it.
2072 * 2) A bunch of interrupts came in and moved
2073 * the page forward again.
2074 */
2075 switch (ret) {
2076 case RB_PAGE_HEAD:
2077 case RB_PAGE_NORMAL:
2078 /* OK */
2079 break;
2080 default:
2081 RB_WARN_ON(cpu_buffer, 1);
2082 return -1;
2083 }
2084
2085 /*
2086 * It is possible that an interrupt came in,
2087 * set the head up, then more interrupts came in
2088 * and moved it again. When we get back here,
2089 * the page would have been set to NORMAL but we
2090 * just set it back to HEAD.
2091 *
2092 * How do you detect this? Well, if that happened
2093 * the tail page would have moved.
2094 */
2095 if (ret == RB_PAGE_NORMAL) {
2096 struct buffer_page *buffer_tail_page;
2097
2098 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2099 /*
2100 * If the tail had moved passed next, then we need
2101 * to reset the pointer.
2102 */
2103 if (buffer_tail_page != tail_page &&
2104 buffer_tail_page != next_page)
2105 rb_head_page_set_normal(cpu_buffer, new_head,
2106 next_page,
2107 RB_PAGE_HEAD);
2108 }
2109
2110 /*
2111 * If this was the outer most commit (the one that
2112 * changed the original pointer from HEAD to UPDATE),
2113 * then it is up to us to reset it to NORMAL.
2114 */
2115 if (type == RB_PAGE_HEAD) {
2116 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2117 tail_page,
2118 RB_PAGE_UPDATE);
2119 if (RB_WARN_ON(cpu_buffer,
2120 ret != RB_PAGE_UPDATE))
2121 return -1;
2122 }
2123
2124 return 0;
2125}
2126
2127static inline void
2128rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2129 unsigned long tail, struct rb_event_info *info)
2130{
2131 struct buffer_page *tail_page = info->tail_page;
2132 struct ring_buffer_event *event;
2133 unsigned long length = info->length;
2134
2135 /*
2136 * Only the event that crossed the page boundary
2137 * must fill the old tail_page with padding.
2138 */
2139 if (tail >= BUF_PAGE_SIZE) {
2140 /*
2141 * If the page was filled, then we still need
2142 * to update the real_end. Reset it to zero
2143 * and the reader will ignore it.
2144 */
2145 if (tail == BUF_PAGE_SIZE)
2146 tail_page->real_end = 0;
2147
2148 local_sub(length, &tail_page->write);
2149 return;
2150 }
2151
2152 event = __rb_page_index(tail_page, tail);
2153
2154 /* account for padding bytes */
2155 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2156
2157 /*
2158 * Save the original length to the meta data.
2159 * This will be used by the reader to add lost event
2160 * counter.
2161 */
2162 tail_page->real_end = tail;
2163
2164 /*
2165 * If this event is bigger than the minimum size, then
2166 * we need to be careful that we don't subtract the
2167 * write counter enough to allow another writer to slip
2168 * in on this page.
2169 * We put in a discarded commit instead, to make sure
2170 * that this space is not used again.
2171 *
2172 * If we are less than the minimum size, we don't need to
2173 * worry about it.
2174 */
2175 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2176 /* No room for any events */
2177
2178 /* Mark the rest of the page with padding */
2179 rb_event_set_padding(event);
2180
2181 /* Set the write back to the previous setting */
2182 local_sub(length, &tail_page->write);
2183 return;
2184 }
2185
2186 /* Put in a discarded event */
2187 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2188 event->type_len = RINGBUF_TYPE_PADDING;
2189 /* time delta must be non zero */
2190 event->time_delta = 1;
2191
2192 /* Set write to end of buffer */
2193 length = (tail + length) - BUF_PAGE_SIZE;
2194 local_sub(length, &tail_page->write);
2195}
2196
2197static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2198
2199/*
2200 * This is the slow path, force gcc not to inline it.
2201 */
2202static noinline struct ring_buffer_event *
2203rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2204 unsigned long tail, struct rb_event_info *info)
2205{
2206 struct buffer_page *tail_page = info->tail_page;
2207 struct buffer_page *commit_page = cpu_buffer->commit_page;
2208 struct ring_buffer *buffer = cpu_buffer->buffer;
2209 struct buffer_page *next_page;
2210 int ret;
2211
2212 next_page = tail_page;
2213
2214 rb_inc_page(cpu_buffer, &next_page);
2215
2216 /*
2217 * If for some reason, we had an interrupt storm that made
2218 * it all the way around the buffer, bail, and warn
2219 * about it.
2220 */
2221 if (unlikely(next_page == commit_page)) {
2222 local_inc(&cpu_buffer->commit_overrun);
2223 goto out_reset;
2224 }
2225
2226 /*
2227 * This is where the fun begins!
2228 *
2229 * We are fighting against races between a reader that
2230 * could be on another CPU trying to swap its reader
2231 * page with the buffer head.
2232 *
2233 * We are also fighting against interrupts coming in and
2234 * moving the head or tail on us as well.
2235 *
2236 * If the next page is the head page then we have filled
2237 * the buffer, unless the commit page is still on the
2238 * reader page.
2239 */
2240 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2241
2242 /*
2243 * If the commit is not on the reader page, then
2244 * move the header page.
2245 */
2246 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2247 /*
2248 * If we are not in overwrite mode,
2249 * this is easy, just stop here.
2250 */
2251 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2252 local_inc(&cpu_buffer->dropped_events);
2253 goto out_reset;
2254 }
2255
2256 ret = rb_handle_head_page(cpu_buffer,
2257 tail_page,
2258 next_page);
2259 if (ret < 0)
2260 goto out_reset;
2261 if (ret)
2262 goto out_again;
2263 } else {
2264 /*
2265 * We need to be careful here too. The
2266 * commit page could still be on the reader
2267 * page. We could have a small buffer, and
2268 * have filled up the buffer with events
2269 * from interrupts and such, and wrapped.
2270 *
2271 * Note, if the tail page is also the on the
2272 * reader_page, we let it move out.
2273 */
2274 if (unlikely((cpu_buffer->commit_page !=
2275 cpu_buffer->tail_page) &&
2276 (cpu_buffer->commit_page ==
2277 cpu_buffer->reader_page))) {
2278 local_inc(&cpu_buffer->commit_overrun);
2279 goto out_reset;
2280 }
2281 }
2282 }
2283
2284 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2285
2286 out_again:
2287
2288 rb_reset_tail(cpu_buffer, tail, info);
2289
2290 /* Commit what we have for now. */
2291 rb_end_commit(cpu_buffer);
2292 /* rb_end_commit() decs committing */
2293 local_inc(&cpu_buffer->committing);
2294
2295 /* fail and let the caller try again */
2296 return ERR_PTR(-EAGAIN);
2297
2298 out_reset:
2299 /* reset write */
2300 rb_reset_tail(cpu_buffer, tail, info);
2301
2302 return NULL;
2303}
2304
2305/* Slow path, do not inline */
2306static noinline struct ring_buffer_event *
2307rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2308{
2309 if (abs)
2310 event->type_len = RINGBUF_TYPE_TIME_STAMP;
2311 else
2312 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2313
2314 /* Not the first event on the page, or not delta? */
2315 if (abs || rb_event_index(event)) {
2316 event->time_delta = delta & TS_MASK;
2317 event->array[0] = delta >> TS_SHIFT;
2318 } else {
2319 /* nope, just zero it */
2320 event->time_delta = 0;
2321 event->array[0] = 0;
2322 }
2323
2324 return skip_time_extend(event);
2325}
2326
2327static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2328 struct ring_buffer_event *event);
2329
2330/**
2331 * rb_update_event - update event type and data
2332 * @event: the event to update
2333 * @type: the type of event
2334 * @length: the size of the event field in the ring buffer
2335 *
2336 * Update the type and data fields of the event. The length
2337 * is the actual size that is written to the ring buffer,
2338 * and with this, we can determine what to place into the
2339 * data field.
2340 */
2341static void
2342rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2343 struct ring_buffer_event *event,
2344 struct rb_event_info *info)
2345{
2346 unsigned length = info->length;
2347 u64 delta = info->delta;
2348
2349 /* Only a commit updates the timestamp */
2350 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2351 delta = 0;
2352
2353 /*
2354 * If we need to add a timestamp, then we
2355 * add it to the start of the reserved space.
2356 */
2357 if (unlikely(info->add_timestamp)) {
2358 bool abs = ring_buffer_time_stamp_abs(cpu_buffer->buffer);
2359
2360 event = rb_add_time_stamp(event, info->delta, abs);
2361 length -= RB_LEN_TIME_EXTEND;
2362 delta = 0;
2363 }
2364
2365 event->time_delta = delta;
2366 length -= RB_EVNT_HDR_SIZE;
2367 if (length > RB_MAX_SMALL_DATA) {
2368 event->type_len = 0;
2369 event->array[0] = length;
2370 } else
2371 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2372}
2373
2374static unsigned rb_calculate_event_length(unsigned length)
2375{
2376 struct ring_buffer_event event; /* Used only for sizeof array */
2377
2378 /* zero length can cause confusions */
2379 if (!length)
2380 length++;
2381
2382 if (length > RB_MAX_SMALL_DATA)
2383 length += sizeof(event.array[0]);
2384
2385 length += RB_EVNT_HDR_SIZE;
2386 length = ALIGN(length, RB_ALIGNMENT);
2387
2388 /*
2389 * In case the time delta is larger than the 27 bits for it
2390 * in the header, we need to add a timestamp. If another
2391 * event comes in when trying to discard this one to increase
2392 * the length, then the timestamp will be added in the allocated
2393 * space of this event. If length is bigger than the size needed
2394 * for the TIME_EXTEND, then padding has to be used. The events
2395 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2396 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2397 * As length is a multiple of 4, we only need to worry if it
2398 * is 12 (RB_LEN_TIME_EXTEND + 4).
2399 */
2400 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2401 length += RB_ALIGNMENT;
2402
2403 return length;
2404}
2405
2406#ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2407static inline bool sched_clock_stable(void)
2408{
2409 return true;
2410}
2411#endif
2412
2413static inline int
2414rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2415 struct ring_buffer_event *event)
2416{
2417 unsigned long new_index, old_index;
2418 struct buffer_page *bpage;
2419 unsigned long index;
2420 unsigned long addr;
2421
2422 new_index = rb_event_index(event);
2423 old_index = new_index + rb_event_ts_length(event);
2424 addr = (unsigned long)event;
2425 addr &= PAGE_MASK;
2426
2427 bpage = READ_ONCE(cpu_buffer->tail_page);
2428
2429 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2430 unsigned long write_mask =
2431 local_read(&bpage->write) & ~RB_WRITE_MASK;
2432 unsigned long event_length = rb_event_length(event);
2433 /*
2434 * This is on the tail page. It is possible that
2435 * a write could come in and move the tail page
2436 * and write to the next page. That is fine
2437 * because we just shorten what is on this page.
2438 */
2439 old_index += write_mask;
2440 new_index += write_mask;
2441 index = local_cmpxchg(&bpage->write, old_index, new_index);
2442 if (index == old_index) {
2443 /* update counters */
2444 local_sub(event_length, &cpu_buffer->entries_bytes);
2445 return 1;
2446 }
2447 }
2448
2449 /* could not discard */
2450 return 0;
2451}
2452
2453static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2454{
2455 local_inc(&cpu_buffer->committing);
2456 local_inc(&cpu_buffer->commits);
2457}
2458
2459static __always_inline void
2460rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2461{
2462 unsigned long max_count;
2463
2464 /*
2465 * We only race with interrupts and NMIs on this CPU.
2466 * If we own the commit event, then we can commit
2467 * all others that interrupted us, since the interruptions
2468 * are in stack format (they finish before they come
2469 * back to us). This allows us to do a simple loop to
2470 * assign the commit to the tail.
2471 */
2472 again:
2473 max_count = cpu_buffer->nr_pages * 100;
2474
2475 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2476 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2477 return;
2478 if (RB_WARN_ON(cpu_buffer,
2479 rb_is_reader_page(cpu_buffer->tail_page)))
2480 return;
2481 local_set(&cpu_buffer->commit_page->page->commit,
2482 rb_page_write(cpu_buffer->commit_page));
2483 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2484 /* Only update the write stamp if the page has an event */
2485 if (rb_page_write(cpu_buffer->commit_page))
2486 cpu_buffer->write_stamp =
2487 cpu_buffer->commit_page->page->time_stamp;
2488 /* add barrier to keep gcc from optimizing too much */
2489 barrier();
2490 }
2491 while (rb_commit_index(cpu_buffer) !=
2492 rb_page_write(cpu_buffer->commit_page)) {
2493
2494 local_set(&cpu_buffer->commit_page->page->commit,
2495 rb_page_write(cpu_buffer->commit_page));
2496 RB_WARN_ON(cpu_buffer,
2497 local_read(&cpu_buffer->commit_page->page->commit) &
2498 ~RB_WRITE_MASK);
2499 barrier();
2500 }
2501
2502 /* again, keep gcc from optimizing */
2503 barrier();
2504
2505 /*
2506 * If an interrupt came in just after the first while loop
2507 * and pushed the tail page forward, we will be left with
2508 * a dangling commit that will never go forward.
2509 */
2510 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2511 goto again;
2512}
2513
2514static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2515{
2516 unsigned long commits;
2517
2518 if (RB_WARN_ON(cpu_buffer,
2519 !local_read(&cpu_buffer->committing)))
2520 return;
2521
2522 again:
2523 commits = local_read(&cpu_buffer->commits);
2524 /* synchronize with interrupts */
2525 barrier();
2526 if (local_read(&cpu_buffer->committing) == 1)
2527 rb_set_commit_to_write(cpu_buffer);
2528
2529 local_dec(&cpu_buffer->committing);
2530
2531 /* synchronize with interrupts */
2532 barrier();
2533
2534 /*
2535 * Need to account for interrupts coming in between the
2536 * updating of the commit page and the clearing of the
2537 * committing counter.
2538 */
2539 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2540 !local_read(&cpu_buffer->committing)) {
2541 local_inc(&cpu_buffer->committing);
2542 goto again;
2543 }
2544}
2545
2546static inline void rb_event_discard(struct ring_buffer_event *event)
2547{
2548 if (extended_time(event))
2549 event = skip_time_extend(event);
2550
2551 /* array[0] holds the actual length for the discarded event */
2552 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2553 event->type_len = RINGBUF_TYPE_PADDING;
2554 /* time delta must be non zero */
2555 if (!event->time_delta)
2556 event->time_delta = 1;
2557}
2558
2559static __always_inline bool
2560rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2561 struct ring_buffer_event *event)
2562{
2563 unsigned long addr = (unsigned long)event;
2564 unsigned long index;
2565
2566 index = rb_event_index(event);
2567 addr &= PAGE_MASK;
2568
2569 return cpu_buffer->commit_page->page == (void *)addr &&
2570 rb_commit_index(cpu_buffer) == index;
2571}
2572
2573static __always_inline void
2574rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2575 struct ring_buffer_event *event)
2576{
2577 u64 delta;
2578
2579 /*
2580 * The event first in the commit queue updates the
2581 * time stamp.
2582 */
2583 if (rb_event_is_commit(cpu_buffer, event)) {
2584 /*
2585 * A commit event that is first on a page
2586 * updates the write timestamp with the page stamp
2587 */
2588 if (!rb_event_index(event))
2589 cpu_buffer->write_stamp =
2590 cpu_buffer->commit_page->page->time_stamp;
2591 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2592 delta = ring_buffer_event_time_stamp(event);
2593 cpu_buffer->write_stamp += delta;
2594 } else if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
2595 delta = ring_buffer_event_time_stamp(event);
2596 cpu_buffer->write_stamp = delta;
2597 } else
2598 cpu_buffer->write_stamp += event->time_delta;
2599 }
2600}
2601
2602static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2603 struct ring_buffer_event *event)
2604{
2605 local_inc(&cpu_buffer->entries);
2606 rb_update_write_stamp(cpu_buffer, event);
2607 rb_end_commit(cpu_buffer);
2608}
2609
2610static __always_inline void
2611rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2612{
2613 size_t nr_pages;
2614 size_t dirty;
2615 size_t full;
2616
2617 if (buffer->irq_work.waiters_pending) {
2618 buffer->irq_work.waiters_pending = false;
2619 /* irq_work_queue() supplies it's own memory barriers */
2620 irq_work_queue(&buffer->irq_work.work);
2621 }
2622
2623 if (cpu_buffer->irq_work.waiters_pending) {
2624 cpu_buffer->irq_work.waiters_pending = false;
2625 /* irq_work_queue() supplies it's own memory barriers */
2626 irq_work_queue(&cpu_buffer->irq_work.work);
2627 }
2628
2629 if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
2630 return;
2631
2632 if (cpu_buffer->reader_page == cpu_buffer->commit_page)
2633 return;
2634
2635 if (!cpu_buffer->irq_work.full_waiters_pending)
2636 return;
2637
2638 cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
2639
2640 full = cpu_buffer->shortest_full;
2641 nr_pages = cpu_buffer->nr_pages;
2642 dirty = ring_buffer_nr_dirty_pages(buffer, cpu_buffer->cpu);
2643 if (full && nr_pages && (dirty * 100) <= full * nr_pages)
2644 return;
2645
2646 cpu_buffer->irq_work.wakeup_full = true;
2647 cpu_buffer->irq_work.full_waiters_pending = false;
2648 /* irq_work_queue() supplies it's own memory barriers */
2649 irq_work_queue(&cpu_buffer->irq_work.work);
2650}
2651
2652/*
2653 * The lock and unlock are done within a preempt disable section.
2654 * The current_context per_cpu variable can only be modified
2655 * by the current task between lock and unlock. But it can
2656 * be modified more than once via an interrupt. To pass this
2657 * information from the lock to the unlock without having to
2658 * access the 'in_interrupt()' functions again (which do show
2659 * a bit of overhead in something as critical as function tracing,
2660 * we use a bitmask trick.
2661 *
2662 * bit 0 = NMI context
2663 * bit 1 = IRQ context
2664 * bit 2 = SoftIRQ context
2665 * bit 3 = normal context.
2666 *
2667 * This works because this is the order of contexts that can
2668 * preempt other contexts. A SoftIRQ never preempts an IRQ
2669 * context.
2670 *
2671 * When the context is determined, the corresponding bit is
2672 * checked and set (if it was set, then a recursion of that context
2673 * happened).
2674 *
2675 * On unlock, we need to clear this bit. To do so, just subtract
2676 * 1 from the current_context and AND it to itself.
2677 *
2678 * (binary)
2679 * 101 - 1 = 100
2680 * 101 & 100 = 100 (clearing bit zero)
2681 *
2682 * 1010 - 1 = 1001
2683 * 1010 & 1001 = 1000 (clearing bit 1)
2684 *
2685 * The least significant bit can be cleared this way, and it
2686 * just so happens that it is the same bit corresponding to
2687 * the current context.
2688 */
2689
2690static __always_inline int
2691trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2692{
2693 unsigned int val = cpu_buffer->current_context;
2694 unsigned long pc = preempt_count();
2695 int bit;
2696
2697 if (!(pc & (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_OFFSET)))
2698 bit = RB_CTX_NORMAL;
2699 else
2700 bit = pc & NMI_MASK ? RB_CTX_NMI :
2701 pc & HARDIRQ_MASK ? RB_CTX_IRQ : RB_CTX_SOFTIRQ;
2702
2703 if (unlikely(val & (1 << (bit + cpu_buffer->nest))))
2704 return 1;
2705
2706 val |= (1 << (bit + cpu_buffer->nest));
2707 cpu_buffer->current_context = val;
2708
2709 return 0;
2710}
2711
2712static __always_inline void
2713trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2714{
2715 cpu_buffer->current_context &=
2716 cpu_buffer->current_context - (1 << cpu_buffer->nest);
2717}
2718
2719/* The recursive locking above uses 4 bits */
2720#define NESTED_BITS 4
2721
2722/**
2723 * ring_buffer_nest_start - Allow to trace while nested
2724 * @buffer: The ring buffer to modify
2725 *
2726 * The ring buffer has a safety mechanism to prevent recursion.
2727 * But there may be a case where a trace needs to be done while
2728 * tracing something else. In this case, calling this function
2729 * will allow this function to nest within a currently active
2730 * ring_buffer_lock_reserve().
2731 *
2732 * Call this function before calling another ring_buffer_lock_reserve() and
2733 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
2734 */
2735void ring_buffer_nest_start(struct ring_buffer *buffer)
2736{
2737 struct ring_buffer_per_cpu *cpu_buffer;
2738 int cpu;
2739
2740 /* Enabled by ring_buffer_nest_end() */
2741 preempt_disable_notrace();
2742 cpu = raw_smp_processor_id();
2743 cpu_buffer = buffer->buffers[cpu];
2744 /* This is the shift value for the above recursive locking */
2745 cpu_buffer->nest += NESTED_BITS;
2746}
2747
2748/**
2749 * ring_buffer_nest_end - Allow to trace while nested
2750 * @buffer: The ring buffer to modify
2751 *
2752 * Must be called after ring_buffer_nest_start() and after the
2753 * ring_buffer_unlock_commit().
2754 */
2755void ring_buffer_nest_end(struct ring_buffer *buffer)
2756{
2757 struct ring_buffer_per_cpu *cpu_buffer;
2758 int cpu;
2759
2760 /* disabled by ring_buffer_nest_start() */
2761 cpu = raw_smp_processor_id();
2762 cpu_buffer = buffer->buffers[cpu];
2763 /* This is the shift value for the above recursive locking */
2764 cpu_buffer->nest -= NESTED_BITS;
2765 preempt_enable_notrace();
2766}
2767
2768/**
2769 * ring_buffer_unlock_commit - commit a reserved
2770 * @buffer: The buffer to commit to
2771 * @event: The event pointer to commit.
2772 *
2773 * This commits the data to the ring buffer, and releases any locks held.
2774 *
2775 * Must be paired with ring_buffer_lock_reserve.
2776 */
2777int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2778 struct ring_buffer_event *event)
2779{
2780 struct ring_buffer_per_cpu *cpu_buffer;
2781 int cpu = raw_smp_processor_id();
2782
2783 cpu_buffer = buffer->buffers[cpu];
2784
2785 rb_commit(cpu_buffer, event);
2786
2787 rb_wakeups(buffer, cpu_buffer);
2788
2789 trace_recursive_unlock(cpu_buffer);
2790
2791 preempt_enable_notrace();
2792
2793 return 0;
2794}
2795EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2796
2797static noinline void
2798rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2799 struct rb_event_info *info)
2800{
2801 WARN_ONCE(info->delta > (1ULL << 59),
2802 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2803 (unsigned long long)info->delta,
2804 (unsigned long long)info->ts,
2805 (unsigned long long)cpu_buffer->write_stamp,
2806 sched_clock_stable() ? "" :
2807 "If you just came from a suspend/resume,\n"
2808 "please switch to the trace global clock:\n"
2809 " echo global > /sys/kernel/debug/tracing/trace_clock\n"
2810 "or add trace_clock=global to the kernel command line\n");
2811 info->add_timestamp = 1;
2812}
2813
2814static struct ring_buffer_event *
2815__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2816 struct rb_event_info *info)
2817{
2818 struct ring_buffer_event *event;
2819 struct buffer_page *tail_page;
2820 unsigned long tail, write;
2821
2822 /*
2823 * If the time delta since the last event is too big to
2824 * hold in the time field of the event, then we append a
2825 * TIME EXTEND event ahead of the data event.
2826 */
2827 if (unlikely(info->add_timestamp))
2828 info->length += RB_LEN_TIME_EXTEND;
2829
2830 /* Don't let the compiler play games with cpu_buffer->tail_page */
2831 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2832 write = local_add_return(info->length, &tail_page->write);
2833
2834 /* set write to only the index of the write */
2835 write &= RB_WRITE_MASK;
2836 tail = write - info->length;
2837
2838 /*
2839 * If this is the first commit on the page, then it has the same
2840 * timestamp as the page itself.
2841 */
2842 if (!tail && !ring_buffer_time_stamp_abs(cpu_buffer->buffer))
2843 info->delta = 0;
2844
2845 /* See if we shot pass the end of this buffer page */
2846 if (unlikely(write > BUF_PAGE_SIZE))
2847 return rb_move_tail(cpu_buffer, tail, info);
2848
2849 /* We reserved something on the buffer */
2850
2851 event = __rb_page_index(tail_page, tail);
2852 rb_update_event(cpu_buffer, event, info);
2853
2854 local_inc(&tail_page->entries);
2855
2856 /*
2857 * If this is the first commit on the page, then update
2858 * its timestamp.
2859 */
2860 if (!tail)
2861 tail_page->page->time_stamp = info->ts;
2862
2863 /* account for these added bytes */
2864 local_add(info->length, &cpu_buffer->entries_bytes);
2865
2866 return event;
2867}
2868
2869static __always_inline struct ring_buffer_event *
2870rb_reserve_next_event(struct ring_buffer *buffer,
2871 struct ring_buffer_per_cpu *cpu_buffer,
2872 unsigned long length)
2873{
2874 struct ring_buffer_event *event;
2875 struct rb_event_info info;
2876 int nr_loops = 0;
2877 u64 diff;
2878
2879 rb_start_commit(cpu_buffer);
2880
2881#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2882 /*
2883 * Due to the ability to swap a cpu buffer from a buffer
2884 * it is possible it was swapped before we committed.
2885 * (committing stops a swap). We check for it here and
2886 * if it happened, we have to fail the write.
2887 */
2888 barrier();
2889 if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
2890 local_dec(&cpu_buffer->committing);
2891 local_dec(&cpu_buffer->commits);
2892 return NULL;
2893 }
2894#endif
2895
2896 info.length = rb_calculate_event_length(length);
2897 again:
2898 info.add_timestamp = 0;
2899 info.delta = 0;
2900
2901 /*
2902 * We allow for interrupts to reenter here and do a trace.
2903 * If one does, it will cause this original code to loop
2904 * back here. Even with heavy interrupts happening, this
2905 * should only happen a few times in a row. If this happens
2906 * 1000 times in a row, there must be either an interrupt
2907 * storm or we have something buggy.
2908 * Bail!
2909 */
2910 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2911 goto out_fail;
2912
2913 info.ts = rb_time_stamp(cpu_buffer->buffer);
2914 diff = info.ts - cpu_buffer->write_stamp;
2915
2916 /* make sure this diff is calculated here */
2917 barrier();
2918
2919 if (ring_buffer_time_stamp_abs(buffer)) {
2920 info.delta = info.ts;
2921 rb_handle_timestamp(cpu_buffer, &info);
2922 } else /* Did the write stamp get updated already? */
2923 if (likely(info.ts >= cpu_buffer->write_stamp)) {
2924 info.delta = diff;
2925 if (unlikely(test_time_stamp(info.delta)))
2926 rb_handle_timestamp(cpu_buffer, &info);
2927 }
2928
2929 event = __rb_reserve_next(cpu_buffer, &info);
2930
2931 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2932 if (info.add_timestamp)
2933 info.length -= RB_LEN_TIME_EXTEND;
2934 goto again;
2935 }
2936
2937 if (!event)
2938 goto out_fail;
2939
2940 return event;
2941
2942 out_fail:
2943 rb_end_commit(cpu_buffer);
2944 return NULL;
2945}
2946
2947/**
2948 * ring_buffer_lock_reserve - reserve a part of the buffer
2949 * @buffer: the ring buffer to reserve from
2950 * @length: the length of the data to reserve (excluding event header)
2951 *
2952 * Returns a reserved event on the ring buffer to copy directly to.
2953 * The user of this interface will need to get the body to write into
2954 * and can use the ring_buffer_event_data() interface.
2955 *
2956 * The length is the length of the data needed, not the event length
2957 * which also includes the event header.
2958 *
2959 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2960 * If NULL is returned, then nothing has been allocated or locked.
2961 */
2962struct ring_buffer_event *
2963ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2964{
2965 struct ring_buffer_per_cpu *cpu_buffer;
2966 struct ring_buffer_event *event;
2967 int cpu;
2968
2969 /* If we are tracing schedule, we don't want to recurse */
2970 preempt_disable_notrace();
2971
2972 if (unlikely(atomic_read(&buffer->record_disabled)))
2973 goto out;
2974
2975 cpu = raw_smp_processor_id();
2976
2977 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2978 goto out;
2979
2980 cpu_buffer = buffer->buffers[cpu];
2981
2982 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2983 goto out;
2984
2985 if (unlikely(length > BUF_MAX_DATA_SIZE))
2986 goto out;
2987
2988 if (unlikely(trace_recursive_lock(cpu_buffer)))
2989 goto out;
2990
2991 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2992 if (!event)
2993 goto out_unlock;
2994
2995 return event;
2996
2997 out_unlock:
2998 trace_recursive_unlock(cpu_buffer);
2999 out:
3000 preempt_enable_notrace();
3001 return NULL;
3002}
3003EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3004
3005/*
3006 * Decrement the entries to the page that an event is on.
3007 * The event does not even need to exist, only the pointer
3008 * to the page it is on. This may only be called before the commit
3009 * takes place.
3010 */
3011static inline void
3012rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3013 struct ring_buffer_event *event)
3014{
3015 unsigned long addr = (unsigned long)event;
3016 struct buffer_page *bpage = cpu_buffer->commit_page;
3017 struct buffer_page *start;
3018
3019 addr &= PAGE_MASK;
3020
3021 /* Do the likely case first */
3022 if (likely(bpage->page == (void *)addr)) {
3023 local_dec(&bpage->entries);
3024 return;
3025 }
3026
3027 /*
3028 * Because the commit page may be on the reader page we
3029 * start with the next page and check the end loop there.
3030 */
3031 rb_inc_page(cpu_buffer, &bpage);
3032 start = bpage;
3033 do {
3034 if (bpage->page == (void *)addr) {
3035 local_dec(&bpage->entries);
3036 return;
3037 }
3038 rb_inc_page(cpu_buffer, &bpage);
3039 } while (bpage != start);
3040
3041 /* commit not part of this buffer?? */
3042 RB_WARN_ON(cpu_buffer, 1);
3043}
3044
3045/**
3046 * ring_buffer_commit_discard - discard an event that has not been committed
3047 * @buffer: the ring buffer
3048 * @event: non committed event to discard
3049 *
3050 * Sometimes an event that is in the ring buffer needs to be ignored.
3051 * This function lets the user discard an event in the ring buffer
3052 * and then that event will not be read later.
3053 *
3054 * This function only works if it is called before the item has been
3055 * committed. It will try to free the event from the ring buffer
3056 * if another event has not been added behind it.
3057 *
3058 * If another event has been added behind it, it will set the event
3059 * up as discarded, and perform the commit.
3060 *
3061 * If this function is called, do not call ring_buffer_unlock_commit on
3062 * the event.
3063 */
3064void ring_buffer_discard_commit(struct ring_buffer *buffer,
3065 struct ring_buffer_event *event)
3066{
3067 struct ring_buffer_per_cpu *cpu_buffer;
3068 int cpu;
3069
3070 /* The event is discarded regardless */
3071 rb_event_discard(event);
3072
3073 cpu = smp_processor_id();
3074 cpu_buffer = buffer->buffers[cpu];
3075
3076 /*
3077 * This must only be called if the event has not been
3078 * committed yet. Thus we can assume that preemption
3079 * is still disabled.
3080 */
3081 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3082
3083 rb_decrement_entry(cpu_buffer, event);
3084 if (rb_try_to_discard(cpu_buffer, event))
3085 goto out;
3086
3087 /*
3088 * The commit is still visible by the reader, so we
3089 * must still update the timestamp.
3090 */
3091 rb_update_write_stamp(cpu_buffer, event);
3092 out:
3093 rb_end_commit(cpu_buffer);
3094
3095 trace_recursive_unlock(cpu_buffer);
3096
3097 preempt_enable_notrace();
3098
3099}
3100EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3101
3102/**
3103 * ring_buffer_write - write data to the buffer without reserving
3104 * @buffer: The ring buffer to write to.
3105 * @length: The length of the data being written (excluding the event header)
3106 * @data: The data to write to the buffer.
3107 *
3108 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3109 * one function. If you already have the data to write to the buffer, it
3110 * may be easier to simply call this function.
3111 *
3112 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3113 * and not the length of the event which would hold the header.
3114 */
3115int ring_buffer_write(struct ring_buffer *buffer,
3116 unsigned long length,
3117 void *data)
3118{
3119 struct ring_buffer_per_cpu *cpu_buffer;
3120 struct ring_buffer_event *event;
3121 void *body;
3122 int ret = -EBUSY;
3123 int cpu;
3124
3125 preempt_disable_notrace();
3126
3127 if (atomic_read(&buffer->record_disabled))
3128 goto out;
3129
3130 cpu = raw_smp_processor_id();
3131
3132 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3133 goto out;
3134
3135 cpu_buffer = buffer->buffers[cpu];
3136
3137 if (atomic_read(&cpu_buffer->record_disabled))
3138 goto out;
3139
3140 if (length > BUF_MAX_DATA_SIZE)
3141 goto out;
3142
3143 if (unlikely(trace_recursive_lock(cpu_buffer)))
3144 goto out;
3145
3146 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3147 if (!event)
3148 goto out_unlock;
3149
3150 body = rb_event_data(event);
3151
3152 memcpy(body, data, length);
3153
3154 rb_commit(cpu_buffer, event);
3155
3156 rb_wakeups(buffer, cpu_buffer);
3157
3158 ret = 0;
3159
3160 out_unlock:
3161 trace_recursive_unlock(cpu_buffer);
3162
3163 out:
3164 preempt_enable_notrace();
3165
3166 return ret;
3167}
3168EXPORT_SYMBOL_GPL(ring_buffer_write);
3169
3170static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3171{
3172 struct buffer_page *reader = cpu_buffer->reader_page;
3173 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3174 struct buffer_page *commit = cpu_buffer->commit_page;
3175
3176 /* In case of error, head will be NULL */
3177 if (unlikely(!head))
3178 return true;
3179
3180 return reader->read == rb_page_commit(reader) &&
3181 (commit == reader ||
3182 (commit == head &&
3183 head->read == rb_page_commit(commit)));
3184}
3185
3186/**
3187 * ring_buffer_record_disable - stop all writes into the buffer
3188 * @buffer: The ring buffer to stop writes to.
3189 *
3190 * This prevents all writes to the buffer. Any attempt to write
3191 * to the buffer after this will fail and return NULL.
3192 *
3193 * The caller should call synchronize_rcu() after this.
3194 */
3195void ring_buffer_record_disable(struct ring_buffer *buffer)
3196{
3197 atomic_inc(&buffer->record_disabled);
3198}
3199EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3200
3201/**
3202 * ring_buffer_record_enable - enable writes to the buffer
3203 * @buffer: The ring buffer to enable writes
3204 *
3205 * Note, multiple disables will need the same number of enables
3206 * to truly enable the writing (much like preempt_disable).
3207 */
3208void ring_buffer_record_enable(struct ring_buffer *buffer)
3209{
3210 atomic_dec(&buffer->record_disabled);
3211}
3212EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3213
3214/**
3215 * ring_buffer_record_off - stop all writes into the buffer
3216 * @buffer: The ring buffer to stop writes to.
3217 *
3218 * This prevents all writes to the buffer. Any attempt to write
3219 * to the buffer after this will fail and return NULL.
3220 *
3221 * This is different than ring_buffer_record_disable() as
3222 * it works like an on/off switch, where as the disable() version
3223 * must be paired with a enable().
3224 */
3225void ring_buffer_record_off(struct ring_buffer *buffer)
3226{
3227 unsigned int rd;
3228 unsigned int new_rd;
3229
3230 do {
3231 rd = atomic_read(&buffer->record_disabled);
3232 new_rd = rd | RB_BUFFER_OFF;
3233 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3234}
3235EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3236
3237/**
3238 * ring_buffer_record_on - restart writes into the buffer
3239 * @buffer: The ring buffer to start writes to.
3240 *
3241 * This enables all writes to the buffer that was disabled by
3242 * ring_buffer_record_off().
3243 *
3244 * This is different than ring_buffer_record_enable() as
3245 * it works like an on/off switch, where as the enable() version
3246 * must be paired with a disable().
