projects/linux·blob 4563243e · refs

7593 lines · 201.5 KiB

1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Generic ring buffer
4 *
5 * Copyright (C) 2008 Steven Rostedt <[email protected]>
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/cacheflush.h>
13#include <linux/trace_seq.h>
14#include <linux/spinlock.h>
15#include <linux/irq_work.h>
16#include <linux/security.h>
17#include <linux/uaccess.h>
18#include <linux/hardirq.h>
19#include <linux/kthread.h> /* for self test */
20#include <linux/module.h>
21#include <linux/percpu.h>
22#include <linux/mutex.h>
23#include <linux/delay.h>
24#include <linux/slab.h>
25#include <linux/init.h>
26#include <linux/hash.h>
27#include <linux/list.h>
28#include <linux/cpu.h>
29#include <linux/oom.h>
30#include <linux/mm.h>
31
32#include <asm/local64.h>
33#include <asm/local.h>
34
35#include "trace.h"
36
37/*
38 * The "absolute" timestamp in the buffer is only 59 bits.
39 * If a clock has the 5 MSBs set, it needs to be saved and
40 * reinserted.
41 */
42#define TS_MSB (0xf8ULL << 56)
43#define ABS_TS_MASK (~TS_MSB)
44
45static void update_pages_handler(struct work_struct *work);
46
47#define RING_BUFFER_META_MAGIC 0xBADFEED
48
49struct ring_buffer_meta {
50 int magic;
51 int struct_size;
52 unsigned long text_addr;
53 unsigned long data_addr;
54 unsigned long first_buffer;
55 unsigned long head_buffer;
56 unsigned long commit_buffer;
57 __u32 subbuf_size;
58 __u32 nr_subbufs;
59 int buffers[];
60};
61
62/*
63 * The ring buffer header is special. We must manually up keep it.
64 */
65int ring_buffer_print_entry_header(struct trace_seq *s)
66{
67 trace_seq_puts(s, "# compressed entry header\n");
68 trace_seq_puts(s, "\ttype_len : 5 bits\n");
69 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
70 trace_seq_puts(s, "\tarray : 32 bits\n");
71 trace_seq_putc(s, '\n');
72 trace_seq_printf(s, "\tpadding : type == %d\n",
73 RINGBUF_TYPE_PADDING);
74 trace_seq_printf(s, "\ttime_extend : type == %d\n",
75 RINGBUF_TYPE_TIME_EXTEND);
76 trace_seq_printf(s, "\ttime_stamp : type == %d\n",
77 RINGBUF_TYPE_TIME_STAMP);
78 trace_seq_printf(s, "\tdata max type_len == %d\n",
79 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
80
81 return !trace_seq_has_overflowed(s);
82}
83
84/*
85 * The ring buffer is made up of a list of pages. A separate list of pages is
86 * allocated for each CPU. A writer may only write to a buffer that is
87 * associated with the CPU it is currently executing on. A reader may read
88 * from any per cpu buffer.
89 *
90 * The reader is special. For each per cpu buffer, the reader has its own
91 * reader page. When a reader has read the entire reader page, this reader
92 * page is swapped with another page in the ring buffer.
93 *
94 * Now, as long as the writer is off the reader page, the reader can do what
95 * ever it wants with that page. The writer will never write to that page
96 * again (as long as it is out of the ring buffer).
97 *
98 * Here's some silly ASCII art.
99 *
100 * +------+
101 * |reader| RING BUFFER
102 * |page |
103 * +------+ +---+ +---+ +---+
104 * | |-->| |-->| |
105 * +---+ +---+ +---+
106 * ^ |
107 * | |
108 * +---------------+
109 *
110 *
111 * +------+
112 * |reader| RING BUFFER
113 * |page |------------------v
114 * +------+ +---+ +---+ +---+
115 * | |-->| |-->| |
116 * +---+ +---+ +---+
117 * ^ |
118 * | |
119 * +---------------+
120 *
121 *
122 * +------+
123 * |reader| RING BUFFER
124 * |page |------------------v
125 * +------+ +---+ +---+ +---+
126 * ^ | |-->| |-->| |
127 * | +---+ +---+ +---+
128 * | |
129 * | |
130 * +------------------------------+
131 *
132 *
133 * +------+
134 * |buffer| RING BUFFER
135 * |page |------------------v
136 * +------+ +---+ +---+ +---+
137 * ^ | | | |-->| |
138 * | New +---+ +---+ +---+
139 * | Reader------^ |
140 * | page |
141 * +------------------------------+
142 *
143 *
144 * After we make this swap, the reader can hand this page off to the splice
145 * code and be done with it. It can even allocate a new page if it needs to
146 * and swap that into the ring buffer.
147 *
148 * We will be using cmpxchg soon to make all this lockless.
149 *
150 */
151
152/* Used for individual buffers (after the counter) */
153#define RB_BUFFER_OFF (1 << 20)
154
155#define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
156
157#define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
158#define RB_ALIGNMENT 4U
159#define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
160#define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
161
162#ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
163# define RB_FORCE_8BYTE_ALIGNMENT 0
164# define RB_ARCH_ALIGNMENT RB_ALIGNMENT
165#else
166# define RB_FORCE_8BYTE_ALIGNMENT 1
167# define RB_ARCH_ALIGNMENT 8U
168#endif
169
170#define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
171
172/* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
173#define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
174
175enum {
176 RB_LEN_TIME_EXTEND = 8,
177 RB_LEN_TIME_STAMP = 8,
178};
179
180#define skip_time_extend(event) \
181 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
182
183#define extended_time(event) \
184 (event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
185
186static inline bool rb_null_event(struct ring_buffer_event *event)
187{
188 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
189}
190
191static void rb_event_set_padding(struct ring_buffer_event *event)
192{
193 /* padding has a NULL time_delta */
194 event->type_len = RINGBUF_TYPE_PADDING;
195 event->time_delta = 0;
196}
197
198static unsigned
199rb_event_data_length(struct ring_buffer_event *event)
200{
201 unsigned length;
202
203 if (event->type_len)
204 length = event->type_len * RB_ALIGNMENT;
205 else
206 length = event->array[0];
207 return length + RB_EVNT_HDR_SIZE;
208}
209
210/*
211 * Return the length of the given event. Will return
212 * the length of the time extend if the event is a
213 * time extend.
214 */
215static inline unsigned
216rb_event_length(struct ring_buffer_event *event)
217{
218 switch (event->type_len) {
219 case RINGBUF_TYPE_PADDING:
220 if (rb_null_event(event))
221 /* undefined */
222 return -1;
223 return event->array[0] + RB_EVNT_HDR_SIZE;
224
225 case RINGBUF_TYPE_TIME_EXTEND:
226 return RB_LEN_TIME_EXTEND;
227
228 case RINGBUF_TYPE_TIME_STAMP:
229 return RB_LEN_TIME_STAMP;
230
231 case RINGBUF_TYPE_DATA:
232 return rb_event_data_length(event);
233 default:
234 WARN_ON_ONCE(1);
235 }
236 /* not hit */
237 return 0;
238}
239
240/*
241 * Return total length of time extend and data,
242 * or just the event length for all other events.
243 */
244static inline unsigned
245rb_event_ts_length(struct ring_buffer_event *event)
246{
247 unsigned len = 0;
248
249 if (extended_time(event)) {
250 /* time extends include the data event after it */
251 len = RB_LEN_TIME_EXTEND;
252 event = skip_time_extend(event);
253 }
254 return len + rb_event_length(event);
255}
256
257/**
258 * ring_buffer_event_length - return the length of the event
259 * @event: the event to get the length of
260 *
261 * Returns the size of the data load of a data event.
262 * If the event is something other than a data event, it
263 * returns the size of the event itself. With the exception
264 * of a TIME EXTEND, where it still returns the size of the
265 * data load of the data event after it.
266 */
267unsigned ring_buffer_event_length(struct ring_buffer_event *event)
268{
269 unsigned length;
270
271 if (extended_time(event))
272 event = skip_time_extend(event);
273
274 length = rb_event_length(event);
275 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
276 return length;
277 length -= RB_EVNT_HDR_SIZE;
278 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
279 length -= sizeof(event->array[0]);
280 return length;
281}
282EXPORT_SYMBOL_GPL(ring_buffer_event_length);
283
284/* inline for ring buffer fast paths */
285static __always_inline void *
286rb_event_data(struct ring_buffer_event *event)
287{
288 if (extended_time(event))
289 event = skip_time_extend(event);
290 WARN_ON_ONCE(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
291 /* If length is in len field, then array[0] has the data */
292 if (event->type_len)
293 return (void *)&event->array[0];
294 /* Otherwise length is in array[0] and array[1] has the data */
295 return (void *)&event->array[1];
296}
297
298/**
299 * ring_buffer_event_data - return the data of the event
300 * @event: the event to get the data from
301 */
302void *ring_buffer_event_data(struct ring_buffer_event *event)
303{
304 return rb_event_data(event);
305}
306EXPORT_SYMBOL_GPL(ring_buffer_event_data);
307
308#define for_each_buffer_cpu(buffer, cpu) \
309 for_each_cpu(cpu, buffer->cpumask)
310
311#define for_each_online_buffer_cpu(buffer, cpu) \
312 for_each_cpu_and(cpu, buffer->cpumask, cpu_online_mask)
313
314#define TS_SHIFT 27
315#define TS_MASK ((1ULL << TS_SHIFT) - 1)
316#define TS_DELTA_TEST (~TS_MASK)
317
318static u64 rb_event_time_stamp(struct ring_buffer_event *event)
319{
320 u64 ts;
321
322 ts = event->array[0];
323 ts <<= TS_SHIFT;
324 ts += event->time_delta;
325
326 return ts;
327}
328
329/* Flag when events were overwritten */
330#define RB_MISSED_EVENTS (1 << 31)
331/* Missed count stored at end */
332#define RB_MISSED_STORED (1 << 30)
333
334#define RB_MISSED_MASK (3 << 30)
335
336struct buffer_data_page {
337 u64 time_stamp; /* page time stamp */
338 local_t commit; /* write committed index */
339 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
340};
341
342struct buffer_data_read_page {
343 unsigned order; /* order of the page */
344 struct buffer_data_page *data; /* actual data, stored in this page */
345};
346
347/*
348 * Note, the buffer_page list must be first. The buffer pages
349 * are allocated in cache lines, which means that each buffer
350 * page will be at the beginning of a cache line, and thus
351 * the least significant bits will be zero. We use this to
352 * add flags in the list struct pointers, to make the ring buffer
353 * lockless.
354 */
355struct buffer_page {
356 struct list_head list; /* list of buffer pages */
357 local_t write; /* index for next write */
358 unsigned read; /* index for next read */
359 local_t entries; /* entries on this page */
360 unsigned long real_end; /* real end of data */
361 unsigned order; /* order of the page */
362 u32 id:30; /* ID for external mapping */
363 u32 range:1; /* Mapped via a range */
364 struct buffer_data_page *page; /* Actual data page */
365};
366
367/*
368 * The buffer page counters, write and entries, must be reset
369 * atomically when crossing page boundaries. To synchronize this
370 * update, two counters are inserted into the number. One is
371 * the actual counter for the write position or count on the page.
372 *
373 * The other is a counter of updaters. Before an update happens
374 * the update partition of the counter is incremented. This will
375 * allow the updater to update the counter atomically.
376 *
377 * The counter is 20 bits, and the state data is 12.
378 */
379#define RB_WRITE_MASK 0xfffff
380#define RB_WRITE_INTCNT (1 << 20)
381
382static void rb_init_page(struct buffer_data_page *bpage)
383{
384 local_set(&bpage->commit, 0);
385}
386
387static __always_inline unsigned int rb_page_commit(struct buffer_page *bpage)
388{
389 return local_read(&bpage->page->commit);
390}
391
392static void free_buffer_page(struct buffer_page *bpage)
393{
394 /* Range pages are not to be freed */
395 if (!bpage->range)
396 free_pages((unsigned long)bpage->page, bpage->order);
397 kfree(bpage);
398}
399
400/*
401 * We need to fit the time_stamp delta into 27 bits.
402 */
403static inline bool test_time_stamp(u64 delta)
404{
405 return !!(delta & TS_DELTA_TEST);
406}
407
408struct rb_irq_work {
409 struct irq_work work;
410 wait_queue_head_t waiters;
411 wait_queue_head_t full_waiters;
412 atomic_t seq;
413 bool waiters_pending;
414 bool full_waiters_pending;
415 bool wakeup_full;
416};
417
418/*
419 * Structure to hold event state and handle nested events.
420 */
421struct rb_event_info {
422 u64 ts;
423 u64 delta;
424 u64 before;
425 u64 after;
426 unsigned long length;
427 struct buffer_page *tail_page;
428 int add_timestamp;
429};
430
431/*
432 * Used for the add_timestamp
433 * NONE
434 * EXTEND - wants a time extend
435 * ABSOLUTE - the buffer requests all events to have absolute time stamps
436 * FORCE - force a full time stamp.
437 */
438enum {
439 RB_ADD_STAMP_NONE = 0,
440 RB_ADD_STAMP_EXTEND = BIT(1),
441 RB_ADD_STAMP_ABSOLUTE = BIT(2),
442 RB_ADD_STAMP_FORCE = BIT(3)
443};
444/*
445 * Used for which event context the event is in.
446 * TRANSITION = 0
447 * NMI = 1
448 * IRQ = 2
449 * SOFTIRQ = 3
450 * NORMAL = 4
451 *
452 * See trace_recursive_lock() comment below for more details.
453 */
454enum {
455 RB_CTX_TRANSITION,
456 RB_CTX_NMI,
457 RB_CTX_IRQ,
458 RB_CTX_SOFTIRQ,
459 RB_CTX_NORMAL,
460 RB_CTX_MAX
461};
462
463struct rb_time_struct {
464 local64_t time;
465};
466typedef struct rb_time_struct rb_time_t;
467
468#define MAX_NEST 5
469
470/*
471 * head_page == tail_page && head == tail then buffer is empty.
472 */
473struct ring_buffer_per_cpu {
474 int cpu;
475 atomic_t record_disabled;
476 atomic_t resize_disabled;
477 struct trace_buffer *buffer;
478 raw_spinlock_t reader_lock; /* serialize readers */
479 arch_spinlock_t lock;
480 struct lock_class_key lock_key;
481 struct buffer_data_page *free_page;
482 unsigned long nr_pages;
483 unsigned int current_context;
484 struct list_head *pages;
485 struct buffer_page *head_page; /* read from head */
486 struct buffer_page *tail_page; /* write to tail */
487 struct buffer_page *commit_page; /* committed pages */
488 struct buffer_page *reader_page;
489 unsigned long lost_events;
490 unsigned long last_overrun;
491 unsigned long nest;
492 local_t entries_bytes;
493 local_t entries;
494 local_t overrun;
495 local_t commit_overrun;
496 local_t dropped_events;
497 local_t committing;
498 local_t commits;
499 local_t pages_touched;
500 local_t pages_lost;
501 local_t pages_read;
502 long last_pages_touch;
503 size_t shortest_full;
504 unsigned long read;
505 unsigned long read_bytes;
506 rb_time_t write_stamp;
507 rb_time_t before_stamp;
508 u64 event_stamp[MAX_NEST];
509 u64 read_stamp;
510 /* pages removed since last reset */
511 unsigned long pages_removed;
512
513 unsigned int mapped;
514 unsigned int user_mapped; /* user space mapping */
515 struct mutex mapping_lock;
516 unsigned long *subbuf_ids; /* ID to subbuf VA */
517 struct trace_buffer_meta *meta_page;
518 struct ring_buffer_meta *ring_meta;
519
520 /* ring buffer pages to update, > 0 to add, < 0 to remove */
521 long nr_pages_to_update;
522 struct list_head new_pages; /* new pages to add */
523 struct work_struct update_pages_work;
524 struct completion update_done;
525
526 struct rb_irq_work irq_work;
527};
528
529struct trace_buffer {
530 unsigned flags;
531 int cpus;
532 atomic_t record_disabled;
533 atomic_t resizing;
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 unsigned long range_addr_start;
549 unsigned long range_addr_end;
550
551 long last_text_delta;
552 long last_data_delta;
553
554 unsigned int subbuf_size;
555 unsigned int subbuf_order;
556 unsigned int max_data_size;
557};
558
559struct ring_buffer_iter {
560 struct ring_buffer_per_cpu *cpu_buffer;
561 unsigned long head;
562 unsigned long next_event;
563 struct buffer_page *head_page;
564 struct buffer_page *cache_reader_page;
565 unsigned long cache_read;
566 unsigned long cache_pages_removed;
567 u64 read_stamp;
568 u64 page_stamp;
569 struct ring_buffer_event *event;
570 size_t event_size;
571 int missed_events;
572};
573
574int ring_buffer_print_page_header(struct trace_buffer *buffer, struct trace_seq *s)
575{
576 struct buffer_data_page field;
577
578 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
579 "offset:0;\tsize:%u;\tsigned:%u;\n",
580 (unsigned int)sizeof(field.time_stamp),
581 (unsigned int)is_signed_type(u64));
582
583 trace_seq_printf(s, "\tfield: local_t commit;\t"
584 "offset:%u;\tsize:%u;\tsigned:%u;\n",
585 (unsigned int)offsetof(typeof(field), commit),
586 (unsigned int)sizeof(field.commit),
587 (unsigned int)is_signed_type(long));
588
589 trace_seq_printf(s, "\tfield: int overwrite;\t"
590 "offset:%u;\tsize:%u;\tsigned:%u;\n",
591 (unsigned int)offsetof(typeof(field), commit),
592 1,
593 (unsigned int)is_signed_type(long));
594
595 trace_seq_printf(s, "\tfield: char data;\t"
596 "offset:%u;\tsize:%u;\tsigned:%u;\n",
597 (unsigned int)offsetof(typeof(field), data),
598 (unsigned int)buffer->subbuf_size,
599 (unsigned int)is_signed_type(char));
600
601 return !trace_seq_has_overflowed(s);
602}
603
604static inline void rb_time_read(rb_time_t *t, u64 *ret)
605{
606 *ret = local64_read(&t->time);
607}
608static void rb_time_set(rb_time_t *t, u64 val)
609{
610 local64_set(&t->time, val);
611}
612
613/*
614 * Enable this to make sure that the event passed to
615 * ring_buffer_event_time_stamp() is not committed and also
616 * is on the buffer that it passed in.
617 */
618//#define RB_VERIFY_EVENT
619#ifdef RB_VERIFY_EVENT
620static struct list_head *rb_list_head(struct list_head *list);
621static void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
622 void *event)
623{
624 struct buffer_page *page = cpu_buffer->commit_page;
625 struct buffer_page *tail_page = READ_ONCE(cpu_buffer->tail_page);
626 struct list_head *next;
627 long commit, write;
628 unsigned long addr = (unsigned long)event;
629 bool done = false;
630 int stop = 0;
631
632 /* Make sure the event exists and is not committed yet */
633 do {
634 if (page == tail_page || WARN_ON_ONCE(stop++ > 100))
635 done = true;
636 commit = local_read(&page->page->commit);
637 write = local_read(&page->write);
638 if (addr >= (unsigned long)&page->page->data[commit] &&
639 addr < (unsigned long)&page->page->data[write])
640 return;
641
642 next = rb_list_head(page->list.next);
643 page = list_entry(next, struct buffer_page, list);
644 } while (!done);
645 WARN_ON_ONCE(1);
646}
647#else
648static inline void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
649 void *event)
650{
651}
652#endif
653
654/*
655 * The absolute time stamp drops the 5 MSBs and some clocks may
656 * require them. The rb_fix_abs_ts() will take a previous full
657 * time stamp, and add the 5 MSB of that time stamp on to the
658 * saved absolute time stamp. Then they are compared in case of
659 * the unlikely event that the latest time stamp incremented
660 * the 5 MSB.
661 */
662static inline u64 rb_fix_abs_ts(u64 abs, u64 save_ts)
663{
664 if (save_ts & TS_MSB) {
665 abs |= save_ts & TS_MSB;
666 /* Check for overflow */
667 if (unlikely(abs < save_ts))
668 abs += 1ULL << 59;
669 }
670 return abs;
671}
672
673static inline u64 rb_time_stamp(struct trace_buffer *buffer);
674
675/**
676 * ring_buffer_event_time_stamp - return the event's current time stamp
677 * @buffer: The buffer that the event is on
678 * @event: the event to get the time stamp of
679 *
680 * Note, this must be called after @event is reserved, and before it is
681 * committed to the ring buffer. And must be called from the same
682 * context where the event was reserved (normal, softirq, irq, etc).
683 *
684 * Returns the time stamp associated with the current event.
685 * If the event has an extended time stamp, then that is used as
686 * the time stamp to return.
687 * In the highly unlikely case that the event was nested more than
688 * the max nesting, then the write_stamp of the buffer is returned,
689 * otherwise current time is returned, but that really neither of
690 * the last two cases should ever happen.
691 */
692u64 ring_buffer_event_time_stamp(struct trace_buffer *buffer,
693 struct ring_buffer_event *event)
694{
695 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[smp_processor_id()];
696 unsigned int nest;
697 u64 ts;
698
699 /* If the event includes an absolute time, then just use that */
700 if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
701 ts = rb_event_time_stamp(event);
702 return rb_fix_abs_ts(ts, cpu_buffer->tail_page->page->time_stamp);
703 }
704
705 nest = local_read(&cpu_buffer->committing);
706 verify_event(cpu_buffer, event);
707 if (WARN_ON_ONCE(!nest))
708 goto fail;
709
710 /* Read the current saved nesting level time stamp */
711 if (likely(--nest < MAX_NEST))
712 return cpu_buffer->event_stamp[nest];
713
714 /* Shouldn't happen, warn if it does */
715 WARN_ONCE(1, "nest (%d) greater than max", nest);
716
717 fail:
718 rb_time_read(&cpu_buffer->write_stamp, &ts);
719
720 return ts;
721}
722
723/**
724 * ring_buffer_nr_dirty_pages - get the number of used pages in the ring buffer
725 * @buffer: The ring_buffer to get the number of pages from
726 * @cpu: The cpu of the ring_buffer to get the number of pages from
727 *
728 * Returns the number of pages that have content in the ring buffer.
729 */
730size_t ring_buffer_nr_dirty_pages(struct trace_buffer *buffer, int cpu)
731{
732 size_t read;
733 size_t lost;
734 size_t cnt;
735
736 read = local_read(&buffer->buffers[cpu]->pages_read);
737 lost = local_read(&buffer->buffers[cpu]->pages_lost);
738 cnt = local_read(&buffer->buffers[cpu]->pages_touched);
739
740 if (WARN_ON_ONCE(cnt < lost))
741 return 0;
742
743 cnt -= lost;
744
745 /* The reader can read an empty page, but not more than that */
746 if (cnt < read) {
747 WARN_ON_ONCE(read > cnt + 1);
748 return 0;
749 }
750
751 return cnt - read;
752}
753
754static __always_inline bool full_hit(struct trace_buffer *buffer, int cpu, int full)
755{
756 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
757 size_t nr_pages;
758 size_t dirty;
759
760 nr_pages = cpu_buffer->nr_pages;
761 if (!nr_pages || !full)
762 return true;
763
764 /*
765 * Add one as dirty will never equal nr_pages, as the sub-buffer
766 * that the writer is on is not counted as dirty.
767 * This is needed if "buffer_percent" is set to 100.
768 */
769 dirty = ring_buffer_nr_dirty_pages(buffer, cpu) + 1;
770
771 return (dirty * 100) >= (full * nr_pages);
772}
773
774/*
775 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
776 *
777 * Schedules a delayed work to wake up any task that is blocked on the
778 * ring buffer waiters queue.
779 */
780static void rb_wake_up_waiters(struct irq_work *work)
781{
782 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
783
784 /* For waiters waiting for the first wake up */
785 (void)atomic_fetch_inc_release(&rbwork->seq);
786
787 wake_up_all(&rbwork->waiters);
788 if (rbwork->full_waiters_pending || rbwork->wakeup_full) {
789 /* Only cpu_buffer sets the above flags */
790 struct ring_buffer_per_cpu *cpu_buffer =
791 container_of(rbwork, struct ring_buffer_per_cpu, irq_work);
792
793 /* Called from interrupt context */
794 raw_spin_lock(&cpu_buffer->reader_lock);
795 rbwork->wakeup_full = false;
796 rbwork->full_waiters_pending = false;
797
798 /* Waking up all waiters, they will reset the shortest full */
799 cpu_buffer->shortest_full = 0;
800 raw_spin_unlock(&cpu_buffer->reader_lock);
801
802 wake_up_all(&rbwork->full_waiters);
803 }
804}
805
806/**
807 * ring_buffer_wake_waiters - wake up any waiters on this ring buffer
808 * @buffer: The ring buffer to wake waiters on
809 * @cpu: The CPU buffer to wake waiters on
810 *
811 * In the case of a file that represents a ring buffer is closing,
812 * it is prudent to wake up any waiters that are on this.
813 */
814void ring_buffer_wake_waiters(struct trace_buffer *buffer, int cpu)
815{
816 struct ring_buffer_per_cpu *cpu_buffer;
817 struct rb_irq_work *rbwork;
818
819 if (!buffer)
820 return;
821
822 if (cpu == RING_BUFFER_ALL_CPUS) {
823
824 /* Wake up individual ones too. One level recursion */
825 for_each_buffer_cpu(buffer, cpu)
826 ring_buffer_wake_waiters(buffer, cpu);
827
828 rbwork = &buffer->irq_work;
829 } else {
830 if (WARN_ON_ONCE(!buffer->buffers))
831 return;
832 if (WARN_ON_ONCE(cpu >= nr_cpu_ids))
833 return;
834
835 cpu_buffer = buffer->buffers[cpu];
836 /* The CPU buffer may not have been initialized yet */
837 if (!cpu_buffer)
838 return;
839 rbwork = &cpu_buffer->irq_work;
840 }
841
842 /* This can be called in any context */
843 irq_work_queue(&rbwork->work);
844}
845
846static bool rb_watermark_hit(struct trace_buffer *buffer, int cpu, int full)
847{
848 struct ring_buffer_per_cpu *cpu_buffer;
849 bool ret = false;
850
851 /* Reads of all CPUs always waits for any data */
852 if (cpu == RING_BUFFER_ALL_CPUS)
853 return !ring_buffer_empty(buffer);
854
855 cpu_buffer = buffer->buffers[cpu];
856
857 if (!ring_buffer_empty_cpu(buffer, cpu)) {
858 unsigned long flags;
859 bool pagebusy;
860
861 if (!full)
862 return true;
863
864 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
865 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
866 ret = !pagebusy && full_hit(buffer, cpu, full);
867
868 if (!ret && (!cpu_buffer->shortest_full ||
869 cpu_buffer->shortest_full > full)) {
870 cpu_buffer->shortest_full = full;
871 }
872 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
873 }
874 return ret;
875}
876
877static inline bool
878rb_wait_cond(struct rb_irq_work *rbwork, struct trace_buffer *buffer,
879 int cpu, int full, ring_buffer_cond_fn cond, void *data)
880{
881 if (rb_watermark_hit(buffer, cpu, full))
882 return true;
883
884 if (cond(data))
885 return true;
886
887 /*
888 * The events can happen in critical sections where
889 * checking a work queue can cause deadlocks.
890 * After adding a task to the queue, this flag is set
891 * only to notify events to try to wake up the queue
892 * using irq_work.
893 *
894 * We don't clear it even if the buffer is no longer
895 * empty. The flag only causes the next event to run
896 * irq_work to do the work queue wake up. The worse
897 * that can happen if we race with !trace_empty() is that
898 * an event will cause an irq_work to try to wake up
899 * an empty queue.
900 *
901 * There's no reason to protect this flag either, as
902 * the work queue and irq_work logic will do the necessary
903 * synchronization for the wake ups. The only thing
904 * that is necessary is that the wake up happens after
905 * a task has been queued. It's OK for spurious wake ups.
906 */
907 if (full)
908 rbwork->full_waiters_pending = true;
909 else
910 rbwork->waiters_pending = true;
911
912 return false;
913}
914
915struct rb_wait_data {
916 struct rb_irq_work *irq_work;
917 int seq;
918};
919
920/*
921 * The default wait condition for ring_buffer_wait() is to just to exit the
922 * wait loop the first time it is woken up.
923 */
924static bool rb_wait_once(void *data)
925{
926 struct rb_wait_data *rdata = data;
927 struct rb_irq_work *rbwork = rdata->irq_work;
928
929 return atomic_read_acquire(&rbwork->seq) != rdata->seq;
930}
931
932/**
933 * ring_buffer_wait - wait for input to the ring buffer
934 * @buffer: buffer to wait on
935 * @cpu: the cpu buffer to wait on
936 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
937 * @cond: condition function to break out of wait (NULL to run once)
938 * @data: the data to pass to @cond.
939 *
940 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
941 * as data is added to any of the @buffer's cpu buffers. Otherwise
942 * it will wait for data to be added to a specific cpu buffer.
943 */
944int ring_buffer_wait(struct trace_buffer *buffer, int cpu, int full,
945 ring_buffer_cond_fn cond, void *data)
946{
947 struct ring_buffer_per_cpu *cpu_buffer;
948 struct wait_queue_head *waitq;
949 struct rb_irq_work *rbwork;
950 struct rb_wait_data rdata;
951 int ret = 0;
952
953 /*
954 * Depending on what the caller is waiting for, either any
955 * data in any cpu buffer, or a specific buffer, put the
956 * caller on the appropriate wait queue.
957 */
958 if (cpu == RING_BUFFER_ALL_CPUS) {
959 rbwork = &buffer->irq_work;
960 /* Full only makes sense on per cpu reads */
961 full = 0;
962 } else {
963 if (!cpumask_test_cpu(cpu, buffer->cpumask))
964 return -ENODEV;
965 cpu_buffer = buffer->buffers[cpu];
966 rbwork = &cpu_buffer->irq_work;
967 }
968
969 if (full)
970 waitq = &rbwork->full_waiters;
971 else
972 waitq = &rbwork->waiters;
973
974 /* Set up to exit loop as soon as it is woken */
975 if (!cond) {
976 cond = rb_wait_once;
977 rdata.irq_work = rbwork;
978 rdata.seq = atomic_read_acquire(&rbwork->seq);
979 data = &rdata;
980 }
981
982 ret = wait_event_interruptible((*waitq),
983 rb_wait_cond(rbwork, buffer, cpu, full, cond, data));
984
985 return ret;
986}
987
988/**
989 * ring_buffer_poll_wait - poll on buffer input
990 * @buffer: buffer to wait on
991 * @cpu: the cpu buffer to wait on
992 * @filp: the file descriptor
993 * @poll_table: The poll descriptor
994 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
995 *
996 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
997 * as data is added to any of the @buffer's cpu buffers. Otherwise
998 * it will wait for data to be added to a specific cpu buffer.
999 *
1000 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
1001 * zero otherwise.
