Loading...
1/*
2 * Generic ring buffer
3 *
4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
5 */
6#include <linux/ring_buffer.h>
7#include <linux/trace_clock.h>
8#include <linux/spinlock.h>
9#include <linux/debugfs.h>
10#include <linux/uaccess.h>
11#include <linux/hardirq.h>
12#include <linux/kmemcheck.h>
13#include <linux/module.h>
14#include <linux/percpu.h>
15#include <linux/mutex.h>
16#include <linux/slab.h>
17#include <linux/init.h>
18#include <linux/hash.h>
19#include <linux/list.h>
20#include <linux/cpu.h>
21#include <linux/fs.h>
22
23#include <asm/local.h>
24#include "trace.h"
25
26/*
27 * The ring buffer header is special. We must manually up keep it.
28 */
29int ring_buffer_print_entry_header(struct trace_seq *s)
30{
31 int ret;
32
33 ret = trace_seq_printf(s, "# compressed entry header\n");
34 ret = trace_seq_printf(s, "\ttype_len : 5 bits\n");
35 ret = trace_seq_printf(s, "\ttime_delta : 27 bits\n");
36 ret = trace_seq_printf(s, "\tarray : 32 bits\n");
37 ret = trace_seq_printf(s, "\n");
38 ret = trace_seq_printf(s, "\tpadding : type == %d\n",
39 RINGBUF_TYPE_PADDING);
40 ret = trace_seq_printf(s, "\ttime_extend : type == %d\n",
41 RINGBUF_TYPE_TIME_EXTEND);
42 ret = trace_seq_printf(s, "\tdata max type_len == %d\n",
43 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
44
45 return ret;
46}
47
48/*
49 * The ring buffer is made up of a list of pages. A separate list of pages is
50 * allocated for each CPU. A writer may only write to a buffer that is
51 * associated with the CPU it is currently executing on. A reader may read
52 * from any per cpu buffer.
53 *
54 * The reader is special. For each per cpu buffer, the reader has its own
55 * reader page. When a reader has read the entire reader page, this reader
56 * page is swapped with another page in the ring buffer.
57 *
58 * Now, as long as the writer is off the reader page, the reader can do what
59 * ever it wants with that page. The writer will never write to that page
60 * again (as long as it is out of the ring buffer).
61 *
62 * Here's some silly ASCII art.
63 *
64 * +------+
65 * |reader| RING BUFFER
66 * |page |
67 * +------+ +---+ +---+ +---+
68 * | |-->| |-->| |
69 * +---+ +---+ +---+
70 * ^ |
71 * | |
72 * +---------------+
73 *
74 *
75 * +------+
76 * |reader| RING BUFFER
77 * |page |------------------v
78 * +------+ +---+ +---+ +---+
79 * | |-->| |-->| |
80 * +---+ +---+ +---+
81 * ^ |
82 * | |
83 * +---------------+
84 *
85 *
86 * +------+
87 * |reader| RING BUFFER
88 * |page |------------------v
89 * +------+ +---+ +---+ +---+
90 * ^ | |-->| |-->| |
91 * | +---+ +---+ +---+
92 * | |
93 * | |
94 * +------------------------------+
95 *
96 *
97 * +------+
98 * |buffer| RING BUFFER
99 * |page |------------------v
100 * +------+ +---+ +---+ +---+
101 * ^ | | | |-->| |
102 * | New +---+ +---+ +---+
103 * | Reader------^ |
104 * | page |
105 * +------------------------------+
106 *
107 *
108 * After we make this swap, the reader can hand this page off to the splice
109 * code and be done with it. It can even allocate a new page if it needs to
110 * and swap that into the ring buffer.
111 *
112 * We will be using cmpxchg soon to make all this lockless.
113 *
114 */
115
116/*
117 * A fast way to enable or disable all ring buffers is to
118 * call tracing_on or tracing_off. Turning off the ring buffers
119 * prevents all ring buffers from being recorded to.
120 * Turning this switch on, makes it OK to write to the
121 * ring buffer, if the ring buffer is enabled itself.
122 *
123 * There's three layers that must be on in order to write
124 * to the ring buffer.
125 *
126 * 1) This global flag must be set.
127 * 2) The ring buffer must be enabled for recording.
128 * 3) The per cpu buffer must be enabled for recording.
129 *
130 * In case of an anomaly, this global flag has a bit set that
131 * will permantly disable all ring buffers.
132 */
133
134/*
135 * Global flag to disable all recording to ring buffers
136 * This has two bits: ON, DISABLED
137 *
138 * ON DISABLED
139 * ---- ----------
140 * 0 0 : ring buffers are off
141 * 1 0 : ring buffers are on
142 * X 1 : ring buffers are permanently disabled
143 */
144
145enum {
146 RB_BUFFERS_ON_BIT = 0,
147 RB_BUFFERS_DISABLED_BIT = 1,
148};
149
150enum {
151 RB_BUFFERS_ON = 1 << RB_BUFFERS_ON_BIT,
152 RB_BUFFERS_DISABLED = 1 << RB_BUFFERS_DISABLED_BIT,
153};
154
155static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON;
156
157#define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
158
159/**
160 * tracing_on - enable all tracing buffers
161 *
162 * This function enables all tracing buffers that may have been
163 * disabled with tracing_off.
164 */
165void tracing_on(void)
166{
167 set_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
168}
169EXPORT_SYMBOL_GPL(tracing_on);
170
171/**
172 * tracing_off - turn off all tracing buffers
173 *
174 * This function stops all tracing buffers from recording data.
175 * It does not disable any overhead the tracers themselves may
176 * be causing. This function simply causes all recording to
177 * the ring buffers to fail.
178 */
179void tracing_off(void)
180{
181 clear_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
182}
183EXPORT_SYMBOL_GPL(tracing_off);
184
185/**
186 * tracing_off_permanent - permanently disable ring buffers
187 *
188 * This function, once called, will disable all ring buffers
189 * permanently.
190 */
191void tracing_off_permanent(void)
192{
193 set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags);
194}
195
196/**
197 * tracing_is_on - show state of ring buffers enabled
198 */
199int tracing_is_on(void)
200{
201 return ring_buffer_flags == RB_BUFFERS_ON;
202}
203EXPORT_SYMBOL_GPL(tracing_is_on);
204
205#define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
206#define RB_ALIGNMENT 4U
207#define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
208#define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
209
210#if !defined(CONFIG_64BIT) || defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
211# define RB_FORCE_8BYTE_ALIGNMENT 0
212# define RB_ARCH_ALIGNMENT RB_ALIGNMENT
213#else
214# define RB_FORCE_8BYTE_ALIGNMENT 1
215# define RB_ARCH_ALIGNMENT 8U
216#endif
217
218/* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
219#define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
220
221enum {
222 RB_LEN_TIME_EXTEND = 8,
223 RB_LEN_TIME_STAMP = 16,
224};
225
226#define skip_time_extend(event) \
227 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
228
229static inline int rb_null_event(struct ring_buffer_event *event)
230{
231 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
232}
233
234static void rb_event_set_padding(struct ring_buffer_event *event)
235{
236 /* padding has a NULL time_delta */
237 event->type_len = RINGBUF_TYPE_PADDING;
238 event->time_delta = 0;
239}
240
241static unsigned
242rb_event_data_length(struct ring_buffer_event *event)
243{
244 unsigned length;
245
246 if (event->type_len)
247 length = event->type_len * RB_ALIGNMENT;
248 else
249 length = event->array[0];
250 return length + RB_EVNT_HDR_SIZE;
251}
252
253/*
254 * Return the length of the given event. Will return
255 * the length of the time extend if the event is a
256 * time extend.
257 */
258static inline unsigned
259rb_event_length(struct ring_buffer_event *event)
260{
261 switch (event->type_len) {
262 case RINGBUF_TYPE_PADDING:
263 if (rb_null_event(event))
264 /* undefined */
265 return -1;
266 return event->array[0] + RB_EVNT_HDR_SIZE;
267
268 case RINGBUF_TYPE_TIME_EXTEND:
269 return RB_LEN_TIME_EXTEND;
270
271 case RINGBUF_TYPE_TIME_STAMP:
272 return RB_LEN_TIME_STAMP;
273
274 case RINGBUF_TYPE_DATA:
275 return rb_event_data_length(event);
276 default:
277 BUG();
278 }
279 /* not hit */
280 return 0;
281}
282
283/*
284 * Return total length of time extend and data,
285 * or just the event length for all other events.
286 */
287static inline unsigned
288rb_event_ts_length(struct ring_buffer_event *event)
289{
290 unsigned len = 0;
291
292 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
293 /* time extends include the data event after it */
294 len = RB_LEN_TIME_EXTEND;
295 event = skip_time_extend(event);
296 }
297 return len + rb_event_length(event);
298}
299
300/**
301 * ring_buffer_event_length - return the length of the event
302 * @event: the event to get the length of
303 *
304 * Returns the size of the data load of a data event.
305 * If the event is something other than a data event, it
306 * returns the size of the event itself. With the exception
307 * of a TIME EXTEND, where it still returns the size of the
308 * data load of the data event after it.
309 */
310unsigned ring_buffer_event_length(struct ring_buffer_event *event)
311{
312 unsigned length;
313
314 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
315 event = skip_time_extend(event);
316
317 length = rb_event_length(event);
318 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
319 return length;
320 length -= RB_EVNT_HDR_SIZE;
321 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
322 length -= sizeof(event->array[0]);
323 return length;
324}
325EXPORT_SYMBOL_GPL(ring_buffer_event_length);
326
327/* inline for ring buffer fast paths */
328static void *
329rb_event_data(struct ring_buffer_event *event)
330{
331 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
332 event = skip_time_extend(event);
333 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
334 /* If length is in len field, then array[0] has the data */
335 if (event->type_len)
336 return (void *)&event->array[0];
337 /* Otherwise length is in array[0] and array[1] has the data */
338 return (void *)&event->array[1];
339}
340
341/**
342 * ring_buffer_event_data - return the data of the event
343 * @event: the event to get the data from
344 */
345void *ring_buffer_event_data(struct ring_buffer_event *event)
346{
347 return rb_event_data(event);
348}
349EXPORT_SYMBOL_GPL(ring_buffer_event_data);
350
351#define for_each_buffer_cpu(buffer, cpu) \
352 for_each_cpu(cpu, buffer->cpumask)
353
354#define TS_SHIFT 27
355#define TS_MASK ((1ULL << TS_SHIFT) - 1)
356#define TS_DELTA_TEST (~TS_MASK)
357
358/* Flag when events were overwritten */
359#define RB_MISSED_EVENTS (1 << 31)
360/* Missed count stored at end */
361#define RB_MISSED_STORED (1 << 30)
362
363struct buffer_data_page {
364 u64 time_stamp; /* page time stamp */
365 local_t commit; /* write committed index */
366 unsigned char data[]; /* data of buffer page */
367};
368
369/*
370 * Note, the buffer_page list must be first. The buffer pages
371 * are allocated in cache lines, which means that each buffer
372 * page will be at the beginning of a cache line, and thus
373 * the least significant bits will be zero. We use this to
374 * add flags in the list struct pointers, to make the ring buffer
375 * lockless.
376 */
377struct buffer_page {
378 struct list_head list; /* list of buffer pages */
379 local_t write; /* index for next write */
380 unsigned read; /* index for next read */
381 local_t entries; /* entries on this page */
382 unsigned long real_end; /* real end of data */
383 struct buffer_data_page *page; /* Actual data page */
384};
385
386/*
387 * The buffer page counters, write and entries, must be reset
388 * atomically when crossing page boundaries. To synchronize this
389 * update, two counters are inserted into the number. One is
390 * the actual counter for the write position or count on the page.
391 *
392 * The other is a counter of updaters. Before an update happens
393 * the update partition of the counter is incremented. This will
394 * allow the updater to update the counter atomically.
395 *
396 * The counter is 20 bits, and the state data is 12.
397 */
398#define RB_WRITE_MASK 0xfffff
399#define RB_WRITE_INTCNT (1 << 20)
400
401static void rb_init_page(struct buffer_data_page *bpage)
402{
403 local_set(&bpage->commit, 0);
404}
405
406/**
407 * ring_buffer_page_len - the size of data on the page.
408 * @page: The page to read
409 *
410 * Returns the amount of data on the page, including buffer page header.
411 */
412size_t ring_buffer_page_len(void *page)
413{
414 return local_read(&((struct buffer_data_page *)page)->commit)
415 + BUF_PAGE_HDR_SIZE;
416}
417
418/*
419 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
420 * this issue out.
421 */
422static void free_buffer_page(struct buffer_page *bpage)
423{
424 free_page((unsigned long)bpage->page);
425 kfree(bpage);
426}
427
428/*
429 * We need to fit the time_stamp delta into 27 bits.
430 */
431static inline int test_time_stamp(u64 delta)
432{
433 if (delta & TS_DELTA_TEST)
434 return 1;
435 return 0;
436}
437
438#define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
439
440/* Max payload is BUF_PAGE_SIZE - header (8bytes) */
441#define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
442
443int ring_buffer_print_page_header(struct trace_seq *s)
444{
445 struct buffer_data_page field;
446 int ret;
447
448 ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t"
449 "offset:0;\tsize:%u;\tsigned:%u;\n",
450 (unsigned int)sizeof(field.time_stamp),
451 (unsigned int)is_signed_type(u64));
452
453 ret = trace_seq_printf(s, "\tfield: local_t commit;\t"
454 "offset:%u;\tsize:%u;\tsigned:%u;\n",
455 (unsigned int)offsetof(typeof(field), commit),
456 (unsigned int)sizeof(field.commit),
457 (unsigned int)is_signed_type(long));
458
459 ret = trace_seq_printf(s, "\tfield: int overwrite;\t"
460 "offset:%u;\tsize:%u;\tsigned:%u;\n",
461 (unsigned int)offsetof(typeof(field), commit),
462 1,
463 (unsigned int)is_signed_type(long));
464
465 ret = trace_seq_printf(s, "\tfield: char data;\t"
466 "offset:%u;\tsize:%u;\tsigned:%u;\n",
467 (unsigned int)offsetof(typeof(field), data),
468 (unsigned int)BUF_PAGE_SIZE,
469 (unsigned int)is_signed_type(char));
470
471 return ret;
472}
473
474/*
475 * head_page == tail_page && head == tail then buffer is empty.
476 */
477struct ring_buffer_per_cpu {
478 int cpu;
479 atomic_t record_disabled;
480 struct ring_buffer *buffer;
481 spinlock_t reader_lock; /* serialize readers */
482 arch_spinlock_t lock;
483 struct lock_class_key lock_key;
484 struct list_head *pages;
485 struct buffer_page *head_page; /* read from head */
486 struct buffer_page *tail_page; /* write to tail */
487 struct buffer_page *commit_page; /* committed pages */
488 struct buffer_page *reader_page;
489 unsigned long lost_events;
490 unsigned long last_overrun;
491 local_t commit_overrun;
492 local_t overrun;
493 local_t entries;
494 local_t committing;
495 local_t commits;
496 unsigned long read;
497 u64 write_stamp;
498 u64 read_stamp;
499};
500
501struct ring_buffer {
502 unsigned pages;
503 unsigned flags;
504 int cpus;
505 atomic_t record_disabled;
506 cpumask_var_t cpumask;
507
508 struct lock_class_key *reader_lock_key;
509
510 struct mutex mutex;
511
512 struct ring_buffer_per_cpu **buffers;
513
514#ifdef CONFIG_HOTPLUG_CPU
515 struct notifier_block cpu_notify;
516#endif
517 u64 (*clock)(void);
518};
519
520struct ring_buffer_iter {
521 struct ring_buffer_per_cpu *cpu_buffer;
522 unsigned long head;
523 struct buffer_page *head_page;
524 struct buffer_page *cache_reader_page;
525 unsigned long cache_read;
526 u64 read_stamp;
527};
528
529/* buffer may be either ring_buffer or ring_buffer_per_cpu */
530#define RB_WARN_ON(b, cond) \
531 ({ \
532 int _____ret = unlikely(cond); \
533 if (_____ret) { \
534 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
535 struct ring_buffer_per_cpu *__b = \
536 (void *)b; \
537 atomic_inc(&__b->buffer->record_disabled); \
538 } else \
539 atomic_inc(&b->record_disabled); \
540 WARN_ON(1); \
541 } \
542 _____ret; \
543 })
544
545/* Up this if you want to test the TIME_EXTENTS and normalization */
546#define DEBUG_SHIFT 0
547
548static inline u64 rb_time_stamp(struct ring_buffer *buffer)
549{
550 /* shift to debug/test normalization and TIME_EXTENTS */
551 return buffer->clock() << DEBUG_SHIFT;
552}
553
554u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
555{
556 u64 time;
557
558 preempt_disable_notrace();
559 time = rb_time_stamp(buffer);
560 preempt_enable_no_resched_notrace();
561
562 return time;
563}
564EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
565
566void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
567 int cpu, u64 *ts)
568{
569 /* Just stupid testing the normalize function and deltas */
570 *ts >>= DEBUG_SHIFT;
571}
572EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
573
574/*
575 * Making the ring buffer lockless makes things tricky.
576 * Although writes only happen on the CPU that they are on,
577 * and they only need to worry about interrupts. Reads can
578 * happen on any CPU.
579 *
580 * The reader page is always off the ring buffer, but when the
581 * reader finishes with a page, it needs to swap its page with
582 * a new one from the buffer. The reader needs to take from
583 * the head (writes go to the tail). But if a writer is in overwrite
584 * mode and wraps, it must push the head page forward.
585 *
586 * Here lies the problem.
587 *
588 * The reader must be careful to replace only the head page, and
589 * not another one. As described at the top of the file in the
590 * ASCII art, the reader sets its old page to point to the next
591 * page after head. It then sets the page after head to point to
592 * the old reader page. But if the writer moves the head page
593 * during this operation, the reader could end up with the tail.
594 *
595 * We use cmpxchg to help prevent this race. We also do something
596 * special with the page before head. We set the LSB to 1.
597 *
598 * When the writer must push the page forward, it will clear the
599 * bit that points to the head page, move the head, and then set
600 * the bit that points to the new head page.
601 *
602 * We also don't want an interrupt coming in and moving the head
603 * page on another writer. Thus we use the second LSB to catch
604 * that too. Thus:
605 *
606 * head->list->prev->next bit 1 bit 0
607 * ------- -------
608 * Normal page 0 0
609 * Points to head page 0 1
610 * New head page 1 0
611 *
612 * Note we can not trust the prev pointer of the head page, because:
613 *
614 * +----+ +-----+ +-----+
615 * | |------>| T |---X--->| N |
616 * | |<------| | | |
617 * +----+ +-----+ +-----+
618 * ^ ^ |
619 * | +-----+ | |
620 * +----------| R |----------+ |
621 * | |<-----------+
622 * +-----+
623 *
624 * Key: ---X--> HEAD flag set in pointer
625 * T Tail page
626 * R Reader page
627 * N Next page
628 *
629 * (see __rb_reserve_next() to see where this happens)
630 *
631 * What the above shows is that the reader just swapped out
632 * the reader page with a page in the buffer, but before it
633 * could make the new header point back to the new page added
634 * it was preempted by a writer. The writer moved forward onto
635 * the new page added by the reader and is about to move forward
636 * again.
637 *
638 * You can see, it is legitimate for the previous pointer of
639 * the head (or any page) not to point back to itself. But only
640 * temporarially.
641 */
642
643#define RB_PAGE_NORMAL 0UL
644#define RB_PAGE_HEAD 1UL
645#define RB_PAGE_UPDATE 2UL
646
647
648#define RB_FLAG_MASK 3UL
649
650/* PAGE_MOVED is not part of the mask */
651#define RB_PAGE_MOVED 4UL
652
653/*
654 * rb_list_head - remove any bit
655 */
656static struct list_head *rb_list_head(struct list_head *list)
657{
658 unsigned long val = (unsigned long)list;
659
660 return (struct list_head *)(val & ~RB_FLAG_MASK);
661}
662
663/*
664 * rb_is_head_page - test if the given page is the head page
665 *
666 * Because the reader may move the head_page pointer, we can
667 * not trust what the head page is (it may be pointing to
668 * the reader page). But if the next page is a header page,
669 * its flags will be non zero.
670 */
671static inline int
672rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
673 struct buffer_page *page, struct list_head *list)
674{
675 unsigned long val;
676
677 val = (unsigned long)list->next;
678
679 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
680 return RB_PAGE_MOVED;
681
682 return val & RB_FLAG_MASK;
683}
684
685/*
686 * rb_is_reader_page
687 *
688 * The unique thing about the reader page, is that, if the
689 * writer is ever on it, the previous pointer never points
690 * back to the reader page.
691 */
692static int rb_is_reader_page(struct buffer_page *page)
693{
694 struct list_head *list = page->list.prev;
695
696 return rb_list_head(list->next) != &page->list;
697}
698
699/*
700 * rb_set_list_to_head - set a list_head to be pointing to head.
701 */
702static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
703 struct list_head *list)
704{
705 unsigned long *ptr;
706
707 ptr = (unsigned long *)&list->next;
708 *ptr |= RB_PAGE_HEAD;
709 *ptr &= ~RB_PAGE_UPDATE;
710}
711
712/*
713 * rb_head_page_activate - sets up head page
714 */
715static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
716{
717 struct buffer_page *head;
718
719 head = cpu_buffer->head_page;
720 if (!head)
721 return;
722
723 /*
724 * Set the previous list pointer to have the HEAD flag.
725 */
726 rb_set_list_to_head(cpu_buffer, head->list.prev);
727}
728
729static void rb_list_head_clear(struct list_head *list)
730{
731 unsigned long *ptr = (unsigned long *)&list->next;
732
733 *ptr &= ~RB_FLAG_MASK;
734}
735
736/*
737 * rb_head_page_dactivate - clears head page ptr (for free list)
738 */
739static void
740rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
741{
742 struct list_head *hd;
743
744 /* Go through the whole list and clear any pointers found. */
745 rb_list_head_clear(cpu_buffer->pages);
746
747 list_for_each(hd, cpu_buffer->pages)
748 rb_list_head_clear(hd);
749}
750
751static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
752 struct buffer_page *head,
753 struct buffer_page *prev,
754 int old_flag, int new_flag)
755{
756 struct list_head *list;
757 unsigned long val = (unsigned long)&head->list;
758 unsigned long ret;
759
760 list = &prev->list;
761
762 val &= ~RB_FLAG_MASK;
763
764 ret = cmpxchg((unsigned long *)&list->next,
765 val | old_flag, val | new_flag);
766
767 /* check if the reader took the page */
768 if ((ret & ~RB_FLAG_MASK) != val)
769 return RB_PAGE_MOVED;
770
771 return ret & RB_FLAG_MASK;
772}
773
774static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
775 struct buffer_page *head,
776 struct buffer_page *prev,
777 int old_flag)
778{
779 return rb_head_page_set(cpu_buffer, head, prev,
780 old_flag, RB_PAGE_UPDATE);
781}
782
783static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
784 struct buffer_page *head,
785 struct buffer_page *prev,
786 int old_flag)
787{
788 return rb_head_page_set(cpu_buffer, head, prev,
789 old_flag, RB_PAGE_HEAD);
790}
791
792static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
793 struct buffer_page *head,
794 struct buffer_page *prev,
795 int old_flag)
796{
797 return rb_head_page_set(cpu_buffer, head, prev,
798 old_flag, RB_PAGE_NORMAL);
799}
800
801static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
802 struct buffer_page **bpage)
803{
804 struct list_head *p = rb_list_head((*bpage)->list.next);
805
806 *bpage = list_entry(p, struct buffer_page, list);
807}
808
809static struct buffer_page *
810rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
811{
812 struct buffer_page *head;
813 struct buffer_page *page;
814 struct list_head *list;
815 int i;
816
817 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
818 return NULL;
819
820 /* sanity check */
821 list = cpu_buffer->pages;
822 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
823 return NULL;
824
825 page = head = cpu_buffer->head_page;
826 /*
827 * It is possible that the writer moves the header behind
828 * where we started, and we miss in one loop.
829 * A second loop should grab the header, but we'll do
830 * three loops just because I'm paranoid.
831 */
832 for (i = 0; i < 3; i++) {
833 do {
834 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
835 cpu_buffer->head_page = page;
836 return page;
837 }
838 rb_inc_page(cpu_buffer, &page);
839 } while (page != head);
840 }
841
842 RB_WARN_ON(cpu_buffer, 1);
843
844 return NULL;
845}
846
847static int rb_head_page_replace(struct buffer_page *old,
848 struct buffer_page *new)
849{
850 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
851 unsigned long val;
852 unsigned long ret;
853
854 val = *ptr & ~RB_FLAG_MASK;
855 val |= RB_PAGE_HEAD;
856
857 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
858
859 return ret == val;
860}
861
862/*
863 * rb_tail_page_update - move the tail page forward
864 *
865 * Returns 1 if moved tail page, 0 if someone else did.
866 */
867static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
868 struct buffer_page *tail_page,
869 struct buffer_page *next_page)
870{
871 struct buffer_page *old_tail;
872 unsigned long old_entries;
873 unsigned long old_write;
874 int ret = 0;
875
876 /*
877 * The tail page now needs to be moved forward.
878 *
879 * We need to reset the tail page, but without messing
880 * with possible erasing of data brought in by interrupts
881 * that have moved the tail page and are currently on it.
882 *
883 * We add a counter to the write field to denote this.
884 */
885 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
886 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
887
888 /*
889 * Just make sure we have seen our old_write and synchronize
890 * with any interrupts that come in.
891 */
892 barrier();
893
894 /*
895 * If the tail page is still the same as what we think
896 * it is, then it is up to us to update the tail
897 * pointer.
898 */
899 if (tail_page == cpu_buffer->tail_page) {
900 /* Zero the write counter */
901 unsigned long val = old_write & ~RB_WRITE_MASK;
902 unsigned long eval = old_entries & ~RB_WRITE_MASK;
903
904 /*
905 * This will only succeed if an interrupt did
906 * not come in and change it. In which case, we
907 * do not want to modify it.
908 *
909 * We add (void) to let the compiler know that we do not care
910 * about the return value of these functions. We use the
911 * cmpxchg to only update if an interrupt did not already
912 * do it for us. If the cmpxchg fails, we don't care.
913 */
914 (void)local_cmpxchg(&next_page->write, old_write, val);
915 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
916
917 /*
918 * No need to worry about races with clearing out the commit.
919 * it only can increment when a commit takes place. But that
920 * only happens in the outer most nested commit.
921 */
922 local_set(&next_page->page->commit, 0);
923
924 old_tail = cmpxchg(&cpu_buffer->tail_page,
925 tail_page, next_page);
926
927 if (old_tail == tail_page)
928 ret = 1;
929 }
930
931 return ret;
932}
933
934static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
935 struct buffer_page *bpage)
936{
937 unsigned long val = (unsigned long)bpage;
938
939 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
940 return 1;
941
942 return 0;
943}
944
945/**
946 * rb_check_list - make sure a pointer to a list has the last bits zero
947 */
948static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
949 struct list_head *list)
950{
951 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
952 return 1;
953 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
954 return 1;
955 return 0;
956}
957
958/**
959 * check_pages - integrity check of buffer pages
960 * @cpu_buffer: CPU buffer with pages to test
961 *
962 * As a safety measure we check to make sure the data pages have not
963 * been corrupted.
964 */
965static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
966{
967 struct list_head *head = cpu_buffer->pages;
968 struct buffer_page *bpage, *tmp;
969
970 rb_head_page_deactivate(cpu_buffer);
971
972 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
973 return -1;
974 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
975 return -1;
976
977 if (rb_check_list(cpu_buffer, head))
978 return -1;
979
980 list_for_each_entry_safe(bpage, tmp, head, list) {
981 if (RB_WARN_ON(cpu_buffer,
982 bpage->list.next->prev != &bpage->list))
983 return -1;
984 if (RB_WARN_ON(cpu_buffer,
985 bpage->list.prev->next != &bpage->list))
986 return -1;
987 if (rb_check_list(cpu_buffer, &bpage->list))
988 return -1;
989 }
990
991 rb_head_page_activate(cpu_buffer);
992
993 return 0;
994}
995
996static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
997 unsigned nr_pages)
998{
999 struct buffer_page *bpage, *tmp;
1000 LIST_HEAD(pages);
1001 unsigned i;
1002
1003 WARN_ON(!nr_pages);
1004
1005 for (i = 0; i < nr_pages; i++) {
1006 struct page *page;
1007 /*
1008 * __GFP_NORETRY flag makes sure that the allocation fails
1009 * gracefully without invoking oom-killer and the system is
1010 * not destabilized.
1011 */
1012 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1013 GFP_KERNEL | __GFP_NORETRY,
1014 cpu_to_node(cpu_buffer->cpu));
1015 if (!bpage)
1016 goto free_pages;
1017
1018 rb_check_bpage(cpu_buffer, bpage);
1019
1020 list_add(&bpage->list, &pages);
1021
1022 page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu),
1023 GFP_KERNEL | __GFP_NORETRY, 0);
1024 if (!page)
1025 goto free_pages;
1026 bpage->page = page_address(page);
1027 rb_init_page(bpage->page);
1028 }
1029
1030 /*
1031 * The ring buffer page list is a circular list that does not
1032 * start and end with a list head. All page list items point to
1033 * other pages.
1034 */
1035 cpu_buffer->pages = pages.next;
1036 list_del(&pages);
1037
1038 rb_check_pages(cpu_buffer);
1039
1040 return 0;
1041
1042 free_pages:
1043 list_for_each_entry_safe(bpage, tmp, &pages, list) {
1044 list_del_init(&bpage->list);
1045 free_buffer_page(bpage);
1046 }
1047 return -ENOMEM;
1048}
1049
1050static struct ring_buffer_per_cpu *
1051rb_allocate_cpu_buffer(struct ring_buffer *buffer, int cpu)
1052{
1053 struct ring_buffer_per_cpu *cpu_buffer;
1054 struct buffer_page *bpage;
1055 struct page *page;
1056 int ret;
1057
1058 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1059 GFP_KERNEL, cpu_to_node(cpu));
1060 if (!cpu_buffer)
1061 return NULL;
1062
1063 cpu_buffer->cpu = cpu;
1064 cpu_buffer->buffer = buffer;
1065 spin_lock_init(&cpu_buffer->reader_lock);
1066 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1067 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1068
1069 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1070 GFP_KERNEL, cpu_to_node(cpu));
1071 if (!bpage)
1072 goto fail_free_buffer;
1073
1074 rb_check_bpage(cpu_buffer, bpage);
1075
1076 cpu_buffer->reader_page = bpage;
1077 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1078 if (!page)
1079 goto fail_free_reader;
1080 bpage->page = page_address(page);
1081 rb_init_page(bpage->page);
1082
1083 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1084
1085 ret = rb_allocate_pages(cpu_buffer, buffer->pages);
1086 if (ret < 0)
1087 goto fail_free_reader;
1088
1089 cpu_buffer->head_page
1090 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1091 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1092
1093 rb_head_page_activate(cpu_buffer);
1094
1095 return cpu_buffer;
1096
1097 fail_free_reader:
1098 free_buffer_page(cpu_buffer->reader_page);
1099
1100 fail_free_buffer:
1101 kfree(cpu_buffer);
1102 return NULL;
1103}
1104
1105static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1106{
1107 struct list_head *head = cpu_buffer->pages;
1108 struct buffer_page *bpage, *tmp;
1109
1110 free_buffer_page(cpu_buffer->reader_page);
1111
1112 rb_head_page_deactivate(cpu_buffer);
1113
1114 if (head) {
1115 list_for_each_entry_safe(bpage, tmp, head, list) {
1116 list_del_init(&bpage->list);
1117 free_buffer_page(bpage);
1118 }
1119 bpage = list_entry(head, struct buffer_page, list);
1120 free_buffer_page(bpage);
1121 }
1122
1123 kfree(cpu_buffer);
1124}
1125
1126#ifdef CONFIG_HOTPLUG_CPU
1127static int rb_cpu_notify(struct notifier_block *self,
1128 unsigned long action, void *hcpu);
1129#endif
1130
1131/**
1132 * ring_buffer_alloc - allocate a new ring_buffer
1133 * @size: the size in bytes per cpu that is needed.
1134 * @flags: attributes to set for the ring buffer.
1135 *
1136 * Currently the only flag that is available is the RB_FL_OVERWRITE
1137 * flag. This flag means that the buffer will overwrite old data
1138 * when the buffer wraps. If this flag is not set, the buffer will
1139 * drop data when the tail hits the head.
1140 */
1141struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1142 struct lock_class_key *key)
1143{
1144 struct ring_buffer *buffer;
1145 int bsize;
1146 int cpu;
1147
1148 /* keep it in its own cache line */
1149 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1150 GFP_KERNEL);
1151 if (!buffer)
1152 return NULL;
1153
1154 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1155 goto fail_free_buffer;
1156
1157 buffer->pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1158 buffer->flags = flags;
1159 buffer->clock = trace_clock_local;
1160 buffer->reader_lock_key = key;
1161
1162 /* need at least two pages */
1163 if (buffer->pages < 2)
1164 buffer->pages = 2;
1165
1166 /*
1167 * In case of non-hotplug cpu, if the ring-buffer is allocated
1168 * in early initcall, it will not be notified of secondary cpus.
1169 * In that off case, we need to allocate for all possible cpus.
1170 */
1171#ifdef CONFIG_HOTPLUG_CPU
1172 get_online_cpus();
1173 cpumask_copy(buffer->cpumask, cpu_online_mask);
1174#else
1175 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1176#endif
1177 buffer->cpus = nr_cpu_ids;
1178
1179 bsize = sizeof(void *) * nr_cpu_ids;
1180 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1181 GFP_KERNEL);
1182 if (!buffer->buffers)
1183 goto fail_free_cpumask;
1184
1185 for_each_buffer_cpu(buffer, cpu) {
1186 buffer->buffers[cpu] =
1187 rb_allocate_cpu_buffer(buffer, cpu);
1188 if (!buffer->buffers[cpu])
1189 goto fail_free_buffers;
1190 }
1191
1192#ifdef CONFIG_HOTPLUG_CPU
1193 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1194 buffer->cpu_notify.priority = 0;
1195 register_cpu_notifier(&buffer->cpu_notify);
1196#endif
1197
1198 put_online_cpus();
1199 mutex_init(&buffer->mutex);
1200
1201 return buffer;
1202
1203 fail_free_buffers:
1204 for_each_buffer_cpu(buffer, cpu) {
1205 if (buffer->buffers[cpu])
1206 rb_free_cpu_buffer(buffer->buffers[cpu]);
1207 }
1208 kfree(buffer->buffers);
1209
1210 fail_free_cpumask:
1211 free_cpumask_var(buffer->cpumask);
1212 put_online_cpus();
1213
1214 fail_free_buffer:
1215 kfree(buffer);
1216 return NULL;
1217}
1218EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1219
1220/**
1221 * ring_buffer_free - free a ring buffer.
1222 * @buffer: the buffer to free.
1223 */
1224void
1225ring_buffer_free(struct ring_buffer *buffer)
1226{
1227 int cpu;
1228
1229 get_online_cpus();
1230
1231#ifdef CONFIG_HOTPLUG_CPU
1232 unregister_cpu_notifier(&buffer->cpu_notify);
1233#endif
1234
1235 for_each_buffer_cpu(buffer, cpu)
1236 rb_free_cpu_buffer(buffer->buffers[cpu]);
1237
1238 put_online_cpus();
1239
1240 kfree(buffer->buffers);
1241 free_cpumask_var(buffer->cpumask);
1242
1243 kfree(buffer);
1244}
1245EXPORT_SYMBOL_GPL(ring_buffer_free);
1246
1247void ring_buffer_set_clock(struct ring_buffer *buffer,
1248 u64 (*clock)(void))
1249{
1250 buffer->clock = clock;
1251}
1252
1253static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1254
1255static void
1256rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned nr_pages)
1257{
1258 struct buffer_page *bpage;
1259 struct list_head *p;
1260 unsigned i;
1261
1262 spin_lock_irq(&cpu_buffer->reader_lock);
1263 rb_head_page_deactivate(cpu_buffer);
1264
1265 for (i = 0; i < nr_pages; i++) {
1266 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1267 goto out;
1268 p = cpu_buffer->pages->next;
1269 bpage = list_entry(p, struct buffer_page, list);
1270 list_del_init(&bpage->list);
1271 free_buffer_page(bpage);
1272 }
1273 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1274 goto out;
1275
1276 rb_reset_cpu(cpu_buffer);
1277 rb_check_pages(cpu_buffer);
1278
1279out:
1280 spin_unlock_irq(&cpu_buffer->reader_lock);
1281}
1282
1283static void
1284rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer,
1285 struct list_head *pages, unsigned nr_pages)
1286{
1287 struct buffer_page *bpage;
1288 struct list_head *p;
1289 unsigned i;
1290
1291 spin_lock_irq(&cpu_buffer->reader_lock);
1292 rb_head_page_deactivate(cpu_buffer);
1293
1294 for (i = 0; i < nr_pages; i++) {
1295 if (RB_WARN_ON(cpu_buffer, list_empty(pages)))
1296 goto out;
1297 p = pages->next;
1298 bpage = list_entry(p, struct buffer_page, list);
1299 list_del_init(&bpage->list);
1300 list_add_tail(&bpage->list, cpu_buffer->pages);
1301 }
1302 rb_reset_cpu(cpu_buffer);
1303 rb_check_pages(cpu_buffer);
1304
1305out:
1306 spin_unlock_irq(&cpu_buffer->reader_lock);
1307}
1308
1309/**
1310 * ring_buffer_resize - resize the ring buffer
1311 * @buffer: the buffer to resize.
