Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Generic ring buffer
4 *
5 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 */
7#include <linux/trace_recursion.h>
8#include <linux/trace_events.h>
9#include <linux/ring_buffer.h>
10#include <linux/trace_clock.h>
11#include <linux/sched/clock.h>
12#include <linux/trace_seq.h>
13#include <linux/spinlock.h>
14#include <linux/irq_work.h>
15#include <linux/security.h>
16#include <linux/uaccess.h>
17#include <linux/hardirq.h>
18#include <linux/kthread.h> /* for self test */
19#include <linux/module.h>
20#include <linux/percpu.h>
21#include <linux/mutex.h>
22#include <linux/delay.h>
23#include <linux/slab.h>
24#include <linux/init.h>
25#include <linux/hash.h>
26#include <linux/list.h>
27#include <linux/cpu.h>
28#include <linux/oom.h>
29
30#include <asm/local64.h>
31#include <asm/local.h>
32
33/*
34 * The "absolute" timestamp in the buffer is only 59 bits.
35 * If a clock has the 5 MSBs set, it needs to be saved and
36 * reinserted.
37 */
38#define TS_MSB (0xf8ULL << 56)
39#define ABS_TS_MASK (~TS_MSB)
40
41static void update_pages_handler(struct work_struct *work);
42
43/*
44 * The ring buffer header is special. We must manually up keep it.
45 */
46int ring_buffer_print_entry_header(struct trace_seq *s)
47{
48 trace_seq_puts(s, "# compressed entry header\n");
49 trace_seq_puts(s, "\ttype_len : 5 bits\n");
50 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
51 trace_seq_puts(s, "\tarray : 32 bits\n");
52 trace_seq_putc(s, '\n');
53 trace_seq_printf(s, "\tpadding : type == %d\n",
54 RINGBUF_TYPE_PADDING);
55 trace_seq_printf(s, "\ttime_extend : type == %d\n",
56 RINGBUF_TYPE_TIME_EXTEND);
57 trace_seq_printf(s, "\ttime_stamp : type == %d\n",
58 RINGBUF_TYPE_TIME_STAMP);
59 trace_seq_printf(s, "\tdata max type_len == %d\n",
60 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
61
62 return !trace_seq_has_overflowed(s);
63}
64
65/*
66 * The ring buffer is made up of a list of pages. A separate list of pages is
67 * allocated for each CPU. A writer may only write to a buffer that is
68 * associated with the CPU it is currently executing on. A reader may read
69 * from any per cpu buffer.
70 *
71 * The reader is special. For each per cpu buffer, the reader has its own
72 * reader page. When a reader has read the entire reader page, this reader
73 * page is swapped with another page in the ring buffer.
74 *
75 * Now, as long as the writer is off the reader page, the reader can do what
76 * ever it wants with that page. The writer will never write to that page
77 * again (as long as it is out of the ring buffer).
78 *
79 * Here's some silly ASCII art.
80 *
81 * +------+
82 * |reader| RING BUFFER
83 * |page |
84 * +------+ +---+ +---+ +---+
85 * | |-->| |-->| |
86 * +---+ +---+ +---+
87 * ^ |
88 * | |
89 * +---------------+
90 *
91 *
92 * +------+
93 * |reader| RING BUFFER
94 * |page |------------------v
95 * +------+ +---+ +---+ +---+
96 * | |-->| |-->| |
97 * +---+ +---+ +---+
98 * ^ |
99 * | |
100 * +---------------+
101 *
102 *
103 * +------+
104 * |reader| RING BUFFER
105 * |page |------------------v
106 * +------+ +---+ +---+ +---+
107 * ^ | |-->| |-->| |
108 * | +---+ +---+ +---+
109 * | |
110 * | |
111 * +------------------------------+
112 *
113 *
114 * +------+
115 * |buffer| RING BUFFER
116 * |page |------------------v
117 * +------+ +---+ +---+ +---+
118 * ^ | | | |-->| |
119 * | New +---+ +---+ +---+
120 * | Reader------^ |
121 * | page |
122 * +------------------------------+
123 *
124 *
125 * After we make this swap, the reader can hand this page off to the splice
126 * code and be done with it. It can even allocate a new page if it needs to
127 * and swap that into the ring buffer.
128 *
129 * We will be using cmpxchg soon to make all this lockless.
130 *
131 */
132
133/* Used for individual buffers (after the counter) */
134#define RB_BUFFER_OFF (1 << 20)
135
136#define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
137
138#define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
139#define RB_ALIGNMENT 4U
140#define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
141#define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
142
143#ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
144# define RB_FORCE_8BYTE_ALIGNMENT 0
145# define RB_ARCH_ALIGNMENT RB_ALIGNMENT
146#else
147# define RB_FORCE_8BYTE_ALIGNMENT 1
148# define RB_ARCH_ALIGNMENT 8U
149#endif
150
151#define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
152
153/* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
154#define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
155
156enum {
157 RB_LEN_TIME_EXTEND = 8,
158 RB_LEN_TIME_STAMP = 8,
159};
160
161#define skip_time_extend(event) \
162 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
163
164#define extended_time(event) \
165 (event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
166
167static inline bool rb_null_event(struct ring_buffer_event *event)
168{
169 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
170}
171
172static void rb_event_set_padding(struct ring_buffer_event *event)
173{
174 /* padding has a NULL time_delta */
175 event->type_len = RINGBUF_TYPE_PADDING;
176 event->time_delta = 0;
177}
178
179static unsigned
180rb_event_data_length(struct ring_buffer_event *event)
181{
182 unsigned length;
183
184 if (event->type_len)
185 length = event->type_len * RB_ALIGNMENT;
186 else
187 length = event->array[0];
188 return length + RB_EVNT_HDR_SIZE;
189}
190
191/*
192 * Return the length of the given event. Will return
193 * the length of the time extend if the event is a
194 * time extend.
195 */
196static inline unsigned
197rb_event_length(struct ring_buffer_event *event)
198{
199 switch (event->type_len) {
200 case RINGBUF_TYPE_PADDING:
201 if (rb_null_event(event))
202 /* undefined */
203 return -1;
204 return event->array[0] + RB_EVNT_HDR_SIZE;
205
206 case RINGBUF_TYPE_TIME_EXTEND:
207 return RB_LEN_TIME_EXTEND;
208
209 case RINGBUF_TYPE_TIME_STAMP:
210 return RB_LEN_TIME_STAMP;
211
212 case RINGBUF_TYPE_DATA:
213 return rb_event_data_length(event);
214 default:
215 WARN_ON_ONCE(1);
216 }
217 /* not hit */
218 return 0;
219}
220
221/*
222 * Return total length of time extend and data,
223 * or just the event length for all other events.
224 */
225static inline unsigned
226rb_event_ts_length(struct ring_buffer_event *event)
227{
228 unsigned len = 0;
229
230 if (extended_time(event)) {
231 /* time extends include the data event after it */
232 len = RB_LEN_TIME_EXTEND;
233 event = skip_time_extend(event);
234 }
235 return len + rb_event_length(event);
236}
237
238/**
239 * ring_buffer_event_length - return the length of the event
240 * @event: the event to get the length of
241 *
242 * Returns the size of the data load of a data event.
243 * If the event is something other than a data event, it
244 * returns the size of the event itself. With the exception
245 * of a TIME EXTEND, where it still returns the size of the
246 * data load of the data event after it.
247 */
248unsigned ring_buffer_event_length(struct ring_buffer_event *event)
249{
250 unsigned length;
251
252 if (extended_time(event))
253 event = skip_time_extend(event);
254
255 length = rb_event_length(event);
256 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
257 return length;
258 length -= RB_EVNT_HDR_SIZE;
259 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
260 length -= sizeof(event->array[0]);
261 return length;
262}
263EXPORT_SYMBOL_GPL(ring_buffer_event_length);
264
265/* inline for ring buffer fast paths */
266static __always_inline void *
267rb_event_data(struct ring_buffer_event *event)
268{
269 if (extended_time(event))
270 event = skip_time_extend(event);
271 WARN_ON_ONCE(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
272 /* If length is in len field, then array[0] has the data */
273 if (event->type_len)
274 return (void *)&event->array[0];
275 /* Otherwise length is in array[0] and array[1] has the data */
276 return (void *)&event->array[1];
277}
278
279/**
280 * ring_buffer_event_data - return the data of the event
281 * @event: the event to get the data from
282 */
283void *ring_buffer_event_data(struct ring_buffer_event *event)
284{
285 return rb_event_data(event);
286}
287EXPORT_SYMBOL_GPL(ring_buffer_event_data);
288
289#define for_each_buffer_cpu(buffer, cpu) \
290 for_each_cpu(cpu, buffer->cpumask)
291
292#define for_each_online_buffer_cpu(buffer, cpu) \
293 for_each_cpu_and(cpu, buffer->cpumask, cpu_online_mask)
294
295#define TS_SHIFT 27
296#define TS_MASK ((1ULL << TS_SHIFT) - 1)
297#define TS_DELTA_TEST (~TS_MASK)
298
299static u64 rb_event_time_stamp(struct ring_buffer_event *event)
300{
301 u64 ts;
302
303 ts = event->array[0];
304 ts <<= TS_SHIFT;
305 ts += event->time_delta;
306
307 return ts;
308}
309
310/* Flag when events were overwritten */
311#define RB_MISSED_EVENTS (1 << 31)
312/* Missed count stored at end */
313#define RB_MISSED_STORED (1 << 30)
314
315struct buffer_data_page {
316 u64 time_stamp; /* page time stamp */
317 local_t commit; /* write committed index */
318 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
319};
320
321struct buffer_data_read_page {
322 unsigned order; /* order of the page */
323 struct buffer_data_page *data; /* actual data, stored in this page */
324};
325
326/*
327 * Note, the buffer_page list must be first. The buffer pages
328 * are allocated in cache lines, which means that each buffer
329 * page will be at the beginning of a cache line, and thus
330 * the least significant bits will be zero. We use this to
331 * add flags in the list struct pointers, to make the ring buffer
332 * lockless.
333 */
334struct buffer_page {
335 struct list_head list; /* list of buffer pages */
336 local_t write; /* index for next write */
337 unsigned read; /* index for next read */
338 local_t entries; /* entries on this page */
339 unsigned long real_end; /* real end of data */
340 unsigned order; /* order of the page */
341 struct buffer_data_page *page; /* Actual data page */
342};
343
344/*
345 * The buffer page counters, write and entries, must be reset
346 * atomically when crossing page boundaries. To synchronize this
347 * update, two counters are inserted into the number. One is
348 * the actual counter for the write position or count on the page.
349 *
350 * The other is a counter of updaters. Before an update happens
351 * the update partition of the counter is incremented. This will
352 * allow the updater to update the counter atomically.
353 *
354 * The counter is 20 bits, and the state data is 12.
355 */
356#define RB_WRITE_MASK 0xfffff
357#define RB_WRITE_INTCNT (1 << 20)
358
359static void rb_init_page(struct buffer_data_page *bpage)
360{
361 local_set(&bpage->commit, 0);
362}
363
364static __always_inline unsigned int rb_page_commit(struct buffer_page *bpage)
365{
366 return local_read(&bpage->page->commit);
367}
368
369static void free_buffer_page(struct buffer_page *bpage)
370{
371 free_pages((unsigned long)bpage->page, bpage->order);
372 kfree(bpage);
373}
374
375/*
376 * We need to fit the time_stamp delta into 27 bits.
377 */
378static inline bool test_time_stamp(u64 delta)
379{
380 return !!(delta & TS_DELTA_TEST);
381}
382
383struct rb_irq_work {
384 struct irq_work work;
385 wait_queue_head_t waiters;
386 wait_queue_head_t full_waiters;
387 atomic_t seq;
388 bool waiters_pending;
389 bool full_waiters_pending;
390 bool wakeup_full;
391};
392
393/*
394 * Structure to hold event state and handle nested events.
395 */
396struct rb_event_info {
397 u64 ts;
398 u64 delta;
399 u64 before;
400 u64 after;
401 unsigned long length;
402 struct buffer_page *tail_page;
403 int add_timestamp;
404};
405
406/*
407 * Used for the add_timestamp
408 * NONE
409 * EXTEND - wants a time extend
410 * ABSOLUTE - the buffer requests all events to have absolute time stamps
411 * FORCE - force a full time stamp.
412 */
413enum {
414 RB_ADD_STAMP_NONE = 0,
415 RB_ADD_STAMP_EXTEND = BIT(1),
416 RB_ADD_STAMP_ABSOLUTE = BIT(2),
417 RB_ADD_STAMP_FORCE = BIT(3)
418};
419/*
420 * Used for which event context the event is in.
421 * TRANSITION = 0
422 * NMI = 1
423 * IRQ = 2
424 * SOFTIRQ = 3
425 * NORMAL = 4
426 *
427 * See trace_recursive_lock() comment below for more details.
428 */
429enum {
430 RB_CTX_TRANSITION,
431 RB_CTX_NMI,
432 RB_CTX_IRQ,
433 RB_CTX_SOFTIRQ,
434 RB_CTX_NORMAL,
435 RB_CTX_MAX
436};
437
438struct rb_time_struct {
439 local64_t time;
440};
441typedef struct rb_time_struct rb_time_t;
442
443#define MAX_NEST 5
444
445/*
446 * head_page == tail_page && head == tail then buffer is empty.
447 */
448struct ring_buffer_per_cpu {
449 int cpu;
450 atomic_t record_disabled;
451 atomic_t resize_disabled;
452 struct trace_buffer *buffer;
453 raw_spinlock_t reader_lock; /* serialize readers */
454 arch_spinlock_t lock;
455 struct lock_class_key lock_key;
456 struct buffer_data_page *free_page;
457 unsigned long nr_pages;
458 unsigned int current_context;
459 struct list_head *pages;
460 struct buffer_page *head_page; /* read from head */
461 struct buffer_page *tail_page; /* write to tail */
462 struct buffer_page *commit_page; /* committed pages */
463 struct buffer_page *reader_page;
464 unsigned long lost_events;
465 unsigned long last_overrun;
466 unsigned long nest;
467 local_t entries_bytes;
468 local_t entries;
469 local_t overrun;
470 local_t commit_overrun;
471 local_t dropped_events;
472 local_t committing;
473 local_t commits;
474 local_t pages_touched;
475 local_t pages_lost;
476 local_t pages_read;
477 long last_pages_touch;
478 size_t shortest_full;
479 unsigned long read;
480 unsigned long read_bytes;
481 rb_time_t write_stamp;
482 rb_time_t before_stamp;
483 u64 event_stamp[MAX_NEST];
484 u64 read_stamp;
485 /* pages removed since last reset */
486 unsigned long pages_removed;
487 /* ring buffer pages to update, > 0 to add, < 0 to remove */
488 long nr_pages_to_update;
489 struct list_head new_pages; /* new pages to add */
490 struct work_struct update_pages_work;
491 struct completion update_done;
492
493 struct rb_irq_work irq_work;
494};
495
496struct trace_buffer {
497 unsigned flags;
498 int cpus;
499 atomic_t record_disabled;
500 atomic_t resizing;
501 cpumask_var_t cpumask;
502
503 struct lock_class_key *reader_lock_key;
504
505 struct mutex mutex;
506
507 struct ring_buffer_per_cpu **buffers;
508
509 struct hlist_node node;
510 u64 (*clock)(void);
511
512 struct rb_irq_work irq_work;
513 bool time_stamp_abs;
514
515 unsigned int subbuf_size;
516 unsigned int subbuf_order;
517 unsigned int max_data_size;
518};
519
520struct ring_buffer_iter {
521 struct ring_buffer_per_cpu *cpu_buffer;
522 unsigned long head;
523 unsigned long next_event;
524 struct buffer_page *head_page;
525 struct buffer_page *cache_reader_page;
526 unsigned long cache_read;
527 unsigned long cache_pages_removed;
528 u64 read_stamp;
529 u64 page_stamp;
530 struct ring_buffer_event *event;
531 size_t event_size;
532 int missed_events;
533};
534
535int ring_buffer_print_page_header(struct trace_buffer *buffer, struct trace_seq *s)
536{
537 struct buffer_data_page field;
538
539 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
540 "offset:0;\tsize:%u;\tsigned:%u;\n",
541 (unsigned int)sizeof(field.time_stamp),
542 (unsigned int)is_signed_type(u64));
543
544 trace_seq_printf(s, "\tfield: local_t commit;\t"
545 "offset:%u;\tsize:%u;\tsigned:%u;\n",
546 (unsigned int)offsetof(typeof(field), commit),
547 (unsigned int)sizeof(field.commit),
548 (unsigned int)is_signed_type(long));
549
550 trace_seq_printf(s, "\tfield: int overwrite;\t"
551 "offset:%u;\tsize:%u;\tsigned:%u;\n",
552 (unsigned int)offsetof(typeof(field), commit),
553 1,
554 (unsigned int)is_signed_type(long));
555
556 trace_seq_printf(s, "\tfield: char data;\t"
557 "offset:%u;\tsize:%u;\tsigned:%u;\n",
558 (unsigned int)offsetof(typeof(field), data),
559 (unsigned int)buffer->subbuf_size,
560 (unsigned int)is_signed_type(char));
561
562 return !trace_seq_has_overflowed(s);
563}
564
565static inline void rb_time_read(rb_time_t *t, u64 *ret)
566{
567 *ret = local64_read(&t->time);
568}
569static void rb_time_set(rb_time_t *t, u64 val)
570{
571 local64_set(&t->time, val);
572}
573
574/*
575 * Enable this to make sure that the event passed to
576 * ring_buffer_event_time_stamp() is not committed and also
577 * is on the buffer that it passed in.
578 */
579//#define RB_VERIFY_EVENT
580#ifdef RB_VERIFY_EVENT
581static struct list_head *rb_list_head(struct list_head *list);
582static void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
583 void *event)
584{
585 struct buffer_page *page = cpu_buffer->commit_page;
586 struct buffer_page *tail_page = READ_ONCE(cpu_buffer->tail_page);
587 struct list_head *next;
588 long commit, write;
589 unsigned long addr = (unsigned long)event;
590 bool done = false;
591 int stop = 0;
592
593 /* Make sure the event exists and is not committed yet */
594 do {
595 if (page == tail_page || WARN_ON_ONCE(stop++ > 100))
596 done = true;
597 commit = local_read(&page->page->commit);
598 write = local_read(&page->write);
599 if (addr >= (unsigned long)&page->page->data[commit] &&
600 addr < (unsigned long)&page->page->data[write])
601 return;
602
603 next = rb_list_head(page->list.next);
604 page = list_entry(next, struct buffer_page, list);
605 } while (!done);
606 WARN_ON_ONCE(1);
607}
608#else
609static inline void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
610 void *event)
611{
612}
613#endif
614
615/*
616 * The absolute time stamp drops the 5 MSBs and some clocks may
617 * require them. The rb_fix_abs_ts() will take a previous full
618 * time stamp, and add the 5 MSB of that time stamp on to the
619 * saved absolute time stamp. Then they are compared in case of
620 * the unlikely event that the latest time stamp incremented
621 * the 5 MSB.
622 */
623static inline u64 rb_fix_abs_ts(u64 abs, u64 save_ts)
624{
625 if (save_ts & TS_MSB) {
626 abs |= save_ts & TS_MSB;
627 /* Check for overflow */
628 if (unlikely(abs < save_ts))
629 abs += 1ULL << 59;
630 }
631 return abs;
632}
633
634static inline u64 rb_time_stamp(struct trace_buffer *buffer);
635
636/**
637 * ring_buffer_event_time_stamp - return the event's current time stamp
638 * @buffer: The buffer that the event is on
639 * @event: the event to get the time stamp of
640 *
641 * Note, this must be called after @event is reserved, and before it is
642 * committed to the ring buffer. And must be called from the same
643 * context where the event was reserved (normal, softirq, irq, etc).
644 *
645 * Returns the time stamp associated with the current event.
646 * If the event has an extended time stamp, then that is used as
647 * the time stamp to return.
648 * In the highly unlikely case that the event was nested more than
649 * the max nesting, then the write_stamp of the buffer is returned,
650 * otherwise current time is returned, but that really neither of
651 * the last two cases should ever happen.
652 */
653u64 ring_buffer_event_time_stamp(struct trace_buffer *buffer,
654 struct ring_buffer_event *event)
655{
656 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[smp_processor_id()];
657 unsigned int nest;
658 u64 ts;
659
660 /* If the event includes an absolute time, then just use that */
661 if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
662 ts = rb_event_time_stamp(event);
663 return rb_fix_abs_ts(ts, cpu_buffer->tail_page->page->time_stamp);
664 }
665
666 nest = local_read(&cpu_buffer->committing);
667 verify_event(cpu_buffer, event);
668 if (WARN_ON_ONCE(!nest))
669 goto fail;
670
671 /* Read the current saved nesting level time stamp */
672 if (likely(--nest < MAX_NEST))
673 return cpu_buffer->event_stamp[nest];
674
675 /* Shouldn't happen, warn if it does */
676 WARN_ONCE(1, "nest (%d) greater than max", nest);
677
678 fail:
679 rb_time_read(&cpu_buffer->write_stamp, &ts);
680
681 return ts;
682}
683
684/**
685 * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
686 * @buffer: The ring_buffer to get the number of pages from
687 * @cpu: The cpu of the ring_buffer to get the number of pages from
688 *
689 * Returns the number of pages used by a per_cpu buffer of the ring buffer.
690 */
691size_t ring_buffer_nr_pages(struct trace_buffer *buffer, int cpu)
692{
693 return buffer->buffers[cpu]->nr_pages;
694}
695
696/**
697 * ring_buffer_nr_dirty_pages - get the number of used pages in the ring buffer
698 * @buffer: The ring_buffer to get the number of pages from
699 * @cpu: The cpu of the ring_buffer to get the number of pages from
700 *
701 * Returns the number of pages that have content in the ring buffer.
702 */
703size_t ring_buffer_nr_dirty_pages(struct trace_buffer *buffer, int cpu)
704{
705 size_t read;
706 size_t lost;
707 size_t cnt;
708
709 read = local_read(&buffer->buffers[cpu]->pages_read);
710 lost = local_read(&buffer->buffers[cpu]->pages_lost);
711 cnt = local_read(&buffer->buffers[cpu]->pages_touched);
712
713 if (WARN_ON_ONCE(cnt < lost))
714 return 0;
715
716 cnt -= lost;
717
718 /* The reader can read an empty page, but not more than that */
719 if (cnt < read) {
720 WARN_ON_ONCE(read > cnt + 1);
721 return 0;
722 }
723
724 return cnt - read;
725}
726
727static __always_inline bool full_hit(struct trace_buffer *buffer, int cpu, int full)
728{
729 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
730 size_t nr_pages;
731 size_t dirty;
732
733 nr_pages = cpu_buffer->nr_pages;
734 if (!nr_pages || !full)
735 return true;
736
737 /*
738 * Add one as dirty will never equal nr_pages, as the sub-buffer
739 * that the writer is on is not counted as dirty.
740 * This is needed if "buffer_percent" is set to 100.
741 */
742 dirty = ring_buffer_nr_dirty_pages(buffer, cpu) + 1;
743
744 return (dirty * 100) >= (full * nr_pages);
745}
746
747/*
748 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
749 *
750 * Schedules a delayed work to wake up any task that is blocked on the
751 * ring buffer waiters queue.
752 */
753static void rb_wake_up_waiters(struct irq_work *work)
754{
755 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
756
757 /* For waiters waiting for the first wake up */
758 (void)atomic_fetch_inc_release(&rbwork->seq);
759
760 wake_up_all(&rbwork->waiters);
761 if (rbwork->full_waiters_pending || rbwork->wakeup_full) {
762 /* Only cpu_buffer sets the above flags */
763 struct ring_buffer_per_cpu *cpu_buffer =
764 container_of(rbwork, struct ring_buffer_per_cpu, irq_work);
765
766 /* Called from interrupt context */
767 raw_spin_lock(&cpu_buffer->reader_lock);
768 rbwork->wakeup_full = false;
769 rbwork->full_waiters_pending = false;
770
771 /* Waking up all waiters, they will reset the shortest full */
772 cpu_buffer->shortest_full = 0;
773 raw_spin_unlock(&cpu_buffer->reader_lock);
774
775 wake_up_all(&rbwork->full_waiters);
776 }
777}
778
779/**
780 * ring_buffer_wake_waiters - wake up any waiters on this ring buffer
781 * @buffer: The ring buffer to wake waiters on
782 * @cpu: The CPU buffer to wake waiters on
783 *
784 * In the case of a file that represents a ring buffer is closing,
785 * it is prudent to wake up any waiters that are on this.
786 */
787void ring_buffer_wake_waiters(struct trace_buffer *buffer, int cpu)
788{
789 struct ring_buffer_per_cpu *cpu_buffer;
790 struct rb_irq_work *rbwork;
791
792 if (!buffer)
793 return;
794
795 if (cpu == RING_BUFFER_ALL_CPUS) {
796
797 /* Wake up individual ones too. One level recursion */
798 for_each_buffer_cpu(buffer, cpu)
799 ring_buffer_wake_waiters(buffer, cpu);
800
801 rbwork = &buffer->irq_work;
802 } else {
803 if (WARN_ON_ONCE(!buffer->buffers))
804 return;
805 if (WARN_ON_ONCE(cpu >= nr_cpu_ids))
806 return;
807
808 cpu_buffer = buffer->buffers[cpu];
809 /* The CPU buffer may not have been initialized yet */
810 if (!cpu_buffer)
811 return;
812 rbwork = &cpu_buffer->irq_work;
813 }
814
815 /* This can be called in any context */
816 irq_work_queue(&rbwork->work);
817}
818
819static bool rb_watermark_hit(struct trace_buffer *buffer, int cpu, int full)
820{
821 struct ring_buffer_per_cpu *cpu_buffer;
822 bool ret = false;
823
824 /* Reads of all CPUs always waits for any data */
825 if (cpu == RING_BUFFER_ALL_CPUS)
826 return !ring_buffer_empty(buffer);
827
828 cpu_buffer = buffer->buffers[cpu];
829
830 if (!ring_buffer_empty_cpu(buffer, cpu)) {
831 unsigned long flags;
832 bool pagebusy;
833
834 if (!full)
835 return true;
836
837 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
838 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
839 ret = !pagebusy && full_hit(buffer, cpu, full);
840
841 if (!ret && (!cpu_buffer->shortest_full ||
842 cpu_buffer->shortest_full > full)) {
843 cpu_buffer->shortest_full = full;
844 }
845 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
846 }
847 return ret;
848}
849
850static inline bool
851rb_wait_cond(struct rb_irq_work *rbwork, struct trace_buffer *buffer,
852 int cpu, int full, ring_buffer_cond_fn cond, void *data)
853{
854 if (rb_watermark_hit(buffer, cpu, full))
855 return true;
856
857 if (cond(data))
858 return true;
859
860 /*
861 * The events can happen in critical sections where
862 * checking a work queue can cause deadlocks.
863 * After adding a task to the queue, this flag is set
864 * only to notify events to try to wake up the queue
865 * using irq_work.
866 *
867 * We don't clear it even if the buffer is no longer
868 * empty. The flag only causes the next event to run
869 * irq_work to do the work queue wake up. The worse
870 * that can happen if we race with !trace_empty() is that
871 * an event will cause an irq_work to try to wake up
872 * an empty queue.
873 *
874 * There's no reason to protect this flag either, as
875 * the work queue and irq_work logic will do the necessary
876 * synchronization for the wake ups. The only thing
877 * that is necessary is that the wake up happens after
878 * a task has been queued. It's OK for spurious wake ups.
879 */
880 if (full)
881 rbwork->full_waiters_pending = true;
882 else
883 rbwork->waiters_pending = true;
884
885 return false;
886}
887
888struct rb_wait_data {
889 struct rb_irq_work *irq_work;
890 int seq;
891};
892
893/*
894 * The default wait condition for ring_buffer_wait() is to just to exit the
895 * wait loop the first time it is woken up.
896 */
897static bool rb_wait_once(void *data)
898{
899 struct rb_wait_data *rdata = data;
900 struct rb_irq_work *rbwork = rdata->irq_work;
901
902 return atomic_read_acquire(&rbwork->seq) != rdata->seq;
903}
904
905/**
906 * ring_buffer_wait - wait for input to the ring buffer
907 * @buffer: buffer to wait on
908 * @cpu: the cpu buffer to wait on
909 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
910 * @cond: condition function to break out of wait (NULL to run once)
911 * @data: the data to pass to @cond.
912 *
913 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
914 * as data is added to any of the @buffer's cpu buffers. Otherwise
915 * it will wait for data to be added to a specific cpu buffer.
916 */
917int ring_buffer_wait(struct trace_buffer *buffer, int cpu, int full,
918 ring_buffer_cond_fn cond, void *data)
919{
920 struct ring_buffer_per_cpu *cpu_buffer;
921 struct wait_queue_head *waitq;
922 struct rb_irq_work *rbwork;
923 struct rb_wait_data rdata;
924 int ret = 0;
925
926 /*
927 * Depending on what the caller is waiting for, either any
928 * data in any cpu buffer, or a specific buffer, put the
929 * caller on the appropriate wait queue.
930 */
931 if (cpu == RING_BUFFER_ALL_CPUS) {
932 rbwork = &buffer->irq_work;
933 /* Full only makes sense on per cpu reads */
934 full = 0;
935 } else {
936 if (!cpumask_test_cpu(cpu, buffer->cpumask))
937 return -ENODEV;
938 cpu_buffer = buffer->buffers[cpu];
939 rbwork = &cpu_buffer->irq_work;
940 }
941
942 if (full)
943 waitq = &rbwork->full_waiters;
944 else
945 waitq = &rbwork->waiters;
946
947 /* Set up to exit loop as soon as it is woken */
948 if (!cond) {
949 cond = rb_wait_once;
950 rdata.irq_work = rbwork;
951 rdata.seq = atomic_read_acquire(&rbwork->seq);
952 data = &rdata;
953 }
954
955 ret = wait_event_interruptible((*waitq),
956 rb_wait_cond(rbwork, buffer, cpu, full, cond, data));
957
958 return ret;
959}
960
961/**
962 * ring_buffer_poll_wait - poll on buffer input
963 * @buffer: buffer to wait on
964 * @cpu: the cpu buffer to wait on
965 * @filp: the file descriptor
966 * @poll_table: The poll descriptor
967 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
968 *
969 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
970 * as data is added to any of the @buffer's cpu buffers. Otherwise
971 * it will wait for data to be added to a specific cpu buffer.
972 *
973 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
974 * zero otherwise.
975 */
976__poll_t ring_buffer_poll_wait(struct trace_buffer *buffer, int cpu,
977 struct file *filp, poll_table *poll_table, int full)
978{
979 struct ring_buffer_per_cpu *cpu_buffer;
980 struct rb_irq_work *rbwork;
981
982 if (cpu == RING_BUFFER_ALL_CPUS) {
983 rbwork = &buffer->irq_work;
984 full = 0;
985 } else {
986 if (!cpumask_test_cpu(cpu, buffer->cpumask))
987 return EPOLLERR;
988
989 cpu_buffer = buffer->buffers[cpu];
990 rbwork = &cpu_buffer->irq_work;
991 }
992
993 if (full) {
994 poll_wait(filp, &rbwork->full_waiters, poll_table);
995
996 if (rb_watermark_hit(buffer, cpu, full))
997 return EPOLLIN | EPOLLRDNORM;
998 /*
999 * Only allow full_waiters_pending update to be seen after
1000 * the shortest_full is set (in rb_watermark_hit). If the
1001 * writer sees the full_waiters_pending flag set, it will
1002 * compare the amount in the ring buffer to shortest_full.
1003 * If the amount in the ring buffer is greater than the
1004 * shortest_full percent, it will call the irq_work handler
1005 * to wake up this list. The irq_handler will reset shortest_full
1006 * back to zero. That's done under the reader_lock, but
1007 * the below smp_mb() makes sure that the update to
1008 * full_waiters_pending doesn't leak up into the above.
1009 */
1010 smp_mb();
1011 rbwork->full_waiters_pending = true;
1012 return 0;
1013 }
1014
1015 poll_wait(filp, &rbwork->waiters, poll_table);
1016 rbwork->waiters_pending = true;
1017
1018 /*
1019 * There's a tight race between setting the waiters_pending and
1020 * checking if the ring buffer is empty. Once the waiters_pending bit
1021 * is set, the next event will wake the task up, but we can get stuck
1022 * if there's only a single event in.
1023 *
1024 * FIXME: Ideally, we need a memory barrier on the writer side as well,
1025 * but adding a memory barrier to all events will cause too much of a
1026 * performance hit in the fast path. We only need a memory barrier when
1027 * the buffer goes from empty to having content. But as this race is
1028 * extremely small, and it's not a problem if another event comes in, we
1029 * will fix it later.