3247 */
3248void ring_buffer_record_on(struct ring_buffer *buffer)
3249{
3250 unsigned int rd;
3251 unsigned int new_rd;
3252
3253 do {
3254 rd = atomic_read(&buffer->record_disabled);
3255 new_rd = rd & ~RB_BUFFER_OFF;
3256 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3257}
3258EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3259
3260/**
3261 * ring_buffer_record_is_on - return true if the ring buffer can write
3262 * @buffer: The ring buffer to see if write is enabled
3263 *
3264 * Returns true if the ring buffer is in a state that it accepts writes.
3265 */
3266bool ring_buffer_record_is_on(struct ring_buffer *buffer)
3267{
3268 return !atomic_read(&buffer->record_disabled);
3269}
3270
3271/**
3272 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
3273 * @buffer: The ring buffer to see if write is set enabled
3274 *
3275 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
3276 * Note that this does NOT mean it is in a writable state.
3277 *
3278 * It may return true when the ring buffer has been disabled by
3279 * ring_buffer_record_disable(), as that is a temporary disabling of
3280 * the ring buffer.
3281 */
3282bool ring_buffer_record_is_set_on(struct ring_buffer *buffer)
3283{
3284 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
3285}
3286
3287/**
3288 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3289 * @buffer: The ring buffer to stop writes to.
3290 * @cpu: The CPU buffer to stop
3291 *
3292 * This prevents all writes to the buffer. Any attempt to write
3293 * to the buffer after this will fail and return NULL.
3294 *
3295 * The caller should call synchronize_rcu() after this.
3296 */
3297void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3298{
3299 struct ring_buffer_per_cpu *cpu_buffer;
3300
3301 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3302 return;
3303
3304 cpu_buffer = buffer->buffers[cpu];
3305 atomic_inc(&cpu_buffer->record_disabled);
3306}
3307EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3308
3309/**
3310 * ring_buffer_record_enable_cpu - enable writes to the buffer
3311 * @buffer: The ring buffer to enable writes
3312 * @cpu: The CPU to enable.
3313 *
3314 * Note, multiple disables will need the same number of enables
3315 * to truly enable the writing (much like preempt_disable).
3316 */
3317void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3318{
3319 struct ring_buffer_per_cpu *cpu_buffer;
3320
3321 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3322 return;
3323
3324 cpu_buffer = buffer->buffers[cpu];
3325 atomic_dec(&cpu_buffer->record_disabled);
3326}
3327EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3328
3329/*
3330 * The total entries in the ring buffer is the running counter
3331 * of entries entered into the ring buffer, minus the sum of
3332 * the entries read from the ring buffer and the number of
3333 * entries that were overwritten.
3334 */
3335static inline unsigned long
3336rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3337{
3338 return local_read(&cpu_buffer->entries) -
3339 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3340}
3341
3342/**
3343 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3344 * @buffer: The ring buffer
3345 * @cpu: The per CPU buffer to read from.
3346 */
3347u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3348{
3349 unsigned long flags;
3350 struct ring_buffer_per_cpu *cpu_buffer;
3351 struct buffer_page *bpage;
3352 u64 ret = 0;
3353
3354 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3355 return 0;
3356
3357 cpu_buffer = buffer->buffers[cpu];
3358 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3359 /*
3360 * if the tail is on reader_page, oldest time stamp is on the reader
3361 * page
3362 */
3363 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3364 bpage = cpu_buffer->reader_page;
3365 else
3366 bpage = rb_set_head_page(cpu_buffer);
3367 if (bpage)
3368 ret = bpage->page->time_stamp;
3369 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3370
3371 return ret;
3372}
3373EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3374
3375/**
3376 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3377 * @buffer: The ring buffer
3378 * @cpu: The per CPU buffer to read from.
3379 */
3380unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3381{
3382 struct ring_buffer_per_cpu *cpu_buffer;
3383 unsigned long ret;
3384
3385 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3386 return 0;
3387
3388 cpu_buffer = buffer->buffers[cpu];
3389 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3390
3391 return ret;
3392}
3393EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3394
3395/**
3396 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3397 * @buffer: The ring buffer
3398 * @cpu: The per CPU buffer to get the entries from.
3399 */
3400unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3401{
3402 struct ring_buffer_per_cpu *cpu_buffer;
3403
3404 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3405 return 0;
3406
3407 cpu_buffer = buffer->buffers[cpu];
3408
3409 return rb_num_of_entries(cpu_buffer);
3410}
3411EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3412
3413/**
3414 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3415 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3416 * @buffer: The ring buffer
3417 * @cpu: The per CPU buffer to get the number of overruns from
3418 */
3419unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3420{
3421 struct ring_buffer_per_cpu *cpu_buffer;
3422 unsigned long ret;
3423
3424 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3425 return 0;
3426
3427 cpu_buffer = buffer->buffers[cpu];
3428 ret = local_read(&cpu_buffer->overrun);
3429
3430 return ret;
3431}
3432EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3433
3434/**
3435 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3436 * commits failing due to the buffer wrapping around while there are uncommitted
3437 * events, such as during an interrupt storm.
3438 * @buffer: The ring buffer
3439 * @cpu: The per CPU buffer to get the number of overruns from
3440 */
3441unsigned long
3442ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3443{
3444 struct ring_buffer_per_cpu *cpu_buffer;
3445 unsigned long ret;
3446
3447 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3448 return 0;
3449
3450 cpu_buffer = buffer->buffers[cpu];
3451 ret = local_read(&cpu_buffer->commit_overrun);
3452
3453 return ret;
3454}
3455EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3456
3457/**
3458 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3459 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3460 * @buffer: The ring buffer
3461 * @cpu: The per CPU buffer to get the number of overruns from
3462 */
3463unsigned long
3464ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3465{
3466 struct ring_buffer_per_cpu *cpu_buffer;
3467 unsigned long ret;
3468
3469 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3470 return 0;
3471
3472 cpu_buffer = buffer->buffers[cpu];
3473 ret = local_read(&cpu_buffer->dropped_events);
3474
3475 return ret;
3476}
3477EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3478
3479/**
3480 * ring_buffer_read_events_cpu - get the number of events successfully read
3481 * @buffer: The ring buffer
3482 * @cpu: The per CPU buffer to get the number of events read
3483 */
3484unsigned long
3485ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3486{
3487 struct ring_buffer_per_cpu *cpu_buffer;
3488
3489 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3490 return 0;
3491
3492 cpu_buffer = buffer->buffers[cpu];
3493 return cpu_buffer->read;
3494}
3495EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3496
3497/**
3498 * ring_buffer_entries - get the number of entries in a buffer
3499 * @buffer: The ring buffer
3500 *
3501 * Returns the total number of entries in the ring buffer
3502 * (all CPU entries)
3503 */
3504unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3505{
3506 struct ring_buffer_per_cpu *cpu_buffer;
3507 unsigned long entries = 0;
3508 int cpu;
3509
3510 /* if you care about this being correct, lock the buffer */
3511 for_each_buffer_cpu(buffer, cpu) {
3512 cpu_buffer = buffer->buffers[cpu];
3513 entries += rb_num_of_entries(cpu_buffer);
3514 }
3515
3516 return entries;
3517}
3518EXPORT_SYMBOL_GPL(ring_buffer_entries);
3519
3520/**
3521 * ring_buffer_overruns - get the number of overruns in buffer
3522 * @buffer: The ring buffer
3523 *
3524 * Returns the total number of overruns in the ring buffer
3525 * (all CPU entries)
3526 */
3527unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3528{
3529 struct ring_buffer_per_cpu *cpu_buffer;
3530 unsigned long overruns = 0;
3531 int cpu;
3532
3533 /* if you care about this being correct, lock the buffer */
3534 for_each_buffer_cpu(buffer, cpu) {
3535 cpu_buffer = buffer->buffers[cpu];
3536 overruns += local_read(&cpu_buffer->overrun);
3537 }
3538
3539 return overruns;
3540}
3541EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3542
3543static void rb_iter_reset(struct ring_buffer_iter *iter)
3544{
3545 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3546
3547 /* Iterator usage is expected to have record disabled */
3548 iter->head_page = cpu_buffer->reader_page;
3549 iter->head = cpu_buffer->reader_page->read;
3550
3551 iter->cache_reader_page = iter->head_page;
3552 iter->cache_read = cpu_buffer->read;
3553
3554 if (iter->head)
3555 iter->read_stamp = cpu_buffer->read_stamp;
3556 else
3557 iter->read_stamp = iter->head_page->page->time_stamp;
3558}
3559
3560/**
3561 * ring_buffer_iter_reset - reset an iterator
3562 * @iter: The iterator to reset
3563 *
3564 * Resets the iterator, so that it will start from the beginning
3565 * again.
3566 */
3567void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3568{
3569 struct ring_buffer_per_cpu *cpu_buffer;
3570 unsigned long flags;
3571
3572 if (!iter)
3573 return;
3574
3575 cpu_buffer = iter->cpu_buffer;
3576
3577 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3578 rb_iter_reset(iter);
3579 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3580}
3581EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3582
3583/**
3584 * ring_buffer_iter_empty - check if an iterator has no more to read
3585 * @iter: The iterator to check
3586 */
3587int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3588{
3589 struct ring_buffer_per_cpu *cpu_buffer;
3590 struct buffer_page *reader;
3591 struct buffer_page *head_page;
3592 struct buffer_page *commit_page;
3593 unsigned commit;
3594
3595 cpu_buffer = iter->cpu_buffer;
3596
3597 /* Remember, trace recording is off when iterator is in use */
3598 reader = cpu_buffer->reader_page;
3599 head_page = cpu_buffer->head_page;
3600 commit_page = cpu_buffer->commit_page;
3601 commit = rb_page_commit(commit_page);
3602
3603 return ((iter->head_page == commit_page && iter->head == commit) ||
3604 (iter->head_page == reader && commit_page == head_page &&
3605 head_page->read == commit &&
3606 iter->head == rb_page_commit(cpu_buffer->reader_page)));
3607}
3608EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3609
3610static void
3611rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3612 struct ring_buffer_event *event)
3613{
3614 u64 delta;
3615
3616 switch (event->type_len) {
3617 case RINGBUF_TYPE_PADDING:
3618 return;
3619
3620 case RINGBUF_TYPE_TIME_EXTEND:
3621 delta = ring_buffer_event_time_stamp(event);
3622 cpu_buffer->read_stamp += delta;
3623 return;
3624
3625 case RINGBUF_TYPE_TIME_STAMP:
3626 delta = ring_buffer_event_time_stamp(event);
3627 cpu_buffer->read_stamp = delta;
3628 return;
3629
3630 case RINGBUF_TYPE_DATA:
3631 cpu_buffer->read_stamp += event->time_delta;
3632 return;
3633
3634 default:
3635 BUG();
3636 }
3637 return;
3638}
3639
3640static void
3641rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3642 struct ring_buffer_event *event)
3643{
3644 u64 delta;
3645
3646 switch (event->type_len) {
3647 case RINGBUF_TYPE_PADDING:
3648 return;
3649
3650 case RINGBUF_TYPE_TIME_EXTEND:
3651 delta = ring_buffer_event_time_stamp(event);
3652 iter->read_stamp += delta;
3653 return;
3654
3655 case RINGBUF_TYPE_TIME_STAMP:
3656 delta = ring_buffer_event_time_stamp(event);
3657 iter->read_stamp = delta;
3658 return;
3659
3660 case RINGBUF_TYPE_DATA:
3661 iter->read_stamp += event->time_delta;
3662 return;
3663
3664 default:
3665 BUG();
3666 }
3667 return;
3668}
3669
3670static struct buffer_page *
3671rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3672{
3673 struct buffer_page *reader = NULL;
3674 unsigned long overwrite;
3675 unsigned long flags;
3676 int nr_loops = 0;
3677 int ret;
3678
3679 local_irq_save(flags);
3680 arch_spin_lock(&cpu_buffer->lock);
3681
3682 again:
3683 /*
3684 * This should normally only loop twice. But because the
3685 * start of the reader inserts an empty page, it causes
3686 * a case where we will loop three times. There should be no
3687 * reason to loop four times (that I know of).
3688 */
3689 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3690 reader = NULL;
3691 goto out;
3692 }
3693
3694 reader = cpu_buffer->reader_page;
3695
3696 /* If there's more to read, return this page */
3697 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3698 goto out;
3699
3700 /* Never should we have an index greater than the size */
3701 if (RB_WARN_ON(cpu_buffer,
3702 cpu_buffer->reader_page->read > rb_page_size(reader)))
3703 goto out;
3704
3705 /* check if we caught up to the tail */
3706 reader = NULL;
3707 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3708 goto out;
3709
3710 /* Don't bother swapping if the ring buffer is empty */
3711 if (rb_num_of_entries(cpu_buffer) == 0)
3712 goto out;
3713
3714 /*
3715 * Reset the reader page to size zero.
3716 */
3717 local_set(&cpu_buffer->reader_page->write, 0);
3718 local_set(&cpu_buffer->reader_page->entries, 0);
3719 local_set(&cpu_buffer->reader_page->page->commit, 0);
3720 cpu_buffer->reader_page->real_end = 0;
3721
3722 spin:
3723 /*
3724 * Splice the empty reader page into the list around the head.
3725 */
3726 reader = rb_set_head_page(cpu_buffer);
3727 if (!reader)
3728 goto out;
3729 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3730 cpu_buffer->reader_page->list.prev = reader->list.prev;
3731
3732 /*
3733 * cpu_buffer->pages just needs to point to the buffer, it
3734 * has no specific buffer page to point to. Lets move it out
3735 * of our way so we don't accidentally swap it.
3736 */
3737 cpu_buffer->pages = reader->list.prev;
3738
3739 /* The reader page will be pointing to the new head */
3740 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3741
3742 /*
3743 * We want to make sure we read the overruns after we set up our
3744 * pointers to the next object. The writer side does a
3745 * cmpxchg to cross pages which acts as the mb on the writer
3746 * side. Note, the reader will constantly fail the swap
3747 * while the writer is updating the pointers, so this
3748 * guarantees that the overwrite recorded here is the one we
3749 * want to compare with the last_overrun.
3750 */
3751 smp_mb();
3752 overwrite = local_read(&(cpu_buffer->overrun));
3753
3754 /*
3755 * Here's the tricky part.
3756 *
3757 * We need to move the pointer past the header page.
3758 * But we can only do that if a writer is not currently
3759 * moving it. The page before the header page has the
3760 * flag bit '1' set if it is pointing to the page we want.
3761 * but if the writer is in the process of moving it
3762 * than it will be '2' or already moved '0'.
3763 */
3764
3765 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3766
3767 /*
3768 * If we did not convert it, then we must try again.
3769 */
3770 if (!ret)
3771 goto spin;
3772
3773 /*
3774 * Yay! We succeeded in replacing the page.
3775 *
3776 * Now make the new head point back to the reader page.
3777 */
3778 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3779 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3780
3781 local_inc(&cpu_buffer->pages_read);
3782
3783 /* Finally update the reader page to the new head */
3784 cpu_buffer->reader_page = reader;
3785 cpu_buffer->reader_page->read = 0;
3786
3787 if (overwrite != cpu_buffer->last_overrun) {
3788 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3789 cpu_buffer->last_overrun = overwrite;
3790 }
3791
3792 goto again;
3793
3794 out:
3795 /* Update the read_stamp on the first event */
3796 if (reader && reader->read == 0)
3797 cpu_buffer->read_stamp = reader->page->time_stamp;
3798
3799 arch_spin_unlock(&cpu_buffer->lock);
3800 local_irq_restore(flags);
3801
3802 return reader;
3803}
3804
3805static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3806{
3807 struct ring_buffer_event *event;
3808 struct buffer_page *reader;
3809 unsigned length;
3810
3811 reader = rb_get_reader_page(cpu_buffer);
3812
3813 /* This function should not be called when buffer is empty */
3814 if (RB_WARN_ON(cpu_buffer, !reader))
3815 return;
3816
3817 event = rb_reader_event(cpu_buffer);
3818
3819 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3820 cpu_buffer->read++;
3821
3822 rb_update_read_stamp(cpu_buffer, event);
3823
3824 length = rb_event_length(event);
3825 cpu_buffer->reader_page->read += length;
3826}
3827
3828static void rb_advance_iter(struct ring_buffer_iter *iter)
3829{
3830 struct ring_buffer_per_cpu *cpu_buffer;
3831 struct ring_buffer_event *event;
3832 unsigned length;
3833
3834 cpu_buffer = iter->cpu_buffer;
3835
3836 /*
3837 * Check if we are at the end of the buffer.
3838 */
3839 if (iter->head >= rb_page_size(iter->head_page)) {
3840 /* discarded commits can make the page empty */
3841 if (iter->head_page == cpu_buffer->commit_page)
3842 return;
3843 rb_inc_iter(iter);
3844 return;
3845 }
3846
3847 event = rb_iter_head_event(iter);
3848
3849 length = rb_event_length(event);
3850
3851 /*
3852 * This should not be called to advance the header if we are
3853 * at the tail of the buffer.
3854 */
3855 if (RB_WARN_ON(cpu_buffer,
3856 (iter->head_page == cpu_buffer->commit_page) &&
3857 (iter->head + length > rb_commit_index(cpu_buffer))))
3858 return;
3859
3860 rb_update_iter_read_stamp(iter, event);
3861
3862 iter->head += length;
3863
3864 /* check for end of page padding */
3865 if ((iter->head >= rb_page_size(iter->head_page)) &&
3866 (iter->head_page != cpu_buffer->commit_page))
3867 rb_inc_iter(iter);
3868}
3869
3870static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3871{
3872 return cpu_buffer->lost_events;
3873}
3874
3875static struct ring_buffer_event *
3876rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3877 unsigned long *lost_events)
3878{
3879 struct ring_buffer_event *event;
3880 struct buffer_page *reader;
3881 int nr_loops = 0;
3882
3883 if (ts)
3884 *ts = 0;
3885 again:
3886 /*
3887 * We repeat when a time extend is encountered.
3888 * Since the time extend is always attached to a data event,
3889 * we should never loop more than once.
3890 * (We never hit the following condition more than twice).
3891 */
3892 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3893 return NULL;
3894
3895 reader = rb_get_reader_page(cpu_buffer);
3896 if (!reader)
3897 return NULL;
3898
3899 event = rb_reader_event(cpu_buffer);
3900
3901 switch (event->type_len) {
3902 case RINGBUF_TYPE_PADDING:
3903 if (rb_null_event(event))
3904 RB_WARN_ON(cpu_buffer, 1);
3905 /*
3906 * Because the writer could be discarding every
3907 * event it creates (which would probably be bad)
3908 * if we were to go back to "again" then we may never
3909 * catch up, and will trigger the warn on, or lock
3910 * the box. Return the padding, and we will release
3911 * the current locks, and try again.
3912 */
3913 return event;
3914
3915 case RINGBUF_TYPE_TIME_EXTEND:
3916 /* Internal data, OK to advance */
3917 rb_advance_reader(cpu_buffer);
3918 goto again;
3919
3920 case RINGBUF_TYPE_TIME_STAMP:
3921 if (ts) {
3922 *ts = ring_buffer_event_time_stamp(event);
3923 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3924 cpu_buffer->cpu, ts);
3925 }
3926 /* Internal data, OK to advance */
3927 rb_advance_reader(cpu_buffer);
3928 goto again;
3929
3930 case RINGBUF_TYPE_DATA:
3931 if (ts && !(*ts)) {
3932 *ts = cpu_buffer->read_stamp + event->time_delta;
3933 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3934 cpu_buffer->cpu, ts);
3935 }
3936 if (lost_events)
3937 *lost_events = rb_lost_events(cpu_buffer);
3938 return event;
3939
3940 default:
3941 BUG();
3942 }
3943
3944 return NULL;
3945}
3946EXPORT_SYMBOL_GPL(ring_buffer_peek);
3947
3948static struct ring_buffer_event *
3949rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3950{
3951 struct ring_buffer *buffer;
3952 struct ring_buffer_per_cpu *cpu_buffer;
3953 struct ring_buffer_event *event;
3954 int nr_loops = 0;
3955
3956 if (ts)
3957 *ts = 0;
3958
3959 cpu_buffer = iter->cpu_buffer;
3960 buffer = cpu_buffer->buffer;
3961
3962 /*
3963 * Check if someone performed a consuming read to
3964 * the buffer. A consuming read invalidates the iterator
3965 * and we need to reset the iterator in this case.
3966 */
3967 if (unlikely(iter->cache_read != cpu_buffer->read ||
3968 iter->cache_reader_page != cpu_buffer->reader_page))
3969 rb_iter_reset(iter);
3970
3971 again:
3972 if (ring_buffer_iter_empty(iter))
3973 return NULL;
3974
3975 /*
3976 * We repeat when a time extend is encountered or we hit
3977 * the end of the page. Since the time extend is always attached
3978 * to a data event, we should never loop more than three times.
3979 * Once for going to next page, once on time extend, and
3980 * finally once to get the event.
3981 * (We never hit the following condition more than thrice).
3982 */
3983 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3984 return NULL;
3985
3986 if (rb_per_cpu_empty(cpu_buffer))
3987 return NULL;
3988
3989 if (iter->head >= rb_page_size(iter->head_page)) {
3990 rb_inc_iter(iter);
3991 goto again;
3992 }
3993
3994 event = rb_iter_head_event(iter);
3995
3996 switch (event->type_len) {
3997 case RINGBUF_TYPE_PADDING:
3998 if (rb_null_event(event)) {
3999 rb_inc_iter(iter);
4000 goto again;
4001 }
4002 rb_advance_iter(iter);
4003 return event;
4004
4005 case RINGBUF_TYPE_TIME_EXTEND:
4006 /* Internal data, OK to advance */
4007 rb_advance_iter(iter);
4008 goto again;
4009
4010 case RINGBUF_TYPE_TIME_STAMP:
4011 if (ts) {
4012 *ts = ring_buffer_event_time_stamp(event);
4013 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4014 cpu_buffer->cpu, ts);
4015 }
4016 /* Internal data, OK to advance */
4017 rb_advance_iter(iter);
4018 goto again;
4019
4020 case RINGBUF_TYPE_DATA:
4021 if (ts && !(*ts)) {
4022 *ts = iter->read_stamp + event->time_delta;
4023 ring_buffer_normalize_time_stamp(buffer,
4024 cpu_buffer->cpu, ts);
4025 }
4026 return event;
4027
4028 default:
4029 BUG();
4030 }
4031
4032 return NULL;
4033}
4034EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4035
4036static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4037{
4038 if (likely(!in_nmi())) {
4039 raw_spin_lock(&cpu_buffer->reader_lock);
4040 return true;
4041 }
4042
4043 /*
4044 * If an NMI die dumps out the content of the ring buffer
4045 * trylock must be used to prevent a deadlock if the NMI
4046 * preempted a task that holds the ring buffer locks. If
4047 * we get the lock then all is fine, if not, then continue
4048 * to do the read, but this can corrupt the ring buffer,
4049 * so it must be permanently disabled from future writes.
4050 * Reading from NMI is a oneshot deal.
4051 */
4052 if (raw_spin_trylock(&cpu_buffer->reader_lock))
4053 return true;
4054
4055 /* Continue without locking, but disable the ring buffer */
4056 atomic_inc(&cpu_buffer->record_disabled);
4057 return false;
4058}
4059
4060static inline void
4061rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4062{
4063 if (likely(locked))
4064 raw_spin_unlock(&cpu_buffer->reader_lock);
4065 return;
4066}
4067
4068/**
4069 * ring_buffer_peek - peek at the next event to be read
4070 * @buffer: The ring buffer to read
4071 * @cpu: The cpu to peak at
4072 * @ts: The timestamp counter of this event.
4073 * @lost_events: a variable to store if events were lost (may be NULL)
4074 *
4075 * This will return the event that will be read next, but does
4076 * not consume the data.
4077 */
4078struct ring_buffer_event *
4079ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
4080 unsigned long *lost_events)
4081{
4082 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4083 struct ring_buffer_event *event;
4084 unsigned long flags;
4085 bool dolock;
4086
4087 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4088 return NULL;
4089
4090 again:
4091 local_irq_save(flags);
4092 dolock = rb_reader_lock(cpu_buffer);
4093 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4094 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4095 rb_advance_reader(cpu_buffer);
4096 rb_reader_unlock(cpu_buffer, dolock);
4097 local_irq_restore(flags);
4098
4099 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4100 goto again;
4101
4102 return event;
4103}
4104
4105/**
4106 * ring_buffer_iter_peek - peek at the next event to be read
4107 * @iter: The ring buffer iterator
4108 * @ts: The timestamp counter of this event.
4109 *
4110 * This will return the event that will be read next, but does
4111 * not increment the iterator.
4112 */
4113struct ring_buffer_event *
4114ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4115{
4116 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4117 struct ring_buffer_event *event;
4118 unsigned long flags;
4119
4120 again:
4121 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4122 event = rb_iter_peek(iter, ts);
4123 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4124
4125 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4126 goto again;
4127
4128 return event;
4129}
4130
4131/**
4132 * ring_buffer_consume - return an event and consume it
4133 * @buffer: The ring buffer to get the next event from
4134 * @cpu: the cpu to read the buffer from
4135 * @ts: a variable to store the timestamp (may be NULL)
4136 * @lost_events: a variable to store if events were lost (may be NULL)
4137 *
4138 * Returns the next event in the ring buffer, and that event is consumed.
4139 * Meaning, that sequential reads will keep returning a different event,
4140 * and eventually empty the ring buffer if the producer is slower.
4141 */
4142struct ring_buffer_event *
4143ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
4144 unsigned long *lost_events)
4145{
4146 struct ring_buffer_per_cpu *cpu_buffer;
4147 struct ring_buffer_event *event = NULL;
4148 unsigned long flags;
4149 bool dolock;
4150
4151 again:
4152 /* might be called in atomic */
4153 preempt_disable();
4154
4155 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4156 goto out;
4157
4158 cpu_buffer = buffer->buffers[cpu];
4159 local_irq_save(flags);
4160 dolock = rb_reader_lock(cpu_buffer);
4161
4162 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4163 if (event) {
4164 cpu_buffer->lost_events = 0;
4165 rb_advance_reader(cpu_buffer);
4166 }
4167
4168 rb_reader_unlock(cpu_buffer, dolock);
4169 local_irq_restore(flags);
4170
4171 out:
4172 preempt_enable();
4173
4174 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4175 goto again;
4176
4177 return event;
4178}
4179EXPORT_SYMBOL_GPL(ring_buffer_consume);
4180
4181/**
4182 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4183 * @buffer: The ring buffer to read from
4184 * @cpu: The cpu buffer to iterate over
4185 * @flags: gfp flags to use for memory allocation
4186 *
4187 * This performs the initial preparations necessary to iterate
4188 * through the buffer. Memory is allocated, buffer recording
4189 * is disabled, and the iterator pointer is returned to the caller.
4190 *
4191 * Disabling buffer recording prevents the reading from being
4192 * corrupted. This is not a consuming read, so a producer is not
4193 * expected.
4194 *
4195 * After a sequence of ring_buffer_read_prepare calls, the user is
4196 * expected to make at least one call to ring_buffer_read_prepare_sync.
4197 * Afterwards, ring_buffer_read_start is invoked to get things going
4198 * for real.
4199 *
4200 * This overall must be paired with ring_buffer_read_finish.
4201 */
4202struct ring_buffer_iter *
4203ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu, gfp_t flags)
4204{
4205 struct ring_buffer_per_cpu *cpu_buffer;
4206 struct ring_buffer_iter *iter;
4207
4208 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4209 return NULL;
4210
4211 iter = kmalloc(sizeof(*iter), flags);
4212 if (!iter)
4213 return NULL;
4214
4215 cpu_buffer = buffer->buffers[cpu];
4216
4217 iter->cpu_buffer = cpu_buffer;
4218
4219 atomic_inc(&buffer->resize_disabled);
4220 atomic_inc(&cpu_buffer->record_disabled);
4221
4222 return iter;
4223}
4224EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4225
4226/**
4227 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4228 *
4229 * All previously invoked ring_buffer_read_prepare calls to prepare
4230 * iterators will be synchronized. Afterwards, read_buffer_read_start
4231 * calls on those iterators are allowed.
4232 */
4233void
4234ring_buffer_read_prepare_sync(void)
4235{
4236 synchronize_rcu();
4237}
4238EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4239
4240/**
4241 * ring_buffer_read_start - start a non consuming read of the buffer
4242 * @iter: The iterator returned by ring_buffer_read_prepare
4243 *
4244 * This finalizes the startup of an iteration through the buffer.
4245 * The iterator comes from a call to ring_buffer_read_prepare and
4246 * an intervening ring_buffer_read_prepare_sync must have been
4247 * performed.
4248 *
4249 * Must be paired with ring_buffer_read_finish.
4250 */
4251void
4252ring_buffer_read_start(struct ring_buffer_iter *iter)
4253{
4254 struct ring_buffer_per_cpu *cpu_buffer;
4255 unsigned long flags;
4256
4257 if (!iter)
4258 return;
4259
4260 cpu_buffer = iter->cpu_buffer;
4261
4262 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4263 arch_spin_lock(&cpu_buffer->lock);
4264 rb_iter_reset(iter);
4265 arch_spin_unlock(&cpu_buffer->lock);
4266 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4267}
4268EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4269
4270/**
4271 * ring_buffer_read_finish - finish reading the iterator of the buffer
4272 * @iter: The iterator retrieved by ring_buffer_start
4273 *
4274 * This re-enables the recording to the buffer, and frees the
4275 * iterator.
4276 */
4277void
4278ring_buffer_read_finish(struct ring_buffer_iter *iter)
4279{
4280 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4281 unsigned long flags;
4282
4283 /*
4284 * Ring buffer is disabled from recording, here's a good place
4285 * to check the integrity of the ring buffer.
4286 * Must prevent readers from trying to read, as the check
4287 * clears the HEAD page and readers require it.
4288 */
4289 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4290 rb_check_pages(cpu_buffer);
4291 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4292
4293 atomic_dec(&cpu_buffer->record_disabled);
4294 atomic_dec(&cpu_buffer->buffer->resize_disabled);
4295 kfree(iter);
4296}
4297EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4298
4299/**
4300 * ring_buffer_read - read the next item in the ring buffer by the iterator
4301 * @iter: The ring buffer iterator
4302 * @ts: The time stamp of the event read.
4303 *
4304 * This reads the next event in the ring buffer and increments the iterator.
4305 */
4306struct ring_buffer_event *
4307ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4308{
4309 struct ring_buffer_event *event;
4310 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4311 unsigned long flags;
4312
4313 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4314 again:
4315 event = rb_iter_peek(iter, ts);
4316 if (!event)
4317 goto out;
4318
4319 if (event->type_len == RINGBUF_TYPE_PADDING)
4320 goto again;
4321
4322 rb_advance_iter(iter);
4323 out:
4324 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4325
4326 return event;
4327}
4328EXPORT_SYMBOL_GPL(ring_buffer_read);
4329
4330/**
4331 * ring_buffer_size - return the size of the ring buffer (in bytes)
4332 * @buffer: The ring buffer.
4333 */
4334unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4335{
4336 /*
4337 * Earlier, this method returned
4338 * BUF_PAGE_SIZE * buffer->nr_pages
4339 * Since the nr_pages field is now removed, we have converted this to
4340 * return the per cpu buffer value.
4341 */
4342 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4343 return 0;
4344
4345 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4346}
4347EXPORT_SYMBOL_GPL(ring_buffer_size);
4348
4349static void
4350rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4351{
4352 rb_head_page_deactivate(cpu_buffer);
4353
4354 cpu_buffer->head_page
4355 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4356 local_set(&cpu_buffer->head_page->write, 0);
4357 local_set(&cpu_buffer->head_page->entries, 0);
4358 local_set(&cpu_buffer->head_page->page->commit, 0);
4359
4360 cpu_buffer->head_page->read = 0;
4361
4362 cpu_buffer->tail_page = cpu_buffer->head_page;
4363 cpu_buffer->commit_page = cpu_buffer->head_page;
4364
4365 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4366 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4367 local_set(&cpu_buffer->reader_page->write, 0);
4368 local_set(&cpu_buffer->reader_page->entries, 0);
4369 local_set(&cpu_buffer->reader_page->page->commit, 0);
4370 cpu_buffer->reader_page->read = 0;
4371
4372 local_set(&cpu_buffer->entries_bytes, 0);
4373 local_set(&cpu_buffer->overrun, 0);
4374 local_set(&cpu_buffer->commit_overrun, 0);
4375 local_set(&cpu_buffer->dropped_events, 0);
4376 local_set(&cpu_buffer->entries, 0);
4377 local_set(&cpu_buffer->committing, 0);
4378 local_set(&cpu_buffer->commits, 0);
4379 local_set(&cpu_buffer->pages_touched, 0);
4380 local_set(&cpu_buffer->pages_read, 0);
4381 cpu_buffer->last_pages_touch = 0;
4382 cpu_buffer->shortest_full = 0;
4383 cpu_buffer->read = 0;
4384 cpu_buffer->read_bytes = 0;
4385
4386 cpu_buffer->write_stamp = 0;
4387 cpu_buffer->read_stamp = 0;
4388
4389 cpu_buffer->lost_events = 0;
4390 cpu_buffer->last_overrun = 0;
4391
4392 rb_head_page_activate(cpu_buffer);
4393}
4394
4395/**
4396 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4397 * @buffer: The ring buffer to reset a per cpu buffer of
4398 * @cpu: The CPU buffer to be reset
4399 */
4400void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4401{
4402 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4403 unsigned long flags;
4404
4405 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4406 return;
4407
4408 atomic_inc(&buffer->resize_disabled);
4409 atomic_inc(&cpu_buffer->record_disabled);
4410
4411 /* Make sure all commits have finished */
4412 synchronize_rcu();
4413
4414 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4415
4416 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4417 goto out;
4418
4419 arch_spin_lock(&cpu_buffer->lock);
4420
4421 rb_reset_cpu(cpu_buffer);
4422
4423 arch_spin_unlock(&cpu_buffer->lock);
4424
4425 out:
4426 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4427
4428 atomic_dec(&cpu_buffer->record_disabled);
4429 atomic_dec(&buffer->resize_disabled);
4430}
4431EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4432
4433/**
4434 * ring_buffer_reset - reset a ring buffer
4435 * @buffer: The ring buffer to reset all cpu buffers
4436 */
4437void ring_buffer_reset(struct ring_buffer *buffer)
4438{
4439 int cpu;
4440
4441 for_each_buffer_cpu(buffer, cpu)
4442 ring_buffer_reset_cpu(buffer, cpu);
4443}
4444EXPORT_SYMBOL_GPL(ring_buffer_reset);
4445
4446/**
4447 * rind_buffer_empty - is the ring buffer empty?
4448 * @buffer: The ring buffer to test
4449 */
4450bool ring_buffer_empty(struct ring_buffer *buffer)
4451{
4452 struct ring_buffer_per_cpu *cpu_buffer;
4453 unsigned long flags;
4454 bool dolock;
4455 int cpu;
4456 int ret;
4457
4458 /* yes this is racy, but if you don't like the race, lock the buffer */
4459 for_each_buffer_cpu(buffer, cpu) {
4460 cpu_buffer = buffer->buffers[cpu];
4461 local_irq_save(flags);
4462 dolock = rb_reader_lock(cpu_buffer);
4463 ret = rb_per_cpu_empty(cpu_buffer);
4464 rb_reader_unlock(cpu_buffer, dolock);
4465 local_irq_restore(flags);
4466
4467 if (!ret)
4468 return false;
4469 }
4470
4471 return true;
4472}
4473EXPORT_SYMBOL_GPL(ring_buffer_empty);
4474
4475/**
4476 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4477 * @buffer: The ring buffer
4478 * @cpu: The CPU buffer to test
4479 */
4480bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4481{
4482 struct ring_buffer_per_cpu *cpu_buffer;
4483 unsigned long flags;
4484 bool dolock;
4485 int ret;
4486
4487 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4488 return true;
4489
4490 cpu_buffer = buffer->buffers[cpu];
4491 local_irq_save(flags);
4492 dolock = rb_reader_lock(cpu_buffer);
4493 ret = rb_per_cpu_empty(cpu_buffer);
4494 rb_reader_unlock(cpu_buffer, dolock);
4495 local_irq_restore(flags);
4496
4497 return ret;
4498}
4499EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4500
4501#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4502/**
4503 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4504 * @buffer_a: One buffer to swap with
4505 * @buffer_b: The other buffer to swap with
4506 *
4507 * This function is useful for tracers that want to take a "snapshot"
4508 * of a CPU buffer and has another back up buffer lying around.