1312 * @size: the new size.
1313 *
1314 * Minimum size is 2 * BUF_PAGE_SIZE.
1315 *
1316 * Returns -1 on failure.
1317 */
1318int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size)
1319{
1320 struct ring_buffer_per_cpu *cpu_buffer;
1321 unsigned nr_pages, rm_pages, new_pages;
1322 struct buffer_page *bpage, *tmp;
1323 unsigned long buffer_size;
1324 LIST_HEAD(pages);
1325 int i, cpu;
1326
1327 /*
1328 * Always succeed at resizing a non-existent buffer:
1329 */
1330 if (!buffer)
1331 return size;
1332
1333 size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1334 size *= BUF_PAGE_SIZE;
1335 buffer_size = buffer->pages * BUF_PAGE_SIZE;
1336
1337 /* we need a minimum of two pages */
1338 if (size < BUF_PAGE_SIZE * 2)
1339 size = BUF_PAGE_SIZE * 2;
1340
1341 if (size == buffer_size)
1342 return size;
1343
1344 atomic_inc(&buffer->record_disabled);
1345
1346 /* Make sure all writers are done with this buffer. */
1347 synchronize_sched();
1348
1349 mutex_lock(&buffer->mutex);
1350 get_online_cpus();
1351
1352 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1353
1354 if (size < buffer_size) {
1355
1356 /* easy case, just free pages */
1357 if (RB_WARN_ON(buffer, nr_pages >= buffer->pages))
1358 goto out_fail;
1359
1360 rm_pages = buffer->pages - nr_pages;
1361
1362 for_each_buffer_cpu(buffer, cpu) {
1363 cpu_buffer = buffer->buffers[cpu];
1364 rb_remove_pages(cpu_buffer, rm_pages);
1365 }
1366 goto out;
1367 }
1368
1369 /*
1370 * This is a bit more difficult. We only want to add pages
1371 * when we can allocate enough for all CPUs. We do this
1372 * by allocating all the pages and storing them on a local
1373 * link list. If we succeed in our allocation, then we
1374 * add these pages to the cpu_buffers. Otherwise we just free
1375 * them all and return -ENOMEM;
1376 */
1377 if (RB_WARN_ON(buffer, nr_pages <= buffer->pages))
1378 goto out_fail;
1379
1380 new_pages = nr_pages - buffer->pages;
1381
1382 for_each_buffer_cpu(buffer, cpu) {
1383 for (i = 0; i < new_pages; i++) {
1384 struct page *page;
1385 /*
1386 * __GFP_NORETRY flag makes sure that the allocation
1387 * fails gracefully without invoking oom-killer and
1388 * the system is not destabilized.
1389 */
1390 bpage = kzalloc_node(ALIGN(sizeof(*bpage),
1391 cache_line_size()),
1392 GFP_KERNEL | __GFP_NORETRY,
1393 cpu_to_node(cpu));
1394 if (!bpage)
1395 goto free_pages;
1396 list_add(&bpage->list, &pages);
1397 page = alloc_pages_node(cpu_to_node(cpu),
1398 GFP_KERNEL | __GFP_NORETRY, 0);
1399 if (!page)
1400 goto free_pages;
1401 bpage->page = page_address(page);
1402 rb_init_page(bpage->page);
1403 }
1404 }
1405
1406 for_each_buffer_cpu(buffer, cpu) {
1407 cpu_buffer = buffer->buffers[cpu];
1408 rb_insert_pages(cpu_buffer, &pages, new_pages);
1409 }
1410
1411 if (RB_WARN_ON(buffer, !list_empty(&pages)))
1412 goto out_fail;
1413
1414 out:
1415 buffer->pages = nr_pages;
1416 put_online_cpus();
1417 mutex_unlock(&buffer->mutex);
1418
1419 atomic_dec(&buffer->record_disabled);
1420
1421 return size;
1422
1423 free_pages:
1424 list_for_each_entry_safe(bpage, tmp, &pages, list) {
1425 list_del_init(&bpage->list);
1426 free_buffer_page(bpage);
1427 }
1428 put_online_cpus();
1429 mutex_unlock(&buffer->mutex);
1430 atomic_dec(&buffer->record_disabled);
1431 return -ENOMEM;
1432
1433 /*
1434 * Something went totally wrong, and we are too paranoid
1435 * to even clean up the mess.
1436 */
1437 out_fail:
1438 put_online_cpus();
1439 mutex_unlock(&buffer->mutex);
1440 atomic_dec(&buffer->record_disabled);
1441 return -1;
1442}
1443EXPORT_SYMBOL_GPL(ring_buffer_resize);
1444
1445void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1446{
1447 mutex_lock(&buffer->mutex);
1448 if (val)
1449 buffer->flags |= RB_FL_OVERWRITE;
1450 else
1451 buffer->flags &= ~RB_FL_OVERWRITE;
1452 mutex_unlock(&buffer->mutex);
1453}
1454EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1455
1456static inline void *
1457__rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1458{
1459 return bpage->data + index;
1460}
1461
1462static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1463{
1464 return bpage->page->data + index;
1465}
1466
1467static inline struct ring_buffer_event *
1468rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1469{
1470 return __rb_page_index(cpu_buffer->reader_page,
1471 cpu_buffer->reader_page->read);
1472}
1473
1474static inline struct ring_buffer_event *
1475rb_iter_head_event(struct ring_buffer_iter *iter)
1476{
1477 return __rb_page_index(iter->head_page, iter->head);
1478}
1479
1480static inline unsigned long rb_page_write(struct buffer_page *bpage)
1481{
1482 return local_read(&bpage->write) & RB_WRITE_MASK;
1483}
1484
1485static inline unsigned rb_page_commit(struct buffer_page *bpage)
1486{
1487 return local_read(&bpage->page->commit);
1488}
1489
1490static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1491{
1492 return local_read(&bpage->entries) & RB_WRITE_MASK;
1493}
1494
1495/* Size is determined by what has been committed */
1496static inline unsigned rb_page_size(struct buffer_page *bpage)
1497{
1498 return rb_page_commit(bpage);
1499}
1500
1501static inline unsigned
1502rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1503{
1504 return rb_page_commit(cpu_buffer->commit_page);
1505}
1506
1507static inline unsigned
1508rb_event_index(struct ring_buffer_event *event)
1509{
1510 unsigned long addr = (unsigned long)event;
1511
1512 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1513}
1514
1515static inline int
1516rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1517 struct ring_buffer_event *event)
1518{
1519 unsigned long addr = (unsigned long)event;
1520 unsigned long index;
1521
1522 index = rb_event_index(event);
1523 addr &= PAGE_MASK;
1524
1525 return cpu_buffer->commit_page->page == (void *)addr &&
1526 rb_commit_index(cpu_buffer) == index;
1527}
1528
1529static void
1530rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1531{
1532 unsigned long max_count;
1533
1534 /*
1535 * We only race with interrupts and NMIs on this CPU.
1536 * If we own the commit event, then we can commit
1537 * all others that interrupted us, since the interruptions
1538 * are in stack format (they finish before they come
1539 * back to us). This allows us to do a simple loop to
1540 * assign the commit to the tail.
1541 */
1542 again:
1543 max_count = cpu_buffer->buffer->pages * 100;
1544
1545 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1546 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1547 return;
1548 if (RB_WARN_ON(cpu_buffer,
1549 rb_is_reader_page(cpu_buffer->tail_page)))
1550 return;
1551 local_set(&cpu_buffer->commit_page->page->commit,
1552 rb_page_write(cpu_buffer->commit_page));
1553 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1554 cpu_buffer->write_stamp =
1555 cpu_buffer->commit_page->page->time_stamp;
1556 /* add barrier to keep gcc from optimizing too much */
1557 barrier();
1558 }
1559 while (rb_commit_index(cpu_buffer) !=
1560 rb_page_write(cpu_buffer->commit_page)) {
1561
1562 local_set(&cpu_buffer->commit_page->page->commit,
1563 rb_page_write(cpu_buffer->commit_page));
1564 RB_WARN_ON(cpu_buffer,
1565 local_read(&cpu_buffer->commit_page->page->commit) &
1566 ~RB_WRITE_MASK);
1567 barrier();
1568 }
1569
1570 /* again, keep gcc from optimizing */
1571 barrier();
1572
1573 /*
1574 * If an interrupt came in just after the first while loop
1575 * and pushed the tail page forward, we will be left with
1576 * a dangling commit that will never go forward.
1577 */
1578 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1579 goto again;
1580}
1581
1582static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1583{
1584 cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1585 cpu_buffer->reader_page->read = 0;
1586}
1587
1588static void rb_inc_iter(struct ring_buffer_iter *iter)
1589{
1590 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1591
1592 /*
1593 * The iterator could be on the reader page (it starts there).
1594 * But the head could have moved, since the reader was
1595 * found. Check for this case and assign the iterator
1596 * to the head page instead of next.
1597 */
1598 if (iter->head_page == cpu_buffer->reader_page)
1599 iter->head_page = rb_set_head_page(cpu_buffer);
1600 else
1601 rb_inc_page(cpu_buffer, &iter->head_page);
1602
1603 iter->read_stamp = iter->head_page->page->time_stamp;
1604 iter->head = 0;
1605}
1606
1607/* Slow path, do not inline */
1608static noinline struct ring_buffer_event *
1609rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
1610{
1611 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
1612
1613 /* Not the first event on the page? */
1614 if (rb_event_index(event)) {
1615 event->time_delta = delta & TS_MASK;
1616 event->array[0] = delta >> TS_SHIFT;
1617 } else {
1618 /* nope, just zero it */
1619 event->time_delta = 0;
1620 event->array[0] = 0;
1621 }
1622
1623 return skip_time_extend(event);
1624}
1625
1626/**
1627 * ring_buffer_update_event - update event type and data
1628 * @event: the even to update
1629 * @type: the type of event
1630 * @length: the size of the event field in the ring buffer
1631 *
1632 * Update the type and data fields of the event. The length
1633 * is the actual size that is written to the ring buffer,
1634 * and with this, we can determine what to place into the
1635 * data field.
1636 */
1637static void
1638rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
1639 struct ring_buffer_event *event, unsigned length,
1640 int add_timestamp, u64 delta)
1641{
1642 /* Only a commit updates the timestamp */
1643 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
1644 delta = 0;
1645
1646 /*
1647 * If we need to add a timestamp, then we
1648 * add it to the start of the resevered space.
1649 */
1650 if (unlikely(add_timestamp)) {
1651 event = rb_add_time_stamp(event, delta);
1652 length -= RB_LEN_TIME_EXTEND;
1653 delta = 0;
1654 }
1655
1656 event->time_delta = delta;
1657 length -= RB_EVNT_HDR_SIZE;
1658 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
1659 event->type_len = 0;
1660 event->array[0] = length;
1661 } else
1662 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
1663}
1664
1665/*
1666 * rb_handle_head_page - writer hit the head page
1667 *
1668 * Returns: +1 to retry page
1669 * 0 to continue
1670 * -1 on error
1671 */
1672static int
1673rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1674 struct buffer_page *tail_page,
1675 struct buffer_page *next_page)
1676{
1677 struct buffer_page *new_head;
1678 int entries;
1679 int type;
1680 int ret;
1681
1682 entries = rb_page_entries(next_page);
1683
1684 /*
1685 * The hard part is here. We need to move the head
1686 * forward, and protect against both readers on
1687 * other CPUs and writers coming in via interrupts.
1688 */
1689 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1690 RB_PAGE_HEAD);
1691
1692 /*
1693 * type can be one of four:
1694 * NORMAL - an interrupt already moved it for us
1695 * HEAD - we are the first to get here.
1696 * UPDATE - we are the interrupt interrupting
1697 * a current move.
1698 * MOVED - a reader on another CPU moved the next
1699 * pointer to its reader page. Give up
1700 * and try again.
1701 */
1702
1703 switch (type) {
1704 case RB_PAGE_HEAD:
1705 /*
1706 * We changed the head to UPDATE, thus
1707 * it is our responsibility to update
1708 * the counters.
1709 */
1710 local_add(entries, &cpu_buffer->overrun);
1711
1712 /*
1713 * The entries will be zeroed out when we move the
1714 * tail page.
1715 */
1716
1717 /* still more to do */
1718 break;
1719
1720 case RB_PAGE_UPDATE:
1721 /*
1722 * This is an interrupt that interrupt the
1723 * previous update. Still more to do.
1724 */
1725 break;
1726 case RB_PAGE_NORMAL:
1727 /*
1728 * An interrupt came in before the update
1729 * and processed this for us.
1730 * Nothing left to do.
1731 */
1732 return 1;
1733 case RB_PAGE_MOVED:
1734 /*
1735 * The reader is on another CPU and just did
1736 * a swap with our next_page.
1737 * Try again.
1738 */
1739 return 1;
1740 default:
1741 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1742 return -1;
1743 }
1744
1745 /*
1746 * Now that we are here, the old head pointer is
1747 * set to UPDATE. This will keep the reader from
1748 * swapping the head page with the reader page.
1749 * The reader (on another CPU) will spin till
1750 * we are finished.
1751 *
1752 * We just need to protect against interrupts
1753 * doing the job. We will set the next pointer
1754 * to HEAD. After that, we set the old pointer
1755 * to NORMAL, but only if it was HEAD before.
1756 * otherwise we are an interrupt, and only
1757 * want the outer most commit to reset it.
1758 */
1759 new_head = next_page;
1760 rb_inc_page(cpu_buffer, &new_head);
1761
1762 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1763 RB_PAGE_NORMAL);
1764
1765 /*
1766 * Valid returns are:
1767 * HEAD - an interrupt came in and already set it.
1768 * NORMAL - One of two things:
1769 * 1) We really set it.
1770 * 2) A bunch of interrupts came in and moved
1771 * the page forward again.
1772 */
1773 switch (ret) {
1774 case RB_PAGE_HEAD:
1775 case RB_PAGE_NORMAL:
1776 /* OK */
1777 break;
1778 default:
1779 RB_WARN_ON(cpu_buffer, 1);
1780 return -1;
1781 }
1782
1783 /*
1784 * It is possible that an interrupt came in,
1785 * set the head up, then more interrupts came in
1786 * and moved it again. When we get back here,
1787 * the page would have been set to NORMAL but we
1788 * just set it back to HEAD.
1789 *
1790 * How do you detect this? Well, if that happened
1791 * the tail page would have moved.
1792 */
1793 if (ret == RB_PAGE_NORMAL) {
1794 /*
1795 * If the tail had moved passed next, then we need
1796 * to reset the pointer.
1797 */
1798 if (cpu_buffer->tail_page != tail_page &&
1799 cpu_buffer->tail_page != next_page)
1800 rb_head_page_set_normal(cpu_buffer, new_head,
1801 next_page,
1802 RB_PAGE_HEAD);
1803 }
1804
1805 /*
1806 * If this was the outer most commit (the one that
1807 * changed the original pointer from HEAD to UPDATE),
1808 * then it is up to us to reset it to NORMAL.
1809 */
1810 if (type == RB_PAGE_HEAD) {
1811 ret = rb_head_page_set_normal(cpu_buffer, next_page,
1812 tail_page,
1813 RB_PAGE_UPDATE);
1814 if (RB_WARN_ON(cpu_buffer,
1815 ret != RB_PAGE_UPDATE))
1816 return -1;
1817 }
1818
1819 return 0;
1820}
1821
1822static unsigned rb_calculate_event_length(unsigned length)
1823{
1824 struct ring_buffer_event event; /* Used only for sizeof array */
1825
1826 /* zero length can cause confusions */
1827 if (!length)
1828 length = 1;
1829
1830 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
1831 length += sizeof(event.array[0]);
1832
1833 length += RB_EVNT_HDR_SIZE;
1834 length = ALIGN(length, RB_ARCH_ALIGNMENT);
1835
1836 return length;
1837}
1838
1839static inline void
1840rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
1841 struct buffer_page *tail_page,
1842 unsigned long tail, unsigned long length)
1843{
1844 struct ring_buffer_event *event;
1845
1846 /*
1847 * Only the event that crossed the page boundary
1848 * must fill the old tail_page with padding.
1849 */
1850 if (tail >= BUF_PAGE_SIZE) {
1851 /*
1852 * If the page was filled, then we still need
1853 * to update the real_end. Reset it to zero
1854 * and the reader will ignore it.
1855 */
1856 if (tail == BUF_PAGE_SIZE)
1857 tail_page->real_end = 0;
1858
1859 local_sub(length, &tail_page->write);
1860 return;
1861 }
1862
1863 event = __rb_page_index(tail_page, tail);
1864 kmemcheck_annotate_bitfield(event, bitfield);
1865
1866 /*
1867 * Save the original length to the meta data.
1868 * This will be used by the reader to add lost event
1869 * counter.
1870 */
1871 tail_page->real_end = tail;
1872
1873 /*
1874 * If this event is bigger than the minimum size, then
1875 * we need to be careful that we don't subtract the
1876 * write counter enough to allow another writer to slip
1877 * in on this page.
1878 * We put in a discarded commit instead, to make sure
1879 * that this space is not used again.
1880 *
1881 * If we are less than the minimum size, we don't need to
1882 * worry about it.
1883 */
1884 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
1885 /* No room for any events */
1886
1887 /* Mark the rest of the page with padding */
1888 rb_event_set_padding(event);
1889
1890 /* Set the write back to the previous setting */
1891 local_sub(length, &tail_page->write);
1892 return;
1893 }
1894
1895 /* Put in a discarded event */
1896 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
1897 event->type_len = RINGBUF_TYPE_PADDING;
1898 /* time delta must be non zero */
1899 event->time_delta = 1;
1900
1901 /* Set write to end of buffer */
1902 length = (tail + length) - BUF_PAGE_SIZE;
1903 local_sub(length, &tail_page->write);
1904}
1905
1906/*
1907 * This is the slow path, force gcc not to inline it.
1908 */
1909static noinline struct ring_buffer_event *
1910rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
1911 unsigned long length, unsigned long tail,
1912 struct buffer_page *tail_page, u64 ts)
1913{
1914 struct buffer_page *commit_page = cpu_buffer->commit_page;
1915 struct ring_buffer *buffer = cpu_buffer->buffer;
1916 struct buffer_page *next_page;
1917 int ret;
1918
1919 next_page = tail_page;
1920
1921 rb_inc_page(cpu_buffer, &next_page);
1922
1923 /*
1924 * If for some reason, we had an interrupt storm that made
1925 * it all the way around the buffer, bail, and warn
1926 * about it.
1927 */
1928 if (unlikely(next_page == commit_page)) {
1929 local_inc(&cpu_buffer->commit_overrun);
1930 goto out_reset;
1931 }
1932
1933 /*
1934 * This is where the fun begins!
1935 *
1936 * We are fighting against races between a reader that
1937 * could be on another CPU trying to swap its reader
1938 * page with the buffer head.
1939 *
1940 * We are also fighting against interrupts coming in and
1941 * moving the head or tail on us as well.
1942 *
1943 * If the next page is the head page then we have filled
1944 * the buffer, unless the commit page is still on the
1945 * reader page.
1946 */
1947 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
1948
1949 /*
1950 * If the commit is not on the reader page, then
1951 * move the header page.
1952 */
1953 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
1954 /*
1955 * If we are not in overwrite mode,
1956 * this is easy, just stop here.
1957 */
1958 if (!(buffer->flags & RB_FL_OVERWRITE))
1959 goto out_reset;
1960
1961 ret = rb_handle_head_page(cpu_buffer,
1962 tail_page,
1963 next_page);
1964 if (ret < 0)
1965 goto out_reset;
1966 if (ret)
1967 goto out_again;
1968 } else {
1969 /*
1970 * We need to be careful here too. The
1971 * commit page could still be on the reader
1972 * page. We could have a small buffer, and
1973 * have filled up the buffer with events
1974 * from interrupts and such, and wrapped.
1975 *
1976 * Note, if the tail page is also the on the
1977 * reader_page, we let it move out.
1978 */
1979 if (unlikely((cpu_buffer->commit_page !=
1980 cpu_buffer->tail_page) &&
1981 (cpu_buffer->commit_page ==
1982 cpu_buffer->reader_page))) {
1983 local_inc(&cpu_buffer->commit_overrun);
1984 goto out_reset;
1985 }
1986 }
1987 }
1988
1989 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
1990 if (ret) {
1991 /*
1992 * Nested commits always have zero deltas, so
1993 * just reread the time stamp
1994 */
1995 ts = rb_time_stamp(buffer);
1996 next_page->page->time_stamp = ts;
1997 }
1998
1999 out_again:
2000
2001 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2002
2003 /* fail and let the caller try again */
2004 return ERR_PTR(-EAGAIN);
2005
2006 out_reset:
2007 /* reset write */
2008 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2009
2010 return NULL;
2011}
2012
2013static struct ring_buffer_event *
2014__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2015 unsigned long length, u64 ts,
2016 u64 delta, int add_timestamp)
2017{
2018 struct buffer_page *tail_page;
2019 struct ring_buffer_event *event;
2020 unsigned long tail, write;
2021
2022 /*
2023 * If the time delta since the last event is too big to
2024 * hold in the time field of the event, then we append a
2025 * TIME EXTEND event ahead of the data event.
2026 */
2027 if (unlikely(add_timestamp))
2028 length += RB_LEN_TIME_EXTEND;
2029
2030 tail_page = cpu_buffer->tail_page;
2031 write = local_add_return(length, &tail_page->write);
2032
2033 /* set write to only the index of the write */
2034 write &= RB_WRITE_MASK;
2035 tail = write - length;
2036
2037 /* See if we shot pass the end of this buffer page */
2038 if (unlikely(write > BUF_PAGE_SIZE))
2039 return rb_move_tail(cpu_buffer, length, tail,
2040 tail_page, ts);
2041
2042 /* We reserved something on the buffer */
2043
2044 event = __rb_page_index(tail_page, tail);
2045 kmemcheck_annotate_bitfield(event, bitfield);
2046 rb_update_event(cpu_buffer, event, length, add_timestamp, delta);
2047
2048 local_inc(&tail_page->entries);
2049
2050 /*
2051 * If this is the first commit on the page, then update
2052 * its timestamp.
2053 */
2054 if (!tail)
2055 tail_page->page->time_stamp = ts;
2056
2057 return event;
2058}
2059
2060static inline int
2061rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2062 struct ring_buffer_event *event)
2063{
2064 unsigned long new_index, old_index;
2065 struct buffer_page *bpage;
2066 unsigned long index;
2067 unsigned long addr;
2068
2069 new_index = rb_event_index(event);
2070 old_index = new_index + rb_event_ts_length(event);
2071 addr = (unsigned long)event;
2072 addr &= PAGE_MASK;
2073
2074 bpage = cpu_buffer->tail_page;
2075
2076 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2077 unsigned long write_mask =
2078 local_read(&bpage->write) & ~RB_WRITE_MASK;
2079 /*
2080 * This is on the tail page. It is possible that
2081 * a write could come in and move the tail page
2082 * and write to the next page. That is fine
2083 * because we just shorten what is on this page.
2084 */
2085 old_index += write_mask;
2086 new_index += write_mask;
2087 index = local_cmpxchg(&bpage->write, old_index, new_index);
2088 if (index == old_index)
2089 return 1;
2090 }
2091
2092 /* could not discard */
2093 return 0;
2094}
2095
2096static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2097{
2098 local_inc(&cpu_buffer->committing);
2099 local_inc(&cpu_buffer->commits);
2100}
2101
2102static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2103{
2104 unsigned long commits;
2105
2106 if (RB_WARN_ON(cpu_buffer,
2107 !local_read(&cpu_buffer->committing)))
2108 return;
2109
2110 again:
2111 commits = local_read(&cpu_buffer->commits);
2112 /* synchronize with interrupts */
2113 barrier();
2114 if (local_read(&cpu_buffer->committing) == 1)
2115 rb_set_commit_to_write(cpu_buffer);
2116
2117 local_dec(&cpu_buffer->committing);
2118
2119 /* synchronize with interrupts */
2120 barrier();
2121
2122 /*
2123 * Need to account for interrupts coming in between the
2124 * updating of the commit page and the clearing of the
2125 * committing counter.
2126 */
2127 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2128 !local_read(&cpu_buffer->committing)) {
2129 local_inc(&cpu_buffer->committing);
2130 goto again;
2131 }
2132}
2133
2134static struct ring_buffer_event *
2135rb_reserve_next_event(struct ring_buffer *buffer,
2136 struct ring_buffer_per_cpu *cpu_buffer,
2137 unsigned long length)
2138{
2139 struct ring_buffer_event *event;
2140 u64 ts, delta;
2141 int nr_loops = 0;
2142 int add_timestamp;
2143 u64 diff;
2144
2145 rb_start_commit(cpu_buffer);
2146
2147#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2148 /*
2149 * Due to the ability to swap a cpu buffer from a buffer
2150 * it is possible it was swapped before we committed.
2151 * (committing stops a swap). We check for it here and
2152 * if it happened, we have to fail the write.
2153 */
2154 barrier();
2155 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2156 local_dec(&cpu_buffer->committing);
2157 local_dec(&cpu_buffer->commits);
2158 return NULL;
2159 }
2160#endif
2161
2162 length = rb_calculate_event_length(length);
2163 again:
2164 add_timestamp = 0;
2165 delta = 0;
2166
2167 /*
2168 * We allow for interrupts to reenter here and do a trace.
2169 * If one does, it will cause this original code to loop
2170 * back here. Even with heavy interrupts happening, this
2171 * should only happen a few times in a row. If this happens
2172 * 1000 times in a row, there must be either an interrupt
2173 * storm or we have something buggy.
2174 * Bail!
2175 */
2176 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2177 goto out_fail;
2178
2179 ts = rb_time_stamp(cpu_buffer->buffer);
2180 diff = ts - cpu_buffer->write_stamp;
2181
2182 /* make sure this diff is calculated here */
2183 barrier();
2184
2185 /* Did the write stamp get updated already? */
2186 if (likely(ts >= cpu_buffer->write_stamp)) {
2187 delta = diff;
2188 if (unlikely(test_time_stamp(delta))) {
2189 int local_clock_stable = 1;
2190#ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2191 local_clock_stable = sched_clock_stable;
2192#endif
2193 WARN_ONCE(delta > (1ULL << 59),
2194 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2195 (unsigned long long)delta,
2196 (unsigned long long)ts,
2197 (unsigned long long)cpu_buffer->write_stamp,
2198 local_clock_stable ? "" :
2199 "If you just came from a suspend/resume,\n"
2200 "please switch to the trace global clock:\n"
2201 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2202 add_timestamp = 1;
2203 }
2204 }
2205
2206 event = __rb_reserve_next(cpu_buffer, length, ts,
2207 delta, add_timestamp);
2208 if (unlikely(PTR_ERR(event) == -EAGAIN))
2209 goto again;
2210
2211 if (!event)
2212 goto out_fail;
2213
2214 return event;
2215
2216 out_fail:
2217 rb_end_commit(cpu_buffer);
2218 return NULL;
2219}
2220
2221#ifdef CONFIG_TRACING
2222
2223#define TRACE_RECURSIVE_DEPTH 16
2224
2225/* Keep this code out of the fast path cache */
2226static noinline void trace_recursive_fail(void)
2227{
2228 /* Disable all tracing before we do anything else */
2229 tracing_off_permanent();
2230
2231 printk_once(KERN_WARNING "Tracing recursion: depth[%ld]:"
2232 "HC[%lu]:SC[%lu]:NMI[%lu]\n",
2233 trace_recursion_buffer(),
2234 hardirq_count() >> HARDIRQ_SHIFT,
2235 softirq_count() >> SOFTIRQ_SHIFT,
2236 in_nmi());
2237
2238 WARN_ON_ONCE(1);
2239}
2240
2241static inline int trace_recursive_lock(void)
2242{
2243 trace_recursion_inc();
2244
2245 if (likely(trace_recursion_buffer() < TRACE_RECURSIVE_DEPTH))
2246 return 0;
2247
2248 trace_recursive_fail();
2249
2250 return -1;
2251}
2252
2253static inline void trace_recursive_unlock(void)
2254{
2255 WARN_ON_ONCE(!trace_recursion_buffer());
2256
2257 trace_recursion_dec();
2258}
2259
2260#else
2261
2262#define trace_recursive_lock() (0)
2263#define trace_recursive_unlock() do { } while (0)
2264
2265#endif
2266
2267/**
2268 * ring_buffer_lock_reserve - reserve a part of the buffer
2269 * @buffer: the ring buffer to reserve from
2270 * @length: the length of the data to reserve (excluding event header)
2271 *
2272 * Returns a reseverd event on the ring buffer to copy directly to.
2273 * The user of this interface will need to get the body to write into
2274 * and can use the ring_buffer_event_data() interface.
2275 *
2276 * The length is the length of the data needed, not the event length
2277 * which also includes the event header.
2278 *
2279 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2280 * If NULL is returned, then nothing has been allocated or locked.
2281 */
2282struct ring_buffer_event *
2283ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2284{
2285 struct ring_buffer_per_cpu *cpu_buffer;
2286 struct ring_buffer_event *event;
2287 int cpu;
2288
2289 if (ring_buffer_flags != RB_BUFFERS_ON)
2290 return NULL;
2291
2292 /* If we are tracing schedule, we don't want to recurse */
2293 preempt_disable_notrace();
2294
2295 if (atomic_read(&buffer->record_disabled))
2296 goto out_nocheck;
2297
2298 if (trace_recursive_lock())
2299 goto out_nocheck;
2300
2301 cpu = raw_smp_processor_id();
2302
2303 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2304 goto out;
2305
2306 cpu_buffer = buffer->buffers[cpu];
2307
2308 if (atomic_read(&cpu_buffer->record_disabled))
2309 goto out;
2310
2311 if (length > BUF_MAX_DATA_SIZE)
2312 goto out;
2313
2314 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2315 if (!event)
2316 goto out;
2317
2318 return event;
2319
2320 out:
2321 trace_recursive_unlock();
2322
2323 out_nocheck:
2324 preempt_enable_notrace();
2325 return NULL;
2326}
2327EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2328
2329static void
2330rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2331 struct ring_buffer_event *event)
2332{
2333 u64 delta;
2334
2335 /*
2336 * The event first in the commit queue updates the
2337 * time stamp.
2338 */
2339 if (rb_event_is_commit(cpu_buffer, event)) {
2340 /*
2341 * A commit event that is first on a page
2342 * updates the write timestamp with the page stamp
2343 */
2344 if (!rb_event_index(event))
2345 cpu_buffer->write_stamp =
2346 cpu_buffer->commit_page->page->time_stamp;
2347 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2348 delta = event->array[0];
2349 delta <<= TS_SHIFT;
2350 delta += event->time_delta;
2351 cpu_buffer->write_stamp += delta;
2352 } else
2353 cpu_buffer->write_stamp += event->time_delta;
2354 }
2355}
2356
2357static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2358 struct ring_buffer_event *event)
2359{
2360 local_inc(&cpu_buffer->entries);
2361 rb_update_write_stamp(cpu_buffer, event);
2362 rb_end_commit(cpu_buffer);
2363}
2364
2365/**
2366 * ring_buffer_unlock_commit - commit a reserved
2367 * @buffer: The buffer to commit to
2368 * @event: The event pointer to commit.
2369 *
2370 * This commits the data to the ring buffer, and releases any locks held.
2371 *
2372 * Must be paired with ring_buffer_lock_reserve.
2373 */
2374int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2375 struct ring_buffer_event *event)
2376{
2377 struct ring_buffer_per_cpu *cpu_buffer;
2378 int cpu = raw_smp_processor_id();
2379
2380 cpu_buffer = buffer->buffers[cpu];
2381
2382 rb_commit(cpu_buffer, event);
2383
2384 trace_recursive_unlock();
2385
2386 preempt_enable_notrace();
2387
2388 return 0;
2389}
2390EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2391
2392static inline void rb_event_discard(struct ring_buffer_event *event)
2393{
2394 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2395 event = skip_time_extend(event);
2396
2397 /* array[0] holds the actual length for the discarded event */
2398 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2399 event->type_len = RINGBUF_TYPE_PADDING;
2400 /* time delta must be non zero */
2401 if (!event->time_delta)
2402 event->time_delta = 1;
2403}
2404
2405/*
2406 * Decrement the entries to the page that an event is on.
2407 * The event does not even need to exist, only the pointer
2408 * to the page it is on. This may only be called before the commit
2409 * takes place.
2410 */
2411static inline void
2412rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2413 struct ring_buffer_event *event)
2414{
2415 unsigned long addr = (unsigned long)event;
2416 struct buffer_page *bpage = cpu_buffer->commit_page;
2417 struct buffer_page *start;
2418
2419 addr &= PAGE_MASK;
2420
2421 /* Do the likely case first */
2422 if (likely(bpage->page == (void *)addr)) {
2423 local_dec(&bpage->entries);
2424 return;
2425 }
2426
2427 /*
2428 * Because the commit page may be on the reader page we
2429 * start with the next page and check the end loop there.
2430 */
2431 rb_inc_page(cpu_buffer, &bpage);
2432 start = bpage;
2433 do {
2434 if (bpage->page == (void *)addr) {
2435 local_dec(&bpage->entries);
2436 return;
2437 }
2438 rb_inc_page(cpu_buffer, &bpage);
2439 } while (bpage != start);
2440
2441 /* commit not part of this buffer?? */
2442 RB_WARN_ON(cpu_buffer, 1);
2443}
2444
2445/**
2446 * ring_buffer_commit_discard - discard an event that has not been committed
2447 * @buffer: the ring buffer
2448 * @event: non committed event to discard
2449 *
2450 * Sometimes an event that is in the ring buffer needs to be ignored.
2451 * This function lets the user discard an event in the ring buffer
2452 * and then that event will not be read later.
2453 *
2454 * This function only works if it is called before the the item has been
2455 * committed. It will try to free the event from the ring buffer
2456 * if another event has not been added behind it.
2457 *
2458 * If another event has been added behind it, it will set the event
2459 * up as discarded, and perform the commit.
2460 *
2461 * If this function is called, do not call ring_buffer_unlock_commit on
2462 * the event.
2463 */
2464void ring_buffer_discard_commit(struct ring_buffer *buffer,
2465 struct ring_buffer_event *event)
2466{
2467 struct ring_buffer_per_cpu *cpu_buffer;
2468 int cpu;
2469
2470 /* The event is discarded regardless */
2471 rb_event_discard(event);
2472
2473 cpu = smp_processor_id();
2474 cpu_buffer = buffer->buffers[cpu];
2475
2476 /*
2477 * This must only be called if the event has not been
2478 * committed yet. Thus we can assume that preemption
2479 * is still disabled.
2480 */
2481 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2482
2483 rb_decrement_entry(cpu_buffer, event);
2484 if (rb_try_to_discard(cpu_buffer, event))
2485 goto out;
2486
2487 /*
2488 * The commit is still visible by the reader, so we
2489 * must still update the timestamp.
2490 */
2491 rb_update_write_stamp(cpu_buffer, event);
2492 out:
2493 rb_end_commit(cpu_buffer);
2494
2495 trace_recursive_unlock();
2496
2497 preempt_enable_notrace();
2498
2499}
2500EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2501
2502/**
2503 * ring_buffer_write - write data to the buffer without reserving
2504 * @buffer: The ring buffer to write to.
2505 * @length: The length of the data being written (excluding the event header)
2506 * @data: The data to write to the buffer.
2507 *
2508 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2509 * one function. If you already have the data to write to the buffer, it
2510 * may be easier to simply call this function.
2511 *
2512 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2513 * and not the length of the event which would hold the header.