1030 */
1031 smp_mb();
1032
1033 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
1034 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
1035 return EPOLLIN | EPOLLRDNORM;
1036 return 0;
1037}
1038
1039/* buffer may be either ring_buffer or ring_buffer_per_cpu */
1040#define RB_WARN_ON(b, cond) \
1041 ({ \
1042 int _____ret = unlikely(cond); \
1043 if (_____ret) { \
1044 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
1045 struct ring_buffer_per_cpu *__b = \
1046 (void *)b; \
1047 atomic_inc(&__b->buffer->record_disabled); \
1048 } else \
1049 atomic_inc(&b->record_disabled); \
1050 WARN_ON(1); \
1051 } \
1052 _____ret; \
1053 })
1054
1055/* Up this if you want to test the TIME_EXTENTS and normalization */
1056#define DEBUG_SHIFT 0
1057
1058static inline u64 rb_time_stamp(struct trace_buffer *buffer)
1059{
1060 u64 ts;
1061
1062 /* Skip retpolines :-( */
1063 if (IS_ENABLED(CONFIG_MITIGATION_RETPOLINE) && likely(buffer->clock == trace_clock_local))
1064 ts = trace_clock_local();
1065 else
1066 ts = buffer->clock();
1067
1068 /* shift to debug/test normalization and TIME_EXTENTS */
1069 return ts << DEBUG_SHIFT;
1070}
1071
1072u64 ring_buffer_time_stamp(struct trace_buffer *buffer)
1073{
1074 u64 time;
1075
1076 preempt_disable_notrace();
1077 time = rb_time_stamp(buffer);
1078 preempt_enable_notrace();
1079
1080 return time;
1081}
1082EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
1083
1084void ring_buffer_normalize_time_stamp(struct trace_buffer *buffer,
1085 int cpu, u64 *ts)
1086{
1087 /* Just stupid testing the normalize function and deltas */
1088 *ts >>= DEBUG_SHIFT;
1089}
1090EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
1091
1092/*
1093 * Making the ring buffer lockless makes things tricky.
1094 * Although writes only happen on the CPU that they are on,
1095 * and they only need to worry about interrupts. Reads can
1096 * happen on any CPU.
1097 *
1098 * The reader page is always off the ring buffer, but when the
1099 * reader finishes with a page, it needs to swap its page with
1100 * a new one from the buffer. The reader needs to take from
1101 * the head (writes go to the tail). But if a writer is in overwrite
1102 * mode and wraps, it must push the head page forward.
1103 *
1104 * Here lies the problem.
1105 *
1106 * The reader must be careful to replace only the head page, and
1107 * not another one. As described at the top of the file in the
1108 * ASCII art, the reader sets its old page to point to the next
1109 * page after head. It then sets the page after head to point to
1110 * the old reader page. But if the writer moves the head page
1111 * during this operation, the reader could end up with the tail.
1112 *
1113 * We use cmpxchg to help prevent this race. We also do something
1114 * special with the page before head. We set the LSB to 1.
1115 *
1116 * When the writer must push the page forward, it will clear the
1117 * bit that points to the head page, move the head, and then set
1118 * the bit that points to the new head page.
1119 *
1120 * We also don't want an interrupt coming in and moving the head
1121 * page on another writer. Thus we use the second LSB to catch
1122 * that too. Thus:
1123 *
1124 * head->list->prev->next bit 1 bit 0
1125 * ------- -------
1126 * Normal page 0 0
1127 * Points to head page 0 1
1128 * New head page 1 0
1129 *
1130 * Note we can not trust the prev pointer of the head page, because:
1131 *
1132 * +----+ +-----+ +-----+
1133 * | |------>| T |---X--->| N |
1134 * | |<------| | | |
1135 * +----+ +-----+ +-----+
1136 * ^ ^ |
1137 * | +-----+ | |
1138 * +----------| R |----------+ |
1139 * | |<-----------+
1140 * +-----+
1141 *
1142 * Key: ---X--> HEAD flag set in pointer
1143 * T Tail page
1144 * R Reader page
1145 * N Next page
1146 *
1147 * (see __rb_reserve_next() to see where this happens)
1148 *
1149 * What the above shows is that the reader just swapped out
1150 * the reader page with a page in the buffer, but before it
1151 * could make the new header point back to the new page added
1152 * it was preempted by a writer. The writer moved forward onto
1153 * the new page added by the reader and is about to move forward
1154 * again.
1155 *
1156 * You can see, it is legitimate for the previous pointer of
1157 * the head (or any page) not to point back to itself. But only
1158 * temporarily.
1159 */
1160
1161#define RB_PAGE_NORMAL 0UL
1162#define RB_PAGE_HEAD 1UL
1163#define RB_PAGE_UPDATE 2UL
1164
1165
1166#define RB_FLAG_MASK 3UL
1167
1168/* PAGE_MOVED is not part of the mask */
1169#define RB_PAGE_MOVED 4UL
1170
1171/*
1172 * rb_list_head - remove any bit
1173 */
1174static struct list_head *rb_list_head(struct list_head *list)
1175{
1176 unsigned long val = (unsigned long)list;
1177
1178 return (struct list_head *)(val & ~RB_FLAG_MASK);
1179}
1180
1181/*
1182 * rb_is_head_page - test if the given page is the head page
1183 *
1184 * Because the reader may move the head_page pointer, we can
1185 * not trust what the head page is (it may be pointing to
1186 * the reader page). But if the next page is a header page,
1187 * its flags will be non zero.
1188 */
1189static inline int
1190rb_is_head_page(struct buffer_page *page, struct list_head *list)
1191{
1192 unsigned long val;
1193
1194 val = (unsigned long)list->next;
1195
1196 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
1197 return RB_PAGE_MOVED;
1198
1199 return val & RB_FLAG_MASK;
1200}
1201
1202/*
1203 * rb_is_reader_page
1204 *
1205 * The unique thing about the reader page, is that, if the
1206 * writer is ever on it, the previous pointer never points
1207 * back to the reader page.
1208 */
1209static bool rb_is_reader_page(struct buffer_page *page)
1210{
1211 struct list_head *list = page->list.prev;
1212
1213 return rb_list_head(list->next) != &page->list;
1214}
1215
1216/*
1217 * rb_set_list_to_head - set a list_head to be pointing to head.
1218 */
1219static void rb_set_list_to_head(struct list_head *list)
1220{
1221 unsigned long *ptr;
1222
1223 ptr = (unsigned long *)&list->next;
1224 *ptr |= RB_PAGE_HEAD;
1225 *ptr &= ~RB_PAGE_UPDATE;
1226}
1227
1228/*
1229 * rb_head_page_activate - sets up head page
1230 */
1231static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
1232{
1233 struct buffer_page *head;
1234
1235 head = cpu_buffer->head_page;
1236 if (!head)
1237 return;
1238
1239 /*
1240 * Set the previous list pointer to have the HEAD flag.
1241 */
1242 rb_set_list_to_head(head->list.prev);
1243}
1244
1245static void rb_list_head_clear(struct list_head *list)
1246{
1247 unsigned long *ptr = (unsigned long *)&list->next;
1248
1249 *ptr &= ~RB_FLAG_MASK;
1250}
1251
1252/*
1253 * rb_head_page_deactivate - clears head page ptr (for free list)
1254 */
1255static void
1256rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
1257{
1258 struct list_head *hd;
1259
1260 /* Go through the whole list and clear any pointers found. */
1261 rb_list_head_clear(cpu_buffer->pages);
1262
1263 list_for_each(hd, cpu_buffer->pages)
1264 rb_list_head_clear(hd);
1265}
1266
1267static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
1268 struct buffer_page *head,
1269 struct buffer_page *prev,
1270 int old_flag, int new_flag)
1271{
1272 struct list_head *list;
1273 unsigned long val = (unsigned long)&head->list;
1274 unsigned long ret;
1275
1276 list = &prev->list;
1277
1278 val &= ~RB_FLAG_MASK;
1279
1280 ret = cmpxchg((unsigned long *)&list->next,
1281 val | old_flag, val | new_flag);
1282
1283 /* check if the reader took the page */
1284 if ((ret & ~RB_FLAG_MASK) != val)
1285 return RB_PAGE_MOVED;
1286
1287 return ret & RB_FLAG_MASK;
1288}
1289
1290static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
1291 struct buffer_page *head,
1292 struct buffer_page *prev,
1293 int old_flag)
1294{
1295 return rb_head_page_set(cpu_buffer, head, prev,
1296 old_flag, RB_PAGE_UPDATE);
1297}
1298
1299static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
1300 struct buffer_page *head,
1301 struct buffer_page *prev,
1302 int old_flag)
1303{
1304 return rb_head_page_set(cpu_buffer, head, prev,
1305 old_flag, RB_PAGE_HEAD);
1306}
1307
1308static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
1309 struct buffer_page *head,
1310 struct buffer_page *prev,
1311 int old_flag)
1312{
1313 return rb_head_page_set(cpu_buffer, head, prev,
1314 old_flag, RB_PAGE_NORMAL);
1315}
1316
1317static inline void rb_inc_page(struct buffer_page **bpage)
1318{
1319 struct list_head *p = rb_list_head((*bpage)->list.next);
1320
1321 *bpage = list_entry(p, struct buffer_page, list);
1322}
1323
1324static struct buffer_page *
1325rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1326{
1327 struct buffer_page *head;
1328 struct buffer_page *page;
1329 struct list_head *list;
1330 int i;
1331
1332 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1333 return NULL;
1334
1335 /* sanity check */
1336 list = cpu_buffer->pages;
1337 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1338 return NULL;
1339
1340 page = head = cpu_buffer->head_page;
1341 /*
1342 * It is possible that the writer moves the header behind
1343 * where we started, and we miss in one loop.
1344 * A second loop should grab the header, but we'll do
1345 * three loops just because I'm paranoid.
1346 */
1347 for (i = 0; i < 3; i++) {
1348 do {
1349 if (rb_is_head_page(page, page->list.prev)) {
1350 cpu_buffer->head_page = page;
1351 return page;
1352 }
1353 rb_inc_page(&page);
1354 } while (page != head);
1355 }
1356
1357 RB_WARN_ON(cpu_buffer, 1);
1358
1359 return NULL;
1360}
1361
1362static bool rb_head_page_replace(struct buffer_page *old,
1363 struct buffer_page *new)
1364{
1365 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1366 unsigned long val;
1367
1368 val = *ptr & ~RB_FLAG_MASK;
1369 val |= RB_PAGE_HEAD;
1370
1371 return try_cmpxchg(ptr, &val, (unsigned long)&new->list);
1372}
1373
1374/*
1375 * rb_tail_page_update - move the tail page forward
1376 */
1377static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1378 struct buffer_page *tail_page,
1379 struct buffer_page *next_page)
1380{
1381 unsigned long old_entries;
1382 unsigned long old_write;
1383
1384 /*
1385 * The tail page now needs to be moved forward.
1386 *
1387 * We need to reset the tail page, but without messing
1388 * with possible erasing of data brought in by interrupts
1389 * that have moved the tail page and are currently on it.
1390 *
1391 * We add a counter to the write field to denote this.
1392 */
1393 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1394 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1395
1396 /*
1397 * Just make sure we have seen our old_write and synchronize
1398 * with any interrupts that come in.
1399 */
1400 barrier();
1401
1402 /*
1403 * If the tail page is still the same as what we think
1404 * it is, then it is up to us to update the tail
1405 * pointer.
1406 */
1407 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1408 /* Zero the write counter */
1409 unsigned long val = old_write & ~RB_WRITE_MASK;
1410 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1411
1412 /*
1413 * This will only succeed if an interrupt did
1414 * not come in and change it. In which case, we
1415 * do not want to modify it.
1416 *
1417 * We add (void) to let the compiler know that we do not care
1418 * about the return value of these functions. We use the
1419 * cmpxchg to only update if an interrupt did not already
1420 * do it for us. If the cmpxchg fails, we don't care.
1421 */
1422 (void)local_cmpxchg(&next_page->write, old_write, val);
1423 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1424
1425 /*
1426 * No need to worry about races with clearing out the commit.
1427 * it only can increment when a commit takes place. But that
1428 * only happens in the outer most nested commit.
1429 */
1430 local_set(&next_page->page->commit, 0);
1431
1432 /* Either we update tail_page or an interrupt does */
1433 if (try_cmpxchg(&cpu_buffer->tail_page, &tail_page, next_page))
1434 local_inc(&cpu_buffer->pages_touched);
1435 }
1436}
1437
1438static void rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1439 struct buffer_page *bpage)
1440{
1441 unsigned long val = (unsigned long)bpage;
1442
1443 RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK);
1444}
1445
1446/**
1447 * rb_check_pages - integrity check of buffer pages
1448 * @cpu_buffer: CPU buffer with pages to test
1449 *
1450 * As a safety measure we check to make sure the data pages have not
1451 * been corrupted.
1452 *
1453 * Callers of this function need to guarantee that the list of pages doesn't get
1454 * modified during the check. In particular, if it's possible that the function
1455 * is invoked with concurrent readers which can swap in a new reader page then
1456 * the caller should take cpu_buffer->reader_lock.
1457 */
1458static void rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1459{
1460 struct list_head *head = rb_list_head(cpu_buffer->pages);
1461 struct list_head *tmp;
1462
1463 if (RB_WARN_ON(cpu_buffer,
1464 rb_list_head(rb_list_head(head->next)->prev) != head))
1465 return;
1466
1467 if (RB_WARN_ON(cpu_buffer,
1468 rb_list_head(rb_list_head(head->prev)->next) != head))
1469 return;
1470
1471 for (tmp = rb_list_head(head->next); tmp != head; tmp = rb_list_head(tmp->next)) {
1472 if (RB_WARN_ON(cpu_buffer,
1473 rb_list_head(rb_list_head(tmp->next)->prev) != tmp))
1474 return;
1475
1476 if (RB_WARN_ON(cpu_buffer,
1477 rb_list_head(rb_list_head(tmp->prev)->next) != tmp))
1478 return;
1479 }
1480}
1481
1482static int __rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1483 long nr_pages, struct list_head *pages)
1484{
1485 struct buffer_page *bpage, *tmp;
1486 bool user_thread = current->mm != NULL;
1487 gfp_t mflags;
1488 long i;
1489
1490 /*
1491 * Check if the available memory is there first.
1492 * Note, si_mem_available() only gives us a rough estimate of available
1493 * memory. It may not be accurate. But we don't care, we just want
1494 * to prevent doing any allocation when it is obvious that it is
1495 * not going to succeed.
1496 */
1497 i = si_mem_available();
1498 if (i < nr_pages)
1499 return -ENOMEM;
1500
1501 /*
1502 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1503 * gracefully without invoking oom-killer and the system is not
1504 * destabilized.
1505 */
1506 mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1507
1508 /*
1509 * If a user thread allocates too much, and si_mem_available()
1510 * reports there's enough memory, even though there is not.
1511 * Make sure the OOM killer kills this thread. This can happen
1512 * even with RETRY_MAYFAIL because another task may be doing
1513 * an allocation after this task has taken all memory.
1514 * This is the task the OOM killer needs to take out during this
1515 * loop, even if it was triggered by an allocation somewhere else.
1516 */
1517 if (user_thread)
1518 set_current_oom_origin();
1519 for (i = 0; i < nr_pages; i++) {
1520 struct page *page;
1521
1522 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1523 mflags, cpu_to_node(cpu_buffer->cpu));
1524 if (!bpage)
1525 goto free_pages;
1526
1527 rb_check_bpage(cpu_buffer, bpage);
1528
1529 list_add(&bpage->list, pages);
1530
1531 page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu),
1532 mflags | __GFP_ZERO,
1533 cpu_buffer->buffer->subbuf_order);
1534 if (!page)
1535 goto free_pages;
1536 bpage->page = page_address(page);
1537 bpage->order = cpu_buffer->buffer->subbuf_order;
1538 rb_init_page(bpage->page);
1539
1540 if (user_thread && fatal_signal_pending(current))
1541 goto free_pages;
1542 }
1543 if (user_thread)
1544 clear_current_oom_origin();
1545
1546 return 0;
1547
1548free_pages:
1549 list_for_each_entry_safe(bpage, tmp, pages, list) {
1550 list_del_init(&bpage->list);
1551 free_buffer_page(bpage);
1552 }
1553 if (user_thread)
1554 clear_current_oom_origin();
1555
1556 return -ENOMEM;
1557}
1558
1559static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1560 unsigned long nr_pages)
1561{
1562 LIST_HEAD(pages);
1563
1564 WARN_ON(!nr_pages);
1565
1566 if (__rb_allocate_pages(cpu_buffer, nr_pages, &pages))
1567 return -ENOMEM;
1568
1569 /*
1570 * The ring buffer page list is a circular list that does not
1571 * start and end with a list head. All page list items point to
1572 * other pages.
1573 */
1574 cpu_buffer->pages = pages.next;
1575 list_del(&pages);
1576
1577 cpu_buffer->nr_pages = nr_pages;
1578
1579 rb_check_pages(cpu_buffer);
1580
1581 return 0;
1582}
1583
1584static struct ring_buffer_per_cpu *
1585rb_allocate_cpu_buffer(struct trace_buffer *buffer, long nr_pages, int cpu)
1586{
1587 struct ring_buffer_per_cpu *cpu_buffer;
1588 struct buffer_page *bpage;
1589 struct page *page;
1590 int ret;
1591
1592 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1593 GFP_KERNEL, cpu_to_node(cpu));
1594 if (!cpu_buffer)
1595 return NULL;
1596
1597 cpu_buffer->cpu = cpu;
1598 cpu_buffer->buffer = buffer;
1599 raw_spin_lock_init(&cpu_buffer->reader_lock);
1600 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1601 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1602 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1603 init_completion(&cpu_buffer->update_done);
1604 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1605 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1606 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1607
1608 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1609 GFP_KERNEL, cpu_to_node(cpu));
1610 if (!bpage)
1611 goto fail_free_buffer;
1612
1613 rb_check_bpage(cpu_buffer, bpage);
1614
1615 cpu_buffer->reader_page = bpage;
1616
1617 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL | __GFP_ZERO,
1618 cpu_buffer->buffer->subbuf_order);
1619 if (!page)
1620 goto fail_free_reader;
1621 bpage->page = page_address(page);
1622 rb_init_page(bpage->page);
1623
1624 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1625 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1626
1627 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1628 if (ret < 0)
1629 goto fail_free_reader;
1630
1631 cpu_buffer->head_page
1632 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1633 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1634
1635 rb_head_page_activate(cpu_buffer);
1636
1637 return cpu_buffer;
1638
1639 fail_free_reader:
1640 free_buffer_page(cpu_buffer->reader_page);
1641
1642 fail_free_buffer:
1643 kfree(cpu_buffer);
1644 return NULL;
1645}
1646
1647static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1648{
1649 struct list_head *head = cpu_buffer->pages;
1650 struct buffer_page *bpage, *tmp;
1651
1652 irq_work_sync(&cpu_buffer->irq_work.work);
1653
1654 free_buffer_page(cpu_buffer->reader_page);
1655
1656 if (head) {
1657 rb_head_page_deactivate(cpu_buffer);
1658
1659 list_for_each_entry_safe(bpage, tmp, head, list) {
1660 list_del_init(&bpage->list);
1661 free_buffer_page(bpage);
1662 }
1663 bpage = list_entry(head, struct buffer_page, list);
1664 free_buffer_page(bpage);
1665 }
1666
1667 free_page((unsigned long)cpu_buffer->free_page);
1668
1669 kfree(cpu_buffer);
1670}
1671
1672/**
1673 * __ring_buffer_alloc - allocate a new ring_buffer
1674 * @size: the size in bytes per cpu that is needed.
1675 * @flags: attributes to set for the ring buffer.
1676 * @key: ring buffer reader_lock_key.
1677 *
1678 * Currently the only flag that is available is the RB_FL_OVERWRITE
1679 * flag. This flag means that the buffer will overwrite old data
1680 * when the buffer wraps. If this flag is not set, the buffer will
1681 * drop data when the tail hits the head.
1682 */
1683struct trace_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1684 struct lock_class_key *key)
1685{
1686 struct trace_buffer *buffer;
1687 long nr_pages;
1688 int bsize;
1689 int cpu;
1690 int ret;
1691
1692 /* keep it in its own cache line */
1693 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1694 GFP_KERNEL);
1695 if (!buffer)
1696 return NULL;
1697
1698 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1699 goto fail_free_buffer;
1700
1701 /* Default buffer page size - one system page */
1702 buffer->subbuf_order = 0;
1703 buffer->subbuf_size = PAGE_SIZE - BUF_PAGE_HDR_SIZE;
1704
1705 /* Max payload is buffer page size - header (8bytes) */
1706 buffer->max_data_size = buffer->subbuf_size - (sizeof(u32) * 2);
1707
1708 nr_pages = DIV_ROUND_UP(size, buffer->subbuf_size);
1709 buffer->flags = flags;
1710 buffer->clock = trace_clock_local;
1711 buffer->reader_lock_key = key;
1712
1713 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1714 init_waitqueue_head(&buffer->irq_work.waiters);
1715
1716 /* need at least two pages */
1717 if (nr_pages < 2)
1718 nr_pages = 2;
1719
1720 buffer->cpus = nr_cpu_ids;
1721
1722 bsize = sizeof(void *) * nr_cpu_ids;
1723 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1724 GFP_KERNEL);
1725 if (!buffer->buffers)
1726 goto fail_free_cpumask;
1727
1728 cpu = raw_smp_processor_id();
1729 cpumask_set_cpu(cpu, buffer->cpumask);
1730 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1731 if (!buffer->buffers[cpu])
1732 goto fail_free_buffers;
1733
1734 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1735 if (ret < 0)
1736 goto fail_free_buffers;
1737
1738 mutex_init(&buffer->mutex);
1739
1740 return buffer;
1741
1742 fail_free_buffers:
1743 for_each_buffer_cpu(buffer, cpu) {
1744 if (buffer->buffers[cpu])
1745 rb_free_cpu_buffer(buffer->buffers[cpu]);
1746 }
1747 kfree(buffer->buffers);
1748
1749 fail_free_cpumask:
1750 free_cpumask_var(buffer->cpumask);
1751
1752 fail_free_buffer:
1753 kfree(buffer);
1754 return NULL;
1755}
1756EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1757
1758/**
1759 * ring_buffer_free - free a ring buffer.
1760 * @buffer: the buffer to free.
1761 */
1762void
1763ring_buffer_free(struct trace_buffer *buffer)
1764{
1765 int cpu;
1766
1767 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1768
1769 irq_work_sync(&buffer->irq_work.work);
1770
1771 for_each_buffer_cpu(buffer, cpu)
1772 rb_free_cpu_buffer(buffer->buffers[cpu]);
1773
1774 kfree(buffer->buffers);
1775 free_cpumask_var(buffer->cpumask);
1776
1777 kfree(buffer);
1778}
1779EXPORT_SYMBOL_GPL(ring_buffer_free);
1780
1781void ring_buffer_set_clock(struct trace_buffer *buffer,
1782 u64 (*clock)(void))
1783{
1784 buffer->clock = clock;
1785}
1786
1787void ring_buffer_set_time_stamp_abs(struct trace_buffer *buffer, bool abs)
1788{
1789 buffer->time_stamp_abs = abs;
1790}
1791
1792bool ring_buffer_time_stamp_abs(struct trace_buffer *buffer)
1793{
1794 return buffer->time_stamp_abs;
1795}
1796
1797static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1798
1799static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1800{
1801 return local_read(&bpage->entries) & RB_WRITE_MASK;
1802}
1803
1804static inline unsigned long rb_page_write(struct buffer_page *bpage)
1805{
1806 return local_read(&bpage->write) & RB_WRITE_MASK;
1807}
1808
1809static bool
1810rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1811{
1812 struct list_head *tail_page, *to_remove, *next_page;
1813 struct buffer_page *to_remove_page, *tmp_iter_page;
1814 struct buffer_page *last_page, *first_page;
1815 unsigned long nr_removed;
1816 unsigned long head_bit;
1817 int page_entries;
1818
1819 head_bit = 0;
1820
1821 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1822 atomic_inc(&cpu_buffer->record_disabled);
1823 /*
1824 * We don't race with the readers since we have acquired the reader
1825 * lock. We also don't race with writers after disabling recording.
1826 * This makes it easy to figure out the first and the last page to be
1827 * removed from the list. We unlink all the pages in between including
1828 * the first and last pages. This is done in a busy loop so that we
1829 * lose the least number of traces.
1830 * The pages are freed after we restart recording and unlock readers.
1831 */
1832 tail_page = &cpu_buffer->tail_page->list;
1833
1834 /*
1835 * tail page might be on reader page, we remove the next page
1836 * from the ring buffer
1837 */
1838 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1839 tail_page = rb_list_head(tail_page->next);
1840 to_remove = tail_page;
1841
1842 /* start of pages to remove */
1843 first_page = list_entry(rb_list_head(to_remove->next),
1844 struct buffer_page, list);
1845
1846 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1847 to_remove = rb_list_head(to_remove)->next;
1848 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1849 }
1850 /* Read iterators need to reset themselves when some pages removed */
1851 cpu_buffer->pages_removed += nr_removed;
1852
1853 next_page = rb_list_head(to_remove)->next;
1854
1855 /*
1856 * Now we remove all pages between tail_page and next_page.
1857 * Make sure that we have head_bit value preserved for the
1858 * next page
1859 */
1860 tail_page->next = (struct list_head *)((unsigned long)next_page |
1861 head_bit);
1862 next_page = rb_list_head(next_page);
1863 next_page->prev = tail_page;
1864
1865 /* make sure pages points to a valid page in the ring buffer */
1866 cpu_buffer->pages = next_page;
1867
1868 /* update head page */
1869 if (head_bit)
1870 cpu_buffer->head_page = list_entry(next_page,
1871 struct buffer_page, list);
1872
1873 /* pages are removed, resume tracing and then free the pages */
1874 atomic_dec(&cpu_buffer->record_disabled);
1875 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1876
1877 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1878
1879 /* last buffer page to remove */
1880 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1881 list);
1882 tmp_iter_page = first_page;
1883
1884 do {
1885 cond_resched();
1886
1887 to_remove_page = tmp_iter_page;
1888 rb_inc_page(&tmp_iter_page);
1889
1890 /* update the counters */
1891 page_entries = rb_page_entries(to_remove_page);
1892 if (page_entries) {
1893 /*
1894 * If something was added to this page, it was full
1895 * since it is not the tail page. So we deduct the
1896 * bytes consumed in ring buffer from here.
1897 * Increment overrun to account for the lost events.
1898 */
1899 local_add(page_entries, &cpu_buffer->overrun);
1900 local_sub(rb_page_commit(to_remove_page), &cpu_buffer->entries_bytes);
1901 local_inc(&cpu_buffer->pages_lost);
1902 }
1903
1904 /*
1905 * We have already removed references to this list item, just
1906 * free up the buffer_page and its page
1907 */
1908 free_buffer_page(to_remove_page);
1909 nr_removed--;
1910
1911 } while (to_remove_page != last_page);
1912
1913 RB_WARN_ON(cpu_buffer, nr_removed);
1914
1915 return nr_removed == 0;
1916}
1917
1918static bool
1919rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1920{
1921 struct list_head *pages = &cpu_buffer->new_pages;
1922 unsigned long flags;
1923 bool success;
1924 int retries;
1925
1926 /* Can be called at early boot up, where interrupts must not been enabled */
1927 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
1928 /*
1929 * We are holding the reader lock, so the reader page won't be swapped
1930 * in the ring buffer. Now we are racing with the writer trying to
1931 * move head page and the tail page.
1932 * We are going to adapt the reader page update process where:
1933 * 1. We first splice the start and end of list of new pages between
1934 * the head page and its previous page.
1935 * 2. We cmpxchg the prev_page->next to point from head page to the
1936 * start of new pages list.
1937 * 3. Finally, we update the head->prev to the end of new list.
1938 *
1939 * We will try this process 10 times, to make sure that we don't keep
1940 * spinning.
1941 */
1942 retries = 10;
1943 success = false;
1944 while (retries--) {
1945 struct list_head *head_page, *prev_page;
1946 struct list_head *last_page, *first_page;
1947 struct list_head *head_page_with_bit;
1948 struct buffer_page *hpage = rb_set_head_page(cpu_buffer);
1949
1950 if (!hpage)
1951 break;
1952 head_page = &hpage->list;
1953 prev_page = head_page->prev;
1954
1955 first_page = pages->next;
1956 last_page = pages->prev;
1957
1958 head_page_with_bit = (struct list_head *)
1959 ((unsigned long)head_page | RB_PAGE_HEAD);
1960
1961 last_page->next = head_page_with_bit;
1962 first_page->prev = prev_page;
1963
1964 /* caution: head_page_with_bit gets updated on cmpxchg failure */
1965 if (try_cmpxchg(&prev_page->next,
1966 &head_page_with_bit, first_page)) {
1967 /*
1968 * yay, we replaced the page pointer to our new list,
1969 * now, we just have to update to head page's prev
1970 * pointer to point to end of list
1971 */
1972 head_page->prev = last_page;
1973 success = true;
1974 break;
1975 }
1976 }
1977
1978 if (success)
1979 INIT_LIST_HEAD(pages);
1980 /*
1981 * If we weren't successful in adding in new pages, warn and stop
1982 * tracing
1983 */
1984 RB_WARN_ON(cpu_buffer, !success);
1985 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
1986
1987 /* free pages if they weren't inserted */
1988 if (!success) {
1989 struct buffer_page *bpage, *tmp;
1990 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1991 list) {
1992 list_del_init(&bpage->list);
1993 free_buffer_page(bpage);
1994 }
1995 }
1996 return success;
1997}
1998
1999static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
2000{
2001 bool success;
2002
2003 if (cpu_buffer->nr_pages_to_update > 0)
2004 success = rb_insert_pages(cpu_buffer);
2005 else
2006 success = rb_remove_pages(cpu_buffer,
2007 -cpu_buffer->nr_pages_to_update);
2008
2009 if (success)
2010 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
2011}
2012
2013static void update_pages_handler(struct work_struct *work)
2014{
2015 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
2016 struct ring_buffer_per_cpu, update_pages_work);
2017 rb_update_pages(cpu_buffer);
2018 complete(&cpu_buffer->update_done);
2019}
2020
2021/**
2022 * ring_buffer_resize - resize the ring buffer
2023 * @buffer: the buffer to resize.
2024 * @size: the new size.
2025 * @cpu_id: the cpu buffer to resize
2026 *
2027 * Minimum size is 2 * buffer->subbuf_size.
2028 *
2029 * Returns 0 on success and < 0 on failure.
2030 */
2031int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
2032 int cpu_id)
2033{
2034 struct ring_buffer_per_cpu *cpu_buffer;
2035 unsigned long nr_pages;
2036 int cpu, err;
2037
2038 /*
2039 * Always succeed at resizing a non-existent buffer:
2040 */
2041 if (!buffer)
2042 return 0;
2043
2044 /* Make sure the requested buffer exists */
2045 if (cpu_id != RING_BUFFER_ALL_CPUS &&
2046 !cpumask_test_cpu(cpu_id, buffer->cpumask))
2047 return 0;
2048
2049 nr_pages = DIV_ROUND_UP(size, buffer->subbuf_size);
2050
2051 /* we need a minimum of two pages */
2052 if (nr_pages < 2)
2053 nr_pages = 2;
2054
2055 /* prevent another thread from changing buffer sizes */
2056 mutex_lock(&buffer->mutex);
2057 atomic_inc(&buffer->resizing);
2058
2059 if (cpu_id == RING_BUFFER_ALL_CPUS) {
2060 /*
2061 * Don't succeed if resizing is disabled, as a reader might be
2062 * manipulating the ring buffer and is expecting a sane state while
2063 * this is true.
2064 */
2065 for_each_buffer_cpu(buffer, cpu) {
2066 cpu_buffer = buffer->buffers[cpu];
2067 if (atomic_read(&cpu_buffer->resize_disabled)) {
2068 err = -EBUSY;
2069 goto out_err_unlock;
2070 }
2071 }
2072
2073 /* calculate the pages to update */
2074 for_each_buffer_cpu(buffer, cpu) {
2075 cpu_buffer = buffer->buffers[cpu];
2076
2077 cpu_buffer->nr_pages_to_update = nr_pages -
2078 cpu_buffer->nr_pages;
2079 /*
2080 * nothing more to do for removing pages or no update
2081 */
2082 if (cpu_buffer->nr_pages_to_update <= 0)
2083 continue;
2084 /*
2085 * to add pages, make sure all new pages can be
2086 * allocated without receiving ENOMEM
2087 */
2088 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2089 if (__rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2090 &cpu_buffer->new_pages)) {
2091 /* not enough memory for new pages */
2092 err = -ENOMEM;
2093 goto out_err;
2094 }
2095
2096 cond_resched();
2097 }
2098
2099 cpus_read_lock();
2100 /*
2101 * Fire off all the required work handlers
2102 * We can't schedule on offline CPUs, but it's not necessary
2103 * since we can change their buffer sizes without any race.
2104 */
2105 for_each_buffer_cpu(buffer, cpu) {
2106 cpu_buffer = buffer->buffers[cpu];
2107 if (!cpu_buffer->nr_pages_to_update)
2108 continue;
2109
2110 /* Can't run something on an offline CPU. */
2111 if (!cpu_online(cpu)) {
2112 rb_update_pages(cpu_buffer);
2113 cpu_buffer->nr_pages_to_update = 0;
2114 } else {
2115 /* Run directly if possible. */
2116 migrate_disable();
2117 if (cpu != smp_processor_id()) {
2118 migrate_enable();
2119 schedule_work_on(cpu,
2120 &cpu_buffer->update_pages_work);
2121 } else {
2122 update_pages_handler(&cpu_buffer->update_pages_work);
2123 migrate_enable();
2124 }
2125 }
2126 }
2127
2128 /* wait for all the updates to complete */
2129 for_each_buffer_cpu(buffer, cpu) {
2130 cpu_buffer = buffer->buffers[cpu];
2131 if (!cpu_buffer->nr_pages_to_update)
2132 continue;
2133
2134 if (cpu_online(cpu))
2135 wait_for_completion(&cpu_buffer->update_done);
2136 cpu_buffer->nr_pages_to_update = 0;
2137 }
2138
2139 cpus_read_unlock();
2140 } else {
2141 cpu_buffer = buffer->buffers[cpu_id];
2142
2143 if (nr_pages == cpu_buffer->nr_pages)
2144 goto out;
2145
2146 /*
2147 * Don't succeed if resizing is disabled, as a reader might be
2148 * manipulating the ring buffer and is expecting a sane state while
2149 * this is true.