1002 */
1003__poll_t ring_buffer_poll_wait(struct trace_buffer *buffer, int cpu,
1004 struct file *filp, poll_table *poll_table, int full)
1005{
1006 struct ring_buffer_per_cpu *cpu_buffer;
1007 struct rb_irq_work *rbwork;
1008
1009 if (cpu == RING_BUFFER_ALL_CPUS) {
1010 rbwork = &buffer->irq_work;
1011 full = 0;
1012 } else {
1013 if (!cpumask_test_cpu(cpu, buffer->cpumask))
1014 return EPOLLERR;
1015
1016 cpu_buffer = buffer->buffers[cpu];
1017 rbwork = &cpu_buffer->irq_work;
1018 }
1019
1020 if (full) {
1021 poll_wait(filp, &rbwork->full_waiters, poll_table);
1022
1023 if (rb_watermark_hit(buffer, cpu, full))
1024 return EPOLLIN | EPOLLRDNORM;
1025 /*
1026 * Only allow full_waiters_pending update to be seen after
1027 * the shortest_full is set (in rb_watermark_hit). If the
1028 * writer sees the full_waiters_pending flag set, it will
1029 * compare the amount in the ring buffer to shortest_full.
1030 * If the amount in the ring buffer is greater than the
1031 * shortest_full percent, it will call the irq_work handler
1032 * to wake up this list. The irq_handler will reset shortest_full
1033 * back to zero. That's done under the reader_lock, but
1034 * the below smp_mb() makes sure that the update to
1035 * full_waiters_pending doesn't leak up into the above.
1036 */
1037 smp_mb();
1038 rbwork->full_waiters_pending = true;
1039 return 0;
1040 }
1041
1042 poll_wait(filp, &rbwork->waiters, poll_table);
1043 rbwork->waiters_pending = true;
1044
1045 /*
1046 * There's a tight race between setting the waiters_pending and
1047 * checking if the ring buffer is empty. Once the waiters_pending bit
1048 * is set, the next event will wake the task up, but we can get stuck
1049 * if there's only a single event in.
1050 *
1051 * FIXME: Ideally, we need a memory barrier on the writer side as well,
1052 * but adding a memory barrier to all events will cause too much of a
1053 * performance hit in the fast path. We only need a memory barrier when
1054 * the buffer goes from empty to having content. But as this race is
1055 * extremely small, and it's not a problem if another event comes in, we
1056 * will fix it later.
1057 */
1058 smp_mb();
1059
1060 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
1061 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
1062 return EPOLLIN | EPOLLRDNORM;
1063 return 0;
1064}
1065
1066/* buffer may be either ring_buffer or ring_buffer_per_cpu */
1067#define RB_WARN_ON(b, cond) \
1068 ({ \
1069 int _____ret = unlikely(cond); \
1070 if (_____ret) { \
1071 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
1072 struct ring_buffer_per_cpu *__b = \
1073 (void *)b; \
1074 atomic_inc(&__b->buffer->record_disabled); \
1075 } else \
1076 atomic_inc(&b->record_disabled); \
1077 WARN_ON(1); \
1078 } \
1079 _____ret; \
1080 })
1081
1082/* Up this if you want to test the TIME_EXTENTS and normalization */
1083#define DEBUG_SHIFT 0
1084
1085static inline u64 rb_time_stamp(struct trace_buffer *buffer)
1086{
1087 u64 ts;
1088
1089 /* Skip retpolines :-( */
1090 if (IS_ENABLED(CONFIG_MITIGATION_RETPOLINE) && likely(buffer->clock == trace_clock_local))
1091 ts = trace_clock_local();
1092 else
1093 ts = buffer->clock();
1094
1095 /* shift to debug/test normalization and TIME_EXTENTS */
1096 return ts << DEBUG_SHIFT;
1097}
1098
1099u64 ring_buffer_time_stamp(struct trace_buffer *buffer)
1100{
1101 u64 time;
1102
1103 preempt_disable_notrace();
1104 time = rb_time_stamp(buffer);
1105 preempt_enable_notrace();
1106
1107 return time;
1108}
1109EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
1110
1111void ring_buffer_normalize_time_stamp(struct trace_buffer *buffer,
1112 int cpu, u64 *ts)
1113{
1114 /* Just stupid testing the normalize function and deltas */
1115 *ts >>= DEBUG_SHIFT;
1116}
1117EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
1118
1119/*
1120 * Making the ring buffer lockless makes things tricky.
1121 * Although writes only happen on the CPU that they are on,
1122 * and they only need to worry about interrupts. Reads can
1123 * happen on any CPU.
1124 *
1125 * The reader page is always off the ring buffer, but when the
1126 * reader finishes with a page, it needs to swap its page with
1127 * a new one from the buffer. The reader needs to take from
1128 * the head (writes go to the tail). But if a writer is in overwrite
1129 * mode and wraps, it must push the head page forward.
1130 *
1131 * Here lies the problem.
1132 *
1133 * The reader must be careful to replace only the head page, and
1134 * not another one. As described at the top of the file in the
1135 * ASCII art, the reader sets its old page to point to the next
1136 * page after head. It then sets the page after head to point to
1137 * the old reader page. But if the writer moves the head page
1138 * during this operation, the reader could end up with the tail.
1139 *
1140 * We use cmpxchg to help prevent this race. We also do something
1141 * special with the page before head. We set the LSB to 1.
1142 *
1143 * When the writer must push the page forward, it will clear the
1144 * bit that points to the head page, move the head, and then set
1145 * the bit that points to the new head page.
1146 *
1147 * We also don't want an interrupt coming in and moving the head
1148 * page on another writer. Thus we use the second LSB to catch
1149 * that too. Thus:
1150 *
1151 * head->list->prev->next bit 1 bit 0
1152 * ------- -------
1153 * Normal page 0 0
1154 * Points to head page 0 1
1155 * New head page 1 0
1156 *
1157 * Note we can not trust the prev pointer of the head page, because:
1158 *
1159 * +----+ +-----+ +-----+
1160 * | |------>| T |---X--->| N |
1161 * | |<------| | | |
1162 * +----+ +-----+ +-----+
1163 * ^ ^ |
1164 * | +-----+ | |
1165 * +----------| R |----------+ |
1166 * | |<-----------+
1167 * +-----+
1168 *
1169 * Key: ---X--> HEAD flag set in pointer
1170 * T Tail page
1171 * R Reader page
1172 * N Next page
1173 *
1174 * (see __rb_reserve_next() to see where this happens)
1175 *
1176 * What the above shows is that the reader just swapped out
1177 * the reader page with a page in the buffer, but before it
1178 * could make the new header point back to the new page added
1179 * it was preempted by a writer. The writer moved forward onto
1180 * the new page added by the reader and is about to move forward
1181 * again.
1182 *
1183 * You can see, it is legitimate for the previous pointer of
1184 * the head (or any page) not to point back to itself. But only
1185 * temporarily.
1186 */
1187
1188#define RB_PAGE_NORMAL 0UL
1189#define RB_PAGE_HEAD 1UL
1190#define RB_PAGE_UPDATE 2UL
1191
1192
1193#define RB_FLAG_MASK 3UL
1194
1195/* PAGE_MOVED is not part of the mask */
1196#define RB_PAGE_MOVED 4UL
1197
1198/*
1199 * rb_list_head - remove any bit
1200 */
1201static struct list_head *rb_list_head(struct list_head *list)
1202{
1203 unsigned long val = (unsigned long)list;
1204
1205 return (struct list_head *)(val & ~RB_FLAG_MASK);
1206}
1207
1208/*
1209 * rb_is_head_page - test if the given page is the head page
1210 *
1211 * Because the reader may move the head_page pointer, we can
1212 * not trust what the head page is (it may be pointing to
1213 * the reader page). But if the next page is a header page,
1214 * its flags will be non zero.
1215 */
1216static inline int
1217rb_is_head_page(struct buffer_page *page, struct list_head *list)
1218{
1219 unsigned long val;
1220
1221 val = (unsigned long)list->next;
1222
1223 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
1224 return RB_PAGE_MOVED;
1225
1226 return val & RB_FLAG_MASK;
1227}
1228
1229/*
1230 * rb_is_reader_page
1231 *
1232 * The unique thing about the reader page, is that, if the
1233 * writer is ever on it, the previous pointer never points
1234 * back to the reader page.
1235 */
1236static bool rb_is_reader_page(struct buffer_page *page)
1237{
1238 struct list_head *list = page->list.prev;
1239
1240 return rb_list_head(list->next) != &page->list;
1241}
1242
1243/*
1244 * rb_set_list_to_head - set a list_head to be pointing to head.
1245 */
1246static void rb_set_list_to_head(struct list_head *list)
1247{
1248 unsigned long *ptr;
1249
1250 ptr = (unsigned long *)&list->next;
1251 *ptr |= RB_PAGE_HEAD;
1252 *ptr &= ~RB_PAGE_UPDATE;
1253}
1254
1255/*
1256 * rb_head_page_activate - sets up head page
1257 */
1258static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
1259{
1260 struct buffer_page *head;
1261
1262 head = cpu_buffer->head_page;
1263 if (!head)
1264 return;
1265
1266 /*
1267 * Set the previous list pointer to have the HEAD flag.
1268 */
1269 rb_set_list_to_head(head->list.prev);
1270
1271 if (cpu_buffer->ring_meta) {
1272 struct ring_buffer_meta *meta = cpu_buffer->ring_meta;
1273 meta->head_buffer = (unsigned long)head->page;
1274 }
1275}
1276
1277static void rb_list_head_clear(struct list_head *list)
1278{
1279 unsigned long *ptr = (unsigned long *)&list->next;
1280
1281 *ptr &= ~RB_FLAG_MASK;
1282}
1283
1284/*
1285 * rb_head_page_deactivate - clears head page ptr (for free list)
1286 */
1287static void
1288rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
1289{
1290 struct list_head *hd;
1291
1292 /* Go through the whole list and clear any pointers found. */
1293 rb_list_head_clear(cpu_buffer->pages);
1294
1295 list_for_each(hd, cpu_buffer->pages)
1296 rb_list_head_clear(hd);
1297}
1298
1299static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
1300 struct buffer_page *head,
1301 struct buffer_page *prev,
1302 int old_flag, int new_flag)
1303{
1304 struct list_head *list;
1305 unsigned long val = (unsigned long)&head->list;
1306 unsigned long ret;
1307
1308 list = &prev->list;
1309
1310 val &= ~RB_FLAG_MASK;
1311
1312 ret = cmpxchg((unsigned long *)&list->next,
1313 val | old_flag, val | new_flag);
1314
1315 /* check if the reader took the page */
1316 if ((ret & ~RB_FLAG_MASK) != val)
1317 return RB_PAGE_MOVED;
1318
1319 return ret & RB_FLAG_MASK;
1320}
1321
1322static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
1323 struct buffer_page *head,
1324 struct buffer_page *prev,
1325 int old_flag)
1326{
1327 return rb_head_page_set(cpu_buffer, head, prev,
1328 old_flag, RB_PAGE_UPDATE);
1329}
1330
1331static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
1332 struct buffer_page *head,
1333 struct buffer_page *prev,
1334 int old_flag)
1335{
1336 return rb_head_page_set(cpu_buffer, head, prev,
1337 old_flag, RB_PAGE_HEAD);
1338}
1339
1340static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
1341 struct buffer_page *head,
1342 struct buffer_page *prev,
1343 int old_flag)
1344{
1345 return rb_head_page_set(cpu_buffer, head, prev,
1346 old_flag, RB_PAGE_NORMAL);
1347}
1348
1349static inline void rb_inc_page(struct buffer_page **bpage)
1350{
1351 struct list_head *p = rb_list_head((*bpage)->list.next);
1352
1353 *bpage = list_entry(p, struct buffer_page, list);
1354}
1355
1356static struct buffer_page *
1357rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1358{
1359 struct buffer_page *head;
1360 struct buffer_page *page;
1361 struct list_head *list;
1362 int i;
1363
1364 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1365 return NULL;
1366
1367 /* sanity check */
1368 list = cpu_buffer->pages;
1369 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1370 return NULL;
1371
1372 page = head = cpu_buffer->head_page;
1373 /*
1374 * It is possible that the writer moves the header behind
1375 * where we started, and we miss in one loop.
1376 * A second loop should grab the header, but we'll do
1377 * three loops just because I'm paranoid.
1378 */
1379 for (i = 0; i < 3; i++) {
1380 do {
1381 if (rb_is_head_page(page, page->list.prev)) {
1382 cpu_buffer->head_page = page;
1383 return page;
1384 }
1385 rb_inc_page(&page);
1386 } while (page != head);
1387 }
1388
1389 RB_WARN_ON(cpu_buffer, 1);
1390
1391 return NULL;
1392}
1393
1394static bool rb_head_page_replace(struct buffer_page *old,
1395 struct buffer_page *new)
1396{
1397 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1398 unsigned long val;
1399
1400 val = *ptr & ~RB_FLAG_MASK;
1401 val |= RB_PAGE_HEAD;
1402
1403 return try_cmpxchg(ptr, &val, (unsigned long)&new->list);
1404}
1405
1406/*
1407 * rb_tail_page_update - move the tail page forward
1408 */
1409static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1410 struct buffer_page *tail_page,
1411 struct buffer_page *next_page)
1412{
1413 unsigned long old_entries;
1414 unsigned long old_write;
1415
1416 /*
1417 * The tail page now needs to be moved forward.
1418 *
1419 * We need to reset the tail page, but without messing
1420 * with possible erasing of data brought in by interrupts
1421 * that have moved the tail page and are currently on it.
1422 *
1423 * We add a counter to the write field to denote this.
1424 */
1425 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1426 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1427
1428 /*
1429 * Just make sure we have seen our old_write and synchronize
1430 * with any interrupts that come in.
1431 */
1432 barrier();
1433
1434 /*
1435 * If the tail page is still the same as what we think
1436 * it is, then it is up to us to update the tail
1437 * pointer.
1438 */
1439 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1440 /* Zero the write counter */
1441 unsigned long val = old_write & ~RB_WRITE_MASK;
1442 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1443
1444 /*
1445 * This will only succeed if an interrupt did
1446 * not come in and change it. In which case, we
1447 * do not want to modify it.
1448 *
1449 * We add (void) to let the compiler know that we do not care
1450 * about the return value of these functions. We use the
1451 * cmpxchg to only update if an interrupt did not already
1452 * do it for us. If the cmpxchg fails, we don't care.
1453 */
1454 (void)local_cmpxchg(&next_page->write, old_write, val);
1455 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1456
1457 /*
1458 * No need to worry about races with clearing out the commit.
1459 * it only can increment when a commit takes place. But that
1460 * only happens in the outer most nested commit.
1461 */
1462 local_set(&next_page->page->commit, 0);
1463
1464 /* Either we update tail_page or an interrupt does */
1465 if (try_cmpxchg(&cpu_buffer->tail_page, &tail_page, next_page))
1466 local_inc(&cpu_buffer->pages_touched);
1467 }
1468}
1469
1470static void rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1471 struct buffer_page *bpage)
1472{
1473 unsigned long val = (unsigned long)bpage;
1474
1475 RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK);
1476}
1477
1478/**
1479 * rb_check_pages - integrity check of buffer pages
1480 * @cpu_buffer: CPU buffer with pages to test
1481 *
1482 * As a safety measure we check to make sure the data pages have not
1483 * been corrupted.
1484 *
1485 * Callers of this function need to guarantee that the list of pages doesn't get
1486 * modified during the check. In particular, if it's possible that the function
1487 * is invoked with concurrent readers which can swap in a new reader page then
1488 * the caller should take cpu_buffer->reader_lock.
1489 */
1490static void rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1491{
1492 struct list_head *head = rb_list_head(cpu_buffer->pages);
1493 struct list_head *tmp;
1494
1495 if (RB_WARN_ON(cpu_buffer,
1496 rb_list_head(rb_list_head(head->next)->prev) != head))
1497 return;
1498
1499 if (RB_WARN_ON(cpu_buffer,
1500 rb_list_head(rb_list_head(head->prev)->next) != head))
1501 return;
1502
1503 for (tmp = rb_list_head(head->next); tmp != head; tmp = rb_list_head(tmp->next)) {
1504 if (RB_WARN_ON(cpu_buffer,
1505 rb_list_head(rb_list_head(tmp->next)->prev) != tmp))
1506 return;
1507
1508 if (RB_WARN_ON(cpu_buffer,
1509 rb_list_head(rb_list_head(tmp->prev)->next) != tmp))
1510 return;
1511 }
1512}
1513
1514/*
1515 * Take an address, add the meta data size as well as the array of
1516 * array subbuffer indexes, then align it to a subbuffer size.
1517 *
1518 * This is used to help find the next per cpu subbuffer within a mapped range.
1519 */
1520static unsigned long
1521rb_range_align_subbuf(unsigned long addr, int subbuf_size, int nr_subbufs)
1522{
1523 addr += sizeof(struct ring_buffer_meta) +
1524 sizeof(int) * nr_subbufs;
1525 return ALIGN(addr, subbuf_size);
1526}
1527
1528/*
1529 * Return the ring_buffer_meta for a given @cpu.
1530 */
1531static void *rb_range_meta(struct trace_buffer *buffer, int nr_pages, int cpu)
1532{
1533 int subbuf_size = buffer->subbuf_size + BUF_PAGE_HDR_SIZE;
1534 unsigned long ptr = buffer->range_addr_start;
1535 struct ring_buffer_meta *meta;
1536 int nr_subbufs;
1537
1538 if (!ptr)
1539 return NULL;
1540
1541 /* When nr_pages passed in is zero, the first meta has already been initialized */
1542 if (!nr_pages) {
1543 meta = (struct ring_buffer_meta *)ptr;
1544 nr_subbufs = meta->nr_subbufs;
1545 } else {
1546 meta = NULL;
1547 /* Include the reader page */
1548 nr_subbufs = nr_pages + 1;
1549 }
1550
1551 /*
1552 * The first chunk may not be subbuffer aligned, where as
1553 * the rest of the chunks are.
1554 */
1555 if (cpu) {
1556 ptr = rb_range_align_subbuf(ptr, subbuf_size, nr_subbufs);
1557 ptr += subbuf_size * nr_subbufs;
1558
1559 /* We can use multiplication to find chunks greater than 1 */
1560 if (cpu > 1) {
1561 unsigned long size;
1562 unsigned long p;
1563
1564 /* Save the beginning of this CPU chunk */
1565 p = ptr;
1566 ptr = rb_range_align_subbuf(ptr, subbuf_size, nr_subbufs);
1567 ptr += subbuf_size * nr_subbufs;
1568
1569 /* Now all chunks after this are the same size */
1570 size = ptr - p;
1571 ptr += size * (cpu - 2);
1572 }
1573 }
1574 return (void *)ptr;
1575}
1576
1577/* Return the start of subbufs given the meta pointer */
1578static void *rb_subbufs_from_meta(struct ring_buffer_meta *meta)
1579{
1580 int subbuf_size = meta->subbuf_size;
1581 unsigned long ptr;
1582
1583 ptr = (unsigned long)meta;
1584 ptr = rb_range_align_subbuf(ptr, subbuf_size, meta->nr_subbufs);
1585
1586 return (void *)ptr;
1587}
1588
1589/*
1590 * Return a specific sub-buffer for a given @cpu defined by @idx.
1591 */
1592static void *rb_range_buffer(struct ring_buffer_per_cpu *cpu_buffer, int idx)
1593{
1594 struct ring_buffer_meta *meta;
1595 unsigned long ptr;
1596 int subbuf_size;
1597
1598 meta = rb_range_meta(cpu_buffer->buffer, 0, cpu_buffer->cpu);
1599 if (!meta)
1600 return NULL;
1601
1602 if (WARN_ON_ONCE(idx >= meta->nr_subbufs))
1603 return NULL;
1604
1605 subbuf_size = meta->subbuf_size;
1606
1607 /* Map this buffer to the order that's in meta->buffers[] */
1608 idx = meta->buffers[idx];
1609
1610 ptr = (unsigned long)rb_subbufs_from_meta(meta);
1611
1612 ptr += subbuf_size * idx;
1613 if (ptr + subbuf_size > cpu_buffer->buffer->range_addr_end)
1614 return NULL;
1615
1616 return (void *)ptr;
1617}
1618
1619/*
1620 * See if the existing memory contains valid ring buffer data.
1621 * As the previous kernel must be the same as this kernel, all
1622 * the calculations (size of buffers and number of buffers)
1623 * must be the same.
1624 */
1625static bool rb_meta_valid(struct ring_buffer_meta *meta, int cpu,
1626 struct trace_buffer *buffer, int nr_pages)
1627{
1628 int subbuf_size = PAGE_SIZE;
1629 struct buffer_data_page *subbuf;
1630 unsigned long buffers_start;
1631 unsigned long buffers_end;
1632 int i;
1633
1634 /* Check the meta magic and meta struct size */
1635 if (meta->magic != RING_BUFFER_META_MAGIC ||
1636 meta->struct_size != sizeof(*meta)) {
1637 pr_info("Ring buffer boot meta[%d] mismatch of magic or struct size\n", cpu);
1638 return false;
1639 }
1640
1641 /* The subbuffer's size and number of subbuffers must match */
1642 if (meta->subbuf_size != subbuf_size ||
1643 meta->nr_subbufs != nr_pages + 1) {
1644 pr_info("Ring buffer boot meta [%d] mismatch of subbuf_size/nr_pages\n", cpu);
1645 return false;
1646 }
1647
1648 buffers_start = meta->first_buffer;
1649 buffers_end = meta->first_buffer + (subbuf_size * meta->nr_subbufs);
1650
1651 /* Is the head and commit buffers within the range of buffers? */
1652 if (meta->head_buffer < buffers_start ||
1653 meta->head_buffer >= buffers_end) {
1654 pr_info("Ring buffer boot meta [%d] head buffer out of range\n", cpu);
1655 return false;
1656 }
1657
1658 if (meta->commit_buffer < buffers_start ||
1659 meta->commit_buffer >= buffers_end) {
1660 pr_info("Ring buffer boot meta [%d] commit buffer out of range\n", cpu);
1661 return false;
1662 }
1663
1664 subbuf = rb_subbufs_from_meta(meta);
1665
1666 /* Is the meta buffers and the subbufs themselves have correct data? */
1667 for (i = 0; i < meta->nr_subbufs; i++) {
1668 if (meta->buffers[i] < 0 ||
1669 meta->buffers[i] >= meta->nr_subbufs) {
1670 pr_info("Ring buffer boot meta [%d] array out of range\n", cpu);
1671 return false;
1672 }
1673
1674 if ((unsigned)local_read(&subbuf->commit) > subbuf_size) {
1675 pr_info("Ring buffer boot meta [%d] buffer invalid commit\n", cpu);
1676 return false;
1677 }
1678
1679 subbuf = (void *)subbuf + subbuf_size;
1680 }
1681
1682 return true;
1683}
1684
1685static int rb_meta_subbuf_idx(struct ring_buffer_meta *meta, void *subbuf);
1686
1687static int rb_read_data_buffer(struct buffer_data_page *dpage, int tail, int cpu,
1688 unsigned long long *timestamp, u64 *delta_ptr)
1689{
1690 struct ring_buffer_event *event;
1691 u64 ts, delta;
1692 int events = 0;
1693 int e;
1694
1695 *delta_ptr = 0;
1696 *timestamp = 0;
1697
1698 ts = dpage->time_stamp;
1699
1700 for (e = 0; e < tail; e += rb_event_length(event)) {
1701
1702 event = (struct ring_buffer_event *)(dpage->data + e);
1703
1704 switch (event->type_len) {
1705
1706 case RINGBUF_TYPE_TIME_EXTEND:
1707 delta = rb_event_time_stamp(event);
1708 ts += delta;
1709 break;
1710
1711 case RINGBUF_TYPE_TIME_STAMP:
1712 delta = rb_event_time_stamp(event);
1713 delta = rb_fix_abs_ts(delta, ts);
1714 if (delta < ts) {
1715 *delta_ptr = delta;
1716 *timestamp = ts;
1717 return -1;
1718 }
1719 ts = delta;
1720 break;
1721
1722 case RINGBUF_TYPE_PADDING:
1723 if (event->time_delta == 1)
1724 break;
1725 fallthrough;
1726 case RINGBUF_TYPE_DATA:
1727 events++;
1728 ts += event->time_delta;
1729 break;
1730
1731 default:
1732 return -1;
1733 }
1734 }
1735 *timestamp = ts;
1736 return events;
1737}
1738
1739static int rb_validate_buffer(struct buffer_data_page *dpage, int cpu)
1740{
1741 unsigned long long ts;
1742 u64 delta;
1743 int tail;
1744
1745 tail = local_read(&dpage->commit);
1746 return rb_read_data_buffer(dpage, tail, cpu, &ts, &delta);
1747}
1748
1749/* If the meta data has been validated, now validate the events */
1750static void rb_meta_validate_events(struct ring_buffer_per_cpu *cpu_buffer)
1751{
1752 struct ring_buffer_meta *meta = cpu_buffer->ring_meta;
1753 struct buffer_page *head_page;
1754 unsigned long entry_bytes = 0;
1755 unsigned long entries = 0;
1756 int ret;
1757 int i;
1758
1759 if (!meta || !meta->head_buffer)
1760 return;
1761
1762 /* Do the reader page first */
1763 ret = rb_validate_buffer(cpu_buffer->reader_page->page, cpu_buffer->cpu);
1764 if (ret < 0) {
1765 pr_info("Ring buffer reader page is invalid\n");
1766 goto invalid;
1767 }
1768 entries += ret;
1769 entry_bytes += local_read(&cpu_buffer->reader_page->page->commit);
1770 local_set(&cpu_buffer->reader_page->entries, ret);
1771
1772 head_page = cpu_buffer->head_page;
1773
1774 /* If both the head and commit are on the reader_page then we are done. */
1775 if (head_page == cpu_buffer->reader_page &&
1776 head_page == cpu_buffer->commit_page)
1777 goto done;
1778
1779 /* Iterate until finding the commit page */
1780 for (i = 0; i < meta->nr_subbufs + 1; i++, rb_inc_page(&head_page)) {
1781
1782 /* Reader page has already been done */
1783 if (head_page == cpu_buffer->reader_page)
1784 continue;
1785
1786 ret = rb_validate_buffer(head_page->page, cpu_buffer->cpu);
1787 if (ret < 0) {
1788 pr_info("Ring buffer meta [%d] invalid buffer page\n",
1789 cpu_buffer->cpu);
1790 goto invalid;
1791 }
1792 entries += ret;
1793 entry_bytes += local_read(&head_page->page->commit);
1794 local_set(&cpu_buffer->head_page->entries, ret);
1795
1796 if (head_page == cpu_buffer->commit_page)
1797 break;
1798 }
1799
1800 if (head_page != cpu_buffer->commit_page) {
1801 pr_info("Ring buffer meta [%d] commit page not found\n",
1802 cpu_buffer->cpu);
1803 goto invalid;
1804 }
1805 done:
1806 local_set(&cpu_buffer->entries, entries);
1807 local_set(&cpu_buffer->entries_bytes, entry_bytes);
1808
1809 pr_info("Ring buffer meta [%d] is from previous boot!\n", cpu_buffer->cpu);
1810 return;
1811
1812 invalid:
1813 /* The content of the buffers are invalid, reset the meta data */
1814 meta->head_buffer = 0;
1815 meta->commit_buffer = 0;
1816
1817 /* Reset the reader page */
1818 local_set(&cpu_buffer->reader_page->entries, 0);
1819 local_set(&cpu_buffer->reader_page->page->commit, 0);
1820
1821 /* Reset all the subbuffers */
1822 for (i = 0; i < meta->nr_subbufs - 1; i++, rb_inc_page(&head_page)) {
1823 local_set(&head_page->entries, 0);
1824 local_set(&head_page->page->commit, 0);
1825 }
1826}
1827
1828/* Used to calculate data delta */
1829static char rb_data_ptr[] = "";
1830
1831#define THIS_TEXT_PTR ((unsigned long)rb_meta_init_text_addr)
1832#define THIS_DATA_PTR ((unsigned long)rb_data_ptr)
1833
1834static void rb_meta_init_text_addr(struct ring_buffer_meta *meta)
1835{
1836 meta->text_addr = THIS_TEXT_PTR;
1837 meta->data_addr = THIS_DATA_PTR;
1838}
1839
1840static void rb_range_meta_init(struct trace_buffer *buffer, int nr_pages)
1841{
1842 struct ring_buffer_meta *meta;
1843 unsigned long delta;
1844 void *subbuf;
1845 int cpu;
1846 int i;
1847
1848 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
1849 void *next_meta;
1850
1851 meta = rb_range_meta(buffer, nr_pages, cpu);
1852
1853 if (rb_meta_valid(meta, cpu, buffer, nr_pages)) {
1854 /* Make the mappings match the current address */
1855 subbuf = rb_subbufs_from_meta(meta);
1856 delta = (unsigned long)subbuf - meta->first_buffer;
1857 meta->first_buffer += delta;
1858 meta->head_buffer += delta;
1859 meta->commit_buffer += delta;
1860 buffer->last_text_delta = THIS_TEXT_PTR - meta->text_addr;
1861 buffer->last_data_delta = THIS_DATA_PTR - meta->data_addr;
1862 continue;
1863 }
1864
1865 if (cpu < nr_cpu_ids - 1)
1866 next_meta = rb_range_meta(buffer, nr_pages, cpu + 1);
1867 else
1868 next_meta = (void *)buffer->range_addr_end;
1869
1870 memset(meta, 0, next_meta - (void *)meta);
1871
1872 meta->magic = RING_BUFFER_META_MAGIC;
1873 meta->struct_size = sizeof(*meta);
1874
1875 meta->nr_subbufs = nr_pages + 1;
1876 meta->subbuf_size = PAGE_SIZE;
1877
1878 subbuf = rb_subbufs_from_meta(meta);
1879
1880 meta->first_buffer = (unsigned long)subbuf;
1881 rb_meta_init_text_addr(meta);
1882
1883 /*
1884 * The buffers[] array holds the order of the sub-buffers
1885 * that are after the meta data. The sub-buffers may
1886 * be swapped out when read and inserted into a different
1887 * location of the ring buffer. Although their addresses
1888 * remain the same, the buffers[] array contains the
1889 * index into the sub-buffers holding their actual order.