4509 * it is expected that the tracer handles the cpu buffer not being
4510 * used at the moment.
4511 */
4512int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4513 struct ring_buffer *buffer_b, int cpu)
4514{
4515 struct ring_buffer_per_cpu *cpu_buffer_a;
4516 struct ring_buffer_per_cpu *cpu_buffer_b;
4517 int ret = -EINVAL;
4518
4519 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4520 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4521 goto out;
4522
4523 cpu_buffer_a = buffer_a->buffers[cpu];
4524 cpu_buffer_b = buffer_b->buffers[cpu];
4525
4526 /* At least make sure the two buffers are somewhat the same */
4527 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4528 goto out;
4529
4530 ret = -EAGAIN;
4531
4532 if (atomic_read(&buffer_a->record_disabled))
4533 goto out;
4534
4535 if (atomic_read(&buffer_b->record_disabled))
4536 goto out;
4537
4538 if (atomic_read(&cpu_buffer_a->record_disabled))
4539 goto out;
4540
4541 if (atomic_read(&cpu_buffer_b->record_disabled))
4542 goto out;
4543
4544 /*
4545 * We can't do a synchronize_rcu here because this
4546 * function can be called in atomic context.
4547 * Normally this will be called from the same CPU as cpu.
4548 * If not it's up to the caller to protect this.
4549 */
4550 atomic_inc(&cpu_buffer_a->record_disabled);
4551 atomic_inc(&cpu_buffer_b->record_disabled);
4552
4553 ret = -EBUSY;
4554 if (local_read(&cpu_buffer_a->committing))
4555 goto out_dec;
4556 if (local_read(&cpu_buffer_b->committing))
4557 goto out_dec;
4558
4559 buffer_a->buffers[cpu] = cpu_buffer_b;
4560 buffer_b->buffers[cpu] = cpu_buffer_a;
4561
4562 cpu_buffer_b->buffer = buffer_a;
4563 cpu_buffer_a->buffer = buffer_b;
4564
4565 ret = 0;
4566
4567out_dec:
4568 atomic_dec(&cpu_buffer_a->record_disabled);
4569 atomic_dec(&cpu_buffer_b->record_disabled);
4570out:
4571 return ret;
4572}
4573EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4574#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4575
4576/**
4577 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4578 * @buffer: the buffer to allocate for.
4579 * @cpu: the cpu buffer to allocate.
4580 *
4581 * This function is used in conjunction with ring_buffer_read_page.
4582 * When reading a full page from the ring buffer, these functions
4583 * can be used to speed up the process. The calling function should
4584 * allocate a few pages first with this function. Then when it
4585 * needs to get pages from the ring buffer, it passes the result
4586 * of this function into ring_buffer_read_page, which will swap
4587 * the page that was allocated, with the read page of the buffer.
4588 *
4589 * Returns:
4590 * The page allocated, or ERR_PTR
4591 */
4592void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4593{
4594 struct ring_buffer_per_cpu *cpu_buffer;
4595 struct buffer_data_page *bpage = NULL;
4596 unsigned long flags;
4597 struct page *page;
4598
4599 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4600 return ERR_PTR(-ENODEV);
4601
4602 cpu_buffer = buffer->buffers[cpu];
4603 local_irq_save(flags);
4604 arch_spin_lock(&cpu_buffer->lock);
4605
4606 if (cpu_buffer->free_page) {
4607 bpage = cpu_buffer->free_page;
4608 cpu_buffer->free_page = NULL;
4609 }
4610
4611 arch_spin_unlock(&cpu_buffer->lock);
4612 local_irq_restore(flags);
4613
4614 if (bpage)
4615 goto out;
4616
4617 page = alloc_pages_node(cpu_to_node(cpu),
4618 GFP_KERNEL | __GFP_NORETRY, 0);
4619 if (!page)
4620 return ERR_PTR(-ENOMEM);
4621
4622 bpage = page_address(page);
4623
4624 out:
4625 rb_init_page(bpage);
4626
4627 return bpage;
4628}
4629EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4630
4631/**
4632 * ring_buffer_free_read_page - free an allocated read page
4633 * @buffer: the buffer the page was allocate for
4634 * @cpu: the cpu buffer the page came from
4635 * @data: the page to free
4636 *
4637 * Free a page allocated from ring_buffer_alloc_read_page.
4638 */
4639void ring_buffer_free_read_page(struct ring_buffer *buffer, int cpu, void *data)
4640{
4641 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4642 struct buffer_data_page *bpage = data;
4643 struct page *page = virt_to_page(bpage);
4644 unsigned long flags;
4645
4646 /* If the page is still in use someplace else, we can't reuse it */
4647 if (page_ref_count(page) > 1)
4648 goto out;
4649
4650 local_irq_save(flags);
4651 arch_spin_lock(&cpu_buffer->lock);
4652
4653 if (!cpu_buffer->free_page) {
4654 cpu_buffer->free_page = bpage;
4655 bpage = NULL;
4656 }
4657
4658 arch_spin_unlock(&cpu_buffer->lock);
4659 local_irq_restore(flags);
4660
4661 out:
4662 free_page((unsigned long)bpage);
4663}
4664EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4665
4666/**
4667 * ring_buffer_read_page - extract a page from the ring buffer
4668 * @buffer: buffer to extract from
4669 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4670 * @len: amount to extract
4671 * @cpu: the cpu of the buffer to extract
4672 * @full: should the extraction only happen when the page is full.
4673 *
4674 * This function will pull out a page from the ring buffer and consume it.
4675 * @data_page must be the address of the variable that was returned
4676 * from ring_buffer_alloc_read_page. This is because the page might be used
4677 * to swap with a page in the ring buffer.
4678 *
4679 * for example:
4680 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4681 * if (IS_ERR(rpage))
4682 * return PTR_ERR(rpage);
4683 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4684 * if (ret >= 0)
4685 * process_page(rpage, ret);
4686 *
4687 * When @full is set, the function will not return true unless
4688 * the writer is off the reader page.
4689 *
4690 * Note: it is up to the calling functions to handle sleeps and wakeups.
4691 * The ring buffer can be used anywhere in the kernel and can not
4692 * blindly call wake_up. The layer that uses the ring buffer must be
4693 * responsible for that.
4694 *
4695 * Returns:
4696 * >=0 if data has been transferred, returns the offset of consumed data.
4697 * <0 if no data has been transferred.
4698 */
4699int ring_buffer_read_page(struct ring_buffer *buffer,
4700 void **data_page, size_t len, int cpu, int full)
4701{
4702 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4703 struct ring_buffer_event *event;
4704 struct buffer_data_page *bpage;
4705 struct buffer_page *reader;
4706 unsigned long missed_events;
4707 unsigned long flags;
4708 unsigned int commit;
4709 unsigned int read;
4710 u64 save_timestamp;
4711 int ret = -1;
4712
4713 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4714 goto out;
4715
4716 /*
4717 * If len is not big enough to hold the page header, then
4718 * we can not copy anything.
4719 */
4720 if (len <= BUF_PAGE_HDR_SIZE)
4721 goto out;
4722
4723 len -= BUF_PAGE_HDR_SIZE;
4724
4725 if (!data_page)
4726 goto out;
4727
4728 bpage = *data_page;
4729 if (!bpage)
4730 goto out;
4731
4732 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4733
4734 reader = rb_get_reader_page(cpu_buffer);
4735 if (!reader)
4736 goto out_unlock;
4737
4738 event = rb_reader_event(cpu_buffer);
4739
4740 read = reader->read;
4741 commit = rb_page_commit(reader);
4742
4743 /* Check if any events were dropped */
4744 missed_events = cpu_buffer->lost_events;
4745
4746 /*
4747 * If this page has been partially read or
4748 * if len is not big enough to read the rest of the page or
4749 * a writer is still on the page, then
4750 * we must copy the data from the page to the buffer.
4751 * Otherwise, we can simply swap the page with the one passed in.
4752 */
4753 if (read || (len < (commit - read)) ||
4754 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4755 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4756 unsigned int rpos = read;
4757 unsigned int pos = 0;
4758 unsigned int size;
4759
4760 if (full)
4761 goto out_unlock;
4762
4763 if (len > (commit - read))
4764 len = (commit - read);
4765
4766 /* Always keep the time extend and data together */
4767 size = rb_event_ts_length(event);
4768
4769 if (len < size)
4770 goto out_unlock;
4771
4772 /* save the current timestamp, since the user will need it */
4773 save_timestamp = cpu_buffer->read_stamp;
4774
4775 /* Need to copy one event at a time */
4776 do {
4777 /* We need the size of one event, because
4778 * rb_advance_reader only advances by one event,
4779 * whereas rb_event_ts_length may include the size of
4780 * one or two events.
4781 * We have already ensured there's enough space if this
4782 * is a time extend. */
4783 size = rb_event_length(event);
4784 memcpy(bpage->data + pos, rpage->data + rpos, size);
4785
4786 len -= size;
4787
4788 rb_advance_reader(cpu_buffer);
4789 rpos = reader->read;
4790 pos += size;
4791
4792 if (rpos >= commit)
4793 break;
4794
4795 event = rb_reader_event(cpu_buffer);
4796 /* Always keep the time extend and data together */
4797 size = rb_event_ts_length(event);
4798 } while (len >= size);
4799
4800 /* update bpage */
4801 local_set(&bpage->commit, pos);
4802 bpage->time_stamp = save_timestamp;
4803
4804 /* we copied everything to the beginning */
4805 read = 0;
4806 } else {
4807 /* update the entry counter */
4808 cpu_buffer->read += rb_page_entries(reader);
4809 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4810
4811 /* swap the pages */
4812 rb_init_page(bpage);
4813 bpage = reader->page;
4814 reader->page = *data_page;
4815 local_set(&reader->write, 0);
4816 local_set(&reader->entries, 0);
4817 reader->read = 0;
4818 *data_page = bpage;
4819
4820 /*
4821 * Use the real_end for the data size,
4822 * This gives us a chance to store the lost events
4823 * on the page.
4824 */
4825 if (reader->real_end)
4826 local_set(&bpage->commit, reader->real_end);
4827 }
4828 ret = read;
4829
4830 cpu_buffer->lost_events = 0;
4831
4832 commit = local_read(&bpage->commit);
4833 /*
4834 * Set a flag in the commit field if we lost events
4835 */
4836 if (missed_events) {
4837 /* If there is room at the end of the page to save the
4838 * missed events, then record it there.
4839 */
4840 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4841 memcpy(&bpage->data[commit], &missed_events,
4842 sizeof(missed_events));
4843 local_add(RB_MISSED_STORED, &bpage->commit);
4844 commit += sizeof(missed_events);
4845 }
4846 local_add(RB_MISSED_EVENTS, &bpage->commit);
4847 }
4848
4849 /*
4850 * This page may be off to user land. Zero it out here.
4851 */
4852 if (commit < BUF_PAGE_SIZE)
4853 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4854
4855 out_unlock:
4856 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4857
4858 out:
4859 return ret;
4860}
4861EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4862
4863/*
4864 * We only allocate new buffers, never free them if the CPU goes down.
4865 * If we were to free the buffer, then the user would lose any trace that was in
4866 * the buffer.
4867 */
4868int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
4869{
4870 struct ring_buffer *buffer;
4871 long nr_pages_same;
4872 int cpu_i;
4873 unsigned long nr_pages;
4874
4875 buffer = container_of(node, struct ring_buffer, node);
4876 if (cpumask_test_cpu(cpu, buffer->cpumask))
4877 return 0;
4878
4879 nr_pages = 0;
4880 nr_pages_same = 1;
4881 /* check if all cpu sizes are same */
4882 for_each_buffer_cpu(buffer, cpu_i) {
4883 /* fill in the size from first enabled cpu */
4884 if (nr_pages == 0)
4885 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4886 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4887 nr_pages_same = 0;
4888 break;
4889 }
4890 }
4891 /* allocate minimum pages, user can later expand it */
4892 if (!nr_pages_same)
4893 nr_pages = 2;
4894 buffer->buffers[cpu] =
4895 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4896 if (!buffer->buffers[cpu]) {
4897 WARN(1, "failed to allocate ring buffer on CPU %u\n",
4898 cpu);
4899 return -ENOMEM;
4900 }
4901 smp_wmb();
4902 cpumask_set_cpu(cpu, buffer->cpumask);
4903 return 0;
4904}
4905
4906#ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4907/*
4908 * This is a basic integrity check of the ring buffer.
4909 * Late in the boot cycle this test will run when configured in.
4910 * It will kick off a thread per CPU that will go into a loop
4911 * writing to the per cpu ring buffer various sizes of data.
4912 * Some of the data will be large items, some small.
4913 *
4914 * Another thread is created that goes into a spin, sending out
4915 * IPIs to the other CPUs to also write into the ring buffer.
4916 * this is to test the nesting ability of the buffer.
4917 *
4918 * Basic stats are recorded and reported. If something in the
4919 * ring buffer should happen that's not expected, a big warning
4920 * is displayed and all ring buffers are disabled.
4921 */
4922static struct task_struct *rb_threads[NR_CPUS] __initdata;
4923
4924struct rb_test_data {
4925 struct ring_buffer *buffer;
4926 unsigned long events;
4927 unsigned long bytes_written;
4928 unsigned long bytes_alloc;
4929 unsigned long bytes_dropped;
4930 unsigned long events_nested;
4931 unsigned long bytes_written_nested;
4932 unsigned long bytes_alloc_nested;
4933 unsigned long bytes_dropped_nested;
4934 int min_size_nested;
4935 int max_size_nested;
4936 int max_size;
4937 int min_size;
4938 int cpu;
4939 int cnt;
4940};
4941
4942static struct rb_test_data rb_data[NR_CPUS] __initdata;
4943
4944/* 1 meg per cpu */
4945#define RB_TEST_BUFFER_SIZE 1048576
4946
4947static char rb_string[] __initdata =
4948 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4949 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4950 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4951
4952static bool rb_test_started __initdata;
4953
4954struct rb_item {
4955 int size;
4956 char str[];
4957};
4958
4959static __init int rb_write_something(struct rb_test_data *data, bool nested)
4960{
4961 struct ring_buffer_event *event;
4962 struct rb_item *item;
4963 bool started;
4964 int event_len;
4965 int size;
4966 int len;
4967 int cnt;
4968
4969 /* Have nested writes different that what is written */
4970 cnt = data->cnt + (nested ? 27 : 0);
4971
4972 /* Multiply cnt by ~e, to make some unique increment */
4973 size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
4974
4975 len = size + sizeof(struct rb_item);
4976
4977 started = rb_test_started;
4978 /* read rb_test_started before checking buffer enabled */
4979 smp_rmb();
4980
4981 event = ring_buffer_lock_reserve(data->buffer, len);
4982 if (!event) {
4983 /* Ignore dropped events before test starts. */
4984 if (started) {
4985 if (nested)
4986 data->bytes_dropped += len;
4987 else
4988 data->bytes_dropped_nested += len;
4989 }
4990 return len;
4991 }
4992
4993 event_len = ring_buffer_event_length(event);
4994
4995 if (RB_WARN_ON(data->buffer, event_len < len))
4996 goto out;
4997
4998 item = ring_buffer_event_data(event);
4999 item->size = size;
5000 memcpy(item->str, rb_string, size);
5001
5002 if (nested) {
5003 data->bytes_alloc_nested += event_len;
5004 data->bytes_written_nested += len;
5005 data->events_nested++;
5006 if (!data->min_size_nested || len < data->min_size_nested)
5007 data->min_size_nested = len;
5008 if (len > data->max_size_nested)
5009 data->max_size_nested = len;
5010 } else {
5011 data->bytes_alloc += event_len;
5012 data->bytes_written += len;
5013 data->events++;
5014 if (!data->min_size || len < data->min_size)
5015 data->max_size = len;
5016 if (len > data->max_size)
5017 data->max_size = len;
5018 }
5019
5020 out:
5021 ring_buffer_unlock_commit(data->buffer, event);
5022
5023 return 0;
5024}
5025
5026static __init int rb_test(void *arg)
5027{
5028 struct rb_test_data *data = arg;
5029
5030 while (!kthread_should_stop()) {
5031 rb_write_something(data, false);
5032 data->cnt++;
5033
5034 set_current_state(TASK_INTERRUPTIBLE);
5035 /* Now sleep between a min of 100-300us and a max of 1ms */
5036 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
5037 }
5038
5039 return 0;
5040}
5041
5042static __init void rb_ipi(void *ignore)
5043{
5044 struct rb_test_data *data;
5045 int cpu = smp_processor_id();
5046
5047 data = &rb_data[cpu];
5048 rb_write_something(data, true);
5049}
5050
5051static __init int rb_hammer_test(void *arg)
5052{
5053 while (!kthread_should_stop()) {
5054
5055 /* Send an IPI to all cpus to write data! */
5056 smp_call_function(rb_ipi, NULL, 1);
5057 /* No sleep, but for non preempt, let others run */
5058 schedule();
5059 }
5060
5061 return 0;
5062}
5063
5064static __init int test_ringbuffer(void)
5065{
5066 struct task_struct *rb_hammer;
5067 struct ring_buffer *buffer;
5068 int cpu;
5069 int ret = 0;
5070
5071 pr_info("Running ring buffer tests...\n");
5072
5073 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
5074 if (WARN_ON(!buffer))
5075 return 0;
5076
5077 /* Disable buffer so that threads can't write to it yet */
5078 ring_buffer_record_off(buffer);
5079
5080 for_each_online_cpu(cpu) {
5081 rb_data[cpu].buffer = buffer;
5082 rb_data[cpu].cpu = cpu;
5083 rb_data[cpu].cnt = cpu;
5084 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
5085 "rbtester/%d", cpu);
5086 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
5087 pr_cont("FAILED\n");
5088 ret = PTR_ERR(rb_threads[cpu]);
5089 goto out_free;
5090 }
5091
5092 kthread_bind(rb_threads[cpu], cpu);
5093 wake_up_process(rb_threads[cpu]);
5094 }
5095
5096 /* Now create the rb hammer! */
5097 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
5098 if (WARN_ON(IS_ERR(rb_hammer))) {
5099 pr_cont("FAILED\n");
5100 ret = PTR_ERR(rb_hammer);
5101 goto out_free;
5102 }
5103
5104 ring_buffer_record_on(buffer);
5105 /*
5106 * Show buffer is enabled before setting rb_test_started.
5107 * Yes there's a small race window where events could be
5108 * dropped and the thread wont catch it. But when a ring
5109 * buffer gets enabled, there will always be some kind of
5110 * delay before other CPUs see it. Thus, we don't care about
5111 * those dropped events. We care about events dropped after
5112 * the threads see that the buffer is active.
5113 */
5114 smp_wmb();
5115 rb_test_started = true;
5116
5117 set_current_state(TASK_INTERRUPTIBLE);
5118 /* Just run for 10 seconds */;
5119 schedule_timeout(10 * HZ);
5120
5121 kthread_stop(rb_hammer);
5122
5123 out_free:
5124 for_each_online_cpu(cpu) {
5125 if (!rb_threads[cpu])
5126 break;
5127 kthread_stop(rb_threads[cpu]);
5128 }
5129 if (ret) {
5130 ring_buffer_free(buffer);
5131 return ret;
5132 }
5133
5134 /* Report! */
5135 pr_info("finished\n");
5136 for_each_online_cpu(cpu) {
5137 struct ring_buffer_event *event;
5138 struct rb_test_data *data = &rb_data[cpu];
5139 struct rb_item *item;
5140 unsigned long total_events;
5141 unsigned long total_dropped;
5142 unsigned long total_written;
5143 unsigned long total_alloc;
5144 unsigned long total_read = 0;
5145 unsigned long total_size = 0;
5146 unsigned long total_len = 0;
5147 unsigned long total_lost = 0;
5148 unsigned long lost;
5149 int big_event_size;
5150 int small_event_size;
5151
5152 ret = -1;
5153
5154 total_events = data->events + data->events_nested;
5155 total_written = data->bytes_written + data->bytes_written_nested;
5156 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
5157 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
5158
5159 big_event_size = data->max_size + data->max_size_nested;
5160 small_event_size = data->min_size + data->min_size_nested;
5161
5162 pr_info("CPU %d:\n", cpu);
5163 pr_info(" events: %ld\n", total_events);
5164 pr_info(" dropped bytes: %ld\n", total_dropped);
5165 pr_info(" alloced bytes: %ld\n", total_alloc);
5166 pr_info(" written bytes: %ld\n", total_written);
5167 pr_info(" biggest event: %d\n", big_event_size);
5168 pr_info(" smallest event: %d\n", small_event_size);
5169
5170 if (RB_WARN_ON(buffer, total_dropped))
5171 break;
5172
5173 ret = 0;
5174
5175 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
5176 total_lost += lost;
5177 item = ring_buffer_event_data(event);
5178 total_len += ring_buffer_event_length(event);
5179 total_size += item->size + sizeof(struct rb_item);
5180 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
5181 pr_info("FAILED!\n");
5182 pr_info("buffer had: %.*s\n", item->size, item->str);
5183 pr_info("expected: %.*s\n", item->size, rb_string);
5184 RB_WARN_ON(buffer, 1);
5185 ret = -1;
5186 break;
5187 }
5188 total_read++;
5189 }
5190 if (ret)
5191 break;
5192
5193 ret = -1;
5194
5195 pr_info(" read events: %ld\n", total_read);
5196 pr_info(" lost events: %ld\n", total_lost);
5197 pr_info(" total events: %ld\n", total_lost + total_read);
5198 pr_info(" recorded len bytes: %ld\n", total_len);
5199 pr_info(" recorded size bytes: %ld\n", total_size);
5200 if (total_lost)
5201 pr_info(" With dropped events, record len and size may not match\n"
5202 " alloced and written from above\n");
5203 if (!total_lost) {
5204 if (RB_WARN_ON(buffer, total_len != total_alloc ||
5205 total_size != total_written))
5206 break;
5207 }
5208 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5209 break;
5210
5211 ret = 0;
5212 }
5213 if (!ret)
5214 pr_info("Ring buffer PASSED!\n");
5215
5216 ring_buffer_free(buffer);
5217 return 0;
5218}
5219
5220late_initcall(test_ringbuffer);
5221#endif /* CONFIG_RING_BUFFER_STARTUP_TEST */
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Generic ring buffer
4 *
5 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 */
7#include <linux/trace_recursion.h>
8#include <linux/trace_events.h>
9#include <linux/ring_buffer.h>
10#include <linux/trace_clock.h>
11#include <linux/sched/clock.h>
12#include <linux/trace_seq.h>
13#include <linux/spinlock.h>
14#include <linux/irq_work.h>
15#include <linux/security.h>
16#include <linux/uaccess.h>
17#include <linux/hardirq.h>
18#include <linux/kthread.h> /* for self test */
19#include <linux/module.h>
20#include <linux/percpu.h>
21#include <linux/mutex.h>
22#include <linux/delay.h>
23#include <linux/slab.h>
24#include <linux/init.h>
25#include <linux/hash.h>
26#include <linux/list.h>
27#include <linux/cpu.h>
28#include <linux/oom.h>
29
30#include <asm/local.h>
31
32/*
33 * The "absolute" timestamp in the buffer is only 59 bits.
34 * If a clock has the 5 MSBs set, it needs to be saved and
35 * reinserted.
36 */
37#define TS_MSB (0xf8ULL << 56)
38#define ABS_TS_MASK (~TS_MSB)
39
40static void update_pages_handler(struct work_struct *work);
41
42/*
43 * The ring buffer header is special. We must manually up keep it.
44 */
45int ring_buffer_print_entry_header(struct trace_seq *s)
46{
47 trace_seq_puts(s, "# compressed entry header\n");
48 trace_seq_puts(s, "\ttype_len : 5 bits\n");
49 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
50 trace_seq_puts(s, "\tarray : 32 bits\n");
51 trace_seq_putc(s, '\n');
52 trace_seq_printf(s, "\tpadding : type == %d\n",
53 RINGBUF_TYPE_PADDING);
54 trace_seq_printf(s, "\ttime_extend : type == %d\n",
55 RINGBUF_TYPE_TIME_EXTEND);
56 trace_seq_printf(s, "\ttime_stamp : type == %d\n",
57 RINGBUF_TYPE_TIME_STAMP);
58 trace_seq_printf(s, "\tdata max type_len == %d\n",
59 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
60
61 return !trace_seq_has_overflowed(s);
62}
63
64/*
65 * The ring buffer is made up of a list of pages. A separate list of pages is
66 * allocated for each CPU. A writer may only write to a buffer that is
67 * associated with the CPU it is currently executing on. A reader may read
68 * from any per cpu buffer.
69 *
70 * The reader is special. For each per cpu buffer, the reader has its own
71 * reader page. When a reader has read the entire reader page, this reader
72 * page is swapped with another page in the ring buffer.
73 *
74 * Now, as long as the writer is off the reader page, the reader can do what
75 * ever it wants with that page. The writer will never write to that page
76 * again (as long as it is out of the ring buffer).
77 *
78 * Here's some silly ASCII art.
79 *
80 * +------+
81 * |reader| RING BUFFER
82 * |page |
83 * +------+ +---+ +---+ +---+
84 * | |-->| |-->| |
85 * +---+ +---+ +---+
86 * ^ |
87 * | |
88 * +---------------+
89 *
90 *
91 * +------+
92 * |reader| RING BUFFER
93 * |page |------------------v
94 * +------+ +---+ +---+ +---+
95 * | |-->| |-->| |
96 * +---+ +---+ +---+
97 * ^ |
98 * | |
99 * +---------------+
100 *
101 *
102 * +------+
103 * |reader| RING BUFFER
104 * |page |------------------v
105 * +------+ +---+ +---+ +---+
106 * ^ | |-->| |-->| |
107 * | +---+ +---+ +---+
108 * | |
109 * | |
110 * +------------------------------+
111 *
112 *
113 * +------+
114 * |buffer| RING BUFFER
115 * |page |------------------v
116 * +------+ +---+ +---+ +---+
117 * ^ | | | |-->| |
118 * | New +---+ +---+ +---+
119 * | Reader------^ |
120 * | page |
121 * +------------------------------+
122 *
123 *
124 * After we make this swap, the reader can hand this page off to the splice
125 * code and be done with it. It can even allocate a new page if it needs to
126 * and swap that into the ring buffer.
127 *
128 * We will be using cmpxchg soon to make all this lockless.
129 *
130 */
131
132/* Used for individual buffers (after the counter) */
133#define RB_BUFFER_OFF (1 << 20)
134
135#define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
136
137#define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
138#define RB_ALIGNMENT 4U
139#define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
140#define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
141
142#ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
143# define RB_FORCE_8BYTE_ALIGNMENT 0
144# define RB_ARCH_ALIGNMENT RB_ALIGNMENT
145#else
146# define RB_FORCE_8BYTE_ALIGNMENT 1
147# define RB_ARCH_ALIGNMENT 8U
148#endif
149
150#define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
151
152/* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
153#define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
154
155enum {
156 RB_LEN_TIME_EXTEND = 8,
157 RB_LEN_TIME_STAMP = 8,
158};
159
160#define skip_time_extend(event) \
161 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
162
163#define extended_time(event) \
164 (event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
165
166static inline int rb_null_event(struct ring_buffer_event *event)
167{
168 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
169}
170
171static void rb_event_set_padding(struct ring_buffer_event *event)
172{
173 /* padding has a NULL time_delta */
174 event->type_len = RINGBUF_TYPE_PADDING;
175 event->time_delta = 0;
176}
177
178static unsigned
179rb_event_data_length(struct ring_buffer_event *event)
180{
181 unsigned length;
182
183 if (event->type_len)
184 length = event->type_len * RB_ALIGNMENT;
185 else
186 length = event->array[0];
187 return length + RB_EVNT_HDR_SIZE;
188}
189
190/*
191 * Return the length of the given event. Will return
192 * the length of the time extend if the event is a
193 * time extend.
194 */
195static inline unsigned
196rb_event_length(struct ring_buffer_event *event)
197{
198 switch (event->type_len) {
199 case RINGBUF_TYPE_PADDING:
200 if (rb_null_event(event))
201 /* undefined */
202 return -1;
203 return event->array[0] + RB_EVNT_HDR_SIZE;
204
205 case RINGBUF_TYPE_TIME_EXTEND:
206 return RB_LEN_TIME_EXTEND;
207
208 case RINGBUF_TYPE_TIME_STAMP:
209 return RB_LEN_TIME_STAMP;
210
211 case RINGBUF_TYPE_DATA:
212 return rb_event_data_length(event);
213 default:
214 WARN_ON_ONCE(1);
215 }
216 /* not hit */
217 return 0;
218}
219
220/*
221 * Return total length of time extend and data,
222 * or just the event length for all other events.
223 */
224static inline unsigned
225rb_event_ts_length(struct ring_buffer_event *event)
226{
227 unsigned len = 0;
228
229 if (extended_time(event)) {
230 /* time extends include the data event after it */
231 len = RB_LEN_TIME_EXTEND;
232 event = skip_time_extend(event);
233 }
234 return len + rb_event_length(event);
235}
236
237/**
238 * ring_buffer_event_length - return the length of the event
239 * @event: the event to get the length of
240 *
241 * Returns the size of the data load of a data event.
242 * If the event is something other than a data event, it
243 * returns the size of the event itself. With the exception
244 * of a TIME EXTEND, where it still returns the size of the
245 * data load of the data event after it.
246 */
247unsigned ring_buffer_event_length(struct ring_buffer_event *event)
248{
249 unsigned length;
250
251 if (extended_time(event))
252 event = skip_time_extend(event);
253
254 length = rb_event_length(event);
255 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
256 return length;
257 length -= RB_EVNT_HDR_SIZE;
258 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
259 length -= sizeof(event->array[0]);
260 return length;
261}
262EXPORT_SYMBOL_GPL(ring_buffer_event_length);
263
264/* inline for ring buffer fast paths */
265static __always_inline void *
266rb_event_data(struct ring_buffer_event *event)
267{
268 if (extended_time(event))
269 event = skip_time_extend(event);
270 WARN_ON_ONCE(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
271 /* If length is in len field, then array[0] has the data */
272 if (event->type_len)
273 return (void *)&event->array[0];
274 /* Otherwise length is in array[0] and array[1] has the data */
275 return (void *)&event->array[1];
276}
277
278/**
279 * ring_buffer_event_data - return the data of the event
280 * @event: the event to get the data from
281 */
282void *ring_buffer_event_data(struct ring_buffer_event *event)
283{
284 return rb_event_data(event);
285}
286EXPORT_SYMBOL_GPL(ring_buffer_event_data);
287
288#define for_each_buffer_cpu(buffer, cpu) \
289 for_each_cpu(cpu, buffer->cpumask)
290
291#define for_each_online_buffer_cpu(buffer, cpu) \
292 for_each_cpu_and(cpu, buffer->cpumask, cpu_online_mask)
293
294#define TS_SHIFT 27
295#define TS_MASK ((1ULL << TS_SHIFT) - 1)
296#define TS_DELTA_TEST (~TS_MASK)
297
298static u64 rb_event_time_stamp(struct ring_buffer_event *event)
299{
300 u64 ts;
301
302 ts = event->array[0];
303 ts <<= TS_SHIFT;
304 ts += event->time_delta;
305
306 return ts;
307}
308
309/* Flag when events were overwritten */
310#define RB_MISSED_EVENTS (1 << 31)
311/* Missed count stored at end */
312#define RB_MISSED_STORED (1 << 30)
313
314struct buffer_data_page {
315 u64 time_stamp; /* page time stamp */
316 local_t commit; /* write committed index */
317 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
318};
319
320/*
321 * Note, the buffer_page list must be first. The buffer pages
322 * are allocated in cache lines, which means that each buffer
323 * page will be at the beginning of a cache line, and thus
324 * the least significant bits will be zero. We use this to
325 * add flags in the list struct pointers, to make the ring buffer
326 * lockless.
327 */
328struct buffer_page {
329 struct list_head list; /* list of buffer pages */
330 local_t write; /* index for next write */
331 unsigned read; /* index for next read */
332 local_t entries; /* entries on this page */
333 unsigned long real_end; /* real end of data */
334 struct buffer_data_page *page; /* Actual data page */
335};
336
337/*
338 * The buffer page counters, write and entries, must be reset
339 * atomically when crossing page boundaries. To synchronize this
340 * update, two counters are inserted into the number. One is
341 * the actual counter for the write position or count on the page.
342 *
343 * The other is a counter of updaters. Before an update happens
344 * the update partition of the counter is incremented. This will
345 * allow the updater to update the counter atomically.
346 *
347 * The counter is 20 bits, and the state data is 12.
348 */
349#define RB_WRITE_MASK 0xfffff
350#define RB_WRITE_INTCNT (1 << 20)
351
352static void rb_init_page(struct buffer_data_page *bpage)
353{
354 local_set(&bpage->commit, 0);
355}
356
357/*
358 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
359 * this issue out.
360 */
361static void free_buffer_page(struct buffer_page *bpage)
362{
363 free_page((unsigned long)bpage->page);
364 kfree(bpage);
365}
366
367/*
368 * We need to fit the time_stamp delta into 27 bits.
369 */
370static inline int test_time_stamp(u64 delta)
371{
372 if (delta & TS_DELTA_TEST)
373 return 1;
374 return 0;
375}
376
377#define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
378
379/* Max payload is BUF_PAGE_SIZE - header (8bytes) */
380#define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
381
382int ring_buffer_print_page_header(struct trace_seq *s)
383{
384 struct buffer_data_page field;
385
386 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
387 "offset:0;\tsize:%u;\tsigned:%u;\n",
388 (unsigned int)sizeof(field.time_stamp),
389 (unsigned int)is_signed_type(u64));
390
391 trace_seq_printf(s, "\tfield: local_t commit;\t"
392 "offset:%u;\tsize:%u;\tsigned:%u;\n",
393 (unsigned int)offsetof(typeof(field), commit),
394 (unsigned int)sizeof(field.commit),
395 (unsigned int)is_signed_type(long));
396
397 trace_seq_printf(s, "\tfield: int overwrite;\t"
398 "offset:%u;\tsize:%u;\tsigned:%u;\n",
399 (unsigned int)offsetof(typeof(field), commit),
400 1,
401 (unsigned int)is_signed_type(long));
402
403 trace_seq_printf(s, "\tfield: char data;\t"
404 "offset:%u;\tsize:%u;\tsigned:%u;\n",
405 (unsigned int)offsetof(typeof(field), data),
406 (unsigned int)BUF_PAGE_SIZE,
407 (unsigned int)is_signed_type(char));
408
409 return !trace_seq_has_overflowed(s);
410}
411
412struct rb_irq_work {
413 struct irq_work work;
414 wait_queue_head_t waiters;
415 wait_queue_head_t full_waiters;
416 long wait_index;
417 bool waiters_pending;
418 bool full_waiters_pending;
419 bool wakeup_full;
420};
421
422/*
423 * Structure to hold event state and handle nested events.
424 */
425struct rb_event_info {
426 u64 ts;
427 u64 delta;
428 u64 before;
429 u64 after;
430 unsigned long length;
431 struct buffer_page *tail_page;
432 int add_timestamp;
433};
434
435/*
436 * Used for the add_timestamp
437 * NONE
438 * EXTEND - wants a time extend
439 * ABSOLUTE - the buffer requests all events to have absolute time stamps
440 * FORCE - force a full time stamp.
441 */
442enum {
443 RB_ADD_STAMP_NONE = 0,
444 RB_ADD_STAMP_EXTEND = BIT(1),
445 RB_ADD_STAMP_ABSOLUTE = BIT(2),
446 RB_ADD_STAMP_FORCE = BIT(3)
447};
448/*
449 * Used for which event context the event is in.
450 * TRANSITION = 0
451 * NMI = 1
452 * IRQ = 2
453 * SOFTIRQ = 3
454 * NORMAL = 4
455 *
456 * See trace_recursive_lock() comment below for more details.