2150 */
2151 if (atomic_read(&cpu_buffer->resize_disabled)) {
2152 err = -EBUSY;
2153 goto out_err_unlock;
2154 }
2155
2156 cpu_buffer->nr_pages_to_update = nr_pages -
2157 cpu_buffer->nr_pages;
2158
2159 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2160 if (cpu_buffer->nr_pages_to_update > 0 &&
2161 __rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2162 &cpu_buffer->new_pages)) {
2163 err = -ENOMEM;
2164 goto out_err;
2165 }
2166
2167 cpus_read_lock();
2168
2169 /* Can't run something on an offline CPU. */
2170 if (!cpu_online(cpu_id))
2171 rb_update_pages(cpu_buffer);
2172 else {
2173 /* Run directly if possible. */
2174 migrate_disable();
2175 if (cpu_id == smp_processor_id()) {
2176 rb_update_pages(cpu_buffer);
2177 migrate_enable();
2178 } else {
2179 migrate_enable();
2180 schedule_work_on(cpu_id,
2181 &cpu_buffer->update_pages_work);
2182 wait_for_completion(&cpu_buffer->update_done);
2183 }
2184 }
2185
2186 cpu_buffer->nr_pages_to_update = 0;
2187 cpus_read_unlock();
2188 }
2189
2190 out:
2191 /*
2192 * The ring buffer resize can happen with the ring buffer
2193 * enabled, so that the update disturbs the tracing as little
2194 * as possible. But if the buffer is disabled, we do not need
2195 * to worry about that, and we can take the time to verify
2196 * that the buffer is not corrupt.
2197 */
2198 if (atomic_read(&buffer->record_disabled)) {
2199 atomic_inc(&buffer->record_disabled);
2200 /*
2201 * Even though the buffer was disabled, we must make sure
2202 * that it is truly disabled before calling rb_check_pages.
2203 * There could have been a race between checking
2204 * record_disable and incrementing it.
2205 */
2206 synchronize_rcu();
2207 for_each_buffer_cpu(buffer, cpu) {
2208 unsigned long flags;
2209
2210 cpu_buffer = buffer->buffers[cpu];
2211 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2212 rb_check_pages(cpu_buffer);
2213 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2214 }
2215 atomic_dec(&buffer->record_disabled);
2216 }
2217
2218 atomic_dec(&buffer->resizing);
2219 mutex_unlock(&buffer->mutex);
2220 return 0;
2221
2222 out_err:
2223 for_each_buffer_cpu(buffer, cpu) {
2224 struct buffer_page *bpage, *tmp;
2225
2226 cpu_buffer = buffer->buffers[cpu];
2227 cpu_buffer->nr_pages_to_update = 0;
2228
2229 if (list_empty(&cpu_buffer->new_pages))
2230 continue;
2231
2232 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2233 list) {
2234 list_del_init(&bpage->list);
2235 free_buffer_page(bpage);
2236 }
2237 }
2238 out_err_unlock:
2239 atomic_dec(&buffer->resizing);
2240 mutex_unlock(&buffer->mutex);
2241 return err;
2242}
2243EXPORT_SYMBOL_GPL(ring_buffer_resize);
2244
2245void ring_buffer_change_overwrite(struct trace_buffer *buffer, int val)
2246{
2247 mutex_lock(&buffer->mutex);
2248 if (val)
2249 buffer->flags |= RB_FL_OVERWRITE;
2250 else
2251 buffer->flags &= ~RB_FL_OVERWRITE;
2252 mutex_unlock(&buffer->mutex);
2253}
2254EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
2255
2256static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
2257{
2258 return bpage->page->data + index;
2259}
2260
2261static __always_inline struct ring_buffer_event *
2262rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
2263{
2264 return __rb_page_index(cpu_buffer->reader_page,
2265 cpu_buffer->reader_page->read);
2266}
2267
2268static struct ring_buffer_event *
2269rb_iter_head_event(struct ring_buffer_iter *iter)
2270{
2271 struct ring_buffer_event *event;
2272 struct buffer_page *iter_head_page = iter->head_page;
2273 unsigned long commit;
2274 unsigned length;
2275
2276 if (iter->head != iter->next_event)
2277 return iter->event;
2278
2279 /*
2280 * When the writer goes across pages, it issues a cmpxchg which
2281 * is a mb(), which will synchronize with the rmb here.
2282 * (see rb_tail_page_update() and __rb_reserve_next())
2283 */
2284 commit = rb_page_commit(iter_head_page);
2285 smp_rmb();
2286
2287 /* An event needs to be at least 8 bytes in size */
2288 if (iter->head > commit - 8)
2289 goto reset;
2290
2291 event = __rb_page_index(iter_head_page, iter->head);
2292 length = rb_event_length(event);
2293
2294 /*
2295 * READ_ONCE() doesn't work on functions and we don't want the
2296 * compiler doing any crazy optimizations with length.
2297 */
2298 barrier();
2299
2300 if ((iter->head + length) > commit || length > iter->event_size)
2301 /* Writer corrupted the read? */
2302 goto reset;
2303
2304 memcpy(iter->event, event, length);
2305 /*
2306 * If the page stamp is still the same after this rmb() then the
2307 * event was safely copied without the writer entering the page.
2308 */
2309 smp_rmb();
2310
2311 /* Make sure the page didn't change since we read this */
2312 if (iter->page_stamp != iter_head_page->page->time_stamp ||
2313 commit > rb_page_commit(iter_head_page))
2314 goto reset;
2315
2316 iter->next_event = iter->head + length;
2317 return iter->event;
2318 reset:
2319 /* Reset to the beginning */
2320 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2321 iter->head = 0;
2322 iter->next_event = 0;
2323 iter->missed_events = 1;
2324 return NULL;
2325}
2326
2327/* Size is determined by what has been committed */
2328static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
2329{
2330 return rb_page_commit(bpage);
2331}
2332
2333static __always_inline unsigned
2334rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
2335{
2336 return rb_page_commit(cpu_buffer->commit_page);
2337}
2338
2339static __always_inline unsigned
2340rb_event_index(struct ring_buffer_per_cpu *cpu_buffer, struct ring_buffer_event *event)
2341{
2342 unsigned long addr = (unsigned long)event;
2343
2344 addr &= (PAGE_SIZE << cpu_buffer->buffer->subbuf_order) - 1;
2345
2346 return addr - BUF_PAGE_HDR_SIZE;
2347}
2348
2349static void rb_inc_iter(struct ring_buffer_iter *iter)
2350{
2351 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2352
2353 /*
2354 * The iterator could be on the reader page (it starts there).
2355 * But the head could have moved, since the reader was
2356 * found. Check for this case and assign the iterator
2357 * to the head page instead of next.
2358 */
2359 if (iter->head_page == cpu_buffer->reader_page)
2360 iter->head_page = rb_set_head_page(cpu_buffer);
2361 else
2362 rb_inc_page(&iter->head_page);
2363
2364 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2365 iter->head = 0;
2366 iter->next_event = 0;
2367}
2368
2369/*
2370 * rb_handle_head_page - writer hit the head page
2371 *
2372 * Returns: +1 to retry page
2373 * 0 to continue
2374 * -1 on error
2375 */
2376static int
2377rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2378 struct buffer_page *tail_page,
2379 struct buffer_page *next_page)
2380{
2381 struct buffer_page *new_head;
2382 int entries;
2383 int type;
2384 int ret;
2385
2386 entries = rb_page_entries(next_page);
2387
2388 /*
2389 * The hard part is here. We need to move the head
2390 * forward, and protect against both readers on
2391 * other CPUs and writers coming in via interrupts.
2392 */
2393 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2394 RB_PAGE_HEAD);
2395
2396 /*
2397 * type can be one of four:
2398 * NORMAL - an interrupt already moved it for us
2399 * HEAD - we are the first to get here.
2400 * UPDATE - we are the interrupt interrupting
2401 * a current move.
2402 * MOVED - a reader on another CPU moved the next
2403 * pointer to its reader page. Give up
2404 * and try again.
2405 */
2406
2407 switch (type) {
2408 case RB_PAGE_HEAD:
2409 /*
2410 * We changed the head to UPDATE, thus
2411 * it is our responsibility to update
2412 * the counters.
2413 */
2414 local_add(entries, &cpu_buffer->overrun);
2415 local_sub(rb_page_commit(next_page), &cpu_buffer->entries_bytes);
2416 local_inc(&cpu_buffer->pages_lost);
2417
2418 /*
2419 * The entries will be zeroed out when we move the
2420 * tail page.
2421 */
2422
2423 /* still more to do */
2424 break;
2425
2426 case RB_PAGE_UPDATE:
2427 /*
2428 * This is an interrupt that interrupt the
2429 * previous update. Still more to do.
2430 */
2431 break;
2432 case RB_PAGE_NORMAL:
2433 /*
2434 * An interrupt came in before the update
2435 * and processed this for us.
2436 * Nothing left to do.
2437 */
2438 return 1;
2439 case RB_PAGE_MOVED:
2440 /*
2441 * The reader is on another CPU and just did
2442 * a swap with our next_page.
2443 * Try again.
2444 */
2445 return 1;
2446 default:
2447 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2448 return -1;
2449 }
2450
2451 /*
2452 * Now that we are here, the old head pointer is
2453 * set to UPDATE. This will keep the reader from
2454 * swapping the head page with the reader page.
2455 * The reader (on another CPU) will spin till
2456 * we are finished.
2457 *
2458 * We just need to protect against interrupts
2459 * doing the job. We will set the next pointer
2460 * to HEAD. After that, we set the old pointer
2461 * to NORMAL, but only if it was HEAD before.
2462 * otherwise we are an interrupt, and only
2463 * want the outer most commit to reset it.
2464 */
2465 new_head = next_page;
2466 rb_inc_page(&new_head);
2467
2468 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2469 RB_PAGE_NORMAL);
2470
2471 /*
2472 * Valid returns are:
2473 * HEAD - an interrupt came in and already set it.
2474 * NORMAL - One of two things:
2475 * 1) We really set it.
2476 * 2) A bunch of interrupts came in and moved
2477 * the page forward again.
2478 */
2479 switch (ret) {
2480 case RB_PAGE_HEAD:
2481 case RB_PAGE_NORMAL:
2482 /* OK */
2483 break;
2484 default:
2485 RB_WARN_ON(cpu_buffer, 1);
2486 return -1;
2487 }
2488
2489 /*
2490 * It is possible that an interrupt came in,
2491 * set the head up, then more interrupts came in
2492 * and moved it again. When we get back here,
2493 * the page would have been set to NORMAL but we
2494 * just set it back to HEAD.
2495 *
2496 * How do you detect this? Well, if that happened
2497 * the tail page would have moved.
2498 */
2499 if (ret == RB_PAGE_NORMAL) {
2500 struct buffer_page *buffer_tail_page;
2501
2502 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2503 /*
2504 * If the tail had moved passed next, then we need
2505 * to reset the pointer.
2506 */
2507 if (buffer_tail_page != tail_page &&
2508 buffer_tail_page != next_page)
2509 rb_head_page_set_normal(cpu_buffer, new_head,
2510 next_page,
2511 RB_PAGE_HEAD);
2512 }
2513
2514 /*
2515 * If this was the outer most commit (the one that
2516 * changed the original pointer from HEAD to UPDATE),
2517 * then it is up to us to reset it to NORMAL.
2518 */
2519 if (type == RB_PAGE_HEAD) {
2520 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2521 tail_page,
2522 RB_PAGE_UPDATE);
2523 if (RB_WARN_ON(cpu_buffer,
2524 ret != RB_PAGE_UPDATE))
2525 return -1;
2526 }
2527
2528 return 0;
2529}
2530
2531static inline void
2532rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2533 unsigned long tail, struct rb_event_info *info)
2534{
2535 unsigned long bsize = READ_ONCE(cpu_buffer->buffer->subbuf_size);
2536 struct buffer_page *tail_page = info->tail_page;
2537 struct ring_buffer_event *event;
2538 unsigned long length = info->length;
2539
2540 /*
2541 * Only the event that crossed the page boundary
2542 * must fill the old tail_page with padding.
2543 */
2544 if (tail >= bsize) {
2545 /*
2546 * If the page was filled, then we still need
2547 * to update the real_end. Reset it to zero
2548 * and the reader will ignore it.
2549 */
2550 if (tail == bsize)
2551 tail_page->real_end = 0;
2552
2553 local_sub(length, &tail_page->write);
2554 return;
2555 }
2556
2557 event = __rb_page_index(tail_page, tail);
2558
2559 /*
2560 * Save the original length to the meta data.
2561 * This will be used by the reader to add lost event
2562 * counter.
2563 */
2564 tail_page->real_end = tail;
2565
2566 /*
2567 * If this event is bigger than the minimum size, then
2568 * we need to be careful that we don't subtract the
2569 * write counter enough to allow another writer to slip
2570 * in on this page.
2571 * We put in a discarded commit instead, to make sure
2572 * that this space is not used again, and this space will
2573 * not be accounted into 'entries_bytes'.
2574 *
2575 * If we are less than the minimum size, we don't need to
2576 * worry about it.
2577 */
2578 if (tail > (bsize - RB_EVNT_MIN_SIZE)) {
2579 /* No room for any events */
2580
2581 /* Mark the rest of the page with padding */
2582 rb_event_set_padding(event);
2583
2584 /* Make sure the padding is visible before the write update */
2585 smp_wmb();
2586
2587 /* Set the write back to the previous setting */
2588 local_sub(length, &tail_page->write);
2589 return;
2590 }
2591
2592 /* Put in a discarded event */
2593 event->array[0] = (bsize - tail) - RB_EVNT_HDR_SIZE;
2594 event->type_len = RINGBUF_TYPE_PADDING;
2595 /* time delta must be non zero */
2596 event->time_delta = 1;
2597
2598 /* account for padding bytes */
2599 local_add(bsize - tail, &cpu_buffer->entries_bytes);
2600
2601 /* Make sure the padding is visible before the tail_page->write update */
2602 smp_wmb();
2603
2604 /* Set write to end of buffer */
2605 length = (tail + length) - bsize;
2606 local_sub(length, &tail_page->write);
2607}
2608
2609static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2610
2611/*
2612 * This is the slow path, force gcc not to inline it.
2613 */
2614static noinline struct ring_buffer_event *
2615rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2616 unsigned long tail, struct rb_event_info *info)
2617{
2618 struct buffer_page *tail_page = info->tail_page;
2619 struct buffer_page *commit_page = cpu_buffer->commit_page;
2620 struct trace_buffer *buffer = cpu_buffer->buffer;
2621 struct buffer_page *next_page;
2622 int ret;
2623
2624 next_page = tail_page;
2625
2626 rb_inc_page(&next_page);
2627
2628 /*
2629 * If for some reason, we had an interrupt storm that made
2630 * it all the way around the buffer, bail, and warn
2631 * about it.
2632 */
2633 if (unlikely(next_page == commit_page)) {
2634 local_inc(&cpu_buffer->commit_overrun);
2635 goto out_reset;
2636 }
2637
2638 /*
2639 * This is where the fun begins!
2640 *
2641 * We are fighting against races between a reader that
2642 * could be on another CPU trying to swap its reader
2643 * page with the buffer head.
2644 *
2645 * We are also fighting against interrupts coming in and
2646 * moving the head or tail on us as well.
2647 *
2648 * If the next page is the head page then we have filled
2649 * the buffer, unless the commit page is still on the
2650 * reader page.
2651 */
2652 if (rb_is_head_page(next_page, &tail_page->list)) {
2653
2654 /*
2655 * If the commit is not on the reader page, then
2656 * move the header page.
2657 */
2658 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2659 /*
2660 * If we are not in overwrite mode,
2661 * this is easy, just stop here.
2662 */
2663 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2664 local_inc(&cpu_buffer->dropped_events);
2665 goto out_reset;
2666 }
2667
2668 ret = rb_handle_head_page(cpu_buffer,
2669 tail_page,
2670 next_page);
2671 if (ret < 0)
2672 goto out_reset;
2673 if (ret)
2674 goto out_again;
2675 } else {
2676 /*
2677 * We need to be careful here too. The
2678 * commit page could still be on the reader
2679 * page. We could have a small buffer, and
2680 * have filled up the buffer with events
2681 * from interrupts and such, and wrapped.
2682 *
2683 * Note, if the tail page is also on the
2684 * reader_page, we let it move out.
2685 */
2686 if (unlikely((cpu_buffer->commit_page !=
2687 cpu_buffer->tail_page) &&
2688 (cpu_buffer->commit_page ==
2689 cpu_buffer->reader_page))) {
2690 local_inc(&cpu_buffer->commit_overrun);
2691 goto out_reset;
2692 }
2693 }
2694 }
2695
2696 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2697
2698 out_again:
2699
2700 rb_reset_tail(cpu_buffer, tail, info);
2701
2702 /* Commit what we have for now. */
2703 rb_end_commit(cpu_buffer);
2704 /* rb_end_commit() decs committing */
2705 local_inc(&cpu_buffer->committing);
2706
2707 /* fail and let the caller try again */
2708 return ERR_PTR(-EAGAIN);
2709
2710 out_reset:
2711 /* reset write */
2712 rb_reset_tail(cpu_buffer, tail, info);
2713
2714 return NULL;
2715}
2716
2717/* Slow path */
2718static struct ring_buffer_event *
2719rb_add_time_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2720 struct ring_buffer_event *event, u64 delta, bool abs)
2721{
2722 if (abs)
2723 event->type_len = RINGBUF_TYPE_TIME_STAMP;
2724 else
2725 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2726
2727 /* Not the first event on the page, or not delta? */
2728 if (abs || rb_event_index(cpu_buffer, event)) {
2729 event->time_delta = delta & TS_MASK;
2730 event->array[0] = delta >> TS_SHIFT;
2731 } else {
2732 /* nope, just zero it */
2733 event->time_delta = 0;
2734 event->array[0] = 0;
2735 }
2736
2737 return skip_time_extend(event);
2738}
2739
2740#ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2741static inline bool sched_clock_stable(void)
2742{
2743 return true;
2744}
2745#endif
2746
2747static void
2748rb_check_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2749 struct rb_event_info *info)
2750{
2751 u64 write_stamp;
2752
2753 WARN_ONCE(1, "Delta way too big! %llu ts=%llu before=%llu after=%llu write stamp=%llu\n%s",
2754 (unsigned long long)info->delta,
2755 (unsigned long long)info->ts,
2756 (unsigned long long)info->before,
2757 (unsigned long long)info->after,
2758 (unsigned long long)({rb_time_read(&cpu_buffer->write_stamp, &write_stamp); write_stamp;}),
2759 sched_clock_stable() ? "" :
2760 "If you just came from a suspend/resume,\n"
2761 "please switch to the trace global clock:\n"
2762 " echo global > /sys/kernel/tracing/trace_clock\n"
2763 "or add trace_clock=global to the kernel command line\n");
2764}
2765
2766static void rb_add_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2767 struct ring_buffer_event **event,
2768 struct rb_event_info *info,
2769 u64 *delta,
2770 unsigned int *length)
2771{
2772 bool abs = info->add_timestamp &
2773 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE);
2774
2775 if (unlikely(info->delta > (1ULL << 59))) {
2776 /*
2777 * Some timers can use more than 59 bits, and when a timestamp
2778 * is added to the buffer, it will lose those bits.
2779 */
2780 if (abs && (info->ts & TS_MSB)) {
2781 info->delta &= ABS_TS_MASK;
2782
2783 /* did the clock go backwards */
2784 } else if (info->before == info->after && info->before > info->ts) {
2785 /* not interrupted */
2786 static int once;
2787
2788 /*
2789 * This is possible with a recalibrating of the TSC.
2790 * Do not produce a call stack, but just report it.
2791 */
2792 if (!once) {
2793 once++;
2794 pr_warn("Ring buffer clock went backwards: %llu -> %llu\n",
2795 info->before, info->ts);
2796 }
2797 } else
2798 rb_check_timestamp(cpu_buffer, info);
2799 if (!abs)
2800 info->delta = 0;
2801 }
2802 *event = rb_add_time_stamp(cpu_buffer, *event, info->delta, abs);
2803 *length -= RB_LEN_TIME_EXTEND;
2804 *delta = 0;
2805}
2806
2807/**
2808 * rb_update_event - update event type and data
2809 * @cpu_buffer: The per cpu buffer of the @event
2810 * @event: the event to update
2811 * @info: The info to update the @event with (contains length and delta)
2812 *
2813 * Update the type and data fields of the @event. The length
2814 * is the actual size that is written to the ring buffer,
2815 * and with this, we can determine what to place into the
2816 * data field.
2817 */
2818static void
2819rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2820 struct ring_buffer_event *event,
2821 struct rb_event_info *info)
2822{
2823 unsigned length = info->length;
2824 u64 delta = info->delta;
2825 unsigned int nest = local_read(&cpu_buffer->committing) - 1;
2826
2827 if (!WARN_ON_ONCE(nest >= MAX_NEST))
2828 cpu_buffer->event_stamp[nest] = info->ts;
2829
2830 /*
2831 * If we need to add a timestamp, then we
2832 * add it to the start of the reserved space.
2833 */
2834 if (unlikely(info->add_timestamp))
2835 rb_add_timestamp(cpu_buffer, &event, info, &delta, &length);
2836
2837 event->time_delta = delta;
2838 length -= RB_EVNT_HDR_SIZE;
2839 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2840 event->type_len = 0;
2841 event->array[0] = length;
2842 } else
2843 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2844}
2845
2846static unsigned rb_calculate_event_length(unsigned length)
2847{
2848 struct ring_buffer_event event; /* Used only for sizeof array */
2849
2850 /* zero length can cause confusions */
2851 if (!length)
2852 length++;
2853
2854 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2855 length += sizeof(event.array[0]);
2856
2857 length += RB_EVNT_HDR_SIZE;
2858 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2859
2860 /*
2861 * In case the time delta is larger than the 27 bits for it
2862 * in the header, we need to add a timestamp. If another
2863 * event comes in when trying to discard this one to increase
2864 * the length, then the timestamp will be added in the allocated
2865 * space of this event. If length is bigger than the size needed
2866 * for the TIME_EXTEND, then padding has to be used. The events
2867 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2868 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2869 * As length is a multiple of 4, we only need to worry if it
2870 * is 12 (RB_LEN_TIME_EXTEND + 4).
2871 */
2872 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2873 length += RB_ALIGNMENT;
2874
2875 return length;
2876}
2877
2878static inline bool
2879rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2880 struct ring_buffer_event *event)
2881{
2882 unsigned long new_index, old_index;
2883 struct buffer_page *bpage;
2884 unsigned long addr;
2885
2886 new_index = rb_event_index(cpu_buffer, event);
2887 old_index = new_index + rb_event_ts_length(event);
2888 addr = (unsigned long)event;
2889 addr &= ~((PAGE_SIZE << cpu_buffer->buffer->subbuf_order) - 1);
2890
2891 bpage = READ_ONCE(cpu_buffer->tail_page);
2892
2893 /*
2894 * Make sure the tail_page is still the same and
2895 * the next write location is the end of this event
2896 */
2897 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2898 unsigned long write_mask =
2899 local_read(&bpage->write) & ~RB_WRITE_MASK;
2900 unsigned long event_length = rb_event_length(event);
2901
2902 /*
2903 * For the before_stamp to be different than the write_stamp
2904 * to make sure that the next event adds an absolute
2905 * value and does not rely on the saved write stamp, which
2906 * is now going to be bogus.
2907 *
2908 * By setting the before_stamp to zero, the next event
2909 * is not going to use the write_stamp and will instead
2910 * create an absolute timestamp. This means there's no
2911 * reason to update the wirte_stamp!
2912 */
2913 rb_time_set(&cpu_buffer->before_stamp, 0);
2914
2915 /*
2916 * If an event were to come in now, it would see that the
2917 * write_stamp and the before_stamp are different, and assume
2918 * that this event just added itself before updating
2919 * the write stamp. The interrupting event will fix the
2920 * write stamp for us, and use an absolute timestamp.
2921 */
2922
2923 /*
2924 * This is on the tail page. It is possible that
2925 * a write could come in and move the tail page
2926 * and write to the next page. That is fine
2927 * because we just shorten what is on this page.
2928 */
2929 old_index += write_mask;
2930 new_index += write_mask;
2931
2932 /* caution: old_index gets updated on cmpxchg failure */
2933 if (local_try_cmpxchg(&bpage->write, &old_index, new_index)) {
2934 /* update counters */
2935 local_sub(event_length, &cpu_buffer->entries_bytes);
2936 return true;
2937 }
2938 }
2939
2940 /* could not discard */
2941 return false;
2942}
2943
2944static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2945{
2946 local_inc(&cpu_buffer->committing);
2947 local_inc(&cpu_buffer->commits);
2948}
2949
2950static __always_inline void
2951rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2952{
2953 unsigned long max_count;
2954
2955 /*
2956 * We only race with interrupts and NMIs on this CPU.
2957 * If we own the commit event, then we can commit
2958 * all others that interrupted us, since the interruptions
2959 * are in stack format (they finish before they come
2960 * back to us). This allows us to do a simple loop to
2961 * assign the commit to the tail.
2962 */
2963 again:
2964 max_count = cpu_buffer->nr_pages * 100;
2965
2966 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2967 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2968 return;
2969 if (RB_WARN_ON(cpu_buffer,
2970 rb_is_reader_page(cpu_buffer->tail_page)))
2971 return;
2972 /*
2973 * No need for a memory barrier here, as the update
2974 * of the tail_page did it for this page.
2975 */
2976 local_set(&cpu_buffer->commit_page->page->commit,
2977 rb_page_write(cpu_buffer->commit_page));
2978 rb_inc_page(&cpu_buffer->commit_page);
2979 /* add barrier to keep gcc from optimizing too much */
2980 barrier();
2981 }
2982 while (rb_commit_index(cpu_buffer) !=
2983 rb_page_write(cpu_buffer->commit_page)) {
2984
2985 /* Make sure the readers see the content of what is committed. */
2986 smp_wmb();
2987 local_set(&cpu_buffer->commit_page->page->commit,
2988 rb_page_write(cpu_buffer->commit_page));
2989 RB_WARN_ON(cpu_buffer,
2990 local_read(&cpu_buffer->commit_page->page->commit) &
2991 ~RB_WRITE_MASK);
2992 barrier();
2993 }
2994
2995 /* again, keep gcc from optimizing */
2996 barrier();
2997
2998 /*
2999 * If an interrupt came in just after the first while loop
3000 * and pushed the tail page forward, we will be left with
3001 * a dangling commit that will never go forward.
3002 */
3003 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
3004 goto again;
3005}
3006
3007static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
3008{
3009 unsigned long commits;
3010
3011 if (RB_WARN_ON(cpu_buffer,
3012 !local_read(&cpu_buffer->committing)))
3013 return;
3014
3015 again:
3016 commits = local_read(&cpu_buffer->commits);
3017 /* synchronize with interrupts */
3018 barrier();
3019 if (local_read(&cpu_buffer->committing) == 1)
3020 rb_set_commit_to_write(cpu_buffer);
3021
3022 local_dec(&cpu_buffer->committing);
3023
3024 /* synchronize with interrupts */
3025 barrier();
3026
3027 /*
3028 * Need to account for interrupts coming in between the
3029 * updating of the commit page and the clearing of the
3030 * committing counter.
3031 */
3032 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
3033 !local_read(&cpu_buffer->committing)) {
3034 local_inc(&cpu_buffer->committing);
3035 goto again;
3036 }
3037}
3038
3039static inline void rb_event_discard(struct ring_buffer_event *event)
3040{
3041 if (extended_time(event))
3042 event = skip_time_extend(event);
3043
3044 /* array[0] holds the actual length for the discarded event */
3045 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
3046 event->type_len = RINGBUF_TYPE_PADDING;
3047 /* time delta must be non zero */
3048 if (!event->time_delta)
3049 event->time_delta = 1;
3050}
3051
3052static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer)
3053{
3054 local_inc(&cpu_buffer->entries);
3055 rb_end_commit(cpu_buffer);
3056}
3057
3058static __always_inline void
3059rb_wakeups(struct trace_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
3060{
3061 if (buffer->irq_work.waiters_pending) {
3062 buffer->irq_work.waiters_pending = false;
3063 /* irq_work_queue() supplies it's own memory barriers */
3064 irq_work_queue(&buffer->irq_work.work);
3065 }
3066
3067 if (cpu_buffer->irq_work.waiters_pending) {
3068 cpu_buffer->irq_work.waiters_pending = false;
3069 /* irq_work_queue() supplies it's own memory barriers */
3070 irq_work_queue(&cpu_buffer->irq_work.work);
3071 }
3072
3073 if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
3074 return;
3075
3076 if (cpu_buffer->reader_page == cpu_buffer->commit_page)
3077 return;
3078
3079 if (!cpu_buffer->irq_work.full_waiters_pending)
3080 return;
3081
3082 cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
3083
3084 if (!full_hit(buffer, cpu_buffer->cpu, cpu_buffer->shortest_full))
3085 return;
3086
3087 cpu_buffer->irq_work.wakeup_full = true;
3088 cpu_buffer->irq_work.full_waiters_pending = false;
3089 /* irq_work_queue() supplies it's own memory barriers */
3090 irq_work_queue(&cpu_buffer->irq_work.work);
3091}
3092
3093#ifdef CONFIG_RING_BUFFER_RECORD_RECURSION
3094# define do_ring_buffer_record_recursion() \
3095 do_ftrace_record_recursion(_THIS_IP_, _RET_IP_)
3096#else
3097# define do_ring_buffer_record_recursion() do { } while (0)
3098#endif
3099
3100/*
3101 * The lock and unlock are done within a preempt disable section.
3102 * The current_context per_cpu variable can only be modified
3103 * by the current task between lock and unlock. But it can
3104 * be modified more than once via an interrupt. To pass this
3105 * information from the lock to the unlock without having to
3106 * access the 'in_interrupt()' functions again (which do show
3107 * a bit of overhead in something as critical as function tracing,
3108 * we use a bitmask trick.
3109 *
3110 * bit 1 = NMI context
3111 * bit 2 = IRQ context
3112 * bit 3 = SoftIRQ context
3113 * bit 4 = normal context.
3114 *
3115 * This works because this is the order of contexts that can
3116 * preempt other contexts. A SoftIRQ never preempts an IRQ
3117 * context.
3118 *
3119 * When the context is determined, the corresponding bit is
3120 * checked and set (if it was set, then a recursion of that context
3121 * happened).
3122 *
3123 * On unlock, we need to clear this bit. To do so, just subtract
3124 * 1 from the current_context and AND it to itself.
3125 *
3126 * (binary)
3127 * 101 - 1 = 100
3128 * 101 & 100 = 100 (clearing bit zero)
3129 *
3130 * 1010 - 1 = 1001
3131 * 1010 & 1001 = 1000 (clearing bit 1)
3132 *
3133 * The least significant bit can be cleared this way, and it
3134 * just so happens that it is the same bit corresponding to
3135 * the current context.
3136 *
3137 * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
3138 * is set when a recursion is detected at the current context, and if
3139 * the TRANSITION bit is already set, it will fail the recursion.
3140 * This is needed because there's a lag between the changing of
3141 * interrupt context and updating the preempt count. In this case,
3142 * a false positive will be found. To handle this, one extra recursion
3143 * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
3144 * bit is already set, then it is considered a recursion and the function
3145 * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
3146 *
3147 * On the trace_recursive_unlock(), the TRANSITION bit will be the first
3148 * to be cleared. Even if it wasn't the context that set it. That is,
3149 * if an interrupt comes in while NORMAL bit is set and the ring buffer
3150 * is called before preempt_count() is updated, since the check will
3151 * be on the NORMAL bit, the TRANSITION bit will then be set. If an
3152 * NMI then comes in, it will set the NMI bit, but when the NMI code
3153 * does the trace_recursive_unlock() it will clear the TRANSITION bit
3154 * and leave the NMI bit set. But this is fine, because the interrupt
3155 * code that set the TRANSITION bit will then clear the NMI bit when it
3156 * calls trace_recursive_unlock(). If another NMI comes in, it will
3157 * set the TRANSITION bit and continue.
3158 *
3159 * Note: The TRANSITION bit only handles a single transition between context.
3160 */
3161
3162static __always_inline bool
3163trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
3164{
3165 unsigned int val = cpu_buffer->current_context;
3166 int bit = interrupt_context_level();
3167
3168 bit = RB_CTX_NORMAL - bit;
3169
3170 if (unlikely(val & (1 << (bit + cpu_buffer->nest)))) {
3171 /*
3172 * It is possible that this was called by transitioning
3173 * between interrupt context, and preempt_count() has not
3174 * been updated yet. In this case, use the TRANSITION bit.
3175 */
3176 bit = RB_CTX_TRANSITION;
3177 if (val & (1 << (bit + cpu_buffer->nest))) {
3178 do_ring_buffer_record_recursion();
3179 return true;
3180 }
3181 }
3182
3183 val |= (1 << (bit + cpu_buffer->nest));
3184 cpu_buffer->current_context = val;
3185
3186 return false;
3187}
3188
3189static __always_inline void
3190trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
3191{
3192 cpu_buffer->current_context &=
3193 cpu_buffer->current_context - (1 << cpu_buffer->nest);
3194}
3195
3196/* The recursive locking above uses 5 bits */
3197#define NESTED_BITS 5
3198
3199/**
3200 * ring_buffer_nest_start - Allow to trace while nested
3201 * @buffer: The ring buffer to modify
3202 *
3203 * The ring buffer has a safety mechanism to prevent recursion.