1890 */
1891 for (i = 0; i < meta->nr_subbufs; i++) {
1892 meta->buffers[i] = i;
1893 rb_init_page(subbuf);
1894 subbuf += meta->subbuf_size;
1895 }
1896 }
1897}
1898
1899static void *rbm_start(struct seq_file *m, loff_t *pos)
1900{
1901 struct ring_buffer_per_cpu *cpu_buffer = m->private;
1902 struct ring_buffer_meta *meta = cpu_buffer->ring_meta;
1903 unsigned long val;
1904
1905 if (!meta)
1906 return NULL;
1907
1908 if (*pos > meta->nr_subbufs)
1909 return NULL;
1910
1911 val = *pos;
1912 val++;
1913
1914 return (void *)val;
1915}
1916
1917static void *rbm_next(struct seq_file *m, void *v, loff_t *pos)
1918{
1919 (*pos)++;
1920
1921 return rbm_start(m, pos);
1922}
1923
1924static int rbm_show(struct seq_file *m, void *v)
1925{
1926 struct ring_buffer_per_cpu *cpu_buffer = m->private;
1927 struct ring_buffer_meta *meta = cpu_buffer->ring_meta;
1928 unsigned long val = (unsigned long)v;
1929
1930 if (val == 1) {
1931 seq_printf(m, "head_buffer: %d\n",
1932 rb_meta_subbuf_idx(meta, (void *)meta->head_buffer));
1933 seq_printf(m, "commit_buffer: %d\n",
1934 rb_meta_subbuf_idx(meta, (void *)meta->commit_buffer));
1935 seq_printf(m, "subbuf_size: %d\n", meta->subbuf_size);
1936 seq_printf(m, "nr_subbufs: %d\n", meta->nr_subbufs);
1937 return 0;
1938 }
1939
1940 val -= 2;
1941 seq_printf(m, "buffer[%ld]: %d\n", val, meta->buffers[val]);
1942
1943 return 0;
1944}
1945
1946static void rbm_stop(struct seq_file *m, void *p)
1947{
1948}
1949
1950static const struct seq_operations rb_meta_seq_ops = {
1951 .start = rbm_start,
1952 .next = rbm_next,
1953 .show = rbm_show,
1954 .stop = rbm_stop,
1955};
1956
1957int ring_buffer_meta_seq_init(struct file *file, struct trace_buffer *buffer, int cpu)
1958{
1959 struct seq_file *m;
1960 int ret;
1961
1962 ret = seq_open(file, &rb_meta_seq_ops);
1963 if (ret)
1964 return ret;
1965
1966 m = file->private_data;
1967 m->private = buffer->buffers[cpu];
1968
1969 return 0;
1970}
1971
1972/* Map the buffer_pages to the previous head and commit pages */
1973static void rb_meta_buffer_update(struct ring_buffer_per_cpu *cpu_buffer,
1974 struct buffer_page *bpage)
1975{
1976 struct ring_buffer_meta *meta = cpu_buffer->ring_meta;
1977
1978 if (meta->head_buffer == (unsigned long)bpage->page)
1979 cpu_buffer->head_page = bpage;
1980
1981 if (meta->commit_buffer == (unsigned long)bpage->page) {
1982 cpu_buffer->commit_page = bpage;
1983 cpu_buffer->tail_page = bpage;
1984 }
1985}
1986
1987static int __rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1988 long nr_pages, struct list_head *pages)
1989{
1990 struct trace_buffer *buffer = cpu_buffer->buffer;
1991 struct ring_buffer_meta *meta = NULL;
1992 struct buffer_page *bpage, *tmp;
1993 bool user_thread = current->mm != NULL;
1994 gfp_t mflags;
1995 long i;
1996
1997 /*
1998 * Check if the available memory is there first.
1999 * Note, si_mem_available() only gives us a rough estimate of available
2000 * memory. It may not be accurate. But we don't care, we just want
2001 * to prevent doing any allocation when it is obvious that it is
2002 * not going to succeed.
2003 */
2004 i = si_mem_available();
2005 if (i < nr_pages)
2006 return -ENOMEM;
2007
2008 /*
2009 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
2010 * gracefully without invoking oom-killer and the system is not
2011 * destabilized.
2012 */
2013 mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
2014
2015 /*
2016 * If a user thread allocates too much, and si_mem_available()
2017 * reports there's enough memory, even though there is not.
2018 * Make sure the OOM killer kills this thread. This can happen
2019 * even with RETRY_MAYFAIL because another task may be doing
2020 * an allocation after this task has taken all memory.
2021 * This is the task the OOM killer needs to take out during this
2022 * loop, even if it was triggered by an allocation somewhere else.
2023 */
2024 if (user_thread)
2025 set_current_oom_origin();
2026
2027 if (buffer->range_addr_start)
2028 meta = rb_range_meta(buffer, nr_pages, cpu_buffer->cpu);
2029
2030 for (i = 0; i < nr_pages; i++) {
2031 struct page *page;
2032
2033 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
2034 mflags, cpu_to_node(cpu_buffer->cpu));
2035 if (!bpage)
2036 goto free_pages;
2037
2038 rb_check_bpage(cpu_buffer, bpage);
2039
2040 /*
2041 * Append the pages as for mapped buffers we want to keep
2042 * the order
2043 */
2044 list_add_tail(&bpage->list, pages);
2045
2046 if (meta) {
2047 /* A range was given. Use that for the buffer page */
2048 bpage->page = rb_range_buffer(cpu_buffer, i + 1);
2049 if (!bpage->page)
2050 goto free_pages;
2051 /* If this is valid from a previous boot */
2052 if (meta->head_buffer)
2053 rb_meta_buffer_update(cpu_buffer, bpage);
2054 bpage->range = 1;
2055 bpage->id = i + 1;
2056 } else {
2057 page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu),
2058 mflags | __GFP_COMP | __GFP_ZERO,
2059 cpu_buffer->buffer->subbuf_order);
2060 if (!page)
2061 goto free_pages;
2062 bpage->page = page_address(page);
2063 rb_init_page(bpage->page);
2064 }
2065 bpage->order = cpu_buffer->buffer->subbuf_order;
2066
2067 if (user_thread && fatal_signal_pending(current))
2068 goto free_pages;
2069 }
2070 if (user_thread)
2071 clear_current_oom_origin();
2072
2073 return 0;
2074
2075free_pages:
2076 list_for_each_entry_safe(bpage, tmp, pages, list) {
2077 list_del_init(&bpage->list);
2078 free_buffer_page(bpage);
2079 }
2080 if (user_thread)
2081 clear_current_oom_origin();
2082
2083 return -ENOMEM;
2084}
2085
2086static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
2087 unsigned long nr_pages)
2088{
2089 LIST_HEAD(pages);
2090
2091 WARN_ON(!nr_pages);
2092
2093 if (__rb_allocate_pages(cpu_buffer, nr_pages, &pages))
2094 return -ENOMEM;
2095
2096 /*
2097 * The ring buffer page list is a circular list that does not
2098 * start and end with a list head. All page list items point to
2099 * other pages.
2100 */
2101 cpu_buffer->pages = pages.next;
2102 list_del(&pages);
2103
2104 cpu_buffer->nr_pages = nr_pages;
2105
2106 rb_check_pages(cpu_buffer);
2107
2108 return 0;
2109}
2110
2111static struct ring_buffer_per_cpu *
2112rb_allocate_cpu_buffer(struct trace_buffer *buffer, long nr_pages, int cpu)
2113{
2114 struct ring_buffer_per_cpu *cpu_buffer;
2115 struct ring_buffer_meta *meta;
2116 struct buffer_page *bpage;
2117 struct page *page;
2118 int ret;
2119
2120 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
2121 GFP_KERNEL, cpu_to_node(cpu));
2122 if (!cpu_buffer)
2123 return NULL;
2124
2125 cpu_buffer->cpu = cpu;
2126 cpu_buffer->buffer = buffer;
2127 raw_spin_lock_init(&cpu_buffer->reader_lock);
2128 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
2129 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
2130 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
2131 init_completion(&cpu_buffer->update_done);
2132 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
2133 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
2134 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
2135 mutex_init(&cpu_buffer->mapping_lock);
2136
2137 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
2138 GFP_KERNEL, cpu_to_node(cpu));
2139 if (!bpage)
2140 goto fail_free_buffer;
2141
2142 rb_check_bpage(cpu_buffer, bpage);
2143
2144 cpu_buffer->reader_page = bpage;
2145
2146 if (buffer->range_addr_start) {
2147 /*
2148 * Range mapped buffers have the same restrictions as memory
2149 * mapped ones do.
2150 */
2151 cpu_buffer->mapped = 1;
2152 cpu_buffer->ring_meta = rb_range_meta(buffer, nr_pages, cpu);
2153 bpage->page = rb_range_buffer(cpu_buffer, 0);
2154 if (!bpage->page)
2155 goto fail_free_reader;
2156 if (cpu_buffer->ring_meta->head_buffer)
2157 rb_meta_buffer_update(cpu_buffer, bpage);
2158 bpage->range = 1;
2159 } else {
2160 page = alloc_pages_node(cpu_to_node(cpu),
2161 GFP_KERNEL | __GFP_COMP | __GFP_ZERO,
2162 cpu_buffer->buffer->subbuf_order);
2163 if (!page)
2164 goto fail_free_reader;
2165 bpage->page = page_address(page);
2166 rb_init_page(bpage->page);
2167 }
2168
2169 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
2170 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2171
2172 ret = rb_allocate_pages(cpu_buffer, nr_pages);
2173 if (ret < 0)
2174 goto fail_free_reader;
2175
2176 rb_meta_validate_events(cpu_buffer);
2177
2178 /* If the boot meta was valid then this has already been updated */
2179 meta = cpu_buffer->ring_meta;
2180 if (!meta || !meta->head_buffer ||
2181 !cpu_buffer->head_page || !cpu_buffer->commit_page || !cpu_buffer->tail_page) {
2182 if (meta && meta->head_buffer &&
2183 (cpu_buffer->head_page || cpu_buffer->commit_page || cpu_buffer->tail_page)) {
2184 pr_warn("Ring buffer meta buffers not all mapped\n");
2185 if (!cpu_buffer->head_page)
2186 pr_warn(" Missing head_page\n");
2187 if (!cpu_buffer->commit_page)
2188 pr_warn(" Missing commit_page\n");
2189 if (!cpu_buffer->tail_page)
2190 pr_warn(" Missing tail_page\n");
2191 }
2192
2193 cpu_buffer->head_page
2194 = list_entry(cpu_buffer->pages, struct buffer_page, list);
2195 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
2196
2197 rb_head_page_activate(cpu_buffer);
2198
2199 if (cpu_buffer->ring_meta)
2200 meta->commit_buffer = meta->head_buffer;
2201 } else {
2202 /* The valid meta buffer still needs to activate the head page */
2203 rb_head_page_activate(cpu_buffer);
2204 }
2205
2206 return cpu_buffer;
2207
2208 fail_free_reader:
2209 free_buffer_page(cpu_buffer->reader_page);
2210
2211 fail_free_buffer:
2212 kfree(cpu_buffer);
2213 return NULL;
2214}
2215
2216static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
2217{
2218 struct list_head *head = cpu_buffer->pages;
2219 struct buffer_page *bpage, *tmp;
2220
2221 irq_work_sync(&cpu_buffer->irq_work.work);
2222
2223 free_buffer_page(cpu_buffer->reader_page);
2224
2225 if (head) {
2226 rb_head_page_deactivate(cpu_buffer);
2227
2228 list_for_each_entry_safe(bpage, tmp, head, list) {
2229 list_del_init(&bpage->list);
2230 free_buffer_page(bpage);
2231 }
2232 bpage = list_entry(head, struct buffer_page, list);
2233 free_buffer_page(bpage);
2234 }
2235
2236 free_page((unsigned long)cpu_buffer->free_page);
2237
2238 kfree(cpu_buffer);
2239}
2240
2241static struct trace_buffer *alloc_buffer(unsigned long size, unsigned flags,
2242 int order, unsigned long start,
2243 unsigned long end,
2244 struct lock_class_key *key)
2245{
2246 struct trace_buffer *buffer;
2247 long nr_pages;
2248 int subbuf_size;
2249 int bsize;
2250 int cpu;
2251 int ret;
2252
2253 /* keep it in its own cache line */
2254 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
2255 GFP_KERNEL);
2256 if (!buffer)
2257 return NULL;
2258
2259 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
2260 goto fail_free_buffer;
2261
2262 buffer->subbuf_order = order;
2263 subbuf_size = (PAGE_SIZE << order);
2264 buffer->subbuf_size = subbuf_size - BUF_PAGE_HDR_SIZE;
2265
2266 /* Max payload is buffer page size - header (8bytes) */
2267 buffer->max_data_size = buffer->subbuf_size - (sizeof(u32) * 2);
2268
2269 buffer->flags = flags;
2270 buffer->clock = trace_clock_local;
2271 buffer->reader_lock_key = key;
2272
2273 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
2274 init_waitqueue_head(&buffer->irq_work.waiters);
2275
2276 buffer->cpus = nr_cpu_ids;
2277
2278 bsize = sizeof(void *) * nr_cpu_ids;
2279 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
2280 GFP_KERNEL);
2281 if (!buffer->buffers)
2282 goto fail_free_cpumask;
2283
2284 /* If start/end are specified, then that overrides size */
2285 if (start && end) {
2286 unsigned long ptr;
2287 int n;
2288
2289 size = end - start;
2290 size = size / nr_cpu_ids;
2291
2292 /*
2293 * The number of sub-buffers (nr_pages) is determined by the
2294 * total size allocated minus the meta data size.
2295 * Then that is divided by the number of per CPU buffers
2296 * needed, plus account for the integer array index that
2297 * will be appended to the meta data.
2298 */
2299 nr_pages = (size - sizeof(struct ring_buffer_meta)) /
2300 (subbuf_size + sizeof(int));
2301 /* Need at least two pages plus the reader page */
2302 if (nr_pages < 3)
2303 goto fail_free_buffers;
2304
2305 again:
2306 /* Make sure that the size fits aligned */
2307 for (n = 0, ptr = start; n < nr_cpu_ids; n++) {
2308 ptr += sizeof(struct ring_buffer_meta) +
2309 sizeof(int) * nr_pages;
2310 ptr = ALIGN(ptr, subbuf_size);
2311 ptr += subbuf_size * nr_pages;
2312 }
2313 if (ptr > end) {
2314 if (nr_pages <= 3)
2315 goto fail_free_buffers;
2316 nr_pages--;
2317 goto again;
2318 }
2319
2320 /* nr_pages should not count the reader page */
2321 nr_pages--;
2322 buffer->range_addr_start = start;
2323 buffer->range_addr_end = end;
2324
2325 rb_range_meta_init(buffer, nr_pages);
2326 } else {
2327
2328 /* need at least two pages */
2329 nr_pages = DIV_ROUND_UP(size, buffer->subbuf_size);
2330 if (nr_pages < 2)
2331 nr_pages = 2;
2332 }
2333
2334 cpu = raw_smp_processor_id();
2335 cpumask_set_cpu(cpu, buffer->cpumask);
2336 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
2337 if (!buffer->buffers[cpu])
2338 goto fail_free_buffers;
2339
2340 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
2341 if (ret < 0)
2342 goto fail_free_buffers;
2343
2344 mutex_init(&buffer->mutex);
2345
2346 return buffer;
2347
2348 fail_free_buffers:
2349 for_each_buffer_cpu(buffer, cpu) {
2350 if (buffer->buffers[cpu])
2351 rb_free_cpu_buffer(buffer->buffers[cpu]);
2352 }
2353 kfree(buffer->buffers);
2354
2355 fail_free_cpumask:
2356 free_cpumask_var(buffer->cpumask);
2357
2358 fail_free_buffer:
2359 kfree(buffer);
2360 return NULL;
2361}
2362
2363/**
2364 * __ring_buffer_alloc - allocate a new ring_buffer
2365 * @size: the size in bytes per cpu that is needed.
2366 * @flags: attributes to set for the ring buffer.
2367 * @key: ring buffer reader_lock_key.
2368 *
2369 * Currently the only flag that is available is the RB_FL_OVERWRITE
2370 * flag. This flag means that the buffer will overwrite old data
2371 * when the buffer wraps. If this flag is not set, the buffer will
2372 * drop data when the tail hits the head.
2373 */
2374struct trace_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
2375 struct lock_class_key *key)
2376{
2377 /* Default buffer page size - one system page */
2378 return alloc_buffer(size, flags, 0, 0, 0,key);
2379
2380}
2381EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
2382
2383/**
2384 * __ring_buffer_alloc_range - allocate a new ring_buffer from existing memory
2385 * @size: the size in bytes per cpu that is needed.
2386 * @flags: attributes to set for the ring buffer.
2387 * @start: start of allocated range
2388 * @range_size: size of allocated range
2389 * @order: sub-buffer order
2390 * @key: ring buffer reader_lock_key.
2391 *
2392 * Currently the only flag that is available is the RB_FL_OVERWRITE
2393 * flag. This flag means that the buffer will overwrite old data
2394 * when the buffer wraps. If this flag is not set, the buffer will
2395 * drop data when the tail hits the head.
2396 */
2397struct trace_buffer *__ring_buffer_alloc_range(unsigned long size, unsigned flags,
2398 int order, unsigned long start,
2399 unsigned long range_size,
2400 struct lock_class_key *key)
2401{
2402 return alloc_buffer(size, flags, order, start, start + range_size, key);
2403}
2404
2405/**
2406 * ring_buffer_last_boot_delta - return the delta offset from last boot
2407 * @buffer: The buffer to return the delta from
2408 * @text: Return text delta
2409 * @data: Return data delta
2410 *
2411 * Returns: The true if the delta is non zero
2412 */
2413bool ring_buffer_last_boot_delta(struct trace_buffer *buffer, long *text,
2414 long *data)
2415{
2416 if (!buffer)
2417 return false;
2418
2419 if (!buffer->last_text_delta)
2420 return false;
2421
2422 *text = buffer->last_text_delta;
2423 *data = buffer->last_data_delta;
2424
2425 return true;
2426}
2427
2428/**
2429 * ring_buffer_free - free a ring buffer.
2430 * @buffer: the buffer to free.
2431 */
2432void
2433ring_buffer_free(struct trace_buffer *buffer)
2434{
2435 int cpu;
2436
2437 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
2438
2439 irq_work_sync(&buffer->irq_work.work);
2440
2441 for_each_buffer_cpu(buffer, cpu)
2442 rb_free_cpu_buffer(buffer->buffers[cpu]);
2443
2444 kfree(buffer->buffers);
2445 free_cpumask_var(buffer->cpumask);
2446
2447 kfree(buffer);
2448}
2449EXPORT_SYMBOL_GPL(ring_buffer_free);
2450
2451void ring_buffer_set_clock(struct trace_buffer *buffer,
2452 u64 (*clock)(void))
2453{
2454 buffer->clock = clock;
2455}
2456
2457void ring_buffer_set_time_stamp_abs(struct trace_buffer *buffer, bool abs)
2458{
2459 buffer->time_stamp_abs = abs;
2460}
2461
2462bool ring_buffer_time_stamp_abs(struct trace_buffer *buffer)
2463{
2464 return buffer->time_stamp_abs;
2465}
2466
2467static inline unsigned long rb_page_entries(struct buffer_page *bpage)
2468{
2469 return local_read(&bpage->entries) & RB_WRITE_MASK;
2470}
2471
2472static inline unsigned long rb_page_write(struct buffer_page *bpage)
2473{
2474 return local_read(&bpage->write) & RB_WRITE_MASK;
2475}
2476
2477static bool
2478rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
2479{
2480 struct list_head *tail_page, *to_remove, *next_page;
2481 struct buffer_page *to_remove_page, *tmp_iter_page;
2482 struct buffer_page *last_page, *first_page;
2483 unsigned long nr_removed;
2484 unsigned long head_bit;
2485 int page_entries;
2486
2487 head_bit = 0;
2488
2489 raw_spin_lock_irq(&cpu_buffer->reader_lock);
2490 atomic_inc(&cpu_buffer->record_disabled);
2491 /*
2492 * We don't race with the readers since we have acquired the reader
2493 * lock. We also don't race with writers after disabling recording.
2494 * This makes it easy to figure out the first and the last page to be
2495 * removed from the list. We unlink all the pages in between including
2496 * the first and last pages. This is done in a busy loop so that we
2497 * lose the least number of traces.
2498 * The pages are freed after we restart recording and unlock readers.
2499 */
2500 tail_page = &cpu_buffer->tail_page->list;
2501
2502 /*
2503 * tail page might be on reader page, we remove the next page
2504 * from the ring buffer
2505 */
2506 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
2507 tail_page = rb_list_head(tail_page->next);
2508 to_remove = tail_page;
2509
2510 /* start of pages to remove */
2511 first_page = list_entry(rb_list_head(to_remove->next),
2512 struct buffer_page, list);
2513
2514 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
2515 to_remove = rb_list_head(to_remove)->next;
2516 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
2517 }
2518 /* Read iterators need to reset themselves when some pages removed */
2519 cpu_buffer->pages_removed += nr_removed;
2520
2521 next_page = rb_list_head(to_remove)->next;
2522
2523 /*
2524 * Now we remove all pages between tail_page and next_page.
2525 * Make sure that we have head_bit value preserved for the
2526 * next page
2527 */
2528 tail_page->next = (struct list_head *)((unsigned long)next_page |
2529 head_bit);
2530 next_page = rb_list_head(next_page);
2531 next_page->prev = tail_page;
2532
2533 /* make sure pages points to a valid page in the ring buffer */
2534 cpu_buffer->pages = next_page;
2535
2536 /* update head page */
2537 if (head_bit)
2538 cpu_buffer->head_page = list_entry(next_page,
2539 struct buffer_page, list);
2540
2541 /* pages are removed, resume tracing and then free the pages */
2542 atomic_dec(&cpu_buffer->record_disabled);
2543 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
2544
2545 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
2546
2547 /* last buffer page to remove */
2548 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
2549 list);
2550 tmp_iter_page = first_page;
2551
2552 do {
2553 cond_resched();
2554
2555 to_remove_page = tmp_iter_page;
2556 rb_inc_page(&tmp_iter_page);
2557
2558 /* update the counters */
2559 page_entries = rb_page_entries(to_remove_page);
2560 if (page_entries) {
2561 /*
2562 * If something was added to this page, it was full
2563 * since it is not the tail page. So we deduct the
2564 * bytes consumed in ring buffer from here.
2565 * Increment overrun to account for the lost events.
2566 */
2567 local_add(page_entries, &cpu_buffer->overrun);
2568 local_sub(rb_page_commit(to_remove_page), &cpu_buffer->entries_bytes);
2569 local_inc(&cpu_buffer->pages_lost);
2570 }
2571
2572 /*
2573 * We have already removed references to this list item, just
2574 * free up the buffer_page and its page
2575 */
2576 free_buffer_page(to_remove_page);
2577 nr_removed--;
2578
2579 } while (to_remove_page != last_page);
2580
2581 RB_WARN_ON(cpu_buffer, nr_removed);
2582
2583 return nr_removed == 0;
2584}
2585
2586static bool
2587rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
2588{
2589 struct list_head *pages = &cpu_buffer->new_pages;
2590 unsigned long flags;
2591 bool success;
2592 int retries;
2593
2594 /* Can be called at early boot up, where interrupts must not been enabled */
2595 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2596 /*
2597 * We are holding the reader lock, so the reader page won't be swapped
2598 * in the ring buffer. Now we are racing with the writer trying to
2599 * move head page and the tail page.
2600 * We are going to adapt the reader page update process where:
2601 * 1. We first splice the start and end of list of new pages between
2602 * the head page and its previous page.
2603 * 2. We cmpxchg the prev_page->next to point from head page to the
2604 * start of new pages list.
2605 * 3. Finally, we update the head->prev to the end of new list.
2606 *
2607 * We will try this process 10 times, to make sure that we don't keep
2608 * spinning.
2609 */
2610 retries = 10;
2611 success = false;
2612 while (retries--) {
2613 struct list_head *head_page, *prev_page;
2614 struct list_head *last_page, *first_page;
2615 struct list_head *head_page_with_bit;
2616 struct buffer_page *hpage = rb_set_head_page(cpu_buffer);
2617
2618 if (!hpage)
2619 break;
2620 head_page = &hpage->list;
2621 prev_page = head_page->prev;
2622
2623 first_page = pages->next;
2624 last_page = pages->prev;
2625
2626 head_page_with_bit = (struct list_head *)
2627 ((unsigned long)head_page | RB_PAGE_HEAD);
2628
2629 last_page->next = head_page_with_bit;
2630 first_page->prev = prev_page;
2631
2632 /* caution: head_page_with_bit gets updated on cmpxchg failure */
2633 if (try_cmpxchg(&prev_page->next,
2634 &head_page_with_bit, first_page)) {
2635 /*
2636 * yay, we replaced the page pointer to our new list,
2637 * now, we just have to update to head page's prev
2638 * pointer to point to end of list
2639 */
2640 head_page->prev = last_page;
2641 success = true;
2642 break;
2643 }
2644 }
2645
2646 if (success)
2647 INIT_LIST_HEAD(pages);
2648 /*
2649 * If we weren't successful in adding in new pages, warn and stop
2650 * tracing
2651 */
2652 RB_WARN_ON(cpu_buffer, !success);
2653 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2654
2655 /* free pages if they weren't inserted */
2656 if (!success) {
2657 struct buffer_page *bpage, *tmp;
2658 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2659 list) {
2660 list_del_init(&bpage->list);
2661 free_buffer_page(bpage);
2662 }
2663 }
2664 return success;
2665}
2666
2667static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
2668{
2669 bool success;
2670
2671 if (cpu_buffer->nr_pages_to_update > 0)
2672 success = rb_insert_pages(cpu_buffer);
2673 else
2674 success = rb_remove_pages(cpu_buffer,
2675 -cpu_buffer->nr_pages_to_update);
2676
2677 if (success)
2678 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
2679}
2680
2681static void update_pages_handler(struct work_struct *work)
2682{
2683 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
2684 struct ring_buffer_per_cpu, update_pages_work);
2685 rb_update_pages(cpu_buffer);
2686 complete(&cpu_buffer->update_done);
2687}
2688
2689/**
2690 * ring_buffer_resize - resize the ring buffer
2691 * @buffer: the buffer to resize.
2692 * @size: the new size.
2693 * @cpu_id: the cpu buffer to resize
2694 *
2695 * Minimum size is 2 * buffer->subbuf_size.
2696 *
2697 * Returns 0 on success and < 0 on failure.
2698 */
2699int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
2700 int cpu_id)
2701{
2702 struct ring_buffer_per_cpu *cpu_buffer;
2703 unsigned long nr_pages;
2704 int cpu, err;
2705
2706 /*
2707 * Always succeed at resizing a non-existent buffer:
2708 */
2709 if (!buffer)
2710 return 0;
2711
2712 /* Make sure the requested buffer exists */
2713 if (cpu_id != RING_BUFFER_ALL_CPUS &&
2714 !cpumask_test_cpu(cpu_id, buffer->cpumask))
2715 return 0;
2716
2717 nr_pages = DIV_ROUND_UP(size, buffer->subbuf_size);
2718
2719 /* we need a minimum of two pages */
2720 if (nr_pages < 2)
2721 nr_pages = 2;
2722
2723 /* prevent another thread from changing buffer sizes */
2724 mutex_lock(&buffer->mutex);
2725 atomic_inc(&buffer->resizing);
2726
2727 if (cpu_id == RING_BUFFER_ALL_CPUS) {
2728 /*
2729 * Don't succeed if resizing is disabled, as a reader might be
2730 * manipulating the ring buffer and is expecting a sane state while
2731 * this is true.
2732 */
2733 for_each_buffer_cpu(buffer, cpu) {
2734 cpu_buffer = buffer->buffers[cpu];
2735 if (atomic_read(&cpu_buffer->resize_disabled)) {
2736 err = -EBUSY;
2737 goto out_err_unlock;
2738 }
2739 }
2740
2741 /* calculate the pages to update */
2742 for_each_buffer_cpu(buffer, cpu) {
2743 cpu_buffer = buffer->buffers[cpu];
2744
2745 cpu_buffer->nr_pages_to_update = nr_pages -
2746 cpu_buffer->nr_pages;
2747 /*
2748 * nothing more to do for removing pages or no update
2749 */
2750 if (cpu_buffer->nr_pages_to_update <= 0)
2751 continue;
2752 /*
2753 * to add pages, make sure all new pages can be
2754 * allocated without receiving ENOMEM
2755 */
2756 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2757 if (__rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2758 &cpu_buffer->new_pages)) {
2759 /* not enough memory for new pages */
2760 err = -ENOMEM;
2761 goto out_err;
2762 }
2763
2764 cond_resched();
2765 }
2766
2767 cpus_read_lock();
2768 /*
2769 * Fire off all the required work handlers
2770 * We can't schedule on offline CPUs, but it's not necessary
2771 * since we can change their buffer sizes without any race.
2772 */
2773 for_each_buffer_cpu(buffer, cpu) {
2774 cpu_buffer = buffer->buffers[cpu];
2775 if (!cpu_buffer->nr_pages_to_update)
2776 continue;
2777
2778 /* Can't run something on an offline CPU. */
2779 if (!cpu_online(cpu)) {
2780 rb_update_pages(cpu_buffer);
2781 cpu_buffer->nr_pages_to_update = 0;
2782 } else {
2783 /* Run directly if possible. */
2784 migrate_disable();
2785 if (cpu != smp_processor_id()) {
2786 migrate_enable();
2787 schedule_work_on(cpu,
2788 &cpu_buffer->update_pages_work);
2789 } else {
2790 update_pages_handler(&cpu_buffer->update_pages_work);
2791 migrate_enable();
2792 }
2793 }
2794 }
2795
2796 /* wait for all the updates to complete */
2797 for_each_buffer_cpu(buffer, cpu) {
2798 cpu_buffer = buffer->buffers[cpu];
2799 if (!cpu_buffer->nr_pages_to_update)
2800 continue;
2801
2802 if (cpu_online(cpu))
2803 wait_for_completion(&cpu_buffer->update_done);
2804 cpu_buffer->nr_pages_to_update = 0;
2805 }
2806
2807 cpus_read_unlock();
2808 } else {
2809 cpu_buffer = buffer->buffers[cpu_id];
2810
2811 if (nr_pages == cpu_buffer->nr_pages)
2812 goto out;
2813
2814 /*
2815 * Don't succeed if resizing is disabled, as a reader might be
2816 * manipulating the ring buffer and is expecting a sane state while
2817 * this is true.
2818 */
2819 if (atomic_read(&cpu_buffer->resize_disabled)) {
2820 err = -EBUSY;
2821 goto out_err_unlock;
2822 }
2823
2824 cpu_buffer->nr_pages_to_update = nr_pages -
2825 cpu_buffer->nr_pages;
2826
2827 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2828 if (cpu_buffer->nr_pages_to_update > 0 &&
2829 __rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2830 &cpu_buffer->new_pages)) {
2831 err = -ENOMEM;
2832 goto out_err;
2833 }
2834
2835 cpus_read_lock();
2836
2837 /* Can't run something on an offline CPU. */
2838 if (!cpu_online(cpu_id))
2839 rb_update_pages(cpu_buffer);
2840 else {
2841 /* Run directly if possible. */
2842 migrate_disable();
2843 if (cpu_id == smp_processor_id()) {
2844 rb_update_pages(cpu_buffer);
2845 migrate_enable();
2846 } else {
2847 migrate_enable();
2848 schedule_work_on(cpu_id,
2849 &cpu_buffer->update_pages_work);
2850 wait_for_completion(&cpu_buffer->update_done);
2851 }
2852 }
2853
2854 cpu_buffer->nr_pages_to_update = 0;
2855 cpus_read_unlock();
2856 }
2857
2858 out:
2859 /*
2860 * The ring buffer resize can happen with the ring buffer
2861 * enabled, so that the update disturbs the tracing as little
2862 * as possible. But if the buffer is disabled, we do not need
2863 * to worry about that, and we can take the time to verify
2864 * that the buffer is not corrupt.
2865 */
2866 if (atomic_read(&buffer->record_disabled)) {
2867 atomic_inc(&buffer->record_disabled);
2868 /*
2869 * Even though the buffer was disabled, we must make sure
2870 * that it is truly disabled before calling rb_check_pages.
2871 * There could have been a race between checking
2872 * record_disable and incrementing it.