2514 */
2515int ring_buffer_write(struct ring_buffer *buffer,
2516 unsigned long length,
2517 void *data)
2518{
2519 struct ring_buffer_per_cpu *cpu_buffer;
2520 struct ring_buffer_event *event;
2521 void *body;
2522 int ret = -EBUSY;
2523 int cpu;
2524
2525 if (ring_buffer_flags != RB_BUFFERS_ON)
2526 return -EBUSY;
2527
2528 preempt_disable_notrace();
2529
2530 if (atomic_read(&buffer->record_disabled))
2531 goto out;
2532
2533 cpu = raw_smp_processor_id();
2534
2535 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2536 goto out;
2537
2538 cpu_buffer = buffer->buffers[cpu];
2539
2540 if (atomic_read(&cpu_buffer->record_disabled))
2541 goto out;
2542
2543 if (length > BUF_MAX_DATA_SIZE)
2544 goto out;
2545
2546 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2547 if (!event)
2548 goto out;
2549
2550 body = rb_event_data(event);
2551
2552 memcpy(body, data, length);
2553
2554 rb_commit(cpu_buffer, event);
2555
2556 ret = 0;
2557 out:
2558 preempt_enable_notrace();
2559
2560 return ret;
2561}
2562EXPORT_SYMBOL_GPL(ring_buffer_write);
2563
2564static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
2565{
2566 struct buffer_page *reader = cpu_buffer->reader_page;
2567 struct buffer_page *head = rb_set_head_page(cpu_buffer);
2568 struct buffer_page *commit = cpu_buffer->commit_page;
2569
2570 /* In case of error, head will be NULL */
2571 if (unlikely(!head))
2572 return 1;
2573
2574 return reader->read == rb_page_commit(reader) &&
2575 (commit == reader ||
2576 (commit == head &&
2577 head->read == rb_page_commit(commit)));
2578}
2579
2580/**
2581 * ring_buffer_record_disable - stop all writes into the buffer
2582 * @buffer: The ring buffer to stop writes to.
2583 *
2584 * This prevents all writes to the buffer. Any attempt to write
2585 * to the buffer after this will fail and return NULL.
2586 *
2587 * The caller should call synchronize_sched() after this.
2588 */
2589void ring_buffer_record_disable(struct ring_buffer *buffer)
2590{
2591 atomic_inc(&buffer->record_disabled);
2592}
2593EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
2594
2595/**
2596 * ring_buffer_record_enable - enable writes to the buffer
2597 * @buffer: The ring buffer to enable writes
2598 *
2599 * Note, multiple disables will need the same number of enables
2600 * to truly enable the writing (much like preempt_disable).
2601 */
2602void ring_buffer_record_enable(struct ring_buffer *buffer)
2603{
2604 atomic_dec(&buffer->record_disabled);
2605}
2606EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
2607
2608/**
2609 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
2610 * @buffer: The ring buffer to stop writes to.
2611 * @cpu: The CPU buffer to stop
2612 *
2613 * This prevents all writes to the buffer. Any attempt to write
2614 * to the buffer after this will fail and return NULL.
2615 *
2616 * The caller should call synchronize_sched() after this.
2617 */
2618void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
2619{
2620 struct ring_buffer_per_cpu *cpu_buffer;
2621
2622 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2623 return;
2624
2625 cpu_buffer = buffer->buffers[cpu];
2626 atomic_inc(&cpu_buffer->record_disabled);
2627}
2628EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
2629
2630/**
2631 * ring_buffer_record_enable_cpu - enable writes to the buffer
2632 * @buffer: The ring buffer to enable writes
2633 * @cpu: The CPU to enable.
2634 *
2635 * Note, multiple disables will need the same number of enables
2636 * to truly enable the writing (much like preempt_disable).
2637 */
2638void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
2639{
2640 struct ring_buffer_per_cpu *cpu_buffer;
2641
2642 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2643 return;
2644
2645 cpu_buffer = buffer->buffers[cpu];
2646 atomic_dec(&cpu_buffer->record_disabled);
2647}
2648EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
2649
2650/*
2651 * The total entries in the ring buffer is the running counter
2652 * of entries entered into the ring buffer, minus the sum of
2653 * the entries read from the ring buffer and the number of
2654 * entries that were overwritten.
2655 */
2656static inline unsigned long
2657rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
2658{
2659 return local_read(&cpu_buffer->entries) -
2660 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
2661}
2662
2663/**
2664 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
2665 * @buffer: The ring buffer
2666 * @cpu: The per CPU buffer to get the entries from.
2667 */
2668unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
2669{
2670 struct ring_buffer_per_cpu *cpu_buffer;
2671
2672 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2673 return 0;
2674
2675 cpu_buffer = buffer->buffers[cpu];
2676
2677 return rb_num_of_entries(cpu_buffer);
2678}
2679EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
2680
2681/**
2682 * ring_buffer_overrun_cpu - get the number of overruns in a cpu_buffer
2683 * @buffer: The ring buffer
2684 * @cpu: The per CPU buffer to get the number of overruns from
2685 */
2686unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
2687{
2688 struct ring_buffer_per_cpu *cpu_buffer;
2689 unsigned long ret;
2690
2691 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2692 return 0;
2693
2694 cpu_buffer = buffer->buffers[cpu];
2695 ret = local_read(&cpu_buffer->overrun);
2696
2697 return ret;
2698}
2699EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
2700
2701/**
2702 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by commits
2703 * @buffer: The ring buffer
2704 * @cpu: The per CPU buffer to get the number of overruns from
2705 */
2706unsigned long
2707ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
2708{
2709 struct ring_buffer_per_cpu *cpu_buffer;
2710 unsigned long ret;
2711
2712 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2713 return 0;
2714
2715 cpu_buffer = buffer->buffers[cpu];
2716 ret = local_read(&cpu_buffer->commit_overrun);
2717
2718 return ret;
2719}
2720EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
2721
2722/**
2723 * ring_buffer_entries - get the number of entries in a buffer
2724 * @buffer: The ring buffer
2725 *
2726 * Returns the total number of entries in the ring buffer
2727 * (all CPU entries)
2728 */
2729unsigned long ring_buffer_entries(struct ring_buffer *buffer)
2730{
2731 struct ring_buffer_per_cpu *cpu_buffer;
2732 unsigned long entries = 0;
2733 int cpu;
2734
2735 /* if you care about this being correct, lock the buffer */
2736 for_each_buffer_cpu(buffer, cpu) {
2737 cpu_buffer = buffer->buffers[cpu];
2738 entries += rb_num_of_entries(cpu_buffer);
2739 }
2740
2741 return entries;
2742}
2743EXPORT_SYMBOL_GPL(ring_buffer_entries);
2744
2745/**
2746 * ring_buffer_overruns - get the number of overruns in buffer
2747 * @buffer: The ring buffer
2748 *
2749 * Returns the total number of overruns in the ring buffer
2750 * (all CPU entries)
2751 */
2752unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
2753{
2754 struct ring_buffer_per_cpu *cpu_buffer;
2755 unsigned long overruns = 0;
2756 int cpu;
2757
2758 /* if you care about this being correct, lock the buffer */
2759 for_each_buffer_cpu(buffer, cpu) {
2760 cpu_buffer = buffer->buffers[cpu];
2761 overruns += local_read(&cpu_buffer->overrun);
2762 }
2763
2764 return overruns;
2765}
2766EXPORT_SYMBOL_GPL(ring_buffer_overruns);
2767
2768static void rb_iter_reset(struct ring_buffer_iter *iter)
2769{
2770 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2771
2772 /* Iterator usage is expected to have record disabled */
2773 if (list_empty(&cpu_buffer->reader_page->list)) {
2774 iter->head_page = rb_set_head_page(cpu_buffer);
2775 if (unlikely(!iter->head_page))
2776 return;
2777 iter->head = iter->head_page->read;
2778 } else {
2779 iter->head_page = cpu_buffer->reader_page;
2780 iter->head = cpu_buffer->reader_page->read;
2781 }
2782 if (iter->head)
2783 iter->read_stamp = cpu_buffer->read_stamp;
2784 else
2785 iter->read_stamp = iter->head_page->page->time_stamp;
2786 iter->cache_reader_page = cpu_buffer->reader_page;
2787 iter->cache_read = cpu_buffer->read;
2788}
2789
2790/**
2791 * ring_buffer_iter_reset - reset an iterator
2792 * @iter: The iterator to reset
2793 *
2794 * Resets the iterator, so that it will start from the beginning
2795 * again.
2796 */
2797void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
2798{
2799 struct ring_buffer_per_cpu *cpu_buffer;
2800 unsigned long flags;
2801
2802 if (!iter)
2803 return;
2804
2805 cpu_buffer = iter->cpu_buffer;
2806
2807 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2808 rb_iter_reset(iter);
2809 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2810}
2811EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
2812
2813/**
2814 * ring_buffer_iter_empty - check if an iterator has no more to read
2815 * @iter: The iterator to check
2816 */
2817int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
2818{
2819 struct ring_buffer_per_cpu *cpu_buffer;
2820
2821 cpu_buffer = iter->cpu_buffer;
2822
2823 return iter->head_page == cpu_buffer->commit_page &&
2824 iter->head == rb_commit_index(cpu_buffer);
2825}
2826EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
2827
2828static void
2829rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2830 struct ring_buffer_event *event)
2831{
2832 u64 delta;
2833
2834 switch (event->type_len) {
2835 case RINGBUF_TYPE_PADDING:
2836 return;
2837
2838 case RINGBUF_TYPE_TIME_EXTEND:
2839 delta = event->array[0];
2840 delta <<= TS_SHIFT;
2841 delta += event->time_delta;
2842 cpu_buffer->read_stamp += delta;
2843 return;
2844
2845 case RINGBUF_TYPE_TIME_STAMP:
2846 /* FIXME: not implemented */
2847 return;
2848
2849 case RINGBUF_TYPE_DATA:
2850 cpu_buffer->read_stamp += event->time_delta;
2851 return;
2852
2853 default:
2854 BUG();
2855 }
2856 return;
2857}
2858
2859static void
2860rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
2861 struct ring_buffer_event *event)
2862{
2863 u64 delta;
2864
2865 switch (event->type_len) {
2866 case RINGBUF_TYPE_PADDING:
2867 return;
2868
2869 case RINGBUF_TYPE_TIME_EXTEND:
2870 delta = event->array[0];
2871 delta <<= TS_SHIFT;
2872 delta += event->time_delta;
2873 iter->read_stamp += delta;
2874 return;
2875
2876 case RINGBUF_TYPE_TIME_STAMP:
2877 /* FIXME: not implemented */
2878 return;
2879
2880 case RINGBUF_TYPE_DATA:
2881 iter->read_stamp += event->time_delta;
2882 return;
2883
2884 default:
2885 BUG();
2886 }
2887 return;
2888}
2889
2890static struct buffer_page *
2891rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
2892{
2893 struct buffer_page *reader = NULL;
2894 unsigned long overwrite;
2895 unsigned long flags;
2896 int nr_loops = 0;
2897 int ret;
2898
2899 local_irq_save(flags);
2900 arch_spin_lock(&cpu_buffer->lock);
2901
2902 again:
2903 /*
2904 * This should normally only loop twice. But because the
2905 * start of the reader inserts an empty page, it causes
2906 * a case where we will loop three times. There should be no
2907 * reason to loop four times (that I know of).
2908 */
2909 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
2910 reader = NULL;
2911 goto out;
2912 }
2913
2914 reader = cpu_buffer->reader_page;
2915
2916 /* If there's more to read, return this page */
2917 if (cpu_buffer->reader_page->read < rb_page_size(reader))
2918 goto out;
2919
2920 /* Never should we have an index greater than the size */
2921 if (RB_WARN_ON(cpu_buffer,
2922 cpu_buffer->reader_page->read > rb_page_size(reader)))
2923 goto out;
2924
2925 /* check if we caught up to the tail */
2926 reader = NULL;
2927 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
2928 goto out;
2929
2930 /*
2931 * Reset the reader page to size zero.
2932 */
2933 local_set(&cpu_buffer->reader_page->write, 0);
2934 local_set(&cpu_buffer->reader_page->entries, 0);
2935 local_set(&cpu_buffer->reader_page->page->commit, 0);
2936 cpu_buffer->reader_page->real_end = 0;
2937
2938 spin:
2939 /*
2940 * Splice the empty reader page into the list around the head.
2941 */
2942 reader = rb_set_head_page(cpu_buffer);
2943 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
2944 cpu_buffer->reader_page->list.prev = reader->list.prev;
2945
2946 /*
2947 * cpu_buffer->pages just needs to point to the buffer, it
2948 * has no specific buffer page to point to. Lets move it out
2949 * of our way so we don't accidentally swap it.
2950 */
2951 cpu_buffer->pages = reader->list.prev;
2952
2953 /* The reader page will be pointing to the new head */
2954 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
2955
2956 /*
2957 * We want to make sure we read the overruns after we set up our
2958 * pointers to the next object. The writer side does a
2959 * cmpxchg to cross pages which acts as the mb on the writer
2960 * side. Note, the reader will constantly fail the swap
2961 * while the writer is updating the pointers, so this
2962 * guarantees that the overwrite recorded here is the one we
2963 * want to compare with the last_overrun.
2964 */
2965 smp_mb();
2966 overwrite = local_read(&(cpu_buffer->overrun));
2967
2968 /*
2969 * Here's the tricky part.
2970 *
2971 * We need to move the pointer past the header page.
2972 * But we can only do that if a writer is not currently
2973 * moving it. The page before the header page has the
2974 * flag bit '1' set if it is pointing to the page we want.
2975 * but if the writer is in the process of moving it
2976 * than it will be '2' or already moved '0'.
2977 */
2978
2979 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
2980
2981 /*
2982 * If we did not convert it, then we must try again.
2983 */
2984 if (!ret)
2985 goto spin;
2986
2987 /*
2988 * Yeah! We succeeded in replacing the page.
2989 *
2990 * Now make the new head point back to the reader page.
2991 */
2992 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
2993 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
2994
2995 /* Finally update the reader page to the new head */
2996 cpu_buffer->reader_page = reader;
2997 rb_reset_reader_page(cpu_buffer);
2998
2999 if (overwrite != cpu_buffer->last_overrun) {
3000 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3001 cpu_buffer->last_overrun = overwrite;
3002 }
3003
3004 goto again;
3005
3006 out:
3007 arch_spin_unlock(&cpu_buffer->lock);
3008 local_irq_restore(flags);
3009
3010 return reader;
3011}
3012
3013static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3014{
3015 struct ring_buffer_event *event;
3016 struct buffer_page *reader;
3017 unsigned length;
3018
3019 reader = rb_get_reader_page(cpu_buffer);
3020
3021 /* This function should not be called when buffer is empty */
3022 if (RB_WARN_ON(cpu_buffer, !reader))
3023 return;
3024
3025 event = rb_reader_event(cpu_buffer);
3026
3027 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3028 cpu_buffer->read++;
3029
3030 rb_update_read_stamp(cpu_buffer, event);
3031
3032 length = rb_event_length(event);
3033 cpu_buffer->reader_page->read += length;
3034}
3035
3036static void rb_advance_iter(struct ring_buffer_iter *iter)
3037{
3038 struct ring_buffer_per_cpu *cpu_buffer;
3039 struct ring_buffer_event *event;
3040 unsigned length;
3041
3042 cpu_buffer = iter->cpu_buffer;
3043
3044 /*
3045 * Check if we are at the end of the buffer.
3046 */
3047 if (iter->head >= rb_page_size(iter->head_page)) {
3048 /* discarded commits can make the page empty */
3049 if (iter->head_page == cpu_buffer->commit_page)
3050 return;
3051 rb_inc_iter(iter);
3052 return;
3053 }
3054
3055 event = rb_iter_head_event(iter);
3056
3057 length = rb_event_length(event);
3058
3059 /*
3060 * This should not be called to advance the header if we are
3061 * at the tail of the buffer.
3062 */
3063 if (RB_WARN_ON(cpu_buffer,
3064 (iter->head_page == cpu_buffer->commit_page) &&
3065 (iter->head + length > rb_commit_index(cpu_buffer))))
3066 return;
3067
3068 rb_update_iter_read_stamp(iter, event);
3069
3070 iter->head += length;
3071
3072 /* check for end of page padding */
3073 if ((iter->head >= rb_page_size(iter->head_page)) &&
3074 (iter->head_page != cpu_buffer->commit_page))
3075 rb_advance_iter(iter);
3076}
3077
3078static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3079{
3080 return cpu_buffer->lost_events;
3081}
3082
3083static struct ring_buffer_event *
3084rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3085 unsigned long *lost_events)
3086{
3087 struct ring_buffer_event *event;
3088 struct buffer_page *reader;
3089 int nr_loops = 0;
3090
3091 again:
3092 /*
3093 * We repeat when a time extend is encountered.
3094 * Since the time extend is always attached to a data event,
3095 * we should never loop more than once.
3096 * (We never hit the following condition more than twice).
3097 */
3098 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3099 return NULL;
3100
3101 reader = rb_get_reader_page(cpu_buffer);
3102 if (!reader)
3103 return NULL;
3104
3105 event = rb_reader_event(cpu_buffer);
3106
3107 switch (event->type_len) {
3108 case RINGBUF_TYPE_PADDING:
3109 if (rb_null_event(event))
3110 RB_WARN_ON(cpu_buffer, 1);
3111 /*
3112 * Because the writer could be discarding every
3113 * event it creates (which would probably be bad)
3114 * if we were to go back to "again" then we may never
3115 * catch up, and will trigger the warn on, or lock
3116 * the box. Return the padding, and we will release
3117 * the current locks, and try again.
3118 */
3119 return event;
3120
3121 case RINGBUF_TYPE_TIME_EXTEND:
3122 /* Internal data, OK to advance */
3123 rb_advance_reader(cpu_buffer);
3124 goto again;
3125
3126 case RINGBUF_TYPE_TIME_STAMP:
3127 /* FIXME: not implemented */
3128 rb_advance_reader(cpu_buffer);
3129 goto again;
3130
3131 case RINGBUF_TYPE_DATA:
3132 if (ts) {
3133 *ts = cpu_buffer->read_stamp + event->time_delta;
3134 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3135 cpu_buffer->cpu, ts);
3136 }
3137 if (lost_events)
3138 *lost_events = rb_lost_events(cpu_buffer);
3139 return event;
3140
3141 default:
3142 BUG();
3143 }
3144
3145 return NULL;
3146}
3147EXPORT_SYMBOL_GPL(ring_buffer_peek);
3148
3149static struct ring_buffer_event *
3150rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3151{
3152 struct ring_buffer *buffer;
3153 struct ring_buffer_per_cpu *cpu_buffer;
3154 struct ring_buffer_event *event;
3155 int nr_loops = 0;
3156
3157 cpu_buffer = iter->cpu_buffer;
3158 buffer = cpu_buffer->buffer;
3159
3160 /*
3161 * Check if someone performed a consuming read to
3162 * the buffer. A consuming read invalidates the iterator
3163 * and we need to reset the iterator in this case.
3164 */
3165 if (unlikely(iter->cache_read != cpu_buffer->read ||
3166 iter->cache_reader_page != cpu_buffer->reader_page))
3167 rb_iter_reset(iter);
3168
3169 again:
3170 if (ring_buffer_iter_empty(iter))
3171 return NULL;
3172
3173 /*
3174 * We repeat when a time extend is encountered.
3175 * Since the time extend is always attached to a data event,
3176 * we should never loop more than once.
3177 * (We never hit the following condition more than twice).
3178 */
3179 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3180 return NULL;
3181
3182 if (rb_per_cpu_empty(cpu_buffer))
3183 return NULL;
3184
3185 if (iter->head >= local_read(&iter->head_page->page->commit)) {
3186 rb_inc_iter(iter);
3187 goto again;
3188 }
3189
3190 event = rb_iter_head_event(iter);
3191
3192 switch (event->type_len) {
3193 case RINGBUF_TYPE_PADDING:
3194 if (rb_null_event(event)) {
3195 rb_inc_iter(iter);
3196 goto again;
3197 }
3198 rb_advance_iter(iter);
3199 return event;
3200
3201 case RINGBUF_TYPE_TIME_EXTEND:
3202 /* Internal data, OK to advance */
3203 rb_advance_iter(iter);
3204 goto again;
3205
3206 case RINGBUF_TYPE_TIME_STAMP:
3207 /* FIXME: not implemented */
3208 rb_advance_iter(iter);
3209 goto again;
3210
3211 case RINGBUF_TYPE_DATA:
3212 if (ts) {
3213 *ts = iter->read_stamp + event->time_delta;
3214 ring_buffer_normalize_time_stamp(buffer,
3215 cpu_buffer->cpu, ts);
3216 }
3217 return event;
3218
3219 default:
3220 BUG();
3221 }
3222
3223 return NULL;
3224}
3225EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3226
3227static inline int rb_ok_to_lock(void)
3228{
3229 /*
3230 * If an NMI die dumps out the content of the ring buffer
3231 * do not grab locks. We also permanently disable the ring
3232 * buffer too. A one time deal is all you get from reading
3233 * the ring buffer from an NMI.
3234 */
3235 if (likely(!in_nmi()))
3236 return 1;
3237
3238 tracing_off_permanent();
3239 return 0;
3240}
3241
3242/**
3243 * ring_buffer_peek - peek at the next event to be read
3244 * @buffer: The ring buffer to read
3245 * @cpu: The cpu to peak at
3246 * @ts: The timestamp counter of this event.
3247 * @lost_events: a variable to store if events were lost (may be NULL)
3248 *
3249 * This will return the event that will be read next, but does
3250 * not consume the data.
3251 */
3252struct ring_buffer_event *
3253ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3254 unsigned long *lost_events)
3255{
3256 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3257 struct ring_buffer_event *event;
3258 unsigned long flags;
3259 int dolock;
3260
3261 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3262 return NULL;
3263
3264 dolock = rb_ok_to_lock();
3265 again:
3266 local_irq_save(flags);
3267 if (dolock)
3268 spin_lock(&cpu_buffer->reader_lock);
3269 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3270 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3271 rb_advance_reader(cpu_buffer);
3272 if (dolock)
3273 spin_unlock(&cpu_buffer->reader_lock);
3274 local_irq_restore(flags);
3275
3276 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3277 goto again;
3278
3279 return event;
3280}
3281
3282/**
3283 * ring_buffer_iter_peek - peek at the next event to be read
3284 * @iter: The ring buffer iterator
3285 * @ts: The timestamp counter of this event.
3286 *
3287 * This will return the event that will be read next, but does
3288 * not increment the iterator.
3289 */
3290struct ring_buffer_event *
3291ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3292{
3293 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3294 struct ring_buffer_event *event;
3295 unsigned long flags;
3296
3297 again:
3298 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3299 event = rb_iter_peek(iter, ts);
3300 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3301
3302 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3303 goto again;
3304
3305 return event;
3306}
3307
3308/**
3309 * ring_buffer_consume - return an event and consume it
3310 * @buffer: The ring buffer to get the next event from
3311 * @cpu: the cpu to read the buffer from
3312 * @ts: a variable to store the timestamp (may be NULL)
3313 * @lost_events: a variable to store if events were lost (may be NULL)
3314 *
3315 * Returns the next event in the ring buffer, and that event is consumed.
3316 * Meaning, that sequential reads will keep returning a different event,
3317 * and eventually empty the ring buffer if the producer is slower.
3318 */
3319struct ring_buffer_event *
3320ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3321 unsigned long *lost_events)
3322{
3323 struct ring_buffer_per_cpu *cpu_buffer;
3324 struct ring_buffer_event *event = NULL;
3325 unsigned long flags;
3326 int dolock;
3327
3328 dolock = rb_ok_to_lock();
3329
3330 again:
3331 /* might be called in atomic */
3332 preempt_disable();
3333
3334 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3335 goto out;
3336
3337 cpu_buffer = buffer->buffers[cpu];
3338 local_irq_save(flags);
3339 if (dolock)
3340 spin_lock(&cpu_buffer->reader_lock);
3341
3342 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3343 if (event) {
3344 cpu_buffer->lost_events = 0;
3345 rb_advance_reader(cpu_buffer);
3346 }
3347
3348 if (dolock)
3349 spin_unlock(&cpu_buffer->reader_lock);
3350 local_irq_restore(flags);
3351
3352 out:
3353 preempt_enable();
3354
3355 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3356 goto again;
3357
3358 return event;
3359}
3360EXPORT_SYMBOL_GPL(ring_buffer_consume);
3361
3362/**
3363 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3364 * @buffer: The ring buffer to read from
3365 * @cpu: The cpu buffer to iterate over
3366 *
3367 * This performs the initial preparations necessary to iterate
3368 * through the buffer. Memory is allocated, buffer recording
3369 * is disabled, and the iterator pointer is returned to the caller.
3370 *
3371 * Disabling buffer recordng prevents the reading from being
3372 * corrupted. This is not a consuming read, so a producer is not
3373 * expected.
3374 *
3375 * After a sequence of ring_buffer_read_prepare calls, the user is
3376 * expected to make at least one call to ring_buffer_prepare_sync.
3377 * Afterwards, ring_buffer_read_start is invoked to get things going
3378 * for real.
3379 *
3380 * This overall must be paired with ring_buffer_finish.
3381 */
3382struct ring_buffer_iter *
3383ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3384{
3385 struct ring_buffer_per_cpu *cpu_buffer;
3386 struct ring_buffer_iter *iter;
3387
3388 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3389 return NULL;
3390
3391 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3392 if (!iter)
3393 return NULL;
3394
3395 cpu_buffer = buffer->buffers[cpu];
3396
3397 iter->cpu_buffer = cpu_buffer;
3398
3399 atomic_inc(&cpu_buffer->record_disabled);
3400
3401 return iter;
3402}
3403EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
3404
3405/**
3406 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
3407 *
3408 * All previously invoked ring_buffer_read_prepare calls to prepare
3409 * iterators will be synchronized. Afterwards, read_buffer_read_start
3410 * calls on those iterators are allowed.
3411 */
3412void
3413ring_buffer_read_prepare_sync(void)
3414{
3415 synchronize_sched();
3416}
3417EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
3418
3419/**
3420 * ring_buffer_read_start - start a non consuming read of the buffer
3421 * @iter: The iterator returned by ring_buffer_read_prepare
3422 *
3423 * This finalizes the startup of an iteration through the buffer.
3424 * The iterator comes from a call to ring_buffer_read_prepare and
3425 * an intervening ring_buffer_read_prepare_sync must have been
3426 * performed.
3427 *
3428 * Must be paired with ring_buffer_finish.
3429 */
3430void
3431ring_buffer_read_start(struct ring_buffer_iter *iter)
3432{
3433 struct ring_buffer_per_cpu *cpu_buffer;
3434 unsigned long flags;
3435
3436 if (!iter)
3437 return;
3438
3439 cpu_buffer = iter->cpu_buffer;
3440
3441 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3442 arch_spin_lock(&cpu_buffer->lock);
3443 rb_iter_reset(iter);
3444 arch_spin_unlock(&cpu_buffer->lock);
3445 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3446}
3447EXPORT_SYMBOL_GPL(ring_buffer_read_start);
3448
3449/**
3450 * ring_buffer_finish - finish reading the iterator of the buffer
3451 * @iter: The iterator retrieved by ring_buffer_start
3452 *
3453 * This re-enables the recording to the buffer, and frees the
3454 * iterator.
3455 */
3456void
3457ring_buffer_read_finish(struct ring_buffer_iter *iter)
3458{
3459 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3460
3461 atomic_dec(&cpu_buffer->record_disabled);
3462 kfree(iter);
3463}
3464EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
3465
3466/**
3467 * ring_buffer_read - read the next item in the ring buffer by the iterator
3468 * @iter: The ring buffer iterator
3469 * @ts: The time stamp of the event read.
3470 *
3471 * This reads the next event in the ring buffer and increments the iterator.
3472 */
3473struct ring_buffer_event *
3474ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
3475{
3476 struct ring_buffer_event *event;
3477 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3478 unsigned long flags;
3479
3480 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3481 again:
3482 event = rb_iter_peek(iter, ts);
3483 if (!event)
3484 goto out;
3485
3486 if (event->type_len == RINGBUF_TYPE_PADDING)
3487 goto again;
3488
3489 rb_advance_iter(iter);
3490 out:
3491 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3492
3493 return event;
3494}
3495EXPORT_SYMBOL_GPL(ring_buffer_read);
3496
3497/**
3498 * ring_buffer_size - return the size of the ring buffer (in bytes)
3499 * @buffer: The ring buffer.
3500 */
3501unsigned long ring_buffer_size(struct ring_buffer *buffer)
3502{
3503 return BUF_PAGE_SIZE * buffer->pages;
3504}
3505EXPORT_SYMBOL_GPL(ring_buffer_size);
3506
3507static void
3508rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
3509{
3510 rb_head_page_deactivate(cpu_buffer);
3511
3512 cpu_buffer->head_page
3513 = list_entry(cpu_buffer->pages, struct buffer_page, list);
3514 local_set(&cpu_buffer->head_page->write, 0);
3515 local_set(&cpu_buffer->head_page->entries, 0);
3516 local_set(&cpu_buffer->head_page->page->commit, 0);
3517
3518 cpu_buffer->head_page->read = 0;
3519
3520 cpu_buffer->tail_page = cpu_buffer->head_page;
3521 cpu_buffer->commit_page = cpu_buffer->head_page;
3522
3523 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
3524 local_set(&cpu_buffer->reader_page->write, 0);
3525 local_set(&cpu_buffer->reader_page->entries, 0);
3526 local_set(&cpu_buffer->reader_page->page->commit, 0);
3527 cpu_buffer->reader_page->read = 0;
3528
3529 local_set(&cpu_buffer->commit_overrun, 0);
3530 local_set(&cpu_buffer->overrun, 0);
3531 local_set(&cpu_buffer->entries, 0);
3532 local_set(&cpu_buffer->committing, 0);
3533 local_set(&cpu_buffer->commits, 0);
3534 cpu_buffer->read = 0;
3535
3536 cpu_buffer->write_stamp = 0;
3537 cpu_buffer->read_stamp = 0;
3538
3539 cpu_buffer->lost_events = 0;
3540 cpu_buffer->last_overrun = 0;
3541
3542 rb_head_page_activate(cpu_buffer);
3543}
3544
3545/**
3546 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
3547 * @buffer: The ring buffer to reset a per cpu buffer of
3548 * @cpu: The CPU buffer to be reset
3549 */
3550void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
3551{
3552 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3553 unsigned long flags;
3554
3555 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3556 return;
3557
3558 atomic_inc(&cpu_buffer->record_disabled);
3559
3560 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3561
3562 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
3563 goto out;
3564
3565 arch_spin_lock(&cpu_buffer->lock);
3566
3567 rb_reset_cpu(cpu_buffer);
3568
3569 arch_spin_unlock(&cpu_buffer->lock);
3570
3571 out:
3572 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3573
3574 atomic_dec(&cpu_buffer->record_disabled);
3575}
3576EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
3577
3578/**
3579 * ring_buffer_reset - reset a ring buffer
3580 * @buffer: The ring buffer to reset all cpu buffers
3581 */
3582void ring_buffer_reset(struct ring_buffer *buffer)
3583{
3584 int cpu;
3585
3586 for_each_buffer_cpu(buffer, cpu)
3587 ring_buffer_reset_cpu(buffer, cpu);
3588}
3589EXPORT_SYMBOL_GPL(ring_buffer_reset);
3590
3591/**
3592 * rind_buffer_empty - is the ring buffer empty?
3593 * @buffer: The ring buffer to test
3594 */
3595int ring_buffer_empty(struct ring_buffer *buffer)
3596{
3597 struct ring_buffer_per_cpu *cpu_buffer;
3598 unsigned long flags;
3599 int dolock;
3600 int cpu;
3601 int ret;
3602
3603 dolock = rb_ok_to_lock();
3604
3605 /* yes this is racy, but if you don't like the race, lock the buffer */
3606 for_each_buffer_cpu(buffer, cpu) {
3607 cpu_buffer = buffer->buffers[cpu];
3608 local_irq_save(flags);
3609 if (dolock)
3610 spin_lock(&cpu_buffer->reader_lock);
3611 ret = rb_per_cpu_empty(cpu_buffer);
3612 if (dolock)
3613 spin_unlock(&cpu_buffer->reader_lock);
3614 local_irq_restore(flags);
3615
3616 if (!ret)
3617 return 0;
3618 }
3619
3620 return 1;
3621}
3622EXPORT_SYMBOL_GPL(ring_buffer_empty);
3623
3624/**
3625 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
3626 * @buffer: The ring buffer
3627 * @cpu: The CPU buffer to test
3628 */
3629int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
3630{
3631 struct ring_buffer_per_cpu *cpu_buffer;
3632 unsigned long flags;
3633 int dolock;
3634 int ret;
3635
3636 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3637 return 1;
3638
3639 dolock = rb_ok_to_lock();
3640
3641 cpu_buffer = buffer->buffers[cpu];
3642 local_irq_save(flags);
3643 if (dolock)
3644 spin_lock(&cpu_buffer->reader_lock);
3645 ret = rb_per_cpu_empty(cpu_buffer);
3646 if (dolock)
3647 spin_unlock(&cpu_buffer->reader_lock);
3648 local_irq_restore(flags);
3649
3650 return ret;
3651}
3652EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
3653
3654#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3655/**
3656 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
3657 * @buffer_a: One buffer to swap with
3658 * @buffer_b: The other buffer to swap with
3659 *
3660 * This function is useful for tracers that want to take a "snapshot"
3661 * of a CPU buffer and has another back up buffer lying around.
3662 * it is expected that the tracer handles the cpu buffer not being
3663 * used at the moment.
3664 */
3665int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
3666 struct ring_buffer *buffer_b, int cpu)
3667{
3668 struct ring_buffer_per_cpu *cpu_buffer_a;
3669 struct ring_buffer_per_cpu *cpu_buffer_b;
3670 int ret = -EINVAL;
3671
3672 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
3673 !cpumask_test_cpu(cpu, buffer_b->cpumask))
3674 goto out;
3675
3676 /* At least make sure the two buffers are somewhat the same */
3677 if (buffer_a->pages != buffer_b->pages)
3678 goto out;
3679
3680 ret = -EAGAIN;
3681
3682 if (ring_buffer_flags != RB_BUFFERS_ON)
3683 goto out;
3684
3685 if (atomic_read(&buffer_a->record_disabled))
3686 goto out;
3687
3688 if (atomic_read(&buffer_b->record_disabled))
3689 goto out;
3690
3691 cpu_buffer_a = buffer_a->buffers[cpu];
3692 cpu_buffer_b = buffer_b->buffers[cpu];
3693
3694 if (atomic_read(&cpu_buffer_a->record_disabled))
3695 goto out;
3696
3697 if (atomic_read(&cpu_buffer_b->record_disabled))
3698 goto out;
3699
3700 /*
3701 * We can't do a synchronize_sched here because this
3702 * function can be called in atomic context.
3703 * Normally this will be called from the same CPU as cpu.
3704 * If not it's up to the caller to protect this.
3705 */
3706 atomic_inc(&cpu_buffer_a->record_disabled);
3707 atomic_inc(&cpu_buffer_b->record_disabled);
3708
3709 ret = -EBUSY;
3710 if (local_read(&cpu_buffer_a->committing))
3711 goto out_dec;
3712 if (local_read(&cpu_buffer_b->committing))
3713 goto out_dec;
3714
3715 buffer_a->buffers[cpu] = cpu_buffer_b;
3716 buffer_b->buffers[cpu] = cpu_buffer_a;
3717
3718 cpu_buffer_b->buffer = buffer_a;
3719 cpu_buffer_a->buffer = buffer_b;
3720
3721 ret = 0;
3722
3723out_dec:
3724 atomic_dec(&cpu_buffer_a->record_disabled);
3725 atomic_dec(&cpu_buffer_b->record_disabled);
3726out:
3727 return ret;
3728}
3729EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
3730#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
3731
3732/**
3733 * ring_buffer_alloc_read_page - allocate a page to read from buffer
3734 * @buffer: the buffer to allocate for.
3735 *
3736 * This function is used in conjunction with ring_buffer_read_page.
3737 * When reading a full page from the ring buffer, these functions
3738 * can be used to speed up the process. The calling function should
3739 * allocate a few pages first with this function. Then when it
3740 * needs to get pages from the ring buffer, it passes the result
3741 * of this function into ring_buffer_read_page, which will swap
3742 * the page that was allocated, with the read page of the buffer.
3743 *
3744 * Returns:
3745 * The page allocated, or NULL on error.
3746 */
3747void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
3748{
3749 struct buffer_data_page *bpage;
3750 struct page *page;
3751
3752 page = alloc_pages_node(cpu_to_node(cpu),
3753 GFP_KERNEL | __GFP_NORETRY, 0);
3754 if (!page)
3755 return NULL;
3756
3757 bpage = page_address(page);
3758
3759 rb_init_page(bpage);
3760
3761 return bpage;
3762}
3763EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
3764
3765/**
3766 * ring_buffer_free_read_page - free an allocated read page
3767 * @buffer: the buffer the page was allocate for
3768 * @data: the page to free
3769 *
3770 * Free a page allocated from ring_buffer_alloc_read_page.