3204 * But there may be a case where a trace needs to be done while
3205 * tracing something else. In this case, calling this function
3206 * will allow this function to nest within a currently active
3207 * ring_buffer_lock_reserve().
3208 *
3209 * Call this function before calling another ring_buffer_lock_reserve() and
3210 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
3211 */
3212void ring_buffer_nest_start(struct trace_buffer *buffer)
3213{
3214 struct ring_buffer_per_cpu *cpu_buffer;
3215 int cpu;
3216
3217 /* Enabled by ring_buffer_nest_end() */
3218 preempt_disable_notrace();
3219 cpu = raw_smp_processor_id();
3220 cpu_buffer = buffer->buffers[cpu];
3221 /* This is the shift value for the above recursive locking */
3222 cpu_buffer->nest += NESTED_BITS;
3223}
3224
3225/**
3226 * ring_buffer_nest_end - Allow to trace while nested
3227 * @buffer: The ring buffer to modify
3228 *
3229 * Must be called after ring_buffer_nest_start() and after the
3230 * ring_buffer_unlock_commit().
3231 */
3232void ring_buffer_nest_end(struct trace_buffer *buffer)
3233{
3234 struct ring_buffer_per_cpu *cpu_buffer;
3235 int cpu;
3236
3237 /* disabled by ring_buffer_nest_start() */
3238 cpu = raw_smp_processor_id();
3239 cpu_buffer = buffer->buffers[cpu];
3240 /* This is the shift value for the above recursive locking */
3241 cpu_buffer->nest -= NESTED_BITS;
3242 preempt_enable_notrace();
3243}
3244
3245/**
3246 * ring_buffer_unlock_commit - commit a reserved
3247 * @buffer: The buffer to commit to
3248 *
3249 * This commits the data to the ring buffer, and releases any locks held.
3250 *
3251 * Must be paired with ring_buffer_lock_reserve.
3252 */
3253int ring_buffer_unlock_commit(struct trace_buffer *buffer)
3254{
3255 struct ring_buffer_per_cpu *cpu_buffer;
3256 int cpu = raw_smp_processor_id();
3257
3258 cpu_buffer = buffer->buffers[cpu];
3259
3260 rb_commit(cpu_buffer);
3261
3262 rb_wakeups(buffer, cpu_buffer);
3263
3264 trace_recursive_unlock(cpu_buffer);
3265
3266 preempt_enable_notrace();
3267
3268 return 0;
3269}
3270EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
3271
3272/* Special value to validate all deltas on a page. */
3273#define CHECK_FULL_PAGE 1L
3274
3275#ifdef CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS
3276
3277static const char *show_irq_str(int bits)
3278{
3279 const char *type[] = {
3280 ".", // 0
3281 "s", // 1
3282 "h", // 2
3283 "Hs", // 3
3284 "n", // 4
3285 "Ns", // 5
3286 "Nh", // 6
3287 "NHs", // 7
3288 };
3289
3290 return type[bits];
3291}
3292
3293/* Assume this is an trace event */
3294static const char *show_flags(struct ring_buffer_event *event)
3295{
3296 struct trace_entry *entry;
3297 int bits = 0;
3298
3299 if (rb_event_data_length(event) - RB_EVNT_HDR_SIZE < sizeof(*entry))
3300 return "X";
3301
3302 entry = ring_buffer_event_data(event);
3303
3304 if (entry->flags & TRACE_FLAG_SOFTIRQ)
3305 bits |= 1;
3306
3307 if (entry->flags & TRACE_FLAG_HARDIRQ)
3308 bits |= 2;
3309
3310 if (entry->flags & TRACE_FLAG_NMI)
3311 bits |= 4;
3312
3313 return show_irq_str(bits);
3314}
3315
3316static const char *show_irq(struct ring_buffer_event *event)
3317{
3318 struct trace_entry *entry;
3319
3320 if (rb_event_data_length(event) - RB_EVNT_HDR_SIZE < sizeof(*entry))
3321 return "";
3322
3323 entry = ring_buffer_event_data(event);
3324 if (entry->flags & TRACE_FLAG_IRQS_OFF)
3325 return "d";
3326 return "";
3327}
3328
3329static const char *show_interrupt_level(void)
3330{
3331 unsigned long pc = preempt_count();
3332 unsigned char level = 0;
3333
3334 if (pc & SOFTIRQ_OFFSET)
3335 level |= 1;
3336
3337 if (pc & HARDIRQ_MASK)
3338 level |= 2;
3339
3340 if (pc & NMI_MASK)
3341 level |= 4;
3342
3343 return show_irq_str(level);
3344}
3345
3346static void dump_buffer_page(struct buffer_data_page *bpage,
3347 struct rb_event_info *info,
3348 unsigned long tail)
3349{
3350 struct ring_buffer_event *event;
3351 u64 ts, delta;
3352 int e;
3353
3354 ts = bpage->time_stamp;
3355 pr_warn(" [%lld] PAGE TIME STAMP\n", ts);
3356
3357 for (e = 0; e < tail; e += rb_event_length(event)) {
3358
3359 event = (struct ring_buffer_event *)(bpage->data + e);
3360
3361 switch (event->type_len) {
3362
3363 case RINGBUF_TYPE_TIME_EXTEND:
3364 delta = rb_event_time_stamp(event);
3365 ts += delta;
3366 pr_warn(" 0x%x: [%lld] delta:%lld TIME EXTEND\n",
3367 e, ts, delta);
3368 break;
3369
3370 case RINGBUF_TYPE_TIME_STAMP:
3371 delta = rb_event_time_stamp(event);
3372 ts = rb_fix_abs_ts(delta, ts);
3373 pr_warn(" 0x%x: [%lld] absolute:%lld TIME STAMP\n",
3374 e, ts, delta);
3375 break;
3376
3377 case RINGBUF_TYPE_PADDING:
3378 ts += event->time_delta;
3379 pr_warn(" 0x%x: [%lld] delta:%d PADDING\n",
3380 e, ts, event->time_delta);
3381 break;
3382
3383 case RINGBUF_TYPE_DATA:
3384 ts += event->time_delta;
3385 pr_warn(" 0x%x: [%lld] delta:%d %s%s\n",
3386 e, ts, event->time_delta,
3387 show_flags(event), show_irq(event));
3388 break;
3389
3390 default:
3391 break;
3392 }
3393 }
3394 pr_warn("expected end:0x%lx last event actually ended at:0x%x\n", tail, e);
3395}
3396
3397static DEFINE_PER_CPU(atomic_t, checking);
3398static atomic_t ts_dump;
3399
3400#define buffer_warn_return(fmt, ...) \
3401 do { \
3402 /* If another report is happening, ignore this one */ \
3403 if (atomic_inc_return(&ts_dump) != 1) { \
3404 atomic_dec(&ts_dump); \
3405 goto out; \
3406 } \
3407 atomic_inc(&cpu_buffer->record_disabled); \
3408 pr_warn(fmt, ##__VA_ARGS__); \
3409 dump_buffer_page(bpage, info, tail); \
3410 atomic_dec(&ts_dump); \
3411 /* There's some cases in boot up that this can happen */ \
3412 if (WARN_ON_ONCE(system_state != SYSTEM_BOOTING)) \
3413 /* Do not re-enable checking */ \
3414 return; \
3415 } while (0)
3416
3417/*
3418 * Check if the current event time stamp matches the deltas on
3419 * the buffer page.
3420 */
3421static void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3422 struct rb_event_info *info,
3423 unsigned long tail)
3424{
3425 struct ring_buffer_event *event;
3426 struct buffer_data_page *bpage;
3427 u64 ts, delta;
3428 bool full = false;
3429 int e;
3430
3431 bpage = info->tail_page->page;
3432
3433 if (tail == CHECK_FULL_PAGE) {
3434 full = true;
3435 tail = local_read(&bpage->commit);
3436 } else if (info->add_timestamp &
3437 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE)) {
3438 /* Ignore events with absolute time stamps */
3439 return;
3440 }
3441
3442 /*
3443 * Do not check the first event (skip possible extends too).
3444 * Also do not check if previous events have not been committed.
3445 */
3446 if (tail <= 8 || tail > local_read(&bpage->commit))
3447 return;
3448
3449 /*
3450 * If this interrupted another event,
3451 */
3452 if (atomic_inc_return(this_cpu_ptr(&checking)) != 1)
3453 goto out;
3454
3455 ts = bpage->time_stamp;
3456
3457 for (e = 0; e < tail; e += rb_event_length(event)) {
3458
3459 event = (struct ring_buffer_event *)(bpage->data + e);
3460
3461 switch (event->type_len) {
3462
3463 case RINGBUF_TYPE_TIME_EXTEND:
3464 delta = rb_event_time_stamp(event);
3465 ts += delta;
3466 break;
3467
3468 case RINGBUF_TYPE_TIME_STAMP:
3469 delta = rb_event_time_stamp(event);
3470 delta = rb_fix_abs_ts(delta, ts);
3471 if (delta < ts) {
3472 buffer_warn_return("[CPU: %d]ABSOLUTE TIME WENT BACKWARDS: last ts: %lld absolute ts: %lld\n",
3473 cpu_buffer->cpu, ts, delta);
3474 }
3475 ts = delta;
3476 break;
3477
3478 case RINGBUF_TYPE_PADDING:
3479 if (event->time_delta == 1)
3480 break;
3481 fallthrough;
3482 case RINGBUF_TYPE_DATA:
3483 ts += event->time_delta;
3484 break;
3485
3486 default:
3487 RB_WARN_ON(cpu_buffer, 1);
3488 }
3489 }
3490 if ((full && ts > info->ts) ||
3491 (!full && ts + info->delta != info->ts)) {
3492 buffer_warn_return("[CPU: %d]TIME DOES NOT MATCH expected:%lld actual:%lld delta:%lld before:%lld after:%lld%s context:%s\n",
3493 cpu_buffer->cpu,
3494 ts + info->delta, info->ts, info->delta,
3495 info->before, info->after,
3496 full ? " (full)" : "", show_interrupt_level());
3497 }
3498out:
3499 atomic_dec(this_cpu_ptr(&checking));
3500}
3501#else
3502static inline void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3503 struct rb_event_info *info,
3504 unsigned long tail)
3505{
3506}
3507#endif /* CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS */
3508
3509static struct ring_buffer_event *
3510__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
3511 struct rb_event_info *info)
3512{
3513 struct ring_buffer_event *event;
3514 struct buffer_page *tail_page;
3515 unsigned long tail, write, w;
3516
3517 /* Don't let the compiler play games with cpu_buffer->tail_page */
3518 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
3519
3520 /*A*/ w = local_read(&tail_page->write) & RB_WRITE_MASK;
3521 barrier();
3522 rb_time_read(&cpu_buffer->before_stamp, &info->before);
3523 rb_time_read(&cpu_buffer->write_stamp, &info->after);
3524 barrier();
3525 info->ts = rb_time_stamp(cpu_buffer->buffer);
3526
3527 if ((info->add_timestamp & RB_ADD_STAMP_ABSOLUTE)) {
3528 info->delta = info->ts;
3529 } else {
3530 /*
3531 * If interrupting an event time update, we may need an
3532 * absolute timestamp.
3533 * Don't bother if this is the start of a new page (w == 0).
3534 */
3535 if (!w) {
3536 /* Use the sub-buffer timestamp */
3537 info->delta = 0;
3538 } else if (unlikely(info->before != info->after)) {
3539 info->add_timestamp |= RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND;
3540 info->length += RB_LEN_TIME_EXTEND;
3541 } else {
3542 info->delta = info->ts - info->after;
3543 if (unlikely(test_time_stamp(info->delta))) {
3544 info->add_timestamp |= RB_ADD_STAMP_EXTEND;
3545 info->length += RB_LEN_TIME_EXTEND;
3546 }
3547 }
3548 }
3549
3550 /*B*/ rb_time_set(&cpu_buffer->before_stamp, info->ts);
3551
3552 /*C*/ write = local_add_return(info->length, &tail_page->write);
3553
3554 /* set write to only the index of the write */
3555 write &= RB_WRITE_MASK;
3556
3557 tail = write - info->length;
3558
3559 /* See if we shot pass the end of this buffer page */
3560 if (unlikely(write > cpu_buffer->buffer->subbuf_size)) {
3561 check_buffer(cpu_buffer, info, CHECK_FULL_PAGE);
3562 return rb_move_tail(cpu_buffer, tail, info);
3563 }
3564
3565 if (likely(tail == w)) {
3566 /* Nothing interrupted us between A and C */
3567 /*D*/ rb_time_set(&cpu_buffer->write_stamp, info->ts);
3568 /*
3569 * If something came in between C and D, the write stamp
3570 * may now not be in sync. But that's fine as the before_stamp
3571 * will be different and then next event will just be forced
3572 * to use an absolute timestamp.
3573 */
3574 if (likely(!(info->add_timestamp &
3575 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3576 /* This did not interrupt any time update */
3577 info->delta = info->ts - info->after;
3578 else
3579 /* Just use full timestamp for interrupting event */
3580 info->delta = info->ts;
3581 check_buffer(cpu_buffer, info, tail);
3582 } else {
3583 u64 ts;
3584 /* SLOW PATH - Interrupted between A and C */
3585
3586 /* Save the old before_stamp */
3587 rb_time_read(&cpu_buffer->before_stamp, &info->before);
3588
3589 /*
3590 * Read a new timestamp and update the before_stamp to make
3591 * the next event after this one force using an absolute
3592 * timestamp. This is in case an interrupt were to come in
3593 * between E and F.
3594 */
3595 ts = rb_time_stamp(cpu_buffer->buffer);
3596 rb_time_set(&cpu_buffer->before_stamp, ts);
3597
3598 barrier();
3599 /*E*/ rb_time_read(&cpu_buffer->write_stamp, &info->after);
3600 barrier();
3601 /*F*/ if (write == (local_read(&tail_page->write) & RB_WRITE_MASK) &&
3602 info->after == info->before && info->after < ts) {
3603 /*
3604 * Nothing came after this event between C and F, it is
3605 * safe to use info->after for the delta as it
3606 * matched info->before and is still valid.
3607 */
3608 info->delta = ts - info->after;
3609 } else {
3610 /*
3611 * Interrupted between C and F:
3612 * Lost the previous events time stamp. Just set the
3613 * delta to zero, and this will be the same time as
3614 * the event this event interrupted. And the events that
3615 * came after this will still be correct (as they would
3616 * have built their delta on the previous event.
3617 */
3618 info->delta = 0;
3619 }
3620 info->ts = ts;
3621 info->add_timestamp &= ~RB_ADD_STAMP_FORCE;
3622 }
3623
3624 /*
3625 * If this is the first commit on the page, then it has the same
3626 * timestamp as the page itself.
3627 */
3628 if (unlikely(!tail && !(info->add_timestamp &
3629 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3630 info->delta = 0;
3631
3632 /* We reserved something on the buffer */
3633
3634 event = __rb_page_index(tail_page, tail);
3635 rb_update_event(cpu_buffer, event, info);
3636
3637 local_inc(&tail_page->entries);
3638
3639 /*
3640 * If this is the first commit on the page, then update
3641 * its timestamp.
3642 */
3643 if (unlikely(!tail))
3644 tail_page->page->time_stamp = info->ts;
3645
3646 /* account for these added bytes */
3647 local_add(info->length, &cpu_buffer->entries_bytes);
3648
3649 return event;
3650}
3651
3652static __always_inline struct ring_buffer_event *
3653rb_reserve_next_event(struct trace_buffer *buffer,
3654 struct ring_buffer_per_cpu *cpu_buffer,
3655 unsigned long length)
3656{
3657 struct ring_buffer_event *event;
3658 struct rb_event_info info;
3659 int nr_loops = 0;
3660 int add_ts_default;
3661
3662 /* ring buffer does cmpxchg, make sure it is safe in NMI context */
3663 if (!IS_ENABLED(CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG) &&
3664 (unlikely(in_nmi()))) {
3665 return NULL;
3666 }
3667
3668 rb_start_commit(cpu_buffer);
3669 /* The commit page can not change after this */
3670
3671#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3672 /*
3673 * Due to the ability to swap a cpu buffer from a buffer
3674 * it is possible it was swapped before we committed.
3675 * (committing stops a swap). We check for it here and
3676 * if it happened, we have to fail the write.
3677 */
3678 barrier();
3679 if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
3680 local_dec(&cpu_buffer->committing);
3681 local_dec(&cpu_buffer->commits);
3682 return NULL;
3683 }
3684#endif
3685
3686 info.length = rb_calculate_event_length(length);
3687
3688 if (ring_buffer_time_stamp_abs(cpu_buffer->buffer)) {
3689 add_ts_default = RB_ADD_STAMP_ABSOLUTE;
3690 info.length += RB_LEN_TIME_EXTEND;
3691 if (info.length > cpu_buffer->buffer->max_data_size)
3692 goto out_fail;
3693 } else {
3694 add_ts_default = RB_ADD_STAMP_NONE;
3695 }
3696
3697 again:
3698 info.add_timestamp = add_ts_default;
3699 info.delta = 0;
3700
3701 /*
3702 * We allow for interrupts to reenter here and do a trace.
3703 * If one does, it will cause this original code to loop
3704 * back here. Even with heavy interrupts happening, this
3705 * should only happen a few times in a row. If this happens
3706 * 1000 times in a row, there must be either an interrupt
3707 * storm or we have something buggy.
3708 * Bail!
3709 */
3710 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
3711 goto out_fail;
3712
3713 event = __rb_reserve_next(cpu_buffer, &info);
3714
3715 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
3716 if (info.add_timestamp & (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND))
3717 info.length -= RB_LEN_TIME_EXTEND;
3718 goto again;
3719 }
3720
3721 if (likely(event))
3722 return event;
3723 out_fail:
3724 rb_end_commit(cpu_buffer);
3725 return NULL;
3726}
3727
3728/**
3729 * ring_buffer_lock_reserve - reserve a part of the buffer
3730 * @buffer: the ring buffer to reserve from
3731 * @length: the length of the data to reserve (excluding event header)
3732 *
3733 * Returns a reserved event on the ring buffer to copy directly to.
3734 * The user of this interface will need to get the body to write into
3735 * and can use the ring_buffer_event_data() interface.
3736 *
3737 * The length is the length of the data needed, not the event length
3738 * which also includes the event header.
3739 *
3740 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
3741 * If NULL is returned, then nothing has been allocated or locked.
3742 */
3743struct ring_buffer_event *
3744ring_buffer_lock_reserve(struct trace_buffer *buffer, unsigned long length)
3745{
3746 struct ring_buffer_per_cpu *cpu_buffer;
3747 struct ring_buffer_event *event;
3748 int cpu;
3749
3750 /* If we are tracing schedule, we don't want to recurse */
3751 preempt_disable_notrace();
3752
3753 if (unlikely(atomic_read(&buffer->record_disabled)))
3754 goto out;
3755
3756 cpu = raw_smp_processor_id();
3757
3758 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
3759 goto out;
3760
3761 cpu_buffer = buffer->buffers[cpu];
3762
3763 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
3764 goto out;
3765
3766 if (unlikely(length > buffer->max_data_size))
3767 goto out;
3768
3769 if (unlikely(trace_recursive_lock(cpu_buffer)))
3770 goto out;
3771
3772 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3773 if (!event)
3774 goto out_unlock;
3775
3776 return event;
3777
3778 out_unlock:
3779 trace_recursive_unlock(cpu_buffer);
3780 out:
3781 preempt_enable_notrace();
3782 return NULL;
3783}
3784EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3785
3786/*
3787 * Decrement the entries to the page that an event is on.
3788 * The event does not even need to exist, only the pointer
3789 * to the page it is on. This may only be called before the commit
3790 * takes place.
3791 */
3792static inline void
3793rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3794 struct ring_buffer_event *event)
3795{
3796 unsigned long addr = (unsigned long)event;
3797 struct buffer_page *bpage = cpu_buffer->commit_page;
3798 struct buffer_page *start;
3799
3800 addr &= ~((PAGE_SIZE << cpu_buffer->buffer->subbuf_order) - 1);
3801
3802 /* Do the likely case first */
3803 if (likely(bpage->page == (void *)addr)) {
3804 local_dec(&bpage->entries);
3805 return;
3806 }
3807
3808 /*
3809 * Because the commit page may be on the reader page we
3810 * start with the next page and check the end loop there.
3811 */
3812 rb_inc_page(&bpage);
3813 start = bpage;
3814 do {
3815 if (bpage->page == (void *)addr) {
3816 local_dec(&bpage->entries);
3817 return;
3818 }
3819 rb_inc_page(&bpage);
3820 } while (bpage != start);
3821
3822 /* commit not part of this buffer?? */
3823 RB_WARN_ON(cpu_buffer, 1);
3824}
3825
3826/**
3827 * ring_buffer_discard_commit - discard an event that has not been committed
3828 * @buffer: the ring buffer
3829 * @event: non committed event to discard
3830 *
3831 * Sometimes an event that is in the ring buffer needs to be ignored.
3832 * This function lets the user discard an event in the ring buffer
3833 * and then that event will not be read later.
3834 *
3835 * This function only works if it is called before the item has been
3836 * committed. It will try to free the event from the ring buffer
3837 * if another event has not been added behind it.
3838 *
3839 * If another event has been added behind it, it will set the event
3840 * up as discarded, and perform the commit.
3841 *
3842 * If this function is called, do not call ring_buffer_unlock_commit on
3843 * the event.
3844 */
3845void ring_buffer_discard_commit(struct trace_buffer *buffer,
3846 struct ring_buffer_event *event)
3847{
3848 struct ring_buffer_per_cpu *cpu_buffer;
3849 int cpu;
3850
3851 /* The event is discarded regardless */
3852 rb_event_discard(event);
3853
3854 cpu = smp_processor_id();
3855 cpu_buffer = buffer->buffers[cpu];
3856
3857 /*
3858 * This must only be called if the event has not been
3859 * committed yet. Thus we can assume that preemption
3860 * is still disabled.
3861 */
3862 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3863
3864 rb_decrement_entry(cpu_buffer, event);
3865 if (rb_try_to_discard(cpu_buffer, event))
3866 goto out;
3867
3868 out:
3869 rb_end_commit(cpu_buffer);
3870
3871 trace_recursive_unlock(cpu_buffer);
3872
3873 preempt_enable_notrace();
3874
3875}
3876EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3877
3878/**
3879 * ring_buffer_write - write data to the buffer without reserving
3880 * @buffer: The ring buffer to write to.
3881 * @length: The length of the data being written (excluding the event header)
3882 * @data: The data to write to the buffer.
3883 *
3884 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3885 * one function. If you already have the data to write to the buffer, it
3886 * may be easier to simply call this function.
3887 *
3888 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3889 * and not the length of the event which would hold the header.
3890 */
3891int ring_buffer_write(struct trace_buffer *buffer,
3892 unsigned long length,
3893 void *data)
3894{
3895 struct ring_buffer_per_cpu *cpu_buffer;
3896 struct ring_buffer_event *event;
3897 void *body;
3898 int ret = -EBUSY;
3899 int cpu;
3900
3901 preempt_disable_notrace();
3902
3903 if (atomic_read(&buffer->record_disabled))
3904 goto out;
3905
3906 cpu = raw_smp_processor_id();
3907
3908 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3909 goto out;
3910
3911 cpu_buffer = buffer->buffers[cpu];
3912
3913 if (atomic_read(&cpu_buffer->record_disabled))
3914 goto out;
3915
3916 if (length > buffer->max_data_size)
3917 goto out;
3918
3919 if (unlikely(trace_recursive_lock(cpu_buffer)))
3920 goto out;
3921
3922 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3923 if (!event)
3924 goto out_unlock;
3925
3926 body = rb_event_data(event);
3927
3928 memcpy(body, data, length);
3929
3930 rb_commit(cpu_buffer);
3931
3932 rb_wakeups(buffer, cpu_buffer);
3933
3934 ret = 0;
3935
3936 out_unlock:
3937 trace_recursive_unlock(cpu_buffer);
3938
3939 out:
3940 preempt_enable_notrace();
3941
3942 return ret;
3943}
3944EXPORT_SYMBOL_GPL(ring_buffer_write);
3945
3946static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3947{
3948 struct buffer_page *reader = cpu_buffer->reader_page;
3949 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3950 struct buffer_page *commit = cpu_buffer->commit_page;
3951
3952 /* In case of error, head will be NULL */
3953 if (unlikely(!head))
3954 return true;
3955
3956 /* Reader should exhaust content in reader page */
3957 if (reader->read != rb_page_commit(reader))
3958 return false;
3959
3960 /*
3961 * If writers are committing on the reader page, knowing all
3962 * committed content has been read, the ring buffer is empty.
3963 */
3964 if (commit == reader)
3965 return true;
3966
3967 /*
3968 * If writers are committing on a page other than reader page
3969 * and head page, there should always be content to read.
3970 */
3971 if (commit != head)
3972 return false;
3973
3974 /*
3975 * Writers are committing on the head page, we just need
3976 * to care about there're committed data, and the reader will
3977 * swap reader page with head page when it is to read data.
3978 */
3979 return rb_page_commit(commit) == 0;
3980}
3981
3982/**
3983 * ring_buffer_record_disable - stop all writes into the buffer
3984 * @buffer: The ring buffer to stop writes to.
3985 *
3986 * This prevents all writes to the buffer. Any attempt to write
3987 * to the buffer after this will fail and return NULL.
3988 *
3989 * The caller should call synchronize_rcu() after this.
3990 */
3991void ring_buffer_record_disable(struct trace_buffer *buffer)
3992{
3993 atomic_inc(&buffer->record_disabled);
3994}
3995EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3996
3997/**
3998 * ring_buffer_record_enable - enable writes to the buffer
3999 * @buffer: The ring buffer to enable writes
4000 *
4001 * Note, multiple disables will need the same number of enables
4002 * to truly enable the writing (much like preempt_disable).
4003 */
4004void ring_buffer_record_enable(struct trace_buffer *buffer)
4005{
4006 atomic_dec(&buffer->record_disabled);
4007}
4008EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
4009
4010/**
4011 * ring_buffer_record_off - stop all writes into the buffer
4012 * @buffer: The ring buffer to stop writes to.
4013 *
4014 * This prevents all writes to the buffer. Any attempt to write
4015 * to the buffer after this will fail and return NULL.
4016 *
4017 * This is different than ring_buffer_record_disable() as
4018 * it works like an on/off switch, where as the disable() version
4019 * must be paired with a enable().
4020 */
4021void ring_buffer_record_off(struct trace_buffer *buffer)
4022{
4023 unsigned int rd;
4024 unsigned int new_rd;
4025
4026 rd = atomic_read(&buffer->record_disabled);
4027 do {
4028 new_rd = rd | RB_BUFFER_OFF;
4029 } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
4030}
4031EXPORT_SYMBOL_GPL(ring_buffer_record_off);
4032
4033/**
4034 * ring_buffer_record_on - restart writes into the buffer
4035 * @buffer: The ring buffer to start writes to.
4036 *
4037 * This enables all writes to the buffer that was disabled by
4038 * ring_buffer_record_off().
4039 *
4040 * This is different than ring_buffer_record_enable() as
4041 * it works like an on/off switch, where as the enable() version
4042 * must be paired with a disable().
4043 */
4044void ring_buffer_record_on(struct trace_buffer *buffer)
4045{
4046 unsigned int rd;
4047 unsigned int new_rd;
4048
4049 rd = atomic_read(&buffer->record_disabled);
4050 do {
4051 new_rd = rd & ~RB_BUFFER_OFF;
4052 } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
4053}
4054EXPORT_SYMBOL_GPL(ring_buffer_record_on);
4055
4056/**
4057 * ring_buffer_record_is_on - return true if the ring buffer can write
4058 * @buffer: The ring buffer to see if write is enabled
4059 *
4060 * Returns true if the ring buffer is in a state that it accepts writes.
4061 */
4062bool ring_buffer_record_is_on(struct trace_buffer *buffer)
4063{
4064 return !atomic_read(&buffer->record_disabled);
4065}
4066
4067/**
4068 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
4069 * @buffer: The ring buffer to see if write is set enabled
4070 *
4071 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
4072 * Note that this does NOT mean it is in a writable state.
4073 *
4074 * It may return true when the ring buffer has been disabled by
4075 * ring_buffer_record_disable(), as that is a temporary disabling of
4076 * the ring buffer.
4077 */
4078bool ring_buffer_record_is_set_on(struct trace_buffer *buffer)
4079{
4080 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
4081}
4082
4083/**
4084 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
4085 * @buffer: The ring buffer to stop writes to.
4086 * @cpu: The CPU buffer to stop
4087 *
4088 * This prevents all writes to the buffer. Any attempt to write
4089 * to the buffer after this will fail and return NULL.
4090 *
4091 * The caller should call synchronize_rcu() after this.
4092 */
4093void ring_buffer_record_disable_cpu(struct trace_buffer *buffer, int cpu)
4094{
4095 struct ring_buffer_per_cpu *cpu_buffer;
4096
4097 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4098 return;
4099
4100 cpu_buffer = buffer->buffers[cpu];
4101 atomic_inc(&cpu_buffer->record_disabled);
4102}
4103EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
4104
4105/**
4106 * ring_buffer_record_enable_cpu - enable writes to the buffer
4107 * @buffer: The ring buffer to enable writes
4108 * @cpu: The CPU to enable.
4109 *
4110 * Note, multiple disables will need the same number of enables
4111 * to truly enable the writing (much like preempt_disable).
4112 */
4113void ring_buffer_record_enable_cpu(struct trace_buffer *buffer, int cpu)
4114{
4115 struct ring_buffer_per_cpu *cpu_buffer;
4116
4117 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4118 return;
4119
4120 cpu_buffer = buffer->buffers[cpu];
4121 atomic_dec(&cpu_buffer->record_disabled);
4122}
4123EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
4124
4125/*
4126 * The total entries in the ring buffer is the running counter
4127 * of entries entered into the ring buffer, minus the sum of
4128 * the entries read from the ring buffer and the number of
4129 * entries that were overwritten.
4130 */
4131static inline unsigned long
4132rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
4133{
4134 return local_read(&cpu_buffer->entries) -
4135 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
4136}
4137
4138/**
4139 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
4140 * @buffer: The ring buffer
4141 * @cpu: The per CPU buffer to read from.
4142 */
4143u64 ring_buffer_oldest_event_ts(struct trace_buffer *buffer, int cpu)
4144{
4145 unsigned long flags;
4146 struct ring_buffer_per_cpu *cpu_buffer;
4147 struct buffer_page *bpage;
4148 u64 ret = 0;
4149
4150 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4151 return 0;
4152
4153 cpu_buffer = buffer->buffers[cpu];
4154 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4155 /*
4156 * if the tail is on reader_page, oldest time stamp is on the reader
4157 * page
4158 */
4159 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
4160 bpage = cpu_buffer->reader_page;
4161 else
4162 bpage = rb_set_head_page(cpu_buffer);
4163 if (bpage)
4164 ret = bpage->page->time_stamp;
4165 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4166
4167 return ret;
4168}
4169EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
4170
4171/**
4172 * ring_buffer_bytes_cpu - get the number of bytes unconsumed in a cpu buffer
4173 * @buffer: The ring buffer
4174 * @cpu: The per CPU buffer to read from.
4175 */
4176unsigned long ring_buffer_bytes_cpu(struct trace_buffer *buffer, int cpu)
4177{
4178 struct ring_buffer_per_cpu *cpu_buffer;
4179 unsigned long ret;
4180
4181 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4182 return 0;
4183
4184 cpu_buffer = buffer->buffers[cpu];
4185 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
4186
4187 return ret;
4188}
4189EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
4190
4191/**
4192 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
4193 * @buffer: The ring buffer
4194 * @cpu: The per CPU buffer to get the entries from.
4195 */
4196unsigned long ring_buffer_entries_cpu(struct trace_buffer *buffer, int cpu)
4197{
4198 struct ring_buffer_per_cpu *cpu_buffer;
4199
4200 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4201 return 0;
4202
4203 cpu_buffer = buffer->buffers[cpu];
4204
4205 return rb_num_of_entries(cpu_buffer);
4206}
4207EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
4208
4209/**
4210 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
4211 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
4212 * @buffer: The ring buffer
4213 * @cpu: The per CPU buffer to get the number of overruns from
4214 */
4215unsigned long ring_buffer_overrun_cpu(struct trace_buffer *buffer, int cpu)
4216{
4217 struct ring_buffer_per_cpu *cpu_buffer;
4218 unsigned long ret;
4219
4220 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4221 return 0;
4222
4223 cpu_buffer = buffer->buffers[cpu];
4224 ret = local_read(&cpu_buffer->overrun);
4225
4226 return ret;
4227}
4228EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
4229
4230/**
4231 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
4232 * commits failing due to the buffer wrapping around while there are uncommitted
4233 * events, such as during an interrupt storm.