2873 */
2874 synchronize_rcu();
2875 for_each_buffer_cpu(buffer, cpu) {
2876 unsigned long flags;
2877
2878 cpu_buffer = buffer->buffers[cpu];
2879 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2880 rb_check_pages(cpu_buffer);
2881 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2882 }
2883 atomic_dec(&buffer->record_disabled);
2884 }
2885
2886 atomic_dec(&buffer->resizing);
2887 mutex_unlock(&buffer->mutex);
2888 return 0;
2889
2890 out_err:
2891 for_each_buffer_cpu(buffer, cpu) {
2892 struct buffer_page *bpage, *tmp;
2893
2894 cpu_buffer = buffer->buffers[cpu];
2895 cpu_buffer->nr_pages_to_update = 0;
2896
2897 if (list_empty(&cpu_buffer->new_pages))
2898 continue;
2899
2900 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2901 list) {
2902 list_del_init(&bpage->list);
2903 free_buffer_page(bpage);
2904 }
2905 }
2906 out_err_unlock:
2907 atomic_dec(&buffer->resizing);
2908 mutex_unlock(&buffer->mutex);
2909 return err;
2910}
2911EXPORT_SYMBOL_GPL(ring_buffer_resize);
2912
2913void ring_buffer_change_overwrite(struct trace_buffer *buffer, int val)
2914{
2915 mutex_lock(&buffer->mutex);
2916 if (val)
2917 buffer->flags |= RB_FL_OVERWRITE;
2918 else
2919 buffer->flags &= ~RB_FL_OVERWRITE;
2920 mutex_unlock(&buffer->mutex);
2921}
2922EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
2923
2924static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
2925{
2926 return bpage->page->data + index;
2927}
2928
2929static __always_inline struct ring_buffer_event *
2930rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
2931{
2932 return __rb_page_index(cpu_buffer->reader_page,
2933 cpu_buffer->reader_page->read);
2934}
2935
2936static struct ring_buffer_event *
2937rb_iter_head_event(struct ring_buffer_iter *iter)
2938{
2939 struct ring_buffer_event *event;
2940 struct buffer_page *iter_head_page = iter->head_page;
2941 unsigned long commit;
2942 unsigned length;
2943
2944 if (iter->head != iter->next_event)
2945 return iter->event;
2946
2947 /*
2948 * When the writer goes across pages, it issues a cmpxchg which
2949 * is a mb(), which will synchronize with the rmb here.
2950 * (see rb_tail_page_update() and __rb_reserve_next())
2951 */
2952 commit = rb_page_commit(iter_head_page);
2953 smp_rmb();
2954
2955 /* An event needs to be at least 8 bytes in size */
2956 if (iter->head > commit - 8)
2957 goto reset;
2958
2959 event = __rb_page_index(iter_head_page, iter->head);
2960 length = rb_event_length(event);
2961
2962 /*
2963 * READ_ONCE() doesn't work on functions and we don't want the
2964 * compiler doing any crazy optimizations with length.
2965 */
2966 barrier();
2967
2968 if ((iter->head + length) > commit || length > iter->event_size)
2969 /* Writer corrupted the read? */
2970 goto reset;
2971
2972 memcpy(iter->event, event, length);
2973 /*
2974 * If the page stamp is still the same after this rmb() then the
2975 * event was safely copied without the writer entering the page.
2976 */
2977 smp_rmb();
2978
2979 /* Make sure the page didn't change since we read this */
2980 if (iter->page_stamp != iter_head_page->page->time_stamp ||
2981 commit > rb_page_commit(iter_head_page))
2982 goto reset;
2983
2984 iter->next_event = iter->head + length;
2985 return iter->event;
2986 reset:
2987 /* Reset to the beginning */
2988 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2989 iter->head = 0;
2990 iter->next_event = 0;
2991 iter->missed_events = 1;
2992 return NULL;
2993}
2994
2995/* Size is determined by what has been committed */
2996static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
2997{
2998 return rb_page_commit(bpage) & ~RB_MISSED_MASK;
2999}
3000
3001static __always_inline unsigned
3002rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
3003{
3004 return rb_page_commit(cpu_buffer->commit_page);
3005}
3006
3007static __always_inline unsigned
3008rb_event_index(struct ring_buffer_per_cpu *cpu_buffer, struct ring_buffer_event *event)
3009{
3010 unsigned long addr = (unsigned long)event;
3011
3012 addr &= (PAGE_SIZE << cpu_buffer->buffer->subbuf_order) - 1;
3013
3014 return addr - BUF_PAGE_HDR_SIZE;
3015}
3016
3017static void rb_inc_iter(struct ring_buffer_iter *iter)
3018{
3019 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3020
3021 /*
3022 * The iterator could be on the reader page (it starts there).
3023 * But the head could have moved, since the reader was
3024 * found. Check for this case and assign the iterator
3025 * to the head page instead of next.
3026 */
3027 if (iter->head_page == cpu_buffer->reader_page)
3028 iter->head_page = rb_set_head_page(cpu_buffer);
3029 else
3030 rb_inc_page(&iter->head_page);
3031
3032 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
3033 iter->head = 0;
3034 iter->next_event = 0;
3035}
3036
3037/* Return the index into the sub-buffers for a given sub-buffer */
3038static int rb_meta_subbuf_idx(struct ring_buffer_meta *meta, void *subbuf)
3039{
3040 void *subbuf_array;
3041
3042 subbuf_array = (void *)meta + sizeof(int) * meta->nr_subbufs;
3043 subbuf_array = (void *)ALIGN((unsigned long)subbuf_array, meta->subbuf_size);
3044 return (subbuf - subbuf_array) / meta->subbuf_size;
3045}
3046
3047static void rb_update_meta_head(struct ring_buffer_per_cpu *cpu_buffer,
3048 struct buffer_page *next_page)
3049{
3050 struct ring_buffer_meta *meta = cpu_buffer->ring_meta;
3051 unsigned long old_head = (unsigned long)next_page->page;
3052 unsigned long new_head;
3053
3054 rb_inc_page(&next_page);
3055 new_head = (unsigned long)next_page->page;
3056
3057 /*
3058 * Only move it forward once, if something else came in and
3059 * moved it forward, then we don't want to touch it.
3060 */
3061 (void)cmpxchg(&meta->head_buffer, old_head, new_head);
3062}
3063
3064static void rb_update_meta_reader(struct ring_buffer_per_cpu *cpu_buffer,
3065 struct buffer_page *reader)
3066{
3067 struct ring_buffer_meta *meta = cpu_buffer->ring_meta;
3068 void *old_reader = cpu_buffer->reader_page->page;
3069 void *new_reader = reader->page;
3070 int id;
3071
3072 id = reader->id;
3073 cpu_buffer->reader_page->id = id;
3074 reader->id = 0;
3075
3076 meta->buffers[0] = rb_meta_subbuf_idx(meta, new_reader);
3077 meta->buffers[id] = rb_meta_subbuf_idx(meta, old_reader);
3078
3079 /* The head pointer is the one after the reader */
3080 rb_update_meta_head(cpu_buffer, reader);
3081}
3082
3083/*
3084 * rb_handle_head_page - writer hit the head page
3085 *
3086 * Returns: +1 to retry page
3087 * 0 to continue
3088 * -1 on error
3089 */
3090static int
3091rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
3092 struct buffer_page *tail_page,
3093 struct buffer_page *next_page)
3094{
3095 struct buffer_page *new_head;
3096 int entries;
3097 int type;
3098 int ret;
3099
3100 entries = rb_page_entries(next_page);
3101
3102 /*
3103 * The hard part is here. We need to move the head
3104 * forward, and protect against both readers on
3105 * other CPUs and writers coming in via interrupts.
3106 */
3107 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
3108 RB_PAGE_HEAD);
3109
3110 /*
3111 * type can be one of four:
3112 * NORMAL - an interrupt already moved it for us
3113 * HEAD - we are the first to get here.
3114 * UPDATE - we are the interrupt interrupting
3115 * a current move.
3116 * MOVED - a reader on another CPU moved the next
3117 * pointer to its reader page. Give up
3118 * and try again.
3119 */
3120
3121 switch (type) {
3122 case RB_PAGE_HEAD:
3123 /*
3124 * We changed the head to UPDATE, thus
3125 * it is our responsibility to update
3126 * the counters.
3127 */
3128 local_add(entries, &cpu_buffer->overrun);
3129 local_sub(rb_page_commit(next_page), &cpu_buffer->entries_bytes);
3130 local_inc(&cpu_buffer->pages_lost);
3131
3132 if (cpu_buffer->ring_meta)
3133 rb_update_meta_head(cpu_buffer, next_page);
3134 /*
3135 * The entries will be zeroed out when we move the
3136 * tail page.
3137 */
3138
3139 /* still more to do */
3140 break;
3141
3142 case RB_PAGE_UPDATE:
3143 /*
3144 * This is an interrupt that interrupt the
3145 * previous update. Still more to do.
3146 */
3147 break;
3148 case RB_PAGE_NORMAL:
3149 /*
3150 * An interrupt came in before the update
3151 * and processed this for us.
3152 * Nothing left to do.
3153 */
3154 return 1;
3155 case RB_PAGE_MOVED:
3156 /*
3157 * The reader is on another CPU and just did
3158 * a swap with our next_page.
3159 * Try again.
3160 */
3161 return 1;
3162 default:
3163 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
3164 return -1;
3165 }
3166
3167 /*
3168 * Now that we are here, the old head pointer is
3169 * set to UPDATE. This will keep the reader from
3170 * swapping the head page with the reader page.
3171 * The reader (on another CPU) will spin till
3172 * we are finished.
3173 *
3174 * We just need to protect against interrupts
3175 * doing the job. We will set the next pointer
3176 * to HEAD. After that, we set the old pointer
3177 * to NORMAL, but only if it was HEAD before.
3178 * otherwise we are an interrupt, and only
3179 * want the outer most commit to reset it.
3180 */
3181 new_head = next_page;
3182 rb_inc_page(&new_head);
3183
3184 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
3185 RB_PAGE_NORMAL);
3186
3187 /*
3188 * Valid returns are:
3189 * HEAD - an interrupt came in and already set it.
3190 * NORMAL - One of two things:
3191 * 1) We really set it.
3192 * 2) A bunch of interrupts came in and moved
3193 * the page forward again.
3194 */
3195 switch (ret) {
3196 case RB_PAGE_HEAD:
3197 case RB_PAGE_NORMAL:
3198 /* OK */
3199 break;
3200 default:
3201 RB_WARN_ON(cpu_buffer, 1);
3202 return -1;
3203 }
3204
3205 /*
3206 * It is possible that an interrupt came in,
3207 * set the head up, then more interrupts came in
3208 * and moved it again. When we get back here,
3209 * the page would have been set to NORMAL but we
3210 * just set it back to HEAD.
3211 *
3212 * How do you detect this? Well, if that happened
3213 * the tail page would have moved.
3214 */
3215 if (ret == RB_PAGE_NORMAL) {
3216 struct buffer_page *buffer_tail_page;
3217
3218 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
3219 /*
3220 * If the tail had moved passed next, then we need
3221 * to reset the pointer.
3222 */
3223 if (buffer_tail_page != tail_page &&
3224 buffer_tail_page != next_page)
3225 rb_head_page_set_normal(cpu_buffer, new_head,
3226 next_page,
3227 RB_PAGE_HEAD);
3228 }
3229
3230 /*
3231 * If this was the outer most commit (the one that
3232 * changed the original pointer from HEAD to UPDATE),
3233 * then it is up to us to reset it to NORMAL.
3234 */
3235 if (type == RB_PAGE_HEAD) {
3236 ret = rb_head_page_set_normal(cpu_buffer, next_page,
3237 tail_page,
3238 RB_PAGE_UPDATE);
3239 if (RB_WARN_ON(cpu_buffer,
3240 ret != RB_PAGE_UPDATE))
3241 return -1;
3242 }
3243
3244 return 0;
3245}
3246
3247static inline void
3248rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
3249 unsigned long tail, struct rb_event_info *info)
3250{
3251 unsigned long bsize = READ_ONCE(cpu_buffer->buffer->subbuf_size);
3252 struct buffer_page *tail_page = info->tail_page;
3253 struct ring_buffer_event *event;
3254 unsigned long length = info->length;
3255
3256 /*
3257 * Only the event that crossed the page boundary
3258 * must fill the old tail_page with padding.
3259 */
3260 if (tail >= bsize) {
3261 /*
3262 * If the page was filled, then we still need
3263 * to update the real_end. Reset it to zero
3264 * and the reader will ignore it.
3265 */
3266 if (tail == bsize)
3267 tail_page->real_end = 0;
3268
3269 local_sub(length, &tail_page->write);
3270 return;
3271 }
3272
3273 event = __rb_page_index(tail_page, tail);
3274
3275 /*
3276 * Save the original length to the meta data.
3277 * This will be used by the reader to add lost event
3278 * counter.
3279 */
3280 tail_page->real_end = tail;
3281
3282 /*
3283 * If this event is bigger than the minimum size, then
3284 * we need to be careful that we don't subtract the
3285 * write counter enough to allow another writer to slip
3286 * in on this page.
3287 * We put in a discarded commit instead, to make sure
3288 * that this space is not used again, and this space will
3289 * not be accounted into 'entries_bytes'.
3290 *
3291 * If we are less than the minimum size, we don't need to
3292 * worry about it.
3293 */
3294 if (tail > (bsize - RB_EVNT_MIN_SIZE)) {
3295 /* No room for any events */
3296
3297 /* Mark the rest of the page with padding */
3298 rb_event_set_padding(event);
3299
3300 /* Make sure the padding is visible before the write update */
3301 smp_wmb();
3302
3303 /* Set the write back to the previous setting */
3304 local_sub(length, &tail_page->write);
3305 return;
3306 }
3307
3308 /* Put in a discarded event */
3309 event->array[0] = (bsize - tail) - RB_EVNT_HDR_SIZE;
3310 event->type_len = RINGBUF_TYPE_PADDING;
3311 /* time delta must be non zero */
3312 event->time_delta = 1;
3313
3314 /* account for padding bytes */
3315 local_add(bsize - tail, &cpu_buffer->entries_bytes);
3316
3317 /* Make sure the padding is visible before the tail_page->write update */
3318 smp_wmb();
3319
3320 /* Set write to end of buffer */
3321 length = (tail + length) - bsize;
3322 local_sub(length, &tail_page->write);
3323}
3324
3325static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
3326
3327/*
3328 * This is the slow path, force gcc not to inline it.
3329 */
3330static noinline struct ring_buffer_event *
3331rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
3332 unsigned long tail, struct rb_event_info *info)
3333{
3334 struct buffer_page *tail_page = info->tail_page;
3335 struct buffer_page *commit_page = cpu_buffer->commit_page;
3336 struct trace_buffer *buffer = cpu_buffer->buffer;
3337 struct buffer_page *next_page;
3338 int ret;
3339
3340 next_page = tail_page;
3341
3342 rb_inc_page(&next_page);
3343
3344 /*
3345 * If for some reason, we had an interrupt storm that made
3346 * it all the way around the buffer, bail, and warn
3347 * about it.
3348 */
3349 if (unlikely(next_page == commit_page)) {
3350 local_inc(&cpu_buffer->commit_overrun);
3351 goto out_reset;
3352 }
3353
3354 /*
3355 * This is where the fun begins!
3356 *
3357 * We are fighting against races between a reader that
3358 * could be on another CPU trying to swap its reader
3359 * page with the buffer head.
3360 *
3361 * We are also fighting against interrupts coming in and
3362 * moving the head or tail on us as well.
3363 *
3364 * If the next page is the head page then we have filled
3365 * the buffer, unless the commit page is still on the
3366 * reader page.
3367 */
3368 if (rb_is_head_page(next_page, &tail_page->list)) {
3369
3370 /*
3371 * If the commit is not on the reader page, then
3372 * move the header page.
3373 */
3374 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
3375 /*
3376 * If we are not in overwrite mode,
3377 * this is easy, just stop here.
3378 */
3379 if (!(buffer->flags & RB_FL_OVERWRITE)) {
3380 local_inc(&cpu_buffer->dropped_events);
3381 goto out_reset;
3382 }
3383
3384 ret = rb_handle_head_page(cpu_buffer,
3385 tail_page,
3386 next_page);
3387 if (ret < 0)
3388 goto out_reset;
3389 if (ret)
3390 goto out_again;
3391 } else {
3392 /*
3393 * We need to be careful here too. The
3394 * commit page could still be on the reader
3395 * page. We could have a small buffer, and
3396 * have filled up the buffer with events
3397 * from interrupts and such, and wrapped.
3398 *
3399 * Note, if the tail page is also on the
3400 * reader_page, we let it move out.
3401 */
3402 if (unlikely((cpu_buffer->commit_page !=
3403 cpu_buffer->tail_page) &&
3404 (cpu_buffer->commit_page ==
3405 cpu_buffer->reader_page))) {
3406 local_inc(&cpu_buffer->commit_overrun);
3407 goto out_reset;
3408 }
3409 }
3410 }
3411
3412 rb_tail_page_update(cpu_buffer, tail_page, next_page);
3413
3414 out_again:
3415
3416 rb_reset_tail(cpu_buffer, tail, info);
3417
3418 /* Commit what we have for now. */
3419 rb_end_commit(cpu_buffer);
3420 /* rb_end_commit() decs committing */
3421 local_inc(&cpu_buffer->committing);
3422
3423 /* fail and let the caller try again */
3424 return ERR_PTR(-EAGAIN);
3425
3426 out_reset:
3427 /* reset write */
3428 rb_reset_tail(cpu_buffer, tail, info);
3429
3430 return NULL;
3431}
3432
3433/* Slow path */
3434static struct ring_buffer_event *
3435rb_add_time_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3436 struct ring_buffer_event *event, u64 delta, bool abs)
3437{
3438 if (abs)
3439 event->type_len = RINGBUF_TYPE_TIME_STAMP;
3440 else
3441 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
3442
3443 /* Not the first event on the page, or not delta? */
3444 if (abs || rb_event_index(cpu_buffer, event)) {
3445 event->time_delta = delta & TS_MASK;
3446 event->array[0] = delta >> TS_SHIFT;
3447 } else {
3448 /* nope, just zero it */
3449 event->time_delta = 0;
3450 event->array[0] = 0;
3451 }
3452
3453 return skip_time_extend(event);
3454}
3455
3456#ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
3457static inline bool sched_clock_stable(void)
3458{
3459 return true;
3460}
3461#endif
3462
3463static void
3464rb_check_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
3465 struct rb_event_info *info)
3466{
3467 u64 write_stamp;
3468
3469 WARN_ONCE(1, "Delta way too big! %llu ts=%llu before=%llu after=%llu write stamp=%llu\n%s",
3470 (unsigned long long)info->delta,
3471 (unsigned long long)info->ts,
3472 (unsigned long long)info->before,
3473 (unsigned long long)info->after,
3474 (unsigned long long)({rb_time_read(&cpu_buffer->write_stamp, &write_stamp); write_stamp;}),
3475 sched_clock_stable() ? "" :
3476 "If you just came from a suspend/resume,\n"
3477 "please switch to the trace global clock:\n"
3478 " echo global > /sys/kernel/tracing/trace_clock\n"
3479 "or add trace_clock=global to the kernel command line\n");
3480}
3481
3482static void rb_add_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
3483 struct ring_buffer_event **event,
3484 struct rb_event_info *info,
3485 u64 *delta,
3486 unsigned int *length)
3487{
3488 bool abs = info->add_timestamp &
3489 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE);
3490
3491 if (unlikely(info->delta > (1ULL << 59))) {
3492 /*
3493 * Some timers can use more than 59 bits, and when a timestamp
3494 * is added to the buffer, it will lose those bits.
3495 */
3496 if (abs && (info->ts & TS_MSB)) {
3497 info->delta &= ABS_TS_MASK;
3498
3499 /* did the clock go backwards */
3500 } else if (info->before == info->after && info->before > info->ts) {
3501 /* not interrupted */
3502 static int once;
3503
3504 /*
3505 * This is possible with a recalibrating of the TSC.
3506 * Do not produce a call stack, but just report it.
3507 */
3508 if (!once) {
3509 once++;
3510 pr_warn("Ring buffer clock went backwards: %llu -> %llu\n",
3511 info->before, info->ts);
3512 }
3513 } else
3514 rb_check_timestamp(cpu_buffer, info);
3515 if (!abs)
3516 info->delta = 0;
3517 }
3518 *event = rb_add_time_stamp(cpu_buffer, *event, info->delta, abs);
3519 *length -= RB_LEN_TIME_EXTEND;
3520 *delta = 0;
3521}
3522
3523/**
3524 * rb_update_event - update event type and data
3525 * @cpu_buffer: The per cpu buffer of the @event
3526 * @event: the event to update
3527 * @info: The info to update the @event with (contains length and delta)
3528 *
3529 * Update the type and data fields of the @event. The length
3530 * is the actual size that is written to the ring buffer,
3531 * and with this, we can determine what to place into the
3532 * data field.
3533 */
3534static void
3535rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
3536 struct ring_buffer_event *event,
3537 struct rb_event_info *info)
3538{
3539 unsigned length = info->length;
3540 u64 delta = info->delta;
3541 unsigned int nest = local_read(&cpu_buffer->committing) - 1;
3542
3543 if (!WARN_ON_ONCE(nest >= MAX_NEST))
3544 cpu_buffer->event_stamp[nest] = info->ts;
3545
3546 /*
3547 * If we need to add a timestamp, then we
3548 * add it to the start of the reserved space.
3549 */
3550 if (unlikely(info->add_timestamp))
3551 rb_add_timestamp(cpu_buffer, &event, info, &delta, &length);
3552
3553 event->time_delta = delta;
3554 length -= RB_EVNT_HDR_SIZE;
3555 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
3556 event->type_len = 0;
3557 event->array[0] = length;
3558 } else
3559 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
3560}
3561
3562static unsigned rb_calculate_event_length(unsigned length)
3563{
3564 struct ring_buffer_event event; /* Used only for sizeof array */
3565
3566 /* zero length can cause confusions */
3567 if (!length)
3568 length++;
3569
3570 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
3571 length += sizeof(event.array[0]);
3572
3573 length += RB_EVNT_HDR_SIZE;
3574 length = ALIGN(length, RB_ARCH_ALIGNMENT);
3575
3576 /*
3577 * In case the time delta is larger than the 27 bits for it
3578 * in the header, we need to add a timestamp. If another
3579 * event comes in when trying to discard this one to increase
3580 * the length, then the timestamp will be added in the allocated
3581 * space of this event. If length is bigger than the size needed
3582 * for the TIME_EXTEND, then padding has to be used. The events
3583 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
3584 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
3585 * As length is a multiple of 4, we only need to worry if it
3586 * is 12 (RB_LEN_TIME_EXTEND + 4).
3587 */
3588 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
3589 length += RB_ALIGNMENT;
3590
3591 return length;
3592}
3593
3594static inline bool
3595rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
3596 struct ring_buffer_event *event)
3597{
3598 unsigned long new_index, old_index;
3599 struct buffer_page *bpage;
3600 unsigned long addr;
3601
3602 new_index = rb_event_index(cpu_buffer, event);
3603 old_index = new_index + rb_event_ts_length(event);
3604 addr = (unsigned long)event;
3605 addr &= ~((PAGE_SIZE << cpu_buffer->buffer->subbuf_order) - 1);
3606
3607 bpage = READ_ONCE(cpu_buffer->tail_page);
3608
3609 /*
3610 * Make sure the tail_page is still the same and
3611 * the next write location is the end of this event
3612 */
3613 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
3614 unsigned long write_mask =
3615 local_read(&bpage->write) & ~RB_WRITE_MASK;
3616 unsigned long event_length = rb_event_length(event);
3617
3618 /*
3619 * For the before_stamp to be different than the write_stamp
3620 * to make sure that the next event adds an absolute
3621 * value and does not rely on the saved write stamp, which
3622 * is now going to be bogus.
3623 *
3624 * By setting the before_stamp to zero, the next event
3625 * is not going to use the write_stamp and will instead
3626 * create an absolute timestamp. This means there's no
3627 * reason to update the wirte_stamp!
3628 */
3629 rb_time_set(&cpu_buffer->before_stamp, 0);
3630
3631 /*
3632 * If an event were to come in now, it would see that the
3633 * write_stamp and the before_stamp are different, and assume
3634 * that this event just added itself before updating
3635 * the write stamp. The interrupting event will fix the
3636 * write stamp for us, and use an absolute timestamp.
3637 */
3638
3639 /*
3640 * This is on the tail page. It is possible that
3641 * a write could come in and move the tail page
3642 * and write to the next page. That is fine
3643 * because we just shorten what is on this page.
3644 */
3645 old_index += write_mask;
3646 new_index += write_mask;
3647
3648 /* caution: old_index gets updated on cmpxchg failure */
3649 if (local_try_cmpxchg(&bpage->write, &old_index, new_index)) {
3650 /* update counters */
3651 local_sub(event_length, &cpu_buffer->entries_bytes);
3652 return true;
3653 }
3654 }
3655
3656 /* could not discard */
3657 return false;
3658}
3659
3660static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
3661{
3662 local_inc(&cpu_buffer->committing);
3663 local_inc(&cpu_buffer->commits);
3664}
3665
3666static __always_inline void
3667rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
3668{
3669 unsigned long max_count;
3670
3671 /*
3672 * We only race with interrupts and NMIs on this CPU.
3673 * If we own the commit event, then we can commit
3674 * all others that interrupted us, since the interruptions
3675 * are in stack format (they finish before they come
3676 * back to us). This allows us to do a simple loop to
3677 * assign the commit to the tail.
3678 */
3679 again:
3680 max_count = cpu_buffer->nr_pages * 100;
3681
3682 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
3683 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
3684 return;
3685 if (RB_WARN_ON(cpu_buffer,
3686 rb_is_reader_page(cpu_buffer->tail_page)))
3687 return;
3688 /*
3689 * No need for a memory barrier here, as the update
3690 * of the tail_page did it for this page.
3691 */
3692 local_set(&cpu_buffer->commit_page->page->commit,
3693 rb_page_write(cpu_buffer->commit_page));
3694 rb_inc_page(&cpu_buffer->commit_page);
3695 if (cpu_buffer->ring_meta) {
3696 struct ring_buffer_meta *meta = cpu_buffer->ring_meta;
3697 meta->commit_buffer = (unsigned long)cpu_buffer->commit_page->page;
3698 }
3699 /* add barrier to keep gcc from optimizing too much */
3700 barrier();
3701 }
3702 while (rb_commit_index(cpu_buffer) !=
3703 rb_page_write(cpu_buffer->commit_page)) {
3704
3705 /* Make sure the readers see the content of what is committed. */
3706 smp_wmb();
3707 local_set(&cpu_buffer->commit_page->page->commit,
3708 rb_page_write(cpu_buffer->commit_page));
3709 RB_WARN_ON(cpu_buffer,
3710 local_read(&cpu_buffer->commit_page->page->commit) &
3711 ~RB_WRITE_MASK);
3712 barrier();
3713 }
3714
3715 /* again, keep gcc from optimizing */
3716 barrier();
3717
3718 /*
3719 * If an interrupt came in just after the first while loop
3720 * and pushed the tail page forward, we will be left with
3721 * a dangling commit that will never go forward.
3722 */
3723 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
3724 goto again;
3725}
3726
3727static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
3728{
3729 unsigned long commits;
3730
3731 if (RB_WARN_ON(cpu_buffer,
3732 !local_read(&cpu_buffer->committing)))
3733 return;
3734
3735 again:
3736 commits = local_read(&cpu_buffer->commits);
3737 /* synchronize with interrupts */
3738 barrier();
3739 if (local_read(&cpu_buffer->committing) == 1)
3740 rb_set_commit_to_write(cpu_buffer);
3741
3742 local_dec(&cpu_buffer->committing);
3743
3744 /* synchronize with interrupts */
3745 barrier();
3746
3747 /*
3748 * Need to account for interrupts coming in between the
3749 * updating of the commit page and the clearing of the
3750 * committing counter.
3751 */
3752 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
3753 !local_read(&cpu_buffer->committing)) {
3754 local_inc(&cpu_buffer->committing);
3755 goto again;
3756 }
3757}
3758
3759static inline void rb_event_discard(struct ring_buffer_event *event)
3760{
3761 if (extended_time(event))
3762 event = skip_time_extend(event);
3763
3764 /* array[0] holds the actual length for the discarded event */
3765 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
3766 event->type_len = RINGBUF_TYPE_PADDING;
3767 /* time delta must be non zero */
3768 if (!event->time_delta)
3769 event->time_delta = 1;
3770}
3771
3772static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer)
3773{
3774 local_inc(&cpu_buffer->entries);
3775 rb_end_commit(cpu_buffer);
3776}
3777
3778static __always_inline void
3779rb_wakeups(struct trace_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
3780{
3781 if (buffer->irq_work.waiters_pending) {
3782 buffer->irq_work.waiters_pending = false;
3783 /* irq_work_queue() supplies it's own memory barriers */
3784 irq_work_queue(&buffer->irq_work.work);
3785 }
3786
3787 if (cpu_buffer->irq_work.waiters_pending) {
3788 cpu_buffer->irq_work.waiters_pending = false;
3789 /* irq_work_queue() supplies it's own memory barriers */
3790 irq_work_queue(&cpu_buffer->irq_work.work);
3791 }
3792
3793 if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
3794 return;
3795
3796 if (cpu_buffer->reader_page == cpu_buffer->commit_page)
3797 return;
3798
3799 if (!cpu_buffer->irq_work.full_waiters_pending)
3800 return;
3801
3802 cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
3803
3804 if (!full_hit(buffer, cpu_buffer->cpu, cpu_buffer->shortest_full))
3805 return;
3806
3807 cpu_buffer->irq_work.wakeup_full = true;
3808 cpu_buffer->irq_work.full_waiters_pending = false;
3809 /* irq_work_queue() supplies it's own memory barriers */
3810 irq_work_queue(&cpu_buffer->irq_work.work);
3811}
3812
3813#ifdef CONFIG_RING_BUFFER_RECORD_RECURSION
3814# define do_ring_buffer_record_recursion() \
3815 do_ftrace_record_recursion(_THIS_IP_, _RET_IP_)
3816#else
3817# define do_ring_buffer_record_recursion() do { } while (0)
3818#endif
3819
3820/*
3821 * The lock and unlock are done within a preempt disable section.
3822 * The current_context per_cpu variable can only be modified
3823 * by the current task between lock and unlock. But it can
3824 * be modified more than once via an interrupt. To pass this
3825 * information from the lock to the unlock without having to
3826 * access the 'in_interrupt()' functions again (which do show
3827 * a bit of overhead in something as critical as function tracing,
3828 * we use a bitmask trick.
3829 *
3830 * bit 1 = NMI context
3831 * bit 2 = IRQ context
3832 * bit 3 = SoftIRQ context
3833 * bit 4 = normal context.
3834 *
3835 * This works because this is the order of contexts that can
3836 * preempt other contexts. A SoftIRQ never preempts an IRQ
3837 * context.
3838 *
3839 * When the context is determined, the corresponding bit is
3840 * checked and set (if it was set, then a recursion of that context
3841 * happened).
3842 *
3843 * On unlock, we need to clear this bit. To do so, just subtract
3844 * 1 from the current_context and AND it to itself.