457 */
458enum {
459 RB_CTX_TRANSITION,
460 RB_CTX_NMI,
461 RB_CTX_IRQ,
462 RB_CTX_SOFTIRQ,
463 RB_CTX_NORMAL,
464 RB_CTX_MAX
465};
466
467#if BITS_PER_LONG == 32
468#define RB_TIME_32
469#endif
470
471/* To test on 64 bit machines */
472//#define RB_TIME_32
473
474#ifdef RB_TIME_32
475
476struct rb_time_struct {
477 local_t cnt;
478 local_t top;
479 local_t bottom;
480 local_t msb;
481};
482#else
483#include <asm/local64.h>
484struct rb_time_struct {
485 local64_t time;
486};
487#endif
488typedef struct rb_time_struct rb_time_t;
489
490#define MAX_NEST 5
491
492/*
493 * head_page == tail_page && head == tail then buffer is empty.
494 */
495struct ring_buffer_per_cpu {
496 int cpu;
497 atomic_t record_disabled;
498 atomic_t resize_disabled;
499 struct trace_buffer *buffer;
500 raw_spinlock_t reader_lock; /* serialize readers */
501 arch_spinlock_t lock;
502 struct lock_class_key lock_key;
503 struct buffer_data_page *free_page;
504 unsigned long nr_pages;
505 unsigned int current_context;
506 struct list_head *pages;
507 struct buffer_page *head_page; /* read from head */
508 struct buffer_page *tail_page; /* write to tail */
509 struct buffer_page *commit_page; /* committed pages */
510 struct buffer_page *reader_page;
511 unsigned long lost_events;
512 unsigned long last_overrun;
513 unsigned long nest;
514 local_t entries_bytes;
515 local_t entries;
516 local_t overrun;
517 local_t commit_overrun;
518 local_t dropped_events;
519 local_t committing;
520 local_t commits;
521 local_t pages_touched;
522 local_t pages_lost;
523 local_t pages_read;
524 long last_pages_touch;
525 size_t shortest_full;
526 unsigned long read;
527 unsigned long read_bytes;
528 rb_time_t write_stamp;
529 rb_time_t before_stamp;
530 u64 event_stamp[MAX_NEST];
531 u64 read_stamp;
532 /* ring buffer pages to update, > 0 to add, < 0 to remove */
533 long nr_pages_to_update;
534 struct list_head new_pages; /* new pages to add */
535 struct work_struct update_pages_work;
536 struct completion update_done;
537
538 struct rb_irq_work irq_work;
539};
540
541struct trace_buffer {
542 unsigned flags;
543 int cpus;
544 atomic_t record_disabled;
545 cpumask_var_t cpumask;
546
547 struct lock_class_key *reader_lock_key;
548
549 struct mutex mutex;
550
551 struct ring_buffer_per_cpu **buffers;
552
553 struct hlist_node node;
554 u64 (*clock)(void);
555
556 struct rb_irq_work irq_work;
557 bool time_stamp_abs;
558};
559
560struct ring_buffer_iter {
561 struct ring_buffer_per_cpu *cpu_buffer;
562 unsigned long head;
563 unsigned long next_event;
564 struct buffer_page *head_page;
565 struct buffer_page *cache_reader_page;
566 unsigned long cache_read;
567 u64 read_stamp;
568 u64 page_stamp;
569 struct ring_buffer_event *event;
570 int missed_events;
571};
572
573#ifdef RB_TIME_32
574
575/*
576 * On 32 bit machines, local64_t is very expensive. As the ring
577 * buffer doesn't need all the features of a true 64 bit atomic,
578 * on 32 bit, it uses these functions (64 still uses local64_t).
579 *
580 * For the ring buffer, 64 bit required operations for the time is
581 * the following:
582 *
583 * - Reads may fail if it interrupted a modification of the time stamp.
584 * It will succeed if it did not interrupt another write even if
585 * the read itself is interrupted by a write.
586 * It returns whether it was successful or not.
587 *
588 * - Writes always succeed and will overwrite other writes and writes
589 * that were done by events interrupting the current write.
590 *
591 * - A write followed by a read of the same time stamp will always succeed,
592 * but may not contain the same value.
593 *
594 * - A cmpxchg will fail if it interrupted another write or cmpxchg.
595 * Other than that, it acts like a normal cmpxchg.
596 *
597 * The 60 bit time stamp is broken up by 30 bits in a top and bottom half
598 * (bottom being the least significant 30 bits of the 60 bit time stamp).
599 *
600 * The two most significant bits of each half holds a 2 bit counter (0-3).
601 * Each update will increment this counter by one.
602 * When reading the top and bottom, if the two counter bits match then the
603 * top and bottom together make a valid 60 bit number.
604 */
605#define RB_TIME_SHIFT 30
606#define RB_TIME_VAL_MASK ((1 << RB_TIME_SHIFT) - 1)
607#define RB_TIME_MSB_SHIFT 60
608
609static inline int rb_time_cnt(unsigned long val)
610{
611 return (val >> RB_TIME_SHIFT) & 3;
612}
613
614static inline u64 rb_time_val(unsigned long top, unsigned long bottom)
615{
616 u64 val;
617
618 val = top & RB_TIME_VAL_MASK;
619 val <<= RB_TIME_SHIFT;
620 val |= bottom & RB_TIME_VAL_MASK;
621
622 return val;
623}
624
625static inline bool __rb_time_read(rb_time_t *t, u64 *ret, unsigned long *cnt)
626{
627 unsigned long top, bottom, msb;
628 unsigned long c;
629
630 /*
631 * If the read is interrupted by a write, then the cnt will
632 * be different. Loop until both top and bottom have been read
633 * without interruption.
634 */
635 do {
636 c = local_read(&t->cnt);
637 top = local_read(&t->top);
638 bottom = local_read(&t->bottom);
639 msb = local_read(&t->msb);
640 } while (c != local_read(&t->cnt));
641
642 *cnt = rb_time_cnt(top);
643
644 /* If top and bottom counts don't match, this interrupted a write */
645 if (*cnt != rb_time_cnt(bottom))
646 return false;
647
648 /* The shift to msb will lose its cnt bits */
649 *ret = rb_time_val(top, bottom) | ((u64)msb << RB_TIME_MSB_SHIFT);
650 return true;
651}
652
653static bool rb_time_read(rb_time_t *t, u64 *ret)
654{
655 unsigned long cnt;
656
657 return __rb_time_read(t, ret, &cnt);
658}
659
660static inline unsigned long rb_time_val_cnt(unsigned long val, unsigned long cnt)
661{
662 return (val & RB_TIME_VAL_MASK) | ((cnt & 3) << RB_TIME_SHIFT);
663}
664
665static inline void rb_time_split(u64 val, unsigned long *top, unsigned long *bottom,
666 unsigned long *msb)
667{
668 *top = (unsigned long)((val >> RB_TIME_SHIFT) & RB_TIME_VAL_MASK);
669 *bottom = (unsigned long)(val & RB_TIME_VAL_MASK);
670 *msb = (unsigned long)(val >> RB_TIME_MSB_SHIFT);
671}
672
673static inline void rb_time_val_set(local_t *t, unsigned long val, unsigned long cnt)
674{
675 val = rb_time_val_cnt(val, cnt);
676 local_set(t, val);
677}
678
679static void rb_time_set(rb_time_t *t, u64 val)
680{
681 unsigned long cnt, top, bottom, msb;
682
683 rb_time_split(val, &top, &bottom, &msb);
684
685 /* Writes always succeed with a valid number even if it gets interrupted. */
686 do {
687 cnt = local_inc_return(&t->cnt);
688 rb_time_val_set(&t->top, top, cnt);
689 rb_time_val_set(&t->bottom, bottom, cnt);
690 rb_time_val_set(&t->msb, val >> RB_TIME_MSB_SHIFT, cnt);
691 } while (cnt != local_read(&t->cnt));
692}
693
694static inline bool
695rb_time_read_cmpxchg(local_t *l, unsigned long expect, unsigned long set)
696{
697 unsigned long ret;
698
699 ret = local_cmpxchg(l, expect, set);
700 return ret == expect;
701}
702
703static int rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
704{
705 unsigned long cnt, top, bottom, msb;
706 unsigned long cnt2, top2, bottom2, msb2;
707 u64 val;
708
709 /* The cmpxchg always fails if it interrupted an update */
710 if (!__rb_time_read(t, &val, &cnt2))
711 return false;
712
713 if (val != expect)
714 return false;
715
716 cnt = local_read(&t->cnt);
717 if ((cnt & 3) != cnt2)
718 return false;
719
720 cnt2 = cnt + 1;
721
722 rb_time_split(val, &top, &bottom, &msb);
723 top = rb_time_val_cnt(top, cnt);
724 bottom = rb_time_val_cnt(bottom, cnt);
725
726 rb_time_split(set, &top2, &bottom2, &msb2);
727 top2 = rb_time_val_cnt(top2, cnt2);
728 bottom2 = rb_time_val_cnt(bottom2, cnt2);
729
730 if (!rb_time_read_cmpxchg(&t->cnt, cnt, cnt2))
731 return false;
732 if (!rb_time_read_cmpxchg(&t->msb, msb, msb2))
733 return false;
734 if (!rb_time_read_cmpxchg(&t->top, top, top2))
735 return false;
736 if (!rb_time_read_cmpxchg(&t->bottom, bottom, bottom2))
737 return false;
738 return true;
739}
740
741#else /* 64 bits */
742
743/* local64_t always succeeds */
744
745static inline bool rb_time_read(rb_time_t *t, u64 *ret)
746{
747 *ret = local64_read(&t->time);
748 return true;
749}
750static void rb_time_set(rb_time_t *t, u64 val)
751{
752 local64_set(&t->time, val);
753}
754
755static bool rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
756{
757 u64 val;
758 val = local64_cmpxchg(&t->time, expect, set);
759 return val == expect;
760}
761#endif
762
763/*
764 * Enable this to make sure that the event passed to
765 * ring_buffer_event_time_stamp() is not committed and also
766 * is on the buffer that it passed in.
767 */
768//#define RB_VERIFY_EVENT
769#ifdef RB_VERIFY_EVENT
770static struct list_head *rb_list_head(struct list_head *list);
771static void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
772 void *event)
773{
774 struct buffer_page *page = cpu_buffer->commit_page;
775 struct buffer_page *tail_page = READ_ONCE(cpu_buffer->tail_page);
776 struct list_head *next;
777 long commit, write;
778 unsigned long addr = (unsigned long)event;
779 bool done = false;
780 int stop = 0;
781
782 /* Make sure the event exists and is not committed yet */
783 do {
784 if (page == tail_page || WARN_ON_ONCE(stop++ > 100))
785 done = true;
786 commit = local_read(&page->page->commit);
787 write = local_read(&page->write);
788 if (addr >= (unsigned long)&page->page->data[commit] &&
789 addr < (unsigned long)&page->page->data[write])
790 return;
791
792 next = rb_list_head(page->list.next);
793 page = list_entry(next, struct buffer_page, list);
794 } while (!done);
795 WARN_ON_ONCE(1);
796}
797#else
798static inline void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
799 void *event)
800{
801}
802#endif
803
804/*
805 * The absolute time stamp drops the 5 MSBs and some clocks may
806 * require them. The rb_fix_abs_ts() will take a previous full
807 * time stamp, and add the 5 MSB of that time stamp on to the
808 * saved absolute time stamp. Then they are compared in case of
809 * the unlikely event that the latest time stamp incremented
810 * the 5 MSB.
811 */
812static inline u64 rb_fix_abs_ts(u64 abs, u64 save_ts)
813{
814 if (save_ts & TS_MSB) {
815 abs |= save_ts & TS_MSB;
816 /* Check for overflow */
817 if (unlikely(abs < save_ts))
818 abs += 1ULL << 59;
819 }
820 return abs;
821}
822
823static inline u64 rb_time_stamp(struct trace_buffer *buffer);
824
825/**
826 * ring_buffer_event_time_stamp - return the event's current time stamp
827 * @buffer: The buffer that the event is on
828 * @event: the event to get the time stamp of
829 *
830 * Note, this must be called after @event is reserved, and before it is
831 * committed to the ring buffer. And must be called from the same
832 * context where the event was reserved (normal, softirq, irq, etc).
833 *
834 * Returns the time stamp associated with the current event.
835 * If the event has an extended time stamp, then that is used as
836 * the time stamp to return.
837 * In the highly unlikely case that the event was nested more than
838 * the max nesting, then the write_stamp of the buffer is returned,
839 * otherwise current time is returned, but that really neither of
840 * the last two cases should ever happen.
841 */
842u64 ring_buffer_event_time_stamp(struct trace_buffer *buffer,
843 struct ring_buffer_event *event)
844{
845 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[smp_processor_id()];
846 unsigned int nest;
847 u64 ts;
848
849 /* If the event includes an absolute time, then just use that */
850 if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
851 ts = rb_event_time_stamp(event);
852 return rb_fix_abs_ts(ts, cpu_buffer->tail_page->page->time_stamp);
853 }
854
855 nest = local_read(&cpu_buffer->committing);
856 verify_event(cpu_buffer, event);
857 if (WARN_ON_ONCE(!nest))
858 goto fail;
859
860 /* Read the current saved nesting level time stamp */
861 if (likely(--nest < MAX_NEST))
862 return cpu_buffer->event_stamp[nest];
863
864 /* Shouldn't happen, warn if it does */
865 WARN_ONCE(1, "nest (%d) greater than max", nest);
866
867 fail:
868 /* Can only fail on 32 bit */
869 if (!rb_time_read(&cpu_buffer->write_stamp, &ts))
870 /* Screw it, just read the current time */
871 ts = rb_time_stamp(cpu_buffer->buffer);
872
873 return ts;
874}
875
876/**
877 * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
878 * @buffer: The ring_buffer to get the number of pages from
879 * @cpu: The cpu of the ring_buffer to get the number of pages from
880 *
881 * Returns the number of pages used by a per_cpu buffer of the ring buffer.
882 */
883size_t ring_buffer_nr_pages(struct trace_buffer *buffer, int cpu)
884{
885 return buffer->buffers[cpu]->nr_pages;
886}
887
888/**
889 * ring_buffer_nr_dirty_pages - get the number of used pages in the ring buffer
890 * @buffer: The ring_buffer to get the number of pages from
891 * @cpu: The cpu of the ring_buffer to get the number of pages from
892 *
893 * Returns the number of pages that have content in the ring buffer.
894 */
895size_t ring_buffer_nr_dirty_pages(struct trace_buffer *buffer, int cpu)
896{
897 size_t read;
898 size_t lost;
899 size_t cnt;
900
901 read = local_read(&buffer->buffers[cpu]->pages_read);
902 lost = local_read(&buffer->buffers[cpu]->pages_lost);
903 cnt = local_read(&buffer->buffers[cpu]->pages_touched);
904
905 if (WARN_ON_ONCE(cnt < lost))
906 return 0;
907
908 cnt -= lost;
909
910 /* The reader can read an empty page, but not more than that */
911 if (cnt < read) {
912 WARN_ON_ONCE(read > cnt + 1);
913 return 0;
914 }
915
916 return cnt - read;
917}
918
919static __always_inline bool full_hit(struct trace_buffer *buffer, int cpu, int full)
920{
921 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
922 size_t nr_pages;
923 size_t dirty;
924
925 nr_pages = cpu_buffer->nr_pages;
926 if (!nr_pages || !full)
927 return true;
928
929 dirty = ring_buffer_nr_dirty_pages(buffer, cpu);
930
931 return (dirty * 100) > (full * nr_pages);
932}
933
934/*
935 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
936 *
937 * Schedules a delayed work to wake up any task that is blocked on the
938 * ring buffer waiters queue.
939 */
940static void rb_wake_up_waiters(struct irq_work *work)
941{
942 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
943
944 wake_up_all(&rbwork->waiters);
945 if (rbwork->full_waiters_pending || rbwork->wakeup_full) {
946 rbwork->wakeup_full = false;
947 rbwork->full_waiters_pending = false;
948 wake_up_all(&rbwork->full_waiters);
949 }
950}
951
952/**
953 * ring_buffer_wake_waiters - wake up any waiters on this ring buffer
954 * @buffer: The ring buffer to wake waiters on
955 *
956 * In the case of a file that represents a ring buffer is closing,
957 * it is prudent to wake up any waiters that are on this.
958 */
959void ring_buffer_wake_waiters(struct trace_buffer *buffer, int cpu)
960{
961 struct ring_buffer_per_cpu *cpu_buffer;
962 struct rb_irq_work *rbwork;
963
964 if (!buffer)
965 return;
966
967 if (cpu == RING_BUFFER_ALL_CPUS) {
968
969 /* Wake up individual ones too. One level recursion */
970 for_each_buffer_cpu(buffer, cpu)
971 ring_buffer_wake_waiters(buffer, cpu);
972
973 rbwork = &buffer->irq_work;
974 } else {
975 if (WARN_ON_ONCE(!buffer->buffers))
976 return;
977 if (WARN_ON_ONCE(cpu >= nr_cpu_ids))
978 return;
979
980 cpu_buffer = buffer->buffers[cpu];
981 /* The CPU buffer may not have been initialized yet */
982 if (!cpu_buffer)
983 return;
984 rbwork = &cpu_buffer->irq_work;
985 }
986
987 rbwork->wait_index++;
988 /* make sure the waiters see the new index */
989 smp_wmb();
990
991 rb_wake_up_waiters(&rbwork->work);
992}
993
994/**
995 * ring_buffer_wait - wait for input to the ring buffer
996 * @buffer: buffer to wait on
997 * @cpu: the cpu buffer to wait on
998 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
999 *
1000 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
1001 * as data is added to any of the @buffer's cpu buffers. Otherwise
1002 * it will wait for data to be added to a specific cpu buffer.
1003 */
1004int ring_buffer_wait(struct trace_buffer *buffer, int cpu, int full)
1005{
1006 struct ring_buffer_per_cpu *cpu_buffer;
1007 DEFINE_WAIT(wait);
1008 struct rb_irq_work *work;
1009 long wait_index;
1010 int ret = 0;
1011
1012 /*
1013 * Depending on what the caller is waiting for, either any
1014 * data in any cpu buffer, or a specific buffer, put the
1015 * caller on the appropriate wait queue.
1016 */
1017 if (cpu == RING_BUFFER_ALL_CPUS) {
1018 work = &buffer->irq_work;
1019 /* Full only makes sense on per cpu reads */
1020 full = 0;
1021 } else {
1022 if (!cpumask_test_cpu(cpu, buffer->cpumask))
1023 return -ENODEV;
1024 cpu_buffer = buffer->buffers[cpu];
1025 work = &cpu_buffer->irq_work;
1026 }
1027
1028 wait_index = READ_ONCE(work->wait_index);
1029
1030 while (true) {
1031 if (full)
1032 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
1033 else
1034 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
1035
1036 /*
1037 * The events can happen in critical sections where
1038 * checking a work queue can cause deadlocks.
1039 * After adding a task to the queue, this flag is set
1040 * only to notify events to try to wake up the queue
1041 * using irq_work.
1042 *
1043 * We don't clear it even if the buffer is no longer
1044 * empty. The flag only causes the next event to run
1045 * irq_work to do the work queue wake up. The worse
1046 * that can happen if we race with !trace_empty() is that
1047 * an event will cause an irq_work to try to wake up
1048 * an empty queue.
1049 *
1050 * There's no reason to protect this flag either, as
1051 * the work queue and irq_work logic will do the necessary
1052 * synchronization for the wake ups. The only thing
1053 * that is necessary is that the wake up happens after
1054 * a task has been queued. It's OK for spurious wake ups.
1055 */
1056 if (full)
1057 work->full_waiters_pending = true;
1058 else
1059 work->waiters_pending = true;
1060
1061 if (signal_pending(current)) {
1062 ret = -EINTR;
1063 break;
1064 }
1065
1066 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
1067 break;
1068
1069 if (cpu != RING_BUFFER_ALL_CPUS &&
1070 !ring_buffer_empty_cpu(buffer, cpu)) {
1071 unsigned long flags;
1072 bool pagebusy;
1073 bool done;
1074
1075 if (!full)
1076 break;
1077
1078 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
1079 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
1080 done = !pagebusy && full_hit(buffer, cpu, full);
1081
1082 if (!cpu_buffer->shortest_full ||
1083 cpu_buffer->shortest_full > full)
1084 cpu_buffer->shortest_full = full;
1085 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
1086 if (done)
1087 break;
1088 }
1089
1090 schedule();
1091
1092 /* Make sure to see the new wait index */
1093 smp_rmb();
1094 if (wait_index != work->wait_index)
1095 break;
1096 }
1097
1098 if (full)
1099 finish_wait(&work->full_waiters, &wait);
1100 else
1101 finish_wait(&work->waiters, &wait);
1102
1103 return ret;
1104}
1105
1106/**
1107 * ring_buffer_poll_wait - poll on buffer input
1108 * @buffer: buffer to wait on
1109 * @cpu: the cpu buffer to wait on
1110 * @filp: the file descriptor
1111 * @poll_table: The poll descriptor
1112 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
1113 *
1114 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
1115 * as data is added to any of the @buffer's cpu buffers. Otherwise
1116 * it will wait for data to be added to a specific cpu buffer.
1117 *
1118 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
1119 * zero otherwise.
1120 */
1121__poll_t ring_buffer_poll_wait(struct trace_buffer *buffer, int cpu,
1122 struct file *filp, poll_table *poll_table, int full)
1123{
1124 struct ring_buffer_per_cpu *cpu_buffer;
1125 struct rb_irq_work *work;
1126
1127 if (cpu == RING_BUFFER_ALL_CPUS) {
1128 work = &buffer->irq_work;
1129 full = 0;
1130 } else {
1131 if (!cpumask_test_cpu(cpu, buffer->cpumask))
1132 return -EINVAL;
1133
1134 cpu_buffer = buffer->buffers[cpu];
1135 work = &cpu_buffer->irq_work;
1136 }
1137
1138 if (full) {
1139 poll_wait(filp, &work->full_waiters, poll_table);
1140 work->full_waiters_pending = true;
1141 } else {
1142 poll_wait(filp, &work->waiters, poll_table);
1143 work->waiters_pending = true;
1144 }
1145
1146 /*
1147 * There's a tight race between setting the waiters_pending and
1148 * checking if the ring buffer is empty. Once the waiters_pending bit
1149 * is set, the next event will wake the task up, but we can get stuck
1150 * if there's only a single event in.
1151 *
1152 * FIXME: Ideally, we need a memory barrier on the writer side as well,
1153 * but adding a memory barrier to all events will cause too much of a
1154 * performance hit in the fast path. We only need a memory barrier when
1155 * the buffer goes from empty to having content. But as this race is
1156 * extremely small, and it's not a problem if another event comes in, we
1157 * will fix it later.
1158 */
1159 smp_mb();
1160
1161 if (full)
1162 return full_hit(buffer, cpu, full) ? EPOLLIN | EPOLLRDNORM : 0;
1163
1164 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
1165 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
1166 return EPOLLIN | EPOLLRDNORM;
1167 return 0;
1168}
1169
1170/* buffer may be either ring_buffer or ring_buffer_per_cpu */
1171#define RB_WARN_ON(b, cond) \
1172 ({ \
1173 int _____ret = unlikely(cond); \
1174 if (_____ret) { \
1175 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
1176 struct ring_buffer_per_cpu *__b = \
1177 (void *)b; \
1178 atomic_inc(&__b->buffer->record_disabled); \
1179 } else \
1180 atomic_inc(&b->record_disabled); \
1181 WARN_ON(1); \
1182 } \
1183 _____ret; \
1184 })
1185
1186/* Up this if you want to test the TIME_EXTENTS and normalization */
1187#define DEBUG_SHIFT 0
1188
1189static inline u64 rb_time_stamp(struct trace_buffer *buffer)
1190{
1191 u64 ts;
1192
1193 /* Skip retpolines :-( */
1194 if (IS_ENABLED(CONFIG_RETPOLINE) && likely(buffer->clock == trace_clock_local))
1195 ts = trace_clock_local();
1196 else
1197 ts = buffer->clock();
1198
1199 /* shift to debug/test normalization and TIME_EXTENTS */
1200 return ts << DEBUG_SHIFT;
1201}
1202
1203u64 ring_buffer_time_stamp(struct trace_buffer *buffer)
1204{
1205 u64 time;
1206
1207 preempt_disable_notrace();
1208 time = rb_time_stamp(buffer);
1209 preempt_enable_notrace();
1210
1211 return time;
1212}
1213EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
1214
1215void ring_buffer_normalize_time_stamp(struct trace_buffer *buffer,
1216 int cpu, u64 *ts)
1217{
1218 /* Just stupid testing the normalize function and deltas */
1219 *ts >>= DEBUG_SHIFT;
1220}
1221EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
1222
1223/*
1224 * Making the ring buffer lockless makes things tricky.
1225 * Although writes only happen on the CPU that they are on,
1226 * and they only need to worry about interrupts. Reads can
1227 * happen on any CPU.
1228 *
1229 * The reader page is always off the ring buffer, but when the
1230 * reader finishes with a page, it needs to swap its page with
1231 * a new one from the buffer. The reader needs to take from
1232 * the head (writes go to the tail). But if a writer is in overwrite
1233 * mode and wraps, it must push the head page forward.
1234 *
1235 * Here lies the problem.
1236 *
1237 * The reader must be careful to replace only the head page, and
1238 * not another one. As described at the top of the file in the
1239 * ASCII art, the reader sets its old page to point to the next
1240 * page after head. It then sets the page after head to point to
1241 * the old reader page. But if the writer moves the head page
1242 * during this operation, the reader could end up with the tail.
1243 *
1244 * We use cmpxchg to help prevent this race. We also do something
1245 * special with the page before head. We set the LSB to 1.
1246 *
1247 * When the writer must push the page forward, it will clear the
1248 * bit that points to the head page, move the head, and then set
1249 * the bit that points to the new head page.
1250 *
1251 * We also don't want an interrupt coming in and moving the head
1252 * page on another writer. Thus we use the second LSB to catch
1253 * that too. Thus:
1254 *
1255 * head->list->prev->next bit 1 bit 0
1256 * ------- -------
1257 * Normal page 0 0
1258 * Points to head page 0 1
1259 * New head page 1 0
1260 *
1261 * Note we can not trust the prev pointer of the head page, because:
1262 *
1263 * +----+ +-----+ +-----+
1264 * | |------>| T |---X--->| N |
1265 * | |<------| | | |
1266 * +----+ +-----+ +-----+
1267 * ^ ^ |
1268 * | +-----+ | |
1269 * +----------| R |----------+ |
1270 * | |<-----------+
1271 * +-----+
1272 *
1273 * Key: ---X--> HEAD flag set in pointer
1274 * T Tail page
1275 * R Reader page
1276 * N Next page
1277 *
1278 * (see __rb_reserve_next() to see where this happens)
1279 *
1280 * What the above shows is that the reader just swapped out
1281 * the reader page with a page in the buffer, but before it
1282 * could make the new header point back to the new page added
1283 * it was preempted by a writer. The writer moved forward onto
1284 * the new page added by the reader and is about to move forward
1285 * again.
1286 *
1287 * You can see, it is legitimate for the previous pointer of
1288 * the head (or any page) not to point back to itself. But only
1289 * temporarily.
1290 */
1291
1292#define RB_PAGE_NORMAL 0UL
1293#define RB_PAGE_HEAD 1UL
1294#define RB_PAGE_UPDATE 2UL
1295
1296
1297#define RB_FLAG_MASK 3UL
1298
1299/* PAGE_MOVED is not part of the mask */
1300#define RB_PAGE_MOVED 4UL
1301
1302/*
1303 * rb_list_head - remove any bit
1304 */
1305static struct list_head *rb_list_head(struct list_head *list)
1306{
1307 unsigned long val = (unsigned long)list;
1308
1309 return (struct list_head *)(val & ~RB_FLAG_MASK);
1310}
1311
1312/*
1313 * rb_is_head_page - test if the given page is the head page
1314 *
1315 * Because the reader may move the head_page pointer, we can
1316 * not trust what the head page is (it may be pointing to
1317 * the reader page). But if the next page is a header page,
1318 * its flags will be non zero.
1319 */
1320static inline int
1321rb_is_head_page(struct buffer_page *page, struct list_head *list)
1322{
1323 unsigned long val;
1324
1325 val = (unsigned long)list->next;
1326
1327 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
1328 return RB_PAGE_MOVED;
1329
1330 return val & RB_FLAG_MASK;
1331}
1332
1333/*
1334 * rb_is_reader_page
1335 *
1336 * The unique thing about the reader page, is that, if the
1337 * writer is ever on it, the previous pointer never points
1338 * back to the reader page.
1339 */
1340static bool rb_is_reader_page(struct buffer_page *page)
1341{
1342 struct list_head *list = page->list.prev;
1343
1344 return rb_list_head(list->next) != &page->list;
1345}
1346
1347/*
1348 * rb_set_list_to_head - set a list_head to be pointing to head.
1349 */
1350static void rb_set_list_to_head(struct list_head *list)
1351{
1352 unsigned long *ptr;
1353
1354 ptr = (unsigned long *)&list->next;
1355 *ptr |= RB_PAGE_HEAD;
1356 *ptr &= ~RB_PAGE_UPDATE;
1357}
1358
1359/*
1360 * rb_head_page_activate - sets up head page
1361 */
1362static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
1363{
1364 struct buffer_page *head;
1365
1366 head = cpu_buffer->head_page;
1367 if (!head)
1368 return;
1369
1370 /*
1371 * Set the previous list pointer to have the HEAD flag.
1372 */
1373 rb_set_list_to_head(head->list.prev);
1374}
1375
1376static void rb_list_head_clear(struct list_head *list)
1377{
1378 unsigned long *ptr = (unsigned long *)&list->next;
1379
1380 *ptr &= ~RB_FLAG_MASK;
1381}
1382
1383/*
1384 * rb_head_page_deactivate - clears head page ptr (for free list)
1385 */
1386static void
1387rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
1388{
1389 struct list_head *hd;
1390
1391 /* Go through the whole list and clear any pointers found. */
1392 rb_list_head_clear(cpu_buffer->pages);
1393
1394 list_for_each(hd, cpu_buffer->pages)
1395 rb_list_head_clear(hd);
1396}
1397
1398static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
1399 struct buffer_page *head,
1400 struct buffer_page *prev,
1401 int old_flag, int new_flag)
1402{
1403 struct list_head *list;
1404 unsigned long val = (unsigned long)&head->list;
1405 unsigned long ret;
1406
1407 list = &prev->list;
1408
1409 val &= ~RB_FLAG_MASK;
1410
1411 ret = cmpxchg((unsigned long *)&list->next,
1412 val | old_flag, val | new_flag);
1413
1414 /* check if the reader took the page */
1415 if ((ret & ~RB_FLAG_MASK) != val)
1416 return RB_PAGE_MOVED;
1417
1418 return ret & RB_FLAG_MASK;
1419}
1420
1421static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
1422 struct buffer_page *head,
1423 struct buffer_page *prev,
1424 int old_flag)
1425{
1426 return rb_head_page_set(cpu_buffer, head, prev,
1427 old_flag, RB_PAGE_UPDATE);
1428}
1429
1430static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
1431 struct buffer_page *head,
1432 struct buffer_page *prev,
1433 int old_flag)
1434{
1435 return rb_head_page_set(cpu_buffer, head, prev,
1436 old_flag, RB_PAGE_HEAD);
1437}
1438
1439static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
1440 struct buffer_page *head,
1441 struct buffer_page *prev,
1442 int old_flag)
1443{
1444 return rb_head_page_set(cpu_buffer, head, prev,
1445 old_flag, RB_PAGE_NORMAL);
1446}
1447
1448static inline void rb_inc_page(struct buffer_page **bpage)
1449{
1450 struct list_head *p = rb_list_head((*bpage)->list.next);
1451
1452 *bpage = list_entry(p, struct buffer_page, list);
1453}
1454
1455static struct buffer_page *
1456rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1457{
1458 struct buffer_page *head;
1459 struct buffer_page *page;
1460 struct list_head *list;
1461 int i;
1462
1463 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1464 return NULL;
1465
1466 /* sanity check */
1467 list = cpu_buffer->pages;
1468 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1469 return NULL;
1470
1471 page = head = cpu_buffer->head_page;
1472 /*
1473 * It is possible that the writer moves the header behind
1474 * where we started, and we miss in one loop.
1475 * A second loop should grab the header, but we'll do
1476 * three loops just because I'm paranoid.
1477 */
1478 for (i = 0; i < 3; i++) {
1479 do {
1480 if (rb_is_head_page(page, page->list.prev)) {
1481 cpu_buffer->head_page = page;
1482 return page;
1483 }
1484 rb_inc_page(&page);
1485 } while (page != head);
1486 }
1487
1488 RB_WARN_ON(cpu_buffer, 1);
1489
1490 return NULL;
1491}
1492
1493static int rb_head_page_replace(struct buffer_page *old,
1494 struct buffer_page *new)
1495{
1496 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1497 unsigned long val;
1498 unsigned long ret;
1499
1500 val = *ptr & ~RB_FLAG_MASK;
1501 val |= RB_PAGE_HEAD;
1502
1503 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1504
1505 return ret == val;
1506}
1507
1508/*
1509 * rb_tail_page_update - move the tail page forward
1510 */
1511static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1512 struct buffer_page *tail_page,
1513 struct buffer_page *next_page)
1514{
1515 unsigned long old_entries;
1516 unsigned long old_write;
1517
1518 /*
1519 * The tail page now needs to be moved forward.
1520 *
1521 * We need to reset the tail page, but without messing
1522 * with possible erasing of data brought in by interrupts
1523 * that have moved the tail page and are currently on it.
1524 *
1525 * We add a counter to the write field to denote this.
1526 */
1527 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1528 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1529
1530 local_inc(&cpu_buffer->pages_touched);
1531 /*
1532 * Just make sure we have seen our old_write and synchronize
1533 * with any interrupts that come in.
1534 */
1535 barrier();
1536
1537 /*
1538 * If the tail page is still the same as what we think
1539 * it is, then it is up to us to update the tail
1540 * pointer.
1541 */
1542 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1543 /* Zero the write counter */
1544 unsigned long val = old_write & ~RB_WRITE_MASK;
1545 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1546
1547 /*
1548 * This will only succeed if an interrupt did
1549 * not come in and change it. In which case, we
1550 * do not want to modify it.
1551 *
1552 * We add (void) to let the compiler know that we do not care
1553 * about the return value of these functions. We use the
1554 * cmpxchg to only update if an interrupt did not already
1555 * do it for us. If the cmpxchg fails, we don't care.
1556 */
1557 (void)local_cmpxchg(&next_page->write, old_write, val);
1558 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1559
1560 /*
1561 * No need to worry about races with clearing out the commit.
1562 * it only can increment when a commit takes place. But that
1563 * only happens in the outer most nested commit.
1564 */
1565 local_set(&next_page->page->commit, 0);
1566
1567 /* Again, either we update tail_page or an interrupt does */
1568 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1569 }
1570}
1571
1572static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1573 struct buffer_page *bpage)
1574{
1575 unsigned long val = (unsigned long)bpage;
1576
1577 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1578 return 1;
1579
1580 return 0;
1581}
1582
1583/**
1584 * rb_check_list - make sure a pointer to a list has the last bits zero
1585 */
1586static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1587 struct list_head *list)
1588{
1589 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1590 return 1;
1591 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1592 return 1;
1593 return 0;
1594}
1595
1596/**
1597 * rb_check_pages - integrity check of buffer pages
1598 * @cpu_buffer: CPU buffer with pages to test
1599 *
1600 * As a safety measure we check to make sure the data pages have not
1601 * been corrupted.