4234 * @buffer: The ring buffer
4235 * @cpu: The per CPU buffer to get the number of overruns from
4236 */
4237unsigned long
4238ring_buffer_commit_overrun_cpu(struct trace_buffer *buffer, int cpu)
4239{
4240 struct ring_buffer_per_cpu *cpu_buffer;
4241 unsigned long ret;
4242
4243 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4244 return 0;
4245
4246 cpu_buffer = buffer->buffers[cpu];
4247 ret = local_read(&cpu_buffer->commit_overrun);
4248
4249 return ret;
4250}
4251EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
4252
4253/**
4254 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
4255 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
4256 * @buffer: The ring buffer
4257 * @cpu: The per CPU buffer to get the number of overruns from
4258 */
4259unsigned long
4260ring_buffer_dropped_events_cpu(struct trace_buffer *buffer, int cpu)
4261{
4262 struct ring_buffer_per_cpu *cpu_buffer;
4263 unsigned long ret;
4264
4265 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4266 return 0;
4267
4268 cpu_buffer = buffer->buffers[cpu];
4269 ret = local_read(&cpu_buffer->dropped_events);
4270
4271 return ret;
4272}
4273EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
4274
4275/**
4276 * ring_buffer_read_events_cpu - get the number of events successfully read
4277 * @buffer: The ring buffer
4278 * @cpu: The per CPU buffer to get the number of events read
4279 */
4280unsigned long
4281ring_buffer_read_events_cpu(struct trace_buffer *buffer, int cpu)
4282{
4283 struct ring_buffer_per_cpu *cpu_buffer;
4284
4285 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4286 return 0;
4287
4288 cpu_buffer = buffer->buffers[cpu];
4289 return cpu_buffer->read;
4290}
4291EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
4292
4293/**
4294 * ring_buffer_entries - get the number of entries in a buffer
4295 * @buffer: The ring buffer
4296 *
4297 * Returns the total number of entries in the ring buffer
4298 * (all CPU entries)
4299 */
4300unsigned long ring_buffer_entries(struct trace_buffer *buffer)
4301{
4302 struct ring_buffer_per_cpu *cpu_buffer;
4303 unsigned long entries = 0;
4304 int cpu;
4305
4306 /* if you care about this being correct, lock the buffer */
4307 for_each_buffer_cpu(buffer, cpu) {
4308 cpu_buffer = buffer->buffers[cpu];
4309 entries += rb_num_of_entries(cpu_buffer);
4310 }
4311
4312 return entries;
4313}
4314EXPORT_SYMBOL_GPL(ring_buffer_entries);
4315
4316/**
4317 * ring_buffer_overruns - get the number of overruns in buffer
4318 * @buffer: The ring buffer
4319 *
4320 * Returns the total number of overruns in the ring buffer
4321 * (all CPU entries)
4322 */
4323unsigned long ring_buffer_overruns(struct trace_buffer *buffer)
4324{
4325 struct ring_buffer_per_cpu *cpu_buffer;
4326 unsigned long overruns = 0;
4327 int cpu;
4328
4329 /* if you care about this being correct, lock the buffer */
4330 for_each_buffer_cpu(buffer, cpu) {
4331 cpu_buffer = buffer->buffers[cpu];
4332 overruns += local_read(&cpu_buffer->overrun);
4333 }
4334
4335 return overruns;
4336}
4337EXPORT_SYMBOL_GPL(ring_buffer_overruns);
4338
4339static void rb_iter_reset(struct ring_buffer_iter *iter)
4340{
4341 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4342
4343 /* Iterator usage is expected to have record disabled */
4344 iter->head_page = cpu_buffer->reader_page;
4345 iter->head = cpu_buffer->reader_page->read;
4346 iter->next_event = iter->head;
4347
4348 iter->cache_reader_page = iter->head_page;
4349 iter->cache_read = cpu_buffer->read;
4350 iter->cache_pages_removed = cpu_buffer->pages_removed;
4351
4352 if (iter->head) {
4353 iter->read_stamp = cpu_buffer->read_stamp;
4354 iter->page_stamp = cpu_buffer->reader_page->page->time_stamp;
4355 } else {
4356 iter->read_stamp = iter->head_page->page->time_stamp;
4357 iter->page_stamp = iter->read_stamp;
4358 }
4359}
4360
4361/**
4362 * ring_buffer_iter_reset - reset an iterator
4363 * @iter: The iterator to reset
4364 *
4365 * Resets the iterator, so that it will start from the beginning
4366 * again.
4367 */
4368void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
4369{
4370 struct ring_buffer_per_cpu *cpu_buffer;
4371 unsigned long flags;
4372
4373 if (!iter)
4374 return;
4375
4376 cpu_buffer = iter->cpu_buffer;
4377
4378 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4379 rb_iter_reset(iter);
4380 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4381}
4382EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
4383
4384/**
4385 * ring_buffer_iter_empty - check if an iterator has no more to read
4386 * @iter: The iterator to check
4387 */
4388int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
4389{
4390 struct ring_buffer_per_cpu *cpu_buffer;
4391 struct buffer_page *reader;
4392 struct buffer_page *head_page;
4393 struct buffer_page *commit_page;
4394 struct buffer_page *curr_commit_page;
4395 unsigned commit;
4396 u64 curr_commit_ts;
4397 u64 commit_ts;
4398
4399 cpu_buffer = iter->cpu_buffer;
4400 reader = cpu_buffer->reader_page;
4401 head_page = cpu_buffer->head_page;
4402 commit_page = READ_ONCE(cpu_buffer->commit_page);
4403 commit_ts = commit_page->page->time_stamp;
4404
4405 /*
4406 * When the writer goes across pages, it issues a cmpxchg which
4407 * is a mb(), which will synchronize with the rmb here.
4408 * (see rb_tail_page_update())
4409 */
4410 smp_rmb();
4411 commit = rb_page_commit(commit_page);
4412 /* We want to make sure that the commit page doesn't change */
4413 smp_rmb();
4414
4415 /* Make sure commit page didn't change */
4416 curr_commit_page = READ_ONCE(cpu_buffer->commit_page);
4417 curr_commit_ts = READ_ONCE(curr_commit_page->page->time_stamp);
4418
4419 /* If the commit page changed, then there's more data */
4420 if (curr_commit_page != commit_page ||
4421 curr_commit_ts != commit_ts)
4422 return 0;
4423
4424 /* Still racy, as it may return a false positive, but that's OK */
4425 return ((iter->head_page == commit_page && iter->head >= commit) ||
4426 (iter->head_page == reader && commit_page == head_page &&
4427 head_page->read == commit &&
4428 iter->head == rb_page_commit(cpu_buffer->reader_page)));
4429}
4430EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
4431
4432static void
4433rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
4434 struct ring_buffer_event *event)
4435{
4436 u64 delta;
4437
4438 switch (event->type_len) {
4439 case RINGBUF_TYPE_PADDING:
4440 return;
4441
4442 case RINGBUF_TYPE_TIME_EXTEND:
4443 delta = rb_event_time_stamp(event);
4444 cpu_buffer->read_stamp += delta;
4445 return;
4446
4447 case RINGBUF_TYPE_TIME_STAMP:
4448 delta = rb_event_time_stamp(event);
4449 delta = rb_fix_abs_ts(delta, cpu_buffer->read_stamp);
4450 cpu_buffer->read_stamp = delta;
4451 return;
4452
4453 case RINGBUF_TYPE_DATA:
4454 cpu_buffer->read_stamp += event->time_delta;
4455 return;
4456
4457 default:
4458 RB_WARN_ON(cpu_buffer, 1);
4459 }
4460}
4461
4462static void
4463rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
4464 struct ring_buffer_event *event)
4465{
4466 u64 delta;
4467
4468 switch (event->type_len) {
4469 case RINGBUF_TYPE_PADDING:
4470 return;
4471
4472 case RINGBUF_TYPE_TIME_EXTEND:
4473 delta = rb_event_time_stamp(event);
4474 iter->read_stamp += delta;
4475 return;
4476
4477 case RINGBUF_TYPE_TIME_STAMP:
4478 delta = rb_event_time_stamp(event);
4479 delta = rb_fix_abs_ts(delta, iter->read_stamp);
4480 iter->read_stamp = delta;
4481 return;
4482
4483 case RINGBUF_TYPE_DATA:
4484 iter->read_stamp += event->time_delta;
4485 return;
4486
4487 default:
4488 RB_WARN_ON(iter->cpu_buffer, 1);
4489 }
4490}
4491
4492static struct buffer_page *
4493rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
4494{
4495 struct buffer_page *reader = NULL;
4496 unsigned long bsize = READ_ONCE(cpu_buffer->buffer->subbuf_size);
4497 unsigned long overwrite;
4498 unsigned long flags;
4499 int nr_loops = 0;
4500 bool ret;
4501
4502 local_irq_save(flags);
4503 arch_spin_lock(&cpu_buffer->lock);
4504
4505 again:
4506 /*
4507 * This should normally only loop twice. But because the
4508 * start of the reader inserts an empty page, it causes
4509 * a case where we will loop three times. There should be no
4510 * reason to loop four times (that I know of).
4511 */
4512 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
4513 reader = NULL;
4514 goto out;
4515 }
4516
4517 reader = cpu_buffer->reader_page;
4518
4519 /* If there's more to read, return this page */
4520 if (cpu_buffer->reader_page->read < rb_page_size(reader))
4521 goto out;
4522
4523 /* Never should we have an index greater than the size */
4524 if (RB_WARN_ON(cpu_buffer,
4525 cpu_buffer->reader_page->read > rb_page_size(reader)))
4526 goto out;
4527
4528 /* check if we caught up to the tail */
4529 reader = NULL;
4530 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
4531 goto out;
4532
4533 /* Don't bother swapping if the ring buffer is empty */
4534 if (rb_num_of_entries(cpu_buffer) == 0)
4535 goto out;
4536
4537 /*
4538 * Reset the reader page to size zero.
4539 */
4540 local_set(&cpu_buffer->reader_page->write, 0);
4541 local_set(&cpu_buffer->reader_page->entries, 0);
4542 local_set(&cpu_buffer->reader_page->page->commit, 0);
4543 cpu_buffer->reader_page->real_end = 0;
4544
4545 spin:
4546 /*
4547 * Splice the empty reader page into the list around the head.
4548 */
4549 reader = rb_set_head_page(cpu_buffer);
4550 if (!reader)
4551 goto out;
4552 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
4553 cpu_buffer->reader_page->list.prev = reader->list.prev;
4554
4555 /*
4556 * cpu_buffer->pages just needs to point to the buffer, it
4557 * has no specific buffer page to point to. Lets move it out
4558 * of our way so we don't accidentally swap it.
4559 */
4560 cpu_buffer->pages = reader->list.prev;
4561
4562 /* The reader page will be pointing to the new head */
4563 rb_set_list_to_head(&cpu_buffer->reader_page->list);
4564
4565 /*
4566 * We want to make sure we read the overruns after we set up our
4567 * pointers to the next object. The writer side does a
4568 * cmpxchg to cross pages which acts as the mb on the writer
4569 * side. Note, the reader will constantly fail the swap
4570 * while the writer is updating the pointers, so this
4571 * guarantees that the overwrite recorded here is the one we
4572 * want to compare with the last_overrun.
4573 */
4574 smp_mb();
4575 overwrite = local_read(&(cpu_buffer->overrun));
4576
4577 /*
4578 * Here's the tricky part.
4579 *
4580 * We need to move the pointer past the header page.
4581 * But we can only do that if a writer is not currently
4582 * moving it. The page before the header page has the
4583 * flag bit '1' set if it is pointing to the page we want.
4584 * but if the writer is in the process of moving it
4585 * than it will be '2' or already moved '0'.
4586 */
4587
4588 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
4589
4590 /*
4591 * If we did not convert it, then we must try again.
4592 */
4593 if (!ret)
4594 goto spin;
4595
4596 /*
4597 * Yay! We succeeded in replacing the page.
4598 *
4599 * Now make the new head point back to the reader page.
4600 */
4601 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
4602 rb_inc_page(&cpu_buffer->head_page);
4603
4604 local_inc(&cpu_buffer->pages_read);
4605
4606 /* Finally update the reader page to the new head */
4607 cpu_buffer->reader_page = reader;
4608 cpu_buffer->reader_page->read = 0;
4609
4610 if (overwrite != cpu_buffer->last_overrun) {
4611 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
4612 cpu_buffer->last_overrun = overwrite;
4613 }
4614
4615 goto again;
4616
4617 out:
4618 /* Update the read_stamp on the first event */
4619 if (reader && reader->read == 0)
4620 cpu_buffer->read_stamp = reader->page->time_stamp;
4621
4622 arch_spin_unlock(&cpu_buffer->lock);
4623 local_irq_restore(flags);
4624
4625 /*
4626 * The writer has preempt disable, wait for it. But not forever
4627 * Although, 1 second is pretty much "forever"
4628 */
4629#define USECS_WAIT 1000000
4630 for (nr_loops = 0; nr_loops < USECS_WAIT; nr_loops++) {
4631 /* If the write is past the end of page, a writer is still updating it */
4632 if (likely(!reader || rb_page_write(reader) <= bsize))
4633 break;
4634
4635 udelay(1);
4636
4637 /* Get the latest version of the reader write value */
4638 smp_rmb();
4639 }
4640
4641 /* The writer is not moving forward? Something is wrong */
4642 if (RB_WARN_ON(cpu_buffer, nr_loops == USECS_WAIT))
4643 reader = NULL;
4644
4645 /*
4646 * Make sure we see any padding after the write update
4647 * (see rb_reset_tail()).
4648 *
4649 * In addition, a writer may be writing on the reader page
4650 * if the page has not been fully filled, so the read barrier
4651 * is also needed to make sure we see the content of what is
4652 * committed by the writer (see rb_set_commit_to_write()).
4653 */
4654 smp_rmb();
4655
4656
4657 return reader;
4658}
4659
4660static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
4661{
4662 struct ring_buffer_event *event;
4663 struct buffer_page *reader;
4664 unsigned length;
4665
4666 reader = rb_get_reader_page(cpu_buffer);
4667
4668 /* This function should not be called when buffer is empty */
4669 if (RB_WARN_ON(cpu_buffer, !reader))
4670 return;
4671
4672 event = rb_reader_event(cpu_buffer);
4673
4674 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
4675 cpu_buffer->read++;
4676
4677 rb_update_read_stamp(cpu_buffer, event);
4678
4679 length = rb_event_length(event);
4680 cpu_buffer->reader_page->read += length;
4681 cpu_buffer->read_bytes += length;
4682}
4683
4684static void rb_advance_iter(struct ring_buffer_iter *iter)
4685{
4686 struct ring_buffer_per_cpu *cpu_buffer;
4687
4688 cpu_buffer = iter->cpu_buffer;
4689
4690 /* If head == next_event then we need to jump to the next event */
4691 if (iter->head == iter->next_event) {
4692 /* If the event gets overwritten again, there's nothing to do */
4693 if (rb_iter_head_event(iter) == NULL)
4694 return;
4695 }
4696
4697 iter->head = iter->next_event;
4698
4699 /*
4700 * Check if we are at the end of the buffer.
4701 */
4702 if (iter->next_event >= rb_page_size(iter->head_page)) {
4703 /* discarded commits can make the page empty */
4704 if (iter->head_page == cpu_buffer->commit_page)
4705 return;
4706 rb_inc_iter(iter);
4707 return;
4708 }
4709
4710 rb_update_iter_read_stamp(iter, iter->event);
4711}
4712
4713static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
4714{
4715 return cpu_buffer->lost_events;
4716}
4717
4718static struct ring_buffer_event *
4719rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
4720 unsigned long *lost_events)
4721{
4722 struct ring_buffer_event *event;
4723 struct buffer_page *reader;
4724 int nr_loops = 0;
4725
4726 if (ts)
4727 *ts = 0;
4728 again:
4729 /*
4730 * We repeat when a time extend is encountered.
4731 * Since the time extend is always attached to a data event,
4732 * we should never loop more than once.
4733 * (We never hit the following condition more than twice).
4734 */
4735 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
4736 return NULL;
4737
4738 reader = rb_get_reader_page(cpu_buffer);
4739 if (!reader)
4740 return NULL;
4741
4742 event = rb_reader_event(cpu_buffer);
4743
4744 switch (event->type_len) {
4745 case RINGBUF_TYPE_PADDING:
4746 if (rb_null_event(event))
4747 RB_WARN_ON(cpu_buffer, 1);
4748 /*
4749 * Because the writer could be discarding every
4750 * event it creates (which would probably be bad)
4751 * if we were to go back to "again" then we may never
4752 * catch up, and will trigger the warn on, or lock
4753 * the box. Return the padding, and we will release
4754 * the current locks, and try again.
4755 */
4756 return event;
4757
4758 case RINGBUF_TYPE_TIME_EXTEND:
4759 /* Internal data, OK to advance */
4760 rb_advance_reader(cpu_buffer);
4761 goto again;
4762
4763 case RINGBUF_TYPE_TIME_STAMP:
4764 if (ts) {
4765 *ts = rb_event_time_stamp(event);
4766 *ts = rb_fix_abs_ts(*ts, reader->page->time_stamp);
4767 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4768 cpu_buffer->cpu, ts);
4769 }
4770 /* Internal data, OK to advance */
4771 rb_advance_reader(cpu_buffer);
4772 goto again;
4773
4774 case RINGBUF_TYPE_DATA:
4775 if (ts && !(*ts)) {
4776 *ts = cpu_buffer->read_stamp + event->time_delta;
4777 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4778 cpu_buffer->cpu, ts);
4779 }
4780 if (lost_events)
4781 *lost_events = rb_lost_events(cpu_buffer);
4782 return event;
4783
4784 default:
4785 RB_WARN_ON(cpu_buffer, 1);
4786 }
4787
4788 return NULL;
4789}
4790EXPORT_SYMBOL_GPL(ring_buffer_peek);
4791
4792static struct ring_buffer_event *
4793rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4794{
4795 struct trace_buffer *buffer;
4796 struct ring_buffer_per_cpu *cpu_buffer;
4797 struct ring_buffer_event *event;
4798 int nr_loops = 0;
4799
4800 if (ts)
4801 *ts = 0;
4802
4803 cpu_buffer = iter->cpu_buffer;
4804 buffer = cpu_buffer->buffer;
4805
4806 /*
4807 * Check if someone performed a consuming read to the buffer
4808 * or removed some pages from the buffer. In these cases,
4809 * iterator was invalidated and we need to reset it.
4810 */
4811 if (unlikely(iter->cache_read != cpu_buffer->read ||
4812 iter->cache_reader_page != cpu_buffer->reader_page ||
4813 iter->cache_pages_removed != cpu_buffer->pages_removed))
4814 rb_iter_reset(iter);
4815
4816 again:
4817 if (ring_buffer_iter_empty(iter))
4818 return NULL;
4819
4820 /*
4821 * As the writer can mess with what the iterator is trying
4822 * to read, just give up if we fail to get an event after
4823 * three tries. The iterator is not as reliable when reading
4824 * the ring buffer with an active write as the consumer is.
4825 * Do not warn if the three failures is reached.
4826 */
4827 if (++nr_loops > 3)
4828 return NULL;
4829
4830 if (rb_per_cpu_empty(cpu_buffer))
4831 return NULL;
4832
4833 if (iter->head >= rb_page_size(iter->head_page)) {
4834 rb_inc_iter(iter);
4835 goto again;
4836 }
4837
4838 event = rb_iter_head_event(iter);
4839 if (!event)
4840 goto again;
4841
4842 switch (event->type_len) {
4843 case RINGBUF_TYPE_PADDING:
4844 if (rb_null_event(event)) {
4845 rb_inc_iter(iter);
4846 goto again;
4847 }
4848 rb_advance_iter(iter);
4849 return event;
4850
4851 case RINGBUF_TYPE_TIME_EXTEND:
4852 /* Internal data, OK to advance */
4853 rb_advance_iter(iter);
4854 goto again;
4855
4856 case RINGBUF_TYPE_TIME_STAMP:
4857 if (ts) {
4858 *ts = rb_event_time_stamp(event);
4859 *ts = rb_fix_abs_ts(*ts, iter->head_page->page->time_stamp);
4860 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4861 cpu_buffer->cpu, ts);
4862 }
4863 /* Internal data, OK to advance */
4864 rb_advance_iter(iter);
4865 goto again;
4866
4867 case RINGBUF_TYPE_DATA:
4868 if (ts && !(*ts)) {
4869 *ts = iter->read_stamp + event->time_delta;
4870 ring_buffer_normalize_time_stamp(buffer,
4871 cpu_buffer->cpu, ts);
4872 }
4873 return event;
4874
4875 default:
4876 RB_WARN_ON(cpu_buffer, 1);
4877 }
4878
4879 return NULL;
4880}
4881EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4882
4883static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4884{
4885 if (likely(!in_nmi())) {
4886 raw_spin_lock(&cpu_buffer->reader_lock);
4887 return true;
4888 }
4889
4890 /*
4891 * If an NMI die dumps out the content of the ring buffer
4892 * trylock must be used to prevent a deadlock if the NMI
4893 * preempted a task that holds the ring buffer locks. If
4894 * we get the lock then all is fine, if not, then continue
4895 * to do the read, but this can corrupt the ring buffer,
4896 * so it must be permanently disabled from future writes.
4897 * Reading from NMI is a oneshot deal.
4898 */
4899 if (raw_spin_trylock(&cpu_buffer->reader_lock))
4900 return true;
4901
4902 /* Continue without locking, but disable the ring buffer */
4903 atomic_inc(&cpu_buffer->record_disabled);
4904 return false;
4905}
4906
4907static inline void
4908rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4909{
4910 if (likely(locked))
4911 raw_spin_unlock(&cpu_buffer->reader_lock);
4912}
4913
4914/**
4915 * ring_buffer_peek - peek at the next event to be read
4916 * @buffer: The ring buffer to read
4917 * @cpu: The cpu to peak at
4918 * @ts: The timestamp counter of this event.
4919 * @lost_events: a variable to store if events were lost (may be NULL)
4920 *
4921 * This will return the event that will be read next, but does
4922 * not consume the data.
4923 */
4924struct ring_buffer_event *
4925ring_buffer_peek(struct trace_buffer *buffer, int cpu, u64 *ts,
4926 unsigned long *lost_events)
4927{
4928 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4929 struct ring_buffer_event *event;
4930 unsigned long flags;
4931 bool dolock;
4932
4933 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4934 return NULL;
4935
4936 again:
4937 local_irq_save(flags);
4938 dolock = rb_reader_lock(cpu_buffer);
4939 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4940 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4941 rb_advance_reader(cpu_buffer);
4942 rb_reader_unlock(cpu_buffer, dolock);
4943 local_irq_restore(flags);
4944
4945 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4946 goto again;
4947
4948 return event;
4949}
4950
4951/** ring_buffer_iter_dropped - report if there are dropped events
4952 * @iter: The ring buffer iterator
4953 *
4954 * Returns true if there was dropped events since the last peek.
4955 */
4956bool ring_buffer_iter_dropped(struct ring_buffer_iter *iter)
4957{
4958 bool ret = iter->missed_events != 0;
4959
4960 iter->missed_events = 0;
4961 return ret;
4962}
4963EXPORT_SYMBOL_GPL(ring_buffer_iter_dropped);
4964
4965/**
4966 * ring_buffer_iter_peek - peek at the next event to be read
4967 * @iter: The ring buffer iterator
4968 * @ts: The timestamp counter of this event.
4969 *
4970 * This will return the event that will be read next, but does
4971 * not increment the iterator.
4972 */
4973struct ring_buffer_event *
4974ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4975{
4976 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4977 struct ring_buffer_event *event;
4978 unsigned long flags;
4979
4980 again:
4981 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4982 event = rb_iter_peek(iter, ts);
4983 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4984
4985 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4986 goto again;
4987
4988 return event;
4989}
4990
4991/**
4992 * ring_buffer_consume - return an event and consume it
4993 * @buffer: The ring buffer to get the next event from
4994 * @cpu: the cpu to read the buffer from
4995 * @ts: a variable to store the timestamp (may be NULL)
4996 * @lost_events: a variable to store if events were lost (may be NULL)
4997 *
4998 * Returns the next event in the ring buffer, and that event is consumed.
4999 * Meaning, that sequential reads will keep returning a different event,
5000 * and eventually empty the ring buffer if the producer is slower.
5001 */
5002struct ring_buffer_event *
5003ring_buffer_consume(struct trace_buffer *buffer, int cpu, u64 *ts,
5004 unsigned long *lost_events)
5005{
5006 struct ring_buffer_per_cpu *cpu_buffer;
5007 struct ring_buffer_event *event = NULL;
5008 unsigned long flags;
5009 bool dolock;
5010
5011 again:
5012 /* might be called in atomic */
5013 preempt_disable();
5014
5015 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5016 goto out;
5017
5018 cpu_buffer = buffer->buffers[cpu];
5019 local_irq_save(flags);
5020 dolock = rb_reader_lock(cpu_buffer);
5021
5022 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
5023 if (event) {
5024 cpu_buffer->lost_events = 0;
5025 rb_advance_reader(cpu_buffer);
5026 }
5027
5028 rb_reader_unlock(cpu_buffer, dolock);
5029 local_irq_restore(flags);
5030
5031 out:
5032 preempt_enable();
5033
5034 if (event && event->type_len == RINGBUF_TYPE_PADDING)
5035 goto again;
5036
5037 return event;
5038}
5039EXPORT_SYMBOL_GPL(ring_buffer_consume);
5040
5041/**
5042 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
5043 * @buffer: The ring buffer to read from
5044 * @cpu: The cpu buffer to iterate over
5045 * @flags: gfp flags to use for memory allocation
5046 *
5047 * This performs the initial preparations necessary to iterate
5048 * through the buffer. Memory is allocated, buffer recording
5049 * is disabled, and the iterator pointer is returned to the caller.
5050 *
5051 * Disabling buffer recording prevents the reading from being
5052 * corrupted. This is not a consuming read, so a producer is not
5053 * expected.
5054 *
5055 * After a sequence of ring_buffer_read_prepare calls, the user is
5056 * expected to make at least one call to ring_buffer_read_prepare_sync.
5057 * Afterwards, ring_buffer_read_start is invoked to get things going
5058 * for real.
5059 *
5060 * This overall must be paired with ring_buffer_read_finish.
5061 */
5062struct ring_buffer_iter *
5063ring_buffer_read_prepare(struct trace_buffer *buffer, int cpu, gfp_t flags)
5064{
5065 struct ring_buffer_per_cpu *cpu_buffer;
5066 struct ring_buffer_iter *iter;
5067
5068 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5069 return NULL;
5070
5071 iter = kzalloc(sizeof(*iter), flags);
5072 if (!iter)
5073 return NULL;
5074
5075 /* Holds the entire event: data and meta data */
5076 iter->event_size = buffer->subbuf_size;
5077 iter->event = kmalloc(iter->event_size, flags);
5078 if (!iter->event) {
5079 kfree(iter);
5080 return NULL;
5081 }
5082
5083 cpu_buffer = buffer->buffers[cpu];
5084
5085 iter->cpu_buffer = cpu_buffer;
5086
5087 atomic_inc(&cpu_buffer->resize_disabled);
5088
5089 return iter;
5090}
5091EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
5092
5093/**
5094 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
5095 *
5096 * All previously invoked ring_buffer_read_prepare calls to prepare
5097 * iterators will be synchronized. Afterwards, read_buffer_read_start
5098 * calls on those iterators are allowed.
5099 */
5100void
5101ring_buffer_read_prepare_sync(void)
5102{
5103 synchronize_rcu();
5104}
5105EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
5106
5107/**
5108 * ring_buffer_read_start - start a non consuming read of the buffer
5109 * @iter: The iterator returned by ring_buffer_read_prepare
5110 *
5111 * This finalizes the startup of an iteration through the buffer.
5112 * The iterator comes from a call to ring_buffer_read_prepare and
5113 * an intervening ring_buffer_read_prepare_sync must have been
5114 * performed.
5115 *
5116 * Must be paired with ring_buffer_read_finish.
5117 */
5118void
5119ring_buffer_read_start(struct ring_buffer_iter *iter)
5120{
5121 struct ring_buffer_per_cpu *cpu_buffer;
5122 unsigned long flags;
5123
5124 if (!iter)
5125 return;
5126
5127 cpu_buffer = iter->cpu_buffer;
5128
5129 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5130 arch_spin_lock(&cpu_buffer->lock);
5131 rb_iter_reset(iter);
5132 arch_spin_unlock(&cpu_buffer->lock);
5133 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5134}
5135EXPORT_SYMBOL_GPL(ring_buffer_read_start);
5136
5137/**
5138 * ring_buffer_read_finish - finish reading the iterator of the buffer
5139 * @iter: The iterator retrieved by ring_buffer_start
5140 *
5141 * This re-enables the recording to the buffer, and frees the
5142 * iterator.
5143 */
5144void
5145ring_buffer_read_finish(struct ring_buffer_iter *iter)
5146{
5147 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5148 unsigned long flags;
5149
5150 /*
5151 * Ring buffer is disabled from recording, here's a good place
5152 * to check the integrity of the ring buffer.
5153 * Must prevent readers from trying to read, as the check
5154 * clears the HEAD page and readers require it.
5155 */
5156 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5157 rb_check_pages(cpu_buffer);
5158 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5159
5160 atomic_dec(&cpu_buffer->resize_disabled);
5161 kfree(iter->event);
5162 kfree(iter);
5163}
5164EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
5165
5166/**
5167 * ring_buffer_iter_advance - advance the iterator to the next location
5168 * @iter: The ring buffer iterator
5169 *
5170 * Move the location of the iterator such that the next read will
5171 * be the next location of the iterator.
5172 */
5173void ring_buffer_iter_advance(struct ring_buffer_iter *iter)
5174{
5175 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5176 unsigned long flags;
5177
5178 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5179
5180 rb_advance_iter(iter);
5181
5182 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5183}
5184EXPORT_SYMBOL_GPL(ring_buffer_iter_advance);
5185
5186/**
5187 * ring_buffer_size - return the size of the ring buffer (in bytes)
5188 * @buffer: The ring buffer.
5189 * @cpu: The CPU to get ring buffer size from.
5190 */
5191unsigned long ring_buffer_size(struct trace_buffer *buffer, int cpu)
5192{
5193 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5194 return 0;
5195
5196 return buffer->subbuf_size * buffer->buffers[cpu]->nr_pages;
5197}
5198EXPORT_SYMBOL_GPL(ring_buffer_size);
5199
5200/**
5201 * ring_buffer_max_event_size - return the max data size of an event
5202 * @buffer: The ring buffer.
5203 *
5204 * Returns the maximum size an event can be.
5205 */
5206unsigned long ring_buffer_max_event_size(struct trace_buffer *buffer)
5207{
5208 /* If abs timestamp is requested, events have a timestamp too */
5209 if (ring_buffer_time_stamp_abs(buffer))
5210 return buffer->max_data_size - RB_LEN_TIME_EXTEND;
5211 return buffer->max_data_size;
5212}
5213EXPORT_SYMBOL_GPL(ring_buffer_max_event_size);
5214
5215static void rb_clear_buffer_page(struct buffer_page *page)
5216{
5217 local_set(&page->write, 0);
5218 local_set(&page->entries, 0);
5219 rb_init_page(page->page);
5220 page->read = 0;
5221}
5222
5223static void
5224rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
5225{
5226 struct buffer_page *page;
5227
5228 rb_head_page_deactivate(cpu_buffer);
5229
5230 cpu_buffer->head_page
5231 = list_entry(cpu_buffer->pages, struct buffer_page, list);
5232 rb_clear_buffer_page(cpu_buffer->head_page);
5233 list_for_each_entry(page, cpu_buffer->pages, list) {
5234 rb_clear_buffer_page(page);
5235 }
5236
5237 cpu_buffer->tail_page = cpu_buffer->head_page;
5238 cpu_buffer->commit_page = cpu_buffer->head_page;
5239
5240 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
5241 INIT_LIST_HEAD(&cpu_buffer->new_pages);
5242 rb_clear_buffer_page(cpu_buffer->reader_page);
5243
5244 local_set(&cpu_buffer->entries_bytes, 0);
5245 local_set(&cpu_buffer->overrun, 0);
5246 local_set(&cpu_buffer->commit_overrun, 0);
5247 local_set(&cpu_buffer->dropped_events, 0);
5248 local_set(&cpu_buffer->entries, 0);
5249 local_set(&cpu_buffer->committing, 0);
5250 local_set(&cpu_buffer->commits, 0);
5251 local_set(&cpu_buffer->pages_touched, 0);
5252 local_set(&cpu_buffer->pages_lost, 0);
5253 local_set(&cpu_buffer->pages_read, 0);
5254 cpu_buffer->last_pages_touch = 0;
5255 cpu_buffer->shortest_full = 0;
5256 cpu_buffer->read = 0;
5257 cpu_buffer->read_bytes = 0;
5258
5259 rb_time_set(&cpu_buffer->write_stamp, 0);
5260 rb_time_set(&cpu_buffer->before_stamp, 0);
5261
5262 memset(cpu_buffer->event_stamp, 0, sizeof(cpu_buffer->event_stamp));
5263
5264 cpu_buffer->lost_events = 0;
5265 cpu_buffer->last_overrun = 0;
5266
5267 rb_head_page_activate(cpu_buffer);
5268 cpu_buffer->pages_removed = 0;
5269}
5270
5271/* Must have disabled the cpu buffer then done a synchronize_rcu */
5272static void reset_disabled_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
5273{
5274 unsigned long flags;
5275
5276 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5277
5278 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
5279 goto out;
5280
5281 arch_spin_lock(&cpu_buffer->lock);
5282
5283 rb_reset_cpu(cpu_buffer);
5284
5285 arch_spin_unlock(&cpu_buffer->lock);
5286
5287 out:
5288 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5289}
5290
5291/**
5292 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
5293 * @buffer: The ring buffer to reset a per cpu buffer of
5294 * @cpu: The CPU buffer to be reset
5295 */
5296void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
5297{
5298 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5299
5300 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5301 return;
5302
5303 /* prevent another thread from changing buffer sizes */
5304 mutex_lock(&buffer->mutex);
5305
5306 atomic_inc(&cpu_buffer->resize_disabled);
5307 atomic_inc(&cpu_buffer->record_disabled);
5308
5309 /* Make sure all commits have finished */
5310 synchronize_rcu();
5311
5312 reset_disabled_cpu_buffer(cpu_buffer);
5313
5314 atomic_dec(&cpu_buffer->record_disabled);
5315 atomic_dec(&cpu_buffer->resize_disabled);
5316
5317 mutex_unlock(&buffer->mutex);
5318}
5319EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
5320
5321/* Flag to ensure proper resetting of atomic variables */
5322#define RESET_BIT (1 << 30)
5323
5324/**
5325 * ring_buffer_reset_online_cpus - reset a ring buffer per CPU buffer
5326 * @buffer: The ring buffer to reset a per cpu buffer of
5327 */
5328void ring_buffer_reset_online_cpus(struct trace_buffer *buffer)
5329{
5330 struct ring_buffer_per_cpu *cpu_buffer;
5331 int cpu;
5332
5333 /* prevent another thread from changing buffer sizes */
5334 mutex_lock(&buffer->mutex);
5335
5336 for_each_online_buffer_cpu(buffer, cpu) {
5337 cpu_buffer = buffer->buffers[cpu];
5338
5339 atomic_add(RESET_BIT, &cpu_buffer->resize_disabled);
5340 atomic_inc(&cpu_buffer->record_disabled);
5341 }
5342
5343 /* Make sure all commits have finished */
5344 synchronize_rcu();
5345
5346 for_each_buffer_cpu(buffer, cpu) {
5347 cpu_buffer = buffer->buffers[cpu];
5348
5349 /*
5350 * If a CPU came online during the synchronize_rcu(), then
5351 * ignore it.
