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