3845 *
3846 * (binary)
3847 * 101 - 1 = 100
3848 * 101 & 100 = 100 (clearing bit zero)
3849 *
3850 * 1010 - 1 = 1001
3851 * 1010 & 1001 = 1000 (clearing bit 1)
3852 *
3853 * The least significant bit can be cleared this way, and it
3854 * just so happens that it is the same bit corresponding to
3855 * the current context.
3856 *
3857 * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
3858 * is set when a recursion is detected at the current context, and if
3859 * the TRANSITION bit is already set, it will fail the recursion.
3860 * This is needed because there's a lag between the changing of
3861 * interrupt context and updating the preempt count. In this case,
3862 * a false positive will be found. To handle this, one extra recursion
3863 * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
3864 * bit is already set, then it is considered a recursion and the function
3865 * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
3866 *
3867 * On the trace_recursive_unlock(), the TRANSITION bit will be the first
3868 * to be cleared. Even if it wasn't the context that set it. That is,
3869 * if an interrupt comes in while NORMAL bit is set and the ring buffer
3870 * is called before preempt_count() is updated, since the check will
3871 * be on the NORMAL bit, the TRANSITION bit will then be set. If an
3872 * NMI then comes in, it will set the NMI bit, but when the NMI code
3873 * does the trace_recursive_unlock() it will clear the TRANSITION bit
3874 * and leave the NMI bit set. But this is fine, because the interrupt
3875 * code that set the TRANSITION bit will then clear the NMI bit when it
3876 * calls trace_recursive_unlock(). If another NMI comes in, it will
3877 * set the TRANSITION bit and continue.
3878 *
3879 * Note: The TRANSITION bit only handles a single transition between context.
3880 */
3881
3882static __always_inline bool
3883trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
3884{
3885 unsigned int val = cpu_buffer->current_context;
3886 int bit = interrupt_context_level();
3887
3888 bit = RB_CTX_NORMAL - bit;
3889
3890 if (unlikely(val & (1 << (bit + cpu_buffer->nest)))) {
3891 /*
3892 * It is possible that this was called by transitioning
3893 * between interrupt context, and preempt_count() has not
3894 * been updated yet. In this case, use the TRANSITION bit.
3895 */
3896 bit = RB_CTX_TRANSITION;
3897 if (val & (1 << (bit + cpu_buffer->nest))) {
3898 do_ring_buffer_record_recursion();
3899 return true;
3900 }
3901 }
3902
3903 val |= (1 << (bit + cpu_buffer->nest));
3904 cpu_buffer->current_context = val;
3905
3906 return false;
3907}
3908
3909static __always_inline void
3910trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
3911{
3912 cpu_buffer->current_context &=
3913 cpu_buffer->current_context - (1 << cpu_buffer->nest);
3914}
3915
3916/* The recursive locking above uses 5 bits */
3917#define NESTED_BITS 5
3918
3919/**
3920 * ring_buffer_nest_start - Allow to trace while nested
3921 * @buffer: The ring buffer to modify
3922 *
3923 * The ring buffer has a safety mechanism to prevent recursion.
3924 * But there may be a case where a trace needs to be done while
3925 * tracing something else. In this case, calling this function
3926 * will allow this function to nest within a currently active
3927 * ring_buffer_lock_reserve().
3928 *
3929 * Call this function before calling another ring_buffer_lock_reserve() and
3930 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
3931 */
3932void ring_buffer_nest_start(struct trace_buffer *buffer)
3933{
3934 struct ring_buffer_per_cpu *cpu_buffer;
3935 int cpu;
3936
3937 /* Enabled by ring_buffer_nest_end() */
3938 preempt_disable_notrace();
3939 cpu = raw_smp_processor_id();
3940 cpu_buffer = buffer->buffers[cpu];
3941 /* This is the shift value for the above recursive locking */
3942 cpu_buffer->nest += NESTED_BITS;
3943}
3944
3945/**
3946 * ring_buffer_nest_end - Allow to trace while nested
3947 * @buffer: The ring buffer to modify
3948 *
3949 * Must be called after ring_buffer_nest_start() and after the
3950 * ring_buffer_unlock_commit().
3951 */
3952void ring_buffer_nest_end(struct trace_buffer *buffer)
3953{
3954 struct ring_buffer_per_cpu *cpu_buffer;
3955 int cpu;
3956
3957 /* disabled by ring_buffer_nest_start() */
3958 cpu = raw_smp_processor_id();
3959 cpu_buffer = buffer->buffers[cpu];
3960 /* This is the shift value for the above recursive locking */
3961 cpu_buffer->nest -= NESTED_BITS;
3962 preempt_enable_notrace();
3963}
3964
3965/**
3966 * ring_buffer_unlock_commit - commit a reserved
3967 * @buffer: The buffer to commit to
3968 *
3969 * This commits the data to the ring buffer, and releases any locks held.
3970 *
3971 * Must be paired with ring_buffer_lock_reserve.
3972 */
3973int ring_buffer_unlock_commit(struct trace_buffer *buffer)
3974{
3975 struct ring_buffer_per_cpu *cpu_buffer;
3976 int cpu = raw_smp_processor_id();
3977
3978 cpu_buffer = buffer->buffers[cpu];
3979
3980 rb_commit(cpu_buffer);
3981
3982 rb_wakeups(buffer, cpu_buffer);
3983
3984 trace_recursive_unlock(cpu_buffer);
3985
3986 preempt_enable_notrace();
3987
3988 return 0;
3989}
3990EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
3991
3992/* Special value to validate all deltas on a page. */
3993#define CHECK_FULL_PAGE 1L
3994
3995#ifdef CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS
3996
3997static const char *show_irq_str(int bits)
3998{
3999 const char *type[] = {
4000 ".", // 0
4001 "s", // 1
4002 "h", // 2
4003 "Hs", // 3
4004 "n", // 4
4005 "Ns", // 5
4006 "Nh", // 6
4007 "NHs", // 7
4008 };
4009
4010 return type[bits];
4011}
4012
4013/* Assume this is an trace event */
4014static const char *show_flags(struct ring_buffer_event *event)
4015{
4016 struct trace_entry *entry;
4017 int bits = 0;
4018
4019 if (rb_event_data_length(event) - RB_EVNT_HDR_SIZE < sizeof(*entry))
4020 return "X";
4021
4022 entry = ring_buffer_event_data(event);
4023
4024 if (entry->flags & TRACE_FLAG_SOFTIRQ)
4025 bits |= 1;
4026
4027 if (entry->flags & TRACE_FLAG_HARDIRQ)
4028 bits |= 2;
4029
4030 if (entry->flags & TRACE_FLAG_NMI)
4031 bits |= 4;
4032
4033 return show_irq_str(bits);
4034}
4035
4036static const char *show_irq(struct ring_buffer_event *event)
4037{
4038 struct trace_entry *entry;
4039
4040 if (rb_event_data_length(event) - RB_EVNT_HDR_SIZE < sizeof(*entry))
4041 return "";
4042
4043 entry = ring_buffer_event_data(event);
4044 if (entry->flags & TRACE_FLAG_IRQS_OFF)
4045 return "d";
4046 return "";
4047}
4048
4049static const char *show_interrupt_level(void)
4050{
4051 unsigned long pc = preempt_count();
4052 unsigned char level = 0;
4053
4054 if (pc & SOFTIRQ_OFFSET)
4055 level |= 1;
4056
4057 if (pc & HARDIRQ_MASK)
4058 level |= 2;
4059
4060 if (pc & NMI_MASK)
4061 level |= 4;
4062
4063 return show_irq_str(level);
4064}
4065
4066static void dump_buffer_page(struct buffer_data_page *bpage,
4067 struct rb_event_info *info,
4068 unsigned long tail)
4069{
4070 struct ring_buffer_event *event;
4071 u64 ts, delta;
4072 int e;
4073
4074 ts = bpage->time_stamp;
4075 pr_warn(" [%lld] PAGE TIME STAMP\n", ts);
4076
4077 for (e = 0; e < tail; e += rb_event_length(event)) {
4078
4079 event = (struct ring_buffer_event *)(bpage->data + e);
4080
4081 switch (event->type_len) {
4082
4083 case RINGBUF_TYPE_TIME_EXTEND:
4084 delta = rb_event_time_stamp(event);
4085 ts += delta;
4086 pr_warn(" 0x%x: [%lld] delta:%lld TIME EXTEND\n",
4087 e, ts, delta);
4088 break;
4089
4090 case RINGBUF_TYPE_TIME_STAMP:
4091 delta = rb_event_time_stamp(event);
4092 ts = rb_fix_abs_ts(delta, ts);
4093 pr_warn(" 0x%x: [%lld] absolute:%lld TIME STAMP\n",
4094 e, ts, delta);
4095 break;
4096
4097 case RINGBUF_TYPE_PADDING:
4098 ts += event->time_delta;
4099 pr_warn(" 0x%x: [%lld] delta:%d PADDING\n",
4100 e, ts, event->time_delta);
4101 break;
4102
4103 case RINGBUF_TYPE_DATA:
4104 ts += event->time_delta;
4105 pr_warn(" 0x%x: [%lld] delta:%d %s%s\n",
4106 e, ts, event->time_delta,
4107 show_flags(event), show_irq(event));
4108 break;
4109
4110 default:
4111 break;
4112 }
4113 }
4114 pr_warn("expected end:0x%lx last event actually ended at:0x%x\n", tail, e);
4115}
4116
4117static DEFINE_PER_CPU(atomic_t, checking);
4118static atomic_t ts_dump;
4119
4120#define buffer_warn_return(fmt, ...) \
4121 do { \
4122 /* If another report is happening, ignore this one */ \
4123 if (atomic_inc_return(&ts_dump) != 1) { \
4124 atomic_dec(&ts_dump); \
4125 goto out; \
4126 } \
4127 atomic_inc(&cpu_buffer->record_disabled); \
4128 pr_warn(fmt, ##__VA_ARGS__); \
4129 dump_buffer_page(bpage, info, tail); \
4130 atomic_dec(&ts_dump); \
4131 /* There's some cases in boot up that this can happen */ \
4132 if (WARN_ON_ONCE(system_state != SYSTEM_BOOTING)) \
4133 /* Do not re-enable checking */ \
4134 return; \
4135 } while (0)
4136
4137/*
4138 * Check if the current event time stamp matches the deltas on
4139 * the buffer page.
4140 */
4141static void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
4142 struct rb_event_info *info,
4143 unsigned long tail)
4144{
4145 struct buffer_data_page *bpage;
4146 u64 ts, delta;
4147 bool full = false;
4148 int ret;
4149
4150 bpage = info->tail_page->page;
4151
4152 if (tail == CHECK_FULL_PAGE) {
4153 full = true;
4154 tail = local_read(&bpage->commit);
4155 } else if (info->add_timestamp &
4156 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE)) {
4157 /* Ignore events with absolute time stamps */
4158 return;
4159 }
4160
4161 /*
4162 * Do not check the first event (skip possible extends too).
4163 * Also do not check if previous events have not been committed.
4164 */
4165 if (tail <= 8 || tail > local_read(&bpage->commit))
4166 return;
4167
4168 /*
4169 * If this interrupted another event,
4170 */
4171 if (atomic_inc_return(this_cpu_ptr(&checking)) != 1)
4172 goto out;
4173
4174 ret = rb_read_data_buffer(bpage, tail, cpu_buffer->cpu, &ts, &delta);
4175 if (ret < 0) {
4176 if (delta < ts) {
4177 buffer_warn_return("[CPU: %d]ABSOLUTE TIME WENT BACKWARDS: last ts: %lld absolute ts: %lld\n",
4178 cpu_buffer->cpu, ts, delta);
4179 goto out;
4180 }
4181 }
4182 if ((full && ts > info->ts) ||
4183 (!full && ts + info->delta != info->ts)) {
4184 buffer_warn_return("[CPU: %d]TIME DOES NOT MATCH expected:%lld actual:%lld delta:%lld before:%lld after:%lld%s context:%s\n",
4185 cpu_buffer->cpu,
4186 ts + info->delta, info->ts, info->delta,
4187 info->before, info->after,
4188 full ? " (full)" : "", show_interrupt_level());
4189 }
4190out:
4191 atomic_dec(this_cpu_ptr(&checking));
4192}
4193#else
4194static inline void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
4195 struct rb_event_info *info,
4196 unsigned long tail)
4197{
4198}
4199#endif /* CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS */
4200
4201static struct ring_buffer_event *
4202__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
4203 struct rb_event_info *info)
4204{
4205 struct ring_buffer_event *event;
4206 struct buffer_page *tail_page;
4207 unsigned long tail, write, w;
4208
4209 /* Don't let the compiler play games with cpu_buffer->tail_page */
4210 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
4211
4212 /*A*/ w = local_read(&tail_page->write) & RB_WRITE_MASK;
4213 barrier();
4214 rb_time_read(&cpu_buffer->before_stamp, &info->before);
4215 rb_time_read(&cpu_buffer->write_stamp, &info->after);
4216 barrier();
4217 info->ts = rb_time_stamp(cpu_buffer->buffer);
4218
4219 if ((info->add_timestamp & RB_ADD_STAMP_ABSOLUTE)) {
4220 info->delta = info->ts;
4221 } else {
4222 /*
4223 * If interrupting an event time update, we may need an
4224 * absolute timestamp.
4225 * Don't bother if this is the start of a new page (w == 0).
4226 */
4227 if (!w) {
4228 /* Use the sub-buffer timestamp */
4229 info->delta = 0;
4230 } else if (unlikely(info->before != info->after)) {
4231 info->add_timestamp |= RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND;
4232 info->length += RB_LEN_TIME_EXTEND;
4233 } else {
4234 info->delta = info->ts - info->after;
4235 if (unlikely(test_time_stamp(info->delta))) {
4236 info->add_timestamp |= RB_ADD_STAMP_EXTEND;
4237 info->length += RB_LEN_TIME_EXTEND;
4238 }
4239 }
4240 }
4241
4242 /*B*/ rb_time_set(&cpu_buffer->before_stamp, info->ts);
4243
4244 /*C*/ write = local_add_return(info->length, &tail_page->write);
4245
4246 /* set write to only the index of the write */
4247 write &= RB_WRITE_MASK;
4248
4249 tail = write - info->length;
4250
4251 /* See if we shot pass the end of this buffer page */
4252 if (unlikely(write > cpu_buffer->buffer->subbuf_size)) {
4253 check_buffer(cpu_buffer, info, CHECK_FULL_PAGE);
4254 return rb_move_tail(cpu_buffer, tail, info);
4255 }
4256
4257 if (likely(tail == w)) {
4258 /* Nothing interrupted us between A and C */
4259 /*D*/ rb_time_set(&cpu_buffer->write_stamp, info->ts);
4260 /*
4261 * If something came in between C and D, the write stamp
4262 * may now not be in sync. But that's fine as the before_stamp
4263 * will be different and then next event will just be forced
4264 * to use an absolute timestamp.
4265 */
4266 if (likely(!(info->add_timestamp &
4267 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
4268 /* This did not interrupt any time update */
4269 info->delta = info->ts - info->after;
4270 else
4271 /* Just use full timestamp for interrupting event */
4272 info->delta = info->ts;
4273 check_buffer(cpu_buffer, info, tail);
4274 } else {
4275 u64 ts;
4276 /* SLOW PATH - Interrupted between A and C */
4277
4278 /* Save the old before_stamp */
4279 rb_time_read(&cpu_buffer->before_stamp, &info->before);
4280
4281 /*
4282 * Read a new timestamp and update the before_stamp to make
4283 * the next event after this one force using an absolute
4284 * timestamp. This is in case an interrupt were to come in
4285 * between E and F.
4286 */
4287 ts = rb_time_stamp(cpu_buffer->buffer);
4288 rb_time_set(&cpu_buffer->before_stamp, ts);
4289
4290 barrier();
4291 /*E*/ rb_time_read(&cpu_buffer->write_stamp, &info->after);
4292 barrier();
4293 /*F*/ if (write == (local_read(&tail_page->write) & RB_WRITE_MASK) &&
4294 info->after == info->before && info->after < ts) {
4295 /*
4296 * Nothing came after this event between C and F, it is
4297 * safe to use info->after for the delta as it
4298 * matched info->before and is still valid.
4299 */
4300 info->delta = ts - info->after;
4301 } else {
4302 /*
4303 * Interrupted between C and F:
4304 * Lost the previous events time stamp. Just set the
4305 * delta to zero, and this will be the same time as
4306 * the event this event interrupted. And the events that
4307 * came after this will still be correct (as they would
4308 * have built their delta on the previous event.
4309 */
4310 info->delta = 0;
4311 }
4312 info->ts = ts;
4313 info->add_timestamp &= ~RB_ADD_STAMP_FORCE;
4314 }
4315
4316 /*
4317 * If this is the first commit on the page, then it has the same
4318 * timestamp as the page itself.
4319 */
4320 if (unlikely(!tail && !(info->add_timestamp &
4321 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
4322 info->delta = 0;
4323
4324 /* We reserved something on the buffer */
4325
4326 event = __rb_page_index(tail_page, tail);
4327 rb_update_event(cpu_buffer, event, info);
4328
4329 local_inc(&tail_page->entries);
4330
4331 /*
4332 * If this is the first commit on the page, then update
4333 * its timestamp.
4334 */
4335 if (unlikely(!tail))
4336 tail_page->page->time_stamp = info->ts;
4337
4338 /* account for these added bytes */
4339 local_add(info->length, &cpu_buffer->entries_bytes);
4340
4341 return event;
4342}
4343
4344static __always_inline struct ring_buffer_event *
4345rb_reserve_next_event(struct trace_buffer *buffer,
4346 struct ring_buffer_per_cpu *cpu_buffer,
4347 unsigned long length)
4348{
4349 struct ring_buffer_event *event;
4350 struct rb_event_info info;
4351 int nr_loops = 0;
4352 int add_ts_default;
4353
4354 /* ring buffer does cmpxchg, make sure it is safe in NMI context */
4355 if (!IS_ENABLED(CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG) &&
4356 (unlikely(in_nmi()))) {
4357 return NULL;
4358 }
4359
4360 rb_start_commit(cpu_buffer);
4361 /* The commit page can not change after this */
4362
4363#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4364 /*
4365 * Due to the ability to swap a cpu buffer from a buffer
4366 * it is possible it was swapped before we committed.
4367 * (committing stops a swap). We check for it here and
4368 * if it happened, we have to fail the write.
4369 */
4370 barrier();
4371 if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
4372 local_dec(&cpu_buffer->committing);
4373 local_dec(&cpu_buffer->commits);
4374 return NULL;
4375 }
4376#endif
4377
4378 info.length = rb_calculate_event_length(length);
4379
4380 if (ring_buffer_time_stamp_abs(cpu_buffer->buffer)) {
4381 add_ts_default = RB_ADD_STAMP_ABSOLUTE;
4382 info.length += RB_LEN_TIME_EXTEND;
4383 if (info.length > cpu_buffer->buffer->max_data_size)
4384 goto out_fail;
4385 } else {
4386 add_ts_default = RB_ADD_STAMP_NONE;
4387 }
4388
4389 again:
4390 info.add_timestamp = add_ts_default;
4391 info.delta = 0;
4392
4393 /*
4394 * We allow for interrupts to reenter here and do a trace.
4395 * If one does, it will cause this original code to loop
4396 * back here. Even with heavy interrupts happening, this
4397 * should only happen a few times in a row. If this happens
4398 * 1000 times in a row, there must be either an interrupt
4399 * storm or we have something buggy.
4400 * Bail!
4401 */
4402 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
4403 goto out_fail;
4404
4405 event = __rb_reserve_next(cpu_buffer, &info);
4406
4407 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
4408 if (info.add_timestamp & (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND))
4409 info.length -= RB_LEN_TIME_EXTEND;
4410 goto again;
4411 }
4412
4413 if (likely(event))
4414 return event;
4415 out_fail:
4416 rb_end_commit(cpu_buffer);
4417 return NULL;
4418}
4419
4420/**
4421 * ring_buffer_lock_reserve - reserve a part of the buffer
4422 * @buffer: the ring buffer to reserve from
4423 * @length: the length of the data to reserve (excluding event header)
4424 *
4425 * Returns a reserved event on the ring buffer to copy directly to.
4426 * The user of this interface will need to get the body to write into
4427 * and can use the ring_buffer_event_data() interface.
4428 *
4429 * The length is the length of the data needed, not the event length
4430 * which also includes the event header.
4431 *
4432 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
4433 * If NULL is returned, then nothing has been allocated or locked.
4434 */
4435struct ring_buffer_event *
4436ring_buffer_lock_reserve(struct trace_buffer *buffer, unsigned long length)
4437{
4438 struct ring_buffer_per_cpu *cpu_buffer;
4439 struct ring_buffer_event *event;
4440 int cpu;
4441
4442 /* If we are tracing schedule, we don't want to recurse */
4443 preempt_disable_notrace();
4444
4445 if (unlikely(atomic_read(&buffer->record_disabled)))
4446 goto out;
4447
4448 cpu = raw_smp_processor_id();
4449
4450 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
4451 goto out;
4452
4453 cpu_buffer = buffer->buffers[cpu];
4454
4455 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
4456 goto out;
4457
4458 if (unlikely(length > buffer->max_data_size))
4459 goto out;
4460
4461 if (unlikely(trace_recursive_lock(cpu_buffer)))
4462 goto out;
4463
4464 event = rb_reserve_next_event(buffer, cpu_buffer, length);
4465 if (!event)
4466 goto out_unlock;
4467
4468 return event;
4469
4470 out_unlock:
4471 trace_recursive_unlock(cpu_buffer);
4472 out:
4473 preempt_enable_notrace();
4474 return NULL;
4475}
4476EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
4477
4478/*
4479 * Decrement the entries to the page that an event is on.
4480 * The event does not even need to exist, only the pointer
4481 * to the page it is on. This may only be called before the commit
4482 * takes place.
4483 */
4484static inline void
4485rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
4486 struct ring_buffer_event *event)
4487{
4488 unsigned long addr = (unsigned long)event;
4489 struct buffer_page *bpage = cpu_buffer->commit_page;
4490 struct buffer_page *start;
4491
4492 addr &= ~((PAGE_SIZE << cpu_buffer->buffer->subbuf_order) - 1);
4493
4494 /* Do the likely case first */
4495 if (likely(bpage->page == (void *)addr)) {
4496 local_dec(&bpage->entries);
4497 return;
4498 }
4499
4500 /*
4501 * Because the commit page may be on the reader page we
4502 * start with the next page and check the end loop there.
4503 */
4504 rb_inc_page(&bpage);
4505 start = bpage;
4506 do {
4507 if (bpage->page == (void *)addr) {
4508 local_dec(&bpage->entries);
4509 return;
4510 }
4511 rb_inc_page(&bpage);
4512 } while (bpage != start);
4513
4514 /* commit not part of this buffer?? */
4515 RB_WARN_ON(cpu_buffer, 1);
4516}
4517
4518/**
4519 * ring_buffer_discard_commit - discard an event that has not been committed
4520 * @buffer: the ring buffer
4521 * @event: non committed event to discard
4522 *
4523 * Sometimes an event that is in the ring buffer needs to be ignored.
4524 * This function lets the user discard an event in the ring buffer
4525 * and then that event will not be read later.
4526 *
4527 * This function only works if it is called before the item has been
4528 * committed. It will try to free the event from the ring buffer
4529 * if another event has not been added behind it.
4530 *
4531 * If another event has been added behind it, it will set the event
4532 * up as discarded, and perform the commit.
4533 *
4534 * If this function is called, do not call ring_buffer_unlock_commit on
4535 * the event.
4536 */
4537void ring_buffer_discard_commit(struct trace_buffer *buffer,
4538 struct ring_buffer_event *event)
4539{
4540 struct ring_buffer_per_cpu *cpu_buffer;
4541 int cpu;
4542
4543 /* The event is discarded regardless */
4544 rb_event_discard(event);
4545
4546 cpu = smp_processor_id();
4547 cpu_buffer = buffer->buffers[cpu];
4548
4549 /*
4550 * This must only be called if the event has not been
4551 * committed yet. Thus we can assume that preemption
4552 * is still disabled.
4553 */
4554 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
4555
4556 rb_decrement_entry(cpu_buffer, event);
4557 if (rb_try_to_discard(cpu_buffer, event))
4558 goto out;
4559
4560 out:
4561 rb_end_commit(cpu_buffer);
4562
4563 trace_recursive_unlock(cpu_buffer);
4564
4565 preempt_enable_notrace();
4566
4567}
4568EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
4569
4570/**
4571 * ring_buffer_write - write data to the buffer without reserving
4572 * @buffer: The ring buffer to write to.
4573 * @length: The length of the data being written (excluding the event header)
4574 * @data: The data to write to the buffer.
4575 *
4576 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
4577 * one function. If you already have the data to write to the buffer, it
4578 * may be easier to simply call this function.
4579 *
4580 * Note, like ring_buffer_lock_reserve, the length is the length of the data
4581 * and not the length of the event which would hold the header.
4582 */
4583int ring_buffer_write(struct trace_buffer *buffer,
4584 unsigned long length,
4585 void *data)
4586{
4587 struct ring_buffer_per_cpu *cpu_buffer;
4588 struct ring_buffer_event *event;
4589 void *body;
4590 int ret = -EBUSY;
4591 int cpu;
4592
4593 preempt_disable_notrace();
4594
4595 if (atomic_read(&buffer->record_disabled))
4596 goto out;
4597
4598 cpu = raw_smp_processor_id();
4599
4600 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4601 goto out;
4602
4603 cpu_buffer = buffer->buffers[cpu];
4604
4605 if (atomic_read(&cpu_buffer->record_disabled))
4606 goto out;
4607
4608 if (length > buffer->max_data_size)
4609 goto out;
4610
4611 if (unlikely(trace_recursive_lock(cpu_buffer)))
4612 goto out;
4613
4614 event = rb_reserve_next_event(buffer, cpu_buffer, length);
4615 if (!event)
4616 goto out_unlock;
4617
4618 body = rb_event_data(event);
4619
4620 memcpy(body, data, length);
4621
4622 rb_commit(cpu_buffer);
4623
4624 rb_wakeups(buffer, cpu_buffer);
4625
4626 ret = 0;
4627
4628 out_unlock:
4629 trace_recursive_unlock(cpu_buffer);
4630
4631 out:
4632 preempt_enable_notrace();
4633
4634 return ret;
4635}
4636EXPORT_SYMBOL_GPL(ring_buffer_write);
4637
4638static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
4639{
4640 struct buffer_page *reader = cpu_buffer->reader_page;
4641 struct buffer_page *head = rb_set_head_page(cpu_buffer);
4642 struct buffer_page *commit = cpu_buffer->commit_page;
4643
4644 /* In case of error, head will be NULL */
4645 if (unlikely(!head))
4646 return true;
4647
4648 /* Reader should exhaust content in reader page */
4649 if (reader->read != rb_page_size(reader))
4650 return false;
4651
4652 /*
4653 * If writers are committing on the reader page, knowing all
4654 * committed content has been read, the ring buffer is empty.
4655 */
4656 if (commit == reader)
4657 return true;
4658
4659 /*
4660 * If writers are committing on a page other than reader page
4661 * and head page, there should always be content to read.
4662 */
4663 if (commit != head)
4664 return false;
4665
4666 /*
4667 * Writers are committing on the head page, we just need
4668 * to care about there're committed data, and the reader will
4669 * swap reader page with head page when it is to read data.
4670 */
4671 return rb_page_commit(commit) == 0;
4672}
4673
4674/**
4675 * ring_buffer_record_disable - stop all writes into the buffer
4676 * @buffer: The ring buffer to stop writes to.
4677 *
4678 * This prevents all writes to the buffer. Any attempt to write
4679 * to the buffer after this will fail and return NULL.
4680 *
4681 * The caller should call synchronize_rcu() after this.
4682 */
4683void ring_buffer_record_disable(struct trace_buffer *buffer)
4684{
4685 atomic_inc(&buffer->record_disabled);
4686}
4687EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
4688
4689/**
4690 * ring_buffer_record_enable - enable writes to the buffer
4691 * @buffer: The ring buffer to enable writes
4692 *
4693 * Note, multiple disables will need the same number of enables
4694 * to truly enable the writing (much like preempt_disable).
4695 */
4696void ring_buffer_record_enable(struct trace_buffer *buffer)
4697{
4698 atomic_dec(&buffer->record_disabled);
4699}
4700EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
4701
4702/**
4703 * ring_buffer_record_off - stop all writes into the buffer
4704 * @buffer: The ring buffer to stop writes to.
4705 *
4706 * This prevents all writes to the buffer. Any attempt to write
4707 * to the buffer after this will fail and return NULL.
4708 *
4709 * This is different than ring_buffer_record_disable() as
4710 * it works like an on/off switch, where as the disable() version
4711 * must be paired with a enable().
4712 */
4713void ring_buffer_record_off(struct trace_buffer *buffer)
4714{
4715 unsigned int rd;
4716 unsigned int new_rd;
4717
4718 rd = atomic_read(&buffer->record_disabled);
4719 do {
4720 new_rd = rd | RB_BUFFER_OFF;
4721 } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
4722}
4723EXPORT_SYMBOL_GPL(ring_buffer_record_off);
4724
4725/**
4726 * ring_buffer_record_on - restart writes into the buffer
4727 * @buffer: The ring buffer to start writes to.
4728 *
4729 * This enables all writes to the buffer that was disabled by
4730 * ring_buffer_record_off().
4731 *
4732 * This is different than ring_buffer_record_enable() as
4733 * it works like an on/off switch, where as the enable() version
4734 * must be paired with a disable().
4735 */
4736void ring_buffer_record_on(struct trace_buffer *buffer)
4737{
4738 unsigned int rd;
4739 unsigned int new_rd;
4740
4741 rd = atomic_read(&buffer->record_disabled);
4742 do {
4743 new_rd = rd & ~RB_BUFFER_OFF;
4744 } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
4745}
4746EXPORT_SYMBOL_GPL(ring_buffer_record_on);
4747
4748/**
4749 * ring_buffer_record_is_on - return true if the ring buffer can write
4750 * @buffer: The ring buffer to see if write is enabled
4751 *
4752 * Returns true if the ring buffer is in a state that it accepts writes.
4753 */
4754bool ring_buffer_record_is_on(struct trace_buffer *buffer)
4755{
4756 return !atomic_read(&buffer->record_disabled);
4757}
4758
4759/**
4760 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
4761 * @buffer: The ring buffer to see if write is set enabled
4762 *
4763 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
4764 * Note that this does NOT mean it is in a writable state.
4765 *
4766 * It may return true when the ring buffer has been disabled by
4767 * ring_buffer_record_disable(), as that is a temporary disabling of
4768 * the ring buffer.
4769 */
4770bool ring_buffer_record_is_set_on(struct trace_buffer *buffer)
4771{
4772 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
4773}
4774
4775/**
4776 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
4777 * @buffer: The ring buffer to stop writes to.
4778 * @cpu: The CPU buffer to stop
4779 *
4780 * This prevents all writes to the buffer. Any attempt to write
4781 * to the buffer after this will fail and return NULL.
4782 *
4783 * The caller should call synchronize_rcu() after this.