3771 */
3772void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
3773{
3774 free_page((unsigned long)data);
3775}
3776EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
3777
3778/**
3779 * ring_buffer_read_page - extract a page from the ring buffer
3780 * @buffer: buffer to extract from
3781 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
3782 * @len: amount to extract
3783 * @cpu: the cpu of the buffer to extract
3784 * @full: should the extraction only happen when the page is full.
3785 *
3786 * This function will pull out a page from the ring buffer and consume it.
3787 * @data_page must be the address of the variable that was returned
3788 * from ring_buffer_alloc_read_page. This is because the page might be used
3789 * to swap with a page in the ring buffer.
3790 *
3791 * for example:
3792 * rpage = ring_buffer_alloc_read_page(buffer);
3793 * if (!rpage)
3794 * return error;
3795 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
3796 * if (ret >= 0)
3797 * process_page(rpage, ret);
3798 *
3799 * When @full is set, the function will not return true unless
3800 * the writer is off the reader page.
3801 *
3802 * Note: it is up to the calling functions to handle sleeps and wakeups.
3803 * The ring buffer can be used anywhere in the kernel and can not
3804 * blindly call wake_up. The layer that uses the ring buffer must be
3805 * responsible for that.
3806 *
3807 * Returns:
3808 * >=0 if data has been transferred, returns the offset of consumed data.
3809 * <0 if no data has been transferred.
3810 */
3811int ring_buffer_read_page(struct ring_buffer *buffer,
3812 void **data_page, size_t len, int cpu, int full)
3813{
3814 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3815 struct ring_buffer_event *event;
3816 struct buffer_data_page *bpage;
3817 struct buffer_page *reader;
3818 unsigned long missed_events;
3819 unsigned long flags;
3820 unsigned int commit;
3821 unsigned int read;
3822 u64 save_timestamp;
3823 int ret = -1;
3824
3825 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3826 goto out;
3827
3828 /*
3829 * If len is not big enough to hold the page header, then
3830 * we can not copy anything.
3831 */
3832 if (len <= BUF_PAGE_HDR_SIZE)
3833 goto out;
3834
3835 len -= BUF_PAGE_HDR_SIZE;
3836
3837 if (!data_page)
3838 goto out;
3839
3840 bpage = *data_page;
3841 if (!bpage)
3842 goto out;
3843
3844 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3845
3846 reader = rb_get_reader_page(cpu_buffer);
3847 if (!reader)
3848 goto out_unlock;
3849
3850 event = rb_reader_event(cpu_buffer);
3851
3852 read = reader->read;
3853 commit = rb_page_commit(reader);
3854
3855 /* Check if any events were dropped */
3856 missed_events = cpu_buffer->lost_events;
3857
3858 /*
3859 * If this page has been partially read or
3860 * if len is not big enough to read the rest of the page or
3861 * a writer is still on the page, then
3862 * we must copy the data from the page to the buffer.
3863 * Otherwise, we can simply swap the page with the one passed in.
3864 */
3865 if (read || (len < (commit - read)) ||
3866 cpu_buffer->reader_page == cpu_buffer->commit_page) {
3867 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
3868 unsigned int rpos = read;
3869 unsigned int pos = 0;
3870 unsigned int size;
3871
3872 if (full)
3873 goto out_unlock;
3874
3875 if (len > (commit - read))
3876 len = (commit - read);
3877
3878 /* Always keep the time extend and data together */
3879 size = rb_event_ts_length(event);
3880
3881 if (len < size)
3882 goto out_unlock;
3883
3884 /* save the current timestamp, since the user will need it */
3885 save_timestamp = cpu_buffer->read_stamp;
3886
3887 /* Need to copy one event at a time */
3888 do {
3889 /* We need the size of one event, because
3890 * rb_advance_reader only advances by one event,
3891 * whereas rb_event_ts_length may include the size of
3892 * one or two events.
3893 * We have already ensured there's enough space if this
3894 * is a time extend. */
3895 size = rb_event_length(event);
3896 memcpy(bpage->data + pos, rpage->data + rpos, size);
3897
3898 len -= size;
3899
3900 rb_advance_reader(cpu_buffer);
3901 rpos = reader->read;
3902 pos += size;
3903
3904 if (rpos >= commit)
3905 break;
3906
3907 event = rb_reader_event(cpu_buffer);
3908 /* Always keep the time extend and data together */
3909 size = rb_event_ts_length(event);
3910 } while (len >= size);
3911
3912 /* update bpage */
3913 local_set(&bpage->commit, pos);
3914 bpage->time_stamp = save_timestamp;
3915
3916 /* we copied everything to the beginning */
3917 read = 0;
3918 } else {
3919 /* update the entry counter */
3920 cpu_buffer->read += rb_page_entries(reader);
3921
3922 /* swap the pages */
3923 rb_init_page(bpage);
3924 bpage = reader->page;
3925 reader->page = *data_page;
3926 local_set(&reader->write, 0);
3927 local_set(&reader->entries, 0);
3928 reader->read = 0;
3929 *data_page = bpage;
3930
3931 /*
3932 * Use the real_end for the data size,
3933 * This gives us a chance to store the lost events
3934 * on the page.
3935 */
3936 if (reader->real_end)
3937 local_set(&bpage->commit, reader->real_end);
3938 }
3939 ret = read;
3940
3941 cpu_buffer->lost_events = 0;
3942
3943 commit = local_read(&bpage->commit);
3944 /*
3945 * Set a flag in the commit field if we lost events
3946 */
3947 if (missed_events) {
3948 /* If there is room at the end of the page to save the
3949 * missed events, then record it there.
3950 */
3951 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
3952 memcpy(&bpage->data[commit], &missed_events,
3953 sizeof(missed_events));
3954 local_add(RB_MISSED_STORED, &bpage->commit);
3955 commit += sizeof(missed_events);
3956 }
3957 local_add(RB_MISSED_EVENTS, &bpage->commit);
3958 }
3959
3960 /*
3961 * This page may be off to user land. Zero it out here.
3962 */
3963 if (commit < BUF_PAGE_SIZE)
3964 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
3965
3966 out_unlock:
3967 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3968
3969 out:
3970 return ret;
3971}
3972EXPORT_SYMBOL_GPL(ring_buffer_read_page);
3973
3974#ifdef CONFIG_TRACING
3975static ssize_t
3976rb_simple_read(struct file *filp, char __user *ubuf,
3977 size_t cnt, loff_t *ppos)
3978{
3979 unsigned long *p = filp->private_data;
3980 char buf[64];
3981 int r;
3982
3983 if (test_bit(RB_BUFFERS_DISABLED_BIT, p))
3984 r = sprintf(buf, "permanently disabled\n");
3985 else
3986 r = sprintf(buf, "%d\n", test_bit(RB_BUFFERS_ON_BIT, p));
3987
3988 return simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
3989}
3990
3991static ssize_t
3992rb_simple_write(struct file *filp, const char __user *ubuf,
3993 size_t cnt, loff_t *ppos)
3994{
3995 unsigned long *p = filp->private_data;
3996 unsigned long val;
3997 int ret;
3998
3999 ret = kstrtoul_from_user(ubuf, cnt, 10, &val);
4000 if (ret)
4001 return ret;
4002
4003 if (val)
4004 set_bit(RB_BUFFERS_ON_BIT, p);
4005 else
4006 clear_bit(RB_BUFFERS_ON_BIT, p);
4007
4008 (*ppos)++;
4009
4010 return cnt;
4011}
4012
4013static const struct file_operations rb_simple_fops = {
4014 .open = tracing_open_generic,
4015 .read = rb_simple_read,
4016 .write = rb_simple_write,
4017 .llseek = default_llseek,
4018};
4019
4020
4021static __init int rb_init_debugfs(void)
4022{
4023 struct dentry *d_tracer;
4024
4025 d_tracer = tracing_init_dentry();
4026
4027 trace_create_file("tracing_on", 0644, d_tracer,
4028 &ring_buffer_flags, &rb_simple_fops);
4029
4030 return 0;
4031}
4032
4033fs_initcall(rb_init_debugfs);
4034#endif
4035
4036#ifdef CONFIG_HOTPLUG_CPU
4037static int rb_cpu_notify(struct notifier_block *self,
4038 unsigned long action, void *hcpu)
4039{
4040 struct ring_buffer *buffer =
4041 container_of(self, struct ring_buffer, cpu_notify);
4042 long cpu = (long)hcpu;
4043
4044 switch (action) {
4045 case CPU_UP_PREPARE:
4046 case CPU_UP_PREPARE_FROZEN:
4047 if (cpumask_test_cpu(cpu, buffer->cpumask))
4048 return NOTIFY_OK;
4049
4050 buffer->buffers[cpu] =
4051 rb_allocate_cpu_buffer(buffer, cpu);
4052 if (!buffer->buffers[cpu]) {
4053 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4054 cpu);
4055 return NOTIFY_OK;
4056 }
4057 smp_wmb();
4058 cpumask_set_cpu(cpu, buffer->cpumask);
4059 break;
4060 case CPU_DOWN_PREPARE:
4061 case CPU_DOWN_PREPARE_FROZEN:
4062 /*
4063 * Do nothing.
4064 * If we were to free the buffer, then the user would
4065 * lose any trace that was in the buffer.
4066 */
4067 break;
4068 default:
4069 break;
4070 }
4071 return NOTIFY_OK;
4072}
4073#endif
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/security.h>
15#include <linux/uaccess.h>
16#include <linux/hardirq.h>
17#include <linux/kthread.h> /* for self test */
18#include <linux/module.h>
19#include <linux/percpu.h>
20#include <linux/mutex.h>
21#include <linux/delay.h>
22#include <linux/slab.h>
23#include <linux/init.h>
24#include <linux/hash.h>
25#include <linux/list.h>
26#include <linux/cpu.h>
27#include <linux/oom.h>
28
29#include <asm/local.h>
30
31static void update_pages_handler(struct work_struct *work);
32
33/*
34 * The ring buffer header is special. We must manually up keep it.
35 */
36int ring_buffer_print_entry_header(struct trace_seq *s)
37{
38 trace_seq_puts(s, "# compressed entry header\n");
39 trace_seq_puts(s, "\ttype_len : 5 bits\n");
40 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
41 trace_seq_puts(s, "\tarray : 32 bits\n");
42 trace_seq_putc(s, '\n');
43 trace_seq_printf(s, "\tpadding : type == %d\n",
44 RINGBUF_TYPE_PADDING);
45 trace_seq_printf(s, "\ttime_extend : type == %d\n",
46 RINGBUF_TYPE_TIME_EXTEND);
47 trace_seq_printf(s, "\ttime_stamp : type == %d\n",
48 RINGBUF_TYPE_TIME_STAMP);
49 trace_seq_printf(s, "\tdata max type_len == %d\n",
50 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
51
52 return !trace_seq_has_overflowed(s);
53}
54
55/*
56 * The ring buffer is made up of a list of pages. A separate list of pages is
57 * allocated for each CPU. A writer may only write to a buffer that is
58 * associated with the CPU it is currently executing on. A reader may read
59 * from any per cpu buffer.
60 *
61 * The reader is special. For each per cpu buffer, the reader has its own
62 * reader page. When a reader has read the entire reader page, this reader
63 * page is swapped with another page in the ring buffer.
64 *
65 * Now, as long as the writer is off the reader page, the reader can do what
66 * ever it wants with that page. The writer will never write to that page
67 * again (as long as it is out of the ring buffer).
68 *
69 * Here's some silly ASCII art.
70 *
71 * +------+
72 * |reader| RING BUFFER
73 * |page |
74 * +------+ +---+ +---+ +---+
75 * | |-->| |-->| |
76 * +---+ +---+ +---+
77 * ^ |
78 * | |
79 * +---------------+
80 *
81 *
82 * +------+
83 * |reader| RING BUFFER
84 * |page |------------------v
85 * +------+ +---+ +---+ +---+
86 * | |-->| |-->| |
87 * +---+ +---+ +---+
88 * ^ |
89 * | |
90 * +---------------+
91 *
92 *
93 * +------+
94 * |reader| RING BUFFER
95 * |page |------------------v
96 * +------+ +---+ +---+ +---+
97 * ^ | |-->| |-->| |
98 * | +---+ +---+ +---+
99 * | |
100 * | |
101 * +------------------------------+
102 *
103 *
104 * +------+
105 * |buffer| RING BUFFER
106 * |page |------------------v
107 * +------+ +---+ +---+ +---+
108 * ^ | | | |-->| |
109 * | New +---+ +---+ +---+
110 * | Reader------^ |
111 * | page |
112 * +------------------------------+
113 *
114 *
115 * After we make this swap, the reader can hand this page off to the splice
116 * code and be done with it. It can even allocate a new page if it needs to
117 * and swap that into the ring buffer.
118 *
119 * We will be using cmpxchg soon to make all this lockless.
120 *
121 */
122
123/* Used for individual buffers (after the counter) */
124#define RB_BUFFER_OFF (1 << 20)
125
126#define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
127
128#define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
129#define RB_ALIGNMENT 4U
130#define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
131#define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
132#define RB_ALIGN_DATA __aligned(RB_ALIGNMENT)
133
134/* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
135#define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
136
137enum {
138 RB_LEN_TIME_EXTEND = 8,
139 RB_LEN_TIME_STAMP = 8,
140};
141
142#define skip_time_extend(event) \
143 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
144
145#define extended_time(event) \
146 (event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
147
148static inline int rb_null_event(struct ring_buffer_event *event)
149{
150 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
151}
152
153static void rb_event_set_padding(struct ring_buffer_event *event)
154{
155 /* padding has a NULL time_delta */
156 event->type_len = RINGBUF_TYPE_PADDING;
157 event->time_delta = 0;
158}
159
160static unsigned
161rb_event_data_length(struct ring_buffer_event *event)
162{
163 unsigned length;
164
165 if (event->type_len)
166 length = event->type_len * RB_ALIGNMENT;
167 else
168 length = event->array[0];
169 return length + RB_EVNT_HDR_SIZE;
170}
171
172/*
173 * Return the length of the given event. Will return
174 * the length of the time extend if the event is a
175 * time extend.
176 */
177static inline unsigned
178rb_event_length(struct ring_buffer_event *event)
179{
180 switch (event->type_len) {
181 case RINGBUF_TYPE_PADDING:
182 if (rb_null_event(event))
183 /* undefined */
184 return -1;
185 return event->array[0] + RB_EVNT_HDR_SIZE;
186
187 case RINGBUF_TYPE_TIME_EXTEND:
188 return RB_LEN_TIME_EXTEND;
189
190 case RINGBUF_TYPE_TIME_STAMP:
191 return RB_LEN_TIME_STAMP;
192
193 case RINGBUF_TYPE_DATA:
194 return rb_event_data_length(event);
195 default:
196 WARN_ON_ONCE(1);
197 }
198 /* not hit */
199 return 0;
200}
201
202/*
203 * Return total length of time extend and data,
204 * or just the event length for all other events.
205 */
206static inline unsigned
207rb_event_ts_length(struct ring_buffer_event *event)
208{
209 unsigned len = 0;
210
211 if (extended_time(event)) {
212 /* time extends include the data event after it */
213 len = RB_LEN_TIME_EXTEND;
214 event = skip_time_extend(event);
215 }
216 return len + rb_event_length(event);
217}
218
219/**
220 * ring_buffer_event_length - return the length of the event
221 * @event: the event to get the length of
222 *
223 * Returns the size of the data load of a data event.
224 * If the event is something other than a data event, it
225 * returns the size of the event itself. With the exception
226 * of a TIME EXTEND, where it still returns the size of the
227 * data load of the data event after it.
228 */
229unsigned ring_buffer_event_length(struct ring_buffer_event *event)
230{
231 unsigned length;
232
233 if (extended_time(event))
234 event = skip_time_extend(event);
235
236 length = rb_event_length(event);
237 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
238 return length;
239 length -= RB_EVNT_HDR_SIZE;
240 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
241 length -= sizeof(event->array[0]);
242 return length;
243}
244EXPORT_SYMBOL_GPL(ring_buffer_event_length);
245
246/* inline for ring buffer fast paths */
247static __always_inline void *
248rb_event_data(struct ring_buffer_event *event)
249{
250 if (extended_time(event))
251 event = skip_time_extend(event);
252 WARN_ON_ONCE(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
253 /* If length is in len field, then array[0] has the data */
254 if (event->type_len)
255 return (void *)&event->array[0];
256 /* Otherwise length is in array[0] and array[1] has the data */
257 return (void *)&event->array[1];
258}
259
260/**
261 * ring_buffer_event_data - return the data of the event
262 * @event: the event to get the data from
263 */
264void *ring_buffer_event_data(struct ring_buffer_event *event)
265{
266 return rb_event_data(event);
267}
268EXPORT_SYMBOL_GPL(ring_buffer_event_data);
269
270#define for_each_buffer_cpu(buffer, cpu) \
271 for_each_cpu(cpu, buffer->cpumask)
272
273#define for_each_online_buffer_cpu(buffer, cpu) \
274 for_each_cpu_and(cpu, buffer->cpumask, cpu_online_mask)
275
276#define TS_SHIFT 27
277#define TS_MASK ((1ULL << TS_SHIFT) - 1)
278#define TS_DELTA_TEST (~TS_MASK)
279
280/**
281 * ring_buffer_event_time_stamp - return the event's extended timestamp
282 * @event: the event to get the timestamp of
283 *
284 * Returns the extended timestamp associated with a data event.
285 * An extended time_stamp is a 64-bit timestamp represented
286 * internally in a special way that makes the best use of space
287 * contained within a ring buffer event. This function decodes
288 * it and maps it to a straight u64 value.
289 */
290u64 ring_buffer_event_time_stamp(struct ring_buffer_event *event)
291{
292 u64 ts;
293
294 ts = event->array[0];
295 ts <<= TS_SHIFT;
296 ts += event->time_delta;
297
298 return ts;
299}
300
301/* Flag when events were overwritten */
302#define RB_MISSED_EVENTS (1 << 31)
303/* Missed count stored at end */
304#define RB_MISSED_STORED (1 << 30)
305
306struct buffer_data_page {
307 u64 time_stamp; /* page time stamp */
308 local_t commit; /* write committed index */
309 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
310};
311
312/*
313 * Note, the buffer_page list must be first. The buffer pages
314 * are allocated in cache lines, which means that each buffer
315 * page will be at the beginning of a cache line, and thus
316 * the least significant bits will be zero. We use this to
317 * add flags in the list struct pointers, to make the ring buffer
318 * lockless.
319 */
320struct buffer_page {
321 struct list_head list; /* list of buffer pages */
322 local_t write; /* index for next write */
323 unsigned read; /* index for next read */
324 local_t entries; /* entries on this page */
325 unsigned long real_end; /* real end of data */
326 struct buffer_data_page *page; /* Actual data page */
327};
328
329/*
330 * The buffer page counters, write and entries, must be reset
331 * atomically when crossing page boundaries. To synchronize this
332 * update, two counters are inserted into the number. One is
333 * the actual counter for the write position or count on the page.
334 *
335 * The other is a counter of updaters. Before an update happens
336 * the update partition of the counter is incremented. This will
337 * allow the updater to update the counter atomically.
338 *
339 * The counter is 20 bits, and the state data is 12.
340 */
341#define RB_WRITE_MASK 0xfffff
342#define RB_WRITE_INTCNT (1 << 20)
343
344static void rb_init_page(struct buffer_data_page *bpage)
345{
346 local_set(&bpage->commit, 0);
347}
348
349/*
350 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
351 * this issue out.
352 */
353static void free_buffer_page(struct buffer_page *bpage)
354{
355 free_page((unsigned long)bpage->page);
356 kfree(bpage);
357}
358
359/*
360 * We need to fit the time_stamp delta into 27 bits.
361 */
362static inline int test_time_stamp(u64 delta)
363{
364 if (delta & TS_DELTA_TEST)
365 return 1;
366 return 0;
367}
368
369#define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
370
371/* Max payload is BUF_PAGE_SIZE - header (8bytes) */
372#define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
373
374int ring_buffer_print_page_header(struct trace_seq *s)
375{
376 struct buffer_data_page field;
377
378 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
379 "offset:0;\tsize:%u;\tsigned:%u;\n",
380 (unsigned int)sizeof(field.time_stamp),
381 (unsigned int)is_signed_type(u64));
382
383 trace_seq_printf(s, "\tfield: local_t commit;\t"
384 "offset:%u;\tsize:%u;\tsigned:%u;\n",
385 (unsigned int)offsetof(typeof(field), commit),
386 (unsigned int)sizeof(field.commit),
387 (unsigned int)is_signed_type(long));
388
389 trace_seq_printf(s, "\tfield: int overwrite;\t"
390 "offset:%u;\tsize:%u;\tsigned:%u;\n",
391 (unsigned int)offsetof(typeof(field), commit),
392 1,
393 (unsigned int)is_signed_type(long));
394
395 trace_seq_printf(s, "\tfield: char data;\t"
396 "offset:%u;\tsize:%u;\tsigned:%u;\n",
397 (unsigned int)offsetof(typeof(field), data),
398 (unsigned int)BUF_PAGE_SIZE,
399 (unsigned int)is_signed_type(char));
400
401 return !trace_seq_has_overflowed(s);
402}
403
404struct rb_irq_work {
405 struct irq_work work;
406 wait_queue_head_t waiters;
407 wait_queue_head_t full_waiters;
408 bool waiters_pending;
409 bool full_waiters_pending;
410 bool wakeup_full;
411};
412
413/*
414 * Structure to hold event state and handle nested events.
415 */
416struct rb_event_info {
417 u64 ts;
418 u64 delta;
419 u64 before;
420 u64 after;
421 unsigned long length;
422 struct buffer_page *tail_page;
423 int add_timestamp;
424};
425
426/*
427 * Used for the add_timestamp
428 * NONE
429 * EXTEND - wants a time extend
430 * ABSOLUTE - the buffer requests all events to have absolute time stamps
431 * FORCE - force a full time stamp.
432 */
433enum {
434 RB_ADD_STAMP_NONE = 0,
435 RB_ADD_STAMP_EXTEND = BIT(1),
436 RB_ADD_STAMP_ABSOLUTE = BIT(2),
437 RB_ADD_STAMP_FORCE = BIT(3)
438};
439/*
440 * Used for which event context the event is in.
441 * NMI = 0
442 * IRQ = 1
443 * SOFTIRQ = 2
444 * NORMAL = 3
445 *
446 * See trace_recursive_lock() comment below for more details.
447 */
448enum {
449 RB_CTX_NMI,
450 RB_CTX_IRQ,
451 RB_CTX_SOFTIRQ,
452 RB_CTX_NORMAL,
453 RB_CTX_MAX
454};
455
456#if BITS_PER_LONG == 32
457#define RB_TIME_32
458#endif
459
460/* To test on 64 bit machines */
461//#define RB_TIME_32
462
463#ifdef RB_TIME_32
464
465struct rb_time_struct {
466 local_t cnt;
467 local_t top;
468 local_t bottom;
469};
470#else
471#include <asm/local64.h>
472struct rb_time_struct {
473 local64_t time;
474};
475#endif
476typedef struct rb_time_struct rb_time_t;
477
478/*
479 * head_page == tail_page && head == tail then buffer is empty.
480 */
481struct ring_buffer_per_cpu {
482 int cpu;
483 atomic_t record_disabled;
484 atomic_t resize_disabled;
485 struct trace_buffer *buffer;
486 raw_spinlock_t reader_lock; /* serialize readers */
487 arch_spinlock_t lock;
488 struct lock_class_key lock_key;
489 struct buffer_data_page *free_page;
490 unsigned long nr_pages;
491 unsigned int current_context;
492 struct list_head *pages;
493 struct buffer_page *head_page; /* read from head */
494 struct buffer_page *tail_page; /* write to tail */
495 struct buffer_page *commit_page; /* committed pages */
496 struct buffer_page *reader_page;
497 unsigned long lost_events;
498 unsigned long last_overrun;
499 unsigned long nest;
500 local_t entries_bytes;
501 local_t entries;
502 local_t overrun;
503 local_t commit_overrun;
504 local_t dropped_events;
505 local_t committing;
506 local_t commits;
507 local_t pages_touched;
508 local_t pages_read;
509 long last_pages_touch;
510 size_t shortest_full;
511 unsigned long read;
512 unsigned long read_bytes;
513 rb_time_t write_stamp;
514 rb_time_t before_stamp;
515 u64 read_stamp;
516 /* ring buffer pages to update, > 0 to add, < 0 to remove */
517 long nr_pages_to_update;
518 struct list_head new_pages; /* new pages to add */
519 struct work_struct update_pages_work;
520 struct completion update_done;
521
522 struct rb_irq_work irq_work;
523};
524
525struct trace_buffer {
526 unsigned flags;
527 int cpus;
528 atomic_t record_disabled;
529 cpumask_var_t cpumask;
530
531 struct lock_class_key *reader_lock_key;
532
533 struct mutex mutex;
534
535 struct ring_buffer_per_cpu **buffers;
536
537 struct hlist_node node;
538 u64 (*clock)(void);
539
540 struct rb_irq_work irq_work;
541 bool time_stamp_abs;
542};
543
544struct ring_buffer_iter {
545 struct ring_buffer_per_cpu *cpu_buffer;
546 unsigned long head;
547 unsigned long next_event;
548 struct buffer_page *head_page;
549 struct buffer_page *cache_reader_page;
550 unsigned long cache_read;
551 u64 read_stamp;
552 u64 page_stamp;
553 struct ring_buffer_event *event;
554 int missed_events;
555};
556
557#ifdef RB_TIME_32
558
559/*
560 * On 32 bit machines, local64_t is very expensive. As the ring
561 * buffer doesn't need all the features of a true 64 bit atomic,
562 * on 32 bit, it uses these functions (64 still uses local64_t).
563 *
564 * For the ring buffer, 64 bit required operations for the time is
565 * the following:
566 *
567 * - Only need 59 bits (uses 60 to make it even).
568 * - Reads may fail if it interrupted a modification of the time stamp.
569 * It will succeed if it did not interrupt another write even if
570 * the read itself is interrupted by a write.
571 * It returns whether it was successful or not.
572 *
573 * - Writes always succeed and will overwrite other writes and writes
574 * that were done by events interrupting the current write.
575 *
576 * - A write followed by a read of the same time stamp will always succeed,
577 * but may not contain the same value.
578 *
579 * - A cmpxchg will fail if it interrupted another write or cmpxchg.
580 * Other than that, it acts like a normal cmpxchg.
581 *
582 * The 60 bit time stamp is broken up by 30 bits in a top and bottom half
583 * (bottom being the least significant 30 bits of the 60 bit time stamp).
584 *
585 * The two most significant bits of each half holds a 2 bit counter (0-3).
586 * Each update will increment this counter by one.
587 * When reading the top and bottom, if the two counter bits match then the
588 * top and bottom together make a valid 60 bit number.
589 */
590#define RB_TIME_SHIFT 30
591#define RB_TIME_VAL_MASK ((1 << RB_TIME_SHIFT) - 1)
592
593static inline int rb_time_cnt(unsigned long val)
594{
595 return (val >> RB_TIME_SHIFT) & 3;
596}
597
598static inline u64 rb_time_val(unsigned long top, unsigned long bottom)
599{
600 u64 val;
601
602 val = top & RB_TIME_VAL_MASK;
603 val <<= RB_TIME_SHIFT;
604 val |= bottom & RB_TIME_VAL_MASK;
605
606 return val;
607}
608
609static inline bool __rb_time_read(rb_time_t *t, u64 *ret, unsigned long *cnt)
610{
611 unsigned long top, bottom;
612 unsigned long c;
613
614 /*
615 * If the read is interrupted by a write, then the cnt will
616 * be different. Loop until both top and bottom have been read
617 * without interruption.
618 */
619 do {
620 c = local_read(&t->cnt);
621 top = local_read(&t->top);
622 bottom = local_read(&t->bottom);
623 } while (c != local_read(&t->cnt));
624
625 *cnt = rb_time_cnt(top);
626
627 /* If top and bottom counts don't match, this interrupted a write */
628 if (*cnt != rb_time_cnt(bottom))
629 return false;
630
631 *ret = rb_time_val(top, bottom);
632 return true;
633}
634
635static bool rb_time_read(rb_time_t *t, u64 *ret)
636{
637 unsigned long cnt;
638
639 return __rb_time_read(t, ret, &cnt);
640}
641
642static inline unsigned long rb_time_val_cnt(unsigned long val, unsigned long cnt)
643{
644 return (val & RB_TIME_VAL_MASK) | ((cnt & 3) << RB_TIME_SHIFT);
645}
646
647static inline void rb_time_split(u64 val, unsigned long *top, unsigned long *bottom)
648{
649 *top = (unsigned long)((val >> RB_TIME_SHIFT) & RB_TIME_VAL_MASK);
650 *bottom = (unsigned long)(val & RB_TIME_VAL_MASK);
651}
652
653static inline void rb_time_val_set(local_t *t, unsigned long val, unsigned long cnt)
654{
655 val = rb_time_val_cnt(val, cnt);
656 local_set(t, val);
657}
658
659static void rb_time_set(rb_time_t *t, u64 val)
660{
661 unsigned long cnt, top, bottom;
662
663 rb_time_split(val, &top, &bottom);
664
665 /* Writes always succeed with a valid number even if it gets interrupted. */
666 do {
667 cnt = local_inc_return(&t->cnt);
668 rb_time_val_set(&t->top, top, cnt);
669 rb_time_val_set(&t->bottom, bottom, cnt);
670 } while (cnt != local_read(&t->cnt));
671}
672
673static inline bool
674rb_time_read_cmpxchg(local_t *l, unsigned long expect, unsigned long set)
675{
676 unsigned long ret;
677
678 ret = local_cmpxchg(l, expect, set);
679 return ret == expect;
680}
681
682static int rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
683{
684 unsigned long cnt, top, bottom;
685 unsigned long cnt2, top2, bottom2;
686 u64 val;
687
688 /* The cmpxchg always fails if it interrupted an update */
689 if (!__rb_time_read(t, &val, &cnt2))
690 return false;
691
692 if (val != expect)
693 return false;
694
695 cnt = local_read(&t->cnt);
696 if ((cnt & 3) != cnt2)
697 return false;
698
699 cnt2 = cnt + 1;
700
701 rb_time_split(val, &top, &bottom);
702 top = rb_time_val_cnt(top, cnt);
703 bottom = rb_time_val_cnt(bottom, cnt);
704
705 rb_time_split(set, &top2, &bottom2);
706 top2 = rb_time_val_cnt(top2, cnt2);
707 bottom2 = rb_time_val_cnt(bottom2, cnt2);
708
709 if (!rb_time_read_cmpxchg(&t->cnt, cnt, cnt2))
710 return false;
711 if (!rb_time_read_cmpxchg(&t->top, top, top2))
712 return false;
713 if (!rb_time_read_cmpxchg(&t->bottom, bottom, bottom2))
714 return false;
715 return true;
716}
717
718#else /* 64 bits */
719
720/* local64_t always succeeds */
721
722static inline bool rb_time_read(rb_time_t *t, u64 *ret)
723{
724 *ret = local64_read(&t->time);
725 return true;
726}
727static void rb_time_set(rb_time_t *t, u64 val)
728{
729 local64_set(&t->time, val);
730}
731
732static bool rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
733{
734 u64 val;
735 val = local64_cmpxchg(&t->time, expect, set);
736 return val == expect;
737}
738#endif
739
740/**
741 * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
742 * @buffer: The ring_buffer to get the number of pages from
743 * @cpu: The cpu of the ring_buffer to get the number of pages from
744 *
745 * Returns the number of pages used by a per_cpu buffer of the ring buffer.
746 */
747size_t ring_buffer_nr_pages(struct trace_buffer *buffer, int cpu)
748{
749 return buffer->buffers[cpu]->nr_pages;
750}
751
752/**
753 * ring_buffer_nr_pages_dirty - get the number of used pages in the ring buffer
754 * @buffer: The ring_buffer to get the number of pages from
755 * @cpu: The cpu of the ring_buffer to get the number of pages from
756 *
757 * Returns the number of pages that have content in the ring buffer.
758 */
759size_t ring_buffer_nr_dirty_pages(struct trace_buffer *buffer, int cpu)
760{
761 size_t read;
762 size_t cnt;
763
764 read = local_read(&buffer->buffers[cpu]->pages_read);
765 cnt = local_read(&buffer->buffers[cpu]->pages_touched);
766 /* The reader can read an empty page, but not more than that */
767 if (cnt < read) {
768 WARN_ON_ONCE(read > cnt + 1);
769 return 0;
770 }
771
772 return cnt - read;
773}
774
775/*
776 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
777 *
778 * Schedules a delayed work to wake up any task that is blocked on the
779 * ring buffer waiters queue.
780 */
781static void rb_wake_up_waiters(struct irq_work *work)
782{
783 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
784
785 wake_up_all(&rbwork->waiters);
786 if (rbwork->wakeup_full) {
787 rbwork->wakeup_full = false;
788 wake_up_all(&rbwork->full_waiters);
789 }
790}
791
792/**
793 * ring_buffer_wait - wait for input to the ring buffer
794 * @buffer: buffer to wait on
795 * @cpu: the cpu buffer to wait on
796 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
797 *
798 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
799 * as data is added to any of the @buffer's cpu buffers. Otherwise
800 * it will wait for data to be added to a specific cpu buffer.
801 */
802int ring_buffer_wait(struct trace_buffer *buffer, int cpu, int full)
803{
804 struct ring_buffer_per_cpu *cpu_buffer;
805 DEFINE_WAIT(wait);
806 struct rb_irq_work *work;
807 int ret = 0;
808
809 /*
810 * Depending on what the caller is waiting for, either any
811 * data in any cpu buffer, or a specific buffer, put the
812 * caller on the appropriate wait queue.
813 */
814 if (cpu == RING_BUFFER_ALL_CPUS) {
815 work = &buffer->irq_work;
816 /* Full only makes sense on per cpu reads */
817 full = 0;
818 } else {
819 if (!cpumask_test_cpu(cpu, buffer->cpumask))
820 return -ENODEV;
821 cpu_buffer = buffer->buffers[cpu];
822 work = &cpu_buffer->irq_work;
823 }
824
825
826 while (true) {
827 if (full)
828 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
829 else
830 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
831
832 /*
833 * The events can happen in critical sections where
834 * checking a work queue can cause deadlocks.
835 * After adding a task to the queue, this flag is set
836 * only to notify events to try to wake up the queue
837 * using irq_work.
838 *
839 * We don't clear it even if the buffer is no longer
840 * empty. The flag only causes the next event to run
841 * irq_work to do the work queue wake up. The worse
842 * that can happen if we race with !trace_empty() is that
843 * an event will cause an irq_work to try to wake up
844 * an empty queue.
845 *
846 * There's no reason to protect this flag either, as
847 * the work queue and irq_work logic will do the necessary
848 * synchronization for the wake ups. The only thing
849 * that is necessary is that the wake up happens after
850 * a task has been queued. It's OK for spurious wake ups.
851 */
852 if (full)
853 work->full_waiters_pending = true;
854 else
855 work->waiters_pending = true;
856
857 if (signal_pending(current)) {
858 ret = -EINTR;
859 break;
860 }
861
862 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
863 break;
864
865 if (cpu != RING_BUFFER_ALL_CPUS &&
866 !ring_buffer_empty_cpu(buffer, cpu)) {
867 unsigned long flags;
868 bool pagebusy;
869 size_t nr_pages;
870 size_t dirty;
871
872 if (!full)
873 break;
874
875 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
876 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
877 nr_pages = cpu_buffer->nr_pages;
878 dirty = ring_buffer_nr_dirty_pages(buffer, cpu);
879 if (!cpu_buffer->shortest_full ||
880 cpu_buffer->shortest_full < full)
881 cpu_buffer->shortest_full = full;
882 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
883 if (!pagebusy &&
884 (!nr_pages || (dirty * 100) > full * nr_pages))
885 break;
886 }
887
888 schedule();
889 }
890
891 if (full)
892 finish_wait(&work->full_waiters, &wait);
893 else
894 finish_wait(&work->waiters, &wait);
895
896 return ret;
897}
898
899/**
900 * ring_buffer_poll_wait - poll on buffer input
901 * @buffer: buffer to wait on
902 * @cpu: the cpu buffer to wait on
903 * @filp: the file descriptor
904 * @poll_table: The poll descriptor
905 *
906 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
907 * as data is added to any of the @buffer's cpu buffers. Otherwise
908 * it will wait for data to be added to a specific cpu buffer.
909 *
910 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
911 * zero otherwise.