1602 */
1603static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1604{
1605 struct list_head *head = cpu_buffer->pages;
1606 struct buffer_page *bpage, *tmp;
1607
1608 /* Reset the head page if it exists */
1609 if (cpu_buffer->head_page)
1610 rb_set_head_page(cpu_buffer);
1611
1612 rb_head_page_deactivate(cpu_buffer);
1613
1614 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1615 return -1;
1616 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1617 return -1;
1618
1619 if (rb_check_list(cpu_buffer, head))
1620 return -1;
1621
1622 list_for_each_entry_safe(bpage, tmp, head, list) {
1623 if (RB_WARN_ON(cpu_buffer,
1624 bpage->list.next->prev != &bpage->list))
1625 return -1;
1626 if (RB_WARN_ON(cpu_buffer,
1627 bpage->list.prev->next != &bpage->list))
1628 return -1;
1629 if (rb_check_list(cpu_buffer, &bpage->list))
1630 return -1;
1631 }
1632
1633 rb_head_page_activate(cpu_buffer);
1634
1635 return 0;
1636}
1637
1638static int __rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1639 long nr_pages, struct list_head *pages)
1640{
1641 struct buffer_page *bpage, *tmp;
1642 bool user_thread = current->mm != NULL;
1643 gfp_t mflags;
1644 long i;
1645
1646 /*
1647 * Check if the available memory is there first.
1648 * Note, si_mem_available() only gives us a rough estimate of available
1649 * memory. It may not be accurate. But we don't care, we just want
1650 * to prevent doing any allocation when it is obvious that it is
1651 * not going to succeed.
1652 */
1653 i = si_mem_available();
1654 if (i < nr_pages)
1655 return -ENOMEM;
1656
1657 /*
1658 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1659 * gracefully without invoking oom-killer and the system is not
1660 * destabilized.
1661 */
1662 mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1663
1664 /*
1665 * If a user thread allocates too much, and si_mem_available()
1666 * reports there's enough memory, even though there is not.
1667 * Make sure the OOM killer kills this thread. This can happen
1668 * even with RETRY_MAYFAIL because another task may be doing
1669 * an allocation after this task has taken all memory.
1670 * This is the task the OOM killer needs to take out during this
1671 * loop, even if it was triggered by an allocation somewhere else.
1672 */
1673 if (user_thread)
1674 set_current_oom_origin();
1675 for (i = 0; i < nr_pages; i++) {
1676 struct page *page;
1677
1678 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1679 mflags, cpu_to_node(cpu_buffer->cpu));
1680 if (!bpage)
1681 goto free_pages;
1682
1683 rb_check_bpage(cpu_buffer, bpage);
1684
1685 list_add(&bpage->list, pages);
1686
1687 page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu), mflags, 0);
1688 if (!page)
1689 goto free_pages;
1690 bpage->page = page_address(page);
1691 rb_init_page(bpage->page);
1692
1693 if (user_thread && fatal_signal_pending(current))
1694 goto free_pages;
1695 }
1696 if (user_thread)
1697 clear_current_oom_origin();
1698
1699 return 0;
1700
1701free_pages:
1702 list_for_each_entry_safe(bpage, tmp, pages, list) {
1703 list_del_init(&bpage->list);
1704 free_buffer_page(bpage);
1705 }
1706 if (user_thread)
1707 clear_current_oom_origin();
1708
1709 return -ENOMEM;
1710}
1711
1712static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1713 unsigned long nr_pages)
1714{
1715 LIST_HEAD(pages);
1716
1717 WARN_ON(!nr_pages);
1718
1719 if (__rb_allocate_pages(cpu_buffer, nr_pages, &pages))
1720 return -ENOMEM;
1721
1722 /*
1723 * The ring buffer page list is a circular list that does not
1724 * start and end with a list head. All page list items point to
1725 * other pages.
1726 */
1727 cpu_buffer->pages = pages.next;
1728 list_del(&pages);
1729
1730 cpu_buffer->nr_pages = nr_pages;
1731
1732 rb_check_pages(cpu_buffer);
1733
1734 return 0;
1735}
1736
1737static struct ring_buffer_per_cpu *
1738rb_allocate_cpu_buffer(struct trace_buffer *buffer, long nr_pages, int cpu)
1739{
1740 struct ring_buffer_per_cpu *cpu_buffer;
1741 struct buffer_page *bpage;
1742 struct page *page;
1743 int ret;
1744
1745 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1746 GFP_KERNEL, cpu_to_node(cpu));
1747 if (!cpu_buffer)
1748 return NULL;
1749
1750 cpu_buffer->cpu = cpu;
1751 cpu_buffer->buffer = buffer;
1752 raw_spin_lock_init(&cpu_buffer->reader_lock);
1753 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1754 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1755 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1756 init_completion(&cpu_buffer->update_done);
1757 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1758 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1759 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1760
1761 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1762 GFP_KERNEL, cpu_to_node(cpu));
1763 if (!bpage)
1764 goto fail_free_buffer;
1765
1766 rb_check_bpage(cpu_buffer, bpage);
1767
1768 cpu_buffer->reader_page = bpage;
1769 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1770 if (!page)
1771 goto fail_free_reader;
1772 bpage->page = page_address(page);
1773 rb_init_page(bpage->page);
1774
1775 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1776 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1777
1778 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1779 if (ret < 0)
1780 goto fail_free_reader;
1781
1782 cpu_buffer->head_page
1783 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1784 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1785
1786 rb_head_page_activate(cpu_buffer);
1787
1788 return cpu_buffer;
1789
1790 fail_free_reader:
1791 free_buffer_page(cpu_buffer->reader_page);
1792
1793 fail_free_buffer:
1794 kfree(cpu_buffer);
1795 return NULL;
1796}
1797
1798static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1799{
1800 struct list_head *head = cpu_buffer->pages;
1801 struct buffer_page *bpage, *tmp;
1802
1803 free_buffer_page(cpu_buffer->reader_page);
1804
1805 if (head) {
1806 rb_head_page_deactivate(cpu_buffer);
1807
1808 list_for_each_entry_safe(bpage, tmp, head, list) {
1809 list_del_init(&bpage->list);
1810 free_buffer_page(bpage);
1811 }
1812 bpage = list_entry(head, struct buffer_page, list);
1813 free_buffer_page(bpage);
1814 }
1815
1816 kfree(cpu_buffer);
1817}
1818
1819/**
1820 * __ring_buffer_alloc - allocate a new ring_buffer
1821 * @size: the size in bytes per cpu that is needed.
1822 * @flags: attributes to set for the ring buffer.
1823 * @key: ring buffer reader_lock_key.
1824 *
1825 * Currently the only flag that is available is the RB_FL_OVERWRITE
1826 * flag. This flag means that the buffer will overwrite old data
1827 * when the buffer wraps. If this flag is not set, the buffer will
1828 * drop data when the tail hits the head.
1829 */
1830struct trace_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1831 struct lock_class_key *key)
1832{
1833 struct trace_buffer *buffer;
1834 long nr_pages;
1835 int bsize;
1836 int cpu;
1837 int ret;
1838
1839 /* keep it in its own cache line */
1840 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1841 GFP_KERNEL);
1842 if (!buffer)
1843 return NULL;
1844
1845 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1846 goto fail_free_buffer;
1847
1848 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1849 buffer->flags = flags;
1850 buffer->clock = trace_clock_local;
1851 buffer->reader_lock_key = key;
1852
1853 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1854 init_waitqueue_head(&buffer->irq_work.waiters);
1855
1856 /* need at least two pages */
1857 if (nr_pages < 2)
1858 nr_pages = 2;
1859
1860 buffer->cpus = nr_cpu_ids;
1861
1862 bsize = sizeof(void *) * nr_cpu_ids;
1863 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1864 GFP_KERNEL);
1865 if (!buffer->buffers)
1866 goto fail_free_cpumask;
1867
1868 cpu = raw_smp_processor_id();
1869 cpumask_set_cpu(cpu, buffer->cpumask);
1870 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1871 if (!buffer->buffers[cpu])
1872 goto fail_free_buffers;
1873
1874 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1875 if (ret < 0)
1876 goto fail_free_buffers;
1877
1878 mutex_init(&buffer->mutex);
1879
1880 return buffer;
1881
1882 fail_free_buffers:
1883 for_each_buffer_cpu(buffer, cpu) {
1884 if (buffer->buffers[cpu])
1885 rb_free_cpu_buffer(buffer->buffers[cpu]);
1886 }
1887 kfree(buffer->buffers);
1888
1889 fail_free_cpumask:
1890 free_cpumask_var(buffer->cpumask);
1891
1892 fail_free_buffer:
1893 kfree(buffer);
1894 return NULL;
1895}
1896EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1897
1898/**
1899 * ring_buffer_free - free a ring buffer.
1900 * @buffer: the buffer to free.
1901 */
1902void
1903ring_buffer_free(struct trace_buffer *buffer)
1904{
1905 int cpu;
1906
1907 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1908
1909 for_each_buffer_cpu(buffer, cpu)
1910 rb_free_cpu_buffer(buffer->buffers[cpu]);
1911
1912 kfree(buffer->buffers);
1913 free_cpumask_var(buffer->cpumask);
1914
1915 kfree(buffer);
1916}
1917EXPORT_SYMBOL_GPL(ring_buffer_free);
1918
1919void ring_buffer_set_clock(struct trace_buffer *buffer,
1920 u64 (*clock)(void))
1921{
1922 buffer->clock = clock;
1923}
1924
1925void ring_buffer_set_time_stamp_abs(struct trace_buffer *buffer, bool abs)
1926{
1927 buffer->time_stamp_abs = abs;
1928}
1929
1930bool ring_buffer_time_stamp_abs(struct trace_buffer *buffer)
1931{
1932 return buffer->time_stamp_abs;
1933}
1934
1935static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1936
1937static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1938{
1939 return local_read(&bpage->entries) & RB_WRITE_MASK;
1940}
1941
1942static inline unsigned long rb_page_write(struct buffer_page *bpage)
1943{
1944 return local_read(&bpage->write) & RB_WRITE_MASK;
1945}
1946
1947static int
1948rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1949{
1950 struct list_head *tail_page, *to_remove, *next_page;
1951 struct buffer_page *to_remove_page, *tmp_iter_page;
1952 struct buffer_page *last_page, *first_page;
1953 unsigned long nr_removed;
1954 unsigned long head_bit;
1955 int page_entries;
1956
1957 head_bit = 0;
1958
1959 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1960 atomic_inc(&cpu_buffer->record_disabled);
1961 /*
1962 * We don't race with the readers since we have acquired the reader
1963 * lock. We also don't race with writers after disabling recording.
1964 * This makes it easy to figure out the first and the last page to be
1965 * removed from the list. We unlink all the pages in between including
1966 * the first and last pages. This is done in a busy loop so that we
1967 * lose the least number of traces.
1968 * The pages are freed after we restart recording and unlock readers.
1969 */
1970 tail_page = &cpu_buffer->tail_page->list;
1971
1972 /*
1973 * tail page might be on reader page, we remove the next page
1974 * from the ring buffer
1975 */
1976 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1977 tail_page = rb_list_head(tail_page->next);
1978 to_remove = tail_page;
1979
1980 /* start of pages to remove */
1981 first_page = list_entry(rb_list_head(to_remove->next),
1982 struct buffer_page, list);
1983
1984 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1985 to_remove = rb_list_head(to_remove)->next;
1986 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1987 }
1988
1989 next_page = rb_list_head(to_remove)->next;
1990
1991 /*
1992 * Now we remove all pages between tail_page and next_page.
1993 * Make sure that we have head_bit value preserved for the
1994 * next page
1995 */
1996 tail_page->next = (struct list_head *)((unsigned long)next_page |
1997 head_bit);
1998 next_page = rb_list_head(next_page);
1999 next_page->prev = tail_page;
2000
2001 /* make sure pages points to a valid page in the ring buffer */
2002 cpu_buffer->pages = next_page;
2003
2004 /* update head page */
2005 if (head_bit)
2006 cpu_buffer->head_page = list_entry(next_page,
2007 struct buffer_page, list);
2008
2009 /*
2010 * change read pointer to make sure any read iterators reset
2011 * themselves
2012 */
2013 cpu_buffer->read = 0;
2014
2015 /* pages are removed, resume tracing and then free the pages */
2016 atomic_dec(&cpu_buffer->record_disabled);
2017 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
2018
2019 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
2020
2021 /* last buffer page to remove */
2022 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
2023 list);
2024 tmp_iter_page = first_page;
2025
2026 do {
2027 cond_resched();
2028
2029 to_remove_page = tmp_iter_page;
2030 rb_inc_page(&tmp_iter_page);
2031
2032 /* update the counters */
2033 page_entries = rb_page_entries(to_remove_page);
2034 if (page_entries) {
2035 /*
2036 * If something was added to this page, it was full
2037 * since it is not the tail page. So we deduct the
2038 * bytes consumed in ring buffer from here.
2039 * Increment overrun to account for the lost events.
2040 */
2041 local_add(page_entries, &cpu_buffer->overrun);
2042 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2043 local_inc(&cpu_buffer->pages_lost);
2044 }
2045
2046 /*
2047 * We have already removed references to this list item, just
2048 * free up the buffer_page and its page
2049 */
2050 free_buffer_page(to_remove_page);
2051 nr_removed--;
2052
2053 } while (to_remove_page != last_page);
2054
2055 RB_WARN_ON(cpu_buffer, nr_removed);
2056
2057 return nr_removed == 0;
2058}
2059
2060static int
2061rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
2062{
2063 struct list_head *pages = &cpu_buffer->new_pages;
2064 int retries, success;
2065 unsigned long flags;
2066
2067 /* Can be called at early boot up, where interrupts must not been enabled */
2068 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2069 /*
2070 * We are holding the reader lock, so the reader page won't be swapped
2071 * in the ring buffer. Now we are racing with the writer trying to
2072 * move head page and the tail page.
2073 * We are going to adapt the reader page update process where:
2074 * 1. We first splice the start and end of list of new pages between
2075 * the head page and its previous page.
2076 * 2. We cmpxchg the prev_page->next to point from head page to the
2077 * start of new pages list.
2078 * 3. Finally, we update the head->prev to the end of new list.
2079 *
2080 * We will try this process 10 times, to make sure that we don't keep
2081 * spinning.
2082 */
2083 retries = 10;
2084 success = 0;
2085 while (retries--) {
2086 struct list_head *head_page, *prev_page, *r;
2087 struct list_head *last_page, *first_page;
2088 struct list_head *head_page_with_bit;
2089
2090 head_page = &rb_set_head_page(cpu_buffer)->list;
2091 if (!head_page)
2092 break;
2093 prev_page = head_page->prev;
2094
2095 first_page = pages->next;
2096 last_page = pages->prev;
2097
2098 head_page_with_bit = (struct list_head *)
2099 ((unsigned long)head_page | RB_PAGE_HEAD);
2100
2101 last_page->next = head_page_with_bit;
2102 first_page->prev = prev_page;
2103
2104 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
2105
2106 if (r == head_page_with_bit) {
2107 /*
2108 * yay, we replaced the page pointer to our new list,
2109 * now, we just have to update to head page's prev
2110 * pointer to point to end of list
2111 */
2112 head_page->prev = last_page;
2113 success = 1;
2114 break;
2115 }
2116 }
2117
2118 if (success)
2119 INIT_LIST_HEAD(pages);
2120 /*
2121 * If we weren't successful in adding in new pages, warn and stop
2122 * tracing
2123 */
2124 RB_WARN_ON(cpu_buffer, !success);
2125 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2126
2127 /* free pages if they weren't inserted */
2128 if (!success) {
2129 struct buffer_page *bpage, *tmp;
2130 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2131 list) {
2132 list_del_init(&bpage->list);
2133 free_buffer_page(bpage);
2134 }
2135 }
2136 return success;
2137}
2138
2139static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
2140{
2141 int success;
2142
2143 if (cpu_buffer->nr_pages_to_update > 0)
2144 success = rb_insert_pages(cpu_buffer);
2145 else
2146 success = rb_remove_pages(cpu_buffer,
2147 -cpu_buffer->nr_pages_to_update);
2148
2149 if (success)
2150 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
2151}
2152
2153static void update_pages_handler(struct work_struct *work)
2154{
2155 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
2156 struct ring_buffer_per_cpu, update_pages_work);
2157 rb_update_pages(cpu_buffer);
2158 complete(&cpu_buffer->update_done);
2159}
2160
2161/**
2162 * ring_buffer_resize - resize the ring buffer
2163 * @buffer: the buffer to resize.
2164 * @size: the new size.
2165 * @cpu_id: the cpu buffer to resize
2166 *
2167 * Minimum size is 2 * BUF_PAGE_SIZE.
2168 *
2169 * Returns 0 on success and < 0 on failure.
2170 */
2171int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
2172 int cpu_id)
2173{
2174 struct ring_buffer_per_cpu *cpu_buffer;
2175 unsigned long nr_pages;
2176 int cpu, err;
2177
2178 /*
2179 * Always succeed at resizing a non-existent buffer:
2180 */
2181 if (!buffer)
2182 return 0;
2183
2184 /* Make sure the requested buffer exists */
2185 if (cpu_id != RING_BUFFER_ALL_CPUS &&
2186 !cpumask_test_cpu(cpu_id, buffer->cpumask))
2187 return 0;
2188
2189 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
2190
2191 /* we need a minimum of two pages */
2192 if (nr_pages < 2)
2193 nr_pages = 2;
2194
2195 /* prevent another thread from changing buffer sizes */
2196 mutex_lock(&buffer->mutex);
2197
2198
2199 if (cpu_id == RING_BUFFER_ALL_CPUS) {
2200 /*
2201 * Don't succeed if resizing is disabled, as a reader might be
2202 * manipulating the ring buffer and is expecting a sane state while
2203 * this is true.
2204 */
2205 for_each_buffer_cpu(buffer, cpu) {
2206 cpu_buffer = buffer->buffers[cpu];
2207 if (atomic_read(&cpu_buffer->resize_disabled)) {
2208 err = -EBUSY;
2209 goto out_err_unlock;
2210 }
2211 }
2212
2213 /* calculate the pages to update */
2214 for_each_buffer_cpu(buffer, cpu) {
2215 cpu_buffer = buffer->buffers[cpu];
2216
2217 cpu_buffer->nr_pages_to_update = nr_pages -
2218 cpu_buffer->nr_pages;
2219 /*
2220 * nothing more to do for removing pages or no update
2221 */
2222 if (cpu_buffer->nr_pages_to_update <= 0)
2223 continue;
2224 /*
2225 * to add pages, make sure all new pages can be
2226 * allocated without receiving ENOMEM
2227 */
2228 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2229 if (__rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2230 &cpu_buffer->new_pages)) {
2231 /* not enough memory for new pages */
2232 err = -ENOMEM;
2233 goto out_err;
2234 }
2235 }
2236
2237 cpus_read_lock();
2238 /*
2239 * Fire off all the required work handlers
2240 * We can't schedule on offline CPUs, but it's not necessary
2241 * since we can change their buffer sizes without any race.
2242 */
2243 for_each_buffer_cpu(buffer, cpu) {
2244 cpu_buffer = buffer->buffers[cpu];
2245 if (!cpu_buffer->nr_pages_to_update)
2246 continue;
2247
2248 /* Can't run something on an offline CPU. */
2249 if (!cpu_online(cpu)) {
2250 rb_update_pages(cpu_buffer);
2251 cpu_buffer->nr_pages_to_update = 0;
2252 } else {
2253 /* Run directly if possible. */
2254 migrate_disable();
2255 if (cpu != smp_processor_id()) {
2256 migrate_enable();
2257 schedule_work_on(cpu,
2258 &cpu_buffer->update_pages_work);
2259 } else {
2260 update_pages_handler(&cpu_buffer->update_pages_work);
2261 migrate_enable();
2262 }
2263 }
2264 }
2265
2266 /* wait for all the updates to complete */
2267 for_each_buffer_cpu(buffer, cpu) {
2268 cpu_buffer = buffer->buffers[cpu];
2269 if (!cpu_buffer->nr_pages_to_update)
2270 continue;
2271
2272 if (cpu_online(cpu))
2273 wait_for_completion(&cpu_buffer->update_done);
2274 cpu_buffer->nr_pages_to_update = 0;
2275 }
2276
2277 cpus_read_unlock();
2278 } else {
2279 cpu_buffer = buffer->buffers[cpu_id];
2280
2281 if (nr_pages == cpu_buffer->nr_pages)
2282 goto out;
2283
2284 /*
2285 * Don't succeed if resizing is disabled, as a reader might be
2286 * manipulating the ring buffer and is expecting a sane state while
2287 * this is true.
2288 */
2289 if (atomic_read(&cpu_buffer->resize_disabled)) {
2290 err = -EBUSY;
2291 goto out_err_unlock;
2292 }
2293
2294 cpu_buffer->nr_pages_to_update = nr_pages -
2295 cpu_buffer->nr_pages;
2296
2297 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2298 if (cpu_buffer->nr_pages_to_update > 0 &&
2299 __rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2300 &cpu_buffer->new_pages)) {
2301 err = -ENOMEM;
2302 goto out_err;
2303 }
2304
2305 cpus_read_lock();
2306
2307 /* Can't run something on an offline CPU. */
2308 if (!cpu_online(cpu_id))
2309 rb_update_pages(cpu_buffer);
2310 else {
2311 /* Run directly if possible. */
2312 migrate_disable();
2313 if (cpu_id == smp_processor_id()) {
2314 rb_update_pages(cpu_buffer);
2315 migrate_enable();
2316 } else {
2317 migrate_enable();
2318 schedule_work_on(cpu_id,
2319 &cpu_buffer->update_pages_work);
2320 wait_for_completion(&cpu_buffer->update_done);
2321 }
2322 }
2323
2324 cpu_buffer->nr_pages_to_update = 0;
2325 cpus_read_unlock();
2326 }
2327
2328 out:
2329 /*
2330 * The ring buffer resize can happen with the ring buffer
2331 * enabled, so that the update disturbs the tracing as little
2332 * as possible. But if the buffer is disabled, we do not need
2333 * to worry about that, and we can take the time to verify
2334 * that the buffer is not corrupt.
2335 */
2336 if (atomic_read(&buffer->record_disabled)) {
2337 atomic_inc(&buffer->record_disabled);
2338 /*
2339 * Even though the buffer was disabled, we must make sure
2340 * that it is truly disabled before calling rb_check_pages.
2341 * There could have been a race between checking
2342 * record_disable and incrementing it.
2343 */
2344 synchronize_rcu();
2345 for_each_buffer_cpu(buffer, cpu) {
2346 cpu_buffer = buffer->buffers[cpu];
2347 rb_check_pages(cpu_buffer);
2348 }
2349 atomic_dec(&buffer->record_disabled);
2350 }
2351
2352 mutex_unlock(&buffer->mutex);
2353 return 0;
2354
2355 out_err:
2356 for_each_buffer_cpu(buffer, cpu) {
2357 struct buffer_page *bpage, *tmp;
2358
2359 cpu_buffer = buffer->buffers[cpu];
2360 cpu_buffer->nr_pages_to_update = 0;
2361
2362 if (list_empty(&cpu_buffer->new_pages))
2363 continue;
2364
2365 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2366 list) {
2367 list_del_init(&bpage->list);
2368 free_buffer_page(bpage);
2369 }
2370 }
2371 out_err_unlock:
2372 mutex_unlock(&buffer->mutex);
2373 return err;
2374}
2375EXPORT_SYMBOL_GPL(ring_buffer_resize);
2376
2377void ring_buffer_change_overwrite(struct trace_buffer *buffer, int val)
2378{
2379 mutex_lock(&buffer->mutex);
2380 if (val)
2381 buffer->flags |= RB_FL_OVERWRITE;
2382 else
2383 buffer->flags &= ~RB_FL_OVERWRITE;
2384 mutex_unlock(&buffer->mutex);
2385}
2386EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
2387
2388static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
2389{
2390 return bpage->page->data + index;
2391}
2392
2393static __always_inline struct ring_buffer_event *
2394rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
2395{
2396 return __rb_page_index(cpu_buffer->reader_page,
2397 cpu_buffer->reader_page->read);
2398}
2399
2400static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
2401{
2402 return local_read(&bpage->page->commit);
2403}
2404
2405static struct ring_buffer_event *
2406rb_iter_head_event(struct ring_buffer_iter *iter)
2407{
2408 struct ring_buffer_event *event;
2409 struct buffer_page *iter_head_page = iter->head_page;
2410 unsigned long commit;
2411 unsigned length;
2412
2413 if (iter->head != iter->next_event)
2414 return iter->event;
2415
2416 /*
2417 * When the writer goes across pages, it issues a cmpxchg which
2418 * is a mb(), which will synchronize with the rmb here.
2419 * (see rb_tail_page_update() and __rb_reserve_next())
2420 */
2421 commit = rb_page_commit(iter_head_page);
2422 smp_rmb();
2423 event = __rb_page_index(iter_head_page, iter->head);
2424 length = rb_event_length(event);
2425
2426 /*
2427 * READ_ONCE() doesn't work on functions and we don't want the
2428 * compiler doing any crazy optimizations with length.
2429 */
2430 barrier();
2431
2432 if ((iter->head + length) > commit || length > BUF_MAX_DATA_SIZE)
2433 /* Writer corrupted the read? */
2434 goto reset;
2435
2436 memcpy(iter->event, event, length);
2437 /*
2438 * If the page stamp is still the same after this rmb() then the
2439 * event was safely copied without the writer entering the page.
2440 */
2441 smp_rmb();
2442
2443 /* Make sure the page didn't change since we read this */
2444 if (iter->page_stamp != iter_head_page->page->time_stamp ||
2445 commit > rb_page_commit(iter_head_page))
2446 goto reset;
2447
2448 iter->next_event = iter->head + length;
2449 return iter->event;
2450 reset:
2451 /* Reset to the beginning */
2452 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2453 iter->head = 0;
2454 iter->next_event = 0;
2455 iter->missed_events = 1;
2456 return NULL;
2457}
2458
2459/* Size is determined by what has been committed */
2460static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
2461{
2462 return rb_page_commit(bpage);
2463}
2464
2465static __always_inline unsigned
2466rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
2467{
2468 return rb_page_commit(cpu_buffer->commit_page);
2469}
2470
2471static __always_inline unsigned
2472rb_event_index(struct ring_buffer_event *event)
2473{
2474 unsigned long addr = (unsigned long)event;
2475
2476 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
2477}
2478
2479static void rb_inc_iter(struct ring_buffer_iter *iter)
2480{
2481 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2482
2483 /*
2484 * The iterator could be on the reader page (it starts there).
2485 * But the head could have moved, since the reader was
2486 * found. Check for this case and assign the iterator
2487 * to the head page instead of next.
2488 */
2489 if (iter->head_page == cpu_buffer->reader_page)
2490 iter->head_page = rb_set_head_page(cpu_buffer);
2491 else
2492 rb_inc_page(&iter->head_page);
2493
2494 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2495 iter->head = 0;
2496 iter->next_event = 0;
2497}
2498
2499/*
2500 * rb_handle_head_page - writer hit the head page
2501 *
2502 * Returns: +1 to retry page
2503 * 0 to continue
2504 * -1 on error
2505 */
2506static int
2507rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2508 struct buffer_page *tail_page,
2509 struct buffer_page *next_page)
2510{
2511 struct buffer_page *new_head;
2512 int entries;
2513 int type;
2514 int ret;
2515
2516 entries = rb_page_entries(next_page);
2517
2518 /*
2519 * The hard part is here. We need to move the head
2520 * forward, and protect against both readers on
2521 * other CPUs and writers coming in via interrupts.
2522 */
2523 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2524 RB_PAGE_HEAD);
2525
2526 /*
2527 * type can be one of four:
2528 * NORMAL - an interrupt already moved it for us
2529 * HEAD - we are the first to get here.
2530 * UPDATE - we are the interrupt interrupting
2531 * a current move.
2532 * MOVED - a reader on another CPU moved the next
2533 * pointer to its reader page. Give up
2534 * and try again.
2535 */
2536
2537 switch (type) {
2538 case RB_PAGE_HEAD:
2539 /*
2540 * We changed the head to UPDATE, thus
2541 * it is our responsibility to update
2542 * the counters.
2543 */
2544 local_add(entries, &cpu_buffer->overrun);
2545 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2546 local_inc(&cpu_buffer->pages_lost);
2547
2548 /*
2549 * The entries will be zeroed out when we move the
2550 * tail page.
2551 */
2552
2553 /* still more to do */
2554 break;
2555
2556 case RB_PAGE_UPDATE:
2557 /*
2558 * This is an interrupt that interrupt the
2559 * previous update. Still more to do.
2560 */
2561 break;
2562 case RB_PAGE_NORMAL:
2563 /*
2564 * An interrupt came in before the update
2565 * and processed this for us.
2566 * Nothing left to do.
2567 */
2568 return 1;
2569 case RB_PAGE_MOVED:
2570 /*
2571 * The reader is on another CPU and just did
2572 * a swap with our next_page.
2573 * Try again.
2574 */
2575 return 1;
2576 default:
2577 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2578 return -1;
2579 }
2580
2581 /*
2582 * Now that we are here, the old head pointer is
2583 * set to UPDATE. This will keep the reader from
2584 * swapping the head page with the reader page.
2585 * The reader (on another CPU) will spin till
2586 * we are finished.
2587 *
2588 * We just need to protect against interrupts
2589 * doing the job. We will set the next pointer
2590 * to HEAD. After that, we set the old pointer
2591 * to NORMAL, but only if it was HEAD before.
2592 * otherwise we are an interrupt, and only
2593 * want the outer most commit to reset it.
2594 */
2595 new_head = next_page;
2596 rb_inc_page(&new_head);
2597
2598 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2599 RB_PAGE_NORMAL);
2600
2601 /*
2602 * Valid returns are:
2603 * HEAD - an interrupt came in and already set it.
2604 * NORMAL - One of two things:
2605 * 1) We really set it.
2606 * 2) A bunch of interrupts came in and moved
2607 * the page forward again.
2608 */
2609 switch (ret) {
2610 case RB_PAGE_HEAD:
2611 case RB_PAGE_NORMAL:
2612 /* OK */
2613 break;
2614 default:
2615 RB_WARN_ON(cpu_buffer, 1);
2616 return -1;
2617 }
2618
2619 /*
2620 * It is possible that an interrupt came in,
2621 * set the head up, then more interrupts came in
2622 * and moved it again. When we get back here,
2623 * the page would have been set to NORMAL but we
2624 * just set it back to HEAD.
2625 *
2626 * How do you detect this? Well, if that happened
2627 * the tail page would have moved.
2628 */
2629 if (ret == RB_PAGE_NORMAL) {
2630 struct buffer_page *buffer_tail_page;
2631
2632 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2633 /*
2634 * If the tail had moved passed next, then we need
2635 * to reset the pointer.
2636 */
2637 if (buffer_tail_page != tail_page &&
2638 buffer_tail_page != next_page)
2639 rb_head_page_set_normal(cpu_buffer, new_head,
2640 next_page,
2641 RB_PAGE_HEAD);
2642 }
2643
2644 /*
2645 * If this was the outer most commit (the one that
2646 * changed the original pointer from HEAD to UPDATE),
2647 * then it is up to us to reset it to NORMAL.
2648 */
2649 if (type == RB_PAGE_HEAD) {
2650 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2651 tail_page,
2652 RB_PAGE_UPDATE);
2653 if (RB_WARN_ON(cpu_buffer,
2654 ret != RB_PAGE_UPDATE))
2655 return -1;
2656 }
2657
2658 return 0;
2659}
2660
2661static inline void
2662rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2663 unsigned long tail, struct rb_event_info *info)
2664{
2665 struct buffer_page *tail_page = info->tail_page;
2666 struct ring_buffer_event *event;
2667 unsigned long length = info->length;
2668
2669 /*
2670 * Only the event that crossed the page boundary
2671 * must fill the old tail_page with padding.
2672 */
2673 if (tail >= BUF_PAGE_SIZE) {
2674 /*
2675 * If the page was filled, then we still need
2676 * to update the real_end. Reset it to zero
2677 * and the reader will ignore it.
2678 */
2679 if (tail == BUF_PAGE_SIZE)
2680 tail_page->real_end = 0;
2681
2682 local_sub(length, &tail_page->write);
2683 return;
2684 }
2685
2686 event = __rb_page_index(tail_page, tail);
2687
2688 /* account for padding bytes */
2689 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2690
2691 /*
2692 * Save the original length to the meta data.
2693 * This will be used by the reader to add lost event
2694 * counter.
2695 */
2696 tail_page->real_end = tail;
2697
2698 /*
2699 * If this event is bigger than the minimum size, then
2700 * we need to be careful that we don't subtract the
2701 * write counter enough to allow another writer to slip
2702 * in on this page.
2703 * We put in a discarded commit instead, to make sure
2704 * that this space is not used again.
2705 *
2706 * If we are less than the minimum size, we don't need to
2707 * worry about it.
2708 */
2709 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2710 /* No room for any events */
2711
2712 /* Mark the rest of the page with padding */
2713 rb_event_set_padding(event);
2714
2715 /* Make sure the padding is visible before the write update */
2716 smp_wmb();
2717
2718 /* Set the write back to the previous setting */
2719 local_sub(length, &tail_page->write);
2720 return;
2721 }
2722
2723 /* Put in a discarded event */
2724 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2725 event->type_len = RINGBUF_TYPE_PADDING;
2726 /* time delta must be non zero */
2727 event->time_delta = 1;
2728
2729 /* Make sure the padding is visible before the tail_page->write update */
2730 smp_wmb();
2731
2732 /* Set write to end of buffer */
2733 length = (tail + length) - BUF_PAGE_SIZE;
2734 local_sub(length, &tail_page->write);
2735}
2736
2737static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2738
2739/*
2740 * This is the slow path, force gcc not to inline it.
2741 */
2742static noinline struct ring_buffer_event *
2743rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2744 unsigned long tail, struct rb_event_info *info)
2745{
2746 struct buffer_page *tail_page = info->tail_page;
2747 struct buffer_page *commit_page = cpu_buffer->commit_page;
2748 struct trace_buffer *buffer = cpu_buffer->buffer;
2749 struct buffer_page *next_page;
2750 int ret;
2751
2752 next_page = tail_page;
2753
2754 rb_inc_page(&next_page);
2755
2756 /*
2757 * If for some reason, we had an interrupt storm that made
2758 * it all the way around the buffer, bail, and warn
2759 * about it.
2760 */
2761 if (unlikely(next_page == commit_page)) {
2762 local_inc(&cpu_buffer->commit_overrun);
2763 goto out_reset;
2764 }
2765
2766 /*
2767 * This is where the fun begins!
2768 *
2769 * We are fighting against races between a reader that
2770 * could be on another CPU trying to swap its reader
2771 * page with the buffer head.
2772 *
2773 * We are also fighting against interrupts coming in and
2774 * moving the head or tail on us as well.
2775 *
2776 * If the next page is the head page then we have filled
2777 * the buffer, unless the commit page is still on the
2778 * reader page.
2779 */
2780 if (rb_is_head_page(next_page, &tail_page->list)) {
2781
2782 /*
2783 * If the commit is not on the reader page, then
2784 * move the header page.
2785 */
2786 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2787 /*
2788 * If we are not in overwrite mode,
2789 * this is easy, just stop here.
2790 */
2791 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2792 local_inc(&cpu_buffer->dropped_events);
2793 goto out_reset;
2794 }
2795
2796 ret = rb_handle_head_page(cpu_buffer,
2797 tail_page,
2798 next_page);
2799 if (ret < 0)
2800 goto out_reset;
2801 if (ret)
2802 goto out_again;
2803 } else {
2804 /*
2805 * We need to be careful here too. The
2806 * commit page could still be on the reader
2807 * page. We could have a small buffer, and
2808 * have filled up the buffer with events
2809 * from interrupts and such, and wrapped.
2810 *
2811 * Note, if the tail page is also on the
2812 * reader_page, we let it move out.