4784 */
4785void ring_buffer_record_disable_cpu(struct trace_buffer *buffer, int cpu)
4786{
4787 struct ring_buffer_per_cpu *cpu_buffer;
4788
4789 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4790 return;
4791
4792 cpu_buffer = buffer->buffers[cpu];
4793 atomic_inc(&cpu_buffer->record_disabled);
4794}
4795EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
4796
4797/**
4798 * ring_buffer_record_enable_cpu - enable writes to the buffer
4799 * @buffer: The ring buffer to enable writes
4800 * @cpu: The CPU to enable.
4801 *
4802 * Note, multiple disables will need the same number of enables
4803 * to truly enable the writing (much like preempt_disable).
4804 */
4805void ring_buffer_record_enable_cpu(struct trace_buffer *buffer, int cpu)
4806{
4807 struct ring_buffer_per_cpu *cpu_buffer;
4808
4809 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4810 return;
4811
4812 cpu_buffer = buffer->buffers[cpu];
4813 atomic_dec(&cpu_buffer->record_disabled);
4814}
4815EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
4816
4817/*
4818 * The total entries in the ring buffer is the running counter
4819 * of entries entered into the ring buffer, minus the sum of
4820 * the entries read from the ring buffer and the number of
4821 * entries that were overwritten.
4822 */
4823static inline unsigned long
4824rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
4825{
4826 return local_read(&cpu_buffer->entries) -
4827 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
4828}
4829
4830/**
4831 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
4832 * @buffer: The ring buffer
4833 * @cpu: The per CPU buffer to read from.
4834 */
4835u64 ring_buffer_oldest_event_ts(struct trace_buffer *buffer, int cpu)
4836{
4837 unsigned long flags;
4838 struct ring_buffer_per_cpu *cpu_buffer;
4839 struct buffer_page *bpage;
4840 u64 ret = 0;
4841
4842 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4843 return 0;
4844
4845 cpu_buffer = buffer->buffers[cpu];
4846 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4847 /*
4848 * if the tail is on reader_page, oldest time stamp is on the reader
4849 * page
4850 */
4851 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
4852 bpage = cpu_buffer->reader_page;
4853 else
4854 bpage = rb_set_head_page(cpu_buffer);
4855 if (bpage)
4856 ret = bpage->page->time_stamp;
4857 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4858
4859 return ret;
4860}
4861EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
4862
4863/**
4864 * ring_buffer_bytes_cpu - get the number of bytes unconsumed in a cpu buffer
4865 * @buffer: The ring buffer
4866 * @cpu: The per CPU buffer to read from.
4867 */
4868unsigned long ring_buffer_bytes_cpu(struct trace_buffer *buffer, int cpu)
4869{
4870 struct ring_buffer_per_cpu *cpu_buffer;
4871 unsigned long ret;
4872
4873 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4874 return 0;
4875
4876 cpu_buffer = buffer->buffers[cpu];
4877 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
4878
4879 return ret;
4880}
4881EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
4882
4883/**
4884 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
4885 * @buffer: The ring buffer
4886 * @cpu: The per CPU buffer to get the entries from.
4887 */
4888unsigned long ring_buffer_entries_cpu(struct trace_buffer *buffer, int cpu)
4889{
4890 struct ring_buffer_per_cpu *cpu_buffer;
4891
4892 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4893 return 0;
4894
4895 cpu_buffer = buffer->buffers[cpu];
4896
4897 return rb_num_of_entries(cpu_buffer);
4898}
4899EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
4900
4901/**
4902 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
4903 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
4904 * @buffer: The ring buffer
4905 * @cpu: The per CPU buffer to get the number of overruns from
4906 */
4907unsigned long ring_buffer_overrun_cpu(struct trace_buffer *buffer, int cpu)
4908{
4909 struct ring_buffer_per_cpu *cpu_buffer;
4910 unsigned long ret;
4911
4912 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4913 return 0;
4914
4915 cpu_buffer = buffer->buffers[cpu];
4916 ret = local_read(&cpu_buffer->overrun);
4917
4918 return ret;
4919}
4920EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
4921
4922/**
4923 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
4924 * commits failing due to the buffer wrapping around while there are uncommitted
4925 * events, such as during an interrupt storm.
4926 * @buffer: The ring buffer
4927 * @cpu: The per CPU buffer to get the number of overruns from
4928 */
4929unsigned long
4930ring_buffer_commit_overrun_cpu(struct trace_buffer *buffer, int cpu)
4931{
4932 struct ring_buffer_per_cpu *cpu_buffer;
4933 unsigned long ret;
4934
4935 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4936 return 0;
4937
4938 cpu_buffer = buffer->buffers[cpu];
4939 ret = local_read(&cpu_buffer->commit_overrun);
4940
4941 return ret;
4942}
4943EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
4944
4945/**
4946 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
4947 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
4948 * @buffer: The ring buffer
4949 * @cpu: The per CPU buffer to get the number of overruns from
4950 */
4951unsigned long
4952ring_buffer_dropped_events_cpu(struct trace_buffer *buffer, int cpu)
4953{
4954 struct ring_buffer_per_cpu *cpu_buffer;
4955 unsigned long ret;
4956
4957 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4958 return 0;
4959
4960 cpu_buffer = buffer->buffers[cpu];
4961 ret = local_read(&cpu_buffer->dropped_events);
4962
4963 return ret;
4964}
4965EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
4966
4967/**
4968 * ring_buffer_read_events_cpu - get the number of events successfully read
4969 * @buffer: The ring buffer
4970 * @cpu: The per CPU buffer to get the number of events read
4971 */
4972unsigned long
4973ring_buffer_read_events_cpu(struct trace_buffer *buffer, int cpu)
4974{
4975 struct ring_buffer_per_cpu *cpu_buffer;
4976
4977 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4978 return 0;
4979
4980 cpu_buffer = buffer->buffers[cpu];
4981 return cpu_buffer->read;
4982}
4983EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
4984
4985/**
4986 * ring_buffer_entries - get the number of entries in a buffer
4987 * @buffer: The ring buffer
4988 *
4989 * Returns the total number of entries in the ring buffer
4990 * (all CPU entries)
4991 */
4992unsigned long ring_buffer_entries(struct trace_buffer *buffer)
4993{
4994 struct ring_buffer_per_cpu *cpu_buffer;
4995 unsigned long entries = 0;
4996 int cpu;
4997
4998 /* if you care about this being correct, lock the buffer */
4999 for_each_buffer_cpu(buffer, cpu) {
5000 cpu_buffer = buffer->buffers[cpu];
5001 entries += rb_num_of_entries(cpu_buffer);
5002 }
5003
5004 return entries;
5005}
5006EXPORT_SYMBOL_GPL(ring_buffer_entries);
5007
5008/**
5009 * ring_buffer_overruns - get the number of overruns in buffer
5010 * @buffer: The ring buffer
5011 *
5012 * Returns the total number of overruns in the ring buffer
5013 * (all CPU entries)
5014 */
5015unsigned long ring_buffer_overruns(struct trace_buffer *buffer)
5016{
5017 struct ring_buffer_per_cpu *cpu_buffer;
5018 unsigned long overruns = 0;
5019 int cpu;
5020
5021 /* if you care about this being correct, lock the buffer */
5022 for_each_buffer_cpu(buffer, cpu) {
5023 cpu_buffer = buffer->buffers[cpu];
5024 overruns += local_read(&cpu_buffer->overrun);
5025 }
5026
5027 return overruns;
5028}
5029EXPORT_SYMBOL_GPL(ring_buffer_overruns);
5030
5031static void rb_iter_reset(struct ring_buffer_iter *iter)
5032{
5033 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5034
5035 /* Iterator usage is expected to have record disabled */
5036 iter->head_page = cpu_buffer->reader_page;
5037 iter->head = cpu_buffer->reader_page->read;
5038 iter->next_event = iter->head;
5039
5040 iter->cache_reader_page = iter->head_page;
5041 iter->cache_read = cpu_buffer->read;
5042 iter->cache_pages_removed = cpu_buffer->pages_removed;
5043
5044 if (iter->head) {
5045 iter->read_stamp = cpu_buffer->read_stamp;
5046 iter->page_stamp = cpu_buffer->reader_page->page->time_stamp;
5047 } else {
5048 iter->read_stamp = iter->head_page->page->time_stamp;
5049 iter->page_stamp = iter->read_stamp;
5050 }
5051}
5052
5053/**
5054 * ring_buffer_iter_reset - reset an iterator
5055 * @iter: The iterator to reset
5056 *
5057 * Resets the iterator, so that it will start from the beginning
5058 * again.
5059 */
5060void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
5061{
5062 struct ring_buffer_per_cpu *cpu_buffer;
5063 unsigned long flags;
5064
5065 if (!iter)
5066 return;
5067
5068 cpu_buffer = iter->cpu_buffer;
5069
5070 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5071 rb_iter_reset(iter);
5072 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5073}
5074EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
5075
5076/**
5077 * ring_buffer_iter_empty - check if an iterator has no more to read
5078 * @iter: The iterator to check
5079 */
5080int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
5081{
5082 struct ring_buffer_per_cpu *cpu_buffer;
5083 struct buffer_page *reader;
5084 struct buffer_page *head_page;
5085 struct buffer_page *commit_page;
5086 struct buffer_page *curr_commit_page;
5087 unsigned commit;
5088 u64 curr_commit_ts;
5089 u64 commit_ts;
5090
5091 cpu_buffer = iter->cpu_buffer;
5092 reader = cpu_buffer->reader_page;
5093 head_page = cpu_buffer->head_page;
5094 commit_page = READ_ONCE(cpu_buffer->commit_page);
5095 commit_ts = commit_page->page->time_stamp;
5096
5097 /*
5098 * When the writer goes across pages, it issues a cmpxchg which
5099 * is a mb(), which will synchronize with the rmb here.
5100 * (see rb_tail_page_update())
5101 */
5102 smp_rmb();
5103 commit = rb_page_commit(commit_page);
5104 /* We want to make sure that the commit page doesn't change */
5105 smp_rmb();
5106
5107 /* Make sure commit page didn't change */
5108 curr_commit_page = READ_ONCE(cpu_buffer->commit_page);
5109 curr_commit_ts = READ_ONCE(curr_commit_page->page->time_stamp);
5110
5111 /* If the commit page changed, then there's more data */
5112 if (curr_commit_page != commit_page ||
5113 curr_commit_ts != commit_ts)
5114 return 0;
5115
5116 /* Still racy, as it may return a false positive, but that's OK */
5117 return ((iter->head_page == commit_page && iter->head >= commit) ||
5118 (iter->head_page == reader && commit_page == head_page &&
5119 head_page->read == commit &&
5120 iter->head == rb_page_size(cpu_buffer->reader_page)));
5121}
5122EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
5123
5124static void
5125rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
5126 struct ring_buffer_event *event)
5127{
5128 u64 delta;
5129
5130 switch (event->type_len) {
5131 case RINGBUF_TYPE_PADDING:
5132 return;
5133
5134 case RINGBUF_TYPE_TIME_EXTEND:
5135 delta = rb_event_time_stamp(event);
5136 cpu_buffer->read_stamp += delta;
5137 return;
5138
5139 case RINGBUF_TYPE_TIME_STAMP:
5140 delta = rb_event_time_stamp(event);
5141 delta = rb_fix_abs_ts(delta, cpu_buffer->read_stamp);
5142 cpu_buffer->read_stamp = delta;
5143 return;
5144
5145 case RINGBUF_TYPE_DATA:
5146 cpu_buffer->read_stamp += event->time_delta;
5147 return;
5148
5149 default:
5150 RB_WARN_ON(cpu_buffer, 1);
5151 }
5152}
5153
5154static void
5155rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
5156 struct ring_buffer_event *event)
5157{
5158 u64 delta;
5159
5160 switch (event->type_len) {
5161 case RINGBUF_TYPE_PADDING:
5162 return;
5163
5164 case RINGBUF_TYPE_TIME_EXTEND:
5165 delta = rb_event_time_stamp(event);
5166 iter->read_stamp += delta;
5167 return;
5168
5169 case RINGBUF_TYPE_TIME_STAMP:
5170 delta = rb_event_time_stamp(event);
5171 delta = rb_fix_abs_ts(delta, iter->read_stamp);
5172 iter->read_stamp = delta;
5173 return;
5174
5175 case RINGBUF_TYPE_DATA:
5176 iter->read_stamp += event->time_delta;
5177 return;
5178
5179 default:
5180 RB_WARN_ON(iter->cpu_buffer, 1);
5181 }
5182}
5183
5184static struct buffer_page *
5185rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
5186{
5187 struct buffer_page *reader = NULL;
5188 unsigned long bsize = READ_ONCE(cpu_buffer->buffer->subbuf_size);
5189 unsigned long overwrite;
5190 unsigned long flags;
5191 int nr_loops = 0;
5192 bool ret;
5193
5194 local_irq_save(flags);
5195 arch_spin_lock(&cpu_buffer->lock);
5196
5197 again:
5198 /*
5199 * This should normally only loop twice. But because the
5200 * start of the reader inserts an empty page, it causes
5201 * a case where we will loop three times. There should be no
5202 * reason to loop four times (that I know of).
5203 */
5204 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
5205 reader = NULL;
5206 goto out;
5207 }
5208
5209 reader = cpu_buffer->reader_page;
5210
5211 /* If there's more to read, return this page */
5212 if (cpu_buffer->reader_page->read < rb_page_size(reader))
5213 goto out;
5214
5215 /* Never should we have an index greater than the size */
5216 if (RB_WARN_ON(cpu_buffer,
5217 cpu_buffer->reader_page->read > rb_page_size(reader)))
5218 goto out;
5219
5220 /* check if we caught up to the tail */
5221 reader = NULL;
5222 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
5223 goto out;
5224
5225 /* Don't bother swapping if the ring buffer is empty */
5226 if (rb_num_of_entries(cpu_buffer) == 0)
5227 goto out;
5228
5229 /*
5230 * Reset the reader page to size zero.
5231 */
5232 local_set(&cpu_buffer->reader_page->write, 0);
5233 local_set(&cpu_buffer->reader_page->entries, 0);
5234 local_set(&cpu_buffer->reader_page->page->commit, 0);
5235 cpu_buffer->reader_page->real_end = 0;
5236
5237 spin:
5238 /*
5239 * Splice the empty reader page into the list around the head.
5240 */
5241 reader = rb_set_head_page(cpu_buffer);
5242 if (!reader)
5243 goto out;
5244 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
5245 cpu_buffer->reader_page->list.prev = reader->list.prev;
5246
5247 /*
5248 * cpu_buffer->pages just needs to point to the buffer, it
5249 * has no specific buffer page to point to. Lets move it out
5250 * of our way so we don't accidentally swap it.
5251 */
5252 cpu_buffer->pages = reader->list.prev;
5253
5254 /* The reader page will be pointing to the new head */
5255 rb_set_list_to_head(&cpu_buffer->reader_page->list);
5256
5257 /*
5258 * We want to make sure we read the overruns after we set up our
5259 * pointers to the next object. The writer side does a
5260 * cmpxchg to cross pages which acts as the mb on the writer
5261 * side. Note, the reader will constantly fail the swap
5262 * while the writer is updating the pointers, so this
5263 * guarantees that the overwrite recorded here is the one we
5264 * want to compare with the last_overrun.
5265 */
5266 smp_mb();
5267 overwrite = local_read(&(cpu_buffer->overrun));
5268
5269 /*
5270 * Here's the tricky part.
5271 *
5272 * We need to move the pointer past the header page.
5273 * But we can only do that if a writer is not currently
5274 * moving it. The page before the header page has the
5275 * flag bit '1' set if it is pointing to the page we want.
5276 * but if the writer is in the process of moving it
5277 * than it will be '2' or already moved '0'.
5278 */
5279
5280 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
5281
5282 /*
5283 * If we did not convert it, then we must try again.
5284 */
5285 if (!ret)
5286 goto spin;
5287
5288 if (cpu_buffer->ring_meta)
5289 rb_update_meta_reader(cpu_buffer, reader);
5290
5291 /*
5292 * Yay! We succeeded in replacing the page.
5293 *
5294 * Now make the new head point back to the reader page.
5295 */
5296 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
5297 rb_inc_page(&cpu_buffer->head_page);
5298
5299 local_inc(&cpu_buffer->pages_read);
5300
5301 /* Finally update the reader page to the new head */
5302 cpu_buffer->reader_page = reader;
5303 cpu_buffer->reader_page->read = 0;
5304
5305 if (overwrite != cpu_buffer->last_overrun) {
5306 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
5307 cpu_buffer->last_overrun = overwrite;
5308 }
5309
5310 goto again;
5311
5312 out:
5313 /* Update the read_stamp on the first event */
5314 if (reader && reader->read == 0)
5315 cpu_buffer->read_stamp = reader->page->time_stamp;
5316
5317 arch_spin_unlock(&cpu_buffer->lock);
5318 local_irq_restore(flags);
5319
5320 /*
5321 * The writer has preempt disable, wait for it. But not forever
5322 * Although, 1 second is pretty much "forever"
5323 */
5324#define USECS_WAIT 1000000
5325 for (nr_loops = 0; nr_loops < USECS_WAIT; nr_loops++) {
5326 /* If the write is past the end of page, a writer is still updating it */
5327 if (likely(!reader || rb_page_write(reader) <= bsize))
5328 break;
5329
5330 udelay(1);
5331
5332 /* Get the latest version of the reader write value */
5333 smp_rmb();
5334 }
5335
5336 /* The writer is not moving forward? Something is wrong */
5337 if (RB_WARN_ON(cpu_buffer, nr_loops == USECS_WAIT))
5338 reader = NULL;
5339
5340 /*
5341 * Make sure we see any padding after the write update
5342 * (see rb_reset_tail()).
5343 *
5344 * In addition, a writer may be writing on the reader page
5345 * if the page has not been fully filled, so the read barrier
5346 * is also needed to make sure we see the content of what is
5347 * committed by the writer (see rb_set_commit_to_write()).
5348 */
5349 smp_rmb();
5350
5351
5352 return reader;
5353}
5354
5355static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
5356{
5357 struct ring_buffer_event *event;
5358 struct buffer_page *reader;
5359 unsigned length;
5360
5361 reader = rb_get_reader_page(cpu_buffer);
5362
5363 /* This function should not be called when buffer is empty */
5364 if (RB_WARN_ON(cpu_buffer, !reader))
5365 return;
5366
5367 event = rb_reader_event(cpu_buffer);
5368
5369 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
5370 cpu_buffer->read++;
5371
5372 rb_update_read_stamp(cpu_buffer, event);
5373
5374 length = rb_event_length(event);
5375 cpu_buffer->reader_page->read += length;
5376 cpu_buffer->read_bytes += length;
5377}
5378
5379static void rb_advance_iter(struct ring_buffer_iter *iter)
5380{
5381 struct ring_buffer_per_cpu *cpu_buffer;
5382
5383 cpu_buffer = iter->cpu_buffer;
5384
5385 /* If head == next_event then we need to jump to the next event */
5386 if (iter->head == iter->next_event) {
5387 /* If the event gets overwritten again, there's nothing to do */
5388 if (rb_iter_head_event(iter) == NULL)
5389 return;
5390 }
5391
5392 iter->head = iter->next_event;
5393
5394 /*
5395 * Check if we are at the end of the buffer.
5396 */
5397 if (iter->next_event >= rb_page_size(iter->head_page)) {
5398 /* discarded commits can make the page empty */
5399 if (iter->head_page == cpu_buffer->commit_page)
5400 return;
5401 rb_inc_iter(iter);
5402 return;
5403 }
5404
5405 rb_update_iter_read_stamp(iter, iter->event);
5406}
5407
5408static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
5409{
5410 return cpu_buffer->lost_events;
5411}
5412
5413static struct ring_buffer_event *
5414rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
5415 unsigned long *lost_events)
5416{
5417 struct ring_buffer_event *event;
5418 struct buffer_page *reader;
5419 int nr_loops = 0;
5420
5421 if (ts)
5422 *ts = 0;
5423 again:
5424 /*
5425 * We repeat when a time extend is encountered.
5426 * Since the time extend is always attached to a data event,
5427 * we should never loop more than once.
5428 * (We never hit the following condition more than twice).
5429 */
5430 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
5431 return NULL;
5432
5433 reader = rb_get_reader_page(cpu_buffer);
5434 if (!reader)
5435 return NULL;
5436
5437 event = rb_reader_event(cpu_buffer);
5438
5439 switch (event->type_len) {
5440 case RINGBUF_TYPE_PADDING:
5441 if (rb_null_event(event))
5442 RB_WARN_ON(cpu_buffer, 1);
5443 /*
5444 * Because the writer could be discarding every
5445 * event it creates (which would probably be bad)
5446 * if we were to go back to "again" then we may never
5447 * catch up, and will trigger the warn on, or lock
5448 * the box. Return the padding, and we will release
5449 * the current locks, and try again.
5450 */
5451 return event;
5452
5453 case RINGBUF_TYPE_TIME_EXTEND:
5454 /* Internal data, OK to advance */
5455 rb_advance_reader(cpu_buffer);
5456 goto again;
5457
5458 case RINGBUF_TYPE_TIME_STAMP:
5459 if (ts) {
5460 *ts = rb_event_time_stamp(event);
5461 *ts = rb_fix_abs_ts(*ts, reader->page->time_stamp);
5462 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
5463 cpu_buffer->cpu, ts);
5464 }
5465 /* Internal data, OK to advance */
5466 rb_advance_reader(cpu_buffer);
5467 goto again;
5468
5469 case RINGBUF_TYPE_DATA:
5470 if (ts && !(*ts)) {
5471 *ts = cpu_buffer->read_stamp + event->time_delta;
5472 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
5473 cpu_buffer->cpu, ts);
5474 }
5475 if (lost_events)
5476 *lost_events = rb_lost_events(cpu_buffer);
5477 return event;
5478
5479 default:
5480 RB_WARN_ON(cpu_buffer, 1);
5481 }
5482
5483 return NULL;
5484}
5485EXPORT_SYMBOL_GPL(ring_buffer_peek);
5486
5487static struct ring_buffer_event *
5488rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
5489{
5490 struct trace_buffer *buffer;
5491 struct ring_buffer_per_cpu *cpu_buffer;
5492 struct ring_buffer_event *event;
5493 int nr_loops = 0;
5494
5495 if (ts)
5496 *ts = 0;
5497
5498 cpu_buffer = iter->cpu_buffer;
5499 buffer = cpu_buffer->buffer;
5500
5501 /*
5502 * Check if someone performed a consuming read to the buffer
5503 * or removed some pages from the buffer. In these cases,
5504 * iterator was invalidated and we need to reset it.
5505 */
5506 if (unlikely(iter->cache_read != cpu_buffer->read ||
5507 iter->cache_reader_page != cpu_buffer->reader_page ||
5508 iter->cache_pages_removed != cpu_buffer->pages_removed))
5509 rb_iter_reset(iter);
5510
5511 again:
5512 if (ring_buffer_iter_empty(iter))
5513 return NULL;
5514
5515 /*
5516 * As the writer can mess with what the iterator is trying
5517 * to read, just give up if we fail to get an event after
5518 * three tries. The iterator is not as reliable when reading
5519 * the ring buffer with an active write as the consumer is.
5520 * Do not warn if the three failures is reached.
5521 */
5522 if (++nr_loops > 3)
5523 return NULL;
5524
5525 if (rb_per_cpu_empty(cpu_buffer))
5526 return NULL;
5527
5528 if (iter->head >= rb_page_size(iter->head_page)) {
5529 rb_inc_iter(iter);
5530 goto again;
5531 }
5532
5533 event = rb_iter_head_event(iter);
5534 if (!event)
5535 goto again;
5536
5537 switch (event->type_len) {
5538 case RINGBUF_TYPE_PADDING:
5539 if (rb_null_event(event)) {
5540 rb_inc_iter(iter);
5541 goto again;
5542 }
5543 rb_advance_iter(iter);
5544 return event;
5545
5546 case RINGBUF_TYPE_TIME_EXTEND:
5547 /* Internal data, OK to advance */
5548 rb_advance_iter(iter);
5549 goto again;
5550
5551 case RINGBUF_TYPE_TIME_STAMP:
5552 if (ts) {
5553 *ts = rb_event_time_stamp(event);
5554 *ts = rb_fix_abs_ts(*ts, iter->head_page->page->time_stamp);
5555 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
5556 cpu_buffer->cpu, ts);
5557 }
5558 /* Internal data, OK to advance */
5559 rb_advance_iter(iter);
5560 goto again;
5561
5562 case RINGBUF_TYPE_DATA:
5563 if (ts && !(*ts)) {
5564 *ts = iter->read_stamp + event->time_delta;
5565 ring_buffer_normalize_time_stamp(buffer,
5566 cpu_buffer->cpu, ts);
5567 }
5568 return event;
5569
5570 default:
5571 RB_WARN_ON(cpu_buffer, 1);
5572 }
5573
5574 return NULL;
5575}
5576EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
5577
5578static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
5579{
5580 if (likely(!in_nmi())) {
5581 raw_spin_lock(&cpu_buffer->reader_lock);
5582 return true;
5583 }
5584
5585 /*
5586 * If an NMI die dumps out the content of the ring buffer
5587 * trylock must be used to prevent a deadlock if the NMI
5588 * preempted a task that holds the ring buffer locks. If
5589 * we get the lock then all is fine, if not, then continue
5590 * to do the read, but this can corrupt the ring buffer,
5591 * so it must be permanently disabled from future writes.
5592 * Reading from NMI is a oneshot deal.
5593 */
5594 if (raw_spin_trylock(&cpu_buffer->reader_lock))
5595 return true;
5596
5597 /* Continue without locking, but disable the ring buffer */
5598 atomic_inc(&cpu_buffer->record_disabled);
5599 return false;
5600}
5601
5602static inline void
5603rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
5604{
5605 if (likely(locked))
5606 raw_spin_unlock(&cpu_buffer->reader_lock);
5607}
5608
5609/**
5610 * ring_buffer_peek - peek at the next event to be read
5611 * @buffer: The ring buffer to read
5612 * @cpu: The cpu to peak at
5613 * @ts: The timestamp counter of this event.
5614 * @lost_events: a variable to store if events were lost (may be NULL)
5615 *
5616 * This will return the event that will be read next, but does
5617 * not consume the data.
5618 */
5619struct ring_buffer_event *
5620ring_buffer_peek(struct trace_buffer *buffer, int cpu, u64 *ts,
5621 unsigned long *lost_events)
5622{
5623 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5624 struct ring_buffer_event *event;
5625 unsigned long flags;
5626 bool dolock;
5627
5628 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5629 return NULL;
5630
5631 again:
5632 local_irq_save(flags);
5633 dolock = rb_reader_lock(cpu_buffer);
5634 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
5635 if (event && event->type_len == RINGBUF_TYPE_PADDING)
5636 rb_advance_reader(cpu_buffer);
5637 rb_reader_unlock(cpu_buffer, dolock);
5638 local_irq_restore(flags);
5639
5640 if (event && event->type_len == RINGBUF_TYPE_PADDING)
5641 goto again;
5642
5643 return event;
5644}
5645
5646/** ring_buffer_iter_dropped - report if there are dropped events
5647 * @iter: The ring buffer iterator
5648 *
5649 * Returns true if there was dropped events since the last peek.
5650 */
5651bool ring_buffer_iter_dropped(struct ring_buffer_iter *iter)
5652{
5653 bool ret = iter->missed_events != 0;
5654
5655 iter->missed_events = 0;
5656 return ret;
5657}
5658EXPORT_SYMBOL_GPL(ring_buffer_iter_dropped);
5659
5660/**
5661 * ring_buffer_iter_peek - peek at the next event to be read
5662 * @iter: The ring buffer iterator
5663 * @ts: The timestamp counter of this event.
5664 *
5665 * This will return the event that will be read next, but does
5666 * not increment the iterator.
5667 */
5668struct ring_buffer_event *
5669ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
5670{
5671 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5672 struct ring_buffer_event *event;
5673 unsigned long flags;
5674
5675 again:
5676 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5677 event = rb_iter_peek(iter, ts);
5678 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5679
5680 if (event && event->type_len == RINGBUF_TYPE_PADDING)
5681 goto again;
5682
5683 return event;
5684}
5685
5686/**
5687 * ring_buffer_consume - return an event and consume it
5688 * @buffer: The ring buffer to get the next event from
5689 * @cpu: the cpu to read the buffer from
5690 * @ts: a variable to store the timestamp (may be NULL)
5691 * @lost_events: a variable to store if events were lost (may be NULL)
5692 *
5693 * Returns the next event in the ring buffer, and that event is consumed.
5694 * Meaning, that sequential reads will keep returning a different event,
5695 * and eventually empty the ring buffer if the producer is slower.
5696 */
5697struct ring_buffer_event *
5698ring_buffer_consume(struct trace_buffer *buffer, int cpu, u64 *ts,
5699 unsigned long *lost_events)
5700{
5701 struct ring_buffer_per_cpu *cpu_buffer;
5702 struct ring_buffer_event *event = NULL;
5703 unsigned long flags;
5704 bool dolock;
5705
5706 again:
5707 /* might be called in atomic */
5708 preempt_disable();
5709
5710 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5711 goto out;
5712
5713 cpu_buffer = buffer->buffers[cpu];
5714 local_irq_save(flags);
5715 dolock = rb_reader_lock(cpu_buffer);
5716
5717 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
5718 if (event) {
5719 cpu_buffer->lost_events = 0;
5720 rb_advance_reader(cpu_buffer);
5721 }
5722
5723 rb_reader_unlock(cpu_buffer, dolock);
5724 local_irq_restore(flags);
5725
5726 out:
5727 preempt_enable();
5728
5729 if (event && event->type_len == RINGBUF_TYPE_PADDING)
5730 goto again;
5731
5732 return event;
5733}
5734EXPORT_SYMBOL_GPL(ring_buffer_consume);
5735
5736/**
5737 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
5738 * @buffer: The ring buffer to read from
5739 * @cpu: The cpu buffer to iterate over
5740 * @flags: gfp flags to use for memory allocation
5741 *
5742 * This performs the initial preparations necessary to iterate
5743 * through the buffer. Memory is allocated, buffer resizing
5744 * is disabled, and the iterator pointer is returned to the caller.
5745 *
5746 * After a sequence of ring_buffer_read_prepare calls, the user is
5747 * expected to make at least one call to ring_buffer_read_prepare_sync.
5748 * Afterwards, ring_buffer_read_start is invoked to get things going
5749 * for real.
5750 *
5751 * This overall must be paired with ring_buffer_read_finish.
5752 */
5753struct ring_buffer_iter *
5754ring_buffer_read_prepare(struct trace_buffer *buffer, int cpu, gfp_t flags)
5755{
5756 struct ring_buffer_per_cpu *cpu_buffer;
5757 struct ring_buffer_iter *iter;
5758
5759 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5760 return NULL;
5761
5762 iter = kzalloc(sizeof(*iter), flags);
5763 if (!iter)
5764 return NULL;
5765
5766 /* Holds the entire event: data and meta data */
5767 iter->event_size = buffer->subbuf_size;
5768 iter->event = kmalloc(iter->event_size, flags);
5769 if (!iter->event) {
5770 kfree(iter);
5771 return NULL;
5772 }
5773
5774 cpu_buffer = buffer->buffers[cpu];
5775
5776 iter->cpu_buffer = cpu_buffer;
5777
5778 atomic_inc(&cpu_buffer->resize_disabled);
5779
5780 return iter;
5781}
5782EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
5783
5784/**
5785 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
5786 *
5787 * All previously invoked ring_buffer_read_prepare calls to prepare
5788 * iterators will be synchronized. Afterwards, read_buffer_read_start
5789 * calls on those iterators are allowed.