912 */
913__poll_t ring_buffer_poll_wait(struct trace_buffer *buffer, int cpu,
914 struct file *filp, poll_table *poll_table)
915{
916 struct ring_buffer_per_cpu *cpu_buffer;
917 struct rb_irq_work *work;
918
919 if (cpu == RING_BUFFER_ALL_CPUS)
920 work = &buffer->irq_work;
921 else {
922 if (!cpumask_test_cpu(cpu, buffer->cpumask))
923 return -EINVAL;
924
925 cpu_buffer = buffer->buffers[cpu];
926 work = &cpu_buffer->irq_work;
927 }
928
929 poll_wait(filp, &work->waiters, poll_table);
930 work->waiters_pending = true;
931 /*
932 * There's a tight race between setting the waiters_pending and
933 * checking if the ring buffer is empty. Once the waiters_pending bit
934 * is set, the next event will wake the task up, but we can get stuck
935 * if there's only a single event in.
936 *
937 * FIXME: Ideally, we need a memory barrier on the writer side as well,
938 * but adding a memory barrier to all events will cause too much of a
939 * performance hit in the fast path. We only need a memory barrier when
940 * the buffer goes from empty to having content. But as this race is
941 * extremely small, and it's not a problem if another event comes in, we
942 * will fix it later.
943 */
944 smp_mb();
945
946 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
947 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
948 return EPOLLIN | EPOLLRDNORM;
949 return 0;
950}
951
952/* buffer may be either ring_buffer or ring_buffer_per_cpu */
953#define RB_WARN_ON(b, cond) \
954 ({ \
955 int _____ret = unlikely(cond); \
956 if (_____ret) { \
957 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
958 struct ring_buffer_per_cpu *__b = \
959 (void *)b; \
960 atomic_inc(&__b->buffer->record_disabled); \
961 } else \
962 atomic_inc(&b->record_disabled); \
963 WARN_ON(1); \
964 } \
965 _____ret; \
966 })
967
968/* Up this if you want to test the TIME_EXTENTS and normalization */
969#define DEBUG_SHIFT 0
970
971static inline u64 rb_time_stamp(struct trace_buffer *buffer)
972{
973 u64 ts;
974
975 /* Skip retpolines :-( */
976 if (IS_ENABLED(CONFIG_RETPOLINE) && likely(buffer->clock == trace_clock_local))
977 ts = trace_clock_local();
978 else
979 ts = buffer->clock();
980
981 /* shift to debug/test normalization and TIME_EXTENTS */
982 return ts << DEBUG_SHIFT;
983}
984
985u64 ring_buffer_time_stamp(struct trace_buffer *buffer, int cpu)
986{
987 u64 time;
988
989 preempt_disable_notrace();
990 time = rb_time_stamp(buffer);
991 preempt_enable_notrace();
992
993 return time;
994}
995EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
996
997void ring_buffer_normalize_time_stamp(struct trace_buffer *buffer,
998 int cpu, u64 *ts)
999{
1000 /* Just stupid testing the normalize function and deltas */
1001 *ts >>= DEBUG_SHIFT;
1002}
1003EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
1004
1005/*
1006 * Making the ring buffer lockless makes things tricky.
1007 * Although writes only happen on the CPU that they are on,
1008 * and they only need to worry about interrupts. Reads can
1009 * happen on any CPU.
1010 *
1011 * The reader page is always off the ring buffer, but when the
1012 * reader finishes with a page, it needs to swap its page with
1013 * a new one from the buffer. The reader needs to take from
1014 * the head (writes go to the tail). But if a writer is in overwrite
1015 * mode and wraps, it must push the head page forward.
1016 *
1017 * Here lies the problem.
1018 *
1019 * The reader must be careful to replace only the head page, and
1020 * not another one. As described at the top of the file in the
1021 * ASCII art, the reader sets its old page to point to the next
1022 * page after head. It then sets the page after head to point to
1023 * the old reader page. But if the writer moves the head page
1024 * during this operation, the reader could end up with the tail.
1025 *
1026 * We use cmpxchg to help prevent this race. We also do something
1027 * special with the page before head. We set the LSB to 1.
1028 *
1029 * When the writer must push the page forward, it will clear the
1030 * bit that points to the head page, move the head, and then set
1031 * the bit that points to the new head page.
1032 *
1033 * We also don't want an interrupt coming in and moving the head
1034 * page on another writer. Thus we use the second LSB to catch
1035 * that too. Thus:
1036 *
1037 * head->list->prev->next bit 1 bit 0
1038 * ------- -------
1039 * Normal page 0 0
1040 * Points to head page 0 1
1041 * New head page 1 0
1042 *
1043 * Note we can not trust the prev pointer of the head page, because:
1044 *
1045 * +----+ +-----+ +-----+
1046 * | |------>| T |---X--->| N |
1047 * | |<------| | | |
1048 * +----+ +-----+ +-----+
1049 * ^ ^ |
1050 * | +-----+ | |
1051 * +----------| R |----------+ |
1052 * | |<-----------+
1053 * +-----+
1054 *
1055 * Key: ---X--> HEAD flag set in pointer
1056 * T Tail page
1057 * R Reader page
1058 * N Next page
1059 *
1060 * (see __rb_reserve_next() to see where this happens)
1061 *
1062 * What the above shows is that the reader just swapped out
1063 * the reader page with a page in the buffer, but before it
1064 * could make the new header point back to the new page added
1065 * it was preempted by a writer. The writer moved forward onto
1066 * the new page added by the reader and is about to move forward
1067 * again.
1068 *
1069 * You can see, it is legitimate for the previous pointer of
1070 * the head (or any page) not to point back to itself. But only
1071 * temporarily.
1072 */
1073
1074#define RB_PAGE_NORMAL 0UL
1075#define RB_PAGE_HEAD 1UL
1076#define RB_PAGE_UPDATE 2UL
1077
1078
1079#define RB_FLAG_MASK 3UL
1080
1081/* PAGE_MOVED is not part of the mask */
1082#define RB_PAGE_MOVED 4UL
1083
1084/*
1085 * rb_list_head - remove any bit
1086 */
1087static struct list_head *rb_list_head(struct list_head *list)
1088{
1089 unsigned long val = (unsigned long)list;
1090
1091 return (struct list_head *)(val & ~RB_FLAG_MASK);
1092}
1093
1094/*
1095 * rb_is_head_page - test if the given page is the head page
1096 *
1097 * Because the reader may move the head_page pointer, we can
1098 * not trust what the head page is (it may be pointing to
1099 * the reader page). But if the next page is a header page,
1100 * its flags will be non zero.
1101 */
1102static inline int
1103rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1104 struct buffer_page *page, struct list_head *list)
1105{
1106 unsigned long val;
1107
1108 val = (unsigned long)list->next;
1109
1110 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
1111 return RB_PAGE_MOVED;
1112
1113 return val & RB_FLAG_MASK;
1114}
1115
1116/*
1117 * rb_is_reader_page
1118 *
1119 * The unique thing about the reader page, is that, if the
1120 * writer is ever on it, the previous pointer never points
1121 * back to the reader page.
1122 */
1123static bool rb_is_reader_page(struct buffer_page *page)
1124{
1125 struct list_head *list = page->list.prev;
1126
1127 return rb_list_head(list->next) != &page->list;
1128}
1129
1130/*
1131 * rb_set_list_to_head - set a list_head to be pointing to head.
1132 */
1133static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
1134 struct list_head *list)
1135{
1136 unsigned long *ptr;
1137
1138 ptr = (unsigned long *)&list->next;
1139 *ptr |= RB_PAGE_HEAD;
1140 *ptr &= ~RB_PAGE_UPDATE;
1141}
1142
1143/*
1144 * rb_head_page_activate - sets up head page
1145 */
1146static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
1147{
1148 struct buffer_page *head;
1149
1150 head = cpu_buffer->head_page;
1151 if (!head)
1152 return;
1153
1154 /*
1155 * Set the previous list pointer to have the HEAD flag.
1156 */
1157 rb_set_list_to_head(cpu_buffer, head->list.prev);
1158}
1159
1160static void rb_list_head_clear(struct list_head *list)
1161{
1162 unsigned long *ptr = (unsigned long *)&list->next;
1163
1164 *ptr &= ~RB_FLAG_MASK;
1165}
1166
1167/*
1168 * rb_head_page_deactivate - clears head page ptr (for free list)
1169 */
1170static void
1171rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
1172{
1173 struct list_head *hd;
1174
1175 /* Go through the whole list and clear any pointers found. */
1176 rb_list_head_clear(cpu_buffer->pages);
1177
1178 list_for_each(hd, cpu_buffer->pages)
1179 rb_list_head_clear(hd);
1180}
1181
1182static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
1183 struct buffer_page *head,
1184 struct buffer_page *prev,
1185 int old_flag, int new_flag)
1186{
1187 struct list_head *list;
1188 unsigned long val = (unsigned long)&head->list;
1189 unsigned long ret;
1190
1191 list = &prev->list;
1192
1193 val &= ~RB_FLAG_MASK;
1194
1195 ret = cmpxchg((unsigned long *)&list->next,
1196 val | old_flag, val | new_flag);
1197
1198 /* check if the reader took the page */
1199 if ((ret & ~RB_FLAG_MASK) != val)
1200 return RB_PAGE_MOVED;
1201
1202 return ret & RB_FLAG_MASK;
1203}
1204
1205static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
1206 struct buffer_page *head,
1207 struct buffer_page *prev,
1208 int old_flag)
1209{
1210 return rb_head_page_set(cpu_buffer, head, prev,
1211 old_flag, RB_PAGE_UPDATE);
1212}
1213
1214static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
1215 struct buffer_page *head,
1216 struct buffer_page *prev,
1217 int old_flag)
1218{
1219 return rb_head_page_set(cpu_buffer, head, prev,
1220 old_flag, RB_PAGE_HEAD);
1221}
1222
1223static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
1224 struct buffer_page *head,
1225 struct buffer_page *prev,
1226 int old_flag)
1227{
1228 return rb_head_page_set(cpu_buffer, head, prev,
1229 old_flag, RB_PAGE_NORMAL);
1230}
1231
1232static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
1233 struct buffer_page **bpage)
1234{
1235 struct list_head *p = rb_list_head((*bpage)->list.next);
1236
1237 *bpage = list_entry(p, struct buffer_page, list);
1238}
1239
1240static struct buffer_page *
1241rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1242{
1243 struct buffer_page *head;
1244 struct buffer_page *page;
1245 struct list_head *list;
1246 int i;
1247
1248 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1249 return NULL;
1250
1251 /* sanity check */
1252 list = cpu_buffer->pages;
1253 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1254 return NULL;
1255
1256 page = head = cpu_buffer->head_page;
1257 /*
1258 * It is possible that the writer moves the header behind
1259 * where we started, and we miss in one loop.
1260 * A second loop should grab the header, but we'll do
1261 * three loops just because I'm paranoid.
1262 */
1263 for (i = 0; i < 3; i++) {
1264 do {
1265 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
1266 cpu_buffer->head_page = page;
1267 return page;
1268 }
1269 rb_inc_page(cpu_buffer, &page);
1270 } while (page != head);
1271 }
1272
1273 RB_WARN_ON(cpu_buffer, 1);
1274
1275 return NULL;
1276}
1277
1278static int rb_head_page_replace(struct buffer_page *old,
1279 struct buffer_page *new)
1280{
1281 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1282 unsigned long val;
1283 unsigned long ret;
1284
1285 val = *ptr & ~RB_FLAG_MASK;
1286 val |= RB_PAGE_HEAD;
1287
1288 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1289
1290 return ret == val;
1291}
1292
1293/*
1294 * rb_tail_page_update - move the tail page forward
1295 */
1296static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1297 struct buffer_page *tail_page,
1298 struct buffer_page *next_page)
1299{
1300 unsigned long old_entries;
1301 unsigned long old_write;
1302
1303 /*
1304 * The tail page now needs to be moved forward.
1305 *
1306 * We need to reset the tail page, but without messing
1307 * with possible erasing of data brought in by interrupts
1308 * that have moved the tail page and are currently on it.
1309 *
1310 * We add a counter to the write field to denote this.
1311 */
1312 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1313 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1314
1315 local_inc(&cpu_buffer->pages_touched);
1316 /*
1317 * Just make sure we have seen our old_write and synchronize
1318 * with any interrupts that come in.
1319 */
1320 barrier();
1321
1322 /*
1323 * If the tail page is still the same as what we think
1324 * it is, then it is up to us to update the tail
1325 * pointer.
1326 */
1327 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1328 /* Zero the write counter */
1329 unsigned long val = old_write & ~RB_WRITE_MASK;
1330 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1331
1332 /*
1333 * This will only succeed if an interrupt did
1334 * not come in and change it. In which case, we
1335 * do not want to modify it.
1336 *
1337 * We add (void) to let the compiler know that we do not care
1338 * about the return value of these functions. We use the
1339 * cmpxchg to only update if an interrupt did not already
1340 * do it for us. If the cmpxchg fails, we don't care.
1341 */
1342 (void)local_cmpxchg(&next_page->write, old_write, val);
1343 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1344
1345 /*
1346 * No need to worry about races with clearing out the commit.
1347 * it only can increment when a commit takes place. But that
1348 * only happens in the outer most nested commit.
1349 */
1350 local_set(&next_page->page->commit, 0);
1351
1352 /* Again, either we update tail_page or an interrupt does */
1353 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1354 }
1355}
1356
1357static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1358 struct buffer_page *bpage)
1359{
1360 unsigned long val = (unsigned long)bpage;
1361
1362 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1363 return 1;
1364
1365 return 0;
1366}
1367
1368/**
1369 * rb_check_list - make sure a pointer to a list has the last bits zero
1370 */
1371static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1372 struct list_head *list)
1373{
1374 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1375 return 1;
1376 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1377 return 1;
1378 return 0;
1379}
1380
1381/**
1382 * rb_check_pages - integrity check of buffer pages
1383 * @cpu_buffer: CPU buffer with pages to test
1384 *
1385 * As a safety measure we check to make sure the data pages have not
1386 * been corrupted.
1387 */
1388static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1389{
1390 struct list_head *head = cpu_buffer->pages;
1391 struct buffer_page *bpage, *tmp;
1392
1393 /* Reset the head page if it exists */
1394 if (cpu_buffer->head_page)
1395 rb_set_head_page(cpu_buffer);
1396
1397 rb_head_page_deactivate(cpu_buffer);
1398
1399 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1400 return -1;
1401 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1402 return -1;
1403
1404 if (rb_check_list(cpu_buffer, head))
1405 return -1;
1406
1407 list_for_each_entry_safe(bpage, tmp, head, list) {
1408 if (RB_WARN_ON(cpu_buffer,
1409 bpage->list.next->prev != &bpage->list))
1410 return -1;
1411 if (RB_WARN_ON(cpu_buffer,
1412 bpage->list.prev->next != &bpage->list))
1413 return -1;
1414 if (rb_check_list(cpu_buffer, &bpage->list))
1415 return -1;
1416 }
1417
1418 rb_head_page_activate(cpu_buffer);
1419
1420 return 0;
1421}
1422
1423static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1424{
1425 struct buffer_page *bpage, *tmp;
1426 bool user_thread = current->mm != NULL;
1427 gfp_t mflags;
1428 long i;
1429
1430 /*
1431 * Check if the available memory is there first.
1432 * Note, si_mem_available() only gives us a rough estimate of available
1433 * memory. It may not be accurate. But we don't care, we just want
1434 * to prevent doing any allocation when it is obvious that it is
1435 * not going to succeed.
1436 */
1437 i = si_mem_available();
1438 if (i < nr_pages)
1439 return -ENOMEM;
1440
1441 /*
1442 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1443 * gracefully without invoking oom-killer and the system is not
1444 * destabilized.
1445 */
1446 mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1447
1448 /*
1449 * If a user thread allocates too much, and si_mem_available()
1450 * reports there's enough memory, even though there is not.
1451 * Make sure the OOM killer kills this thread. This can happen
1452 * even with RETRY_MAYFAIL because another task may be doing
1453 * an allocation after this task has taken all memory.
1454 * This is the task the OOM killer needs to take out during this
1455 * loop, even if it was triggered by an allocation somewhere else.
1456 */
1457 if (user_thread)
1458 set_current_oom_origin();
1459 for (i = 0; i < nr_pages; i++) {
1460 struct page *page;
1461
1462 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1463 mflags, cpu_to_node(cpu));
1464 if (!bpage)
1465 goto free_pages;
1466
1467 list_add(&bpage->list, pages);
1468
1469 page = alloc_pages_node(cpu_to_node(cpu), mflags, 0);
1470 if (!page)
1471 goto free_pages;
1472 bpage->page = page_address(page);
1473 rb_init_page(bpage->page);
1474
1475 if (user_thread && fatal_signal_pending(current))
1476 goto free_pages;
1477 }
1478 if (user_thread)
1479 clear_current_oom_origin();
1480
1481 return 0;
1482
1483free_pages:
1484 list_for_each_entry_safe(bpage, tmp, pages, list) {
1485 list_del_init(&bpage->list);
1486 free_buffer_page(bpage);
1487 }
1488 if (user_thread)
1489 clear_current_oom_origin();
1490
1491 return -ENOMEM;
1492}
1493
1494static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1495 unsigned long nr_pages)
1496{
1497 LIST_HEAD(pages);
1498
1499 WARN_ON(!nr_pages);
1500
1501 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1502 return -ENOMEM;
1503
1504 /*
1505 * The ring buffer page list is a circular list that does not
1506 * start and end with a list head. All page list items point to
1507 * other pages.
1508 */
1509 cpu_buffer->pages = pages.next;
1510 list_del(&pages);
1511
1512 cpu_buffer->nr_pages = nr_pages;
1513
1514 rb_check_pages(cpu_buffer);
1515
1516 return 0;
1517}
1518
1519static struct ring_buffer_per_cpu *
1520rb_allocate_cpu_buffer(struct trace_buffer *buffer, long nr_pages, int cpu)
1521{
1522 struct ring_buffer_per_cpu *cpu_buffer;
1523 struct buffer_page *bpage;
1524 struct page *page;
1525 int ret;
1526
1527 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1528 GFP_KERNEL, cpu_to_node(cpu));
1529 if (!cpu_buffer)
1530 return NULL;
1531
1532 cpu_buffer->cpu = cpu;
1533 cpu_buffer->buffer = buffer;
1534 raw_spin_lock_init(&cpu_buffer->reader_lock);
1535 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1536 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1537 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1538 init_completion(&cpu_buffer->update_done);
1539 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1540 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1541 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1542
1543 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1544 GFP_KERNEL, cpu_to_node(cpu));
1545 if (!bpage)
1546 goto fail_free_buffer;
1547
1548 rb_check_bpage(cpu_buffer, bpage);
1549
1550 cpu_buffer->reader_page = bpage;
1551 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1552 if (!page)
1553 goto fail_free_reader;
1554 bpage->page = page_address(page);
1555 rb_init_page(bpage->page);
1556
1557 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1558 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1559
1560 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1561 if (ret < 0)
1562 goto fail_free_reader;
1563
1564 cpu_buffer->head_page
1565 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1566 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1567
1568 rb_head_page_activate(cpu_buffer);
1569
1570 return cpu_buffer;
1571
1572 fail_free_reader:
1573 free_buffer_page(cpu_buffer->reader_page);
1574
1575 fail_free_buffer:
1576 kfree(cpu_buffer);
1577 return NULL;
1578}
1579
1580static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1581{
1582 struct list_head *head = cpu_buffer->pages;
1583 struct buffer_page *bpage, *tmp;
1584
1585 free_buffer_page(cpu_buffer->reader_page);
1586
1587 rb_head_page_deactivate(cpu_buffer);
1588
1589 if (head) {
1590 list_for_each_entry_safe(bpage, tmp, head, list) {
1591 list_del_init(&bpage->list);
1592 free_buffer_page(bpage);
1593 }
1594 bpage = list_entry(head, struct buffer_page, list);
1595 free_buffer_page(bpage);
1596 }
1597
1598 kfree(cpu_buffer);
1599}
1600
1601/**
1602 * __ring_buffer_alloc - allocate a new ring_buffer
1603 * @size: the size in bytes per cpu that is needed.
1604 * @flags: attributes to set for the ring buffer.
1605 * @key: ring buffer reader_lock_key.
1606 *
1607 * Currently the only flag that is available is the RB_FL_OVERWRITE
1608 * flag. This flag means that the buffer will overwrite old data
1609 * when the buffer wraps. If this flag is not set, the buffer will
1610 * drop data when the tail hits the head.
1611 */
1612struct trace_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1613 struct lock_class_key *key)
1614{
1615 struct trace_buffer *buffer;
1616 long nr_pages;
1617 int bsize;
1618 int cpu;
1619 int ret;
1620
1621 /* keep it in its own cache line */
1622 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1623 GFP_KERNEL);
1624 if (!buffer)
1625 return NULL;
1626
1627 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1628 goto fail_free_buffer;
1629
1630 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1631 buffer->flags = flags;
1632 buffer->clock = trace_clock_local;
1633 buffer->reader_lock_key = key;
1634
1635 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1636 init_waitqueue_head(&buffer->irq_work.waiters);
1637
1638 /* need at least two pages */
1639 if (nr_pages < 2)
1640 nr_pages = 2;
1641
1642 buffer->cpus = nr_cpu_ids;
1643
1644 bsize = sizeof(void *) * nr_cpu_ids;
1645 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1646 GFP_KERNEL);
1647 if (!buffer->buffers)
1648 goto fail_free_cpumask;
1649
1650 cpu = raw_smp_processor_id();
1651 cpumask_set_cpu(cpu, buffer->cpumask);
1652 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1653 if (!buffer->buffers[cpu])
1654 goto fail_free_buffers;
1655
1656 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1657 if (ret < 0)
1658 goto fail_free_buffers;
1659
1660 mutex_init(&buffer->mutex);
1661
1662 return buffer;
1663
1664 fail_free_buffers:
1665 for_each_buffer_cpu(buffer, cpu) {
1666 if (buffer->buffers[cpu])
1667 rb_free_cpu_buffer(buffer->buffers[cpu]);
1668 }
1669 kfree(buffer->buffers);
1670
1671 fail_free_cpumask:
1672 free_cpumask_var(buffer->cpumask);
1673
1674 fail_free_buffer:
1675 kfree(buffer);
1676 return NULL;
1677}
1678EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1679
1680/**
1681 * ring_buffer_free - free a ring buffer.
1682 * @buffer: the buffer to free.
1683 */
1684void
1685ring_buffer_free(struct trace_buffer *buffer)
1686{
1687 int cpu;
1688
1689 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1690
1691 for_each_buffer_cpu(buffer, cpu)
1692 rb_free_cpu_buffer(buffer->buffers[cpu]);
1693
1694 kfree(buffer->buffers);
1695 free_cpumask_var(buffer->cpumask);
1696
1697 kfree(buffer);
1698}
1699EXPORT_SYMBOL_GPL(ring_buffer_free);
1700
1701void ring_buffer_set_clock(struct trace_buffer *buffer,
1702 u64 (*clock)(void))
1703{
1704 buffer->clock = clock;
1705}
1706
1707void ring_buffer_set_time_stamp_abs(struct trace_buffer *buffer, bool abs)
1708{
1709 buffer->time_stamp_abs = abs;
1710}
1711
1712bool ring_buffer_time_stamp_abs(struct trace_buffer *buffer)
1713{
1714 return buffer->time_stamp_abs;
1715}
1716
1717static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1718
1719static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1720{
1721 return local_read(&bpage->entries) & RB_WRITE_MASK;
1722}
1723
1724static inline unsigned long rb_page_write(struct buffer_page *bpage)
1725{
1726 return local_read(&bpage->write) & RB_WRITE_MASK;
1727}
1728
1729static int
1730rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1731{
1732 struct list_head *tail_page, *to_remove, *next_page;
1733 struct buffer_page *to_remove_page, *tmp_iter_page;
1734 struct buffer_page *last_page, *first_page;
1735 unsigned long nr_removed;
1736 unsigned long head_bit;
1737 int page_entries;
1738
1739 head_bit = 0;
1740
1741 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1742 atomic_inc(&cpu_buffer->record_disabled);
1743 /*
1744 * We don't race with the readers since we have acquired the reader
1745 * lock. We also don't race with writers after disabling recording.
1746 * This makes it easy to figure out the first and the last page to be
1747 * removed from the list. We unlink all the pages in between including
1748 * the first and last pages. This is done in a busy loop so that we
1749 * lose the least number of traces.
1750 * The pages are freed after we restart recording and unlock readers.
1751 */
1752 tail_page = &cpu_buffer->tail_page->list;
1753
1754 /*
1755 * tail page might be on reader page, we remove the next page
1756 * from the ring buffer
1757 */
1758 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1759 tail_page = rb_list_head(tail_page->next);
1760 to_remove = tail_page;
1761
1762 /* start of pages to remove */
1763 first_page = list_entry(rb_list_head(to_remove->next),
1764 struct buffer_page, list);
1765
1766 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1767 to_remove = rb_list_head(to_remove)->next;
1768 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1769 }
1770
1771 next_page = rb_list_head(to_remove)->next;
1772
1773 /*
1774 * Now we remove all pages between tail_page and next_page.
1775 * Make sure that we have head_bit value preserved for the
1776 * next page
1777 */
1778 tail_page->next = (struct list_head *)((unsigned long)next_page |
1779 head_bit);
1780 next_page = rb_list_head(next_page);
1781 next_page->prev = tail_page;
1782
1783 /* make sure pages points to a valid page in the ring buffer */
1784 cpu_buffer->pages = next_page;
1785
1786 /* update head page */
1787 if (head_bit)
1788 cpu_buffer->head_page = list_entry(next_page,
1789 struct buffer_page, list);
1790
1791 /*
1792 * change read pointer to make sure any read iterators reset
1793 * themselves
1794 */
1795 cpu_buffer->read = 0;
1796
1797 /* pages are removed, resume tracing and then free the pages */
1798 atomic_dec(&cpu_buffer->record_disabled);
1799 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1800
1801 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1802
1803 /* last buffer page to remove */
1804 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1805 list);
1806 tmp_iter_page = first_page;
1807
1808 do {
1809 cond_resched();
1810
1811 to_remove_page = tmp_iter_page;
1812 rb_inc_page(cpu_buffer, &tmp_iter_page);
1813
1814 /* update the counters */
1815 page_entries = rb_page_entries(to_remove_page);
1816 if (page_entries) {
1817 /*
1818 * If something was added to this page, it was full
1819 * since it is not the tail page. So we deduct the
1820 * bytes consumed in ring buffer from here.
1821 * Increment overrun to account for the lost events.
1822 */
1823 local_add(page_entries, &cpu_buffer->overrun);
1824 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1825 }
1826
1827 /*
1828 * We have already removed references to this list item, just
1829 * free up the buffer_page and its page
1830 */
1831 free_buffer_page(to_remove_page);
1832 nr_removed--;
1833
1834 } while (to_remove_page != last_page);
1835
1836 RB_WARN_ON(cpu_buffer, nr_removed);
1837
1838 return nr_removed == 0;
1839}
1840
1841static int
1842rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1843{
1844 struct list_head *pages = &cpu_buffer->new_pages;
1845 int retries, success;
1846
1847 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1848 /*
1849 * We are holding the reader lock, so the reader page won't be swapped
1850 * in the ring buffer. Now we are racing with the writer trying to
1851 * move head page and the tail page.
1852 * We are going to adapt the reader page update process where:
1853 * 1. We first splice the start and end of list of new pages between
1854 * the head page and its previous page.
1855 * 2. We cmpxchg the prev_page->next to point from head page to the
1856 * start of new pages list.
1857 * 3. Finally, we update the head->prev to the end of new list.
1858 *
1859 * We will try this process 10 times, to make sure that we don't keep
1860 * spinning.
1861 */
1862 retries = 10;
1863 success = 0;
1864 while (retries--) {
1865 struct list_head *head_page, *prev_page, *r;
1866 struct list_head *last_page, *first_page;
1867 struct list_head *head_page_with_bit;
1868
1869 head_page = &rb_set_head_page(cpu_buffer)->list;
1870 if (!head_page)
1871 break;
1872 prev_page = head_page->prev;
1873
1874 first_page = pages->next;
1875 last_page = pages->prev;
1876
1877 head_page_with_bit = (struct list_head *)
1878 ((unsigned long)head_page | RB_PAGE_HEAD);
1879
1880 last_page->next = head_page_with_bit;
1881 first_page->prev = prev_page;
1882
1883 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1884
1885 if (r == head_page_with_bit) {
1886 /*
1887 * yay, we replaced the page pointer to our new list,
1888 * now, we just have to update to head page's prev
1889 * pointer to point to end of list
1890 */
1891 head_page->prev = last_page;
1892 success = 1;
1893 break;
1894 }
1895 }
1896
1897 if (success)
1898 INIT_LIST_HEAD(pages);
1899 /*
1900 * If we weren't successful in adding in new pages, warn and stop
1901 * tracing
1902 */
1903 RB_WARN_ON(cpu_buffer, !success);
1904 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1905
1906 /* free pages if they weren't inserted */
1907 if (!success) {
1908 struct buffer_page *bpage, *tmp;
1909 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1910 list) {
1911 list_del_init(&bpage->list);
1912 free_buffer_page(bpage);
1913 }
1914 }
1915 return success;
1916}
1917
1918static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1919{
1920 int success;
1921
1922 if (cpu_buffer->nr_pages_to_update > 0)
1923 success = rb_insert_pages(cpu_buffer);
1924 else
1925 success = rb_remove_pages(cpu_buffer,
1926 -cpu_buffer->nr_pages_to_update);
1927
1928 if (success)
1929 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1930}
1931
1932static void update_pages_handler(struct work_struct *work)
1933{
1934 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1935 struct ring_buffer_per_cpu, update_pages_work);
1936 rb_update_pages(cpu_buffer);
1937 complete(&cpu_buffer->update_done);
1938}
1939
1940/**
1941 * ring_buffer_resize - resize the ring buffer
1942 * @buffer: the buffer to resize.
1943 * @size: the new size.
1944 * @cpu_id: the cpu buffer to resize
1945 *
1946 * Minimum size is 2 * BUF_PAGE_SIZE.
1947 *
1948 * Returns 0 on success and < 0 on failure.
1949 */
1950int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
1951 int cpu_id)
1952{
1953 struct ring_buffer_per_cpu *cpu_buffer;
1954 unsigned long nr_pages;
1955 int cpu, err = 0;
1956
1957 /*
1958 * Always succeed at resizing a non-existent buffer:
1959 */
1960 if (!buffer)
1961 return size;
1962
1963 /* Make sure the requested buffer exists */
1964 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1965 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1966 return size;
1967
1968 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1969
1970 /* we need a minimum of two pages */
1971 if (nr_pages < 2)
1972 nr_pages = 2;
1973
1974 size = nr_pages * BUF_PAGE_SIZE;
1975
1976 /* prevent another thread from changing buffer sizes */
1977 mutex_lock(&buffer->mutex);
1978
1979
1980 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1981 /*
1982 * Don't succeed if resizing is disabled, as a reader might be
1983 * manipulating the ring buffer and is expecting a sane state while
1984 * this is true.
1985 */
1986 for_each_buffer_cpu(buffer, cpu) {
1987 cpu_buffer = buffer->buffers[cpu];
1988 if (atomic_read(&cpu_buffer->resize_disabled)) {
1989 err = -EBUSY;
1990 goto out_err_unlock;
1991 }
1992 }
1993
1994 /* calculate the pages to update */
1995 for_each_buffer_cpu(buffer, cpu) {
1996 cpu_buffer = buffer->buffers[cpu];
1997
1998 cpu_buffer->nr_pages_to_update = nr_pages -
1999 cpu_buffer->nr_pages;
2000 /*
2001 * nothing more to do for removing pages or no update
2002 */
2003 if (cpu_buffer->nr_pages_to_update <= 0)
2004 continue;
2005 /*
2006 * to add pages, make sure all new pages can be
2007 * allocated without receiving ENOMEM
2008 */
2009 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2010 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
2011 &cpu_buffer->new_pages, cpu)) {
2012 /* not enough memory for new pages */
2013 err = -ENOMEM;
2014 goto out_err;
2015 }
2016 }
2017
2018 get_online_cpus();
2019 /*
2020 * Fire off all the required work handlers
2021 * We can't schedule on offline CPUs, but it's not necessary
2022 * since we can change their buffer sizes without any race.
2023 */
2024 for_each_buffer_cpu(buffer, cpu) {
2025 cpu_buffer = buffer->buffers[cpu];
2026 if (!cpu_buffer->nr_pages_to_update)
2027 continue;
2028
2029 /* Can't run something on an offline CPU. */
2030 if (!cpu_online(cpu)) {
2031 rb_update_pages(cpu_buffer);
2032 cpu_buffer->nr_pages_to_update = 0;
2033 } else {
2034 schedule_work_on(cpu,
2035 &cpu_buffer->update_pages_work);
2036 }
2037 }
2038
2039 /* wait for all the updates to complete */
2040 for_each_buffer_cpu(buffer, cpu) {
2041 cpu_buffer = buffer->buffers[cpu];
2042 if (!cpu_buffer->nr_pages_to_update)
2043 continue;
2044
2045 if (cpu_online(cpu))
2046 wait_for_completion(&cpu_buffer->update_done);
2047 cpu_buffer->nr_pages_to_update = 0;
2048 }
2049
2050 put_online_cpus();
2051 } else {
2052 /* Make sure this CPU has been initialized */
2053 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
2054 goto out;
2055
2056 cpu_buffer = buffer->buffers[cpu_id];
2057
2058 if (nr_pages == cpu_buffer->nr_pages)
2059 goto out;
2060
2061 /*
2062 * Don't succeed if resizing is disabled, as a reader might be
2063 * manipulating the ring buffer and is expecting a sane state while
2064 * this is true.
2065 */
2066 if (atomic_read(&cpu_buffer->resize_disabled)) {
2067 err = -EBUSY;
2068 goto out_err_unlock;
2069 }
2070
2071 cpu_buffer->nr_pages_to_update = nr_pages -
2072 cpu_buffer->nr_pages;
2073
2074 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2075 if (cpu_buffer->nr_pages_to_update > 0 &&
2076 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
2077 &cpu_buffer->new_pages, cpu_id)) {
2078 err = -ENOMEM;
2079 goto out_err;
2080 }
2081
2082 get_online_cpus();
2083
2084 /* Can't run something on an offline CPU. */
2085 if (!cpu_online(cpu_id))
2086 rb_update_pages(cpu_buffer);
2087 else {
2088 schedule_work_on(cpu_id,
2089 &cpu_buffer->update_pages_work);
2090 wait_for_completion(&cpu_buffer->update_done);
2091 }
2092
2093 cpu_buffer->nr_pages_to_update = 0;
2094 put_online_cpus();
2095 }
2096
2097 out:
2098 /*
2099 * The ring buffer resize can happen with the ring buffer
2100 * enabled, so that the update disturbs the tracing as little
2101 * as possible. But if the buffer is disabled, we do not need
2102 * to worry about that, and we can take the time to verify
2103 * that the buffer is not corrupt.
2104 */
2105 if (atomic_read(&buffer->record_disabled)) {
2106 atomic_inc(&buffer->record_disabled);
2107 /*
2108 * Even though the buffer was disabled, we must make sure
2109 * that it is truly disabled before calling rb_check_pages.
2110 * There could have been a race between checking
2111 * record_disable and incrementing it.
2112 */
2113 synchronize_rcu();
2114 for_each_buffer_cpu(buffer, cpu) {
2115 cpu_buffer = buffer->buffers[cpu];
2116 rb_check_pages(cpu_buffer);
2117 }
2118 atomic_dec(&buffer->record_disabled);
2119 }
2120
2121 mutex_unlock(&buffer->mutex);
2122 return size;
2123
2124 out_err:
2125 for_each_buffer_cpu(buffer, cpu) {
2126 struct buffer_page *bpage, *tmp;
2127
2128 cpu_buffer = buffer->buffers[cpu];
2129 cpu_buffer->nr_pages_to_update = 0;
2130
2131 if (list_empty(&cpu_buffer->new_pages))
2132 continue;
2133
2134 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2135 list) {
2136 list_del_init(&bpage->list);
2137 free_buffer_page(bpage);
2138 }
2139 }
2140 out_err_unlock:
2141 mutex_unlock(&buffer->mutex);
2142 return err;
2143}
2144EXPORT_SYMBOL_GPL(ring_buffer_resize);
2145
2146void ring_buffer_change_overwrite(struct trace_buffer *buffer, int val)
2147{
2148 mutex_lock(&buffer->mutex);
2149 if (val)
2150 buffer->flags |= RB_FL_OVERWRITE;
2151 else
2152 buffer->flags &= ~RB_FL_OVERWRITE;
2153 mutex_unlock(&buffer->mutex);
2154}
2155EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
2156
2157static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
2158{
2159 return bpage->page->data + index;
2160}
2161
2162static __always_inline struct ring_buffer_event *
2163rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
2164{
2165 return __rb_page_index(cpu_buffer->reader_page,
2166 cpu_buffer->reader_page->read);
2167}
2168
2169static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
2170{
2171 return local_read(&bpage->page->commit);
2172}
2173
2174static struct ring_buffer_event *
2175rb_iter_head_event(struct ring_buffer_iter *iter)
2176{
2177 struct ring_buffer_event *event;
2178 struct buffer_page *iter_head_page = iter->head_page;
2179 unsigned long commit;
2180 unsigned length;
2181
2182 if (iter->head != iter->next_event)
2183 return iter->event;
2184
2185 /*
2186 * When the writer goes across pages, it issues a cmpxchg which
2187 * is a mb(), which will synchronize with the rmb here.