2813 */
2814 if (unlikely((cpu_buffer->commit_page !=
2815 cpu_buffer->tail_page) &&
2816 (cpu_buffer->commit_page ==
2817 cpu_buffer->reader_page))) {
2818 local_inc(&cpu_buffer->commit_overrun);
2819 goto out_reset;
2820 }
2821 }
2822 }
2823
2824 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2825
2826 out_again:
2827
2828 rb_reset_tail(cpu_buffer, tail, info);
2829
2830 /* Commit what we have for now. */
2831 rb_end_commit(cpu_buffer);
2832 /* rb_end_commit() decs committing */
2833 local_inc(&cpu_buffer->committing);
2834
2835 /* fail and let the caller try again */
2836 return ERR_PTR(-EAGAIN);
2837
2838 out_reset:
2839 /* reset write */
2840 rb_reset_tail(cpu_buffer, tail, info);
2841
2842 return NULL;
2843}
2844
2845/* Slow path */
2846static struct ring_buffer_event *
2847rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2848{
2849 if (abs)
2850 event->type_len = RINGBUF_TYPE_TIME_STAMP;
2851 else
2852 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2853
2854 /* Not the first event on the page, or not delta? */
2855 if (abs || rb_event_index(event)) {
2856 event->time_delta = delta & TS_MASK;
2857 event->array[0] = delta >> TS_SHIFT;
2858 } else {
2859 /* nope, just zero it */
2860 event->time_delta = 0;
2861 event->array[0] = 0;
2862 }
2863
2864 return skip_time_extend(event);
2865}
2866
2867#ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2868static inline bool sched_clock_stable(void)
2869{
2870 return true;
2871}
2872#endif
2873
2874static void
2875rb_check_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2876 struct rb_event_info *info)
2877{
2878 u64 write_stamp;
2879
2880 WARN_ONCE(1, "Delta way too big! %llu ts=%llu before=%llu after=%llu write stamp=%llu\n%s",
2881 (unsigned long long)info->delta,
2882 (unsigned long long)info->ts,
2883 (unsigned long long)info->before,
2884 (unsigned long long)info->after,
2885 (unsigned long long)(rb_time_read(&cpu_buffer->write_stamp, &write_stamp) ? write_stamp : 0),
2886 sched_clock_stable() ? "" :
2887 "If you just came from a suspend/resume,\n"
2888 "please switch to the trace global clock:\n"
2889 " echo global > /sys/kernel/debug/tracing/trace_clock\n"
2890 "or add trace_clock=global to the kernel command line\n");
2891}
2892
2893static void rb_add_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2894 struct ring_buffer_event **event,
2895 struct rb_event_info *info,
2896 u64 *delta,
2897 unsigned int *length)
2898{
2899 bool abs = info->add_timestamp &
2900 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE);
2901
2902 if (unlikely(info->delta > (1ULL << 59))) {
2903 /*
2904 * Some timers can use more than 59 bits, and when a timestamp
2905 * is added to the buffer, it will lose those bits.
2906 */
2907 if (abs && (info->ts & TS_MSB)) {
2908 info->delta &= ABS_TS_MASK;
2909
2910 /* did the clock go backwards */
2911 } else if (info->before == info->after && info->before > info->ts) {
2912 /* not interrupted */
2913 static int once;
2914
2915 /*
2916 * This is possible with a recalibrating of the TSC.
2917 * Do not produce a call stack, but just report it.
2918 */
2919 if (!once) {
2920 once++;
2921 pr_warn("Ring buffer clock went backwards: %llu -> %llu\n",
2922 info->before, info->ts);
2923 }
2924 } else
2925 rb_check_timestamp(cpu_buffer, info);
2926 if (!abs)
2927 info->delta = 0;
2928 }
2929 *event = rb_add_time_stamp(*event, info->delta, abs);
2930 *length -= RB_LEN_TIME_EXTEND;
2931 *delta = 0;
2932}
2933
2934/**
2935 * rb_update_event - update event type and data
2936 * @cpu_buffer: The per cpu buffer of the @event
2937 * @event: the event to update
2938 * @info: The info to update the @event with (contains length and delta)
2939 *
2940 * Update the type and data fields of the @event. The length
2941 * is the actual size that is written to the ring buffer,
2942 * and with this, we can determine what to place into the
2943 * data field.
2944 */
2945static void
2946rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2947 struct ring_buffer_event *event,
2948 struct rb_event_info *info)
2949{
2950 unsigned length = info->length;
2951 u64 delta = info->delta;
2952 unsigned int nest = local_read(&cpu_buffer->committing) - 1;
2953
2954 if (!WARN_ON_ONCE(nest >= MAX_NEST))
2955 cpu_buffer->event_stamp[nest] = info->ts;
2956
2957 /*
2958 * If we need to add a timestamp, then we
2959 * add it to the start of the reserved space.
2960 */
2961 if (unlikely(info->add_timestamp))
2962 rb_add_timestamp(cpu_buffer, &event, info, &delta, &length);
2963
2964 event->time_delta = delta;
2965 length -= RB_EVNT_HDR_SIZE;
2966 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2967 event->type_len = 0;
2968 event->array[0] = length;
2969 } else
2970 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2971}
2972
2973static unsigned rb_calculate_event_length(unsigned length)
2974{
2975 struct ring_buffer_event event; /* Used only for sizeof array */
2976
2977 /* zero length can cause confusions */
2978 if (!length)
2979 length++;
2980
2981 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2982 length += sizeof(event.array[0]);
2983
2984 length += RB_EVNT_HDR_SIZE;
2985 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2986
2987 /*
2988 * In case the time delta is larger than the 27 bits for it
2989 * in the header, we need to add a timestamp. If another
2990 * event comes in when trying to discard this one to increase
2991 * the length, then the timestamp will be added in the allocated
2992 * space of this event. If length is bigger than the size needed
2993 * for the TIME_EXTEND, then padding has to be used. The events
2994 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2995 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2996 * As length is a multiple of 4, we only need to worry if it
2997 * is 12 (RB_LEN_TIME_EXTEND + 4).
2998 */
2999 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
3000 length += RB_ALIGNMENT;
3001
3002 return length;
3003}
3004
3005static u64 rb_time_delta(struct ring_buffer_event *event)
3006{
3007 switch (event->type_len) {
3008 case RINGBUF_TYPE_PADDING:
3009 return 0;
3010
3011 case RINGBUF_TYPE_TIME_EXTEND:
3012 return rb_event_time_stamp(event);
3013
3014 case RINGBUF_TYPE_TIME_STAMP:
3015 return 0;
3016
3017 case RINGBUF_TYPE_DATA:
3018 return event->time_delta;
3019 default:
3020 return 0;
3021 }
3022}
3023
3024static inline int
3025rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
3026 struct ring_buffer_event *event)
3027{
3028 unsigned long new_index, old_index;
3029 struct buffer_page *bpage;
3030 unsigned long index;
3031 unsigned long addr;
3032 u64 write_stamp;
3033 u64 delta;
3034
3035 new_index = rb_event_index(event);
3036 old_index = new_index + rb_event_ts_length(event);
3037 addr = (unsigned long)event;
3038 addr &= PAGE_MASK;
3039
3040 bpage = READ_ONCE(cpu_buffer->tail_page);
3041
3042 delta = rb_time_delta(event);
3043
3044 if (!rb_time_read(&cpu_buffer->write_stamp, &write_stamp))
3045 return 0;
3046
3047 /* Make sure the write stamp is read before testing the location */
3048 barrier();
3049
3050 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
3051 unsigned long write_mask =
3052 local_read(&bpage->write) & ~RB_WRITE_MASK;
3053 unsigned long event_length = rb_event_length(event);
3054
3055 /* Something came in, can't discard */
3056 if (!rb_time_cmpxchg(&cpu_buffer->write_stamp,
3057 write_stamp, write_stamp - delta))
3058 return 0;
3059
3060 /*
3061 * It's possible that the event time delta is zero
3062 * (has the same time stamp as the previous event)
3063 * in which case write_stamp and before_stamp could
3064 * be the same. In such a case, force before_stamp
3065 * to be different than write_stamp. It doesn't
3066 * matter what it is, as long as its different.
3067 */
3068 if (!delta)
3069 rb_time_set(&cpu_buffer->before_stamp, 0);
3070
3071 /*
3072 * If an event were to come in now, it would see that the
3073 * write_stamp and the before_stamp are different, and assume
3074 * that this event just added itself before updating
3075 * the write stamp. The interrupting event will fix the
3076 * write stamp for us, and use the before stamp as its delta.
3077 */
3078
3079 /*
3080 * This is on the tail page. It is possible that
3081 * a write could come in and move the tail page
3082 * and write to the next page. That is fine
3083 * because we just shorten what is on this page.
3084 */
3085 old_index += write_mask;
3086 new_index += write_mask;
3087 index = local_cmpxchg(&bpage->write, old_index, new_index);
3088 if (index == old_index) {
3089 /* update counters */
3090 local_sub(event_length, &cpu_buffer->entries_bytes);
3091 return 1;
3092 }
3093 }
3094
3095 /* could not discard */
3096 return 0;
3097}
3098
3099static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
3100{
3101 local_inc(&cpu_buffer->committing);
3102 local_inc(&cpu_buffer->commits);
3103}
3104
3105static __always_inline void
3106rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
3107{
3108 unsigned long max_count;
3109
3110 /*
3111 * We only race with interrupts and NMIs on this CPU.
3112 * If we own the commit event, then we can commit
3113 * all others that interrupted us, since the interruptions
3114 * are in stack format (they finish before they come
3115 * back to us). This allows us to do a simple loop to
3116 * assign the commit to the tail.
3117 */
3118 again:
3119 max_count = cpu_buffer->nr_pages * 100;
3120
3121 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
3122 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
3123 return;
3124 if (RB_WARN_ON(cpu_buffer,
3125 rb_is_reader_page(cpu_buffer->tail_page)))
3126 return;
3127 local_set(&cpu_buffer->commit_page->page->commit,
3128 rb_page_write(cpu_buffer->commit_page));
3129 rb_inc_page(&cpu_buffer->commit_page);
3130 /* add barrier to keep gcc from optimizing too much */
3131 barrier();
3132 }
3133 while (rb_commit_index(cpu_buffer) !=
3134 rb_page_write(cpu_buffer->commit_page)) {
3135
3136 local_set(&cpu_buffer->commit_page->page->commit,
3137 rb_page_write(cpu_buffer->commit_page));
3138 RB_WARN_ON(cpu_buffer,
3139 local_read(&cpu_buffer->commit_page->page->commit) &
3140 ~RB_WRITE_MASK);
3141 barrier();
3142 }
3143
3144 /* again, keep gcc from optimizing */
3145 barrier();
3146
3147 /*
3148 * If an interrupt came in just after the first while loop
3149 * and pushed the tail page forward, we will be left with
3150 * a dangling commit that will never go forward.
3151 */
3152 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
3153 goto again;
3154}
3155
3156static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
3157{
3158 unsigned long commits;
3159
3160 if (RB_WARN_ON(cpu_buffer,
3161 !local_read(&cpu_buffer->committing)))
3162 return;
3163
3164 again:
3165 commits = local_read(&cpu_buffer->commits);
3166 /* synchronize with interrupts */
3167 barrier();
3168 if (local_read(&cpu_buffer->committing) == 1)
3169 rb_set_commit_to_write(cpu_buffer);
3170
3171 local_dec(&cpu_buffer->committing);
3172
3173 /* synchronize with interrupts */
3174 barrier();
3175
3176 /*
3177 * Need to account for interrupts coming in between the
3178 * updating of the commit page and the clearing of the
3179 * committing counter.
3180 */
3181 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
3182 !local_read(&cpu_buffer->committing)) {
3183 local_inc(&cpu_buffer->committing);
3184 goto again;
3185 }
3186}
3187
3188static inline void rb_event_discard(struct ring_buffer_event *event)
3189{
3190 if (extended_time(event))
3191 event = skip_time_extend(event);
3192
3193 /* array[0] holds the actual length for the discarded event */
3194 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
3195 event->type_len = RINGBUF_TYPE_PADDING;
3196 /* time delta must be non zero */
3197 if (!event->time_delta)
3198 event->time_delta = 1;
3199}
3200
3201static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer)
3202{
3203 local_inc(&cpu_buffer->entries);
3204 rb_end_commit(cpu_buffer);
3205}
3206
3207static __always_inline void
3208rb_wakeups(struct trace_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
3209{
3210 if (buffer->irq_work.waiters_pending) {
3211 buffer->irq_work.waiters_pending = false;
3212 /* irq_work_queue() supplies it's own memory barriers */
3213 irq_work_queue(&buffer->irq_work.work);
3214 }
3215
3216 if (cpu_buffer->irq_work.waiters_pending) {
3217 cpu_buffer->irq_work.waiters_pending = false;
3218 /* irq_work_queue() supplies it's own memory barriers */
3219 irq_work_queue(&cpu_buffer->irq_work.work);
3220 }
3221
3222 if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
3223 return;
3224
3225 if (cpu_buffer->reader_page == cpu_buffer->commit_page)
3226 return;
3227
3228 if (!cpu_buffer->irq_work.full_waiters_pending)
3229 return;
3230
3231 cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
3232
3233 if (!full_hit(buffer, cpu_buffer->cpu, cpu_buffer->shortest_full))
3234 return;
3235
3236 cpu_buffer->irq_work.wakeup_full = true;
3237 cpu_buffer->irq_work.full_waiters_pending = false;
3238 /* irq_work_queue() supplies it's own memory barriers */
3239 irq_work_queue(&cpu_buffer->irq_work.work);
3240}
3241
3242#ifdef CONFIG_RING_BUFFER_RECORD_RECURSION
3243# define do_ring_buffer_record_recursion() \
3244 do_ftrace_record_recursion(_THIS_IP_, _RET_IP_)
3245#else
3246# define do_ring_buffer_record_recursion() do { } while (0)
3247#endif
3248
3249/*
3250 * The lock and unlock are done within a preempt disable section.
3251 * The current_context per_cpu variable can only be modified
3252 * by the current task between lock and unlock. But it can
3253 * be modified more than once via an interrupt. To pass this
3254 * information from the lock to the unlock without having to
3255 * access the 'in_interrupt()' functions again (which do show
3256 * a bit of overhead in something as critical as function tracing,
3257 * we use a bitmask trick.
3258 *
3259 * bit 1 = NMI context
3260 * bit 2 = IRQ context
3261 * bit 3 = SoftIRQ context
3262 * bit 4 = normal context.
3263 *
3264 * This works because this is the order of contexts that can
3265 * preempt other contexts. A SoftIRQ never preempts an IRQ
3266 * context.
3267 *
3268 * When the context is determined, the corresponding bit is
3269 * checked and set (if it was set, then a recursion of that context
3270 * happened).
3271 *
3272 * On unlock, we need to clear this bit. To do so, just subtract
3273 * 1 from the current_context and AND it to itself.
3274 *
3275 * (binary)
3276 * 101 - 1 = 100
3277 * 101 & 100 = 100 (clearing bit zero)
3278 *
3279 * 1010 - 1 = 1001
3280 * 1010 & 1001 = 1000 (clearing bit 1)
3281 *
3282 * The least significant bit can be cleared this way, and it
3283 * just so happens that it is the same bit corresponding to
3284 * the current context.
3285 *
3286 * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
3287 * is set when a recursion is detected at the current context, and if
3288 * the TRANSITION bit is already set, it will fail the recursion.
3289 * This is needed because there's a lag between the changing of
3290 * interrupt context and updating the preempt count. In this case,
3291 * a false positive will be found. To handle this, one extra recursion
3292 * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
3293 * bit is already set, then it is considered a recursion and the function
3294 * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
3295 *
3296 * On the trace_recursive_unlock(), the TRANSITION bit will be the first
3297 * to be cleared. Even if it wasn't the context that set it. That is,
3298 * if an interrupt comes in while NORMAL bit is set and the ring buffer
3299 * is called before preempt_count() is updated, since the check will
3300 * be on the NORMAL bit, the TRANSITION bit will then be set. If an
3301 * NMI then comes in, it will set the NMI bit, but when the NMI code
3302 * does the trace_recursive_unlock() it will clear the TRANSITION bit
3303 * and leave the NMI bit set. But this is fine, because the interrupt
3304 * code that set the TRANSITION bit will then clear the NMI bit when it
3305 * calls trace_recursive_unlock(). If another NMI comes in, it will
3306 * set the TRANSITION bit and continue.
3307 *
3308 * Note: The TRANSITION bit only handles a single transition between context.
3309 */
3310
3311static __always_inline int
3312trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
3313{
3314 unsigned int val = cpu_buffer->current_context;
3315 int bit = interrupt_context_level();
3316
3317 bit = RB_CTX_NORMAL - bit;
3318
3319 if (unlikely(val & (1 << (bit + cpu_buffer->nest)))) {
3320 /*
3321 * It is possible that this was called by transitioning
3322 * between interrupt context, and preempt_count() has not
3323 * been updated yet. In this case, use the TRANSITION bit.
3324 */
3325 bit = RB_CTX_TRANSITION;
3326 if (val & (1 << (bit + cpu_buffer->nest))) {
3327 do_ring_buffer_record_recursion();
3328 return 1;
3329 }
3330 }
3331
3332 val |= (1 << (bit + cpu_buffer->nest));
3333 cpu_buffer->current_context = val;
3334
3335 return 0;
3336}
3337
3338static __always_inline void
3339trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
3340{
3341 cpu_buffer->current_context &=
3342 cpu_buffer->current_context - (1 << cpu_buffer->nest);
3343}
3344
3345/* The recursive locking above uses 5 bits */
3346#define NESTED_BITS 5
3347
3348/**
3349 * ring_buffer_nest_start - Allow to trace while nested
3350 * @buffer: The ring buffer to modify
3351 *
3352 * The ring buffer has a safety mechanism to prevent recursion.
3353 * But there may be a case where a trace needs to be done while
3354 * tracing something else. In this case, calling this function
3355 * will allow this function to nest within a currently active
3356 * ring_buffer_lock_reserve().
3357 *
3358 * Call this function before calling another ring_buffer_lock_reserve() and
3359 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
3360 */
3361void ring_buffer_nest_start(struct trace_buffer *buffer)
3362{
3363 struct ring_buffer_per_cpu *cpu_buffer;
3364 int cpu;
3365
3366 /* Enabled by ring_buffer_nest_end() */
3367 preempt_disable_notrace();
3368 cpu = raw_smp_processor_id();
3369 cpu_buffer = buffer->buffers[cpu];
3370 /* This is the shift value for the above recursive locking */
3371 cpu_buffer->nest += NESTED_BITS;
3372}
3373
3374/**
3375 * ring_buffer_nest_end - Allow to trace while nested
3376 * @buffer: The ring buffer to modify
3377 *
3378 * Must be called after ring_buffer_nest_start() and after the
3379 * ring_buffer_unlock_commit().
3380 */
3381void ring_buffer_nest_end(struct trace_buffer *buffer)
3382{
3383 struct ring_buffer_per_cpu *cpu_buffer;
3384 int cpu;
3385
3386 /* disabled by ring_buffer_nest_start() */
3387 cpu = raw_smp_processor_id();
3388 cpu_buffer = buffer->buffers[cpu];
3389 /* This is the shift value for the above recursive locking */
3390 cpu_buffer->nest -= NESTED_BITS;
3391 preempt_enable_notrace();
3392}
3393
3394/**
3395 * ring_buffer_unlock_commit - commit a reserved
3396 * @buffer: The buffer to commit to
3397 * @event: The event pointer to commit.
3398 *
3399 * This commits the data to the ring buffer, and releases any locks held.
3400 *
3401 * Must be paired with ring_buffer_lock_reserve.
3402 */
3403int ring_buffer_unlock_commit(struct trace_buffer *buffer)
3404{
3405 struct ring_buffer_per_cpu *cpu_buffer;
3406 int cpu = raw_smp_processor_id();
3407
3408 cpu_buffer = buffer->buffers[cpu];
3409
3410 rb_commit(cpu_buffer);
3411
3412 rb_wakeups(buffer, cpu_buffer);
3413
3414 trace_recursive_unlock(cpu_buffer);
3415
3416 preempt_enable_notrace();
3417
3418 return 0;
3419}
3420EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
3421
3422/* Special value to validate all deltas on a page. */
3423#define CHECK_FULL_PAGE 1L
3424
3425#ifdef CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS
3426static void dump_buffer_page(struct buffer_data_page *bpage,
3427 struct rb_event_info *info,
3428 unsigned long tail)
3429{
3430 struct ring_buffer_event *event;
3431 u64 ts, delta;
3432 int e;
3433
3434 ts = bpage->time_stamp;
3435 pr_warn(" [%lld] PAGE TIME STAMP\n", ts);
3436
3437 for (e = 0; e < tail; e += rb_event_length(event)) {
3438
3439 event = (struct ring_buffer_event *)(bpage->data + e);
3440
3441 switch (event->type_len) {
3442
3443 case RINGBUF_TYPE_TIME_EXTEND:
3444 delta = rb_event_time_stamp(event);
3445 ts += delta;
3446 pr_warn(" [%lld] delta:%lld TIME EXTEND\n", ts, delta);
3447 break;
3448
3449 case RINGBUF_TYPE_TIME_STAMP:
3450 delta = rb_event_time_stamp(event);
3451 ts = rb_fix_abs_ts(delta, ts);
3452 pr_warn(" [%lld] absolute:%lld TIME STAMP\n", ts, delta);
3453 break;
3454
3455 case RINGBUF_TYPE_PADDING:
3456 ts += event->time_delta;
3457 pr_warn(" [%lld] delta:%d PADDING\n", ts, event->time_delta);
3458 break;
3459
3460 case RINGBUF_TYPE_DATA:
3461 ts += event->time_delta;
3462 pr_warn(" [%lld] delta:%d\n", ts, event->time_delta);
3463 break;
3464
3465 default:
3466 break;
3467 }
3468 }
3469}
3470
3471static DEFINE_PER_CPU(atomic_t, checking);
3472static atomic_t ts_dump;
3473
3474/*
3475 * Check if the current event time stamp matches the deltas on
3476 * the buffer page.
3477 */
3478static void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3479 struct rb_event_info *info,
3480 unsigned long tail)
3481{
3482 struct ring_buffer_event *event;
3483 struct buffer_data_page *bpage;
3484 u64 ts, delta;
3485 bool full = false;
3486 int e;
3487
3488 bpage = info->tail_page->page;
3489
3490 if (tail == CHECK_FULL_PAGE) {
3491 full = true;
3492 tail = local_read(&bpage->commit);
3493 } else if (info->add_timestamp &
3494 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE)) {
3495 /* Ignore events with absolute time stamps */
3496 return;
3497 }
3498
3499 /*
3500 * Do not check the first event (skip possible extends too).
3501 * Also do not check if previous events have not been committed.
3502 */
3503 if (tail <= 8 || tail > local_read(&bpage->commit))
3504 return;
3505
3506 /*
3507 * If this interrupted another event,
3508 */
3509 if (atomic_inc_return(this_cpu_ptr(&checking)) != 1)
3510 goto out;
3511
3512 ts = bpage->time_stamp;
3513
3514 for (e = 0; e < tail; e += rb_event_length(event)) {
3515
3516 event = (struct ring_buffer_event *)(bpage->data + e);
3517
3518 switch (event->type_len) {
3519
3520 case RINGBUF_TYPE_TIME_EXTEND:
3521 delta = rb_event_time_stamp(event);
3522 ts += delta;
3523 break;
3524
3525 case RINGBUF_TYPE_TIME_STAMP:
3526 delta = rb_event_time_stamp(event);
3527 ts = rb_fix_abs_ts(delta, ts);
3528 break;
3529
3530 case RINGBUF_TYPE_PADDING:
3531 if (event->time_delta == 1)
3532 break;
3533 fallthrough;
3534 case RINGBUF_TYPE_DATA:
3535 ts += event->time_delta;
3536 break;
3537
3538 default:
3539 RB_WARN_ON(cpu_buffer, 1);
3540 }
3541 }
3542 if ((full && ts > info->ts) ||
3543 (!full && ts + info->delta != info->ts)) {
3544 /* If another report is happening, ignore this one */
3545 if (atomic_inc_return(&ts_dump) != 1) {
3546 atomic_dec(&ts_dump);
3547 goto out;
3548 }
3549 atomic_inc(&cpu_buffer->record_disabled);
3550 /* There's some cases in boot up that this can happen */
3551 WARN_ON_ONCE(system_state != SYSTEM_BOOTING);
3552 pr_warn("[CPU: %d]TIME DOES NOT MATCH expected:%lld actual:%lld delta:%lld before:%lld after:%lld%s\n",
3553 cpu_buffer->cpu,
3554 ts + info->delta, info->ts, info->delta,
3555 info->before, info->after,
3556 full ? " (full)" : "");
3557 dump_buffer_page(bpage, info, tail);
3558 atomic_dec(&ts_dump);
3559 /* Do not re-enable checking */
3560 return;
3561 }
3562out:
3563 atomic_dec(this_cpu_ptr(&checking));
3564}
3565#else
3566static inline void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3567 struct rb_event_info *info,
3568 unsigned long tail)
3569{
3570}
3571#endif /* CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS */
3572
3573static struct ring_buffer_event *
3574__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
3575 struct rb_event_info *info)
3576{
3577 struct ring_buffer_event *event;
3578 struct buffer_page *tail_page;
3579 unsigned long tail, write, w;
3580 bool a_ok;
3581 bool b_ok;
3582
3583 /* Don't let the compiler play games with cpu_buffer->tail_page */
3584 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
3585
3586 /*A*/ w = local_read(&tail_page->write) & RB_WRITE_MASK;
3587 barrier();
3588 b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3589 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3590 barrier();
3591 info->ts = rb_time_stamp(cpu_buffer->buffer);
3592
3593 if ((info->add_timestamp & RB_ADD_STAMP_ABSOLUTE)) {
3594 info->delta = info->ts;
3595 } else {
3596 /*
3597 * If interrupting an event time update, we may need an
3598 * absolute timestamp.
3599 * Don't bother if this is the start of a new page (w == 0).
3600 */
3601 if (unlikely(!a_ok || !b_ok || (info->before != info->after && w))) {
3602 info->add_timestamp |= RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND;
3603 info->length += RB_LEN_TIME_EXTEND;
3604 } else {
3605 info->delta = info->ts - info->after;
3606 if (unlikely(test_time_stamp(info->delta))) {
3607 info->add_timestamp |= RB_ADD_STAMP_EXTEND;
3608 info->length += RB_LEN_TIME_EXTEND;
3609 }
3610 }
3611 }
3612
3613 /*B*/ rb_time_set(&cpu_buffer->before_stamp, info->ts);
3614
3615 /*C*/ write = local_add_return(info->length, &tail_page->write);
3616
3617 /* set write to only the index of the write */
3618 write &= RB_WRITE_MASK;
3619
3620 tail = write - info->length;
3621
3622 /* See if we shot pass the end of this buffer page */
3623 if (unlikely(write > BUF_PAGE_SIZE)) {
3624 /* before and after may now different, fix it up*/
3625 b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3626 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3627 if (a_ok && b_ok && info->before != info->after)
3628 (void)rb_time_cmpxchg(&cpu_buffer->before_stamp,
3629 info->before, info->after);
3630 if (a_ok && b_ok)
3631 check_buffer(cpu_buffer, info, CHECK_FULL_PAGE);
3632 return rb_move_tail(cpu_buffer, tail, info);
3633 }
3634
3635 if (likely(tail == w)) {
3636 u64 save_before;
3637 bool s_ok;
3638
3639 /* Nothing interrupted us between A and C */
3640 /*D*/ rb_time_set(&cpu_buffer->write_stamp, info->ts);
3641 barrier();
3642 /*E*/ s_ok = rb_time_read(&cpu_buffer->before_stamp, &save_before);
3643 RB_WARN_ON(cpu_buffer, !s_ok);
3644 if (likely(!(info->add_timestamp &
3645 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3646 /* This did not interrupt any time update */
3647 info->delta = info->ts - info->after;
3648 else
3649 /* Just use full timestamp for interrupting event */
3650 info->delta = info->ts;
3651 barrier();
3652 check_buffer(cpu_buffer, info, tail);
3653 if (unlikely(info->ts != save_before)) {
3654 /* SLOW PATH - Interrupted between C and E */
3655
3656 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3657 RB_WARN_ON(cpu_buffer, !a_ok);
3658
3659 /* Write stamp must only go forward */
3660 if (save_before > info->after) {
3661 /*
3662 * We do not care about the result, only that
3663 * it gets updated atomically.
3664 */
3665 (void)rb_time_cmpxchg(&cpu_buffer->write_stamp,
3666 info->after, save_before);
3667 }
3668 }
3669 } else {
3670 u64 ts;
3671 /* SLOW PATH - Interrupted between A and C */
3672 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3673 /* Was interrupted before here, write_stamp must be valid */
3674 RB_WARN_ON(cpu_buffer, !a_ok);
3675 ts = rb_time_stamp(cpu_buffer->buffer);
3676 barrier();
3677 /*E*/ if (write == (local_read(&tail_page->write) & RB_WRITE_MASK) &&
3678 info->after < ts &&
3679 rb_time_cmpxchg(&cpu_buffer->write_stamp,
3680 info->after, ts)) {
3681 /* Nothing came after this event between C and E */
3682 info->delta = ts - info->after;
3683 } else {
3684 /*
3685 * Interrupted between C and E:
3686 * Lost the previous events time stamp. Just set the
3687 * delta to zero, and this will be the same time as
3688 * the event this event interrupted. And the events that
3689 * came after this will still be correct (as they would
3690 * have built their delta on the previous event.
3691 */
3692 info->delta = 0;
3693 }
3694 info->ts = ts;
3695 info->add_timestamp &= ~RB_ADD_STAMP_FORCE;
3696 }
3697
3698 /*
3699 * If this is the first commit on the page, then it has the same
3700 * timestamp as the page itself.
3701 */
3702 if (unlikely(!tail && !(info->add_timestamp &
3703 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3704 info->delta = 0;
3705
3706 /* We reserved something on the buffer */
3707
3708 event = __rb_page_index(tail_page, tail);
3709 rb_update_event(cpu_buffer, event, info);
3710
3711 local_inc(&tail_page->entries);
3712
3713 /*
3714 * If this is the first commit on the page, then update
3715 * its timestamp.
3716 */
3717 if (unlikely(!tail))
3718 tail_page->page->time_stamp = info->ts;
3719
3720 /* account for these added bytes */
3721 local_add(info->length, &cpu_buffer->entries_bytes);
3722
3723 return event;
3724}
3725
3726static __always_inline struct ring_buffer_event *
3727rb_reserve_next_event(struct trace_buffer *buffer,
3728 struct ring_buffer_per_cpu *cpu_buffer,
3729 unsigned long length)
3730{
3731 struct ring_buffer_event *event;
3732 struct rb_event_info info;
3733 int nr_loops = 0;
3734 int add_ts_default;
3735
3736 rb_start_commit(cpu_buffer);
3737 /* The commit page can not change after this */
3738
3739#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3740 /*
3741 * Due to the ability to swap a cpu buffer from a buffer
3742 * it is possible it was swapped before we committed.
3743 * (committing stops a swap). We check for it here and
3744 * if it happened, we have to fail the write.
3745 */
3746 barrier();
3747 if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
3748 local_dec(&cpu_buffer->committing);
3749 local_dec(&cpu_buffer->commits);
3750 return NULL;
3751 }
3752#endif
3753
3754 info.length = rb_calculate_event_length(length);
3755
3756 if (ring_buffer_time_stamp_abs(cpu_buffer->buffer)) {
3757 add_ts_default = RB_ADD_STAMP_ABSOLUTE;
3758 info.length += RB_LEN_TIME_EXTEND;
3759 } else {
3760 add_ts_default = RB_ADD_STAMP_NONE;
3761 }
3762
3763 again:
3764 info.add_timestamp = add_ts_default;
3765 info.delta = 0;
3766
3767 /*
3768 * We allow for interrupts to reenter here and do a trace.
3769 * If one does, it will cause this original code to loop
3770 * back here. Even with heavy interrupts happening, this
3771 * should only happen a few times in a row. If this happens
3772 * 1000 times in a row, there must be either an interrupt
3773 * storm or we have something buggy.
3774 * Bail!
3775 */
3776 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
3777 goto out_fail;
3778
3779 event = __rb_reserve_next(cpu_buffer, &info);
3780
3781 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
3782 if (info.add_timestamp & (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND))
3783 info.length -= RB_LEN_TIME_EXTEND;
3784 goto again;
3785 }
3786
3787 if (likely(event))
3788 return event;
3789 out_fail:
3790 rb_end_commit(cpu_buffer);
3791 return NULL;
3792}
3793
3794/**
3795 * ring_buffer_lock_reserve - reserve a part of the buffer
3796 * @buffer: the ring buffer to reserve from
3797 * @length: the length of the data to reserve (excluding event header)
3798 *
3799 * Returns a reserved event on the ring buffer to copy directly to.
3800 * The user of this interface will need to get the body to write into
3801 * and can use the ring_buffer_event_data() interface.
3802 *
3803 * The length is the length of the data needed, not the event length
3804 * which also includes the event header.
3805 *
3806 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
3807 * If NULL is returned, then nothing has been allocated or locked.
3808 */
3809struct ring_buffer_event *
3810ring_buffer_lock_reserve(struct trace_buffer *buffer, unsigned long length)
3811{
3812 struct ring_buffer_per_cpu *cpu_buffer;
3813 struct ring_buffer_event *event;
3814 int cpu;
3815
3816 /* If we are tracing schedule, we don't want to recurse */
3817 preempt_disable_notrace();
3818
3819 if (unlikely(atomic_read(&buffer->record_disabled)))
3820 goto out;
3821
3822 cpu = raw_smp_processor_id();
3823
3824 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
3825 goto out;
3826
3827 cpu_buffer = buffer->buffers[cpu];
3828
3829 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
3830 goto out;
3831
3832 if (unlikely(length > BUF_MAX_DATA_SIZE))
3833 goto out;
3834
3835 if (unlikely(trace_recursive_lock(cpu_buffer)))
3836 goto out;
3837
3838 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3839 if (!event)
3840 goto out_unlock;
3841
3842 return event;
3843
3844 out_unlock:
3845 trace_recursive_unlock(cpu_buffer);
3846 out:
3847 preempt_enable_notrace();
3848 return NULL;
3849}
3850EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3851
3852/*
3853 * Decrement the entries to the page that an event is on.
3854 * The event does not even need to exist, only the pointer
3855 * to the page it is on. This may only be called before the commit
3856 * takes place.
3857 */
3858static inline void
3859rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3860 struct ring_buffer_event *event)
3861{
3862 unsigned long addr = (unsigned long)event;
3863 struct buffer_page *bpage = cpu_buffer->commit_page;
3864 struct buffer_page *start;
3865
3866 addr &= PAGE_MASK;
3867
3868 /* Do the likely case first */
3869 if (likely(bpage->page == (void *)addr)) {
3870 local_dec(&bpage->entries);
3871 return;
3872 }
3873
3874 /*
3875 * Because the commit page may be on the reader page we
3876 * start with the next page and check the end loop there.
3877 */
3878 rb_inc_page(&bpage);
3879 start = bpage;
3880 do {
3881 if (bpage->page == (void *)addr) {
3882 local_dec(&bpage->entries);
3883 return;
3884 }
3885 rb_inc_page(&bpage);
3886 } while (bpage != start);
3887
3888 /* commit not part of this buffer?? */
3889 RB_WARN_ON(cpu_buffer, 1);
3890}
3891
3892/**
3893 * ring_buffer_discard_commit - discard an event that has not been committed
3894 * @buffer: the ring buffer
3895 * @event: non committed event to discard
3896 *
3897 * Sometimes an event that is in the ring buffer needs to be ignored.
3898 * This function lets the user discard an event in the ring buffer
3899 * and then that event will not be read later.
3900 *
3901 * This function only works if it is called before the item has been
3902 * committed. It will try to free the event from the ring buffer
3903 * if another event has not been added behind it.
3904 *
3905 * If another event has been added behind it, it will set the event
3906 * up as discarded, and perform the commit.
3907 *
3908 * If this function is called, do not call ring_buffer_unlock_commit on
3909 * the event.
3910 */
3911void ring_buffer_discard_commit(struct trace_buffer *buffer,
3912 struct ring_buffer_event *event)
3913{
3914 struct ring_buffer_per_cpu *cpu_buffer;
3915 int cpu;
3916
3917 /* The event is discarded regardless */
3918 rb_event_discard(event);
3919
3920 cpu = smp_processor_id();
3921 cpu_buffer = buffer->buffers[cpu];
3922
3923 /*
3924 * This must only be called if the event has not been
3925 * committed yet. Thus we can assume that preemption
3926 * is still disabled.