5352 */
5353 if (!(atomic_read(&cpu_buffer->resize_disabled) & RESET_BIT))
5354 continue;
5355
5356 reset_disabled_cpu_buffer(cpu_buffer);
5357
5358 atomic_dec(&cpu_buffer->record_disabled);
5359 atomic_sub(RESET_BIT, &cpu_buffer->resize_disabled);
5360 }
5361
5362 mutex_unlock(&buffer->mutex);
5363}
5364
5365/**
5366 * ring_buffer_reset - reset a ring buffer
5367 * @buffer: The ring buffer to reset all cpu buffers
5368 */
5369void ring_buffer_reset(struct trace_buffer *buffer)
5370{
5371 struct ring_buffer_per_cpu *cpu_buffer;
5372 int cpu;
5373
5374 /* prevent another thread from changing buffer sizes */
5375 mutex_lock(&buffer->mutex);
5376
5377 for_each_buffer_cpu(buffer, cpu) {
5378 cpu_buffer = buffer->buffers[cpu];
5379
5380 atomic_inc(&cpu_buffer->resize_disabled);
5381 atomic_inc(&cpu_buffer->record_disabled);
5382 }
5383
5384 /* Make sure all commits have finished */
5385 synchronize_rcu();
5386
5387 for_each_buffer_cpu(buffer, cpu) {
5388 cpu_buffer = buffer->buffers[cpu];
5389
5390 reset_disabled_cpu_buffer(cpu_buffer);
5391
5392 atomic_dec(&cpu_buffer->record_disabled);
5393 atomic_dec(&cpu_buffer->resize_disabled);
5394 }
5395
5396 mutex_unlock(&buffer->mutex);
5397}
5398EXPORT_SYMBOL_GPL(ring_buffer_reset);
5399
5400/**
5401 * ring_buffer_empty - is the ring buffer empty?
5402 * @buffer: The ring buffer to test
5403 */
5404bool ring_buffer_empty(struct trace_buffer *buffer)
5405{
5406 struct ring_buffer_per_cpu *cpu_buffer;
5407 unsigned long flags;
5408 bool dolock;
5409 bool ret;
5410 int cpu;
5411
5412 /* yes this is racy, but if you don't like the race, lock the buffer */
5413 for_each_buffer_cpu(buffer, cpu) {
5414 cpu_buffer = buffer->buffers[cpu];
5415 local_irq_save(flags);
5416 dolock = rb_reader_lock(cpu_buffer);
5417 ret = rb_per_cpu_empty(cpu_buffer);
5418 rb_reader_unlock(cpu_buffer, dolock);
5419 local_irq_restore(flags);
5420
5421 if (!ret)
5422 return false;
5423 }
5424
5425 return true;
5426}
5427EXPORT_SYMBOL_GPL(ring_buffer_empty);
5428
5429/**
5430 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
5431 * @buffer: The ring buffer
5432 * @cpu: The CPU buffer to test
5433 */
5434bool ring_buffer_empty_cpu(struct trace_buffer *buffer, int cpu)
5435{
5436 struct ring_buffer_per_cpu *cpu_buffer;
5437 unsigned long flags;
5438 bool dolock;
5439 bool ret;
5440
5441 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5442 return true;
5443
5444 cpu_buffer = buffer->buffers[cpu];
5445 local_irq_save(flags);
5446 dolock = rb_reader_lock(cpu_buffer);
5447 ret = rb_per_cpu_empty(cpu_buffer);
5448 rb_reader_unlock(cpu_buffer, dolock);
5449 local_irq_restore(flags);
5450
5451 return ret;
5452}
5453EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
5454
5455#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
5456/**
5457 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
5458 * @buffer_a: One buffer to swap with
5459 * @buffer_b: The other buffer to swap with
5460 * @cpu: the CPU of the buffers to swap
5461 *
5462 * This function is useful for tracers that want to take a "snapshot"
5463 * of a CPU buffer and has another back up buffer lying around.
5464 * it is expected that the tracer handles the cpu buffer not being
5465 * used at the moment.
5466 */
5467int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
5468 struct trace_buffer *buffer_b, int cpu)
5469{
5470 struct ring_buffer_per_cpu *cpu_buffer_a;
5471 struct ring_buffer_per_cpu *cpu_buffer_b;
5472 int ret = -EINVAL;
5473
5474 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
5475 !cpumask_test_cpu(cpu, buffer_b->cpumask))
5476 goto out;
5477
5478 cpu_buffer_a = buffer_a->buffers[cpu];
5479 cpu_buffer_b = buffer_b->buffers[cpu];
5480
5481 /* At least make sure the two buffers are somewhat the same */
5482 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
5483 goto out;
5484
5485 if (buffer_a->subbuf_order != buffer_b->subbuf_order)
5486 goto out;
5487
5488 ret = -EAGAIN;
5489
5490 if (atomic_read(&buffer_a->record_disabled))
5491 goto out;
5492
5493 if (atomic_read(&buffer_b->record_disabled))
5494 goto out;
5495
5496 if (atomic_read(&cpu_buffer_a->record_disabled))
5497 goto out;
5498
5499 if (atomic_read(&cpu_buffer_b->record_disabled))
5500 goto out;
5501
5502 /*
5503 * We can't do a synchronize_rcu here because this
5504 * function can be called in atomic context.
5505 * Normally this will be called from the same CPU as cpu.
5506 * If not it's up to the caller to protect this.
5507 */
5508 atomic_inc(&cpu_buffer_a->record_disabled);
5509 atomic_inc(&cpu_buffer_b->record_disabled);
5510
5511 ret = -EBUSY;
5512 if (local_read(&cpu_buffer_a->committing))
5513 goto out_dec;
5514 if (local_read(&cpu_buffer_b->committing))
5515 goto out_dec;
5516
5517 /*
5518 * When resize is in progress, we cannot swap it because
5519 * it will mess the state of the cpu buffer.
5520 */
5521 if (atomic_read(&buffer_a->resizing))
5522 goto out_dec;
5523 if (atomic_read(&buffer_b->resizing))
5524 goto out_dec;
5525
5526 buffer_a->buffers[cpu] = cpu_buffer_b;
5527 buffer_b->buffers[cpu] = cpu_buffer_a;
5528
5529 cpu_buffer_b->buffer = buffer_a;
5530 cpu_buffer_a->buffer = buffer_b;
5531
5532 ret = 0;
5533
5534out_dec:
5535 atomic_dec(&cpu_buffer_a->record_disabled);
5536 atomic_dec(&cpu_buffer_b->record_disabled);
5537out:
5538 return ret;
5539}
5540EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
5541#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
5542
5543/**
5544 * ring_buffer_alloc_read_page - allocate a page to read from buffer
5545 * @buffer: the buffer to allocate for.
5546 * @cpu: the cpu buffer to allocate.
5547 *
5548 * This function is used in conjunction with ring_buffer_read_page.
5549 * When reading a full page from the ring buffer, these functions
5550 * can be used to speed up the process. The calling function should
5551 * allocate a few pages first with this function. Then when it
5552 * needs to get pages from the ring buffer, it passes the result
5553 * of this function into ring_buffer_read_page, which will swap
5554 * the page that was allocated, with the read page of the buffer.
5555 *
5556 * Returns:
5557 * The page allocated, or ERR_PTR
5558 */
5559struct buffer_data_read_page *
5560ring_buffer_alloc_read_page(struct trace_buffer *buffer, int cpu)
5561{
5562 struct ring_buffer_per_cpu *cpu_buffer;
5563 struct buffer_data_read_page *bpage = NULL;
5564 unsigned long flags;
5565 struct page *page;
5566
5567 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5568 return ERR_PTR(-ENODEV);
5569
5570 bpage = kzalloc(sizeof(*bpage), GFP_KERNEL);
5571 if (!bpage)
5572 return ERR_PTR(-ENOMEM);
5573
5574 bpage->order = buffer->subbuf_order;
5575 cpu_buffer = buffer->buffers[cpu];
5576 local_irq_save(flags);
5577 arch_spin_lock(&cpu_buffer->lock);
5578
5579 if (cpu_buffer->free_page) {
5580 bpage->data = cpu_buffer->free_page;
5581 cpu_buffer->free_page = NULL;
5582 }
5583
5584 arch_spin_unlock(&cpu_buffer->lock);
5585 local_irq_restore(flags);
5586
5587 if (bpage->data)
5588 goto out;
5589
5590 page = alloc_pages_node(cpu_to_node(cpu),
5591 GFP_KERNEL | __GFP_NORETRY | __GFP_ZERO,
5592 cpu_buffer->buffer->subbuf_order);
5593 if (!page) {
5594 kfree(bpage);
5595 return ERR_PTR(-ENOMEM);
5596 }
5597
5598 bpage->data = page_address(page);
5599
5600 out:
5601 rb_init_page(bpage->data);
5602
5603 return bpage;
5604}
5605EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
5606
5607/**
5608 * ring_buffer_free_read_page - free an allocated read page
5609 * @buffer: the buffer the page was allocate for
5610 * @cpu: the cpu buffer the page came from
5611 * @data_page: the page to free
5612 *
5613 * Free a page allocated from ring_buffer_alloc_read_page.
5614 */
5615void ring_buffer_free_read_page(struct trace_buffer *buffer, int cpu,
5616 struct buffer_data_read_page *data_page)
5617{
5618 struct ring_buffer_per_cpu *cpu_buffer;
5619 struct buffer_data_page *bpage = data_page->data;
5620 struct page *page = virt_to_page(bpage);
5621 unsigned long flags;
5622
5623 if (!buffer || !buffer->buffers || !buffer->buffers[cpu])
5624 return;
5625
5626 cpu_buffer = buffer->buffers[cpu];
5627
5628 /*
5629 * If the page is still in use someplace else, or order of the page
5630 * is different from the subbuffer order of the buffer -
5631 * we can't reuse it
5632 */
5633 if (page_ref_count(page) > 1 || data_page->order != buffer->subbuf_order)
5634 goto out;
5635
5636 local_irq_save(flags);
5637 arch_spin_lock(&cpu_buffer->lock);
5638
5639 if (!cpu_buffer->free_page) {
5640 cpu_buffer->free_page = bpage;
5641 bpage = NULL;
5642 }
5643
5644 arch_spin_unlock(&cpu_buffer->lock);
5645 local_irq_restore(flags);
5646
5647 out:
5648 free_pages((unsigned long)bpage, data_page->order);
5649 kfree(data_page);
5650}
5651EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
5652
5653/**
5654 * ring_buffer_read_page - extract a page from the ring buffer
5655 * @buffer: buffer to extract from
5656 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
5657 * @len: amount to extract
5658 * @cpu: the cpu of the buffer to extract
5659 * @full: should the extraction only happen when the page is full.
5660 *
5661 * This function will pull out a page from the ring buffer and consume it.
5662 * @data_page must be the address of the variable that was returned
5663 * from ring_buffer_alloc_read_page. This is because the page might be used
5664 * to swap with a page in the ring buffer.
5665 *
5666 * for example:
5667 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
5668 * if (IS_ERR(rpage))
5669 * return PTR_ERR(rpage);
5670 * ret = ring_buffer_read_page(buffer, rpage, len, cpu, 0);
5671 * if (ret >= 0)
5672 * process_page(ring_buffer_read_page_data(rpage), ret);
5673 * ring_buffer_free_read_page(buffer, cpu, rpage);
5674 *
5675 * When @full is set, the function will not return true unless
5676 * the writer is off the reader page.
5677 *
5678 * Note: it is up to the calling functions to handle sleeps and wakeups.
5679 * The ring buffer can be used anywhere in the kernel and can not
5680 * blindly call wake_up. The layer that uses the ring buffer must be
5681 * responsible for that.
5682 *
5683 * Returns:
5684 * >=0 if data has been transferred, returns the offset of consumed data.
5685 * <0 if no data has been transferred.
5686 */
5687int ring_buffer_read_page(struct trace_buffer *buffer,
5688 struct buffer_data_read_page *data_page,
5689 size_t len, int cpu, int full)
5690{
5691 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5692 struct ring_buffer_event *event;
5693 struct buffer_data_page *bpage;
5694 struct buffer_page *reader;
5695 unsigned long missed_events;
5696 unsigned long flags;
5697 unsigned int commit;
5698 unsigned int read;
5699 u64 save_timestamp;
5700 int ret = -1;
5701
5702 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5703 goto out;
5704
5705 /*
5706 * If len is not big enough to hold the page header, then
5707 * we can not copy anything.
5708 */
5709 if (len <= BUF_PAGE_HDR_SIZE)
5710 goto out;
5711
5712 len -= BUF_PAGE_HDR_SIZE;
5713
5714 if (!data_page || !data_page->data)
5715 goto out;
5716 if (data_page->order != buffer->subbuf_order)
5717 goto out;
5718
5719 bpage = data_page->data;
5720 if (!bpage)
5721 goto out;
5722
5723 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5724
5725 reader = rb_get_reader_page(cpu_buffer);
5726 if (!reader)
5727 goto out_unlock;
5728
5729 event = rb_reader_event(cpu_buffer);
5730
5731 read = reader->read;
5732 commit = rb_page_commit(reader);
5733
5734 /* Check if any events were dropped */
5735 missed_events = cpu_buffer->lost_events;
5736
5737 /*
5738 * If this page has been partially read or
5739 * if len is not big enough to read the rest of the page or
5740 * a writer is still on the page, then
5741 * we must copy the data from the page to the buffer.
5742 * Otherwise, we can simply swap the page with the one passed in.
5743 */
5744 if (read || (len < (commit - read)) ||
5745 cpu_buffer->reader_page == cpu_buffer->commit_page) {
5746 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
5747 unsigned int rpos = read;
5748 unsigned int pos = 0;
5749 unsigned int size;
5750
5751 /*
5752 * If a full page is expected, this can still be returned
5753 * if there's been a previous partial read and the
5754 * rest of the page can be read and the commit page is off
5755 * the reader page.
5756 */
5757 if (full &&
5758 (!read || (len < (commit - read)) ||
5759 cpu_buffer->reader_page == cpu_buffer->commit_page))
5760 goto out_unlock;
5761
5762 if (len > (commit - read))
5763 len = (commit - read);
5764
5765 /* Always keep the time extend and data together */
5766 size = rb_event_ts_length(event);
5767
5768 if (len < size)
5769 goto out_unlock;
5770
5771 /* save the current timestamp, since the user will need it */
5772 save_timestamp = cpu_buffer->read_stamp;
5773
5774 /* Need to copy one event at a time */
5775 do {
5776 /* We need the size of one event, because
5777 * rb_advance_reader only advances by one event,
5778 * whereas rb_event_ts_length may include the size of
5779 * one or two events.
5780 * We have already ensured there's enough space if this
5781 * is a time extend. */
5782 size = rb_event_length(event);
5783 memcpy(bpage->data + pos, rpage->data + rpos, size);
5784
5785 len -= size;
5786
5787 rb_advance_reader(cpu_buffer);
5788 rpos = reader->read;
5789 pos += size;
5790
5791 if (rpos >= commit)
5792 break;
5793
5794 event = rb_reader_event(cpu_buffer);
5795 /* Always keep the time extend and data together */
5796 size = rb_event_ts_length(event);
5797 } while (len >= size);
5798
5799 /* update bpage */
5800 local_set(&bpage->commit, pos);
5801 bpage->time_stamp = save_timestamp;
5802
5803 /* we copied everything to the beginning */
5804 read = 0;
5805 } else {
5806 /* update the entry counter */
5807 cpu_buffer->read += rb_page_entries(reader);
5808 cpu_buffer->read_bytes += rb_page_commit(reader);
5809
5810 /* swap the pages */
5811 rb_init_page(bpage);
5812 bpage = reader->page;
5813 reader->page = data_page->data;
5814 local_set(&reader->write, 0);
5815 local_set(&reader->entries, 0);
5816 reader->read = 0;
5817 data_page->data = bpage;
5818
5819 /*
5820 * Use the real_end for the data size,
5821 * This gives us a chance to store the lost events
5822 * on the page.
5823 */
5824 if (reader->real_end)
5825 local_set(&bpage->commit, reader->real_end);
5826 }
5827 ret = read;
5828
5829 cpu_buffer->lost_events = 0;
5830
5831 commit = local_read(&bpage->commit);
5832 /*
5833 * Set a flag in the commit field if we lost events
5834 */
5835 if (missed_events) {
5836 /* If there is room at the end of the page to save the
5837 * missed events, then record it there.
5838 */
5839 if (buffer->subbuf_size - commit >= sizeof(missed_events)) {
5840 memcpy(&bpage->data[commit], &missed_events,
5841 sizeof(missed_events));
5842 local_add(RB_MISSED_STORED, &bpage->commit);
5843 commit += sizeof(missed_events);
5844 }
5845 local_add(RB_MISSED_EVENTS, &bpage->commit);
5846 }
5847
5848 /*
5849 * This page may be off to user land. Zero it out here.
5850 */
5851 if (commit < buffer->subbuf_size)
5852 memset(&bpage->data[commit], 0, buffer->subbuf_size - commit);
5853
5854 out_unlock:
5855 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5856
5857 out:
5858 return ret;
5859}
5860EXPORT_SYMBOL_GPL(ring_buffer_read_page);
5861
5862/**
5863 * ring_buffer_read_page_data - get pointer to the data in the page.
5864 * @page: the page to get the data from
5865 *
5866 * Returns pointer to the actual data in this page.
5867 */
5868void *ring_buffer_read_page_data(struct buffer_data_read_page *page)
5869{
5870 return page->data;
5871}
5872EXPORT_SYMBOL_GPL(ring_buffer_read_page_data);
5873
5874/**
5875 * ring_buffer_subbuf_size_get - get size of the sub buffer.
5876 * @buffer: the buffer to get the sub buffer size from
5877 *
5878 * Returns size of the sub buffer, in bytes.
5879 */
5880int ring_buffer_subbuf_size_get(struct trace_buffer *buffer)
5881{
5882 return buffer->subbuf_size + BUF_PAGE_HDR_SIZE;
5883}
5884EXPORT_SYMBOL_GPL(ring_buffer_subbuf_size_get);
5885
5886/**
5887 * ring_buffer_subbuf_order_get - get order of system sub pages in one buffer page.
5888 * @buffer: The ring_buffer to get the system sub page order from
5889 *
5890 * By default, one ring buffer sub page equals to one system page. This parameter
5891 * is configurable, per ring buffer. The size of the ring buffer sub page can be
5892 * extended, but must be an order of system page size.
5893 *
5894 * Returns the order of buffer sub page size, in system pages:
5895 * 0 means the sub buffer size is 1 system page and so forth.
5896 * In case of an error < 0 is returned.
5897 */
5898int ring_buffer_subbuf_order_get(struct trace_buffer *buffer)
5899{
5900 if (!buffer)
5901 return -EINVAL;
5902
5903 return buffer->subbuf_order;
5904}
5905EXPORT_SYMBOL_GPL(ring_buffer_subbuf_order_get);
5906
5907/**
5908 * ring_buffer_subbuf_order_set - set the size of ring buffer sub page.
5909 * @buffer: The ring_buffer to set the new page size.
5910 * @order: Order of the system pages in one sub buffer page
5911 *
5912 * By default, one ring buffer pages equals to one system page. This API can be
5913 * used to set new size of the ring buffer page. The size must be order of
5914 * system page size, that's why the input parameter @order is the order of
5915 * system pages that are allocated for one ring buffer page:
5916 * 0 - 1 system page
5917 * 1 - 2 system pages
5918 * 3 - 4 system pages
5919 * ...
5920 *
5921 * Returns 0 on success or < 0 in case of an error.
5922 */
5923int ring_buffer_subbuf_order_set(struct trace_buffer *buffer, int order)
5924{
5925 struct ring_buffer_per_cpu *cpu_buffer;
5926 struct buffer_page *bpage, *tmp;
5927 int old_order, old_size;
5928 int nr_pages;
5929 int psize;
5930 int err;
5931 int cpu;
5932
5933 if (!buffer || order < 0)
5934 return -EINVAL;
5935
5936 if (buffer->subbuf_order == order)
5937 return 0;
5938
5939 psize = (1 << order) * PAGE_SIZE;
5940 if (psize <= BUF_PAGE_HDR_SIZE)
5941 return -EINVAL;
5942
5943 /* Size of a subbuf cannot be greater than the write counter */
5944 if (psize > RB_WRITE_MASK + 1)
5945 return -EINVAL;
5946
5947 old_order = buffer->subbuf_order;
5948 old_size = buffer->subbuf_size;
5949
5950 /* prevent another thread from changing buffer sizes */
5951 mutex_lock(&buffer->mutex);
5952 atomic_inc(&buffer->record_disabled);
5953
5954 /* Make sure all commits have finished */
5955 synchronize_rcu();
5956
5957 buffer->subbuf_order = order;
5958 buffer->subbuf_size = psize - BUF_PAGE_HDR_SIZE;
5959
5960 /* Make sure all new buffers are allocated, before deleting the old ones */
5961 for_each_buffer_cpu(buffer, cpu) {
5962
5963 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5964 continue;
5965
5966 cpu_buffer = buffer->buffers[cpu];
5967
5968 /* Update the number of pages to match the new size */
5969 nr_pages = old_size * buffer->buffers[cpu]->nr_pages;
5970 nr_pages = DIV_ROUND_UP(nr_pages, buffer->subbuf_size);
5971
5972 /* we need a minimum of two pages */
5973 if (nr_pages < 2)
5974 nr_pages = 2;
5975
5976 cpu_buffer->nr_pages_to_update = nr_pages;
5977
5978 /* Include the reader page */
5979 nr_pages++;
5980
5981 /* Allocate the new size buffer */
5982 INIT_LIST_HEAD(&cpu_buffer->new_pages);
5983 if (__rb_allocate_pages(cpu_buffer, nr_pages,
5984 &cpu_buffer->new_pages)) {
5985 /* not enough memory for new pages */
5986 err = -ENOMEM;
5987 goto error;
5988 }
5989 }
5990
5991 for_each_buffer_cpu(buffer, cpu) {
5992
5993 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5994 continue;
5995
5996 cpu_buffer = buffer->buffers[cpu];
5997
5998 /* Clear the head bit to make the link list normal to read */
5999 rb_head_page_deactivate(cpu_buffer);
6000
6001 /* Now walk the list and free all the old sub buffers */
6002 list_for_each_entry_safe(bpage, tmp, cpu_buffer->pages, list) {
6003 list_del_init(&bpage->list);
6004 free_buffer_page(bpage);
6005 }
6006 /* The above loop stopped an the last page needing to be freed */
6007 bpage = list_entry(cpu_buffer->pages, struct buffer_page, list);
6008 free_buffer_page(bpage);
6009
6010 /* Free the current reader page */
6011 free_buffer_page(cpu_buffer->reader_page);
6012
6013 /* One page was allocated for the reader page */
6014 cpu_buffer->reader_page = list_entry(cpu_buffer->new_pages.next,
6015 struct buffer_page, list);
6016 list_del_init(&cpu_buffer->reader_page->list);
6017
6018 /* The cpu_buffer pages are a link list with no head */
6019 cpu_buffer->pages = cpu_buffer->new_pages.next;
6020 cpu_buffer->new_pages.next->prev = cpu_buffer->new_pages.prev;
6021 cpu_buffer->new_pages.prev->next = cpu_buffer->new_pages.next;
6022
6023 /* Clear the new_pages list */
6024 INIT_LIST_HEAD(&cpu_buffer->new_pages);
6025
6026 cpu_buffer->head_page
6027 = list_entry(cpu_buffer->pages, struct buffer_page, list);
6028 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
6029
6030 cpu_buffer->nr_pages = cpu_buffer->nr_pages_to_update;
6031 cpu_buffer->nr_pages_to_update = 0;
6032
6033 free_pages((unsigned long)cpu_buffer->free_page, old_order);
6034 cpu_buffer->free_page = NULL;
6035
6036 rb_head_page_activate(cpu_buffer);
6037
6038 rb_check_pages(cpu_buffer);
6039 }
6040
6041 atomic_dec(&buffer->record_disabled);
6042 mutex_unlock(&buffer->mutex);
6043
6044 return 0;
6045
6046error:
6047 buffer->subbuf_order = old_order;
6048 buffer->subbuf_size = old_size;
6049
6050 atomic_dec(&buffer->record_disabled);
6051 mutex_unlock(&buffer->mutex);
6052
6053 for_each_buffer_cpu(buffer, cpu) {
6054 cpu_buffer = buffer->buffers[cpu];
6055
6056 if (!cpu_buffer->nr_pages_to_update)
6057 continue;
6058
6059 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages, list) {
6060 list_del_init(&bpage->list);
6061 free_buffer_page(bpage);
6062 }
6063 }
6064
6065 return err;
6066}
6067EXPORT_SYMBOL_GPL(ring_buffer_subbuf_order_set);
6068
6069/*
6070 * We only allocate new buffers, never free them if the CPU goes down.
6071 * If we were to free the buffer, then the user would lose any trace that was in
6072 * the buffer.
6073 */
6074int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
6075{
6076 struct trace_buffer *buffer;
6077 long nr_pages_same;
6078 int cpu_i;
6079 unsigned long nr_pages;
6080
6081 buffer = container_of(node, struct trace_buffer, node);
6082 if (cpumask_test_cpu(cpu, buffer->cpumask))
6083 return 0;
6084
6085 nr_pages = 0;
6086 nr_pages_same = 1;
6087 /* check if all cpu sizes are same */
6088 for_each_buffer_cpu(buffer, cpu_i) {
6089 /* fill in the size from first enabled cpu */
6090 if (nr_pages == 0)
6091 nr_pages = buffer->buffers[cpu_i]->nr_pages;
6092 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
6093 nr_pages_same = 0;
6094 break;
6095 }
6096 }
6097 /* allocate minimum pages, user can later expand it */
6098 if (!nr_pages_same)
6099 nr_pages = 2;
6100 buffer->buffers[cpu] =
6101 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
6102 if (!buffer->buffers[cpu]) {
6103 WARN(1, "failed to allocate ring buffer on CPU %u\n",
6104 cpu);
6105 return -ENOMEM;
6106 }
6107 smp_wmb();
6108 cpumask_set_cpu(cpu, buffer->cpumask);
6109 return 0;
6110}
6111
6112#ifdef CONFIG_RING_BUFFER_STARTUP_TEST
6113/*
6114 * This is a basic integrity check of the ring buffer.
6115 * Late in the boot cycle this test will run when configured in.
6116 * It will kick off a thread per CPU that will go into a loop
6117 * writing to the per cpu ring buffer various sizes of data.
6118 * Some of the data will be large items, some small.
6119 *
6120 * Another thread is created that goes into a spin, sending out
6121 * IPIs to the other CPUs to also write into the ring buffer.
6122 * this is to test the nesting ability of the buffer.
6123 *
6124 * Basic stats are recorded and reported. If something in the
6125 * ring buffer should happen that's not expected, a big warning
6126 * is displayed and all ring buffers are disabled.
6127 */
6128static struct task_struct *rb_threads[NR_CPUS] __initdata;
6129
6130struct rb_test_data {
6131 struct trace_buffer *buffer;
6132 unsigned long events;
6133 unsigned long bytes_written;
6134 unsigned long bytes_alloc;
6135 unsigned long bytes_dropped;
6136 unsigned long events_nested;
6137 unsigned long bytes_written_nested;
6138 unsigned long bytes_alloc_nested;
6139 unsigned long bytes_dropped_nested;
6140 int min_size_nested;
6141 int max_size_nested;
6142 int max_size;
6143 int min_size;
6144 int cpu;
6145 int cnt;
6146};
6147
6148static struct rb_test_data rb_data[NR_CPUS] __initdata;
6149
6150/* 1 meg per cpu */
6151#define RB_TEST_BUFFER_SIZE 1048576
6152
6153static char rb_string[] __initdata =
6154 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
6155 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
6156 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
6157
6158static bool rb_test_started __initdata;
6159
6160struct rb_item {
6161 int size;
6162 char str[];
6163};
6164
6165static __init int rb_write_something(struct rb_test_data *data, bool nested)
6166{
6167 struct ring_buffer_event *event;
6168 struct rb_item *item;
6169 bool started;
6170 int event_len;
6171 int size;
6172 int len;
6173 int cnt;
6174
6175 /* Have nested writes different that what is written */
6176 cnt = data->cnt + (nested ? 27 : 0);
6177
6178 /* Multiply cnt by ~e, to make some unique increment */
6179 size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
6180
6181 len = size + sizeof(struct rb_item);
6182
6183 started = rb_test_started;
6184 /* read rb_test_started before checking buffer enabled */
6185 smp_rmb();
6186
6187 event = ring_buffer_lock_reserve(data->buffer, len);
6188 if (!event) {
6189 /* Ignore dropped events before test starts. */
6190 if (started) {
6191 if (nested)
6192 data->bytes_dropped += len;
6193 else
6194 data->bytes_dropped_nested += len;
6195 }
6196 return len;
6197 }
6198
6199 event_len = ring_buffer_event_length(event);
6200
6201 if (RB_WARN_ON(data->buffer, event_len < len))
6202 goto out;
6203
6204 item = ring_buffer_event_data(event);
6205 item->size = size;
6206 memcpy(item->str, rb_string, size);
6207
6208 if (nested) {
6209 data->bytes_alloc_nested += event_len;
6210 data->bytes_written_nested += len;
6211 data->events_nested++;
6212 if (!data->min_size_nested || len < data->min_size_nested)
6213 data->min_size_nested = len;
6214 if (len > data->max_size_nested)
6215 data->max_size_nested = len;
6216 } else {
6217 data->bytes_alloc += event_len;
6218 data->bytes_written += len;
6219 data->events++;
6220 if (!data->min_size || len < data->min_size)
6221 data->max_size = len;
6222 if (len > data->max_size)
6223 data->max_size = len;
6224 }
6225
6226 out:
6227 ring_buffer_unlock_commit(data->buffer);
6228
6229 return 0;
6230}
6231
6232static __init int rb_test(void *arg)
6233{
6234 struct rb_test_data *data = arg;
6235
6236 while (!kthread_should_stop()) {
6237 rb_write_something(data, false);
6238 data->cnt++;
6239
6240 set_current_state(TASK_INTERRUPTIBLE);
6241 /* Now sleep between a min of 100-300us and a max of 1ms */
6242 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
6243 }
6244
6245 return 0;
6246}
6247
6248static __init void rb_ipi(void *ignore)
6249{
6250 struct rb_test_data *data;
6251 int cpu = smp_processor_id();
6252
6253 data = &rb_data[cpu];
6254 rb_write_something(data, true);
6255}
6256
6257static __init int rb_hammer_test(void *arg)
6258{
6259 while (!kthread_should_stop()) {
6260
6261 /* Send an IPI to all cpus to write data! */
6262 smp_call_function(rb_ipi, NULL, 1);
6263 /* No sleep, but for non preempt, let others run */
6264 schedule();
6265 }
6266
6267 return 0;
6268}
6269
6270static __init int test_ringbuffer(void)
6271{
6272 struct task_struct *rb_hammer;
6273 struct trace_buffer *buffer;
6274 int cpu;
6275 int ret = 0;
6276
6277 if (security_locked_down(LOCKDOWN_TRACEFS)) {
6278 pr_warn("Lockdown is enabled, skipping ring buffer tests\n");
6279 return 0;
6280 }
6281
6282 pr_info("Running ring buffer tests...\n");
6283
6284 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
6285 if (WARN_ON(!buffer))
6286 return 0;
6287
6288 /* Disable buffer so that threads can't write to it yet */
6289 ring_buffer_record_off(buffer);
6290
6291 for_each_online_cpu(cpu) {
6292 rb_data[cpu].buffer = buffer;
6293 rb_data[cpu].cpu = cpu;
6294 rb_data[cpu].cnt = cpu;
6295 rb_threads[cpu] = kthread_run_on_cpu(rb_test, &rb_data[cpu],
6296 cpu, "rbtester/%u");
6297 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
6298 pr_cont("FAILED\n");
6299 ret = PTR_ERR(rb_threads[cpu]);
6300 goto out_free;
6301 }
6302 }
6303
6304 /* Now create the rb hammer! */
6305 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
6306 if (WARN_ON(IS_ERR(rb_hammer))) {
6307 pr_cont("FAILED\n");
6308 ret = PTR_ERR(rb_hammer);
6309 goto out_free;
6310 }
6311
6312 ring_buffer_record_on(buffer);
6313 /*
6314 * Show buffer is enabled before setting rb_test_started.