2188 * (see rb_tail_page_update() and __rb_reserve_next())
2189 */
2190 commit = rb_page_commit(iter_head_page);
2191 smp_rmb();
2192 event = __rb_page_index(iter_head_page, iter->head);
2193 length = rb_event_length(event);
2194
2195 /*
2196 * READ_ONCE() doesn't work on functions and we don't want the
2197 * compiler doing any crazy optimizations with length.
2198 */
2199 barrier();
2200
2201 if ((iter->head + length) > commit || length > BUF_MAX_DATA_SIZE)
2202 /* Writer corrupted the read? */
2203 goto reset;
2204
2205 memcpy(iter->event, event, length);
2206 /*
2207 * If the page stamp is still the same after this rmb() then the
2208 * event was safely copied without the writer entering the page.
2209 */
2210 smp_rmb();
2211
2212 /* Make sure the page didn't change since we read this */
2213 if (iter->page_stamp != iter_head_page->page->time_stamp ||
2214 commit > rb_page_commit(iter_head_page))
2215 goto reset;
2216
2217 iter->next_event = iter->head + length;
2218 return iter->event;
2219 reset:
2220 /* Reset to the beginning */
2221 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2222 iter->head = 0;
2223 iter->next_event = 0;
2224 iter->missed_events = 1;
2225 return NULL;
2226}
2227
2228/* Size is determined by what has been committed */
2229static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
2230{
2231 return rb_page_commit(bpage);
2232}
2233
2234static __always_inline unsigned
2235rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
2236{
2237 return rb_page_commit(cpu_buffer->commit_page);
2238}
2239
2240static __always_inline unsigned
2241rb_event_index(struct ring_buffer_event *event)
2242{
2243 unsigned long addr = (unsigned long)event;
2244
2245 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
2246}
2247
2248static void rb_inc_iter(struct ring_buffer_iter *iter)
2249{
2250 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2251
2252 /*
2253 * The iterator could be on the reader page (it starts there).
2254 * But the head could have moved, since the reader was
2255 * found. Check for this case and assign the iterator
2256 * to the head page instead of next.
2257 */
2258 if (iter->head_page == cpu_buffer->reader_page)
2259 iter->head_page = rb_set_head_page(cpu_buffer);
2260 else
2261 rb_inc_page(cpu_buffer, &iter->head_page);
2262
2263 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2264 iter->head = 0;
2265 iter->next_event = 0;
2266}
2267
2268/*
2269 * rb_handle_head_page - writer hit the head page
2270 *
2271 * Returns: +1 to retry page
2272 * 0 to continue
2273 * -1 on error
2274 */
2275static int
2276rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2277 struct buffer_page *tail_page,
2278 struct buffer_page *next_page)
2279{
2280 struct buffer_page *new_head;
2281 int entries;
2282 int type;
2283 int ret;
2284
2285 entries = rb_page_entries(next_page);
2286
2287 /*
2288 * The hard part is here. We need to move the head
2289 * forward, and protect against both readers on
2290 * other CPUs and writers coming in via interrupts.
2291 */
2292 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2293 RB_PAGE_HEAD);
2294
2295 /*
2296 * type can be one of four:
2297 * NORMAL - an interrupt already moved it for us
2298 * HEAD - we are the first to get here.
2299 * UPDATE - we are the interrupt interrupting
2300 * a current move.
2301 * MOVED - a reader on another CPU moved the next
2302 * pointer to its reader page. Give up
2303 * and try again.
2304 */
2305
2306 switch (type) {
2307 case RB_PAGE_HEAD:
2308 /*
2309 * We changed the head to UPDATE, thus
2310 * it is our responsibility to update
2311 * the counters.
2312 */
2313 local_add(entries, &cpu_buffer->overrun);
2314 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2315
2316 /*
2317 * The entries will be zeroed out when we move the
2318 * tail page.
2319 */
2320
2321 /* still more to do */
2322 break;
2323
2324 case RB_PAGE_UPDATE:
2325 /*
2326 * This is an interrupt that interrupt the
2327 * previous update. Still more to do.
2328 */
2329 break;
2330 case RB_PAGE_NORMAL:
2331 /*
2332 * An interrupt came in before the update
2333 * and processed this for us.
2334 * Nothing left to do.
2335 */
2336 return 1;
2337 case RB_PAGE_MOVED:
2338 /*
2339 * The reader is on another CPU and just did
2340 * a swap with our next_page.
2341 * Try again.
2342 */
2343 return 1;
2344 default:
2345 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2346 return -1;
2347 }
2348
2349 /*
2350 * Now that we are here, the old head pointer is
2351 * set to UPDATE. This will keep the reader from
2352 * swapping the head page with the reader page.
2353 * The reader (on another CPU) will spin till
2354 * we are finished.
2355 *
2356 * We just need to protect against interrupts
2357 * doing the job. We will set the next pointer
2358 * to HEAD. After that, we set the old pointer
2359 * to NORMAL, but only if it was HEAD before.
2360 * otherwise we are an interrupt, and only
2361 * want the outer most commit to reset it.
2362 */
2363 new_head = next_page;
2364 rb_inc_page(cpu_buffer, &new_head);
2365
2366 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2367 RB_PAGE_NORMAL);
2368
2369 /*
2370 * Valid returns are:
2371 * HEAD - an interrupt came in and already set it.
2372 * NORMAL - One of two things:
2373 * 1) We really set it.
2374 * 2) A bunch of interrupts came in and moved
2375 * the page forward again.
2376 */
2377 switch (ret) {
2378 case RB_PAGE_HEAD:
2379 case RB_PAGE_NORMAL:
2380 /* OK */
2381 break;
2382 default:
2383 RB_WARN_ON(cpu_buffer, 1);
2384 return -1;
2385 }
2386
2387 /*
2388 * It is possible that an interrupt came in,
2389 * set the head up, then more interrupts came in
2390 * and moved it again. When we get back here,
2391 * the page would have been set to NORMAL but we
2392 * just set it back to HEAD.
2393 *
2394 * How do you detect this? Well, if that happened
2395 * the tail page would have moved.
2396 */
2397 if (ret == RB_PAGE_NORMAL) {
2398 struct buffer_page *buffer_tail_page;
2399
2400 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2401 /*
2402 * If the tail had moved passed next, then we need
2403 * to reset the pointer.
2404 */
2405 if (buffer_tail_page != tail_page &&
2406 buffer_tail_page != next_page)
2407 rb_head_page_set_normal(cpu_buffer, new_head,
2408 next_page,
2409 RB_PAGE_HEAD);
2410 }
2411
2412 /*
2413 * If this was the outer most commit (the one that
2414 * changed the original pointer from HEAD to UPDATE),
2415 * then it is up to us to reset it to NORMAL.
2416 */
2417 if (type == RB_PAGE_HEAD) {
2418 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2419 tail_page,
2420 RB_PAGE_UPDATE);
2421 if (RB_WARN_ON(cpu_buffer,
2422 ret != RB_PAGE_UPDATE))
2423 return -1;
2424 }
2425
2426 return 0;
2427}
2428
2429static inline void
2430rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2431 unsigned long tail, struct rb_event_info *info)
2432{
2433 struct buffer_page *tail_page = info->tail_page;
2434 struct ring_buffer_event *event;
2435 unsigned long length = info->length;
2436
2437 /*
2438 * Only the event that crossed the page boundary
2439 * must fill the old tail_page with padding.
2440 */
2441 if (tail >= BUF_PAGE_SIZE) {
2442 /*
2443 * If the page was filled, then we still need
2444 * to update the real_end. Reset it to zero
2445 * and the reader will ignore it.
2446 */
2447 if (tail == BUF_PAGE_SIZE)
2448 tail_page->real_end = 0;
2449
2450 local_sub(length, &tail_page->write);
2451 return;
2452 }
2453
2454 event = __rb_page_index(tail_page, tail);
2455
2456 /* account for padding bytes */
2457 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2458
2459 /*
2460 * Save the original length to the meta data.
2461 * This will be used by the reader to add lost event
2462 * counter.
2463 */
2464 tail_page->real_end = tail;
2465
2466 /*
2467 * If this event is bigger than the minimum size, then
2468 * we need to be careful that we don't subtract the
2469 * write counter enough to allow another writer to slip
2470 * in on this page.
2471 * We put in a discarded commit instead, to make sure
2472 * that this space is not used again.
2473 *
2474 * If we are less than the minimum size, we don't need to
2475 * worry about it.
2476 */
2477 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2478 /* No room for any events */
2479
2480 /* Mark the rest of the page with padding */
2481 rb_event_set_padding(event);
2482
2483 /* Set the write back to the previous setting */
2484 local_sub(length, &tail_page->write);
2485 return;
2486 }
2487
2488 /* Put in a discarded event */
2489 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2490 event->type_len = RINGBUF_TYPE_PADDING;
2491 /* time delta must be non zero */
2492 event->time_delta = 1;
2493
2494 /* Set write to end of buffer */
2495 length = (tail + length) - BUF_PAGE_SIZE;
2496 local_sub(length, &tail_page->write);
2497}
2498
2499static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2500
2501/*
2502 * This is the slow path, force gcc not to inline it.
2503 */
2504static noinline struct ring_buffer_event *
2505rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2506 unsigned long tail, struct rb_event_info *info)
2507{
2508 struct buffer_page *tail_page = info->tail_page;
2509 struct buffer_page *commit_page = cpu_buffer->commit_page;
2510 struct trace_buffer *buffer = cpu_buffer->buffer;
2511 struct buffer_page *next_page;
2512 int ret;
2513
2514 next_page = tail_page;
2515
2516 rb_inc_page(cpu_buffer, &next_page);
2517
2518 /*
2519 * If for some reason, we had an interrupt storm that made
2520 * it all the way around the buffer, bail, and warn
2521 * about it.
2522 */
2523 if (unlikely(next_page == commit_page)) {
2524 local_inc(&cpu_buffer->commit_overrun);
2525 goto out_reset;
2526 }
2527
2528 /*
2529 * This is where the fun begins!
2530 *
2531 * We are fighting against races between a reader that
2532 * could be on another CPU trying to swap its reader
2533 * page with the buffer head.
2534 *
2535 * We are also fighting against interrupts coming in and
2536 * moving the head or tail on us as well.
2537 *
2538 * If the next page is the head page then we have filled
2539 * the buffer, unless the commit page is still on the
2540 * reader page.
2541 */
2542 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2543
2544 /*
2545 * If the commit is not on the reader page, then
2546 * move the header page.
2547 */
2548 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2549 /*
2550 * If we are not in overwrite mode,
2551 * this is easy, just stop here.
2552 */
2553 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2554 local_inc(&cpu_buffer->dropped_events);
2555 goto out_reset;
2556 }
2557
2558 ret = rb_handle_head_page(cpu_buffer,
2559 tail_page,
2560 next_page);
2561 if (ret < 0)
2562 goto out_reset;
2563 if (ret)
2564 goto out_again;
2565 } else {
2566 /*
2567 * We need to be careful here too. The
2568 * commit page could still be on the reader
2569 * page. We could have a small buffer, and
2570 * have filled up the buffer with events
2571 * from interrupts and such, and wrapped.
2572 *
2573 * Note, if the tail page is also the on the
2574 * reader_page, we let it move out.
2575 */
2576 if (unlikely((cpu_buffer->commit_page !=
2577 cpu_buffer->tail_page) &&
2578 (cpu_buffer->commit_page ==
2579 cpu_buffer->reader_page))) {
2580 local_inc(&cpu_buffer->commit_overrun);
2581 goto out_reset;
2582 }
2583 }
2584 }
2585
2586 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2587
2588 out_again:
2589
2590 rb_reset_tail(cpu_buffer, tail, info);
2591
2592 /* Commit what we have for now. */
2593 rb_end_commit(cpu_buffer);
2594 /* rb_end_commit() decs committing */
2595 local_inc(&cpu_buffer->committing);
2596
2597 /* fail and let the caller try again */
2598 return ERR_PTR(-EAGAIN);
2599
2600 out_reset:
2601 /* reset write */
2602 rb_reset_tail(cpu_buffer, tail, info);
2603
2604 return NULL;
2605}
2606
2607/* Slow path */
2608static struct ring_buffer_event *
2609rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2610{
2611 if (abs)
2612 event->type_len = RINGBUF_TYPE_TIME_STAMP;
2613 else
2614 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2615
2616 /* Not the first event on the page, or not delta? */
2617 if (abs || rb_event_index(event)) {
2618 event->time_delta = delta & TS_MASK;
2619 event->array[0] = delta >> TS_SHIFT;
2620 } else {
2621 /* nope, just zero it */
2622 event->time_delta = 0;
2623 event->array[0] = 0;
2624 }
2625
2626 return skip_time_extend(event);
2627}
2628
2629static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2630 struct ring_buffer_event *event);
2631
2632#ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2633static inline bool sched_clock_stable(void)
2634{
2635 return true;
2636}
2637#endif
2638
2639static void
2640rb_check_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2641 struct rb_event_info *info)
2642{
2643 u64 write_stamp;
2644
2645 WARN_ONCE(1, "Delta way too big! %llu ts=%llu before=%llu after=%llu write stamp=%llu\n%s",
2646 (unsigned long long)info->delta,
2647 (unsigned long long)info->ts,
2648 (unsigned long long)info->before,
2649 (unsigned long long)info->after,
2650 (unsigned long long)(rb_time_read(&cpu_buffer->write_stamp, &write_stamp) ? write_stamp : 0),
2651 sched_clock_stable() ? "" :
2652 "If you just came from a suspend/resume,\n"
2653 "please switch to the trace global clock:\n"
2654 " echo global > /sys/kernel/debug/tracing/trace_clock\n"
2655 "or add trace_clock=global to the kernel command line\n");
2656}
2657
2658static void rb_add_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2659 struct ring_buffer_event **event,
2660 struct rb_event_info *info,
2661 u64 *delta,
2662 unsigned int *length)
2663{
2664 bool abs = info->add_timestamp &
2665 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE);
2666
2667 if (unlikely(info->delta > (1ULL << 59))) {
2668 /* did the clock go backwards */
2669 if (info->before == info->after && info->before > info->ts) {
2670 /* not interrupted */
2671 static int once;
2672
2673 /*
2674 * This is possible with a recalibrating of the TSC.
2675 * Do not produce a call stack, but just report it.
2676 */
2677 if (!once) {
2678 once++;
2679 pr_warn("Ring buffer clock went backwards: %llu -> %llu\n",
2680 info->before, info->ts);
2681 }
2682 } else
2683 rb_check_timestamp(cpu_buffer, info);
2684 if (!abs)
2685 info->delta = 0;
2686 }
2687 *event = rb_add_time_stamp(*event, info->delta, abs);
2688 *length -= RB_LEN_TIME_EXTEND;
2689 *delta = 0;
2690}
2691
2692/**
2693 * rb_update_event - update event type and data
2694 * @cpu_buffer: The per cpu buffer of the @event
2695 * @event: the event to update
2696 * @info: The info to update the @event with (contains length and delta)
2697 *
2698 * Update the type and data fields of the @event. The length
2699 * is the actual size that is written to the ring buffer,
2700 * and with this, we can determine what to place into the
2701 * data field.
2702 */
2703static void
2704rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2705 struct ring_buffer_event *event,
2706 struct rb_event_info *info)
2707{
2708 unsigned length = info->length;
2709 u64 delta = info->delta;
2710
2711 /*
2712 * If we need to add a timestamp, then we
2713 * add it to the start of the reserved space.
2714 */
2715 if (unlikely(info->add_timestamp))
2716 rb_add_timestamp(cpu_buffer, &event, info, &delta, &length);
2717
2718 event->time_delta = delta;
2719 length -= RB_EVNT_HDR_SIZE;
2720 if (length > RB_MAX_SMALL_DATA) {
2721 event->type_len = 0;
2722 event->array[0] = length;
2723 } else
2724 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2725}
2726
2727static unsigned rb_calculate_event_length(unsigned length)
2728{
2729 struct ring_buffer_event event; /* Used only for sizeof array */
2730
2731 /* zero length can cause confusions */
2732 if (!length)
2733 length++;
2734
2735 if (length > RB_MAX_SMALL_DATA)
2736 length += sizeof(event.array[0]);
2737
2738 length += RB_EVNT_HDR_SIZE;
2739 length = ALIGN(length, RB_ALIGNMENT);
2740
2741 /*
2742 * In case the time delta is larger than the 27 bits for it
2743 * in the header, we need to add a timestamp. If another
2744 * event comes in when trying to discard this one to increase
2745 * the length, then the timestamp will be added in the allocated
2746 * space of this event. If length is bigger than the size needed
2747 * for the TIME_EXTEND, then padding has to be used. The events
2748 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2749 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2750 * As length is a multiple of 4, we only need to worry if it
2751 * is 12 (RB_LEN_TIME_EXTEND + 4).
2752 */
2753 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2754 length += RB_ALIGNMENT;
2755
2756 return length;
2757}
2758
2759static __always_inline bool
2760rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2761 struct ring_buffer_event *event)
2762{
2763 unsigned long addr = (unsigned long)event;
2764 unsigned long index;
2765
2766 index = rb_event_index(event);
2767 addr &= PAGE_MASK;
2768
2769 return cpu_buffer->commit_page->page == (void *)addr &&
2770 rb_commit_index(cpu_buffer) == index;
2771}
2772
2773static u64 rb_time_delta(struct ring_buffer_event *event)
2774{
2775 switch (event->type_len) {
2776 case RINGBUF_TYPE_PADDING:
2777 return 0;
2778
2779 case RINGBUF_TYPE_TIME_EXTEND:
2780 return ring_buffer_event_time_stamp(event);
2781
2782 case RINGBUF_TYPE_TIME_STAMP:
2783 return 0;
2784
2785 case RINGBUF_TYPE_DATA:
2786 return event->time_delta;
2787 default:
2788 return 0;
2789 }
2790}
2791
2792static inline int
2793rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2794 struct ring_buffer_event *event)
2795{
2796 unsigned long new_index, old_index;
2797 struct buffer_page *bpage;
2798 unsigned long index;
2799 unsigned long addr;
2800 u64 write_stamp;
2801 u64 delta;
2802
2803 new_index = rb_event_index(event);
2804 old_index = new_index + rb_event_ts_length(event);
2805 addr = (unsigned long)event;
2806 addr &= PAGE_MASK;
2807
2808 bpage = READ_ONCE(cpu_buffer->tail_page);
2809
2810 delta = rb_time_delta(event);
2811
2812 if (!rb_time_read(&cpu_buffer->write_stamp, &write_stamp))
2813 return 0;
2814
2815 /* Make sure the write stamp is read before testing the location */
2816 barrier();
2817
2818 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2819 unsigned long write_mask =
2820 local_read(&bpage->write) & ~RB_WRITE_MASK;
2821 unsigned long event_length = rb_event_length(event);
2822
2823 /* Something came in, can't discard */
2824 if (!rb_time_cmpxchg(&cpu_buffer->write_stamp,
2825 write_stamp, write_stamp - delta))
2826 return 0;
2827
2828 /*
2829 * If an event were to come in now, it would see that the
2830 * write_stamp and the before_stamp are different, and assume
2831 * that this event just added itself before updating
2832 * the write stamp. The interrupting event will fix the
2833 * write stamp for us, and use the before stamp as its delta.
2834 */
2835
2836 /*
2837 * This is on the tail page. It is possible that
2838 * a write could come in and move the tail page
2839 * and write to the next page. That is fine
2840 * because we just shorten what is on this page.
2841 */
2842 old_index += write_mask;
2843 new_index += write_mask;
2844 index = local_cmpxchg(&bpage->write, old_index, new_index);
2845 if (index == old_index) {
2846 /* update counters */
2847 local_sub(event_length, &cpu_buffer->entries_bytes);
2848 return 1;
2849 }
2850 }
2851
2852 /* could not discard */
2853 return 0;
2854}
2855
2856static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2857{
2858 local_inc(&cpu_buffer->committing);
2859 local_inc(&cpu_buffer->commits);
2860}
2861
2862static __always_inline void
2863rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2864{
2865 unsigned long max_count;
2866
2867 /*
2868 * We only race with interrupts and NMIs on this CPU.
2869 * If we own the commit event, then we can commit
2870 * all others that interrupted us, since the interruptions
2871 * are in stack format (they finish before they come
2872 * back to us). This allows us to do a simple loop to
2873 * assign the commit to the tail.
2874 */
2875 again:
2876 max_count = cpu_buffer->nr_pages * 100;
2877
2878 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2879 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2880 return;
2881 if (RB_WARN_ON(cpu_buffer,
2882 rb_is_reader_page(cpu_buffer->tail_page)))
2883 return;
2884 local_set(&cpu_buffer->commit_page->page->commit,
2885 rb_page_write(cpu_buffer->commit_page));
2886 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2887 /* add barrier to keep gcc from optimizing too much */
2888 barrier();
2889 }
2890 while (rb_commit_index(cpu_buffer) !=
2891 rb_page_write(cpu_buffer->commit_page)) {
2892
2893 local_set(&cpu_buffer->commit_page->page->commit,
2894 rb_page_write(cpu_buffer->commit_page));
2895 RB_WARN_ON(cpu_buffer,
2896 local_read(&cpu_buffer->commit_page->page->commit) &
2897 ~RB_WRITE_MASK);
2898 barrier();
2899 }
2900
2901 /* again, keep gcc from optimizing */
2902 barrier();
2903
2904 /*
2905 * If an interrupt came in just after the first while loop
2906 * and pushed the tail page forward, we will be left with
2907 * a dangling commit that will never go forward.
2908 */
2909 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2910 goto again;
2911}
2912
2913static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2914{
2915 unsigned long commits;
2916
2917 if (RB_WARN_ON(cpu_buffer,
2918 !local_read(&cpu_buffer->committing)))
2919 return;
2920
2921 again:
2922 commits = local_read(&cpu_buffer->commits);
2923 /* synchronize with interrupts */
2924 barrier();
2925 if (local_read(&cpu_buffer->committing) == 1)
2926 rb_set_commit_to_write(cpu_buffer);
2927
2928 local_dec(&cpu_buffer->committing);
2929
2930 /* synchronize with interrupts */
2931 barrier();
2932
2933 /*
2934 * Need to account for interrupts coming in between the
2935 * updating of the commit page and the clearing of the
2936 * committing counter.
2937 */
2938 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2939 !local_read(&cpu_buffer->committing)) {
2940 local_inc(&cpu_buffer->committing);
2941 goto again;
2942 }
2943}
2944
2945static inline void rb_event_discard(struct ring_buffer_event *event)
2946{
2947 if (extended_time(event))
2948 event = skip_time_extend(event);
2949
2950 /* array[0] holds the actual length for the discarded event */
2951 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2952 event->type_len = RINGBUF_TYPE_PADDING;
2953 /* time delta must be non zero */
2954 if (!event->time_delta)
2955 event->time_delta = 1;
2956}
2957
2958static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2959 struct ring_buffer_event *event)
2960{
2961 local_inc(&cpu_buffer->entries);
2962 rb_end_commit(cpu_buffer);
2963}
2964
2965static __always_inline void
2966rb_wakeups(struct trace_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2967{
2968 size_t nr_pages;
2969 size_t dirty;
2970 size_t full;
2971
2972 if (buffer->irq_work.waiters_pending) {
2973 buffer->irq_work.waiters_pending = false;
2974 /* irq_work_queue() supplies it's own memory barriers */
2975 irq_work_queue(&buffer->irq_work.work);
2976 }
2977
2978 if (cpu_buffer->irq_work.waiters_pending) {
2979 cpu_buffer->irq_work.waiters_pending = false;
2980 /* irq_work_queue() supplies it's own memory barriers */
2981 irq_work_queue(&cpu_buffer->irq_work.work);
2982 }
2983
2984 if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
2985 return;
2986
2987 if (cpu_buffer->reader_page == cpu_buffer->commit_page)
2988 return;
2989
2990 if (!cpu_buffer->irq_work.full_waiters_pending)
2991 return;
2992
2993 cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
2994
2995 full = cpu_buffer->shortest_full;
2996 nr_pages = cpu_buffer->nr_pages;
2997 dirty = ring_buffer_nr_dirty_pages(buffer, cpu_buffer->cpu);
2998 if (full && nr_pages && (dirty * 100) <= full * nr_pages)
2999 return;
3000
3001 cpu_buffer->irq_work.wakeup_full = true;
3002 cpu_buffer->irq_work.full_waiters_pending = false;
3003 /* irq_work_queue() supplies it's own memory barriers */
3004 irq_work_queue(&cpu_buffer->irq_work.work);
3005}
3006
3007/*
3008 * The lock and unlock are done within a preempt disable section.
3009 * The current_context per_cpu variable can only be modified
3010 * by the current task between lock and unlock. But it can
3011 * be modified more than once via an interrupt. To pass this
3012 * information from the lock to the unlock without having to
3013 * access the 'in_interrupt()' functions again (which do show
3014 * a bit of overhead in something as critical as function tracing,
3015 * we use a bitmask trick.
3016 *
3017 * bit 0 = NMI context
3018 * bit 1 = IRQ context
3019 * bit 2 = SoftIRQ context
3020 * bit 3 = normal context.
3021 *
3022 * This works because this is the order of contexts that can
3023 * preempt other contexts. A SoftIRQ never preempts an IRQ
3024 * context.
3025 *
3026 * When the context is determined, the corresponding bit is
3027 * checked and set (if it was set, then a recursion of that context
3028 * happened).
3029 *
3030 * On unlock, we need to clear this bit. To do so, just subtract
3031 * 1 from the current_context and AND it to itself.
3032 *
3033 * (binary)
3034 * 101 - 1 = 100
3035 * 101 & 100 = 100 (clearing bit zero)
3036 *
3037 * 1010 - 1 = 1001
3038 * 1010 & 1001 = 1000 (clearing bit 1)
3039 *
3040 * The least significant bit can be cleared this way, and it
3041 * just so happens that it is the same bit corresponding to
3042 * the current context.
3043 */
3044
3045static __always_inline int
3046trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
3047{
3048 unsigned int val = cpu_buffer->current_context;
3049 unsigned long pc = preempt_count();
3050 int bit;
3051
3052 if (!(pc & (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_OFFSET)))
3053 bit = RB_CTX_NORMAL;
3054 else
3055 bit = pc & NMI_MASK ? RB_CTX_NMI :
3056 pc & HARDIRQ_MASK ? RB_CTX_IRQ : RB_CTX_SOFTIRQ;
3057
3058 if (unlikely(val & (1 << (bit + cpu_buffer->nest))))
3059 return 1;
3060
3061 val |= (1 << (bit + cpu_buffer->nest));
3062 cpu_buffer->current_context = val;
3063
3064 return 0;
3065}
3066
3067static __always_inline void
3068trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
3069{
3070 cpu_buffer->current_context &=
3071 cpu_buffer->current_context - (1 << cpu_buffer->nest);
3072}
3073
3074/* The recursive locking above uses 4 bits */
3075#define NESTED_BITS 4
3076
3077/**
3078 * ring_buffer_nest_start - Allow to trace while nested
3079 * @buffer: The ring buffer to modify
3080 *
3081 * The ring buffer has a safety mechanism to prevent recursion.
3082 * But there may be a case where a trace needs to be done while
3083 * tracing something else. In this case, calling this function
3084 * will allow this function to nest within a currently active
3085 * ring_buffer_lock_reserve().
3086 *
3087 * Call this function before calling another ring_buffer_lock_reserve() and
3088 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
3089 */
3090void ring_buffer_nest_start(struct trace_buffer *buffer)
3091{
3092 struct ring_buffer_per_cpu *cpu_buffer;
3093 int cpu;
3094
3095 /* Enabled by ring_buffer_nest_end() */
3096 preempt_disable_notrace();
3097 cpu = raw_smp_processor_id();
3098 cpu_buffer = buffer->buffers[cpu];
3099 /* This is the shift value for the above recursive locking */
3100 cpu_buffer->nest += NESTED_BITS;
3101}
3102
3103/**
3104 * ring_buffer_nest_end - Allow to trace while nested
3105 * @buffer: The ring buffer to modify
3106 *
3107 * Must be called after ring_buffer_nest_start() and after the
3108 * ring_buffer_unlock_commit().
3109 */
3110void ring_buffer_nest_end(struct trace_buffer *buffer)
3111{
3112 struct ring_buffer_per_cpu *cpu_buffer;
3113 int cpu;
3114
3115 /* disabled by ring_buffer_nest_start() */
3116 cpu = raw_smp_processor_id();
3117 cpu_buffer = buffer->buffers[cpu];
3118 /* This is the shift value for the above recursive locking */
3119 cpu_buffer->nest -= NESTED_BITS;
3120 preempt_enable_notrace();
3121}
3122
3123/**
3124 * ring_buffer_unlock_commit - commit a reserved
3125 * @buffer: The buffer to commit to
3126 * @event: The event pointer to commit.
3127 *
3128 * This commits the data to the ring buffer, and releases any locks held.
3129 *
3130 * Must be paired with ring_buffer_lock_reserve.
3131 */
3132int ring_buffer_unlock_commit(struct trace_buffer *buffer,
3133 struct ring_buffer_event *event)
3134{
3135 struct ring_buffer_per_cpu *cpu_buffer;
3136 int cpu = raw_smp_processor_id();
3137
3138 cpu_buffer = buffer->buffers[cpu];
3139
3140 rb_commit(cpu_buffer, event);
3141
3142 rb_wakeups(buffer, cpu_buffer);
3143
3144 trace_recursive_unlock(cpu_buffer);
3145
3146 preempt_enable_notrace();
3147
3148 return 0;
3149}
3150EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
3151
3152static struct ring_buffer_event *
3153__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
3154 struct rb_event_info *info)
3155{
3156 struct ring_buffer_event *event;
3157 struct buffer_page *tail_page;
3158 unsigned long tail, write, w;
3159 bool a_ok;
3160 bool b_ok;
3161
3162 /* Don't let the compiler play games with cpu_buffer->tail_page */
3163 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
3164
3165 /*A*/ w = local_read(&tail_page->write) & RB_WRITE_MASK;
3166 barrier();
3167 b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3168 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3169 barrier();
3170 info->ts = rb_time_stamp(cpu_buffer->buffer);
3171
3172 if ((info->add_timestamp & RB_ADD_STAMP_ABSOLUTE)) {
3173 info->delta = info->ts;
3174 } else {
3175 /*
3176 * If interrupting an event time update, we may need an
3177 * absolute timestamp.
3178 * Don't bother if this is the start of a new page (w == 0).
3179 */
3180 if (unlikely(!a_ok || !b_ok || (info->before != info->after && w))) {
3181 info->add_timestamp |= RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND;
3182 info->length += RB_LEN_TIME_EXTEND;
3183 } else {
3184 info->delta = info->ts - info->after;
3185 if (unlikely(test_time_stamp(info->delta))) {
3186 info->add_timestamp |= RB_ADD_STAMP_EXTEND;
3187 info->length += RB_LEN_TIME_EXTEND;
3188 }
3189 }
3190 }
3191
3192 /*B*/ rb_time_set(&cpu_buffer->before_stamp, info->ts);
3193
3194 /*C*/ write = local_add_return(info->length, &tail_page->write);
3195
3196 /* set write to only the index of the write */
3197 write &= RB_WRITE_MASK;
3198
3199 tail = write - info->length;
3200
3201 /* See if we shot pass the end of this buffer page */
3202 if (unlikely(write > BUF_PAGE_SIZE)) {
3203 if (tail != w) {
3204 /* before and after may now different, fix it up*/
3205 b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3206 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3207 if (a_ok && b_ok && info->before != info->after)
3208 (void)rb_time_cmpxchg(&cpu_buffer->before_stamp,
3209 info->before, info->after);
3210 }
3211 return rb_move_tail(cpu_buffer, tail, info);
3212 }
3213
3214 if (likely(tail == w)) {
3215 u64 save_before;
3216 bool s_ok;
3217
3218 /* Nothing interrupted us between A and C */
3219 /*D*/ rb_time_set(&cpu_buffer->write_stamp, info->ts);
3220 barrier();
3221 /*E*/ s_ok = rb_time_read(&cpu_buffer->before_stamp, &save_before);
3222 RB_WARN_ON(cpu_buffer, !s_ok);
3223 if (likely(!(info->add_timestamp &
3224 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3225 /* This did not interrupt any time update */
3226 info->delta = info->ts - info->after;
3227 else
3228 /* Just use full timestamp for inerrupting event */
3229 info->delta = info->ts;
3230 barrier();
3231 if (unlikely(info->ts != save_before)) {
3232 /* SLOW PATH - Interrupted between C and E */
3233
3234 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3235 RB_WARN_ON(cpu_buffer, !a_ok);
3236
3237 /* Write stamp must only go forward */
3238 if (save_before > info->after) {
3239 /*
3240 * We do not care about the result, only that
3241 * it gets updated atomically.
3242 */
3243 (void)rb_time_cmpxchg(&cpu_buffer->write_stamp,
3244 info->after, save_before);
3245 }
3246 }
3247 } else {
3248 u64 ts;
3249 /* SLOW PATH - Interrupted between A and C */
3250 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3251 /* Was interrupted before here, write_stamp must be valid */
3252 RB_WARN_ON(cpu_buffer, !a_ok);
3253 ts = rb_time_stamp(cpu_buffer->buffer);
3254 barrier();
3255 /*E*/ if (write == (local_read(&tail_page->write) & RB_WRITE_MASK) &&
3256 info->after < ts) {
3257 /* Nothing came after this event between C and E */
3258 info->delta = ts - info->after;
3259 (void)rb_time_cmpxchg(&cpu_buffer->write_stamp,
3260 info->after, info->ts);
3261 info->ts = ts;
3262 } else {
3263 /*
3264 * Interrupted beween C and E:
3265 * Lost the previous events time stamp. Just set the
3266 * delta to zero, and this will be the same time as
3267 * the event this event interrupted. And the events that
3268 * came after this will still be correct (as they would
3269 * have built their delta on the previous event.
3270 */
3271 info->delta = 0;
3272 }
3273 info->add_timestamp &= ~RB_ADD_STAMP_FORCE;
3274 }
3275
3276 /*
3277 * If this is the first commit on the page, then it has the same
3278 * timestamp as the page itself.
3279 */
3280 if (unlikely(!tail && !(info->add_timestamp &
3281 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3282 info->delta = 0;
3283
3284 /* We reserved something on the buffer */
3285
3286 event = __rb_page_index(tail_page, tail);
3287 rb_update_event(cpu_buffer, event, info);
3288
3289 local_inc(&tail_page->entries);
3290
3291 /*
3292 * If this is the first commit on the page, then update
3293 * its timestamp.
3294 */
3295 if (unlikely(!tail))
3296 tail_page->page->time_stamp = info->ts;
3297
3298 /* account for these added bytes */
3299 local_add(info->length, &cpu_buffer->entries_bytes);
3300
3301 return event;
3302}
3303
3304static __always_inline struct ring_buffer_event *
3305rb_reserve_next_event(struct trace_buffer *buffer,
3306 struct ring_buffer_per_cpu *cpu_buffer,
3307 unsigned long length)
3308{
3309 struct ring_buffer_event *event;
3310 struct rb_event_info info;
3311 int nr_loops = 0;
3312 int add_ts_default;
3313
3314 rb_start_commit(cpu_buffer);
3315 /* The commit page can not change after this */
3316
3317#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3318 /*
3319 * Due to the ability to swap a cpu buffer from a buffer
3320 * it is possible it was swapped before we committed.
3321 * (committing stops a swap). We check for it here and
3322 * if it happened, we have to fail the write.
3323 */
3324 barrier();
3325 if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
3326 local_dec(&cpu_buffer->committing);
3327 local_dec(&cpu_buffer->commits);
3328 return NULL;
3329 }
3330#endif
3331
3332 info.length = rb_calculate_event_length(length);
3333
3334 if (ring_buffer_time_stamp_abs(cpu_buffer->buffer)) {
3335 add_ts_default = RB_ADD_STAMP_ABSOLUTE;
3336 info.length += RB_LEN_TIME_EXTEND;
3337 } else {
3338 add_ts_default = RB_ADD_STAMP_NONE;
3339 }
3340
3341 again:
3342 info.add_timestamp = add_ts_default;
3343 info.delta = 0;
3344
3345 /*
3346 * We allow for interrupts to reenter here and do a trace.