3927 */
3928 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3929
3930 rb_decrement_entry(cpu_buffer, event);
3931 if (rb_try_to_discard(cpu_buffer, event))
3932 goto out;
3933
3934 out:
3935 rb_end_commit(cpu_buffer);
3936
3937 trace_recursive_unlock(cpu_buffer);
3938
3939 preempt_enable_notrace();
3940
3941}
3942EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3943
3944/**
3945 * ring_buffer_write - write data to the buffer without reserving
3946 * @buffer: The ring buffer to write to.
3947 * @length: The length of the data being written (excluding the event header)
3948 * @data: The data to write to the buffer.
3949 *
3950 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3951 * one function. If you already have the data to write to the buffer, it
3952 * may be easier to simply call this function.
3953 *
3954 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3955 * and not the length of the event which would hold the header.
3956 */
3957int ring_buffer_write(struct trace_buffer *buffer,
3958 unsigned long length,
3959 void *data)
3960{
3961 struct ring_buffer_per_cpu *cpu_buffer;
3962 struct ring_buffer_event *event;
3963 void *body;
3964 int ret = -EBUSY;
3965 int cpu;
3966
3967 preempt_disable_notrace();
3968
3969 if (atomic_read(&buffer->record_disabled))
3970 goto out;
3971
3972 cpu = raw_smp_processor_id();
3973
3974 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3975 goto out;
3976
3977 cpu_buffer = buffer->buffers[cpu];
3978
3979 if (atomic_read(&cpu_buffer->record_disabled))
3980 goto out;
3981
3982 if (length > BUF_MAX_DATA_SIZE)
3983 goto out;
3984
3985 if (unlikely(trace_recursive_lock(cpu_buffer)))
3986 goto out;
3987
3988 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3989 if (!event)
3990 goto out_unlock;
3991
3992 body = rb_event_data(event);
3993
3994 memcpy(body, data, length);
3995
3996 rb_commit(cpu_buffer);
3997
3998 rb_wakeups(buffer, cpu_buffer);
3999
4000 ret = 0;
4001
4002 out_unlock:
4003 trace_recursive_unlock(cpu_buffer);
4004
4005 out:
4006 preempt_enable_notrace();
4007
4008 return ret;
4009}
4010EXPORT_SYMBOL_GPL(ring_buffer_write);
4011
4012static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
4013{
4014 struct buffer_page *reader = cpu_buffer->reader_page;
4015 struct buffer_page *head = rb_set_head_page(cpu_buffer);
4016 struct buffer_page *commit = cpu_buffer->commit_page;
4017
4018 /* In case of error, head will be NULL */
4019 if (unlikely(!head))
4020 return true;
4021
4022 /* Reader should exhaust content in reader page */
4023 if (reader->read != rb_page_commit(reader))
4024 return false;
4025
4026 /*
4027 * If writers are committing on the reader page, knowing all
4028 * committed content has been read, the ring buffer is empty.
4029 */
4030 if (commit == reader)
4031 return true;
4032
4033 /*
4034 * If writers are committing on a page other than reader page
4035 * and head page, there should always be content to read.
4036 */
4037 if (commit != head)
4038 return false;
4039
4040 /*
4041 * Writers are committing on the head page, we just need
4042 * to care about there're committed data, and the reader will
4043 * swap reader page with head page when it is to read data.
4044 */
4045 return rb_page_commit(commit) == 0;
4046}
4047
4048/**
4049 * ring_buffer_record_disable - stop all writes into the buffer
4050 * @buffer: The ring buffer to stop writes to.
4051 *
4052 * This prevents all writes to the buffer. Any attempt to write
4053 * to the buffer after this will fail and return NULL.
4054 *
4055 * The caller should call synchronize_rcu() after this.
4056 */
4057void ring_buffer_record_disable(struct trace_buffer *buffer)
4058{
4059 atomic_inc(&buffer->record_disabled);
4060}
4061EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
4062
4063/**
4064 * ring_buffer_record_enable - enable writes to the buffer
4065 * @buffer: The ring buffer to enable writes
4066 *
4067 * Note, multiple disables will need the same number of enables
4068 * to truly enable the writing (much like preempt_disable).
4069 */
4070void ring_buffer_record_enable(struct trace_buffer *buffer)
4071{
4072 atomic_dec(&buffer->record_disabled);
4073}
4074EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
4075
4076/**
4077 * ring_buffer_record_off - stop all writes into the buffer
4078 * @buffer: The ring buffer to stop writes to.
4079 *
4080 * This prevents all writes to the buffer. Any attempt to write
4081 * to the buffer after this will fail and return NULL.
4082 *
4083 * This is different than ring_buffer_record_disable() as
4084 * it works like an on/off switch, where as the disable() version
4085 * must be paired with a enable().
4086 */
4087void ring_buffer_record_off(struct trace_buffer *buffer)
4088{
4089 unsigned int rd;
4090 unsigned int new_rd;
4091
4092 do {
4093 rd = atomic_read(&buffer->record_disabled);
4094 new_rd = rd | RB_BUFFER_OFF;
4095 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
4096}
4097EXPORT_SYMBOL_GPL(ring_buffer_record_off);
4098
4099/**
4100 * ring_buffer_record_on - restart writes into the buffer
4101 * @buffer: The ring buffer to start writes to.
4102 *
4103 * This enables all writes to the buffer that was disabled by
4104 * ring_buffer_record_off().
4105 *
4106 * This is different than ring_buffer_record_enable() as
4107 * it works like an on/off switch, where as the enable() version
4108 * must be paired with a disable().
4109 */
4110void ring_buffer_record_on(struct trace_buffer *buffer)
4111{
4112 unsigned int rd;
4113 unsigned int new_rd;
4114
4115 do {
4116 rd = atomic_read(&buffer->record_disabled);
4117 new_rd = rd & ~RB_BUFFER_OFF;
4118 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
4119}
4120EXPORT_SYMBOL_GPL(ring_buffer_record_on);
4121
4122/**
4123 * ring_buffer_record_is_on - return true if the ring buffer can write
4124 * @buffer: The ring buffer to see if write is enabled
4125 *
4126 * Returns true if the ring buffer is in a state that it accepts writes.
4127 */
4128bool ring_buffer_record_is_on(struct trace_buffer *buffer)
4129{
4130 return !atomic_read(&buffer->record_disabled);
4131}
4132
4133/**
4134 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
4135 * @buffer: The ring buffer to see if write is set enabled
4136 *
4137 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
4138 * Note that this does NOT mean it is in a writable state.
4139 *
4140 * It may return true when the ring buffer has been disabled by
4141 * ring_buffer_record_disable(), as that is a temporary disabling of
4142 * the ring buffer.
4143 */
4144bool ring_buffer_record_is_set_on(struct trace_buffer *buffer)
4145{
4146 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
4147}
4148
4149/**
4150 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
4151 * @buffer: The ring buffer to stop writes to.
4152 * @cpu: The CPU buffer to stop
4153 *
4154 * This prevents all writes to the buffer. Any attempt to write
4155 * to the buffer after this will fail and return NULL.
4156 *
4157 * The caller should call synchronize_rcu() after this.
4158 */
4159void ring_buffer_record_disable_cpu(struct trace_buffer *buffer, int cpu)
4160{
4161 struct ring_buffer_per_cpu *cpu_buffer;
4162
4163 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4164 return;
4165
4166 cpu_buffer = buffer->buffers[cpu];
4167 atomic_inc(&cpu_buffer->record_disabled);
4168}
4169EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
4170
4171/**
4172 * ring_buffer_record_enable_cpu - enable writes to the buffer
4173 * @buffer: The ring buffer to enable writes
4174 * @cpu: The CPU to enable.
4175 *
4176 * Note, multiple disables will need the same number of enables
4177 * to truly enable the writing (much like preempt_disable).
4178 */
4179void ring_buffer_record_enable_cpu(struct trace_buffer *buffer, int cpu)
4180{
4181 struct ring_buffer_per_cpu *cpu_buffer;
4182
4183 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4184 return;
4185
4186 cpu_buffer = buffer->buffers[cpu];
4187 atomic_dec(&cpu_buffer->record_disabled);
4188}
4189EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
4190
4191/*
4192 * The total entries in the ring buffer is the running counter
4193 * of entries entered into the ring buffer, minus the sum of
4194 * the entries read from the ring buffer and the number of
4195 * entries that were overwritten.
4196 */
4197static inline unsigned long
4198rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
4199{
4200 return local_read(&cpu_buffer->entries) -
4201 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
4202}
4203
4204/**
4205 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
4206 * @buffer: The ring buffer
4207 * @cpu: The per CPU buffer to read from.
4208 */
4209u64 ring_buffer_oldest_event_ts(struct trace_buffer *buffer, int cpu)
4210{
4211 unsigned long flags;
4212 struct ring_buffer_per_cpu *cpu_buffer;
4213 struct buffer_page *bpage;
4214 u64 ret = 0;
4215
4216 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4217 return 0;
4218
4219 cpu_buffer = buffer->buffers[cpu];
4220 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4221 /*
4222 * if the tail is on reader_page, oldest time stamp is on the reader
4223 * page
4224 */
4225 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
4226 bpage = cpu_buffer->reader_page;
4227 else
4228 bpage = rb_set_head_page(cpu_buffer);
4229 if (bpage)
4230 ret = bpage->page->time_stamp;
4231 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4232
4233 return ret;
4234}
4235EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
4236
4237/**
4238 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
4239 * @buffer: The ring buffer
4240 * @cpu: The per CPU buffer to read from.
4241 */
4242unsigned long ring_buffer_bytes_cpu(struct trace_buffer *buffer, int cpu)
4243{
4244 struct ring_buffer_per_cpu *cpu_buffer;
4245 unsigned long ret;
4246
4247 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4248 return 0;
4249
4250 cpu_buffer = buffer->buffers[cpu];
4251 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
4252
4253 return ret;
4254}
4255EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
4256
4257/**
4258 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
4259 * @buffer: The ring buffer
4260 * @cpu: The per CPU buffer to get the entries from.
4261 */
4262unsigned long ring_buffer_entries_cpu(struct trace_buffer *buffer, int cpu)
4263{
4264 struct ring_buffer_per_cpu *cpu_buffer;
4265
4266 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4267 return 0;
4268
4269 cpu_buffer = buffer->buffers[cpu];
4270
4271 return rb_num_of_entries(cpu_buffer);
4272}
4273EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
4274
4275/**
4276 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
4277 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
4278 * @buffer: The ring buffer
4279 * @cpu: The per CPU buffer to get the number of overruns from
4280 */
4281unsigned long ring_buffer_overrun_cpu(struct trace_buffer *buffer, int cpu)
4282{
4283 struct ring_buffer_per_cpu *cpu_buffer;
4284 unsigned long ret;
4285
4286 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4287 return 0;
4288
4289 cpu_buffer = buffer->buffers[cpu];
4290 ret = local_read(&cpu_buffer->overrun);
4291
4292 return ret;
4293}
4294EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
4295
4296/**
4297 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
4298 * commits failing due to the buffer wrapping around while there are uncommitted
4299 * events, such as during an interrupt storm.
4300 * @buffer: The ring buffer
4301 * @cpu: The per CPU buffer to get the number of overruns from
4302 */
4303unsigned long
4304ring_buffer_commit_overrun_cpu(struct trace_buffer *buffer, int cpu)
4305{
4306 struct ring_buffer_per_cpu *cpu_buffer;
4307 unsigned long ret;
4308
4309 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4310 return 0;
4311
4312 cpu_buffer = buffer->buffers[cpu];
4313 ret = local_read(&cpu_buffer->commit_overrun);
4314
4315 return ret;
4316}
4317EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
4318
4319/**
4320 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
4321 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
4322 * @buffer: The ring buffer
4323 * @cpu: The per CPU buffer to get the number of overruns from
4324 */
4325unsigned long
4326ring_buffer_dropped_events_cpu(struct trace_buffer *buffer, int cpu)
4327{
4328 struct ring_buffer_per_cpu *cpu_buffer;
4329 unsigned long ret;
4330
4331 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4332 return 0;
4333
4334 cpu_buffer = buffer->buffers[cpu];
4335 ret = local_read(&cpu_buffer->dropped_events);
4336
4337 return ret;
4338}
4339EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
4340
4341/**
4342 * ring_buffer_read_events_cpu - get the number of events successfully read
4343 * @buffer: The ring buffer
4344 * @cpu: The per CPU buffer to get the number of events read
4345 */
4346unsigned long
4347ring_buffer_read_events_cpu(struct trace_buffer *buffer, int cpu)
4348{
4349 struct ring_buffer_per_cpu *cpu_buffer;
4350
4351 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4352 return 0;
4353
4354 cpu_buffer = buffer->buffers[cpu];
4355 return cpu_buffer->read;
4356}
4357EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
4358
4359/**
4360 * ring_buffer_entries - get the number of entries in a buffer
4361 * @buffer: The ring buffer
4362 *
4363 * Returns the total number of entries in the ring buffer
4364 * (all CPU entries)
4365 */
4366unsigned long ring_buffer_entries(struct trace_buffer *buffer)
4367{
4368 struct ring_buffer_per_cpu *cpu_buffer;
4369 unsigned long entries = 0;
4370 int cpu;
4371
4372 /* if you care about this being correct, lock the buffer */
4373 for_each_buffer_cpu(buffer, cpu) {
4374 cpu_buffer = buffer->buffers[cpu];
4375 entries += rb_num_of_entries(cpu_buffer);
4376 }
4377
4378 return entries;
4379}
4380EXPORT_SYMBOL_GPL(ring_buffer_entries);
4381
4382/**
4383 * ring_buffer_overruns - get the number of overruns in buffer
4384 * @buffer: The ring buffer
4385 *
4386 * Returns the total number of overruns in the ring buffer
4387 * (all CPU entries)
4388 */
4389unsigned long ring_buffer_overruns(struct trace_buffer *buffer)
4390{
4391 struct ring_buffer_per_cpu *cpu_buffer;
4392 unsigned long overruns = 0;
4393 int cpu;
4394
4395 /* if you care about this being correct, lock the buffer */
4396 for_each_buffer_cpu(buffer, cpu) {
4397 cpu_buffer = buffer->buffers[cpu];
4398 overruns += local_read(&cpu_buffer->overrun);
4399 }
4400
4401 return overruns;
4402}
4403EXPORT_SYMBOL_GPL(ring_buffer_overruns);
4404
4405static void rb_iter_reset(struct ring_buffer_iter *iter)
4406{
4407 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4408
4409 /* Iterator usage is expected to have record disabled */
4410 iter->head_page = cpu_buffer->reader_page;
4411 iter->head = cpu_buffer->reader_page->read;
4412 iter->next_event = iter->head;
4413
4414 iter->cache_reader_page = iter->head_page;
4415 iter->cache_read = cpu_buffer->read;
4416
4417 if (iter->head) {
4418 iter->read_stamp = cpu_buffer->read_stamp;
4419 iter->page_stamp = cpu_buffer->reader_page->page->time_stamp;
4420 } else {
4421 iter->read_stamp = iter->head_page->page->time_stamp;
4422 iter->page_stamp = iter->read_stamp;
4423 }
4424}
4425
4426/**
4427 * ring_buffer_iter_reset - reset an iterator
4428 * @iter: The iterator to reset
4429 *
4430 * Resets the iterator, so that it will start from the beginning
4431 * again.
4432 */
4433void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
4434{
4435 struct ring_buffer_per_cpu *cpu_buffer;
4436 unsigned long flags;
4437
4438 if (!iter)
4439 return;
4440
4441 cpu_buffer = iter->cpu_buffer;
4442
4443 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4444 rb_iter_reset(iter);
4445 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4446}
4447EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
4448
4449/**
4450 * ring_buffer_iter_empty - check if an iterator has no more to read
4451 * @iter: The iterator to check
4452 */
4453int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
4454{
4455 struct ring_buffer_per_cpu *cpu_buffer;
4456 struct buffer_page *reader;
4457 struct buffer_page *head_page;
4458 struct buffer_page *commit_page;
4459 struct buffer_page *curr_commit_page;
4460 unsigned commit;
4461 u64 curr_commit_ts;
4462 u64 commit_ts;
4463
4464 cpu_buffer = iter->cpu_buffer;
4465 reader = cpu_buffer->reader_page;
4466 head_page = cpu_buffer->head_page;
4467 commit_page = cpu_buffer->commit_page;
4468 commit_ts = commit_page->page->time_stamp;
4469
4470 /*
4471 * When the writer goes across pages, it issues a cmpxchg which
4472 * is a mb(), which will synchronize with the rmb here.
4473 * (see rb_tail_page_update())
4474 */
4475 smp_rmb();
4476 commit = rb_page_commit(commit_page);
4477 /* We want to make sure that the commit page doesn't change */
4478 smp_rmb();
4479
4480 /* Make sure commit page didn't change */
4481 curr_commit_page = READ_ONCE(cpu_buffer->commit_page);
4482 curr_commit_ts = READ_ONCE(curr_commit_page->page->time_stamp);
4483
4484 /* If the commit page changed, then there's more data */
4485 if (curr_commit_page != commit_page ||
4486 curr_commit_ts != commit_ts)
4487 return 0;
4488
4489 /* Still racy, as it may return a false positive, but that's OK */
4490 return ((iter->head_page == commit_page && iter->head >= commit) ||
4491 (iter->head_page == reader && commit_page == head_page &&
4492 head_page->read == commit &&
4493 iter->head == rb_page_commit(cpu_buffer->reader_page)));
4494}
4495EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
4496
4497static void
4498rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
4499 struct ring_buffer_event *event)
4500{
4501 u64 delta;
4502
4503 switch (event->type_len) {
4504 case RINGBUF_TYPE_PADDING:
4505 return;
4506
4507 case RINGBUF_TYPE_TIME_EXTEND:
4508 delta = rb_event_time_stamp(event);
4509 cpu_buffer->read_stamp += delta;
4510 return;
4511
4512 case RINGBUF_TYPE_TIME_STAMP:
4513 delta = rb_event_time_stamp(event);
4514 delta = rb_fix_abs_ts(delta, cpu_buffer->read_stamp);
4515 cpu_buffer->read_stamp = delta;
4516 return;
4517
4518 case RINGBUF_TYPE_DATA:
4519 cpu_buffer->read_stamp += event->time_delta;
4520 return;
4521
4522 default:
4523 RB_WARN_ON(cpu_buffer, 1);
4524 }
4525 return;
4526}
4527
4528static void
4529rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
4530 struct ring_buffer_event *event)
4531{
4532 u64 delta;
4533
4534 switch (event->type_len) {
4535 case RINGBUF_TYPE_PADDING:
4536 return;
4537
4538 case RINGBUF_TYPE_TIME_EXTEND:
4539 delta = rb_event_time_stamp(event);
4540 iter->read_stamp += delta;
4541 return;
4542
4543 case RINGBUF_TYPE_TIME_STAMP:
4544 delta = rb_event_time_stamp(event);
4545 delta = rb_fix_abs_ts(delta, iter->read_stamp);
4546 iter->read_stamp = delta;
4547 return;
4548
4549 case RINGBUF_TYPE_DATA:
4550 iter->read_stamp += event->time_delta;
4551 return;
4552
4553 default:
4554 RB_WARN_ON(iter->cpu_buffer, 1);
4555 }
4556 return;
4557}
4558
4559static struct buffer_page *
4560rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
4561{
4562 struct buffer_page *reader = NULL;
4563 unsigned long overwrite;
4564 unsigned long flags;
4565 int nr_loops = 0;
4566 int ret;
4567
4568 local_irq_save(flags);
4569 arch_spin_lock(&cpu_buffer->lock);
4570
4571 again:
4572 /*
4573 * This should normally only loop twice. But because the
4574 * start of the reader inserts an empty page, it causes
4575 * a case where we will loop three times. There should be no
4576 * reason to loop four times (that I know of).
4577 */
4578 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
4579 reader = NULL;
4580 goto out;
4581 }
4582
4583 reader = cpu_buffer->reader_page;
4584
4585 /* If there's more to read, return this page */
4586 if (cpu_buffer->reader_page->read < rb_page_size(reader))
4587 goto out;
4588
4589 /* Never should we have an index greater than the size */
4590 if (RB_WARN_ON(cpu_buffer,
4591 cpu_buffer->reader_page->read > rb_page_size(reader)))
4592 goto out;
4593
4594 /* check if we caught up to the tail */
4595 reader = NULL;
4596 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
4597 goto out;
4598
4599 /* Don't bother swapping if the ring buffer is empty */
4600 if (rb_num_of_entries(cpu_buffer) == 0)
4601 goto out;
4602
4603 /*
4604 * Reset the reader page to size zero.
4605 */
4606 local_set(&cpu_buffer->reader_page->write, 0);
4607 local_set(&cpu_buffer->reader_page->entries, 0);
4608 local_set(&cpu_buffer->reader_page->page->commit, 0);
4609 cpu_buffer->reader_page->real_end = 0;
4610
4611 spin:
4612 /*
4613 * Splice the empty reader page into the list around the head.
4614 */
4615 reader = rb_set_head_page(cpu_buffer);
4616 if (!reader)
4617 goto out;
4618 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
4619 cpu_buffer->reader_page->list.prev = reader->list.prev;
4620
4621 /*
4622 * cpu_buffer->pages just needs to point to the buffer, it
4623 * has no specific buffer page to point to. Lets move it out
4624 * of our way so we don't accidentally swap it.
4625 */
4626 cpu_buffer->pages = reader->list.prev;
4627
4628 /* The reader page will be pointing to the new head */
4629 rb_set_list_to_head(&cpu_buffer->reader_page->list);
4630
4631 /*
4632 * We want to make sure we read the overruns after we set up our
4633 * pointers to the next object. The writer side does a
4634 * cmpxchg to cross pages which acts as the mb on the writer
4635 * side. Note, the reader will constantly fail the swap
4636 * while the writer is updating the pointers, so this
4637 * guarantees that the overwrite recorded here is the one we
4638 * want to compare with the last_overrun.
4639 */
4640 smp_mb();
4641 overwrite = local_read(&(cpu_buffer->overrun));
4642
4643 /*
4644 * Here's the tricky part.
4645 *
4646 * We need to move the pointer past the header page.
4647 * But we can only do that if a writer is not currently
4648 * moving it. The page before the header page has the
4649 * flag bit '1' set if it is pointing to the page we want.
4650 * but if the writer is in the process of moving it
4651 * than it will be '2' or already moved '0'.
4652 */
4653
4654 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
4655
4656 /*
4657 * If we did not convert it, then we must try again.
4658 */
4659 if (!ret)
4660 goto spin;
4661
4662 /*
4663 * Yay! We succeeded in replacing the page.
4664 *
4665 * Now make the new head point back to the reader page.
4666 */
4667 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
4668 rb_inc_page(&cpu_buffer->head_page);
4669
4670 local_inc(&cpu_buffer->pages_read);
4671
4672 /* Finally update the reader page to the new head */
4673 cpu_buffer->reader_page = reader;
4674 cpu_buffer->reader_page->read = 0;
4675
4676 if (overwrite != cpu_buffer->last_overrun) {
4677 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
4678 cpu_buffer->last_overrun = overwrite;
4679 }
4680
4681 goto again;
4682
4683 out:
4684 /* Update the read_stamp on the first event */
4685 if (reader && reader->read == 0)
4686 cpu_buffer->read_stamp = reader->page->time_stamp;
4687
4688 arch_spin_unlock(&cpu_buffer->lock);
4689 local_irq_restore(flags);
4690
4691 /*
4692 * The writer has preempt disable, wait for it. But not forever
4693 * Although, 1 second is pretty much "forever"
4694 */
4695#define USECS_WAIT 1000000
4696 for (nr_loops = 0; nr_loops < USECS_WAIT; nr_loops++) {
4697 /* If the write is past the end of page, a writer is still updating it */
4698 if (likely(!reader || rb_page_write(reader) <= BUF_PAGE_SIZE))
4699 break;
4700
4701 udelay(1);
4702
4703 /* Get the latest version of the reader write value */
4704 smp_rmb();
4705 }
4706
4707 /* The writer is not moving forward? Something is wrong */
4708 if (RB_WARN_ON(cpu_buffer, nr_loops == USECS_WAIT))
4709 reader = NULL;
4710
4711 /*
4712 * Make sure we see any padding after the write update
4713 * (see rb_reset_tail())
4714 */
4715 smp_rmb();
4716
4717
4718 return reader;
4719}
4720
4721static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
4722{
4723 struct ring_buffer_event *event;
4724 struct buffer_page *reader;
4725 unsigned length;
4726
4727 reader = rb_get_reader_page(cpu_buffer);
4728
4729 /* This function should not be called when buffer is empty */
4730 if (RB_WARN_ON(cpu_buffer, !reader))
4731 return;
4732
4733 event = rb_reader_event(cpu_buffer);
4734
4735 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
4736 cpu_buffer->read++;
4737
4738 rb_update_read_stamp(cpu_buffer, event);
4739
4740 length = rb_event_length(event);
4741 cpu_buffer->reader_page->read += length;
4742}
4743
4744static void rb_advance_iter(struct ring_buffer_iter *iter)
4745{
4746 struct ring_buffer_per_cpu *cpu_buffer;
4747
4748 cpu_buffer = iter->cpu_buffer;
4749
4750 /* If head == next_event then we need to jump to the next event */
4751 if (iter->head == iter->next_event) {
4752 /* If the event gets overwritten again, there's nothing to do */
4753 if (rb_iter_head_event(iter) == NULL)
4754 return;
4755 }
4756
4757 iter->head = iter->next_event;
4758
4759 /*
4760 * Check if we are at the end of the buffer.
4761 */
4762 if (iter->next_event >= rb_page_size(iter->head_page)) {
4763 /* discarded commits can make the page empty */
4764 if (iter->head_page == cpu_buffer->commit_page)
4765 return;
4766 rb_inc_iter(iter);
4767 return;
4768 }
4769
4770 rb_update_iter_read_stamp(iter, iter->event);
4771}
4772
4773static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
4774{
4775 return cpu_buffer->lost_events;
4776}
4777
4778static struct ring_buffer_event *
4779rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
4780 unsigned long *lost_events)
4781{
4782 struct ring_buffer_event *event;
4783 struct buffer_page *reader;
4784 int nr_loops = 0;
4785
4786 if (ts)
4787 *ts = 0;
4788 again:
4789 /*
4790 * We repeat when a time extend is encountered.
4791 * Since the time extend is always attached to a data event,
4792 * we should never loop more than once.
4793 * (We never hit the following condition more than twice).
4794 */
4795 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
4796 return NULL;
4797
4798 reader = rb_get_reader_page(cpu_buffer);
4799 if (!reader)
4800 return NULL;
4801
4802 event = rb_reader_event(cpu_buffer);
4803
4804 switch (event->type_len) {
4805 case RINGBUF_TYPE_PADDING:
4806 if (rb_null_event(event))
4807 RB_WARN_ON(cpu_buffer, 1);
4808 /*
4809 * Because the writer could be discarding every
4810 * event it creates (which would probably be bad)
4811 * if we were to go back to "again" then we may never
4812 * catch up, and will trigger the warn on, or lock
4813 * the box. Return the padding, and we will release
4814 * the current locks, and try again.
4815 */
4816 return event;
4817
4818 case RINGBUF_TYPE_TIME_EXTEND:
4819 /* Internal data, OK to advance */
4820 rb_advance_reader(cpu_buffer);
4821 goto again;
4822
4823 case RINGBUF_TYPE_TIME_STAMP:
4824 if (ts) {
4825 *ts = rb_event_time_stamp(event);
4826 *ts = rb_fix_abs_ts(*ts, reader->page->time_stamp);
4827 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4828 cpu_buffer->cpu, ts);
4829 }
4830 /* Internal data, OK to advance */
4831 rb_advance_reader(cpu_buffer);
4832 goto again;
4833
4834 case RINGBUF_TYPE_DATA:
4835 if (ts && !(*ts)) {
4836 *ts = cpu_buffer->read_stamp + event->time_delta;
4837 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4838 cpu_buffer->cpu, ts);
4839 }
4840 if (lost_events)
4841 *lost_events = rb_lost_events(cpu_buffer);
4842 return event;
4843
4844 default:
4845 RB_WARN_ON(cpu_buffer, 1);
4846 }
4847
4848 return NULL;
4849}
4850EXPORT_SYMBOL_GPL(ring_buffer_peek);
4851
4852static struct ring_buffer_event *
4853rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4854{
4855 struct trace_buffer *buffer;
4856 struct ring_buffer_per_cpu *cpu_buffer;
4857 struct ring_buffer_event *event;
4858 int nr_loops = 0;
4859
4860 if (ts)
4861 *ts = 0;
4862
4863 cpu_buffer = iter->cpu_buffer;
4864 buffer = cpu_buffer->buffer;
4865
4866 /*
4867 * Check if someone performed a consuming read to
4868 * the buffer. A consuming read invalidates the iterator
4869 * and we need to reset the iterator in this case.
4870 */
4871 if (unlikely(iter->cache_read != cpu_buffer->read ||
4872 iter->cache_reader_page != cpu_buffer->reader_page))
4873 rb_iter_reset(iter);
4874
4875 again:
4876 if (ring_buffer_iter_empty(iter))
4877 return NULL;
4878
4879 /*
4880 * As the writer can mess with what the iterator is trying
4881 * to read, just give up if we fail to get an event after
4882 * three tries. The iterator is not as reliable when reading
4883 * the ring buffer with an active write as the consumer is.
4884 * Do not warn if the three failures is reached.
4885 */
4886 if (++nr_loops > 3)
4887 return NULL;
4888
4889 if (rb_per_cpu_empty(cpu_buffer))
4890 return NULL;
4891
4892 if (iter->head >= rb_page_size(iter->head_page)) {
4893 rb_inc_iter(iter);
4894 goto again;
4895 }
4896
4897 event = rb_iter_head_event(iter);
4898 if (!event)
4899 goto again;
4900
4901 switch (event->type_len) {
4902 case RINGBUF_TYPE_PADDING:
4903 if (rb_null_event(event)) {
4904 rb_inc_iter(iter);
4905 goto again;
4906 }
4907 rb_advance_iter(iter);
4908 return event;
4909
4910 case RINGBUF_TYPE_TIME_EXTEND:
4911 /* Internal data, OK to advance */
4912 rb_advance_iter(iter);
4913 goto again;
4914
4915 case RINGBUF_TYPE_TIME_STAMP:
4916 if (ts) {
4917 *ts = rb_event_time_stamp(event);
4918 *ts = rb_fix_abs_ts(*ts, iter->head_page->page->time_stamp);
4919 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4920 cpu_buffer->cpu, ts);
4921 }
4922 /* Internal data, OK to advance */
4923 rb_advance_iter(iter);
4924 goto again;
4925
4926 case RINGBUF_TYPE_DATA:
4927 if (ts && !(*ts)) {
4928 *ts = iter->read_stamp + event->time_delta;
4929 ring_buffer_normalize_time_stamp(buffer,
4930 cpu_buffer->cpu, ts);
4931 }
4932 return event;
4933
4934 default:
4935 RB_WARN_ON(cpu_buffer, 1);
4936 }
4937
4938 return NULL;
4939}
4940EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4941
4942static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4943{
4944 if (likely(!in_nmi())) {
4945 raw_spin_lock(&cpu_buffer->reader_lock);
4946 return true;
4947 }
4948
4949 /*
4950 * If an NMI die dumps out the content of the ring buffer
4951 * trylock must be used to prevent a deadlock if the NMI
4952 * preempted a task that holds the ring buffer locks. If
4953 * we get the lock then all is fine, if not, then continue
4954 * to do the read, but this can corrupt the ring buffer,
4955 * so it must be permanently disabled from future writes.
4956 * Reading from NMI is a oneshot deal.
4957 */
4958 if (raw_spin_trylock(&cpu_buffer->reader_lock))
4959 return true;
4960
4961 /* Continue without locking, but disable the ring buffer */
4962 atomic_inc(&cpu_buffer->record_disabled);
4963 return false;
4964}
4965
4966static inline void
4967rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4968{
4969 if (likely(locked))
4970 raw_spin_unlock(&cpu_buffer->reader_lock);
4971 return;
4972}
4973
4974/**
4975 * ring_buffer_peek - peek at the next event to be read
4976 * @buffer: The ring buffer to read
4977 * @cpu: The cpu to peak at
4978 * @ts: The timestamp counter of this event.
4979 * @lost_events: a variable to store if events were lost (may be NULL)
4980 *
4981 * This will return the event that will be read next, but does
4982 * not consume the data.
4983 */
4984struct ring_buffer_event *
4985ring_buffer_peek(struct trace_buffer *buffer, int cpu, u64 *ts,
4986 unsigned long *lost_events)
4987{
4988 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4989 struct ring_buffer_event *event;
4990 unsigned long flags;
4991 bool dolock;
4992
4993 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4994 return NULL;
4995
4996 again:
4997 local_irq_save(flags);
4998 dolock = rb_reader_lock(cpu_buffer);
4999 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
5000 if (event && event->type_len == RINGBUF_TYPE_PADDING)
5001 rb_advance_reader(cpu_buffer);
5002 rb_reader_unlock(cpu_buffer, dolock);
5003 local_irq_restore(flags);
5004
5005 if (event && event->type_len == RINGBUF_TYPE_PADDING)
5006 goto again;
5007
5008 return event;
5009}
5010
5011/** ring_buffer_iter_dropped - report if there are dropped events
5012 * @iter: The ring buffer iterator
5013 *
5014 * Returns true if there was dropped events since the last peek.
5015 */
5016bool ring_buffer_iter_dropped(struct ring_buffer_iter *iter)
5017{
5018 bool ret = iter->missed_events != 0;
5019
5020 iter->missed_events = 0;
5021 return ret;
5022}
5023EXPORT_SYMBOL_GPL(ring_buffer_iter_dropped);
5024
5025/**
5026 * ring_buffer_iter_peek - peek at the next event to be read
5027 * @iter: The ring buffer iterator
5028 * @ts: The timestamp counter of this event.
5029 *
5030 * This will return the event that will be read next, but does
5031 * not increment the iterator.
5032 */
5033struct ring_buffer_event *
5034ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
5035{
5036 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5037 struct ring_buffer_event *event;
5038 unsigned long flags;
5039
5040 again:
5041 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5042 event = rb_iter_peek(iter, ts);
5043 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5044
5045 if (event && event->type_len == RINGBUF_TYPE_PADDING)
5046 goto again;
5047
5048 return event;
5049}
5050
5051/**
5052 * ring_buffer_consume - return an event and consume it
5053 * @buffer: The ring buffer to get the next event from
5054 * @cpu: the cpu to read the buffer from
5055 * @ts: a variable to store the timestamp (may be NULL)
5056 * @lost_events: a variable to store if events were lost (may be NULL)
5057 *
5058 * Returns the next event in the ring buffer, and that event is consumed.
5059 * Meaning, that sequential reads will keep returning a different event,
5060 * and eventually empty the ring buffer if the producer is slower.
5061 */
5062struct ring_buffer_event *
5063ring_buffer_consume(struct trace_buffer *buffer, int cpu, u64 *ts,
5064 unsigned long *lost_events)
5065{
5066 struct ring_buffer_per_cpu *cpu_buffer;
5067 struct ring_buffer_event *event = NULL;
5068 unsigned long flags;
5069 bool dolock;
5070
5071 again:
5072 /* might be called in atomic */
5073 preempt_disable();
5074
5075 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5076 goto out;
5077
5078 cpu_buffer = buffer->buffers[cpu];
5079 local_irq_save(flags);
5080 dolock = rb_reader_lock(cpu_buffer);
5081
5082 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
5083 if (event) {
5084 cpu_buffer->lost_events = 0;
5085 rb_advance_reader(cpu_buffer);
5086 }
5087
5088 rb_reader_unlock(cpu_buffer, dolock);
5089 local_irq_restore(flags);
5090
5091 out:
5092 preempt_enable();
5093
5094 if (event && event->type_len == RINGBUF_TYPE_PADDING)
5095 goto again;
5096
5097 return event;
5098}
5099EXPORT_SYMBOL_GPL(ring_buffer_consume);
5100
5101/**
5102 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
5103 * @buffer: The ring buffer to read from
5104 * @cpu: The cpu buffer to iterate over
5105 * @flags: gfp flags to use for memory allocation
5106 *
5107 * This performs the initial preparations necessary to iterate
5108 * through the buffer. Memory is allocated, buffer recording
5109 * is disabled, and the iterator pointer is returned to the caller.