6315 * Yes there's a small race window where events could be
6316 * dropped and the thread wont catch it. But when a ring
6317 * buffer gets enabled, there will always be some kind of
6318 * delay before other CPUs see it. Thus, we don't care about
6319 * those dropped events. We care about events dropped after
6320 * the threads see that the buffer is active.
6321 */
6322 smp_wmb();
6323 rb_test_started = true;
6324
6325 set_current_state(TASK_INTERRUPTIBLE);
6326 /* Just run for 10 seconds */;
6327 schedule_timeout(10 * HZ);
6328
6329 kthread_stop(rb_hammer);
6330
6331 out_free:
6332 for_each_online_cpu(cpu) {
6333 if (!rb_threads[cpu])
6334 break;
6335 kthread_stop(rb_threads[cpu]);
6336 }
6337 if (ret) {
6338 ring_buffer_free(buffer);
6339 return ret;
6340 }
6341
6342 /* Report! */
6343 pr_info("finished\n");
6344 for_each_online_cpu(cpu) {
6345 struct ring_buffer_event *event;
6346 struct rb_test_data *data = &rb_data[cpu];
6347 struct rb_item *item;
6348 unsigned long total_events;
6349 unsigned long total_dropped;
6350 unsigned long total_written;
6351 unsigned long total_alloc;
6352 unsigned long total_read = 0;
6353 unsigned long total_size = 0;
6354 unsigned long total_len = 0;
6355 unsigned long total_lost = 0;
6356 unsigned long lost;
6357 int big_event_size;
6358 int small_event_size;
6359
6360 ret = -1;
6361
6362 total_events = data->events + data->events_nested;
6363 total_written = data->bytes_written + data->bytes_written_nested;
6364 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
6365 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
6366
6367 big_event_size = data->max_size + data->max_size_nested;
6368 small_event_size = data->min_size + data->min_size_nested;
6369
6370 pr_info("CPU %d:\n", cpu);
6371 pr_info(" events: %ld\n", total_events);
6372 pr_info(" dropped bytes: %ld\n", total_dropped);
6373 pr_info(" alloced bytes: %ld\n", total_alloc);
6374 pr_info(" written bytes: %ld\n", total_written);
6375 pr_info(" biggest event: %d\n", big_event_size);
6376 pr_info(" smallest event: %d\n", small_event_size);
6377
6378 if (RB_WARN_ON(buffer, total_dropped))
6379 break;
6380
6381 ret = 0;
6382
6383 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
6384 total_lost += lost;
6385 item = ring_buffer_event_data(event);
6386 total_len += ring_buffer_event_length(event);
6387 total_size += item->size + sizeof(struct rb_item);
6388 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
6389 pr_info("FAILED!\n");
6390 pr_info("buffer had: %.*s\n", item->size, item->str);
6391 pr_info("expected: %.*s\n", item->size, rb_string);
6392 RB_WARN_ON(buffer, 1);
6393 ret = -1;
6394 break;
6395 }
6396 total_read++;
6397 }
6398 if (ret)
6399 break;
6400
6401 ret = -1;
6402
6403 pr_info(" read events: %ld\n", total_read);
6404 pr_info(" lost events: %ld\n", total_lost);
6405 pr_info(" total events: %ld\n", total_lost + total_read);
6406 pr_info(" recorded len bytes: %ld\n", total_len);
6407 pr_info(" recorded size bytes: %ld\n", total_size);
6408 if (total_lost) {
6409 pr_info(" With dropped events, record len and size may not match\n"
6410 " alloced and written from above\n");
6411 } else {
6412 if (RB_WARN_ON(buffer, total_len != total_alloc ||
6413 total_size != total_written))
6414 break;
6415 }
6416 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
6417 break;
6418
6419 ret = 0;
6420 }
6421 if (!ret)
6422 pr_info("Ring buffer PASSED!\n");
6423
6424 ring_buffer_free(buffer);
6425 return 0;
6426}
6427
6428late_initcall(test_ringbuffer);
6429#endif /* CONFIG_RING_BUFFER_STARTUP_TEST */
1/*
2 * Generic ring buffer
3 *
4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
5 */
6#include <linux/trace_events.h>
7#include <linux/ring_buffer.h>
8#include <linux/trace_clock.h>
9#include <linux/trace_seq.h>
10#include <linux/spinlock.h>
11#include <linux/irq_work.h>
12#include <linux/uaccess.h>
13#include <linux/hardirq.h>
14#include <linux/kthread.h> /* for self test */
15#include <linux/kmemcheck.h>
16#include <linux/module.h>
17#include <linux/percpu.h>
18#include <linux/mutex.h>
19#include <linux/delay.h>
20#include <linux/slab.h>
21#include <linux/init.h>
22#include <linux/hash.h>
23#include <linux/list.h>
24#include <linux/cpu.h>
25
26#include <asm/local.h>
27
28static void update_pages_handler(struct work_struct *work);
29
30/*
31 * The ring buffer header is special. We must manually up keep it.
32 */
33int ring_buffer_print_entry_header(struct trace_seq *s)
34{
35 trace_seq_puts(s, "# compressed entry header\n");
36 trace_seq_puts(s, "\ttype_len : 5 bits\n");
37 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
38 trace_seq_puts(s, "\tarray : 32 bits\n");
39 trace_seq_putc(s, '\n');
40 trace_seq_printf(s, "\tpadding : type == %d\n",
41 RINGBUF_TYPE_PADDING);
42 trace_seq_printf(s, "\ttime_extend : type == %d\n",
43 RINGBUF_TYPE_TIME_EXTEND);
44 trace_seq_printf(s, "\tdata max type_len == %d\n",
45 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
46
47 return !trace_seq_has_overflowed(s);
48}
49
50/*
51 * The ring buffer is made up of a list of pages. A separate list of pages is
52 * allocated for each CPU. A writer may only write to a buffer that is
53 * associated with the CPU it is currently executing on. A reader may read
54 * from any per cpu buffer.
55 *
56 * The reader is special. For each per cpu buffer, the reader has its own
57 * reader page. When a reader has read the entire reader page, this reader
58 * page is swapped with another page in the ring buffer.
59 *
60 * Now, as long as the writer is off the reader page, the reader can do what
61 * ever it wants with that page. The writer will never write to that page
62 * again (as long as it is out of the ring buffer).
63 *
64 * Here's some silly ASCII art.
65 *
66 * +------+
67 * |reader| RING BUFFER
68 * |page |
69 * +------+ +---+ +---+ +---+
70 * | |-->| |-->| |
71 * +---+ +---+ +---+
72 * ^ |
73 * | |
74 * +---------------+
75 *
76 *
77 * +------+
78 * |reader| RING BUFFER
79 * |page |------------------v
80 * +------+ +---+ +---+ +---+
81 * | |-->| |-->| |
82 * +---+ +---+ +---+
83 * ^ |
84 * | |
85 * +---------------+
86 *
87 *
88 * +------+
89 * |reader| RING BUFFER
90 * |page |------------------v
91 * +------+ +---+ +---+ +---+
92 * ^ | |-->| |-->| |
93 * | +---+ +---+ +---+
94 * | |
95 * | |
96 * +------------------------------+
97 *
98 *
99 * +------+
100 * |buffer| RING BUFFER
101 * |page |------------------v
102 * +------+ +---+ +---+ +---+
103 * ^ | | | |-->| |
104 * | New +---+ +---+ +---+
105 * | Reader------^ |
106 * | page |
107 * +------------------------------+
108 *
109 *
110 * After we make this swap, the reader can hand this page off to the splice
111 * code and be done with it. It can even allocate a new page if it needs to
112 * and swap that into the ring buffer.
113 *
114 * We will be using cmpxchg soon to make all this lockless.
115 *
116 */
117
118/* Used for individual buffers (after the counter) */
119#define RB_BUFFER_OFF (1 << 20)
120
121#define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
122
123#define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
124#define RB_ALIGNMENT 4U
125#define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
126#define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
127
128#ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
129# define RB_FORCE_8BYTE_ALIGNMENT 0
130# define RB_ARCH_ALIGNMENT RB_ALIGNMENT
131#else
132# define RB_FORCE_8BYTE_ALIGNMENT 1
133# define RB_ARCH_ALIGNMENT 8U
134#endif
135
136#define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
137
138/* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
139#define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
140
141enum {
142 RB_LEN_TIME_EXTEND = 8,
143 RB_LEN_TIME_STAMP = 16,
144};
145
146#define skip_time_extend(event) \
147 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
148
149static inline int rb_null_event(struct ring_buffer_event *event)
150{
151 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
152}
153
154static void rb_event_set_padding(struct ring_buffer_event *event)
155{
156 /* padding has a NULL time_delta */
157 event->type_len = RINGBUF_TYPE_PADDING;
158 event->time_delta = 0;
159}
160
161static unsigned
162rb_event_data_length(struct ring_buffer_event *event)
163{
164 unsigned length;
165
166 if (event->type_len)
167 length = event->type_len * RB_ALIGNMENT;
168 else
169 length = event->array[0];
170 return length + RB_EVNT_HDR_SIZE;
171}
172
173/*
174 * Return the length of the given event. Will return
175 * the length of the time extend if the event is a
176 * time extend.
177 */
178static inline unsigned
179rb_event_length(struct ring_buffer_event *event)
180{
181 switch (event->type_len) {
182 case RINGBUF_TYPE_PADDING:
183 if (rb_null_event(event))
184 /* undefined */
185 return -1;
186 return event->array[0] + RB_EVNT_HDR_SIZE;
187
188 case RINGBUF_TYPE_TIME_EXTEND:
189 return RB_LEN_TIME_EXTEND;
190
191 case RINGBUF_TYPE_TIME_STAMP:
192 return RB_LEN_TIME_STAMP;
193
194 case RINGBUF_TYPE_DATA:
195 return rb_event_data_length(event);
196 default:
197 BUG();
198 }
199 /* not hit */
200 return 0;
201}
202
203/*
204 * Return total length of time extend and data,
205 * or just the event length for all other events.
206 */
207static inline unsigned
208rb_event_ts_length(struct ring_buffer_event *event)
209{
210 unsigned len = 0;
211
212 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
213 /* time extends include the data event after it */
214 len = RB_LEN_TIME_EXTEND;
215 event = skip_time_extend(event);
216 }
217 return len + rb_event_length(event);
218}
219
220/**
221 * ring_buffer_event_length - return the length of the event
222 * @event: the event to get the length of
223 *
224 * Returns the size of the data load of a data event.
225 * If the event is something other than a data event, it
226 * returns the size of the event itself. With the exception
227 * of a TIME EXTEND, where it still returns the size of the
228 * data load of the data event after it.
229 */
230unsigned ring_buffer_event_length(struct ring_buffer_event *event)
231{
232 unsigned length;
233
234 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
235 event = skip_time_extend(event);
236
237 length = rb_event_length(event);
238 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
239 return length;
240 length -= RB_EVNT_HDR_SIZE;
241 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
242 length -= sizeof(event->array[0]);
243 return length;
244}
245EXPORT_SYMBOL_GPL(ring_buffer_event_length);
246
247/* inline for ring buffer fast paths */
248static __always_inline void *
249rb_event_data(struct ring_buffer_event *event)
250{
251 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
252 event = skip_time_extend(event);
253 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
254 /* If length is in len field, then array[0] has the data */
255 if (event->type_len)
256 return (void *)&event->array[0];
257 /* Otherwise length is in array[0] and array[1] has the data */
258 return (void *)&event->array[1];
259}
260
261/**
262 * ring_buffer_event_data - return the data of the event
263 * @event: the event to get the data from
264 */
265void *ring_buffer_event_data(struct ring_buffer_event *event)
266{
267 return rb_event_data(event);
268}
269EXPORT_SYMBOL_GPL(ring_buffer_event_data);
270
271#define for_each_buffer_cpu(buffer, cpu) \
272 for_each_cpu(cpu, buffer->cpumask)
273
274#define TS_SHIFT 27
275#define TS_MASK ((1ULL << TS_SHIFT) - 1)
276#define TS_DELTA_TEST (~TS_MASK)
277
278/* Flag when events were overwritten */
279#define RB_MISSED_EVENTS (1 << 31)
280/* Missed count stored at end */
281#define RB_MISSED_STORED (1 << 30)
282
283struct buffer_data_page {
284 u64 time_stamp; /* page time stamp */
285 local_t commit; /* write committed index */
286 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
287};
288
289/*
290 * Note, the buffer_page list must be first. The buffer pages
291 * are allocated in cache lines, which means that each buffer
292 * page will be at the beginning of a cache line, and thus
293 * the least significant bits will be zero. We use this to
294 * add flags in the list struct pointers, to make the ring buffer
295 * lockless.
296 */
297struct buffer_page {
298 struct list_head list; /* list of buffer pages */
299 local_t write; /* index for next write */
300 unsigned read; /* index for next read */
301 local_t entries; /* entries on this page */
302 unsigned long real_end; /* real end of data */
303 struct buffer_data_page *page; /* Actual data page */
304};
305
306/*
307 * The buffer page counters, write and entries, must be reset
308 * atomically when crossing page boundaries. To synchronize this
309 * update, two counters are inserted into the number. One is
310 * the actual counter for the write position or count on the page.
311 *
312 * The other is a counter of updaters. Before an update happens
313 * the update partition of the counter is incremented. This will
314 * allow the updater to update the counter atomically.
315 *
316 * The counter is 20 bits, and the state data is 12.
317 */
318#define RB_WRITE_MASK 0xfffff
319#define RB_WRITE_INTCNT (1 << 20)
320
321static void rb_init_page(struct buffer_data_page *bpage)
322{
323 local_set(&bpage->commit, 0);
324}
325
326/**
327 * ring_buffer_page_len - the size of data on the page.
328 * @page: The page to read
329 *
330 * Returns the amount of data on the page, including buffer page header.
331 */
332size_t ring_buffer_page_len(void *page)
333{
334 return local_read(&((struct buffer_data_page *)page)->commit)
335 + BUF_PAGE_HDR_SIZE;
336}
337
338/*
339 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
340 * this issue out.
341 */
342static void free_buffer_page(struct buffer_page *bpage)
343{
344 free_page((unsigned long)bpage->page);
345 kfree(bpage);
346}
347
348/*
349 * We need to fit the time_stamp delta into 27 bits.
350 */
351static inline int test_time_stamp(u64 delta)
352{
353 if (delta & TS_DELTA_TEST)
354 return 1;
355 return 0;
356}
357
358#define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
359
360/* Max payload is BUF_PAGE_SIZE - header (8bytes) */
361#define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
362
363int ring_buffer_print_page_header(struct trace_seq *s)
364{
365 struct buffer_data_page field;
366
367 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
368 "offset:0;\tsize:%u;\tsigned:%u;\n",
369 (unsigned int)sizeof(field.time_stamp),
370 (unsigned int)is_signed_type(u64));
371
372 trace_seq_printf(s, "\tfield: local_t commit;\t"
373 "offset:%u;\tsize:%u;\tsigned:%u;\n",
374 (unsigned int)offsetof(typeof(field), commit),
375 (unsigned int)sizeof(field.commit),
376 (unsigned int)is_signed_type(long));
377
378 trace_seq_printf(s, "\tfield: int overwrite;\t"
379 "offset:%u;\tsize:%u;\tsigned:%u;\n",
380 (unsigned int)offsetof(typeof(field), commit),
381 1,
382 (unsigned int)is_signed_type(long));
383
384 trace_seq_printf(s, "\tfield: char data;\t"
385 "offset:%u;\tsize:%u;\tsigned:%u;\n",
386 (unsigned int)offsetof(typeof(field), data),
387 (unsigned int)BUF_PAGE_SIZE,
388 (unsigned int)is_signed_type(char));
389
390 return !trace_seq_has_overflowed(s);
391}
392
393struct rb_irq_work {
394 struct irq_work work;
395 wait_queue_head_t waiters;
396 wait_queue_head_t full_waiters;
397 bool waiters_pending;
398 bool full_waiters_pending;
399 bool wakeup_full;
400};
401
402/*
403 * Structure to hold event state and handle nested events.
404 */
405struct rb_event_info {
406 u64 ts;
407 u64 delta;
408 unsigned long length;
409 struct buffer_page *tail_page;
410 int add_timestamp;
411};
412
413/*
414 * Used for which event context the event is in.
415 * NMI = 0
416 * IRQ = 1
417 * SOFTIRQ = 2
418 * NORMAL = 3
419 *
420 * See trace_recursive_lock() comment below for more details.
421 */
422enum {
423 RB_CTX_NMI,
424 RB_CTX_IRQ,
425 RB_CTX_SOFTIRQ,
426 RB_CTX_NORMAL,
427 RB_CTX_MAX
428};
429
430/*
431 * head_page == tail_page && head == tail then buffer is empty.
432 */
433struct ring_buffer_per_cpu {
434 int cpu;
435 atomic_t record_disabled;
436 struct ring_buffer *buffer;
437 raw_spinlock_t reader_lock; /* serialize readers */
438 arch_spinlock_t lock;
439 struct lock_class_key lock_key;
440 unsigned long nr_pages;
441 unsigned int current_context;
442 struct list_head *pages;
443 struct buffer_page *head_page; /* read from head */
444 struct buffer_page *tail_page; /* write to tail */
445 struct buffer_page *commit_page; /* committed pages */
446 struct buffer_page *reader_page;
447 unsigned long lost_events;
448 unsigned long last_overrun;
449 local_t entries_bytes;
450 local_t entries;
451 local_t overrun;
452 local_t commit_overrun;
453 local_t dropped_events;
454 local_t committing;
455 local_t commits;
456 unsigned long read;
457 unsigned long read_bytes;
458 u64 write_stamp;
459 u64 read_stamp;
460 /* ring buffer pages to update, > 0 to add, < 0 to remove */
461 long nr_pages_to_update;
462 struct list_head new_pages; /* new pages to add */
463 struct work_struct update_pages_work;
464 struct completion update_done;
465
466 struct rb_irq_work irq_work;
467};
468
469struct ring_buffer {
470 unsigned flags;
471 int cpus;
472 atomic_t record_disabled;
473 atomic_t resize_disabled;
474 cpumask_var_t cpumask;
475
476 struct lock_class_key *reader_lock_key;
477
478 struct mutex mutex;
479
480 struct ring_buffer_per_cpu **buffers;
481
482 struct hlist_node node;
483 u64 (*clock)(void);
484
485 struct rb_irq_work irq_work;
486};
487
488struct ring_buffer_iter {
489 struct ring_buffer_per_cpu *cpu_buffer;
490 unsigned long head;
491 struct buffer_page *head_page;
492 struct buffer_page *cache_reader_page;
493 unsigned long cache_read;
494 u64 read_stamp;
495};
496
497/*
498 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
499 *
500 * Schedules a delayed work to wake up any task that is blocked on the
501 * ring buffer waiters queue.
502 */
503static void rb_wake_up_waiters(struct irq_work *work)
504{
505 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
506
507 wake_up_all(&rbwork->waiters);
508 if (rbwork->wakeup_full) {
509 rbwork->wakeup_full = false;
510 wake_up_all(&rbwork->full_waiters);
511 }
512}
513
514/**
515 * ring_buffer_wait - wait for input to the ring buffer
516 * @buffer: buffer to wait on
517 * @cpu: the cpu buffer to wait on
518 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
519 *
520 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
521 * as data is added to any of the @buffer's cpu buffers. Otherwise
522 * it will wait for data to be added to a specific cpu buffer.
523 */
524int ring_buffer_wait(struct ring_buffer *buffer, int cpu, bool full)
525{
526 struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
527 DEFINE_WAIT(wait);
528 struct rb_irq_work *work;
529 int ret = 0;
530
531 /*
532 * Depending on what the caller is waiting for, either any
533 * data in any cpu buffer, or a specific buffer, put the
534 * caller on the appropriate wait queue.
535 */
536 if (cpu == RING_BUFFER_ALL_CPUS) {
537 work = &buffer->irq_work;
538 /* Full only makes sense on per cpu reads */
539 full = false;
540 } else {
541 if (!cpumask_test_cpu(cpu, buffer->cpumask))
542 return -ENODEV;
543 cpu_buffer = buffer->buffers[cpu];
544 work = &cpu_buffer->irq_work;
545 }
546
547
548 while (true) {
549 if (full)
550 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
551 else
552 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
553
554 /*
555 * The events can happen in critical sections where
556 * checking a work queue can cause deadlocks.
557 * After adding a task to the queue, this flag is set
558 * only to notify events to try to wake up the queue
559 * using irq_work.
560 *
561 * We don't clear it even if the buffer is no longer
562 * empty. The flag only causes the next event to run
563 * irq_work to do the work queue wake up. The worse
564 * that can happen if we race with !trace_empty() is that
565 * an event will cause an irq_work to try to wake up
566 * an empty queue.
567 *
568 * There's no reason to protect this flag either, as
569 * the work queue and irq_work logic will do the necessary
570 * synchronization for the wake ups. The only thing
571 * that is necessary is that the wake up happens after
572 * a task has been queued. It's OK for spurious wake ups.
573 */
574 if (full)
575 work->full_waiters_pending = true;
576 else
577 work->waiters_pending = true;
578
579 if (signal_pending(current)) {
580 ret = -EINTR;
581 break;
582 }
583
584 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
585 break;
586
587 if (cpu != RING_BUFFER_ALL_CPUS &&
588 !ring_buffer_empty_cpu(buffer, cpu)) {
589 unsigned long flags;
590 bool pagebusy;
591
592 if (!full)
593 break;
594
595 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
596 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
597 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
598
599 if (!pagebusy)
600 break;
601 }
602
603 schedule();
604 }
605
606 if (full)
607 finish_wait(&work->full_waiters, &wait);
608 else
609 finish_wait(&work->waiters, &wait);
610
611 return ret;
612}
613
614/**
615 * ring_buffer_poll_wait - poll on buffer input
616 * @buffer: buffer to wait on
617 * @cpu: the cpu buffer to wait on
618 * @filp: the file descriptor
619 * @poll_table: The poll descriptor
620 *
621 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
622 * as data is added to any of the @buffer's cpu buffers. Otherwise
623 * it will wait for data to be added to a specific cpu buffer.
624 *
625 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
626 * zero otherwise.
627 */
628int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
629 struct file *filp, poll_table *poll_table)
630{
631 struct ring_buffer_per_cpu *cpu_buffer;
632 struct rb_irq_work *work;
633
634 if (cpu == RING_BUFFER_ALL_CPUS)
635 work = &buffer->irq_work;
636 else {
637 if (!cpumask_test_cpu(cpu, buffer->cpumask))
638 return -EINVAL;
639
640 cpu_buffer = buffer->buffers[cpu];
641 work = &cpu_buffer->irq_work;
642 }
643
644 poll_wait(filp, &work->waiters, poll_table);
645 work->waiters_pending = true;
646 /*
647 * There's a tight race between setting the waiters_pending and
648 * checking if the ring buffer is empty. Once the waiters_pending bit
649 * is set, the next event will wake the task up, but we can get stuck
650 * if there's only a single event in.
651 *
652 * FIXME: Ideally, we need a memory barrier on the writer side as well,
653 * but adding a memory barrier to all events will cause too much of a
654 * performance hit in the fast path. We only need a memory barrier when
655 * the buffer goes from empty to having content. But as this race is
656 * extremely small, and it's not a problem if another event comes in, we
657 * will fix it later.
658 */
659 smp_mb();
660
661 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
662 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
663 return POLLIN | POLLRDNORM;
664 return 0;
665}
666
667/* buffer may be either ring_buffer or ring_buffer_per_cpu */
668#define RB_WARN_ON(b, cond) \
669 ({ \
670 int _____ret = unlikely(cond); \
671 if (_____ret) { \
672 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
673 struct ring_buffer_per_cpu *__b = \
674 (void *)b; \
675 atomic_inc(&__b->buffer->record_disabled); \
676 } else \
677 atomic_inc(&b->record_disabled); \
678 WARN_ON(1); \
679 } \
680 _____ret; \
681 })
682
683/* Up this if you want to test the TIME_EXTENTS and normalization */
684#define DEBUG_SHIFT 0
685
686static inline u64 rb_time_stamp(struct ring_buffer *buffer)
687{
688 /* shift to debug/test normalization and TIME_EXTENTS */
689 return buffer->clock() << DEBUG_SHIFT;
690}
691
692u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
693{
694 u64 time;
695
696 preempt_disable_notrace();
697 time = rb_time_stamp(buffer);
698 preempt_enable_no_resched_notrace();
699
700 return time;
701}
702EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
703
704void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
705 int cpu, u64 *ts)
706{
707 /* Just stupid testing the normalize function and deltas */
708 *ts >>= DEBUG_SHIFT;
709}
710EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
711
712/*
713 * Making the ring buffer lockless makes things tricky.
714 * Although writes only happen on the CPU that they are on,
715 * and they only need to worry about interrupts. Reads can
716 * happen on any CPU.
717 *
718 * The reader page is always off the ring buffer, but when the
719 * reader finishes with a page, it needs to swap its page with
720 * a new one from the buffer. The reader needs to take from
721 * the head (writes go to the tail). But if a writer is in overwrite
722 * mode and wraps, it must push the head page forward.
723 *
724 * Here lies the problem.
725 *
726 * The reader must be careful to replace only the head page, and
727 * not another one. As described at the top of the file in the
728 * ASCII art, the reader sets its old page to point to the next
729 * page after head. It then sets the page after head to point to
730 * the old reader page. But if the writer moves the head page
731 * during this operation, the reader could end up with the tail.
732 *
733 * We use cmpxchg to help prevent this race. We also do something
734 * special with the page before head. We set the LSB to 1.
735 *
736 * When the writer must push the page forward, it will clear the
737 * bit that points to the head page, move the head, and then set
738 * the bit that points to the new head page.
739 *
740 * We also don't want an interrupt coming in and moving the head
741 * page on another writer. Thus we use the second LSB to catch
742 * that too. Thus:
743 *
744 * head->list->prev->next bit 1 bit 0
745 * ------- -------
746 * Normal page 0 0
747 * Points to head page 0 1
748 * New head page 1 0
749 *
750 * Note we can not trust the prev pointer of the head page, because:
751 *
752 * +----+ +-----+ +-----+
753 * | |------>| T |---X--->| N |
754 * | |<------| | | |
755 * +----+ +-----+ +-----+
756 * ^ ^ |
757 * | +-----+ | |
758 * +----------| R |----------+ |
759 * | |<-----------+
760 * +-----+
761 *
762 * Key: ---X--> HEAD flag set in pointer
763 * T Tail page
764 * R Reader page
765 * N Next page
766 *
767 * (see __rb_reserve_next() to see where this happens)
768 *
769 * What the above shows is that the reader just swapped out
770 * the reader page with a page in the buffer, but before it
771 * could make the new header point back to the new page added
772 * it was preempted by a writer. The writer moved forward onto
773 * the new page added by the reader and is about to move forward
774 * again.
775 *
776 * You can see, it is legitimate for the previous pointer of
777 * the head (or any page) not to point back to itself. But only
778 * temporarially.
779 */
780
781#define RB_PAGE_NORMAL 0UL
782#define RB_PAGE_HEAD 1UL
783#define RB_PAGE_UPDATE 2UL
784
785
786#define RB_FLAG_MASK 3UL
787
788/* PAGE_MOVED is not part of the mask */
789#define RB_PAGE_MOVED 4UL
790
791/*
792 * rb_list_head - remove any bit
793 */
794static struct list_head *rb_list_head(struct list_head *list)
795{
796 unsigned long val = (unsigned long)list;
797
798 return (struct list_head *)(val & ~RB_FLAG_MASK);
799}
800
801/*
802 * rb_is_head_page - test if the given page is the head page
803 *
804 * Because the reader may move the head_page pointer, we can
805 * not trust what the head page is (it may be pointing to
806 * the reader page). But if the next page is a header page,
807 * its flags will be non zero.
808 */
809static inline int
810rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
811 struct buffer_page *page, struct list_head *list)
812{
813 unsigned long val;
814
815 val = (unsigned long)list->next;
816
817 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
818 return RB_PAGE_MOVED;
819
820 return val & RB_FLAG_MASK;
821}
822
823/*
824 * rb_is_reader_page
825 *
826 * The unique thing about the reader page, is that, if the
827 * writer is ever on it, the previous pointer never points
828 * back to the reader page.
829 */
830static bool rb_is_reader_page(struct buffer_page *page)
831{
832 struct list_head *list = page->list.prev;
833
834 return rb_list_head(list->next) != &page->list;
835}
836
837/*
838 * rb_set_list_to_head - set a list_head to be pointing to head.
839 */
840static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
841 struct list_head *list)
842{
843 unsigned long *ptr;
844
845 ptr = (unsigned long *)&list->next;
846 *ptr |= RB_PAGE_HEAD;
847 *ptr &= ~RB_PAGE_UPDATE;
848}
849
850/*
851 * rb_head_page_activate - sets up head page
852 */
853static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
854{
855 struct buffer_page *head;
856
857 head = cpu_buffer->head_page;
858 if (!head)
859 return;
860
861 /*
862 * Set the previous list pointer to have the HEAD flag.
863 */
864 rb_set_list_to_head(cpu_buffer, head->list.prev);
865}
866
867static void rb_list_head_clear(struct list_head *list)
868{
869 unsigned long *ptr = (unsigned long *)&list->next;
870
871 *ptr &= ~RB_FLAG_MASK;
872}
873
874/*
875 * rb_head_page_dactivate - clears head page ptr (for free list)
876 */
877static void
878rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
879{
880 struct list_head *hd;
881
882 /* Go through the whole list and clear any pointers found. */
883 rb_list_head_clear(cpu_buffer->pages);
884
885 list_for_each(hd, cpu_buffer->pages)
886 rb_list_head_clear(hd);
887}
888
889static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
890 struct buffer_page *head,
891 struct buffer_page *prev,
892 int old_flag, int new_flag)
893{
894 struct list_head *list;
895 unsigned long val = (unsigned long)&head->list;
896 unsigned long ret;
897
898 list = &prev->list;
899
900 val &= ~RB_FLAG_MASK;
901
902 ret = cmpxchg((unsigned long *)&list->next,
903 val | old_flag, val | new_flag);
904
905 /* check if the reader took the page */
906 if ((ret & ~RB_FLAG_MASK) != val)
907 return RB_PAGE_MOVED;
908
909 return ret & RB_FLAG_MASK;
910}
911
912static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
913 struct buffer_page *head,
914 struct buffer_page *prev,
915 int old_flag)
916{
917 return rb_head_page_set(cpu_buffer, head, prev,
918 old_flag, RB_PAGE_UPDATE);
919}
920
921static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
922 struct buffer_page *head,
923 struct buffer_page *prev,
924 int old_flag)
925{
926 return rb_head_page_set(cpu_buffer, head, prev,
927 old_flag, RB_PAGE_HEAD);
928}
929
930static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
931 struct buffer_page *head,
932 struct buffer_page *prev,
933 int old_flag)
934{
935 return rb_head_page_set(cpu_buffer, head, prev,
936 old_flag, RB_PAGE_NORMAL);
937}
938
939static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
940 struct buffer_page **bpage)
941{
942 struct list_head *p = rb_list_head((*bpage)->list.next);
943
944 *bpage = list_entry(p, struct buffer_page, list);
945}
946
947static struct buffer_page *
948rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
949{
950 struct buffer_page *head;
951 struct buffer_page *page;
952 struct list_head *list;
953 int i;
954
955 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
956 return NULL;
957
958 /* sanity check */
959 list = cpu_buffer->pages;
960 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
961 return NULL;
962
963 page = head = cpu_buffer->head_page;
964 /*
965 * It is possible that the writer moves the header behind
966 * where we started, and we miss in one loop.
967 * A second loop should grab the header, but we'll do
968 * three loops just because I'm paranoid.
969 */
970 for (i = 0; i < 3; i++) {
971 do {
972 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
973 cpu_buffer->head_page = page;
974 return page;
975 }
976 rb_inc_page(cpu_buffer, &page);
977 } while (page != head);
978 }
979
980 RB_WARN_ON(cpu_buffer, 1);
981
982 return NULL;
983}
984
985static int rb_head_page_replace(struct buffer_page *old,
986 struct buffer_page *new)
987{
988 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
989 unsigned long val;
990 unsigned long ret;
991
992 val = *ptr & ~RB_FLAG_MASK;
993 val |= RB_PAGE_HEAD;
994
995 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
996
997 return ret == val;
998}
999
1000/*
1001 * rb_tail_page_update - move the tail page forward
1002 */
1003static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1004 struct buffer_page *tail_page,
1005 struct buffer_page *next_page)
1006{
1007 unsigned long old_entries;
1008 unsigned long old_write;
1009
1010 /*
1011 * The tail page now needs to be moved forward.
1012 *
1013 * We need to reset the tail page, but without messing
1014 * with possible erasing of data brought in by interrupts
1015 * that have moved the tail page and are currently on it.
1016 *
1017 * We add a counter to the write field to denote this.
1018 */
1019 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1020 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1021
1022 /*
1023 * Just make sure we have seen our old_write and synchronize
1024 * with any interrupts that come in.