3347 * If one does, it will cause this original code to loop
3348 * back here. Even with heavy interrupts happening, this
3349 * should only happen a few times in a row. If this happens
3350 * 1000 times in a row, there must be either an interrupt
3351 * storm or we have something buggy.
3352 * Bail!
3353 */
3354 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
3355 goto out_fail;
3356
3357 event = __rb_reserve_next(cpu_buffer, &info);
3358
3359 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
3360 if (info.add_timestamp & (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND))
3361 info.length -= RB_LEN_TIME_EXTEND;
3362 goto again;
3363 }
3364
3365 if (likely(event))
3366 return event;
3367 out_fail:
3368 rb_end_commit(cpu_buffer);
3369 return NULL;
3370}
3371
3372/**
3373 * ring_buffer_lock_reserve - reserve a part of the buffer
3374 * @buffer: the ring buffer to reserve from
3375 * @length: the length of the data to reserve (excluding event header)
3376 *
3377 * Returns a reserved event on the ring buffer to copy directly to.
3378 * The user of this interface will need to get the body to write into
3379 * and can use the ring_buffer_event_data() interface.
3380 *
3381 * The length is the length of the data needed, not the event length
3382 * which also includes the event header.
3383 *
3384 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
3385 * If NULL is returned, then nothing has been allocated or locked.
3386 */
3387struct ring_buffer_event *
3388ring_buffer_lock_reserve(struct trace_buffer *buffer, unsigned long length)
3389{
3390 struct ring_buffer_per_cpu *cpu_buffer;
3391 struct ring_buffer_event *event;
3392 int cpu;
3393
3394 /* If we are tracing schedule, we don't want to recurse */
3395 preempt_disable_notrace();
3396
3397 if (unlikely(atomic_read(&buffer->record_disabled)))
3398 goto out;
3399
3400 cpu = raw_smp_processor_id();
3401
3402 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
3403 goto out;
3404
3405 cpu_buffer = buffer->buffers[cpu];
3406
3407 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
3408 goto out;
3409
3410 if (unlikely(length > BUF_MAX_DATA_SIZE))
3411 goto out;
3412
3413 if (unlikely(trace_recursive_lock(cpu_buffer)))
3414 goto out;
3415
3416 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3417 if (!event)
3418 goto out_unlock;
3419
3420 return event;
3421
3422 out_unlock:
3423 trace_recursive_unlock(cpu_buffer);
3424 out:
3425 preempt_enable_notrace();
3426 return NULL;
3427}
3428EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3429
3430/*
3431 * Decrement the entries to the page that an event is on.
3432 * The event does not even need to exist, only the pointer
3433 * to the page it is on. This may only be called before the commit
3434 * takes place.
3435 */
3436static inline void
3437rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3438 struct ring_buffer_event *event)
3439{
3440 unsigned long addr = (unsigned long)event;
3441 struct buffer_page *bpage = cpu_buffer->commit_page;
3442 struct buffer_page *start;
3443
3444 addr &= PAGE_MASK;
3445
3446 /* Do the likely case first */
3447 if (likely(bpage->page == (void *)addr)) {
3448 local_dec(&bpage->entries);
3449 return;
3450 }
3451
3452 /*
3453 * Because the commit page may be on the reader page we
3454 * start with the next page and check the end loop there.
3455 */
3456 rb_inc_page(cpu_buffer, &bpage);
3457 start = bpage;
3458 do {
3459 if (bpage->page == (void *)addr) {
3460 local_dec(&bpage->entries);
3461 return;
3462 }
3463 rb_inc_page(cpu_buffer, &bpage);
3464 } while (bpage != start);
3465
3466 /* commit not part of this buffer?? */
3467 RB_WARN_ON(cpu_buffer, 1);
3468}
3469
3470/**
3471 * ring_buffer_commit_discard - discard an event that has not been committed
3472 * @buffer: the ring buffer
3473 * @event: non committed event to discard
3474 *
3475 * Sometimes an event that is in the ring buffer needs to be ignored.
3476 * This function lets the user discard an event in the ring buffer
3477 * and then that event will not be read later.
3478 *
3479 * This function only works if it is called before the item has been
3480 * committed. It will try to free the event from the ring buffer
3481 * if another event has not been added behind it.
3482 *
3483 * If another event has been added behind it, it will set the event
3484 * up as discarded, and perform the commit.
3485 *
3486 * If this function is called, do not call ring_buffer_unlock_commit on
3487 * the event.
3488 */
3489void ring_buffer_discard_commit(struct trace_buffer *buffer,
3490 struct ring_buffer_event *event)
3491{
3492 struct ring_buffer_per_cpu *cpu_buffer;
3493 int cpu;
3494
3495 /* The event is discarded regardless */
3496 rb_event_discard(event);
3497
3498 cpu = smp_processor_id();
3499 cpu_buffer = buffer->buffers[cpu];
3500
3501 /*
3502 * This must only be called if the event has not been
3503 * committed yet. Thus we can assume that preemption
3504 * is still disabled.
3505 */
3506 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3507
3508 rb_decrement_entry(cpu_buffer, event);
3509 if (rb_try_to_discard(cpu_buffer, event))
3510 goto out;
3511
3512 out:
3513 rb_end_commit(cpu_buffer);
3514
3515 trace_recursive_unlock(cpu_buffer);
3516
3517 preempt_enable_notrace();
3518
3519}
3520EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3521
3522/**
3523 * ring_buffer_write - write data to the buffer without reserving
3524 * @buffer: The ring buffer to write to.
3525 * @length: The length of the data being written (excluding the event header)
3526 * @data: The data to write to the buffer.
3527 *
3528 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3529 * one function. If you already have the data to write to the buffer, it
3530 * may be easier to simply call this function.
3531 *
3532 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3533 * and not the length of the event which would hold the header.
3534 */
3535int ring_buffer_write(struct trace_buffer *buffer,
3536 unsigned long length,
3537 void *data)
3538{
3539 struct ring_buffer_per_cpu *cpu_buffer;
3540 struct ring_buffer_event *event;
3541 void *body;
3542 int ret = -EBUSY;
3543 int cpu;
3544
3545 preempt_disable_notrace();
3546
3547 if (atomic_read(&buffer->record_disabled))
3548 goto out;
3549
3550 cpu = raw_smp_processor_id();
3551
3552 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3553 goto out;
3554
3555 cpu_buffer = buffer->buffers[cpu];
3556
3557 if (atomic_read(&cpu_buffer->record_disabled))
3558 goto out;
3559
3560 if (length > BUF_MAX_DATA_SIZE)
3561 goto out;
3562
3563 if (unlikely(trace_recursive_lock(cpu_buffer)))
3564 goto out;
3565
3566 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3567 if (!event)
3568 goto out_unlock;
3569
3570 body = rb_event_data(event);
3571
3572 memcpy(body, data, length);
3573
3574 rb_commit(cpu_buffer, event);
3575
3576 rb_wakeups(buffer, cpu_buffer);
3577
3578 ret = 0;
3579
3580 out_unlock:
3581 trace_recursive_unlock(cpu_buffer);
3582
3583 out:
3584 preempt_enable_notrace();
3585
3586 return ret;
3587}
3588EXPORT_SYMBOL_GPL(ring_buffer_write);
3589
3590static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3591{
3592 struct buffer_page *reader = cpu_buffer->reader_page;
3593 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3594 struct buffer_page *commit = cpu_buffer->commit_page;
3595
3596 /* In case of error, head will be NULL */
3597 if (unlikely(!head))
3598 return true;
3599
3600 return reader->read == rb_page_commit(reader) &&
3601 (commit == reader ||
3602 (commit == head &&
3603 head->read == rb_page_commit(commit)));
3604}
3605
3606/**
3607 * ring_buffer_record_disable - stop all writes into the buffer
3608 * @buffer: The ring buffer to stop writes to.
3609 *
3610 * This prevents all writes to the buffer. Any attempt to write
3611 * to the buffer after this will fail and return NULL.
3612 *
3613 * The caller should call synchronize_rcu() after this.
3614 */
3615void ring_buffer_record_disable(struct trace_buffer *buffer)
3616{
3617 atomic_inc(&buffer->record_disabled);
3618}
3619EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3620
3621/**
3622 * ring_buffer_record_enable - enable writes to the buffer
3623 * @buffer: The ring buffer to enable writes
3624 *
3625 * Note, multiple disables will need the same number of enables
3626 * to truly enable the writing (much like preempt_disable).
3627 */
3628void ring_buffer_record_enable(struct trace_buffer *buffer)
3629{
3630 atomic_dec(&buffer->record_disabled);
3631}
3632EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3633
3634/**
3635 * ring_buffer_record_off - stop all writes into the buffer
3636 * @buffer: The ring buffer to stop writes to.
3637 *
3638 * This prevents all writes to the buffer. Any attempt to write
3639 * to the buffer after this will fail and return NULL.
3640 *
3641 * This is different than ring_buffer_record_disable() as
3642 * it works like an on/off switch, where as the disable() version
3643 * must be paired with a enable().
3644 */
3645void ring_buffer_record_off(struct trace_buffer *buffer)
3646{
3647 unsigned int rd;
3648 unsigned int new_rd;
3649
3650 do {
3651 rd = atomic_read(&buffer->record_disabled);
3652 new_rd = rd | RB_BUFFER_OFF;
3653 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3654}
3655EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3656
3657/**
3658 * ring_buffer_record_on - restart writes into the buffer
3659 * @buffer: The ring buffer to start writes to.
3660 *
3661 * This enables all writes to the buffer that was disabled by
3662 * ring_buffer_record_off().
3663 *
3664 * This is different than ring_buffer_record_enable() as
3665 * it works like an on/off switch, where as the enable() version
3666 * must be paired with a disable().
3667 */
3668void ring_buffer_record_on(struct trace_buffer *buffer)
3669{
3670 unsigned int rd;
3671 unsigned int new_rd;
3672
3673 do {
3674 rd = atomic_read(&buffer->record_disabled);
3675 new_rd = rd & ~RB_BUFFER_OFF;
3676 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3677}
3678EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3679
3680/**
3681 * ring_buffer_record_is_on - return true if the ring buffer can write
3682 * @buffer: The ring buffer to see if write is enabled
3683 *
3684 * Returns true if the ring buffer is in a state that it accepts writes.
3685 */
3686bool ring_buffer_record_is_on(struct trace_buffer *buffer)
3687{
3688 return !atomic_read(&buffer->record_disabled);
3689}
3690
3691/**
3692 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
3693 * @buffer: The ring buffer to see if write is set enabled
3694 *
3695 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
3696 * Note that this does NOT mean it is in a writable state.
3697 *
3698 * It may return true when the ring buffer has been disabled by
3699 * ring_buffer_record_disable(), as that is a temporary disabling of
3700 * the ring buffer.
3701 */
3702bool ring_buffer_record_is_set_on(struct trace_buffer *buffer)
3703{
3704 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
3705}
3706
3707/**
3708 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3709 * @buffer: The ring buffer to stop writes to.
3710 * @cpu: The CPU buffer to stop
3711 *
3712 * This prevents all writes to the buffer. Any attempt to write
3713 * to the buffer after this will fail and return NULL.
3714 *
3715 * The caller should call synchronize_rcu() after this.
3716 */
3717void ring_buffer_record_disable_cpu(struct trace_buffer *buffer, int cpu)
3718{
3719 struct ring_buffer_per_cpu *cpu_buffer;
3720
3721 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3722 return;
3723
3724 cpu_buffer = buffer->buffers[cpu];
3725 atomic_inc(&cpu_buffer->record_disabled);
3726}
3727EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3728
3729/**
3730 * ring_buffer_record_enable_cpu - enable writes to the buffer
3731 * @buffer: The ring buffer to enable writes
3732 * @cpu: The CPU to enable.
3733 *
3734 * Note, multiple disables will need the same number of enables
3735 * to truly enable the writing (much like preempt_disable).
3736 */
3737void ring_buffer_record_enable_cpu(struct trace_buffer *buffer, int cpu)
3738{
3739 struct ring_buffer_per_cpu *cpu_buffer;
3740
3741 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3742 return;
3743
3744 cpu_buffer = buffer->buffers[cpu];
3745 atomic_dec(&cpu_buffer->record_disabled);
3746}
3747EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3748
3749/*
3750 * The total entries in the ring buffer is the running counter
3751 * of entries entered into the ring buffer, minus the sum of
3752 * the entries read from the ring buffer and the number of
3753 * entries that were overwritten.
3754 */
3755static inline unsigned long
3756rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3757{
3758 return local_read(&cpu_buffer->entries) -
3759 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3760}
3761
3762/**
3763 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3764 * @buffer: The ring buffer
3765 * @cpu: The per CPU buffer to read from.
3766 */
3767u64 ring_buffer_oldest_event_ts(struct trace_buffer *buffer, int cpu)
3768{
3769 unsigned long flags;
3770 struct ring_buffer_per_cpu *cpu_buffer;
3771 struct buffer_page *bpage;
3772 u64 ret = 0;
3773
3774 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3775 return 0;
3776
3777 cpu_buffer = buffer->buffers[cpu];
3778 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3779 /*
3780 * if the tail is on reader_page, oldest time stamp is on the reader
3781 * page
3782 */
3783 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3784 bpage = cpu_buffer->reader_page;
3785 else
3786 bpage = rb_set_head_page(cpu_buffer);
3787 if (bpage)
3788 ret = bpage->page->time_stamp;
3789 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3790
3791 return ret;
3792}
3793EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3794
3795/**
3796 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3797 * @buffer: The ring buffer
3798 * @cpu: The per CPU buffer to read from.
3799 */
3800unsigned long ring_buffer_bytes_cpu(struct trace_buffer *buffer, int cpu)
3801{
3802 struct ring_buffer_per_cpu *cpu_buffer;
3803 unsigned long ret;
3804
3805 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3806 return 0;
3807
3808 cpu_buffer = buffer->buffers[cpu];
3809 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3810
3811 return ret;
3812}
3813EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3814
3815/**
3816 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3817 * @buffer: The ring buffer
3818 * @cpu: The per CPU buffer to get the entries from.
3819 */
3820unsigned long ring_buffer_entries_cpu(struct trace_buffer *buffer, int cpu)
3821{
3822 struct ring_buffer_per_cpu *cpu_buffer;
3823
3824 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3825 return 0;
3826
3827 cpu_buffer = buffer->buffers[cpu];
3828
3829 return rb_num_of_entries(cpu_buffer);
3830}
3831EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3832
3833/**
3834 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3835 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3836 * @buffer: The ring buffer
3837 * @cpu: The per CPU buffer to get the number of overruns from
3838 */
3839unsigned long ring_buffer_overrun_cpu(struct trace_buffer *buffer, int cpu)
3840{
3841 struct ring_buffer_per_cpu *cpu_buffer;
3842 unsigned long ret;
3843
3844 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3845 return 0;
3846
3847 cpu_buffer = buffer->buffers[cpu];
3848 ret = local_read(&cpu_buffer->overrun);
3849
3850 return ret;
3851}
3852EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3853
3854/**
3855 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3856 * commits failing due to the buffer wrapping around while there are uncommitted
3857 * events, such as during an interrupt storm.
3858 * @buffer: The ring buffer
3859 * @cpu: The per CPU buffer to get the number of overruns from
3860 */
3861unsigned long
3862ring_buffer_commit_overrun_cpu(struct trace_buffer *buffer, int cpu)
3863{
3864 struct ring_buffer_per_cpu *cpu_buffer;
3865 unsigned long ret;
3866
3867 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3868 return 0;
3869
3870 cpu_buffer = buffer->buffers[cpu];
3871 ret = local_read(&cpu_buffer->commit_overrun);
3872
3873 return ret;
3874}
3875EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3876
3877/**
3878 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3879 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3880 * @buffer: The ring buffer
3881 * @cpu: The per CPU buffer to get the number of overruns from
3882 */
3883unsigned long
3884ring_buffer_dropped_events_cpu(struct trace_buffer *buffer, int cpu)
3885{
3886 struct ring_buffer_per_cpu *cpu_buffer;
3887 unsigned long ret;
3888
3889 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3890 return 0;
3891
3892 cpu_buffer = buffer->buffers[cpu];
3893 ret = local_read(&cpu_buffer->dropped_events);
3894
3895 return ret;
3896}
3897EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3898
3899/**
3900 * ring_buffer_read_events_cpu - get the number of events successfully read
3901 * @buffer: The ring buffer
3902 * @cpu: The per CPU buffer to get the number of events read
3903 */
3904unsigned long
3905ring_buffer_read_events_cpu(struct trace_buffer *buffer, int cpu)
3906{
3907 struct ring_buffer_per_cpu *cpu_buffer;
3908
3909 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3910 return 0;
3911
3912 cpu_buffer = buffer->buffers[cpu];
3913 return cpu_buffer->read;
3914}
3915EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3916
3917/**
3918 * ring_buffer_entries - get the number of entries in a buffer
3919 * @buffer: The ring buffer
3920 *
3921 * Returns the total number of entries in the ring buffer
3922 * (all CPU entries)
3923 */
3924unsigned long ring_buffer_entries(struct trace_buffer *buffer)
3925{
3926 struct ring_buffer_per_cpu *cpu_buffer;
3927 unsigned long entries = 0;
3928 int cpu;
3929
3930 /* if you care about this being correct, lock the buffer */
3931 for_each_buffer_cpu(buffer, cpu) {
3932 cpu_buffer = buffer->buffers[cpu];
3933 entries += rb_num_of_entries(cpu_buffer);
3934 }
3935
3936 return entries;
3937}
3938EXPORT_SYMBOL_GPL(ring_buffer_entries);
3939
3940/**
3941 * ring_buffer_overruns - get the number of overruns in buffer
3942 * @buffer: The ring buffer
3943 *
3944 * Returns the total number of overruns in the ring buffer
3945 * (all CPU entries)
3946 */
3947unsigned long ring_buffer_overruns(struct trace_buffer *buffer)
3948{
3949 struct ring_buffer_per_cpu *cpu_buffer;
3950 unsigned long overruns = 0;
3951 int cpu;
3952
3953 /* if you care about this being correct, lock the buffer */
3954 for_each_buffer_cpu(buffer, cpu) {
3955 cpu_buffer = buffer->buffers[cpu];
3956 overruns += local_read(&cpu_buffer->overrun);
3957 }
3958
3959 return overruns;
3960}
3961EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3962
3963static void rb_iter_reset(struct ring_buffer_iter *iter)
3964{
3965 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3966
3967 /* Iterator usage is expected to have record disabled */
3968 iter->head_page = cpu_buffer->reader_page;
3969 iter->head = cpu_buffer->reader_page->read;
3970 iter->next_event = iter->head;
3971
3972 iter->cache_reader_page = iter->head_page;
3973 iter->cache_read = cpu_buffer->read;
3974
3975 if (iter->head) {
3976 iter->read_stamp = cpu_buffer->read_stamp;
3977 iter->page_stamp = cpu_buffer->reader_page->page->time_stamp;
3978 } else {
3979 iter->read_stamp = iter->head_page->page->time_stamp;
3980 iter->page_stamp = iter->read_stamp;
3981 }
3982}
3983
3984/**
3985 * ring_buffer_iter_reset - reset an iterator
3986 * @iter: The iterator to reset
3987 *
3988 * Resets the iterator, so that it will start from the beginning
3989 * again.
3990 */
3991void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3992{
3993 struct ring_buffer_per_cpu *cpu_buffer;
3994 unsigned long flags;
3995
3996 if (!iter)
3997 return;
3998
3999 cpu_buffer = iter->cpu_buffer;
4000
4001 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4002 rb_iter_reset(iter);
4003 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4004}
4005EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
4006
4007/**
4008 * ring_buffer_iter_empty - check if an iterator has no more to read
4009 * @iter: The iterator to check
4010 */
4011int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
4012{
4013 struct ring_buffer_per_cpu *cpu_buffer;
4014 struct buffer_page *reader;
4015 struct buffer_page *head_page;
4016 struct buffer_page *commit_page;
4017 struct buffer_page *curr_commit_page;
4018 unsigned commit;
4019 u64 curr_commit_ts;
4020 u64 commit_ts;
4021
4022 cpu_buffer = iter->cpu_buffer;
4023 reader = cpu_buffer->reader_page;
4024 head_page = cpu_buffer->head_page;
4025 commit_page = cpu_buffer->commit_page;
4026 commit_ts = commit_page->page->time_stamp;
4027
4028 /*
4029 * When the writer goes across pages, it issues a cmpxchg which
4030 * is a mb(), which will synchronize with the rmb here.
4031 * (see rb_tail_page_update())
4032 */
4033 smp_rmb();
4034 commit = rb_page_commit(commit_page);
4035 /* We want to make sure that the commit page doesn't change */
4036 smp_rmb();
4037
4038 /* Make sure commit page didn't change */
4039 curr_commit_page = READ_ONCE(cpu_buffer->commit_page);
4040 curr_commit_ts = READ_ONCE(curr_commit_page->page->time_stamp);
4041
4042 /* If the commit page changed, then there's more data */
4043 if (curr_commit_page != commit_page ||
4044 curr_commit_ts != commit_ts)
4045 return 0;
4046
4047 /* Still racy, as it may return a false positive, but that's OK */
4048 return ((iter->head_page == commit_page && iter->head >= commit) ||
4049 (iter->head_page == reader && commit_page == head_page &&
4050 head_page->read == commit &&
4051 iter->head == rb_page_commit(cpu_buffer->reader_page)));
4052}
4053EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
4054
4055static void
4056rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
4057 struct ring_buffer_event *event)
4058{
4059 u64 delta;
4060
4061 switch (event->type_len) {
4062 case RINGBUF_TYPE_PADDING:
4063 return;
4064
4065 case RINGBUF_TYPE_TIME_EXTEND:
4066 delta = ring_buffer_event_time_stamp(event);
4067 cpu_buffer->read_stamp += delta;
4068 return;
4069
4070 case RINGBUF_TYPE_TIME_STAMP:
4071 delta = ring_buffer_event_time_stamp(event);
4072 cpu_buffer->read_stamp = delta;
4073 return;
4074
4075 case RINGBUF_TYPE_DATA:
4076 cpu_buffer->read_stamp += event->time_delta;
4077 return;
4078
4079 default:
4080 RB_WARN_ON(cpu_buffer, 1);
4081 }
4082 return;
4083}
4084
4085static void
4086rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
4087 struct ring_buffer_event *event)
4088{
4089 u64 delta;
4090
4091 switch (event->type_len) {
4092 case RINGBUF_TYPE_PADDING:
4093 return;
4094
4095 case RINGBUF_TYPE_TIME_EXTEND:
4096 delta = ring_buffer_event_time_stamp(event);
4097 iter->read_stamp += delta;
4098 return;
4099
4100 case RINGBUF_TYPE_TIME_STAMP:
4101 delta = ring_buffer_event_time_stamp(event);
4102 iter->read_stamp = delta;
4103 return;
4104
4105 case RINGBUF_TYPE_DATA:
4106 iter->read_stamp += event->time_delta;
4107 return;
4108
4109 default:
4110 RB_WARN_ON(iter->cpu_buffer, 1);
4111 }
4112 return;
4113}
4114
4115static struct buffer_page *
4116rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
4117{
4118 struct buffer_page *reader = NULL;
4119 unsigned long overwrite;
4120 unsigned long flags;
4121 int nr_loops = 0;
4122 int ret;
4123
4124 local_irq_save(flags);
4125 arch_spin_lock(&cpu_buffer->lock);
4126
4127 again:
4128 /*
4129 * This should normally only loop twice. But because the
4130 * start of the reader inserts an empty page, it causes
4131 * a case where we will loop three times. There should be no
4132 * reason to loop four times (that I know of).
4133 */
4134 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
4135 reader = NULL;
4136 goto out;
4137 }
4138
4139 reader = cpu_buffer->reader_page;
4140
4141 /* If there's more to read, return this page */
4142 if (cpu_buffer->reader_page->read < rb_page_size(reader))
4143 goto out;
4144
4145 /* Never should we have an index greater than the size */
4146 if (RB_WARN_ON(cpu_buffer,
4147 cpu_buffer->reader_page->read > rb_page_size(reader)))
4148 goto out;
4149
4150 /* check if we caught up to the tail */
4151 reader = NULL;
4152 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
4153 goto out;
4154
4155 /* Don't bother swapping if the ring buffer is empty */
4156 if (rb_num_of_entries(cpu_buffer) == 0)
4157 goto out;
4158
4159 /*
4160 * Reset the reader page to size zero.
4161 */
4162 local_set(&cpu_buffer->reader_page->write, 0);
4163 local_set(&cpu_buffer->reader_page->entries, 0);
4164 local_set(&cpu_buffer->reader_page->page->commit, 0);
4165 cpu_buffer->reader_page->real_end = 0;
4166
4167 spin:
4168 /*
4169 * Splice the empty reader page into the list around the head.
4170 */
4171 reader = rb_set_head_page(cpu_buffer);
4172 if (!reader)
4173 goto out;
4174 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
4175 cpu_buffer->reader_page->list.prev = reader->list.prev;
4176
4177 /*
4178 * cpu_buffer->pages just needs to point to the buffer, it
4179 * has no specific buffer page to point to. Lets move it out
4180 * of our way so we don't accidentally swap it.
4181 */
4182 cpu_buffer->pages = reader->list.prev;
4183
4184 /* The reader page will be pointing to the new head */
4185 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
4186
4187 /*
4188 * We want to make sure we read the overruns after we set up our
4189 * pointers to the next object. The writer side does a
4190 * cmpxchg to cross pages which acts as the mb on the writer
4191 * side. Note, the reader will constantly fail the swap
4192 * while the writer is updating the pointers, so this
4193 * guarantees that the overwrite recorded here is the one we
4194 * want to compare with the last_overrun.
4195 */
4196 smp_mb();
4197 overwrite = local_read(&(cpu_buffer->overrun));
4198
4199 /*
4200 * Here's the tricky part.
4201 *
4202 * We need to move the pointer past the header page.
4203 * But we can only do that if a writer is not currently
4204 * moving it. The page before the header page has the
4205 * flag bit '1' set if it is pointing to the page we want.
4206 * but if the writer is in the process of moving it
4207 * than it will be '2' or already moved '0'.
4208 */
4209
4210 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
4211
4212 /*
4213 * If we did not convert it, then we must try again.
4214 */
4215 if (!ret)
4216 goto spin;
4217
4218 /*
4219 * Yay! We succeeded in replacing the page.
4220 *
4221 * Now make the new head point back to the reader page.
4222 */
4223 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
4224 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
4225
4226 local_inc(&cpu_buffer->pages_read);
4227
4228 /* Finally update the reader page to the new head */
4229 cpu_buffer->reader_page = reader;
4230 cpu_buffer->reader_page->read = 0;
4231
4232 if (overwrite != cpu_buffer->last_overrun) {
4233 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
4234 cpu_buffer->last_overrun = overwrite;
4235 }
4236
4237 goto again;
4238
4239 out:
4240 /* Update the read_stamp on the first event */
4241 if (reader && reader->read == 0)
4242 cpu_buffer->read_stamp = reader->page->time_stamp;
4243
4244 arch_spin_unlock(&cpu_buffer->lock);
4245 local_irq_restore(flags);
4246
4247 return reader;
4248}
4249
4250static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
4251{
4252 struct ring_buffer_event *event;
4253 struct buffer_page *reader;
4254 unsigned length;
4255
4256 reader = rb_get_reader_page(cpu_buffer);
4257
4258 /* This function should not be called when buffer is empty */
4259 if (RB_WARN_ON(cpu_buffer, !reader))
4260 return;
4261
4262 event = rb_reader_event(cpu_buffer);
4263
4264 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
4265 cpu_buffer->read++;
4266
4267 rb_update_read_stamp(cpu_buffer, event);
4268
4269 length = rb_event_length(event);
4270 cpu_buffer->reader_page->read += length;
4271}
4272
4273static void rb_advance_iter(struct ring_buffer_iter *iter)
4274{
4275 struct ring_buffer_per_cpu *cpu_buffer;
4276
4277 cpu_buffer = iter->cpu_buffer;
4278
4279 /* If head == next_event then we need to jump to the next event */
4280 if (iter->head == iter->next_event) {
4281 /* If the event gets overwritten again, there's nothing to do */
4282 if (rb_iter_head_event(iter) == NULL)
4283 return;
4284 }
4285
4286 iter->head = iter->next_event;
4287
4288 /*
4289 * Check if we are at the end of the buffer.
4290 */
4291 if (iter->next_event >= rb_page_size(iter->head_page)) {
4292 /* discarded commits can make the page empty */
4293 if (iter->head_page == cpu_buffer->commit_page)
4294 return;
4295 rb_inc_iter(iter);
4296 return;
4297 }
4298
4299 rb_update_iter_read_stamp(iter, iter->event);
4300}
4301
4302static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
4303{
4304 return cpu_buffer->lost_events;
4305}
4306
4307static struct ring_buffer_event *
4308rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
4309 unsigned long *lost_events)
4310{
4311 struct ring_buffer_event *event;
4312 struct buffer_page *reader;
4313 int nr_loops = 0;
4314
4315 if (ts)
4316 *ts = 0;
4317 again:
4318 /*
4319 * We repeat when a time extend is encountered.
4320 * Since the time extend is always attached to a data event,
4321 * we should never loop more than once.
4322 * (We never hit the following condition more than twice).
4323 */
4324 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
4325 return NULL;
4326
4327 reader = rb_get_reader_page(cpu_buffer);
4328 if (!reader)
4329 return NULL;
4330
4331 event = rb_reader_event(cpu_buffer);
4332
4333 switch (event->type_len) {
4334 case RINGBUF_TYPE_PADDING:
4335 if (rb_null_event(event))
4336 RB_WARN_ON(cpu_buffer, 1);
4337 /*
4338 * Because the writer could be discarding every
4339 * event it creates (which would probably be bad)
4340 * if we were to go back to "again" then we may never
4341 * catch up, and will trigger the warn on, or lock
4342 * the box. Return the padding, and we will release
4343 * the current locks, and try again.
4344 */
4345 return event;
4346
4347 case RINGBUF_TYPE_TIME_EXTEND:
4348 /* Internal data, OK to advance */
4349 rb_advance_reader(cpu_buffer);
4350 goto again;
4351
4352 case RINGBUF_TYPE_TIME_STAMP:
4353 if (ts) {
4354 *ts = ring_buffer_event_time_stamp(event);
4355 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4356 cpu_buffer->cpu, ts);
4357 }
4358 /* Internal data, OK to advance */
4359 rb_advance_reader(cpu_buffer);
4360 goto again;
4361
4362 case RINGBUF_TYPE_DATA:
4363 if (ts && !(*ts)) {
4364 *ts = cpu_buffer->read_stamp + event->time_delta;
4365 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4366 cpu_buffer->cpu, ts);
4367 }
4368 if (lost_events)
4369 *lost_events = rb_lost_events(cpu_buffer);
4370 return event;
4371
4372 default:
4373 RB_WARN_ON(cpu_buffer, 1);
4374 }
4375
4376 return NULL;
4377}
4378EXPORT_SYMBOL_GPL(ring_buffer_peek);
4379
4380static struct ring_buffer_event *
4381rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4382{
4383 struct trace_buffer *buffer;
4384 struct ring_buffer_per_cpu *cpu_buffer;
4385 struct ring_buffer_event *event;
4386 int nr_loops = 0;
4387
4388 if (ts)
4389 *ts = 0;
4390
4391 cpu_buffer = iter->cpu_buffer;
4392 buffer = cpu_buffer->buffer;
4393
4394 /*
4395 * Check if someone performed a consuming read to
4396 * the buffer. A consuming read invalidates the iterator
4397 * and we need to reset the iterator in this case.
4398 */
4399 if (unlikely(iter->cache_read != cpu_buffer->read ||
4400 iter->cache_reader_page != cpu_buffer->reader_page))
4401 rb_iter_reset(iter);
4402
4403 again:
4404 if (ring_buffer_iter_empty(iter))
4405 return NULL;
4406
4407 /*
4408 * As the writer can mess with what the iterator is trying
4409 * to read, just give up if we fail to get an event after
4410 * three tries. The iterator is not as reliable when reading
4411 * the ring buffer with an active write as the consumer is.
4412 * Do not warn if the three failures is reached.
4413 */
4414 if (++nr_loops > 3)
4415 return NULL;
4416
4417 if (rb_per_cpu_empty(cpu_buffer))
4418 return NULL;
4419
4420 if (iter->head >= rb_page_size(iter->head_page)) {
4421 rb_inc_iter(iter);
4422 goto again;
4423 }
4424
4425 event = rb_iter_head_event(iter);
4426 if (!event)
4427 goto again;
4428
4429 switch (event->type_len) {
4430 case RINGBUF_TYPE_PADDING:
4431 if (rb_null_event(event)) {
4432 rb_inc_iter(iter);
4433 goto again;
4434 }
4435 rb_advance_iter(iter);
4436 return event;
4437
4438 case RINGBUF_TYPE_TIME_EXTEND:
4439 /* Internal data, OK to advance */
4440 rb_advance_iter(iter);
4441 goto again;
4442
4443 case RINGBUF_TYPE_TIME_STAMP:
4444 if (ts) {
4445 *ts = ring_buffer_event_time_stamp(event);
4446 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4447 cpu_buffer->cpu, ts);
4448 }
4449 /* Internal data, OK to advance */
4450 rb_advance_iter(iter);
4451 goto again;
4452
4453 case RINGBUF_TYPE_DATA:
4454 if (ts && !(*ts)) {
4455 *ts = iter->read_stamp + event->time_delta;
4456 ring_buffer_normalize_time_stamp(buffer,
4457 cpu_buffer->cpu, ts);
4458 }
4459 return event;
4460
4461 default:
4462 RB_WARN_ON(cpu_buffer, 1);
4463 }
4464
4465 return NULL;
4466}
4467EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4468
4469static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4470{
4471 if (likely(!in_nmi())) {
4472 raw_spin_lock(&cpu_buffer->reader_lock);
4473 return true;
4474 }
4475
4476 /*
4477 * If an NMI die dumps out the content of the ring buffer
4478 * trylock must be used to prevent a deadlock if the NMI
4479 * preempted a task that holds the ring buffer locks. If
4480 * we get the lock then all is fine, if not, then continue
4481 * to do the read, but this can corrupt the ring buffer,
4482 * so it must be permanently disabled from future writes.
4483 * Reading from NMI is a oneshot deal.
4484 */
4485 if (raw_spin_trylock(&cpu_buffer->reader_lock))
4486 return true;
4487
4488 /* Continue without locking, but disable the ring buffer */
4489 atomic_inc(&cpu_buffer->record_disabled);
4490 return false;
4491}
4492
4493static inline void
4494rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4495{
4496 if (likely(locked))
4497 raw_spin_unlock(&cpu_buffer->reader_lock);
4498 return;
4499}
4500
4501/**
4502 * ring_buffer_peek - peek at the next event to be read
4503 * @buffer: The ring buffer to read
4504 * @cpu: The cpu to peak at
4505 * @ts: The timestamp counter of this event.
4506 * @lost_events: a variable to store if events were lost (may be NULL)
4507 *
4508 * This will return the event that will be read next, but does
4509 * not consume the data.
4510 */
4511struct ring_buffer_event *
4512ring_buffer_peek(struct trace_buffer *buffer, int cpu, u64 *ts,
4513 unsigned long *lost_events)
4514{
4515 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4516 struct ring_buffer_event *event;
4517 unsigned long flags;
4518 bool dolock;
4519
4520 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4521 return NULL;
4522
4523 again:
4524 local_irq_save(flags);
4525 dolock = rb_reader_lock(cpu_buffer);
4526 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4527 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4528 rb_advance_reader(cpu_buffer);
4529 rb_reader_unlock(cpu_buffer, dolock);
4530 local_irq_restore(flags);
4531
4532 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4533 goto again;
4534
4535 return event;
4536}
4537
4538/** ring_buffer_iter_dropped - report if there are dropped events
4539 * @iter: The ring buffer iterator
4540 *
4541 * Returns true if there was dropped events since the last peek.
4542 */
4543bool ring_buffer_iter_dropped(struct ring_buffer_iter *iter)
4544{
4545 bool ret = iter->missed_events != 0;
4546
4547 iter->missed_events = 0;
4548 return ret;
4549}
4550EXPORT_SYMBOL_GPL(ring_buffer_iter_dropped);
4551
4552/**
4553 * ring_buffer_iter_peek - peek at the next event to be read
4554 * @iter: The ring buffer iterator
4555 * @ts: The timestamp counter of this event.
4556 *
4557 * This will return the event that will be read next, but does
4558 * not increment the iterator.