5790 */
5791void
5792ring_buffer_read_prepare_sync(void)
5793{
5794 synchronize_rcu();
5795}
5796EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
5797
5798/**
5799 * ring_buffer_read_start - start a non consuming read of the buffer
5800 * @iter: The iterator returned by ring_buffer_read_prepare
5801 *
5802 * This finalizes the startup of an iteration through the buffer.
5803 * The iterator comes from a call to ring_buffer_read_prepare and
5804 * an intervening ring_buffer_read_prepare_sync must have been
5805 * performed.
5806 *
5807 * Must be paired with ring_buffer_read_finish.
5808 */
5809void
5810ring_buffer_read_start(struct ring_buffer_iter *iter)
5811{
5812 struct ring_buffer_per_cpu *cpu_buffer;
5813 unsigned long flags;
5814
5815 if (!iter)
5816 return;
5817
5818 cpu_buffer = iter->cpu_buffer;
5819
5820 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5821 arch_spin_lock(&cpu_buffer->lock);
5822 rb_iter_reset(iter);
5823 arch_spin_unlock(&cpu_buffer->lock);
5824 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5825}
5826EXPORT_SYMBOL_GPL(ring_buffer_read_start);
5827
5828/**
5829 * ring_buffer_read_finish - finish reading the iterator of the buffer
5830 * @iter: The iterator retrieved by ring_buffer_start
5831 *
5832 * This re-enables resizing of the buffer, and frees the iterator.
5833 */
5834void
5835ring_buffer_read_finish(struct ring_buffer_iter *iter)
5836{
5837 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5838 unsigned long flags;
5839
5840 /* Use this opportunity to check the integrity of the ring buffer. */
5841 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5842 rb_check_pages(cpu_buffer);
5843 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5844
5845 atomic_dec(&cpu_buffer->resize_disabled);
5846 kfree(iter->event);
5847 kfree(iter);
5848}
5849EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
5850
5851/**
5852 * ring_buffer_iter_advance - advance the iterator to the next location
5853 * @iter: The ring buffer iterator
5854 *
5855 * Move the location of the iterator such that the next read will
5856 * be the next location of the iterator.
5857 */
5858void ring_buffer_iter_advance(struct ring_buffer_iter *iter)
5859{
5860 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5861 unsigned long flags;
5862
5863 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5864
5865 rb_advance_iter(iter);
5866
5867 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5868}
5869EXPORT_SYMBOL_GPL(ring_buffer_iter_advance);
5870
5871/**
5872 * ring_buffer_size - return the size of the ring buffer (in bytes)
5873 * @buffer: The ring buffer.
5874 * @cpu: The CPU to get ring buffer size from.
5875 */
5876unsigned long ring_buffer_size(struct trace_buffer *buffer, int cpu)
5877{
5878 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5879 return 0;
5880
5881 return buffer->subbuf_size * buffer->buffers[cpu]->nr_pages;
5882}
5883EXPORT_SYMBOL_GPL(ring_buffer_size);
5884
5885/**
5886 * ring_buffer_max_event_size - return the max data size of an event
5887 * @buffer: The ring buffer.
5888 *
5889 * Returns the maximum size an event can be.
5890 */
5891unsigned long ring_buffer_max_event_size(struct trace_buffer *buffer)
5892{
5893 /* If abs timestamp is requested, events have a timestamp too */
5894 if (ring_buffer_time_stamp_abs(buffer))
5895 return buffer->max_data_size - RB_LEN_TIME_EXTEND;
5896 return buffer->max_data_size;
5897}
5898EXPORT_SYMBOL_GPL(ring_buffer_max_event_size);
5899
5900static void rb_clear_buffer_page(struct buffer_page *page)
5901{
5902 local_set(&page->write, 0);
5903 local_set(&page->entries, 0);
5904 rb_init_page(page->page);
5905 page->read = 0;
5906}
5907
5908static void rb_update_meta_page(struct ring_buffer_per_cpu *cpu_buffer)
5909{
5910 struct trace_buffer_meta *meta = cpu_buffer->meta_page;
5911
5912 if (!meta)
5913 return;
5914
5915 meta->reader.read = cpu_buffer->reader_page->read;
5916 meta->reader.id = cpu_buffer->reader_page->id;
5917 meta->reader.lost_events = cpu_buffer->lost_events;
5918
5919 meta->entries = local_read(&cpu_buffer->entries);
5920 meta->overrun = local_read(&cpu_buffer->overrun);
5921 meta->read = cpu_buffer->read;
5922
5923 /* Some archs do not have data cache coherency between kernel and user-space */
5924 flush_dcache_folio(virt_to_folio(cpu_buffer->meta_page));
5925}
5926
5927static void
5928rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
5929{
5930 struct buffer_page *page;
5931
5932 rb_head_page_deactivate(cpu_buffer);
5933
5934 cpu_buffer->head_page
5935 = list_entry(cpu_buffer->pages, struct buffer_page, list);
5936 rb_clear_buffer_page(cpu_buffer->head_page);
5937 list_for_each_entry(page, cpu_buffer->pages, list) {
5938 rb_clear_buffer_page(page);
5939 }
5940
5941 cpu_buffer->tail_page = cpu_buffer->head_page;
5942 cpu_buffer->commit_page = cpu_buffer->head_page;
5943
5944 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
5945 INIT_LIST_HEAD(&cpu_buffer->new_pages);
5946 rb_clear_buffer_page(cpu_buffer->reader_page);
5947
5948 local_set(&cpu_buffer->entries_bytes, 0);
5949 local_set(&cpu_buffer->overrun, 0);
5950 local_set(&cpu_buffer->commit_overrun, 0);
5951 local_set(&cpu_buffer->dropped_events, 0);
5952 local_set(&cpu_buffer->entries, 0);
5953 local_set(&cpu_buffer->committing, 0);
5954 local_set(&cpu_buffer->commits, 0);
5955 local_set(&cpu_buffer->pages_touched, 0);
5956 local_set(&cpu_buffer->pages_lost, 0);
5957 local_set(&cpu_buffer->pages_read, 0);
5958 cpu_buffer->last_pages_touch = 0;
5959 cpu_buffer->shortest_full = 0;
5960 cpu_buffer->read = 0;
5961 cpu_buffer->read_bytes = 0;
5962
5963 rb_time_set(&cpu_buffer->write_stamp, 0);
5964 rb_time_set(&cpu_buffer->before_stamp, 0);
5965
5966 memset(cpu_buffer->event_stamp, 0, sizeof(cpu_buffer->event_stamp));
5967
5968 cpu_buffer->lost_events = 0;
5969 cpu_buffer->last_overrun = 0;
5970
5971 rb_head_page_activate(cpu_buffer);
5972 cpu_buffer->pages_removed = 0;
5973
5974 if (cpu_buffer->mapped) {
5975 rb_update_meta_page(cpu_buffer);
5976 if (cpu_buffer->ring_meta) {
5977 struct ring_buffer_meta *meta = cpu_buffer->ring_meta;
5978 meta->commit_buffer = meta->head_buffer;
5979 }
5980 }
5981}
5982
5983/* Must have disabled the cpu buffer then done a synchronize_rcu */
5984static void reset_disabled_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
5985{
5986 unsigned long flags;
5987
5988 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5989
5990 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
5991 goto out;
5992
5993 arch_spin_lock(&cpu_buffer->lock);
5994
5995 rb_reset_cpu(cpu_buffer);
5996
5997 arch_spin_unlock(&cpu_buffer->lock);
5998
5999 out:
6000 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
6001}
6002
6003/**
6004 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
6005 * @buffer: The ring buffer to reset a per cpu buffer of
6006 * @cpu: The CPU buffer to be reset
6007 */
6008void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
6009{
6010 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
6011 struct ring_buffer_meta *meta;
6012
6013 if (!cpumask_test_cpu(cpu, buffer->cpumask))
6014 return;
6015
6016 /* prevent another thread from changing buffer sizes */
6017 mutex_lock(&buffer->mutex);
6018
6019 atomic_inc(&cpu_buffer->resize_disabled);
6020 atomic_inc(&cpu_buffer->record_disabled);
6021
6022 /* Make sure all commits have finished */
6023 synchronize_rcu();
6024
6025 reset_disabled_cpu_buffer(cpu_buffer);
6026
6027 atomic_dec(&cpu_buffer->record_disabled);
6028 atomic_dec(&cpu_buffer->resize_disabled);
6029
6030 /* Make sure persistent meta now uses this buffer's addresses */
6031 meta = rb_range_meta(buffer, 0, cpu_buffer->cpu);
6032 if (meta)
6033 rb_meta_init_text_addr(meta);
6034
6035 mutex_unlock(&buffer->mutex);
6036}
6037EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
6038
6039/* Flag to ensure proper resetting of atomic variables */
6040#define RESET_BIT (1 << 30)
6041
6042/**
6043 * ring_buffer_reset_online_cpus - reset a ring buffer per CPU buffer
6044 * @buffer: The ring buffer to reset a per cpu buffer of
6045 */
6046void ring_buffer_reset_online_cpus(struct trace_buffer *buffer)
6047{
6048 struct ring_buffer_per_cpu *cpu_buffer;
6049 struct ring_buffer_meta *meta;
6050 int cpu;
6051
6052 /* prevent another thread from changing buffer sizes */
6053 mutex_lock(&buffer->mutex);
6054
6055 for_each_online_buffer_cpu(buffer, cpu) {
6056 cpu_buffer = buffer->buffers[cpu];
6057
6058 atomic_add(RESET_BIT, &cpu_buffer->resize_disabled);
6059 atomic_inc(&cpu_buffer->record_disabled);
6060 }
6061
6062 /* Make sure all commits have finished */
6063 synchronize_rcu();
6064
6065 for_each_buffer_cpu(buffer, cpu) {
6066 cpu_buffer = buffer->buffers[cpu];
6067
6068 /*
6069 * If a CPU came online during the synchronize_rcu(), then
6070 * ignore it.
6071 */
6072 if (!(atomic_read(&cpu_buffer->resize_disabled) & RESET_BIT))
6073 continue;
6074
6075 reset_disabled_cpu_buffer(cpu_buffer);
6076
6077 /* Make sure persistent meta now uses this buffer's addresses */
6078 meta = rb_range_meta(buffer, 0, cpu_buffer->cpu);
6079 if (meta)
6080 rb_meta_init_text_addr(meta);
6081
6082 atomic_dec(&cpu_buffer->record_disabled);
6083 atomic_sub(RESET_BIT, &cpu_buffer->resize_disabled);
6084 }
6085
6086 mutex_unlock(&buffer->mutex);
6087}
6088
6089/**
6090 * ring_buffer_reset - reset a ring buffer
6091 * @buffer: The ring buffer to reset all cpu buffers
6092 */
6093void ring_buffer_reset(struct trace_buffer *buffer)
6094{
6095 struct ring_buffer_per_cpu *cpu_buffer;
6096 int cpu;
6097
6098 /* prevent another thread from changing buffer sizes */
6099 mutex_lock(&buffer->mutex);
6100
6101 for_each_buffer_cpu(buffer, cpu) {
6102 cpu_buffer = buffer->buffers[cpu];
6103
6104 atomic_inc(&cpu_buffer->resize_disabled);
6105 atomic_inc(&cpu_buffer->record_disabled);
6106 }
6107
6108 /* Make sure all commits have finished */
6109 synchronize_rcu();
6110
6111 for_each_buffer_cpu(buffer, cpu) {
6112 cpu_buffer = buffer->buffers[cpu];
6113
6114 reset_disabled_cpu_buffer(cpu_buffer);
6115
6116 atomic_dec(&cpu_buffer->record_disabled);
6117 atomic_dec(&cpu_buffer->resize_disabled);
6118 }
6119
6120 mutex_unlock(&buffer->mutex);
6121}
6122EXPORT_SYMBOL_GPL(ring_buffer_reset);
6123
6124/**
6125 * ring_buffer_empty - is the ring buffer empty?
6126 * @buffer: The ring buffer to test
6127 */
6128bool ring_buffer_empty(struct trace_buffer *buffer)
6129{
6130 struct ring_buffer_per_cpu *cpu_buffer;
6131 unsigned long flags;
6132 bool dolock;
6133 bool ret;
6134 int cpu;
6135
6136 /* yes this is racy, but if you don't like the race, lock the buffer */
6137 for_each_buffer_cpu(buffer, cpu) {
6138 cpu_buffer = buffer->buffers[cpu];
6139 local_irq_save(flags);
6140 dolock = rb_reader_lock(cpu_buffer);
6141 ret = rb_per_cpu_empty(cpu_buffer);
6142 rb_reader_unlock(cpu_buffer, dolock);
6143 local_irq_restore(flags);
6144
6145 if (!ret)
6146 return false;
6147 }
6148
6149 return true;
6150}
6151EXPORT_SYMBOL_GPL(ring_buffer_empty);
6152
6153/**
6154 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
6155 * @buffer: The ring buffer
6156 * @cpu: The CPU buffer to test
6157 */
6158bool ring_buffer_empty_cpu(struct trace_buffer *buffer, int cpu)
6159{
6160 struct ring_buffer_per_cpu *cpu_buffer;
6161 unsigned long flags;
6162 bool dolock;
6163 bool ret;
6164
6165 if (!cpumask_test_cpu(cpu, buffer->cpumask))
6166 return true;
6167
6168 cpu_buffer = buffer->buffers[cpu];
6169 local_irq_save(flags);
6170 dolock = rb_reader_lock(cpu_buffer);
6171 ret = rb_per_cpu_empty(cpu_buffer);
6172 rb_reader_unlock(cpu_buffer, dolock);
6173 local_irq_restore(flags);
6174
6175 return ret;
6176}
6177EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
6178
6179#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
6180/**
6181 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
6182 * @buffer_a: One buffer to swap with
6183 * @buffer_b: The other buffer to swap with
6184 * @cpu: the CPU of the buffers to swap
6185 *
6186 * This function is useful for tracers that want to take a "snapshot"
6187 * of a CPU buffer and has another back up buffer lying around.
6188 * it is expected that the tracer handles the cpu buffer not being
6189 * used at the moment.
6190 */
6191int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
6192 struct trace_buffer *buffer_b, int cpu)
6193{
6194 struct ring_buffer_per_cpu *cpu_buffer_a;
6195 struct ring_buffer_per_cpu *cpu_buffer_b;
6196 int ret = -EINVAL;
6197
6198 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
6199 !cpumask_test_cpu(cpu, buffer_b->cpumask))
6200 goto out;
6201
6202 cpu_buffer_a = buffer_a->buffers[cpu];
6203 cpu_buffer_b = buffer_b->buffers[cpu];
6204
6205 /* It's up to the callers to not try to swap mapped buffers */
6206 if (WARN_ON_ONCE(cpu_buffer_a->mapped || cpu_buffer_b->mapped)) {
6207 ret = -EBUSY;
6208 goto out;
6209 }
6210
6211 /* At least make sure the two buffers are somewhat the same */
6212 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
6213 goto out;
6214
6215 if (buffer_a->subbuf_order != buffer_b->subbuf_order)
6216 goto out;
6217
6218 ret = -EAGAIN;
6219
6220 if (atomic_read(&buffer_a->record_disabled))
6221 goto out;
6222
6223 if (atomic_read(&buffer_b->record_disabled))
6224 goto out;
6225
6226 if (atomic_read(&cpu_buffer_a->record_disabled))
6227 goto out;
6228
6229 if (atomic_read(&cpu_buffer_b->record_disabled))
6230 goto out;
6231
6232 /*
6233 * We can't do a synchronize_rcu here because this
6234 * function can be called in atomic context.
6235 * Normally this will be called from the same CPU as cpu.
6236 * If not it's up to the caller to protect this.
6237 */
6238 atomic_inc(&cpu_buffer_a->record_disabled);
6239 atomic_inc(&cpu_buffer_b->record_disabled);
6240
6241 ret = -EBUSY;
6242 if (local_read(&cpu_buffer_a->committing))
6243 goto out_dec;
6244 if (local_read(&cpu_buffer_b->committing))
6245 goto out_dec;
6246
6247 /*
6248 * When resize is in progress, we cannot swap it because
6249 * it will mess the state of the cpu buffer.
6250 */
6251 if (atomic_read(&buffer_a->resizing))
6252 goto out_dec;
6253 if (atomic_read(&buffer_b->resizing))
6254 goto out_dec;
6255
6256 buffer_a->buffers[cpu] = cpu_buffer_b;
6257 buffer_b->buffers[cpu] = cpu_buffer_a;
6258
6259 cpu_buffer_b->buffer = buffer_a;
6260 cpu_buffer_a->buffer = buffer_b;
6261
6262 ret = 0;
6263
6264out_dec:
6265 atomic_dec(&cpu_buffer_a->record_disabled);
6266 atomic_dec(&cpu_buffer_b->record_disabled);
6267out:
6268 return ret;
6269}
6270EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
6271#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
6272
6273/**
6274 * ring_buffer_alloc_read_page - allocate a page to read from buffer
6275 * @buffer: the buffer to allocate for.
6276 * @cpu: the cpu buffer to allocate.
6277 *
6278 * This function is used in conjunction with ring_buffer_read_page.
6279 * When reading a full page from the ring buffer, these functions
6280 * can be used to speed up the process. The calling function should
6281 * allocate a few pages first with this function. Then when it
6282 * needs to get pages from the ring buffer, it passes the result
6283 * of this function into ring_buffer_read_page, which will swap
6284 * the page that was allocated, with the read page of the buffer.
6285 *
6286 * Returns:
6287 * The page allocated, or ERR_PTR
6288 */
6289struct buffer_data_read_page *
6290ring_buffer_alloc_read_page(struct trace_buffer *buffer, int cpu)
6291{
6292 struct ring_buffer_per_cpu *cpu_buffer;
6293 struct buffer_data_read_page *bpage = NULL;
6294 unsigned long flags;
6295 struct page *page;
6296
6297 if (!cpumask_test_cpu(cpu, buffer->cpumask))
6298 return ERR_PTR(-ENODEV);
6299
6300 bpage = kzalloc(sizeof(*bpage), GFP_KERNEL);
6301 if (!bpage)
6302 return ERR_PTR(-ENOMEM);
6303
6304 bpage->order = buffer->subbuf_order;
6305 cpu_buffer = buffer->buffers[cpu];
6306 local_irq_save(flags);
6307 arch_spin_lock(&cpu_buffer->lock);
6308
6309 if (cpu_buffer->free_page) {
6310 bpage->data = cpu_buffer->free_page;
6311 cpu_buffer->free_page = NULL;
6312 }
6313
6314 arch_spin_unlock(&cpu_buffer->lock);
6315 local_irq_restore(flags);
6316
6317 if (bpage->data)
6318 goto out;
6319
6320 page = alloc_pages_node(cpu_to_node(cpu),
6321 GFP_KERNEL | __GFP_NORETRY | __GFP_COMP | __GFP_ZERO,
6322 cpu_buffer->buffer->subbuf_order);
6323 if (!page) {
6324 kfree(bpage);
6325 return ERR_PTR(-ENOMEM);
6326 }
6327
6328 bpage->data = page_address(page);
6329
6330 out:
6331 rb_init_page(bpage->data);
6332
6333 return bpage;
6334}
6335EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
6336
6337/**
6338 * ring_buffer_free_read_page - free an allocated read page
6339 * @buffer: the buffer the page was allocate for
6340 * @cpu: the cpu buffer the page came from
6341 * @data_page: the page to free
6342 *
6343 * Free a page allocated from ring_buffer_alloc_read_page.
6344 */
6345void ring_buffer_free_read_page(struct trace_buffer *buffer, int cpu,
6346 struct buffer_data_read_page *data_page)
6347{
6348 struct ring_buffer_per_cpu *cpu_buffer;
6349 struct buffer_data_page *bpage = data_page->data;
6350 struct page *page = virt_to_page(bpage);
6351 unsigned long flags;
6352
6353 if (!buffer || !buffer->buffers || !buffer->buffers[cpu])
6354 return;
6355
6356 cpu_buffer = buffer->buffers[cpu];
6357
6358 /*
6359 * If the page is still in use someplace else, or order of the page
6360 * is different from the subbuffer order of the buffer -
6361 * we can't reuse it
6362 */
6363 if (page_ref_count(page) > 1 || data_page->order != buffer->subbuf_order)
6364 goto out;
6365
6366 local_irq_save(flags);
6367 arch_spin_lock(&cpu_buffer->lock);
6368
6369 if (!cpu_buffer->free_page) {
6370 cpu_buffer->free_page = bpage;
6371 bpage = NULL;
6372 }
6373
6374 arch_spin_unlock(&cpu_buffer->lock);
6375 local_irq_restore(flags);
6376
6377 out:
6378 free_pages((unsigned long)bpage, data_page->order);
6379 kfree(data_page);
6380}
6381EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
6382
6383/**
6384 * ring_buffer_read_page - extract a page from the ring buffer
6385 * @buffer: buffer to extract from
6386 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
6387 * @len: amount to extract
6388 * @cpu: the cpu of the buffer to extract
6389 * @full: should the extraction only happen when the page is full.
6390 *
6391 * This function will pull out a page from the ring buffer and consume it.
6392 * @data_page must be the address of the variable that was returned
6393 * from ring_buffer_alloc_read_page. This is because the page might be used
6394 * to swap with a page in the ring buffer.
6395 *
6396 * for example:
6397 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
6398 * if (IS_ERR(rpage))
6399 * return PTR_ERR(rpage);
6400 * ret = ring_buffer_read_page(buffer, rpage, len, cpu, 0);
6401 * if (ret >= 0)
6402 * process_page(ring_buffer_read_page_data(rpage), ret);
6403 * ring_buffer_free_read_page(buffer, cpu, rpage);
6404 *
6405 * When @full is set, the function will not return true unless
6406 * the writer is off the reader page.
6407 *
6408 * Note: it is up to the calling functions to handle sleeps and wakeups.
6409 * The ring buffer can be used anywhere in the kernel and can not
6410 * blindly call wake_up. The layer that uses the ring buffer must be
6411 * responsible for that.
6412 *
6413 * Returns:
6414 * >=0 if data has been transferred, returns the offset of consumed data.
6415 * <0 if no data has been transferred.
6416 */
6417int ring_buffer_read_page(struct trace_buffer *buffer,
6418 struct buffer_data_read_page *data_page,
6419 size_t len, int cpu, int full)
6420{
6421 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
6422 struct ring_buffer_event *event;
6423 struct buffer_data_page *bpage;
6424 struct buffer_page *reader;
6425 unsigned long missed_events;
6426 unsigned long flags;
6427 unsigned int commit;
6428 unsigned int read;
6429 u64 save_timestamp;
6430 int ret = -1;
6431
6432 if (!cpumask_test_cpu(cpu, buffer->cpumask))
6433 goto out;
6434
6435 /*
6436 * If len is not big enough to hold the page header, then
6437 * we can not copy anything.
6438 */
6439 if (len <= BUF_PAGE_HDR_SIZE)
6440 goto out;
6441
6442 len -= BUF_PAGE_HDR_SIZE;
6443
6444 if (!data_page || !data_page->data)
6445 goto out;
6446 if (data_page->order != buffer->subbuf_order)
6447 goto out;
6448
6449 bpage = data_page->data;
6450 if (!bpage)
6451 goto out;
6452
6453 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
6454
6455 reader = rb_get_reader_page(cpu_buffer);
6456 if (!reader)
6457 goto out_unlock;
6458
6459 event = rb_reader_event(cpu_buffer);
6460
6461 read = reader->read;
6462 commit = rb_page_size(reader);
6463
6464 /* Check if any events were dropped */
6465 missed_events = cpu_buffer->lost_events;
6466
6467 /*
6468 * If this page has been partially read or
6469 * if len is not big enough to read the rest of the page or
6470 * a writer is still on the page, then
6471 * we must copy the data from the page to the buffer.
6472 * Otherwise, we can simply swap the page with the one passed in.
6473 */
6474 if (read || (len < (commit - read)) ||
6475 cpu_buffer->reader_page == cpu_buffer->commit_page ||
6476 cpu_buffer->mapped) {
6477 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
6478 unsigned int rpos = read;
6479 unsigned int pos = 0;
6480 unsigned int size;
6481
6482 /*
6483 * If a full page is expected, this can still be returned
6484 * if there's been a previous partial read and the
6485 * rest of the page can be read and the commit page is off
6486 * the reader page.
6487 */
6488 if (full &&
6489 (!read || (len < (commit - read)) ||
6490 cpu_buffer->reader_page == cpu_buffer->commit_page))
6491 goto out_unlock;
6492
6493 if (len > (commit - read))
6494 len = (commit - read);
6495
6496 /* Always keep the time extend and data together */
6497 size = rb_event_ts_length(event);
6498
6499 if (len < size)
6500 goto out_unlock;
6501
6502 /* save the current timestamp, since the user will need it */
6503 save_timestamp = cpu_buffer->read_stamp;
6504
6505 /* Need to copy one event at a time */
6506 do {
6507 /* We need the size of one event, because
6508 * rb_advance_reader only advances by one event,
6509 * whereas rb_event_ts_length may include the size of
6510 * one or two events.
6511 * We have already ensured there's enough space if this
6512 * is a time extend. */
6513 size = rb_event_length(event);
6514 memcpy(bpage->data + pos, rpage->data + rpos, size);
6515
6516 len -= size;
6517
6518 rb_advance_reader(cpu_buffer);
6519 rpos = reader->read;
6520 pos += size;
6521
6522 if (rpos >= commit)
6523 break;
6524
6525 event = rb_reader_event(cpu_buffer);
6526 /* Always keep the time extend and data together */
6527 size = rb_event_ts_length(event);
6528 } while (len >= size);
6529
6530 /* update bpage */
6531 local_set(&bpage->commit, pos);
6532 bpage->time_stamp = save_timestamp;
6533
6534 /* we copied everything to the beginning */
6535 read = 0;
6536 } else {
6537 /* update the entry counter */
6538 cpu_buffer->read += rb_page_entries(reader);
6539 cpu_buffer->read_bytes += rb_page_size(reader);
6540
6541 /* swap the pages */
6542 rb_init_page(bpage);
6543 bpage = reader->page;
6544 reader->page = data_page->data;
6545 local_set(&reader->write, 0);
6546 local_set(&reader->entries, 0);
6547 reader->read = 0;
6548 data_page->data = bpage;
6549
6550 /*
6551 * Use the real_end for the data size,
6552 * This gives us a chance to store the lost events
6553 * on the page.
6554 */
6555 if (reader->real_end)
6556 local_set(&bpage->commit, reader->real_end);
6557 }
6558 ret = read;
6559
6560 cpu_buffer->lost_events = 0;
6561
6562 commit = local_read(&bpage->commit);
6563 /*
6564 * Set a flag in the commit field if we lost events
6565 */
6566 if (missed_events) {
6567 /* If there is room at the end of the page to save the
6568 * missed events, then record it there.
6569 */
6570 if (buffer->subbuf_size - commit >= sizeof(missed_events)) {
6571 memcpy(&bpage->data[commit], &missed_events,
6572 sizeof(missed_events));
6573 local_add(RB_MISSED_STORED, &bpage->commit);
6574 commit += sizeof(missed_events);
6575 }
6576 local_add(RB_MISSED_EVENTS, &bpage->commit);
6577 }
6578
6579 /*
6580 * This page may be off to user land. Zero it out here.
6581 */
6582 if (commit < buffer->subbuf_size)
6583 memset(&bpage->data[commit], 0, buffer->subbuf_size - commit);
6584
6585 out_unlock:
6586 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
6587
6588 out:
6589 return ret;
6590}
6591EXPORT_SYMBOL_GPL(ring_buffer_read_page);
6592
6593/**
6594 * ring_buffer_read_page_data - get pointer to the data in the page.
6595 * @page: the page to get the data from
6596 *
6597 * Returns pointer to the actual data in this page.
6598 */
6599void *ring_buffer_read_page_data(struct buffer_data_read_page *page)
6600{
6601 return page->data;
6602}
6603EXPORT_SYMBOL_GPL(ring_buffer_read_page_data);
6604
6605/**
6606 * ring_buffer_subbuf_size_get - get size of the sub buffer.
6607 * @buffer: the buffer to get the sub buffer size from
6608 *
6609 * Returns size of the sub buffer, in bytes.
6610 */
6611int ring_buffer_subbuf_size_get(struct trace_buffer *buffer)
6612{
6613 return buffer->subbuf_size + BUF_PAGE_HDR_SIZE;
6614}
6615EXPORT_SYMBOL_GPL(ring_buffer_subbuf_size_get);
6616
6617/**
6618 * ring_buffer_subbuf_order_get - get order of system sub pages in one buffer page.
6619 * @buffer: The ring_buffer to get the system sub page order from
6620 *
6621 * By default, one ring buffer sub page equals to one system page. This parameter
6622 * is configurable, per ring buffer. The size of the ring buffer sub page can be
6623 * extended, but must be an order of system page size.
6624 *
6625 * Returns the order of buffer sub page size, in system pages:
6626 * 0 means the sub buffer size is 1 system page and so forth.
6627 * In case of an error < 0 is returned.
6628 */
6629int ring_buffer_subbuf_order_get(struct trace_buffer *buffer)
6630{
6631 if (!buffer)
6632 return -EINVAL;
6633
6634 return buffer->subbuf_order;
6635}
6636EXPORT_SYMBOL_GPL(ring_buffer_subbuf_order_get);
6637
6638/**
6639 * ring_buffer_subbuf_order_set - set the size of ring buffer sub page.
6640 * @buffer: The ring_buffer to set the new page size.
6641 * @order: Order of the system pages in one sub buffer page
6642 *
6643 * By default, one ring buffer pages equals to one system page. This API can be
6644 * used to set new size of the ring buffer page. The size must be order of
6645 * system page size, that's why the input parameter @order is the order of
6646 * system pages that are allocated for one ring buffer page:
6647 * 0 - 1 system page
6648 * 1 - 2 system pages
6649 * 3 - 4 system pages
6650 * ...
6651 *
6652 * Returns 0 on success or < 0 in case of an error.