5110 *
5111 * Disabling buffer recording prevents the reading from being
5112 * corrupted. This is not a consuming read, so a producer is not
5113 * expected.
5114 *
5115 * After a sequence of ring_buffer_read_prepare calls, the user is
5116 * expected to make at least one call to ring_buffer_read_prepare_sync.
5117 * Afterwards, ring_buffer_read_start is invoked to get things going
5118 * for real.
5119 *
5120 * This overall must be paired with ring_buffer_read_finish.
5121 */
5122struct ring_buffer_iter *
5123ring_buffer_read_prepare(struct trace_buffer *buffer, int cpu, gfp_t flags)
5124{
5125 struct ring_buffer_per_cpu *cpu_buffer;
5126 struct ring_buffer_iter *iter;
5127
5128 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5129 return NULL;
5130
5131 iter = kzalloc(sizeof(*iter), flags);
5132 if (!iter)
5133 return NULL;
5134
5135 iter->event = kmalloc(BUF_MAX_DATA_SIZE, flags);
5136 if (!iter->event) {
5137 kfree(iter);
5138 return NULL;
5139 }
5140
5141 cpu_buffer = buffer->buffers[cpu];
5142
5143 iter->cpu_buffer = cpu_buffer;
5144
5145 atomic_inc(&cpu_buffer->resize_disabled);
5146
5147 return iter;
5148}
5149EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
5150
5151/**
5152 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
5153 *
5154 * All previously invoked ring_buffer_read_prepare calls to prepare
5155 * iterators will be synchronized. Afterwards, read_buffer_read_start
5156 * calls on those iterators are allowed.
5157 */
5158void
5159ring_buffer_read_prepare_sync(void)
5160{
5161 synchronize_rcu();
5162}
5163EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
5164
5165/**
5166 * ring_buffer_read_start - start a non consuming read of the buffer
5167 * @iter: The iterator returned by ring_buffer_read_prepare
5168 *
5169 * This finalizes the startup of an iteration through the buffer.
5170 * The iterator comes from a call to ring_buffer_read_prepare and
5171 * an intervening ring_buffer_read_prepare_sync must have been
5172 * performed.
5173 *
5174 * Must be paired with ring_buffer_read_finish.
5175 */
5176void
5177ring_buffer_read_start(struct ring_buffer_iter *iter)
5178{
5179 struct ring_buffer_per_cpu *cpu_buffer;
5180 unsigned long flags;
5181
5182 if (!iter)
5183 return;
5184
5185 cpu_buffer = iter->cpu_buffer;
5186
5187 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5188 arch_spin_lock(&cpu_buffer->lock);
5189 rb_iter_reset(iter);
5190 arch_spin_unlock(&cpu_buffer->lock);
5191 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5192}
5193EXPORT_SYMBOL_GPL(ring_buffer_read_start);
5194
5195/**
5196 * ring_buffer_read_finish - finish reading the iterator of the buffer
5197 * @iter: The iterator retrieved by ring_buffer_start
5198 *
5199 * This re-enables the recording to the buffer, and frees the
5200 * iterator.
5201 */
5202void
5203ring_buffer_read_finish(struct ring_buffer_iter *iter)
5204{
5205 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5206 unsigned long flags;
5207
5208 /*
5209 * Ring buffer is disabled from recording, here's a good place
5210 * to check the integrity of the ring buffer.
5211 * Must prevent readers from trying to read, as the check
5212 * clears the HEAD page and readers require it.
5213 */
5214 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5215 rb_check_pages(cpu_buffer);
5216 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5217
5218 atomic_dec(&cpu_buffer->resize_disabled);
5219 kfree(iter->event);
5220 kfree(iter);
5221}
5222EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
5223
5224/**
5225 * ring_buffer_iter_advance - advance the iterator to the next location
5226 * @iter: The ring buffer iterator
5227 *
5228 * Move the location of the iterator such that the next read will
5229 * be the next location of the iterator.
5230 */
5231void ring_buffer_iter_advance(struct ring_buffer_iter *iter)
5232{
5233 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5234 unsigned long flags;
5235
5236 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5237
5238 rb_advance_iter(iter);
5239
5240 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5241}
5242EXPORT_SYMBOL_GPL(ring_buffer_iter_advance);
5243
5244/**
5245 * ring_buffer_size - return the size of the ring buffer (in bytes)
5246 * @buffer: The ring buffer.
5247 * @cpu: The CPU to get ring buffer size from.
5248 */
5249unsigned long ring_buffer_size(struct trace_buffer *buffer, int cpu)
5250{
5251 /*
5252 * Earlier, this method returned
5253 * BUF_PAGE_SIZE * buffer->nr_pages
5254 * Since the nr_pages field is now removed, we have converted this to
5255 * return the per cpu buffer value.
5256 */
5257 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5258 return 0;
5259
5260 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
5261}
5262EXPORT_SYMBOL_GPL(ring_buffer_size);
5263
5264static void
5265rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
5266{
5267 rb_head_page_deactivate(cpu_buffer);
5268
5269 cpu_buffer->head_page
5270 = list_entry(cpu_buffer->pages, struct buffer_page, list);
5271 local_set(&cpu_buffer->head_page->write, 0);
5272 local_set(&cpu_buffer->head_page->entries, 0);
5273 local_set(&cpu_buffer->head_page->page->commit, 0);
5274
5275 cpu_buffer->head_page->read = 0;
5276
5277 cpu_buffer->tail_page = cpu_buffer->head_page;
5278 cpu_buffer->commit_page = cpu_buffer->head_page;
5279
5280 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
5281 INIT_LIST_HEAD(&cpu_buffer->new_pages);
5282 local_set(&cpu_buffer->reader_page->write, 0);
5283 local_set(&cpu_buffer->reader_page->entries, 0);
5284 local_set(&cpu_buffer->reader_page->page->commit, 0);
5285 cpu_buffer->reader_page->read = 0;
5286
5287 local_set(&cpu_buffer->entries_bytes, 0);
5288 local_set(&cpu_buffer->overrun, 0);
5289 local_set(&cpu_buffer->commit_overrun, 0);
5290 local_set(&cpu_buffer->dropped_events, 0);
5291 local_set(&cpu_buffer->entries, 0);
5292 local_set(&cpu_buffer->committing, 0);
5293 local_set(&cpu_buffer->commits, 0);
5294 local_set(&cpu_buffer->pages_touched, 0);
5295 local_set(&cpu_buffer->pages_lost, 0);
5296 local_set(&cpu_buffer->pages_read, 0);
5297 cpu_buffer->last_pages_touch = 0;
5298 cpu_buffer->shortest_full = 0;
5299 cpu_buffer->read = 0;
5300 cpu_buffer->read_bytes = 0;
5301
5302 rb_time_set(&cpu_buffer->write_stamp, 0);
5303 rb_time_set(&cpu_buffer->before_stamp, 0);
5304
5305 memset(cpu_buffer->event_stamp, 0, sizeof(cpu_buffer->event_stamp));
5306
5307 cpu_buffer->lost_events = 0;
5308 cpu_buffer->last_overrun = 0;
5309
5310 rb_head_page_activate(cpu_buffer);
5311}
5312
5313/* Must have disabled the cpu buffer then done a synchronize_rcu */
5314static void reset_disabled_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
5315{
5316 unsigned long flags;
5317
5318 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5319
5320 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
5321 goto out;
5322
5323 arch_spin_lock(&cpu_buffer->lock);
5324
5325 rb_reset_cpu(cpu_buffer);
5326
5327 arch_spin_unlock(&cpu_buffer->lock);
5328
5329 out:
5330 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5331}
5332
5333/**
5334 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
5335 * @buffer: The ring buffer to reset a per cpu buffer of
5336 * @cpu: The CPU buffer to be reset
5337 */
5338void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
5339{
5340 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5341
5342 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5343 return;
5344
5345 /* prevent another thread from changing buffer sizes */
5346 mutex_lock(&buffer->mutex);
5347
5348 atomic_inc(&cpu_buffer->resize_disabled);
5349 atomic_inc(&cpu_buffer->record_disabled);
5350
5351 /* Make sure all commits have finished */
5352 synchronize_rcu();
5353
5354 reset_disabled_cpu_buffer(cpu_buffer);
5355
5356 atomic_dec(&cpu_buffer->record_disabled);
5357 atomic_dec(&cpu_buffer->resize_disabled);
5358
5359 mutex_unlock(&buffer->mutex);
5360}
5361EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
5362
5363/**
5364 * ring_buffer_reset_online_cpus - reset a ring buffer per CPU buffer
5365 * @buffer: The ring buffer to reset a per cpu buffer of
5366 * @cpu: The CPU buffer to be reset
5367 */
5368void ring_buffer_reset_online_cpus(struct trace_buffer *buffer)
5369{
5370 struct ring_buffer_per_cpu *cpu_buffer;
5371 int cpu;
5372
5373 /* prevent another thread from changing buffer sizes */
5374 mutex_lock(&buffer->mutex);
5375
5376 for_each_online_buffer_cpu(buffer, cpu) {
5377 cpu_buffer = buffer->buffers[cpu];
5378
5379 atomic_inc(&cpu_buffer->resize_disabled);
5380 atomic_inc(&cpu_buffer->record_disabled);
5381 }
5382
5383 /* Make sure all commits have finished */
5384 synchronize_rcu();
5385
5386 for_each_online_buffer_cpu(buffer, cpu) {
5387 cpu_buffer = buffer->buffers[cpu];
5388
5389 reset_disabled_cpu_buffer(cpu_buffer);
5390
5391 atomic_dec(&cpu_buffer->record_disabled);
5392 atomic_dec(&cpu_buffer->resize_disabled);
5393 }
5394
5395 mutex_unlock(&buffer->mutex);
5396}
5397
5398/**
5399 * ring_buffer_reset - reset a ring buffer
5400 * @buffer: The ring buffer to reset all cpu buffers
5401 */
5402void ring_buffer_reset(struct trace_buffer *buffer)
5403{
5404 struct ring_buffer_per_cpu *cpu_buffer;
5405 int cpu;
5406
5407 /* prevent another thread from changing buffer sizes */
5408 mutex_lock(&buffer->mutex);
5409
5410 for_each_buffer_cpu(buffer, cpu) {
5411 cpu_buffer = buffer->buffers[cpu];
5412
5413 atomic_inc(&cpu_buffer->resize_disabled);
5414 atomic_inc(&cpu_buffer->record_disabled);
5415 }
5416
5417 /* Make sure all commits have finished */
5418 synchronize_rcu();
5419
5420 for_each_buffer_cpu(buffer, cpu) {
5421 cpu_buffer = buffer->buffers[cpu];
5422
5423 reset_disabled_cpu_buffer(cpu_buffer);
5424
5425 atomic_dec(&cpu_buffer->record_disabled);
5426 atomic_dec(&cpu_buffer->resize_disabled);
5427 }
5428
5429 mutex_unlock(&buffer->mutex);
5430}
5431EXPORT_SYMBOL_GPL(ring_buffer_reset);
5432
5433/**
5434 * ring_buffer_empty - is the ring buffer empty?
5435 * @buffer: The ring buffer to test
5436 */
5437bool ring_buffer_empty(struct trace_buffer *buffer)
5438{
5439 struct ring_buffer_per_cpu *cpu_buffer;
5440 unsigned long flags;
5441 bool dolock;
5442 int cpu;
5443 int ret;
5444
5445 /* yes this is racy, but if you don't like the race, lock the buffer */
5446 for_each_buffer_cpu(buffer, cpu) {
5447 cpu_buffer = buffer->buffers[cpu];
5448 local_irq_save(flags);
5449 dolock = rb_reader_lock(cpu_buffer);
5450 ret = rb_per_cpu_empty(cpu_buffer);
5451 rb_reader_unlock(cpu_buffer, dolock);
5452 local_irq_restore(flags);
5453
5454 if (!ret)
5455 return false;
5456 }
5457
5458 return true;
5459}
5460EXPORT_SYMBOL_GPL(ring_buffer_empty);
5461
5462/**
5463 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
5464 * @buffer: The ring buffer
5465 * @cpu: The CPU buffer to test
5466 */
5467bool ring_buffer_empty_cpu(struct trace_buffer *buffer, int cpu)
5468{
5469 struct ring_buffer_per_cpu *cpu_buffer;
5470 unsigned long flags;
5471 bool dolock;
5472 int ret;
5473
5474 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5475 return true;
5476
5477 cpu_buffer = buffer->buffers[cpu];
5478 local_irq_save(flags);
5479 dolock = rb_reader_lock(cpu_buffer);
5480 ret = rb_per_cpu_empty(cpu_buffer);
5481 rb_reader_unlock(cpu_buffer, dolock);
5482 local_irq_restore(flags);
5483
5484 return ret;
5485}
5486EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
5487
5488#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
5489/**
5490 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
5491 * @buffer_a: One buffer to swap with
5492 * @buffer_b: The other buffer to swap with
5493 * @cpu: the CPU of the buffers to swap
5494 *
5495 * This function is useful for tracers that want to take a "snapshot"
5496 * of a CPU buffer and has another back up buffer lying around.
5497 * it is expected that the tracer handles the cpu buffer not being
5498 * used at the moment.
5499 */
5500int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
5501 struct trace_buffer *buffer_b, int cpu)
5502{
5503 struct ring_buffer_per_cpu *cpu_buffer_a;
5504 struct ring_buffer_per_cpu *cpu_buffer_b;
5505 int ret = -EINVAL;
5506
5507 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
5508 !cpumask_test_cpu(cpu, buffer_b->cpumask))
5509 goto out;
5510
5511 cpu_buffer_a = buffer_a->buffers[cpu];
5512 cpu_buffer_b = buffer_b->buffers[cpu];
5513
5514 /* At least make sure the two buffers are somewhat the same */
5515 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
5516 goto out;
5517
5518 ret = -EAGAIN;
5519
5520 if (atomic_read(&buffer_a->record_disabled))
5521 goto out;
5522
5523 if (atomic_read(&buffer_b->record_disabled))
5524 goto out;
5525
5526 if (atomic_read(&cpu_buffer_a->record_disabled))
5527 goto out;
5528
5529 if (atomic_read(&cpu_buffer_b->record_disabled))
5530 goto out;
5531
5532 /*
5533 * We can't do a synchronize_rcu here because this
5534 * function can be called in atomic context.
5535 * Normally this will be called from the same CPU as cpu.
5536 * If not it's up to the caller to protect this.
5537 */
5538 atomic_inc(&cpu_buffer_a->record_disabled);
5539 atomic_inc(&cpu_buffer_b->record_disabled);
5540
5541 ret = -EBUSY;
5542 if (local_read(&cpu_buffer_a->committing))
5543 goto out_dec;
5544 if (local_read(&cpu_buffer_b->committing))
5545 goto out_dec;
5546
5547 buffer_a->buffers[cpu] = cpu_buffer_b;
5548 buffer_b->buffers[cpu] = cpu_buffer_a;
5549
5550 cpu_buffer_b->buffer = buffer_a;
5551 cpu_buffer_a->buffer = buffer_b;
5552
5553 ret = 0;
5554
5555out_dec:
5556 atomic_dec(&cpu_buffer_a->record_disabled);
5557 atomic_dec(&cpu_buffer_b->record_disabled);
5558out:
5559 return ret;
5560}
5561EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
5562#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
5563
5564/**
5565 * ring_buffer_alloc_read_page - allocate a page to read from buffer
5566 * @buffer: the buffer to allocate for.
5567 * @cpu: the cpu buffer to allocate.
5568 *
5569 * This function is used in conjunction with ring_buffer_read_page.
5570 * When reading a full page from the ring buffer, these functions
5571 * can be used to speed up the process. The calling function should
5572 * allocate a few pages first with this function. Then when it
5573 * needs to get pages from the ring buffer, it passes the result
5574 * of this function into ring_buffer_read_page, which will swap
5575 * the page that was allocated, with the read page of the buffer.
5576 *
5577 * Returns:
5578 * The page allocated, or ERR_PTR
5579 */
5580void *ring_buffer_alloc_read_page(struct trace_buffer *buffer, int cpu)
5581{
5582 struct ring_buffer_per_cpu *cpu_buffer;
5583 struct buffer_data_page *bpage = NULL;
5584 unsigned long flags;
5585 struct page *page;
5586
5587 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5588 return ERR_PTR(-ENODEV);
5589
5590 cpu_buffer = buffer->buffers[cpu];
5591 local_irq_save(flags);
5592 arch_spin_lock(&cpu_buffer->lock);
5593
5594 if (cpu_buffer->free_page) {
5595 bpage = cpu_buffer->free_page;
5596 cpu_buffer->free_page = NULL;
5597 }
5598
5599 arch_spin_unlock(&cpu_buffer->lock);
5600 local_irq_restore(flags);
5601
5602 if (bpage)
5603 goto out;
5604
5605 page = alloc_pages_node(cpu_to_node(cpu),
5606 GFP_KERNEL | __GFP_NORETRY, 0);
5607 if (!page)
5608 return ERR_PTR(-ENOMEM);
5609
5610 bpage = page_address(page);
5611
5612 out:
5613 rb_init_page(bpage);
5614
5615 return bpage;
5616}
5617EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
5618
5619/**
5620 * ring_buffer_free_read_page - free an allocated read page
5621 * @buffer: the buffer the page was allocate for
5622 * @cpu: the cpu buffer the page came from
5623 * @data: the page to free
5624 *
5625 * Free a page allocated from ring_buffer_alloc_read_page.
5626 */
5627void ring_buffer_free_read_page(struct trace_buffer *buffer, int cpu, void *data)
5628{
5629 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5630 struct buffer_data_page *bpage = data;
5631 struct page *page = virt_to_page(bpage);
5632 unsigned long flags;
5633
5634 /* If the page is still in use someplace else, we can't reuse it */
5635 if (page_ref_count(page) > 1)
5636 goto out;
5637
5638 local_irq_save(flags);
5639 arch_spin_lock(&cpu_buffer->lock);
5640
5641 if (!cpu_buffer->free_page) {
5642 cpu_buffer->free_page = bpage;
5643 bpage = NULL;
5644 }
5645
5646 arch_spin_unlock(&cpu_buffer->lock);
5647 local_irq_restore(flags);
5648
5649 out:
5650 free_page((unsigned long)bpage);
5651}
5652EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
5653
5654/**
5655 * ring_buffer_read_page - extract a page from the ring buffer
5656 * @buffer: buffer to extract from
5657 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
5658 * @len: amount to extract
5659 * @cpu: the cpu of the buffer to extract
5660 * @full: should the extraction only happen when the page is full.
5661 *
5662 * This function will pull out a page from the ring buffer and consume it.
5663 * @data_page must be the address of the variable that was returned
5664 * from ring_buffer_alloc_read_page. This is because the page might be used
5665 * to swap with a page in the ring buffer.
5666 *
5667 * for example:
5668 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
5669 * if (IS_ERR(rpage))
5670 * return PTR_ERR(rpage);
5671 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
5672 * if (ret >= 0)
5673 * process_page(rpage, ret);
5674 *
5675 * When @full is set, the function will not return true unless
5676 * the writer is off the reader page.
5677 *
5678 * Note: it is up to the calling functions to handle sleeps and wakeups.
5679 * The ring buffer can be used anywhere in the kernel and can not
5680 * blindly call wake_up. The layer that uses the ring buffer must be
5681 * responsible for that.
5682 *
5683 * Returns:
5684 * >=0 if data has been transferred, returns the offset of consumed data.
5685 * <0 if no data has been transferred.
5686 */
5687int ring_buffer_read_page(struct trace_buffer *buffer,
5688 void **data_page, size_t len, int cpu, int full)
5689{
5690 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5691 struct ring_buffer_event *event;
5692 struct buffer_data_page *bpage;
5693 struct buffer_page *reader;
5694 unsigned long missed_events;
5695 unsigned long flags;
5696 unsigned int commit;
5697 unsigned int read;
5698 u64 save_timestamp;
5699 int ret = -1;
5700
5701 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5702 goto out;
5703
5704 /*
5705 * If len is not big enough to hold the page header, then
5706 * we can not copy anything.
5707 */
5708 if (len <= BUF_PAGE_HDR_SIZE)
5709 goto out;
5710
5711 len -= BUF_PAGE_HDR_SIZE;
5712
5713 if (!data_page)
5714 goto out;
5715
5716 bpage = *data_page;
5717 if (!bpage)
5718 goto out;
5719
5720 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5721
5722 reader = rb_get_reader_page(cpu_buffer);
5723 if (!reader)
5724 goto out_unlock;
5725
5726 event = rb_reader_event(cpu_buffer);
5727
5728 read = reader->read;
5729 commit = rb_page_commit(reader);
5730
5731 /* Check if any events were dropped */
5732 missed_events = cpu_buffer->lost_events;
5733
5734 /*
5735 * If this page has been partially read or
5736 * if len is not big enough to read the rest of the page or
5737 * a writer is still on the page, then
5738 * we must copy the data from the page to the buffer.
5739 * Otherwise, we can simply swap the page with the one passed in.
5740 */
5741 if (read || (len < (commit - read)) ||
5742 cpu_buffer->reader_page == cpu_buffer->commit_page) {
5743 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
5744 unsigned int rpos = read;
5745 unsigned int pos = 0;
5746 unsigned int size;
5747
5748 /*
5749 * If a full page is expected, this can still be returned
5750 * if there's been a previous partial read and the
5751 * rest of the page can be read and the commit page is off
5752 * the reader page.
5753 */
5754 if (full &&
5755 (!read || (len < (commit - read)) ||
5756 cpu_buffer->reader_page == cpu_buffer->commit_page))
5757 goto out_unlock;
5758
5759 if (len > (commit - read))
5760 len = (commit - read);
5761
5762 /* Always keep the time extend and data together */
5763 size = rb_event_ts_length(event);
5764
5765 if (len < size)
5766 goto out_unlock;
5767
5768 /* save the current timestamp, since the user will need it */
5769 save_timestamp = cpu_buffer->read_stamp;
5770
5771 /* Need to copy one event at a time */
5772 do {
5773 /* We need the size of one event, because
5774 * rb_advance_reader only advances by one event,
5775 * whereas rb_event_ts_length may include the size of
5776 * one or two events.
5777 * We have already ensured there's enough space if this
5778 * is a time extend. */
5779 size = rb_event_length(event);
5780 memcpy(bpage->data + pos, rpage->data + rpos, size);
5781
5782 len -= size;
5783
5784 rb_advance_reader(cpu_buffer);
5785 rpos = reader->read;
5786 pos += size;
5787
5788 if (rpos >= commit)
5789 break;
5790
5791 event = rb_reader_event(cpu_buffer);
5792 /* Always keep the time extend and data together */
5793 size = rb_event_ts_length(event);
5794 } while (len >= size);
5795
5796 /* update bpage */
5797 local_set(&bpage->commit, pos);
5798 bpage->time_stamp = save_timestamp;
5799
5800 /* we copied everything to the beginning */
5801 read = 0;
5802 } else {
5803 /* update the entry counter */
5804 cpu_buffer->read += rb_page_entries(reader);
5805 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
5806
5807 /* swap the pages */
5808 rb_init_page(bpage);
5809 bpage = reader->page;
5810 reader->page = *data_page;
5811 local_set(&reader->write, 0);
5812 local_set(&reader->entries, 0);
5813 reader->read = 0;
5814 *data_page = bpage;
5815
5816 /*
5817 * Use the real_end for the data size,
5818 * This gives us a chance to store the lost events
5819 * on the page.
5820 */
5821 if (reader->real_end)
5822 local_set(&bpage->commit, reader->real_end);
5823 }
5824 ret = read;
5825
5826 cpu_buffer->lost_events = 0;
5827
5828 commit = local_read(&bpage->commit);
5829 /*
5830 * Set a flag in the commit field if we lost events
5831 */
5832 if (missed_events) {
5833 /* If there is room at the end of the page to save the
5834 * missed events, then record it there.
5835 */
5836 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
5837 memcpy(&bpage->data[commit], &missed_events,
5838 sizeof(missed_events));
5839 local_add(RB_MISSED_STORED, &bpage->commit);
5840 commit += sizeof(missed_events);
5841 }
5842 local_add(RB_MISSED_EVENTS, &bpage->commit);
5843 }
5844
5845 /*
5846 * This page may be off to user land. Zero it out here.
5847 */
5848 if (commit < BUF_PAGE_SIZE)
5849 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
5850
5851 out_unlock:
5852 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5853
5854 out:
5855 return ret;
5856}
5857EXPORT_SYMBOL_GPL(ring_buffer_read_page);
5858
5859/*
5860 * We only allocate new buffers, never free them if the CPU goes down.
5861 * If we were to free the buffer, then the user would lose any trace that was in
5862 * the buffer.
5863 */
5864int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
5865{
5866 struct trace_buffer *buffer;
5867 long nr_pages_same;
5868 int cpu_i;
5869 unsigned long nr_pages;
5870
5871 buffer = container_of(node, struct trace_buffer, node);
5872 if (cpumask_test_cpu(cpu, buffer->cpumask))
5873 return 0;
5874
5875 nr_pages = 0;
5876 nr_pages_same = 1;
5877 /* check if all cpu sizes are same */
5878 for_each_buffer_cpu(buffer, cpu_i) {
5879 /* fill in the size from first enabled cpu */
5880 if (nr_pages == 0)
5881 nr_pages = buffer->buffers[cpu_i]->nr_pages;
5882 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
5883 nr_pages_same = 0;
5884 break;
5885 }
5886 }
5887 /* allocate minimum pages, user can later expand it */
5888 if (!nr_pages_same)
5889 nr_pages = 2;
5890 buffer->buffers[cpu] =
5891 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
5892 if (!buffer->buffers[cpu]) {
5893 WARN(1, "failed to allocate ring buffer on CPU %u\n",
5894 cpu);
5895 return -ENOMEM;
5896 }
5897 smp_wmb();
5898 cpumask_set_cpu(cpu, buffer->cpumask);
5899 return 0;
5900}
5901
5902#ifdef CONFIG_RING_BUFFER_STARTUP_TEST
5903/*
5904 * This is a basic integrity check of the ring buffer.
5905 * Late in the boot cycle this test will run when configured in.
5906 * It will kick off a thread per CPU that will go into a loop
5907 * writing to the per cpu ring buffer various sizes of data.
5908 * Some of the data will be large items, some small.
5909 *
5910 * Another thread is created that goes into a spin, sending out
5911 * IPIs to the other CPUs to also write into the ring buffer.
5912 * this is to test the nesting ability of the buffer.
5913 *
5914 * Basic stats are recorded and reported. If something in the
5915 * ring buffer should happen that's not expected, a big warning
5916 * is displayed and all ring buffers are disabled.
5917 */
5918static struct task_struct *rb_threads[NR_CPUS] __initdata;
5919
5920struct rb_test_data {
5921 struct trace_buffer *buffer;
5922 unsigned long events;
5923 unsigned long bytes_written;
5924 unsigned long bytes_alloc;
5925 unsigned long bytes_dropped;
5926 unsigned long events_nested;
5927 unsigned long bytes_written_nested;
5928 unsigned long bytes_alloc_nested;
5929 unsigned long bytes_dropped_nested;
5930 int min_size_nested;
5931 int max_size_nested;
5932 int max_size;
5933 int min_size;
5934 int cpu;
5935 int cnt;
5936};
5937
5938static struct rb_test_data rb_data[NR_CPUS] __initdata;
5939
5940/* 1 meg per cpu */
5941#define RB_TEST_BUFFER_SIZE 1048576
5942
5943static char rb_string[] __initdata =
5944 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
5945 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
5946 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
5947
5948static bool rb_test_started __initdata;
5949
5950struct rb_item {
5951 int size;
5952 char str[];
5953};
5954
5955static __init int rb_write_something(struct rb_test_data *data, bool nested)
5956{
5957 struct ring_buffer_event *event;
5958 struct rb_item *item;
5959 bool started;
5960 int event_len;
5961 int size;
5962 int len;
5963 int cnt;
5964
5965 /* Have nested writes different that what is written */
5966 cnt = data->cnt + (nested ? 27 : 0);
5967
5968 /* Multiply cnt by ~e, to make some unique increment */
5969 size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
5970
5971 len = size + sizeof(struct rb_item);
5972
5973 started = rb_test_started;
5974 /* read rb_test_started before checking buffer enabled */
5975 smp_rmb();
5976
5977 event = ring_buffer_lock_reserve(data->buffer, len);
5978 if (!event) {
5979 /* Ignore dropped events before test starts. */
5980 if (started) {
5981 if (nested)
5982 data->bytes_dropped += len;
5983 else
5984 data->bytes_dropped_nested += len;
5985 }
5986 return len;
5987 }
5988
5989 event_len = ring_buffer_event_length(event);
5990
5991 if (RB_WARN_ON(data->buffer, event_len < len))
5992 goto out;
5993
5994 item = ring_buffer_event_data(event);
5995 item->size = size;
5996 memcpy(item->str, rb_string, size);
5997
5998 if (nested) {
5999 data->bytes_alloc_nested += event_len;
6000 data->bytes_written_nested += len;
6001 data->events_nested++;
6002 if (!data->min_size_nested || len < data->min_size_nested)
6003 data->min_size_nested = len;
6004 if (len > data->max_size_nested)
6005 data->max_size_nested = len;
6006 } else {
6007 data->bytes_alloc += event_len;
6008 data->bytes_written += len;
6009 data->events++;
6010 if (!data->min_size || len < data->min_size)
6011 data->max_size = len;
6012 if (len > data->max_size)
6013 data->max_size = len;
6014 }
6015
6016 out:
6017 ring_buffer_unlock_commit(data->buffer);
6018
6019 return 0;
6020}
6021
6022static __init int rb_test(void *arg)
6023{
6024 struct rb_test_data *data = arg;
6025
6026 while (!kthread_should_stop()) {
6027 rb_write_something(data, false);
6028 data->cnt++;
6029
6030 set_current_state(TASK_INTERRUPTIBLE);
6031 /* Now sleep between a min of 100-300us and a max of 1ms */
6032 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
6033 }
6034
6035 return 0;
6036}
6037
6038static __init void rb_ipi(void *ignore)
6039{
6040 struct rb_test_data *data;
6041 int cpu = smp_processor_id();
6042
6043 data = &rb_data[cpu];
6044 rb_write_something(data, true);
6045}
6046
6047static __init int rb_hammer_test(void *arg)
6048{
6049 while (!kthread_should_stop()) {
6050
6051 /* Send an IPI to all cpus to write data! */
6052 smp_call_function(rb_ipi, NULL, 1);
6053 /* No sleep, but for non preempt, let others run */
6054 schedule();
6055 }
6056
6057 return 0;
6058}
6059
6060static __init int test_ringbuffer(void)
6061{
6062 struct task_struct *rb_hammer;
6063 struct trace_buffer *buffer;
6064 int cpu;
6065 int ret = 0;
6066
6067 if (security_locked_down(LOCKDOWN_TRACEFS)) {
6068 pr_warn("Lockdown is enabled, skipping ring buffer tests\n");
6069 return 0;
6070 }
6071
6072 pr_info("Running ring buffer tests...\n");
6073
6074 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
6075 if (WARN_ON(!buffer))
6076 return 0;
6077
6078 /* Disable buffer so that threads can't write to it yet */
6079 ring_buffer_record_off(buffer);
6080
6081 for_each_online_cpu(cpu) {
6082 rb_data[cpu].buffer = buffer;
6083 rb_data[cpu].cpu = cpu;
6084 rb_data[cpu].cnt = cpu;
6085 rb_threads[cpu] = kthread_run_on_cpu(rb_test, &rb_data[cpu],
6086 cpu, "rbtester/%u");
6087 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
6088 pr_cont("FAILED\n");
6089 ret = PTR_ERR(rb_threads[cpu]);
6090 goto out_free;
6091 }
6092 }
6093
6094 /* Now create the rb hammer! */
6095 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
6096 if (WARN_ON(IS_ERR(rb_hammer))) {
6097 pr_cont("FAILED\n");
6098 ret = PTR_ERR(rb_hammer);
6099 goto out_free;
6100 }
6101
6102 ring_buffer_record_on(buffer);
6103 /*
6104 * Show buffer is enabled before setting rb_test_started.
6105 * Yes there's a small race window where events could be
6106 * dropped and the thread wont catch it. But when a ring
6107 * buffer gets enabled, there will always be some kind of
6108 * delay before other CPUs see it. Thus, we don't care about
6109 * those dropped events. We care about events dropped after
6110 * the threads see that the buffer is active.
6111 */
6112 smp_wmb();
6113 rb_test_started = true;
6114
6115 set_current_state(TASK_INTERRUPTIBLE);
6116 /* Just run for 10 seconds */;
6117 schedule_timeout(10 * HZ);
6118
6119 kthread_stop(rb_hammer);
6120
6121 out_free:
6122 for_each_online_cpu(cpu) {
6123 if (!rb_threads[cpu])
6124 break;
6125 kthread_stop(rb_threads[cpu]);
6126 }
6127 if (ret) {
6128 ring_buffer_free(buffer);
6129 return ret;
6130 }
6131
6132 /* Report! */
6133 pr_info("finished\n");
6134 for_each_online_cpu(cpu) {
6135 struct ring_buffer_event *event;
6136 struct rb_test_data *data = &rb_data[cpu];
6137 struct rb_item *item;
6138 unsigned long total_events;
6139 unsigned long total_dropped;
6140 unsigned long total_written;
6141 unsigned long total_alloc;
6142 unsigned long total_read = 0;
6143 unsigned long total_size = 0;
6144 unsigned long total_len = 0;
6145 unsigned long total_lost = 0;
6146 unsigned long lost;
6147 int big_event_size;
6148 int small_event_size;
6149
6150 ret = -1;
6151
6152 total_events = data->events + data->events_nested;
6153 total_written = data->bytes_written + data->bytes_written_nested;
6154 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
6155 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
6156
6157 big_event_size = data->max_size + data->max_size_nested;
6158 small_event_size = data->min_size + data->min_size_nested;
6159
6160 pr_info("CPU %d:\n", cpu);
6161 pr_info(" events: %ld\n", total_events);
6162 pr_info(" dropped bytes: %ld\n", total_dropped);
6163 pr_info(" alloced bytes: %ld\n", total_alloc);
6164 pr_info(" written bytes: %ld\n", total_written);
6165 pr_info(" biggest event: %d\n", big_event_size);
6166 pr_info(" smallest event: %d\n", small_event_size);
6167
6168 if (RB_WARN_ON(buffer, total_dropped))
6169 break;
6170
6171 ret = 0;
6172
6173 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
6174 total_lost += lost;
6175 item = ring_buffer_event_data(event);
6176 total_len += ring_buffer_event_length(event);
6177 total_size += item->size + sizeof(struct rb_item);
6178 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
6179 pr_info("FAILED!\n");
6180 pr_info("buffer had: %.*s\n", item->size, item->str);
6181 pr_info("expected: %.*s\n", item->size, rb_string);
6182 RB_WARN_ON(buffer, 1);
6183 ret = -1;
6184 break;
6185 }
6186 total_read++;
6187 }
6188 if (ret)
6189 break;
6190
6191 ret = -1;
6192
6193 pr_info(" read events: %ld\n", total_read);
6194 pr_info(" lost events: %ld\n", total_lost);
6195 pr_info(" total events: %ld\n", total_lost + total_read);
6196 pr_info(" recorded len bytes: %ld\n", total_len);
6197 pr_info(" recorded size bytes: %ld\n", total_size);
6198 if (total_lost) {
6199 pr_info(" With dropped events, record len and size may not match\n"
6200 " alloced and written from above\n");
6201 } else {
6202 if (RB_WARN_ON(buffer, total_len != total_alloc ||
6203 total_size != total_written))
6204 break;
6205 }
6206 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
6207 break;
6208
6209 ret = 0;
6210 }
6211 if (!ret)
6212 pr_info("Ring buffer PASSED!\n");
6213
6214 ring_buffer_free(buffer);
6215 return 0;
6216}
6217
6218late_initcall(test_ringbuffer);
6219#endif /* CONFIG_RING_BUFFER_STARTUP_TEST */