4559 */
4560struct ring_buffer_event *
4561ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4562{
4563 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4564 struct ring_buffer_event *event;
4565 unsigned long flags;
4566
4567 again:
4568 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4569 event = rb_iter_peek(iter, ts);
4570 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4571
4572 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4573 goto again;
4574
4575 return event;
4576}
4577
4578/**
4579 * ring_buffer_consume - return an event and consume it
4580 * @buffer: The ring buffer to get the next event from
4581 * @cpu: the cpu to read the buffer from
4582 * @ts: a variable to store the timestamp (may be NULL)
4583 * @lost_events: a variable to store if events were lost (may be NULL)
4584 *
4585 * Returns the next event in the ring buffer, and that event is consumed.
4586 * Meaning, that sequential reads will keep returning a different event,
4587 * and eventually empty the ring buffer if the producer is slower.
4588 */
4589struct ring_buffer_event *
4590ring_buffer_consume(struct trace_buffer *buffer, int cpu, u64 *ts,
4591 unsigned long *lost_events)
4592{
4593 struct ring_buffer_per_cpu *cpu_buffer;
4594 struct ring_buffer_event *event = NULL;
4595 unsigned long flags;
4596 bool dolock;
4597
4598 again:
4599 /* might be called in atomic */
4600 preempt_disable();
4601
4602 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4603 goto out;
4604
4605 cpu_buffer = buffer->buffers[cpu];
4606 local_irq_save(flags);
4607 dolock = rb_reader_lock(cpu_buffer);
4608
4609 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4610 if (event) {
4611 cpu_buffer->lost_events = 0;
4612 rb_advance_reader(cpu_buffer);
4613 }
4614
4615 rb_reader_unlock(cpu_buffer, dolock);
4616 local_irq_restore(flags);
4617
4618 out:
4619 preempt_enable();
4620
4621 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4622 goto again;
4623
4624 return event;
4625}
4626EXPORT_SYMBOL_GPL(ring_buffer_consume);
4627
4628/**
4629 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4630 * @buffer: The ring buffer to read from
4631 * @cpu: The cpu buffer to iterate over
4632 * @flags: gfp flags to use for memory allocation
4633 *
4634 * This performs the initial preparations necessary to iterate
4635 * through the buffer. Memory is allocated, buffer recording
4636 * is disabled, and the iterator pointer is returned to the caller.
4637 *
4638 * Disabling buffer recording prevents the reading from being
4639 * corrupted. This is not a consuming read, so a producer is not
4640 * expected.
4641 *
4642 * After a sequence of ring_buffer_read_prepare calls, the user is
4643 * expected to make at least one call to ring_buffer_read_prepare_sync.
4644 * Afterwards, ring_buffer_read_start is invoked to get things going
4645 * for real.
4646 *
4647 * This overall must be paired with ring_buffer_read_finish.
4648 */
4649struct ring_buffer_iter *
4650ring_buffer_read_prepare(struct trace_buffer *buffer, int cpu, gfp_t flags)
4651{
4652 struct ring_buffer_per_cpu *cpu_buffer;
4653 struct ring_buffer_iter *iter;
4654
4655 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4656 return NULL;
4657
4658 iter = kzalloc(sizeof(*iter), flags);
4659 if (!iter)
4660 return NULL;
4661
4662 iter->event = kmalloc(BUF_MAX_DATA_SIZE, flags);
4663 if (!iter->event) {
4664 kfree(iter);
4665 return NULL;
4666 }
4667
4668 cpu_buffer = buffer->buffers[cpu];
4669
4670 iter->cpu_buffer = cpu_buffer;
4671
4672 atomic_inc(&cpu_buffer->resize_disabled);
4673
4674 return iter;
4675}
4676EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4677
4678/**
4679 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4680 *
4681 * All previously invoked ring_buffer_read_prepare calls to prepare
4682 * iterators will be synchronized. Afterwards, read_buffer_read_start
4683 * calls on those iterators are allowed.
4684 */
4685void
4686ring_buffer_read_prepare_sync(void)
4687{
4688 synchronize_rcu();
4689}
4690EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4691
4692/**
4693 * ring_buffer_read_start - start a non consuming read of the buffer
4694 * @iter: The iterator returned by ring_buffer_read_prepare
4695 *
4696 * This finalizes the startup of an iteration through the buffer.
4697 * The iterator comes from a call to ring_buffer_read_prepare and
4698 * an intervening ring_buffer_read_prepare_sync must have been
4699 * performed.
4700 *
4701 * Must be paired with ring_buffer_read_finish.
4702 */
4703void
4704ring_buffer_read_start(struct ring_buffer_iter *iter)
4705{
4706 struct ring_buffer_per_cpu *cpu_buffer;
4707 unsigned long flags;
4708
4709 if (!iter)
4710 return;
4711
4712 cpu_buffer = iter->cpu_buffer;
4713
4714 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4715 arch_spin_lock(&cpu_buffer->lock);
4716 rb_iter_reset(iter);
4717 arch_spin_unlock(&cpu_buffer->lock);
4718 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4719}
4720EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4721
4722/**
4723 * ring_buffer_read_finish - finish reading the iterator of the buffer
4724 * @iter: The iterator retrieved by ring_buffer_start
4725 *
4726 * This re-enables the recording to the buffer, and frees the
4727 * iterator.
4728 */
4729void
4730ring_buffer_read_finish(struct ring_buffer_iter *iter)
4731{
4732 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4733 unsigned long flags;
4734
4735 /*
4736 * Ring buffer is disabled from recording, here's a good place
4737 * to check the integrity of the ring buffer.
4738 * Must prevent readers from trying to read, as the check
4739 * clears the HEAD page and readers require it.
4740 */
4741 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4742 rb_check_pages(cpu_buffer);
4743 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4744
4745 atomic_dec(&cpu_buffer->resize_disabled);
4746 kfree(iter->event);
4747 kfree(iter);
4748}
4749EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4750
4751/**
4752 * ring_buffer_iter_advance - advance the iterator to the next location
4753 * @iter: The ring buffer iterator
4754 *
4755 * Move the location of the iterator such that the next read will
4756 * be the next location of the iterator.
4757 */
4758void ring_buffer_iter_advance(struct ring_buffer_iter *iter)
4759{
4760 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4761 unsigned long flags;
4762
4763 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4764
4765 rb_advance_iter(iter);
4766
4767 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4768}
4769EXPORT_SYMBOL_GPL(ring_buffer_iter_advance);
4770
4771/**
4772 * ring_buffer_size - return the size of the ring buffer (in bytes)
4773 * @buffer: The ring buffer.
4774 * @cpu: The CPU to get ring buffer size from.
4775 */
4776unsigned long ring_buffer_size(struct trace_buffer *buffer, int cpu)
4777{
4778 /*
4779 * Earlier, this method returned
4780 * BUF_PAGE_SIZE * buffer->nr_pages
4781 * Since the nr_pages field is now removed, we have converted this to
4782 * return the per cpu buffer value.
4783 */
4784 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4785 return 0;
4786
4787 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4788}
4789EXPORT_SYMBOL_GPL(ring_buffer_size);
4790
4791static void
4792rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4793{
4794 rb_head_page_deactivate(cpu_buffer);
4795
4796 cpu_buffer->head_page
4797 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4798 local_set(&cpu_buffer->head_page->write, 0);
4799 local_set(&cpu_buffer->head_page->entries, 0);
4800 local_set(&cpu_buffer->head_page->page->commit, 0);
4801
4802 cpu_buffer->head_page->read = 0;
4803
4804 cpu_buffer->tail_page = cpu_buffer->head_page;
4805 cpu_buffer->commit_page = cpu_buffer->head_page;
4806
4807 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4808 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4809 local_set(&cpu_buffer->reader_page->write, 0);
4810 local_set(&cpu_buffer->reader_page->entries, 0);
4811 local_set(&cpu_buffer->reader_page->page->commit, 0);
4812 cpu_buffer->reader_page->read = 0;
4813
4814 local_set(&cpu_buffer->entries_bytes, 0);
4815 local_set(&cpu_buffer->overrun, 0);
4816 local_set(&cpu_buffer->commit_overrun, 0);
4817 local_set(&cpu_buffer->dropped_events, 0);
4818 local_set(&cpu_buffer->entries, 0);
4819 local_set(&cpu_buffer->committing, 0);
4820 local_set(&cpu_buffer->commits, 0);
4821 local_set(&cpu_buffer->pages_touched, 0);
4822 local_set(&cpu_buffer->pages_read, 0);
4823 cpu_buffer->last_pages_touch = 0;
4824 cpu_buffer->shortest_full = 0;
4825 cpu_buffer->read = 0;
4826 cpu_buffer->read_bytes = 0;
4827
4828 rb_time_set(&cpu_buffer->write_stamp, 0);
4829 rb_time_set(&cpu_buffer->before_stamp, 0);
4830
4831 cpu_buffer->lost_events = 0;
4832 cpu_buffer->last_overrun = 0;
4833
4834 rb_head_page_activate(cpu_buffer);
4835}
4836
4837/* Must have disabled the cpu buffer then done a synchronize_rcu */
4838static void reset_disabled_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
4839{
4840 unsigned long flags;
4841
4842 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4843
4844 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4845 goto out;
4846
4847 arch_spin_lock(&cpu_buffer->lock);
4848
4849 rb_reset_cpu(cpu_buffer);
4850
4851 arch_spin_unlock(&cpu_buffer->lock);
4852
4853 out:
4854 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4855}
4856
4857/**
4858 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4859 * @buffer: The ring buffer to reset a per cpu buffer of
4860 * @cpu: The CPU buffer to be reset
4861 */
4862void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
4863{
4864 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4865
4866 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4867 return;
4868
4869 atomic_inc(&cpu_buffer->resize_disabled);
4870 atomic_inc(&cpu_buffer->record_disabled);
4871
4872 /* Make sure all commits have finished */
4873 synchronize_rcu();
4874
4875 reset_disabled_cpu_buffer(cpu_buffer);
4876
4877 atomic_dec(&cpu_buffer->record_disabled);
4878 atomic_dec(&cpu_buffer->resize_disabled);
4879}
4880EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4881
4882/**
4883 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4884 * @buffer: The ring buffer to reset a per cpu buffer of
4885 * @cpu: The CPU buffer to be reset
4886 */
4887void ring_buffer_reset_online_cpus(struct trace_buffer *buffer)
4888{
4889 struct ring_buffer_per_cpu *cpu_buffer;
4890 int cpu;
4891
4892 for_each_online_buffer_cpu(buffer, cpu) {
4893 cpu_buffer = buffer->buffers[cpu];
4894
4895 atomic_inc(&cpu_buffer->resize_disabled);
4896 atomic_inc(&cpu_buffer->record_disabled);
4897 }
4898
4899 /* Make sure all commits have finished */
4900 synchronize_rcu();
4901
4902 for_each_online_buffer_cpu(buffer, cpu) {
4903 cpu_buffer = buffer->buffers[cpu];
4904
4905 reset_disabled_cpu_buffer(cpu_buffer);
4906
4907 atomic_dec(&cpu_buffer->record_disabled);
4908 atomic_dec(&cpu_buffer->resize_disabled);
4909 }
4910}
4911
4912/**
4913 * ring_buffer_reset - reset a ring buffer
4914 * @buffer: The ring buffer to reset all cpu buffers
4915 */
4916void ring_buffer_reset(struct trace_buffer *buffer)
4917{
4918 struct ring_buffer_per_cpu *cpu_buffer;
4919 int cpu;
4920
4921 for_each_buffer_cpu(buffer, cpu) {
4922 cpu_buffer = buffer->buffers[cpu];
4923
4924 atomic_inc(&cpu_buffer->resize_disabled);
4925 atomic_inc(&cpu_buffer->record_disabled);
4926 }
4927
4928 /* Make sure all commits have finished */
4929 synchronize_rcu();
4930
4931 for_each_buffer_cpu(buffer, cpu) {
4932 cpu_buffer = buffer->buffers[cpu];
4933
4934 reset_disabled_cpu_buffer(cpu_buffer);
4935
4936 atomic_dec(&cpu_buffer->record_disabled);
4937 atomic_dec(&cpu_buffer->resize_disabled);
4938 }
4939}
4940EXPORT_SYMBOL_GPL(ring_buffer_reset);
4941
4942/**
4943 * rind_buffer_empty - is the ring buffer empty?
4944 * @buffer: The ring buffer to test
4945 */
4946bool ring_buffer_empty(struct trace_buffer *buffer)
4947{
4948 struct ring_buffer_per_cpu *cpu_buffer;
4949 unsigned long flags;
4950 bool dolock;
4951 int cpu;
4952 int ret;
4953
4954 /* yes this is racy, but if you don't like the race, lock the buffer */
4955 for_each_buffer_cpu(buffer, cpu) {
4956 cpu_buffer = buffer->buffers[cpu];
4957 local_irq_save(flags);
4958 dolock = rb_reader_lock(cpu_buffer);
4959 ret = rb_per_cpu_empty(cpu_buffer);
4960 rb_reader_unlock(cpu_buffer, dolock);
4961 local_irq_restore(flags);
4962
4963 if (!ret)
4964 return false;
4965 }
4966
4967 return true;
4968}
4969EXPORT_SYMBOL_GPL(ring_buffer_empty);
4970
4971/**
4972 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4973 * @buffer: The ring buffer
4974 * @cpu: The CPU buffer to test
4975 */
4976bool ring_buffer_empty_cpu(struct trace_buffer *buffer, int cpu)
4977{
4978 struct ring_buffer_per_cpu *cpu_buffer;
4979 unsigned long flags;
4980 bool dolock;
4981 int ret;
4982
4983 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4984 return true;
4985
4986 cpu_buffer = buffer->buffers[cpu];
4987 local_irq_save(flags);
4988 dolock = rb_reader_lock(cpu_buffer);
4989 ret = rb_per_cpu_empty(cpu_buffer);
4990 rb_reader_unlock(cpu_buffer, dolock);
4991 local_irq_restore(flags);
4992
4993 return ret;
4994}
4995EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4996
4997#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4998/**
4999 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
5000 * @buffer_a: One buffer to swap with
5001 * @buffer_b: The other buffer to swap with
5002 * @cpu: the CPU of the buffers to swap
5003 *
5004 * This function is useful for tracers that want to take a "snapshot"
5005 * of a CPU buffer and has another back up buffer lying around.
5006 * it is expected that the tracer handles the cpu buffer not being
5007 * used at the moment.
5008 */
5009int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
5010 struct trace_buffer *buffer_b, int cpu)
5011{
5012 struct ring_buffer_per_cpu *cpu_buffer_a;
5013 struct ring_buffer_per_cpu *cpu_buffer_b;
5014 int ret = -EINVAL;
5015
5016 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
5017 !cpumask_test_cpu(cpu, buffer_b->cpumask))
5018 goto out;
5019
5020 cpu_buffer_a = buffer_a->buffers[cpu];
5021 cpu_buffer_b = buffer_b->buffers[cpu];
5022
5023 /* At least make sure the two buffers are somewhat the same */
5024 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
5025 goto out;
5026
5027 ret = -EAGAIN;
5028
5029 if (atomic_read(&buffer_a->record_disabled))
5030 goto out;
5031
5032 if (atomic_read(&buffer_b->record_disabled))
5033 goto out;
5034
5035 if (atomic_read(&cpu_buffer_a->record_disabled))
5036 goto out;
5037
5038 if (atomic_read(&cpu_buffer_b->record_disabled))
5039 goto out;
5040
5041 /*
5042 * We can't do a synchronize_rcu here because this
5043 * function can be called in atomic context.
5044 * Normally this will be called from the same CPU as cpu.
5045 * If not it's up to the caller to protect this.
5046 */
5047 atomic_inc(&cpu_buffer_a->record_disabled);
5048 atomic_inc(&cpu_buffer_b->record_disabled);
5049
5050 ret = -EBUSY;
5051 if (local_read(&cpu_buffer_a->committing))
5052 goto out_dec;
5053 if (local_read(&cpu_buffer_b->committing))
5054 goto out_dec;
5055
5056 buffer_a->buffers[cpu] = cpu_buffer_b;
5057 buffer_b->buffers[cpu] = cpu_buffer_a;
5058
5059 cpu_buffer_b->buffer = buffer_a;
5060 cpu_buffer_a->buffer = buffer_b;
5061
5062 ret = 0;
5063
5064out_dec:
5065 atomic_dec(&cpu_buffer_a->record_disabled);
5066 atomic_dec(&cpu_buffer_b->record_disabled);
5067out:
5068 return ret;
5069}
5070EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
5071#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
5072
5073/**
5074 * ring_buffer_alloc_read_page - allocate a page to read from buffer
5075 * @buffer: the buffer to allocate for.
5076 * @cpu: the cpu buffer to allocate.
5077 *
5078 * This function is used in conjunction with ring_buffer_read_page.
5079 * When reading a full page from the ring buffer, these functions
5080 * can be used to speed up the process. The calling function should
5081 * allocate a few pages first with this function. Then when it
5082 * needs to get pages from the ring buffer, it passes the result
5083 * of this function into ring_buffer_read_page, which will swap
5084 * the page that was allocated, with the read page of the buffer.
5085 *
5086 * Returns:
5087 * The page allocated, or ERR_PTR
5088 */
5089void *ring_buffer_alloc_read_page(struct trace_buffer *buffer, int cpu)
5090{
5091 struct ring_buffer_per_cpu *cpu_buffer;
5092 struct buffer_data_page *bpage = NULL;
5093 unsigned long flags;
5094 struct page *page;
5095
5096 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5097 return ERR_PTR(-ENODEV);
5098
5099 cpu_buffer = buffer->buffers[cpu];
5100 local_irq_save(flags);
5101 arch_spin_lock(&cpu_buffer->lock);
5102
5103 if (cpu_buffer->free_page) {
5104 bpage = cpu_buffer->free_page;
5105 cpu_buffer->free_page = NULL;
5106 }
5107
5108 arch_spin_unlock(&cpu_buffer->lock);
5109 local_irq_restore(flags);
5110
5111 if (bpage)
5112 goto out;
5113
5114 page = alloc_pages_node(cpu_to_node(cpu),
5115 GFP_KERNEL | __GFP_NORETRY, 0);
5116 if (!page)
5117 return ERR_PTR(-ENOMEM);
5118
5119 bpage = page_address(page);
5120
5121 out:
5122 rb_init_page(bpage);
5123
5124 return bpage;
5125}
5126EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
5127
5128/**
5129 * ring_buffer_free_read_page - free an allocated read page
5130 * @buffer: the buffer the page was allocate for
5131 * @cpu: the cpu buffer the page came from
5132 * @data: the page to free
5133 *
5134 * Free a page allocated from ring_buffer_alloc_read_page.
5135 */
5136void ring_buffer_free_read_page(struct trace_buffer *buffer, int cpu, void *data)
5137{
5138 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5139 struct buffer_data_page *bpage = data;
5140 struct page *page = virt_to_page(bpage);
5141 unsigned long flags;
5142
5143 /* If the page is still in use someplace else, we can't reuse it */
5144 if (page_ref_count(page) > 1)
5145 goto out;
5146
5147 local_irq_save(flags);
5148 arch_spin_lock(&cpu_buffer->lock);
5149
5150 if (!cpu_buffer->free_page) {
5151 cpu_buffer->free_page = bpage;
5152 bpage = NULL;
5153 }
5154
5155 arch_spin_unlock(&cpu_buffer->lock);
5156 local_irq_restore(flags);
5157
5158 out:
5159 free_page((unsigned long)bpage);
5160}
5161EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
5162
5163/**
5164 * ring_buffer_read_page - extract a page from the ring buffer
5165 * @buffer: buffer to extract from
5166 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
5167 * @len: amount to extract
5168 * @cpu: the cpu of the buffer to extract
5169 * @full: should the extraction only happen when the page is full.
5170 *
5171 * This function will pull out a page from the ring buffer and consume it.
5172 * @data_page must be the address of the variable that was returned
5173 * from ring_buffer_alloc_read_page. This is because the page might be used
5174 * to swap with a page in the ring buffer.
5175 *
5176 * for example:
5177 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
5178 * if (IS_ERR(rpage))
5179 * return PTR_ERR(rpage);
5180 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
5181 * if (ret >= 0)
5182 * process_page(rpage, ret);
5183 *
5184 * When @full is set, the function will not return true unless
5185 * the writer is off the reader page.
5186 *
5187 * Note: it is up to the calling functions to handle sleeps and wakeups.
5188 * The ring buffer can be used anywhere in the kernel and can not
5189 * blindly call wake_up. The layer that uses the ring buffer must be
5190 * responsible for that.
5191 *
5192 * Returns:
5193 * >=0 if data has been transferred, returns the offset of consumed data.
5194 * <0 if no data has been transferred.
5195 */
5196int ring_buffer_read_page(struct trace_buffer *buffer,
5197 void **data_page, size_t len, int cpu, int full)
5198{
5199 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5200 struct ring_buffer_event *event;
5201 struct buffer_data_page *bpage;
5202 struct buffer_page *reader;
5203 unsigned long missed_events;
5204 unsigned long flags;
5205 unsigned int commit;
5206 unsigned int read;
5207 u64 save_timestamp;
5208 int ret = -1;
5209
5210 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5211 goto out;
5212
5213 /*
5214 * If len is not big enough to hold the page header, then
5215 * we can not copy anything.
5216 */
5217 if (len <= BUF_PAGE_HDR_SIZE)
5218 goto out;
5219
5220 len -= BUF_PAGE_HDR_SIZE;
5221
5222 if (!data_page)
5223 goto out;
5224
5225 bpage = *data_page;
5226 if (!bpage)
5227 goto out;
5228
5229 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5230
5231 reader = rb_get_reader_page(cpu_buffer);
5232 if (!reader)
5233 goto out_unlock;
5234
5235 event = rb_reader_event(cpu_buffer);
5236
5237 read = reader->read;
5238 commit = rb_page_commit(reader);
5239
5240 /* Check if any events were dropped */
5241 missed_events = cpu_buffer->lost_events;
5242
5243 /*
5244 * If this page has been partially read or
5245 * if len is not big enough to read the rest of the page or
5246 * a writer is still on the page, then
5247 * we must copy the data from the page to the buffer.
5248 * Otherwise, we can simply swap the page with the one passed in.
5249 */
5250 if (read || (len < (commit - read)) ||
5251 cpu_buffer->reader_page == cpu_buffer->commit_page) {
5252 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
5253 unsigned int rpos = read;
5254 unsigned int pos = 0;
5255 unsigned int size;
5256
5257 if (full)
5258 goto out_unlock;
5259
5260 if (len > (commit - read))
5261 len = (commit - read);
5262
5263 /* Always keep the time extend and data together */
5264 size = rb_event_ts_length(event);
5265
5266 if (len < size)
5267 goto out_unlock;
5268
5269 /* save the current timestamp, since the user will need it */
5270 save_timestamp = cpu_buffer->read_stamp;
5271
5272 /* Need to copy one event at a time */
5273 do {
5274 /* We need the size of one event, because
5275 * rb_advance_reader only advances by one event,
5276 * whereas rb_event_ts_length may include the size of
5277 * one or two events.
5278 * We have already ensured there's enough space if this
5279 * is a time extend. */
5280 size = rb_event_length(event);
5281 memcpy(bpage->data + pos, rpage->data + rpos, size);
5282
5283 len -= size;
5284
5285 rb_advance_reader(cpu_buffer);
5286 rpos = reader->read;
5287 pos += size;
5288
5289 if (rpos >= commit)
5290 break;
5291
5292 event = rb_reader_event(cpu_buffer);
5293 /* Always keep the time extend and data together */
5294 size = rb_event_ts_length(event);
5295 } while (len >= size);
5296
5297 /* update bpage */
5298 local_set(&bpage->commit, pos);
5299 bpage->time_stamp = save_timestamp;
5300
5301 /* we copied everything to the beginning */
5302 read = 0;
5303 } else {
5304 /* update the entry counter */
5305 cpu_buffer->read += rb_page_entries(reader);
5306 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
5307
5308 /* swap the pages */
5309 rb_init_page(bpage);
5310 bpage = reader->page;
5311 reader->page = *data_page;
5312 local_set(&reader->write, 0);
5313 local_set(&reader->entries, 0);
5314 reader->read = 0;
5315 *data_page = bpage;
5316
5317 /*
5318 * Use the real_end for the data size,
5319 * This gives us a chance to store the lost events
5320 * on the page.
5321 */
5322 if (reader->real_end)
5323 local_set(&bpage->commit, reader->real_end);
5324 }
5325 ret = read;
5326
5327 cpu_buffer->lost_events = 0;
5328
5329 commit = local_read(&bpage->commit);
5330 /*
5331 * Set a flag in the commit field if we lost events
5332 */
5333 if (missed_events) {
5334 /* If there is room at the end of the page to save the
5335 * missed events, then record it there.
5336 */
5337 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
5338 memcpy(&bpage->data[commit], &missed_events,
5339 sizeof(missed_events));
5340 local_add(RB_MISSED_STORED, &bpage->commit);
5341 commit += sizeof(missed_events);
5342 }
5343 local_add(RB_MISSED_EVENTS, &bpage->commit);
5344 }
5345
5346 /*
5347 * This page may be off to user land. Zero it out here.
5348 */
5349 if (commit < BUF_PAGE_SIZE)
5350 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
5351
5352 out_unlock:
5353 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5354
5355 out:
5356 return ret;
5357}
5358EXPORT_SYMBOL_GPL(ring_buffer_read_page);
5359
5360/*
5361 * We only allocate new buffers, never free them if the CPU goes down.
5362 * If we were to free the buffer, then the user would lose any trace that was in
5363 * the buffer.
5364 */
5365int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
5366{
5367 struct trace_buffer *buffer;
5368 long nr_pages_same;
5369 int cpu_i;
5370 unsigned long nr_pages;
5371
5372 buffer = container_of(node, struct trace_buffer, node);
5373 if (cpumask_test_cpu(cpu, buffer->cpumask))
5374 return 0;
5375
5376 nr_pages = 0;
5377 nr_pages_same = 1;
5378 /* check if all cpu sizes are same */
5379 for_each_buffer_cpu(buffer, cpu_i) {
5380 /* fill in the size from first enabled cpu */
5381 if (nr_pages == 0)
5382 nr_pages = buffer->buffers[cpu_i]->nr_pages;
5383 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
5384 nr_pages_same = 0;
5385 break;
5386 }
5387 }
5388 /* allocate minimum pages, user can later expand it */
5389 if (!nr_pages_same)
5390 nr_pages = 2;
5391 buffer->buffers[cpu] =
5392 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
5393 if (!buffer->buffers[cpu]) {
5394 WARN(1, "failed to allocate ring buffer on CPU %u\n",
5395 cpu);
5396 return -ENOMEM;
5397 }
5398 smp_wmb();
5399 cpumask_set_cpu(cpu, buffer->cpumask);
5400 return 0;
5401}
5402
5403#ifdef CONFIG_RING_BUFFER_STARTUP_TEST
5404/*
5405 * This is a basic integrity check of the ring buffer.
5406 * Late in the boot cycle this test will run when configured in.
5407 * It will kick off a thread per CPU that will go into a loop
5408 * writing to the per cpu ring buffer various sizes of data.
5409 * Some of the data will be large items, some small.
5410 *
5411 * Another thread is created that goes into a spin, sending out
5412 * IPIs to the other CPUs to also write into the ring buffer.
5413 * this is to test the nesting ability of the buffer.
5414 *
5415 * Basic stats are recorded and reported. If something in the
5416 * ring buffer should happen that's not expected, a big warning
5417 * is displayed and all ring buffers are disabled.
5418 */
5419static struct task_struct *rb_threads[NR_CPUS] __initdata;
5420
5421struct rb_test_data {
5422 struct trace_buffer *buffer;
5423 unsigned long events;
5424 unsigned long bytes_written;
5425 unsigned long bytes_alloc;
5426 unsigned long bytes_dropped;
5427 unsigned long events_nested;
5428 unsigned long bytes_written_nested;
5429 unsigned long bytes_alloc_nested;
5430 unsigned long bytes_dropped_nested;
5431 int min_size_nested;
5432 int max_size_nested;
5433 int max_size;
5434 int min_size;
5435 int cpu;
5436 int cnt;
5437};
5438
5439static struct rb_test_data rb_data[NR_CPUS] __initdata;
5440
5441/* 1 meg per cpu */
5442#define RB_TEST_BUFFER_SIZE 1048576
5443
5444static char rb_string[] __initdata =
5445 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
5446 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
5447 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
5448
5449static bool rb_test_started __initdata;
5450
5451struct rb_item {
5452 int size;
5453 char str[];
5454};
5455
5456static __init int rb_write_something(struct rb_test_data *data, bool nested)
5457{
5458 struct ring_buffer_event *event;
5459 struct rb_item *item;
5460 bool started;
5461 int event_len;
5462 int size;
5463 int len;
5464 int cnt;
5465
5466 /* Have nested writes different that what is written */
5467 cnt = data->cnt + (nested ? 27 : 0);
5468
5469 /* Multiply cnt by ~e, to make some unique increment */
5470 size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
5471
5472 len = size + sizeof(struct rb_item);
5473
5474 started = rb_test_started;
5475 /* read rb_test_started before checking buffer enabled */
5476 smp_rmb();
5477
5478 event = ring_buffer_lock_reserve(data->buffer, len);
5479 if (!event) {
5480 /* Ignore dropped events before test starts. */
5481 if (started) {
5482 if (nested)
5483 data->bytes_dropped += len;
5484 else
5485 data->bytes_dropped_nested += len;
5486 }
5487 return len;
5488 }
5489
5490 event_len = ring_buffer_event_length(event);
5491
5492 if (RB_WARN_ON(data->buffer, event_len < len))
5493 goto out;
5494
5495 item = ring_buffer_event_data(event);
5496 item->size = size;
5497 memcpy(item->str, rb_string, size);
5498
5499 if (nested) {
5500 data->bytes_alloc_nested += event_len;
5501 data->bytes_written_nested += len;
5502 data->events_nested++;
5503 if (!data->min_size_nested || len < data->min_size_nested)
5504 data->min_size_nested = len;
5505 if (len > data->max_size_nested)
5506 data->max_size_nested = len;
5507 } else {
5508 data->bytes_alloc += event_len;
5509 data->bytes_written += len;
5510 data->events++;
5511 if (!data->min_size || len < data->min_size)
5512 data->max_size = len;
5513 if (len > data->max_size)
5514 data->max_size = len;
5515 }
5516
5517 out:
5518 ring_buffer_unlock_commit(data->buffer, event);
5519
5520 return 0;
5521}
5522
5523static __init int rb_test(void *arg)
5524{
5525 struct rb_test_data *data = arg;
5526
5527 while (!kthread_should_stop()) {
5528 rb_write_something(data, false);
5529 data->cnt++;
5530
5531 set_current_state(TASK_INTERRUPTIBLE);
5532 /* Now sleep between a min of 100-300us and a max of 1ms */
5533 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
5534 }
5535
5536 return 0;
5537}
5538
5539static __init void rb_ipi(void *ignore)
5540{
5541 struct rb_test_data *data;
5542 int cpu = smp_processor_id();
5543
5544 data = &rb_data[cpu];
5545 rb_write_something(data, true);
5546}
5547
5548static __init int rb_hammer_test(void *arg)
5549{
5550 while (!kthread_should_stop()) {
5551
5552 /* Send an IPI to all cpus to write data! */
5553 smp_call_function(rb_ipi, NULL, 1);
5554 /* No sleep, but for non preempt, let others run */
5555 schedule();
5556 }
5557
5558 return 0;
5559}
5560
5561static __init int test_ringbuffer(void)
5562{
5563 struct task_struct *rb_hammer;
5564 struct trace_buffer *buffer;
5565 int cpu;
5566 int ret = 0;
5567
5568 if (security_locked_down(LOCKDOWN_TRACEFS)) {
5569 pr_warn("Lockdown is enabled, skipping ring buffer tests\n");
5570 return 0;
5571 }
5572
5573 pr_info("Running ring buffer tests...\n");
5574
5575 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
5576 if (WARN_ON(!buffer))
5577 return 0;
5578
5579 /* Disable buffer so that threads can't write to it yet */
5580 ring_buffer_record_off(buffer);
5581
5582 for_each_online_cpu(cpu) {
5583 rb_data[cpu].buffer = buffer;
5584 rb_data[cpu].cpu = cpu;
5585 rb_data[cpu].cnt = cpu;
5586 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
5587 "rbtester/%d", cpu);
5588 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
5589 pr_cont("FAILED\n");
5590 ret = PTR_ERR(rb_threads[cpu]);
5591 goto out_free;
5592 }
5593
5594 kthread_bind(rb_threads[cpu], cpu);
5595 wake_up_process(rb_threads[cpu]);
5596 }
5597
5598 /* Now create the rb hammer! */
5599 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
5600 if (WARN_ON(IS_ERR(rb_hammer))) {
5601 pr_cont("FAILED\n");
5602 ret = PTR_ERR(rb_hammer);
5603 goto out_free;
5604 }
5605
5606 ring_buffer_record_on(buffer);
5607 /*
5608 * Show buffer is enabled before setting rb_test_started.
5609 * Yes there's a small race window where events could be
5610 * dropped and the thread wont catch it. But when a ring
5611 * buffer gets enabled, there will always be some kind of
5612 * delay before other CPUs see it. Thus, we don't care about
5613 * those dropped events. We care about events dropped after
5614 * the threads see that the buffer is active.
5615 */
5616 smp_wmb();
5617 rb_test_started = true;
5618
5619 set_current_state(TASK_INTERRUPTIBLE);
5620 /* Just run for 10 seconds */;
5621 schedule_timeout(10 * HZ);
5622
5623 kthread_stop(rb_hammer);
5624
5625 out_free:
5626 for_each_online_cpu(cpu) {
5627 if (!rb_threads[cpu])
5628 break;
5629 kthread_stop(rb_threads[cpu]);
5630 }
5631 if (ret) {
5632 ring_buffer_free(buffer);
5633 return ret;
5634 }
5635
5636 /* Report! */
5637 pr_info("finished\n");
5638 for_each_online_cpu(cpu) {
5639 struct ring_buffer_event *event;
5640 struct rb_test_data *data = &rb_data[cpu];
5641 struct rb_item *item;
5642 unsigned long total_events;
5643 unsigned long total_dropped;
5644 unsigned long total_written;
5645 unsigned long total_alloc;
5646 unsigned long total_read = 0;
5647 unsigned long total_size = 0;
5648 unsigned long total_len = 0;
5649 unsigned long total_lost = 0;
5650 unsigned long lost;
5651 int big_event_size;
5652 int small_event_size;
5653
5654 ret = -1;
5655
5656 total_events = data->events + data->events_nested;
5657 total_written = data->bytes_written + data->bytes_written_nested;
5658 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
5659 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
5660
5661 big_event_size = data->max_size + data->max_size_nested;
5662 small_event_size = data->min_size + data->min_size_nested;
5663
5664 pr_info("CPU %d:\n", cpu);
5665 pr_info(" events: %ld\n", total_events);
5666 pr_info(" dropped bytes: %ld\n", total_dropped);
5667 pr_info(" alloced bytes: %ld\n", total_alloc);
5668 pr_info(" written bytes: %ld\n", total_written);
5669 pr_info(" biggest event: %d\n", big_event_size);
5670 pr_info(" smallest event: %d\n", small_event_size);
5671
5672 if (RB_WARN_ON(buffer, total_dropped))
5673 break;
5674
5675 ret = 0;
5676
5677 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
5678 total_lost += lost;
5679 item = ring_buffer_event_data(event);
5680 total_len += ring_buffer_event_length(event);
5681 total_size += item->size + sizeof(struct rb_item);
5682 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
5683 pr_info("FAILED!\n");
5684 pr_info("buffer had: %.*s\n", item->size, item->str);
5685 pr_info("expected: %.*s\n", item->size, rb_string);
5686 RB_WARN_ON(buffer, 1);
5687 ret = -1;
5688 break;
5689 }
5690 total_read++;
5691 }
5692 if (ret)
5693 break;
5694
5695 ret = -1;
5696
5697 pr_info(" read events: %ld\n", total_read);
5698 pr_info(" lost events: %ld\n", total_lost);
5699 pr_info(" total events: %ld\n", total_lost + total_read);
5700 pr_info(" recorded len bytes: %ld\n", total_len);
5701 pr_info(" recorded size bytes: %ld\n", total_size);
5702 if (total_lost)
5703 pr_info(" With dropped events, record len and size may not match\n"
5704 " alloced and written from above\n");
5705 if (!total_lost) {
5706 if (RB_WARN_ON(buffer, total_len != total_alloc ||
5707 total_size != total_written))
5708 break;
5709 }
5710 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5711 break;
5712
5713 ret = 0;
5714 }
5715 if (!ret)
5716 pr_info("Ring buffer PASSED!\n");
5717
5718 ring_buffer_free(buffer);
5719 return 0;
5720}
5721
5722late_initcall(test_ringbuffer);
5723#endif /* CONFIG_RING_BUFFER_STARTUP_TEST */