6653 */
6654int ring_buffer_subbuf_order_set(struct trace_buffer *buffer, int order)
6655{
6656 struct ring_buffer_per_cpu *cpu_buffer;
6657 struct buffer_page *bpage, *tmp;
6658 int old_order, old_size;
6659 int nr_pages;
6660 int psize;
6661 int err;
6662 int cpu;
6663
6664 if (!buffer || order < 0)
6665 return -EINVAL;
6666
6667 if (buffer->subbuf_order == order)
6668 return 0;
6669
6670 psize = (1 << order) * PAGE_SIZE;
6671 if (psize <= BUF_PAGE_HDR_SIZE)
6672 return -EINVAL;
6673
6674 /* Size of a subbuf cannot be greater than the write counter */
6675 if (psize > RB_WRITE_MASK + 1)
6676 return -EINVAL;
6677
6678 old_order = buffer->subbuf_order;
6679 old_size = buffer->subbuf_size;
6680
6681 /* prevent another thread from changing buffer sizes */
6682 mutex_lock(&buffer->mutex);
6683 atomic_inc(&buffer->record_disabled);
6684
6685 /* Make sure all commits have finished */
6686 synchronize_rcu();
6687
6688 buffer->subbuf_order = order;
6689 buffer->subbuf_size = psize - BUF_PAGE_HDR_SIZE;
6690
6691 /* Make sure all new buffers are allocated, before deleting the old ones */
6692 for_each_buffer_cpu(buffer, cpu) {
6693
6694 if (!cpumask_test_cpu(cpu, buffer->cpumask))
6695 continue;
6696
6697 cpu_buffer = buffer->buffers[cpu];
6698
6699 if (cpu_buffer->mapped) {
6700 err = -EBUSY;
6701 goto error;
6702 }
6703
6704 /* Update the number of pages to match the new size */
6705 nr_pages = old_size * buffer->buffers[cpu]->nr_pages;
6706 nr_pages = DIV_ROUND_UP(nr_pages, buffer->subbuf_size);
6707
6708 /* we need a minimum of two pages */
6709 if (nr_pages < 2)
6710 nr_pages = 2;
6711
6712 cpu_buffer->nr_pages_to_update = nr_pages;
6713
6714 /* Include the reader page */
6715 nr_pages++;
6716
6717 /* Allocate the new size buffer */
6718 INIT_LIST_HEAD(&cpu_buffer->new_pages);
6719 if (__rb_allocate_pages(cpu_buffer, nr_pages,
6720 &cpu_buffer->new_pages)) {
6721 /* not enough memory for new pages */
6722 err = -ENOMEM;
6723 goto error;
6724 }
6725 }
6726
6727 for_each_buffer_cpu(buffer, cpu) {
6728
6729 if (!cpumask_test_cpu(cpu, buffer->cpumask))
6730 continue;
6731
6732 cpu_buffer = buffer->buffers[cpu];
6733
6734 /* Clear the head bit to make the link list normal to read */
6735 rb_head_page_deactivate(cpu_buffer);
6736
6737 /* Now walk the list and free all the old sub buffers */
6738 list_for_each_entry_safe(bpage, tmp, cpu_buffer->pages, list) {
6739 list_del_init(&bpage->list);
6740 free_buffer_page(bpage);
6741 }
6742 /* The above loop stopped an the last page needing to be freed */
6743 bpage = list_entry(cpu_buffer->pages, struct buffer_page, list);
6744 free_buffer_page(bpage);
6745
6746 /* Free the current reader page */
6747 free_buffer_page(cpu_buffer->reader_page);
6748
6749 /* One page was allocated for the reader page */
6750 cpu_buffer->reader_page = list_entry(cpu_buffer->new_pages.next,
6751 struct buffer_page, list);
6752 list_del_init(&cpu_buffer->reader_page->list);
6753
6754 /* The cpu_buffer pages are a link list with no head */
6755 cpu_buffer->pages = cpu_buffer->new_pages.next;
6756 cpu_buffer->new_pages.next->prev = cpu_buffer->new_pages.prev;
6757 cpu_buffer->new_pages.prev->next = cpu_buffer->new_pages.next;
6758
6759 /* Clear the new_pages list */
6760 INIT_LIST_HEAD(&cpu_buffer->new_pages);
6761
6762 cpu_buffer->head_page
6763 = list_entry(cpu_buffer->pages, struct buffer_page, list);
6764 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
6765
6766 cpu_buffer->nr_pages = cpu_buffer->nr_pages_to_update;
6767 cpu_buffer->nr_pages_to_update = 0;
6768
6769 free_pages((unsigned long)cpu_buffer->free_page, old_order);
6770 cpu_buffer->free_page = NULL;
6771
6772 rb_head_page_activate(cpu_buffer);
6773
6774 rb_check_pages(cpu_buffer);
6775 }
6776
6777 atomic_dec(&buffer->record_disabled);
6778 mutex_unlock(&buffer->mutex);
6779
6780 return 0;
6781
6782error:
6783 buffer->subbuf_order = old_order;
6784 buffer->subbuf_size = old_size;
6785
6786 atomic_dec(&buffer->record_disabled);
6787 mutex_unlock(&buffer->mutex);
6788
6789 for_each_buffer_cpu(buffer, cpu) {
6790 cpu_buffer = buffer->buffers[cpu];
6791
6792 if (!cpu_buffer->nr_pages_to_update)
6793 continue;
6794
6795 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages, list) {
6796 list_del_init(&bpage->list);
6797 free_buffer_page(bpage);
6798 }
6799 }
6800
6801 return err;
6802}
6803EXPORT_SYMBOL_GPL(ring_buffer_subbuf_order_set);
6804
6805static int rb_alloc_meta_page(struct ring_buffer_per_cpu *cpu_buffer)
6806{
6807 struct page *page;
6808
6809 if (cpu_buffer->meta_page)
6810 return 0;
6811
6812 page = alloc_page(GFP_USER | __GFP_ZERO);
6813 if (!page)
6814 return -ENOMEM;
6815
6816 cpu_buffer->meta_page = page_to_virt(page);
6817
6818 return 0;
6819}
6820
6821static void rb_free_meta_page(struct ring_buffer_per_cpu *cpu_buffer)
6822{
6823 unsigned long addr = (unsigned long)cpu_buffer->meta_page;
6824
6825 free_page(addr);
6826 cpu_buffer->meta_page = NULL;
6827}
6828
6829static void rb_setup_ids_meta_page(struct ring_buffer_per_cpu *cpu_buffer,
6830 unsigned long *subbuf_ids)
6831{
6832 struct trace_buffer_meta *meta = cpu_buffer->meta_page;
6833 unsigned int nr_subbufs = cpu_buffer->nr_pages + 1;
6834 struct buffer_page *first_subbuf, *subbuf;
6835 int id = 0;
6836
6837 subbuf_ids[id] = (unsigned long)cpu_buffer->reader_page->page;
6838 cpu_buffer->reader_page->id = id++;
6839
6840 first_subbuf = subbuf = rb_set_head_page(cpu_buffer);
6841 do {
6842 if (WARN_ON(id >= nr_subbufs))
6843 break;
6844
6845 subbuf_ids[id] = (unsigned long)subbuf->page;
6846 subbuf->id = id;
6847
6848 rb_inc_page(&subbuf);
6849 id++;
6850 } while (subbuf != first_subbuf);
6851
6852 /* install subbuf ID to kern VA translation */
6853 cpu_buffer->subbuf_ids = subbuf_ids;
6854
6855 meta->meta_struct_len = sizeof(*meta);
6856 meta->nr_subbufs = nr_subbufs;
6857 meta->subbuf_size = cpu_buffer->buffer->subbuf_size + BUF_PAGE_HDR_SIZE;
6858 meta->meta_page_size = meta->subbuf_size;
6859
6860 rb_update_meta_page(cpu_buffer);
6861}
6862
6863static struct ring_buffer_per_cpu *
6864rb_get_mapped_buffer(struct trace_buffer *buffer, int cpu)
6865{
6866 struct ring_buffer_per_cpu *cpu_buffer;
6867
6868 if (!cpumask_test_cpu(cpu, buffer->cpumask))
6869 return ERR_PTR(-EINVAL);
6870
6871 cpu_buffer = buffer->buffers[cpu];
6872
6873 mutex_lock(&cpu_buffer->mapping_lock);
6874
6875 if (!cpu_buffer->user_mapped) {
6876 mutex_unlock(&cpu_buffer->mapping_lock);
6877 return ERR_PTR(-ENODEV);
6878 }
6879
6880 return cpu_buffer;
6881}
6882
6883static void rb_put_mapped_buffer(struct ring_buffer_per_cpu *cpu_buffer)
6884{
6885 mutex_unlock(&cpu_buffer->mapping_lock);
6886}
6887
6888/*
6889 * Fast-path for rb_buffer_(un)map(). Called whenever the meta-page doesn't need
6890 * to be set-up or torn-down.
6891 */
6892static int __rb_inc_dec_mapped(struct ring_buffer_per_cpu *cpu_buffer,
6893 bool inc)
6894{
6895 unsigned long flags;
6896
6897 lockdep_assert_held(&cpu_buffer->mapping_lock);
6898
6899 /* mapped is always greater or equal to user_mapped */
6900 if (WARN_ON(cpu_buffer->mapped < cpu_buffer->user_mapped))
6901 return -EINVAL;
6902
6903 if (inc && cpu_buffer->mapped == UINT_MAX)
6904 return -EBUSY;
6905
6906 if (WARN_ON(!inc && cpu_buffer->user_mapped == 0))
6907 return -EINVAL;
6908
6909 mutex_lock(&cpu_buffer->buffer->mutex);
6910 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
6911
6912 if (inc) {
6913 cpu_buffer->user_mapped++;
6914 cpu_buffer->mapped++;
6915 } else {
6916 cpu_buffer->user_mapped--;
6917 cpu_buffer->mapped--;
6918 }
6919
6920 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
6921 mutex_unlock(&cpu_buffer->buffer->mutex);
6922
6923 return 0;
6924}
6925
6926/*
6927 * +--------------+ pgoff == 0
6928 * | meta page |
6929 * +--------------+ pgoff == 1
6930 * | subbuffer 0 |
6931 * | |
6932 * +--------------+ pgoff == (1 + (1 << subbuf_order))
6933 * | subbuffer 1 |
6934 * | |
6935 * ...
6936 */
6937#ifdef CONFIG_MMU
6938static int __rb_map_vma(struct ring_buffer_per_cpu *cpu_buffer,
6939 struct vm_area_struct *vma)
6940{
6941 unsigned long nr_subbufs, nr_pages, nr_vma_pages, pgoff = vma->vm_pgoff;
6942 unsigned int subbuf_pages, subbuf_order;
6943 struct page **pages;
6944 int p = 0, s = 0;
6945 int err;
6946
6947 /* Refuse MP_PRIVATE or writable mappings */
6948 if (vma->vm_flags & VM_WRITE || vma->vm_flags & VM_EXEC ||
6949 !(vma->vm_flags & VM_MAYSHARE))
6950 return -EPERM;
6951
6952 subbuf_order = cpu_buffer->buffer->subbuf_order;
6953 subbuf_pages = 1 << subbuf_order;
6954
6955 if (subbuf_order && pgoff % subbuf_pages)
6956 return -EINVAL;
6957
6958 /*
6959 * Make sure the mapping cannot become writable later. Also tell the VM
6960 * to not touch these pages (VM_DONTCOPY | VM_DONTEXPAND).
6961 */
6962 vm_flags_mod(vma, VM_DONTCOPY | VM_DONTEXPAND | VM_DONTDUMP,
6963 VM_MAYWRITE);
6964
6965 lockdep_assert_held(&cpu_buffer->mapping_lock);
6966
6967 nr_subbufs = cpu_buffer->nr_pages + 1; /* + reader-subbuf */
6968 nr_pages = ((nr_subbufs + 1) << subbuf_order) - pgoff; /* + meta-page */
6969
6970 nr_vma_pages = vma_pages(vma);
6971 if (!nr_vma_pages || nr_vma_pages > nr_pages)
6972 return -EINVAL;
6973
6974 nr_pages = nr_vma_pages;
6975
6976 pages = kcalloc(nr_pages, sizeof(*pages), GFP_KERNEL);
6977 if (!pages)
6978 return -ENOMEM;
6979
6980 if (!pgoff) {
6981 unsigned long meta_page_padding;
6982
6983 pages[p++] = virt_to_page(cpu_buffer->meta_page);
6984
6985 /*
6986 * Pad with the zero-page to align the meta-page with the
6987 * sub-buffers.
6988 */
6989 meta_page_padding = subbuf_pages - 1;
6990 while (meta_page_padding-- && p < nr_pages) {
6991 unsigned long __maybe_unused zero_addr =
6992 vma->vm_start + (PAGE_SIZE * p);
6993
6994 pages[p++] = ZERO_PAGE(zero_addr);
6995 }
6996 } else {
6997 /* Skip the meta-page */
6998 pgoff -= subbuf_pages;
6999
7000 s += pgoff / subbuf_pages;
7001 }
7002
7003 while (p < nr_pages) {
7004 struct page *page = virt_to_page((void *)cpu_buffer->subbuf_ids[s]);
7005 int off = 0;
7006
7007 if (WARN_ON_ONCE(s >= nr_subbufs)) {
7008 err = -EINVAL;
7009 goto out;
7010 }
7011
7012 for (; off < (1 << (subbuf_order)); off++, page++) {
7013 if (p >= nr_pages)
7014 break;
7015
7016 pages[p++] = page;
7017 }
7018 s++;
7019 }
7020
7021 err = vm_insert_pages(vma, vma->vm_start, pages, &nr_pages);
7022
7023out:
7024 kfree(pages);
7025
7026 return err;
7027}
7028#else
7029static int __rb_map_vma(struct ring_buffer_per_cpu *cpu_buffer,
7030 struct vm_area_struct *vma)
7031{
7032 return -EOPNOTSUPP;
7033}
7034#endif
7035
7036int ring_buffer_map(struct trace_buffer *buffer, int cpu,
7037 struct vm_area_struct *vma)
7038{
7039 struct ring_buffer_per_cpu *cpu_buffer;
7040 unsigned long flags, *subbuf_ids;
7041 int err = 0;
7042
7043 if (!cpumask_test_cpu(cpu, buffer->cpumask))
7044 return -EINVAL;
7045
7046 cpu_buffer = buffer->buffers[cpu];
7047
7048 mutex_lock(&cpu_buffer->mapping_lock);
7049
7050 if (cpu_buffer->user_mapped) {
7051 err = __rb_map_vma(cpu_buffer, vma);
7052 if (!err)
7053 err = __rb_inc_dec_mapped(cpu_buffer, true);
7054 mutex_unlock(&cpu_buffer->mapping_lock);
7055 return err;
7056 }
7057
7058 /* prevent another thread from changing buffer/sub-buffer sizes */
7059 mutex_lock(&buffer->mutex);
7060
7061 err = rb_alloc_meta_page(cpu_buffer);
7062 if (err)
7063 goto unlock;
7064
7065 /* subbuf_ids include the reader while nr_pages does not */
7066 subbuf_ids = kcalloc(cpu_buffer->nr_pages + 1, sizeof(*subbuf_ids), GFP_KERNEL);
7067 if (!subbuf_ids) {
7068 rb_free_meta_page(cpu_buffer);
7069 err = -ENOMEM;
7070 goto unlock;
7071 }
7072
7073 atomic_inc(&cpu_buffer->resize_disabled);
7074
7075 /*
7076 * Lock all readers to block any subbuf swap until the subbuf IDs are
7077 * assigned.
7078 */
7079 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
7080 rb_setup_ids_meta_page(cpu_buffer, subbuf_ids);
7081
7082 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
7083
7084 err = __rb_map_vma(cpu_buffer, vma);
7085 if (!err) {
7086 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
7087 /* This is the first time it is mapped by user */
7088 cpu_buffer->mapped++;
7089 cpu_buffer->user_mapped = 1;
7090 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
7091 } else {
7092 kfree(cpu_buffer->subbuf_ids);
7093 cpu_buffer->subbuf_ids = NULL;
7094 rb_free_meta_page(cpu_buffer);
7095 }
7096
7097unlock:
7098 mutex_unlock(&buffer->mutex);
7099 mutex_unlock(&cpu_buffer->mapping_lock);
7100
7101 return err;
7102}
7103
7104int ring_buffer_unmap(struct trace_buffer *buffer, int cpu)
7105{
7106 struct ring_buffer_per_cpu *cpu_buffer;
7107 unsigned long flags;
7108 int err = 0;
7109
7110 if (!cpumask_test_cpu(cpu, buffer->cpumask))
7111 return -EINVAL;
7112
7113 cpu_buffer = buffer->buffers[cpu];
7114
7115 mutex_lock(&cpu_buffer->mapping_lock);
7116
7117 if (!cpu_buffer->user_mapped) {
7118 err = -ENODEV;
7119 goto out;
7120 } else if (cpu_buffer->user_mapped > 1) {
7121 __rb_inc_dec_mapped(cpu_buffer, false);
7122 goto out;
7123 }
7124
7125 mutex_lock(&buffer->mutex);
7126 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
7127
7128 /* This is the last user space mapping */
7129 if (!WARN_ON_ONCE(cpu_buffer->mapped < cpu_buffer->user_mapped))
7130 cpu_buffer->mapped--;
7131 cpu_buffer->user_mapped = 0;
7132
7133 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
7134
7135 kfree(cpu_buffer->subbuf_ids);
7136 cpu_buffer->subbuf_ids = NULL;
7137 rb_free_meta_page(cpu_buffer);
7138 atomic_dec(&cpu_buffer->resize_disabled);
7139
7140 mutex_unlock(&buffer->mutex);
7141
7142out:
7143 mutex_unlock(&cpu_buffer->mapping_lock);
7144
7145 return err;
7146}
7147
7148int ring_buffer_map_get_reader(struct trace_buffer *buffer, int cpu)
7149{
7150 struct ring_buffer_per_cpu *cpu_buffer;
7151 struct buffer_page *reader;
7152 unsigned long missed_events;
7153 unsigned long reader_size;
7154 unsigned long flags;
7155
7156 cpu_buffer = rb_get_mapped_buffer(buffer, cpu);
7157 if (IS_ERR(cpu_buffer))
7158 return (int)PTR_ERR(cpu_buffer);
7159
7160 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
7161
7162consume:
7163 if (rb_per_cpu_empty(cpu_buffer))
7164 goto out;
7165
7166 reader_size = rb_page_size(cpu_buffer->reader_page);
7167
7168 /*
7169 * There are data to be read on the current reader page, we can
7170 * return to the caller. But before that, we assume the latter will read
7171 * everything. Let's update the kernel reader accordingly.
7172 */
7173 if (cpu_buffer->reader_page->read < reader_size) {
7174 while (cpu_buffer->reader_page->read < reader_size)
7175 rb_advance_reader(cpu_buffer);
7176 goto out;
7177 }
7178
7179 reader = rb_get_reader_page(cpu_buffer);
7180 if (WARN_ON(!reader))
7181 goto out;
7182
7183 /* Check if any events were dropped */
7184 missed_events = cpu_buffer->lost_events;
7185
7186 if (cpu_buffer->reader_page != cpu_buffer->commit_page) {
7187 if (missed_events) {
7188 struct buffer_data_page *bpage = reader->page;
7189 unsigned int commit;
7190 /*
7191 * Use the real_end for the data size,
7192 * This gives us a chance to store the lost events
7193 * on the page.
7194 */
7195 if (reader->real_end)
7196 local_set(&bpage->commit, reader->real_end);
7197 /*
7198 * If there is room at the end of the page to save the
7199 * missed events, then record it there.
7200 */
7201 commit = rb_page_size(reader);
7202 if (buffer->subbuf_size - commit >= sizeof(missed_events)) {
7203 memcpy(&bpage->data[commit], &missed_events,
7204 sizeof(missed_events));
7205 local_add(RB_MISSED_STORED, &bpage->commit);
7206 }
7207 local_add(RB_MISSED_EVENTS, &bpage->commit);
7208 }
7209 } else {
7210 /*
7211 * There really shouldn't be any missed events if the commit
7212 * is on the reader page.
7213 */
7214 WARN_ON_ONCE(missed_events);
7215 }
7216
7217 cpu_buffer->lost_events = 0;
7218
7219 goto consume;
7220
7221out:
7222 /* Some archs do not have data cache coherency between kernel and user-space */
7223 flush_dcache_folio(virt_to_folio(cpu_buffer->reader_page->page));
7224
7225 rb_update_meta_page(cpu_buffer);
7226
7227 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
7228 rb_put_mapped_buffer(cpu_buffer);
7229
7230 return 0;
7231}
7232
7233/*
7234 * We only allocate new buffers, never free them if the CPU goes down.
7235 * If we were to free the buffer, then the user would lose any trace that was in
7236 * the buffer.
7237 */
7238int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
7239{
7240 struct trace_buffer *buffer;
7241 long nr_pages_same;
7242 int cpu_i;
7243 unsigned long nr_pages;
7244
7245 buffer = container_of(node, struct trace_buffer, node);
7246 if (cpumask_test_cpu(cpu, buffer->cpumask))
7247 return 0;
7248
7249 nr_pages = 0;
7250 nr_pages_same = 1;
7251 /* check if all cpu sizes are same */
7252 for_each_buffer_cpu(buffer, cpu_i) {
7253 /* fill in the size from first enabled cpu */
7254 if (nr_pages == 0)
7255 nr_pages = buffer->buffers[cpu_i]->nr_pages;
7256 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
7257 nr_pages_same = 0;
7258 break;
7259 }
7260 }
7261 /* allocate minimum pages, user can later expand it */
7262 if (!nr_pages_same)
7263 nr_pages = 2;
7264 buffer->buffers[cpu] =
7265 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
7266 if (!buffer->buffers[cpu]) {
7267 WARN(1, "failed to allocate ring buffer on CPU %u\n",
7268 cpu);
7269 return -ENOMEM;
7270 }
7271 smp_wmb();
7272 cpumask_set_cpu(cpu, buffer->cpumask);
7273 return 0;
7274}
7275
7276#ifdef CONFIG_RING_BUFFER_STARTUP_TEST
7277/*
7278 * This is a basic integrity check of the ring buffer.
7279 * Late in the boot cycle this test will run when configured in.
7280 * It will kick off a thread per CPU that will go into a loop
7281 * writing to the per cpu ring buffer various sizes of data.
7282 * Some of the data will be large items, some small.
7283 *
7284 * Another thread is created that goes into a spin, sending out
7285 * IPIs to the other CPUs to also write into the ring buffer.
7286 * this is to test the nesting ability of the buffer.
7287 *
7288 * Basic stats are recorded and reported. If something in the
7289 * ring buffer should happen that's not expected, a big warning
7290 * is displayed and all ring buffers are disabled.
7291 */
7292static struct task_struct *rb_threads[NR_CPUS] __initdata;
7293
7294struct rb_test_data {
7295 struct trace_buffer *buffer;
7296 unsigned long events;
7297 unsigned long bytes_written;
7298 unsigned long bytes_alloc;
7299 unsigned long bytes_dropped;
7300 unsigned long events_nested;
7301 unsigned long bytes_written_nested;
7302 unsigned long bytes_alloc_nested;
7303 unsigned long bytes_dropped_nested;
7304 int min_size_nested;
7305 int max_size_nested;
7306 int max_size;
7307 int min_size;
7308 int cpu;
7309 int cnt;
7310};
7311
7312static struct rb_test_data rb_data[NR_CPUS] __initdata;
7313
7314/* 1 meg per cpu */
7315#define RB_TEST_BUFFER_SIZE 1048576
7316
7317static char rb_string[] __initdata =
7318 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
7319 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
7320 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
7321
7322static bool rb_test_started __initdata;
7323
7324struct rb_item {
7325 int size;
7326 char str[];
7327};
7328
7329static __init int rb_write_something(struct rb_test_data *data, bool nested)
7330{
7331 struct ring_buffer_event *event;
7332 struct rb_item *item;
7333 bool started;
7334 int event_len;
7335 int size;
7336 int len;
7337 int cnt;
7338
7339 /* Have nested writes different that what is written */
7340 cnt = data->cnt + (nested ? 27 : 0);
7341
7342 /* Multiply cnt by ~e, to make some unique increment */
7343 size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
7344
7345 len = size + sizeof(struct rb_item);
7346
7347 started = rb_test_started;
7348 /* read rb_test_started before checking buffer enabled */
7349 smp_rmb();
7350
7351 event = ring_buffer_lock_reserve(data->buffer, len);
7352 if (!event) {
7353 /* Ignore dropped events before test starts. */
7354 if (started) {
7355 if (nested)
7356 data->bytes_dropped += len;
7357 else
7358 data->bytes_dropped_nested += len;
7359 }
7360 return len;
7361 }
7362
7363 event_len = ring_buffer_event_length(event);
7364
7365 if (RB_WARN_ON(data->buffer, event_len < len))
7366 goto out;
7367
7368 item = ring_buffer_event_data(event);
7369 item->size = size;
7370 memcpy(item->str, rb_string, size);
7371
7372 if (nested) {
7373 data->bytes_alloc_nested += event_len;
7374 data->bytes_written_nested += len;
7375 data->events_nested++;
7376 if (!data->min_size_nested || len < data->min_size_nested)
7377 data->min_size_nested = len;
7378 if (len > data->max_size_nested)
7379 data->max_size_nested = len;
7380 } else {
7381 data->bytes_alloc += event_len;
7382 data->bytes_written += len;
7383 data->events++;
7384 if (!data->min_size || len < data->min_size)
7385 data->max_size = len;
7386 if (len > data->max_size)
7387 data->max_size = len;
7388 }
7389
7390 out:
7391 ring_buffer_unlock_commit(data->buffer);
7392
7393 return 0;
7394}
7395
7396static __init int rb_test(void *arg)
7397{
7398 struct rb_test_data *data = arg;
7399
7400 while (!kthread_should_stop()) {
7401 rb_write_something(data, false);
7402 data->cnt++;
7403
7404 set_current_state(TASK_INTERRUPTIBLE);
7405 /* Now sleep between a min of 100-300us and a max of 1ms */
7406 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
7407 }
7408
7409 return 0;
7410}
7411
7412static __init void rb_ipi(void *ignore)
7413{
7414 struct rb_test_data *data;
7415 int cpu = smp_processor_id();
7416
7417 data = &rb_data[cpu];
7418 rb_write_something(data, true);
7419}
7420
7421static __init int rb_hammer_test(void *arg)
7422{
7423 while (!kthread_should_stop()) {
7424
7425 /* Send an IPI to all cpus to write data! */
7426 smp_call_function(rb_ipi, NULL, 1);
7427 /* No sleep, but for non preempt, let others run */
7428 schedule();
7429 }
7430
7431 return 0;
7432}
7433
7434static __init int test_ringbuffer(void)
7435{
7436 struct task_struct *rb_hammer;
7437 struct trace_buffer *buffer;
7438 int cpu;
7439 int ret = 0;
7440
7441 if (security_locked_down(LOCKDOWN_TRACEFS)) {
7442 pr_warn("Lockdown is enabled, skipping ring buffer tests\n");
7443 return 0;
7444 }
7445
7446 pr_info("Running ring buffer tests...\n");
7447
7448 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
7449 if (WARN_ON(!buffer))
7450 return 0;
7451
7452 /* Disable buffer so that threads can't write to it yet */
7453 ring_buffer_record_off(buffer);
7454
7455 for_each_online_cpu(cpu) {
7456 rb_data[cpu].buffer = buffer;
7457 rb_data[cpu].cpu = cpu;
7458 rb_data[cpu].cnt = cpu;
7459 rb_threads[cpu] = kthread_run_on_cpu(rb_test, &rb_data[cpu],
7460 cpu, "rbtester/%u");
7461 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
7462 pr_cont("FAILED\n");
7463 ret = PTR_ERR(rb_threads[cpu]);
7464 goto out_free;
7465 }
7466 }
7467
7468 /* Now create the rb hammer! */
7469 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
7470 if (WARN_ON(IS_ERR(rb_hammer))) {
7471 pr_cont("FAILED\n");
7472 ret = PTR_ERR(rb_hammer);
7473 goto out_free;
7474 }
7475
7476 ring_buffer_record_on(buffer);
7477 /*
7478 * Show buffer is enabled before setting rb_test_started.
7479 * Yes there's a small race window where events could be
7480 * dropped and the thread wont catch it. But when a ring
7481 * buffer gets enabled, there will always be some kind of
7482 * delay before other CPUs see it. Thus, we don't care about
7483 * those dropped events. We care about events dropped after
7484 * the threads see that the buffer is active.
7485 */
7486 smp_wmb();
7487 rb_test_started = true;
7488
7489 set_current_state(TASK_INTERRUPTIBLE);
7490 /* Just run for 10 seconds */;
7491 schedule_timeout(10 * HZ);
7492
7493 kthread_stop(rb_hammer);
7494
7495 out_free:
7496 for_each_online_cpu(cpu) {
7497 if (!rb_threads[cpu])
7498 break;
7499 kthread_stop(rb_threads[cpu]);
7500 }
7501 if (ret) {
7502 ring_buffer_free(buffer);
7503 return ret;
7504 }
7505
7506 /* Report! */
7507 pr_info("finished\n");
7508 for_each_online_cpu(cpu) {
7509 struct ring_buffer_event *event;
7510 struct rb_test_data *data = &rb_data[cpu];
7511 struct rb_item *item;
7512 unsigned long total_events;
7513 unsigned long total_dropped;
7514 unsigned long total_written;
7515 unsigned long total_alloc;
7516 unsigned long total_read = 0;
7517 unsigned long total_size = 0;
7518 unsigned long total_len = 0;
7519 unsigned long total_lost = 0;
7520 unsigned long lost;
7521 int big_event_size;
7522 int small_event_size;
7523
7524 ret = -1;
7525
7526 total_events = data->events + data->events_nested;
7527 total_written = data->bytes_written + data->bytes_written_nested;
7528 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
7529 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
7530
7531 big_event_size = data->max_size + data->max_size_nested;
7532 small_event_size = data->min_size + data->min_size_nested;
7533
7534 pr_info("CPU %d:\n", cpu);
7535 pr_info(" events: %ld\n", total_events);
7536 pr_info(" dropped bytes: %ld\n", total_dropped);
7537 pr_info(" alloced bytes: %ld\n", total_alloc);
7538 pr_info(" written bytes: %ld\n", total_written);
7539 pr_info(" biggest event: %d\n", big_event_size);
7540 pr_info(" smallest event: %d\n", small_event_size);
7541
7542 if (RB_WARN_ON(buffer, total_dropped))
7543 break;
7544
7545 ret = 0;
7546
7547 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
7548 total_lost += lost;
7549 item = ring_buffer_event_data(event);
7550 total_len += ring_buffer_event_length(event);
7551 total_size += item->size + sizeof(struct rb_item);
7552 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
7553 pr_info("FAILED!\n");
7554 pr_info("buffer had: %.*s\n", item->size, item->str);
7555 pr_info("expected: %.*s\n", item->size, rb_string);
7556 RB_WARN_ON(buffer, 1);
7557 ret = -1;
7558 break;
7559 }
7560 total_read++;
7561 }
7562 if (ret)
7563 break;
7564
7565 ret = -1;
7566
7567 pr_info(" read events: %ld\n", total_read);
7568 pr_info(" lost events: %ld\n", total_lost);
7569 pr_info(" total events: %ld\n", total_lost + total_read);
7570 pr_info(" recorded len bytes: %ld\n", total_len);
7571 pr_info(" recorded size bytes: %ld\n", total_size);
7572 if (total_lost) {
7573 pr_info(" With dropped events, record len and size may not match\n"
7574 " alloced and written from above\n");
7575 } else {
7576 if (RB_WARN_ON(buffer, total_len != total_alloc ||
7577 total_size != total_written))
7578 break;
7579 }
7580 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
7581 break;
7582
7583 ret = 0;
7584 }
7585 if (!ret)
7586 pr_info("Ring buffer PASSED!\n");
7587
7588 ring_buffer_free(buffer);
7589 return 0;
7590}
7591
7592late_initcall(test_ringbuffer);
7593#endif /* CONFIG_RING_BUFFER_STARTUP_TEST */