1025 */
1026 barrier();
1027
1028 /*
1029 * If the tail page is still the same as what we think
1030 * it is, then it is up to us to update the tail
1031 * pointer.
1032 */
1033 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1034 /* Zero the write counter */
1035 unsigned long val = old_write & ~RB_WRITE_MASK;
1036 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1037
1038 /*
1039 * This will only succeed if an interrupt did
1040 * not come in and change it. In which case, we
1041 * do not want to modify it.
1042 *
1043 * We add (void) to let the compiler know that we do not care
1044 * about the return value of these functions. We use the
1045 * cmpxchg to only update if an interrupt did not already
1046 * do it for us. If the cmpxchg fails, we don't care.
1047 */
1048 (void)local_cmpxchg(&next_page->write, old_write, val);
1049 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1050
1051 /*
1052 * No need to worry about races with clearing out the commit.
1053 * it only can increment when a commit takes place. But that
1054 * only happens in the outer most nested commit.
1055 */
1056 local_set(&next_page->page->commit, 0);
1057
1058 /* Again, either we update tail_page or an interrupt does */
1059 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1060 }
1061}
1062
1063static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1064 struct buffer_page *bpage)
1065{
1066 unsigned long val = (unsigned long)bpage;
1067
1068 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1069 return 1;
1070
1071 return 0;
1072}
1073
1074/**
1075 * rb_check_list - make sure a pointer to a list has the last bits zero
1076 */
1077static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1078 struct list_head *list)
1079{
1080 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1081 return 1;
1082 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1083 return 1;
1084 return 0;
1085}
1086
1087/**
1088 * rb_check_pages - integrity check of buffer pages
1089 * @cpu_buffer: CPU buffer with pages to test
1090 *
1091 * As a safety measure we check to make sure the data pages have not
1092 * been corrupted.
1093 */
1094static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1095{
1096 struct list_head *head = cpu_buffer->pages;
1097 struct buffer_page *bpage, *tmp;
1098
1099 /* Reset the head page if it exists */
1100 if (cpu_buffer->head_page)
1101 rb_set_head_page(cpu_buffer);
1102
1103 rb_head_page_deactivate(cpu_buffer);
1104
1105 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1106 return -1;
1107 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1108 return -1;
1109
1110 if (rb_check_list(cpu_buffer, head))
1111 return -1;
1112
1113 list_for_each_entry_safe(bpage, tmp, head, list) {
1114 if (RB_WARN_ON(cpu_buffer,
1115 bpage->list.next->prev != &bpage->list))
1116 return -1;
1117 if (RB_WARN_ON(cpu_buffer,
1118 bpage->list.prev->next != &bpage->list))
1119 return -1;
1120 if (rb_check_list(cpu_buffer, &bpage->list))
1121 return -1;
1122 }
1123
1124 rb_head_page_activate(cpu_buffer);
1125
1126 return 0;
1127}
1128
1129static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1130{
1131 struct buffer_page *bpage, *tmp;
1132 long i;
1133
1134 for (i = 0; i < nr_pages; i++) {
1135 struct page *page;
1136 /*
1137 * __GFP_NORETRY flag makes sure that the allocation fails
1138 * gracefully without invoking oom-killer and the system is
1139 * not destabilized.
1140 */
1141 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1142 GFP_KERNEL | __GFP_NORETRY,
1143 cpu_to_node(cpu));
1144 if (!bpage)
1145 goto free_pages;
1146
1147 list_add(&bpage->list, pages);
1148
1149 page = alloc_pages_node(cpu_to_node(cpu),
1150 GFP_KERNEL | __GFP_NORETRY, 0);
1151 if (!page)
1152 goto free_pages;
1153 bpage->page = page_address(page);
1154 rb_init_page(bpage->page);
1155 }
1156
1157 return 0;
1158
1159free_pages:
1160 list_for_each_entry_safe(bpage, tmp, pages, list) {
1161 list_del_init(&bpage->list);
1162 free_buffer_page(bpage);
1163 }
1164
1165 return -ENOMEM;
1166}
1167
1168static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1169 unsigned long nr_pages)
1170{
1171 LIST_HEAD(pages);
1172
1173 WARN_ON(!nr_pages);
1174
1175 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1176 return -ENOMEM;
1177
1178 /*
1179 * The ring buffer page list is a circular list that does not
1180 * start and end with a list head. All page list items point to
1181 * other pages.
1182 */
1183 cpu_buffer->pages = pages.next;
1184 list_del(&pages);
1185
1186 cpu_buffer->nr_pages = nr_pages;
1187
1188 rb_check_pages(cpu_buffer);
1189
1190 return 0;
1191}
1192
1193static struct ring_buffer_per_cpu *
1194rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1195{
1196 struct ring_buffer_per_cpu *cpu_buffer;
1197 struct buffer_page *bpage;
1198 struct page *page;
1199 int ret;
1200
1201 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1202 GFP_KERNEL, cpu_to_node(cpu));
1203 if (!cpu_buffer)
1204 return NULL;
1205
1206 cpu_buffer->cpu = cpu;
1207 cpu_buffer->buffer = buffer;
1208 raw_spin_lock_init(&cpu_buffer->reader_lock);
1209 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1210 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1211 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1212 init_completion(&cpu_buffer->update_done);
1213 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1214 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1215 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1216
1217 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1218 GFP_KERNEL, cpu_to_node(cpu));
1219 if (!bpage)
1220 goto fail_free_buffer;
1221
1222 rb_check_bpage(cpu_buffer, bpage);
1223
1224 cpu_buffer->reader_page = bpage;
1225 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1226 if (!page)
1227 goto fail_free_reader;
1228 bpage->page = page_address(page);
1229 rb_init_page(bpage->page);
1230
1231 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1232 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1233
1234 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1235 if (ret < 0)
1236 goto fail_free_reader;
1237
1238 cpu_buffer->head_page
1239 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1240 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1241
1242 rb_head_page_activate(cpu_buffer);
1243
1244 return cpu_buffer;
1245
1246 fail_free_reader:
1247 free_buffer_page(cpu_buffer->reader_page);
1248
1249 fail_free_buffer:
1250 kfree(cpu_buffer);
1251 return NULL;
1252}
1253
1254static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1255{
1256 struct list_head *head = cpu_buffer->pages;
1257 struct buffer_page *bpage, *tmp;
1258
1259 free_buffer_page(cpu_buffer->reader_page);
1260
1261 rb_head_page_deactivate(cpu_buffer);
1262
1263 if (head) {
1264 list_for_each_entry_safe(bpage, tmp, head, list) {
1265 list_del_init(&bpage->list);
1266 free_buffer_page(bpage);
1267 }
1268 bpage = list_entry(head, struct buffer_page, list);
1269 free_buffer_page(bpage);
1270 }
1271
1272 kfree(cpu_buffer);
1273}
1274
1275/**
1276 * __ring_buffer_alloc - allocate a new ring_buffer
1277 * @size: the size in bytes per cpu that is needed.
1278 * @flags: attributes to set for the ring buffer.
1279 *
1280 * Currently the only flag that is available is the RB_FL_OVERWRITE
1281 * flag. This flag means that the buffer will overwrite old data
1282 * when the buffer wraps. If this flag is not set, the buffer will
1283 * drop data when the tail hits the head.
1284 */
1285struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1286 struct lock_class_key *key)
1287{
1288 struct ring_buffer *buffer;
1289 long nr_pages;
1290 int bsize;
1291 int cpu;
1292 int ret;
1293
1294 /* keep it in its own cache line */
1295 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1296 GFP_KERNEL);
1297 if (!buffer)
1298 return NULL;
1299
1300 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1301 goto fail_free_buffer;
1302
1303 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1304 buffer->flags = flags;
1305 buffer->clock = trace_clock_local;
1306 buffer->reader_lock_key = key;
1307
1308 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1309 init_waitqueue_head(&buffer->irq_work.waiters);
1310
1311 /* need at least two pages */
1312 if (nr_pages < 2)
1313 nr_pages = 2;
1314
1315 buffer->cpus = nr_cpu_ids;
1316
1317 bsize = sizeof(void *) * nr_cpu_ids;
1318 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1319 GFP_KERNEL);
1320 if (!buffer->buffers)
1321 goto fail_free_cpumask;
1322
1323 cpu = raw_smp_processor_id();
1324 cpumask_set_cpu(cpu, buffer->cpumask);
1325 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1326 if (!buffer->buffers[cpu])
1327 goto fail_free_buffers;
1328
1329 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1330 if (ret < 0)
1331 goto fail_free_buffers;
1332
1333 mutex_init(&buffer->mutex);
1334
1335 return buffer;
1336
1337 fail_free_buffers:
1338 for_each_buffer_cpu(buffer, cpu) {
1339 if (buffer->buffers[cpu])
1340 rb_free_cpu_buffer(buffer->buffers[cpu]);
1341 }
1342 kfree(buffer->buffers);
1343
1344 fail_free_cpumask:
1345 free_cpumask_var(buffer->cpumask);
1346
1347 fail_free_buffer:
1348 kfree(buffer);
1349 return NULL;
1350}
1351EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1352
1353/**
1354 * ring_buffer_free - free a ring buffer.
1355 * @buffer: the buffer to free.
1356 */
1357void
1358ring_buffer_free(struct ring_buffer *buffer)
1359{
1360 int cpu;
1361
1362 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1363
1364 for_each_buffer_cpu(buffer, cpu)
1365 rb_free_cpu_buffer(buffer->buffers[cpu]);
1366
1367 kfree(buffer->buffers);
1368 free_cpumask_var(buffer->cpumask);
1369
1370 kfree(buffer);
1371}
1372EXPORT_SYMBOL_GPL(ring_buffer_free);
1373
1374void ring_buffer_set_clock(struct ring_buffer *buffer,
1375 u64 (*clock)(void))
1376{
1377 buffer->clock = clock;
1378}
1379
1380static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1381
1382static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1383{
1384 return local_read(&bpage->entries) & RB_WRITE_MASK;
1385}
1386
1387static inline unsigned long rb_page_write(struct buffer_page *bpage)
1388{
1389 return local_read(&bpage->write) & RB_WRITE_MASK;
1390}
1391
1392static int
1393rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1394{
1395 struct list_head *tail_page, *to_remove, *next_page;
1396 struct buffer_page *to_remove_page, *tmp_iter_page;
1397 struct buffer_page *last_page, *first_page;
1398 unsigned long nr_removed;
1399 unsigned long head_bit;
1400 int page_entries;
1401
1402 head_bit = 0;
1403
1404 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1405 atomic_inc(&cpu_buffer->record_disabled);
1406 /*
1407 * We don't race with the readers since we have acquired the reader
1408 * lock. We also don't race with writers after disabling recording.
1409 * This makes it easy to figure out the first and the last page to be
1410 * removed from the list. We unlink all the pages in between including
1411 * the first and last pages. This is done in a busy loop so that we
1412 * lose the least number of traces.
1413 * The pages are freed after we restart recording and unlock readers.
1414 */
1415 tail_page = &cpu_buffer->tail_page->list;
1416
1417 /*
1418 * tail page might be on reader page, we remove the next page
1419 * from the ring buffer
1420 */
1421 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1422 tail_page = rb_list_head(tail_page->next);
1423 to_remove = tail_page;
1424
1425 /* start of pages to remove */
1426 first_page = list_entry(rb_list_head(to_remove->next),
1427 struct buffer_page, list);
1428
1429 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1430 to_remove = rb_list_head(to_remove)->next;
1431 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1432 }
1433
1434 next_page = rb_list_head(to_remove)->next;
1435
1436 /*
1437 * Now we remove all pages between tail_page and next_page.
1438 * Make sure that we have head_bit value preserved for the
1439 * next page
1440 */
1441 tail_page->next = (struct list_head *)((unsigned long)next_page |
1442 head_bit);
1443 next_page = rb_list_head(next_page);
1444 next_page->prev = tail_page;
1445
1446 /* make sure pages points to a valid page in the ring buffer */
1447 cpu_buffer->pages = next_page;
1448
1449 /* update head page */
1450 if (head_bit)
1451 cpu_buffer->head_page = list_entry(next_page,
1452 struct buffer_page, list);
1453
1454 /*
1455 * change read pointer to make sure any read iterators reset
1456 * themselves
1457 */
1458 cpu_buffer->read = 0;
1459
1460 /* pages are removed, resume tracing and then free the pages */
1461 atomic_dec(&cpu_buffer->record_disabled);
1462 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1463
1464 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1465
1466 /* last buffer page to remove */
1467 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1468 list);
1469 tmp_iter_page = first_page;
1470
1471 do {
1472 to_remove_page = tmp_iter_page;
1473 rb_inc_page(cpu_buffer, &tmp_iter_page);
1474
1475 /* update the counters */
1476 page_entries = rb_page_entries(to_remove_page);
1477 if (page_entries) {
1478 /*
1479 * If something was added to this page, it was full
1480 * since it is not the tail page. So we deduct the
1481 * bytes consumed in ring buffer from here.
1482 * Increment overrun to account for the lost events.
1483 */
1484 local_add(page_entries, &cpu_buffer->overrun);
1485 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1486 }
1487
1488 /*
1489 * We have already removed references to this list item, just
1490 * free up the buffer_page and its page
1491 */
1492 free_buffer_page(to_remove_page);
1493 nr_removed--;
1494
1495 } while (to_remove_page != last_page);
1496
1497 RB_WARN_ON(cpu_buffer, nr_removed);
1498
1499 return nr_removed == 0;
1500}
1501
1502static int
1503rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1504{
1505 struct list_head *pages = &cpu_buffer->new_pages;
1506 int retries, success;
1507
1508 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1509 /*
1510 * We are holding the reader lock, so the reader page won't be swapped
1511 * in the ring buffer. Now we are racing with the writer trying to
1512 * move head page and the tail page.
1513 * We are going to adapt the reader page update process where:
1514 * 1. We first splice the start and end of list of new pages between
1515 * the head page and its previous page.
1516 * 2. We cmpxchg the prev_page->next to point from head page to the
1517 * start of new pages list.
1518 * 3. Finally, we update the head->prev to the end of new list.
1519 *
1520 * We will try this process 10 times, to make sure that we don't keep
1521 * spinning.
1522 */
1523 retries = 10;
1524 success = 0;
1525 while (retries--) {
1526 struct list_head *head_page, *prev_page, *r;
1527 struct list_head *last_page, *first_page;
1528 struct list_head *head_page_with_bit;
1529
1530 head_page = &rb_set_head_page(cpu_buffer)->list;
1531 if (!head_page)
1532 break;
1533 prev_page = head_page->prev;
1534
1535 first_page = pages->next;
1536 last_page = pages->prev;
1537
1538 head_page_with_bit = (struct list_head *)
1539 ((unsigned long)head_page | RB_PAGE_HEAD);
1540
1541 last_page->next = head_page_with_bit;
1542 first_page->prev = prev_page;
1543
1544 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1545
1546 if (r == head_page_with_bit) {
1547 /*
1548 * yay, we replaced the page pointer to our new list,
1549 * now, we just have to update to head page's prev
1550 * pointer to point to end of list
1551 */
1552 head_page->prev = last_page;
1553 success = 1;
1554 break;
1555 }
1556 }
1557
1558 if (success)
1559 INIT_LIST_HEAD(pages);
1560 /*
1561 * If we weren't successful in adding in new pages, warn and stop
1562 * tracing
1563 */
1564 RB_WARN_ON(cpu_buffer, !success);
1565 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1566
1567 /* free pages if they weren't inserted */
1568 if (!success) {
1569 struct buffer_page *bpage, *tmp;
1570 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1571 list) {
1572 list_del_init(&bpage->list);
1573 free_buffer_page(bpage);
1574 }
1575 }
1576 return success;
1577}
1578
1579static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1580{
1581 int success;
1582
1583 if (cpu_buffer->nr_pages_to_update > 0)
1584 success = rb_insert_pages(cpu_buffer);
1585 else
1586 success = rb_remove_pages(cpu_buffer,
1587 -cpu_buffer->nr_pages_to_update);
1588
1589 if (success)
1590 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1591}
1592
1593static void update_pages_handler(struct work_struct *work)
1594{
1595 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1596 struct ring_buffer_per_cpu, update_pages_work);
1597 rb_update_pages(cpu_buffer);
1598 complete(&cpu_buffer->update_done);
1599}
1600
1601/**
1602 * ring_buffer_resize - resize the ring buffer
1603 * @buffer: the buffer to resize.
1604 * @size: the new size.
1605 * @cpu_id: the cpu buffer to resize
1606 *
1607 * Minimum size is 2 * BUF_PAGE_SIZE.
1608 *
1609 * Returns 0 on success and < 0 on failure.
1610 */
1611int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1612 int cpu_id)
1613{
1614 struct ring_buffer_per_cpu *cpu_buffer;
1615 unsigned long nr_pages;
1616 int cpu, err = 0;
1617
1618 /*
1619 * Always succeed at resizing a non-existent buffer:
1620 */
1621 if (!buffer)
1622 return size;
1623
1624 /* Make sure the requested buffer exists */
1625 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1626 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1627 return size;
1628
1629 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1630
1631 /* we need a minimum of two pages */
1632 if (nr_pages < 2)
1633 nr_pages = 2;
1634
1635 size = nr_pages * BUF_PAGE_SIZE;
1636
1637 /*
1638 * Don't succeed if resizing is disabled, as a reader might be
1639 * manipulating the ring buffer and is expecting a sane state while
1640 * this is true.
1641 */
1642 if (atomic_read(&buffer->resize_disabled))
1643 return -EBUSY;
1644
1645 /* prevent another thread from changing buffer sizes */
1646 mutex_lock(&buffer->mutex);
1647
1648 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1649 /* calculate the pages to update */
1650 for_each_buffer_cpu(buffer, cpu) {
1651 cpu_buffer = buffer->buffers[cpu];
1652
1653 cpu_buffer->nr_pages_to_update = nr_pages -
1654 cpu_buffer->nr_pages;
1655 /*
1656 * nothing more to do for removing pages or no update
1657 */
1658 if (cpu_buffer->nr_pages_to_update <= 0)
1659 continue;
1660 /*
1661 * to add pages, make sure all new pages can be
1662 * allocated without receiving ENOMEM
1663 */
1664 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1665 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1666 &cpu_buffer->new_pages, cpu)) {
1667 /* not enough memory for new pages */
1668 err = -ENOMEM;
1669 goto out_err;
1670 }
1671 }
1672
1673 get_online_cpus();
1674 /*
1675 * Fire off all the required work handlers
1676 * We can't schedule on offline CPUs, but it's not necessary
1677 * since we can change their buffer sizes without any race.
1678 */
1679 for_each_buffer_cpu(buffer, cpu) {
1680 cpu_buffer = buffer->buffers[cpu];
1681 if (!cpu_buffer->nr_pages_to_update)
1682 continue;
1683
1684 /* Can't run something on an offline CPU. */
1685 if (!cpu_online(cpu)) {
1686 rb_update_pages(cpu_buffer);
1687 cpu_buffer->nr_pages_to_update = 0;
1688 } else {
1689 schedule_work_on(cpu,
1690 &cpu_buffer->update_pages_work);
1691 }
1692 }
1693
1694 /* wait for all the updates to complete */
1695 for_each_buffer_cpu(buffer, cpu) {
1696 cpu_buffer = buffer->buffers[cpu];
1697 if (!cpu_buffer->nr_pages_to_update)
1698 continue;
1699
1700 if (cpu_online(cpu))
1701 wait_for_completion(&cpu_buffer->update_done);
1702 cpu_buffer->nr_pages_to_update = 0;
1703 }
1704
1705 put_online_cpus();
1706 } else {
1707 /* Make sure this CPU has been intitialized */
1708 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1709 goto out;
1710
1711 cpu_buffer = buffer->buffers[cpu_id];
1712
1713 if (nr_pages == cpu_buffer->nr_pages)
1714 goto out;
1715
1716 cpu_buffer->nr_pages_to_update = nr_pages -
1717 cpu_buffer->nr_pages;
1718
1719 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1720 if (cpu_buffer->nr_pages_to_update > 0 &&
1721 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1722 &cpu_buffer->new_pages, cpu_id)) {
1723 err = -ENOMEM;
1724 goto out_err;
1725 }
1726
1727 get_online_cpus();
1728
1729 /* Can't run something on an offline CPU. */
1730 if (!cpu_online(cpu_id))
1731 rb_update_pages(cpu_buffer);
1732 else {
1733 schedule_work_on(cpu_id,
1734 &cpu_buffer->update_pages_work);
1735 wait_for_completion(&cpu_buffer->update_done);
1736 }
1737
1738 cpu_buffer->nr_pages_to_update = 0;
1739 put_online_cpus();
1740 }
1741
1742 out:
1743 /*
1744 * The ring buffer resize can happen with the ring buffer
1745 * enabled, so that the update disturbs the tracing as little
1746 * as possible. But if the buffer is disabled, we do not need
1747 * to worry about that, and we can take the time to verify
1748 * that the buffer is not corrupt.
1749 */
1750 if (atomic_read(&buffer->record_disabled)) {
1751 atomic_inc(&buffer->record_disabled);
1752 /*
1753 * Even though the buffer was disabled, we must make sure
1754 * that it is truly disabled before calling rb_check_pages.
1755 * There could have been a race between checking
1756 * record_disable and incrementing it.
1757 */
1758 synchronize_sched();
1759 for_each_buffer_cpu(buffer, cpu) {
1760 cpu_buffer = buffer->buffers[cpu];
1761 rb_check_pages(cpu_buffer);
1762 }
1763 atomic_dec(&buffer->record_disabled);
1764 }
1765
1766 mutex_unlock(&buffer->mutex);
1767 return size;
1768
1769 out_err:
1770 for_each_buffer_cpu(buffer, cpu) {
1771 struct buffer_page *bpage, *tmp;
1772
1773 cpu_buffer = buffer->buffers[cpu];
1774 cpu_buffer->nr_pages_to_update = 0;
1775
1776 if (list_empty(&cpu_buffer->new_pages))
1777 continue;
1778
1779 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1780 list) {
1781 list_del_init(&bpage->list);
1782 free_buffer_page(bpage);
1783 }
1784 }
1785 mutex_unlock(&buffer->mutex);
1786 return err;
1787}
1788EXPORT_SYMBOL_GPL(ring_buffer_resize);
1789
1790void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1791{
1792 mutex_lock(&buffer->mutex);
1793 if (val)
1794 buffer->flags |= RB_FL_OVERWRITE;
1795 else
1796 buffer->flags &= ~RB_FL_OVERWRITE;
1797 mutex_unlock(&buffer->mutex);
1798}
1799EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1800
1801static __always_inline void *
1802__rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1803{
1804 return bpage->data + index;
1805}
1806
1807static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1808{
1809 return bpage->page->data + index;
1810}
1811
1812static __always_inline struct ring_buffer_event *
1813rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1814{
1815 return __rb_page_index(cpu_buffer->reader_page,
1816 cpu_buffer->reader_page->read);
1817}
1818
1819static __always_inline struct ring_buffer_event *
1820rb_iter_head_event(struct ring_buffer_iter *iter)
1821{
1822 return __rb_page_index(iter->head_page, iter->head);
1823}
1824
1825static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
1826{
1827 return local_read(&bpage->page->commit);
1828}
1829
1830/* Size is determined by what has been committed */
1831static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
1832{
1833 return rb_page_commit(bpage);
1834}
1835
1836static __always_inline unsigned
1837rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1838{
1839 return rb_page_commit(cpu_buffer->commit_page);
1840}
1841
1842static __always_inline unsigned
1843rb_event_index(struct ring_buffer_event *event)
1844{
1845 unsigned long addr = (unsigned long)event;
1846
1847 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1848}
1849
1850static void rb_inc_iter(struct ring_buffer_iter *iter)
1851{
1852 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1853
1854 /*
1855 * The iterator could be on the reader page (it starts there).
1856 * But the head could have moved, since the reader was
1857 * found. Check for this case and assign the iterator
1858 * to the head page instead of next.
1859 */
1860 if (iter->head_page == cpu_buffer->reader_page)
1861 iter->head_page = rb_set_head_page(cpu_buffer);
1862 else
1863 rb_inc_page(cpu_buffer, &iter->head_page);
1864
1865 iter->read_stamp = iter->head_page->page->time_stamp;
1866 iter->head = 0;
1867}
1868
1869/*
1870 * rb_handle_head_page - writer hit the head page
1871 *
1872 * Returns: +1 to retry page
1873 * 0 to continue
1874 * -1 on error
1875 */
1876static int
1877rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1878 struct buffer_page *tail_page,
1879 struct buffer_page *next_page)
1880{
1881 struct buffer_page *new_head;
1882 int entries;
1883 int type;
1884 int ret;
1885
1886 entries = rb_page_entries(next_page);
1887
1888 /*
1889 * The hard part is here. We need to move the head
1890 * forward, and protect against both readers on
1891 * other CPUs and writers coming in via interrupts.
1892 */
1893 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1894 RB_PAGE_HEAD);
1895
1896 /*
1897 * type can be one of four:
1898 * NORMAL - an interrupt already moved it for us
1899 * HEAD - we are the first to get here.
1900 * UPDATE - we are the interrupt interrupting
1901 * a current move.
1902 * MOVED - a reader on another CPU moved the next
1903 * pointer to its reader page. Give up
1904 * and try again.
1905 */
1906
1907 switch (type) {
1908 case RB_PAGE_HEAD:
1909 /*
1910 * We changed the head to UPDATE, thus
1911 * it is our responsibility to update
1912 * the counters.
1913 */
1914 local_add(entries, &cpu_buffer->overrun);
1915 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1916
1917 /*
1918 * The entries will be zeroed out when we move the
1919 * tail page.
1920 */
1921
1922 /* still more to do */
1923 break;
1924
1925 case RB_PAGE_UPDATE:
1926 /*
1927 * This is an interrupt that interrupt the
1928 * previous update. Still more to do.
1929 */
1930 break;
1931 case RB_PAGE_NORMAL:
1932 /*
1933 * An interrupt came in before the update
1934 * and processed this for us.
1935 * Nothing left to do.
1936 */
1937 return 1;
1938 case RB_PAGE_MOVED:
1939 /*
1940 * The reader is on another CPU and just did
1941 * a swap with our next_page.
1942 * Try again.
1943 */
1944 return 1;
1945 default:
1946 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1947 return -1;
1948 }
1949
1950 /*
1951 * Now that we are here, the old head pointer is
1952 * set to UPDATE. This will keep the reader from
1953 * swapping the head page with the reader page.
1954 * The reader (on another CPU) will spin till
1955 * we are finished.
1956 *
1957 * We just need to protect against interrupts
1958 * doing the job. We will set the next pointer
1959 * to HEAD. After that, we set the old pointer
1960 * to NORMAL, but only if it was HEAD before.
1961 * otherwise we are an interrupt, and only
1962 * want the outer most commit to reset it.
1963 */
1964 new_head = next_page;
1965 rb_inc_page(cpu_buffer, &new_head);
1966
1967 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1968 RB_PAGE_NORMAL);
1969
1970 /*
1971 * Valid returns are:
1972 * HEAD - an interrupt came in and already set it.
1973 * NORMAL - One of two things:
1974 * 1) We really set it.
1975 * 2) A bunch of interrupts came in and moved
1976 * the page forward again.
1977 */
1978 switch (ret) {
1979 case RB_PAGE_HEAD:
1980 case RB_PAGE_NORMAL:
1981 /* OK */
1982 break;
1983 default:
1984 RB_WARN_ON(cpu_buffer, 1);
1985 return -1;
1986 }
1987
1988 /*
1989 * It is possible that an interrupt came in,
1990 * set the head up, then more interrupts came in
1991 * and moved it again. When we get back here,
1992 * the page would have been set to NORMAL but we
1993 * just set it back to HEAD.
1994 *
1995 * How do you detect this? Well, if that happened
1996 * the tail page would have moved.
1997 */
1998 if (ret == RB_PAGE_NORMAL) {
1999 struct buffer_page *buffer_tail_page;
2000
2001 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2002 /*
2003 * If the tail had moved passed next, then we need
2004 * to reset the pointer.
2005 */
2006 if (buffer_tail_page != tail_page &&
2007 buffer_tail_page != next_page)
2008 rb_head_page_set_normal(cpu_buffer, new_head,
2009 next_page,
2010 RB_PAGE_HEAD);
2011 }
2012
2013 /*
2014 * If this was the outer most commit (the one that
2015 * changed the original pointer from HEAD to UPDATE),
2016 * then it is up to us to reset it to NORMAL.
2017 */
2018 if (type == RB_PAGE_HEAD) {
2019 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2020 tail_page,
2021 RB_PAGE_UPDATE);
2022 if (RB_WARN_ON(cpu_buffer,
2023 ret != RB_PAGE_UPDATE))
2024 return -1;
2025 }
2026
2027 return 0;
2028}
2029
2030static inline void
2031rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2032 unsigned long tail, struct rb_event_info *info)
2033{
2034 struct buffer_page *tail_page = info->tail_page;
2035 struct ring_buffer_event *event;
2036 unsigned long length = info->length;
2037
2038 /*
2039 * Only the event that crossed the page boundary
2040 * must fill the old tail_page with padding.
2041 */
2042 if (tail >= BUF_PAGE_SIZE) {
2043 /*
2044 * If the page was filled, then we still need
2045 * to update the real_end. Reset it to zero
2046 * and the reader will ignore it.
2047 */
2048 if (tail == BUF_PAGE_SIZE)
2049 tail_page->real_end = 0;
2050
2051 local_sub(length, &tail_page->write);
2052 return;
2053 }
2054
2055 event = __rb_page_index(tail_page, tail);
2056 kmemcheck_annotate_bitfield(event, bitfield);
2057
2058 /* account for padding bytes */
2059 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2060
2061 /*
2062 * Save the original length to the meta data.
2063 * This will be used by the reader to add lost event
2064 * counter.
2065 */
2066 tail_page->real_end = tail;
2067
2068 /*
2069 * If this event is bigger than the minimum size, then
2070 * we need to be careful that we don't subtract the
2071 * write counter enough to allow another writer to slip
2072 * in on this page.
2073 * We put in a discarded commit instead, to make sure
2074 * that this space is not used again.
2075 *
2076 * If we are less than the minimum size, we don't need to
2077 * worry about it.
2078 */
2079 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2080 /* No room for any events */
2081
2082 /* Mark the rest of the page with padding */
2083 rb_event_set_padding(event);
2084
2085 /* Set the write back to the previous setting */
2086 local_sub(length, &tail_page->write);
2087 return;
2088 }
2089
2090 /* Put in a discarded event */
2091 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2092 event->type_len = RINGBUF_TYPE_PADDING;
2093 /* time delta must be non zero */
2094 event->time_delta = 1;
2095
2096 /* Set write to end of buffer */
2097 length = (tail + length) - BUF_PAGE_SIZE;
2098 local_sub(length, &tail_page->write);
2099}
2100
2101static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2102
2103/*
2104 * This is the slow path, force gcc not to inline it.
2105 */
2106static noinline struct ring_buffer_event *
2107rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2108 unsigned long tail, struct rb_event_info *info)
2109{
2110 struct buffer_page *tail_page = info->tail_page;
2111 struct buffer_page *commit_page = cpu_buffer->commit_page;
2112 struct ring_buffer *buffer = cpu_buffer->buffer;
2113 struct buffer_page *next_page;
2114 int ret;
2115
2116 next_page = tail_page;
2117
2118 rb_inc_page(cpu_buffer, &next_page);
2119
2120 /*
2121 * If for some reason, we had an interrupt storm that made
2122 * it all the way around the buffer, bail, and warn
2123 * about it.
2124 */
2125 if (unlikely(next_page == commit_page)) {
2126 local_inc(&cpu_buffer->commit_overrun);
2127 goto out_reset;
2128 }
2129
2130 /*
2131 * This is where the fun begins!
2132 *
2133 * We are fighting against races between a reader that
2134 * could be on another CPU trying to swap its reader
2135 * page with the buffer head.
2136 *
2137 * We are also fighting against interrupts coming in and
2138 * moving the head or tail on us as well.
2139 *
2140 * If the next page is the head page then we have filled
2141 * the buffer, unless the commit page is still on the
2142 * reader page.
2143 */
2144 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2145
2146 /*
2147 * If the commit is not on the reader page, then
2148 * move the header page.
2149 */
2150 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2151 /*
2152 * If we are not in overwrite mode,
2153 * this is easy, just stop here.
2154 */
2155 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2156 local_inc(&cpu_buffer->dropped_events);
2157 goto out_reset;
2158 }
2159
2160 ret = rb_handle_head_page(cpu_buffer,
2161 tail_page,
2162 next_page);
2163 if (ret < 0)
2164 goto out_reset;
2165 if (ret)
2166 goto out_again;
2167 } else {
2168 /*
2169 * We need to be careful here too. The
2170 * commit page could still be on the reader
2171 * page. We could have a small buffer, and
2172 * have filled up the buffer with events
2173 * from interrupts and such, and wrapped.
2174 *
2175 * Note, if the tail page is also the on the
2176 * reader_page, we let it move out.
2177 */
2178 if (unlikely((cpu_buffer->commit_page !=
2179 cpu_buffer->tail_page) &&
2180 (cpu_buffer->commit_page ==
2181 cpu_buffer->reader_page))) {
2182 local_inc(&cpu_buffer->commit_overrun);
2183 goto out_reset;
2184 }
2185 }
2186 }
2187
2188 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2189
2190 out_again:
2191
2192 rb_reset_tail(cpu_buffer, tail, info);
2193
2194 /* Commit what we have for now. */
2195 rb_end_commit(cpu_buffer);
2196 /* rb_end_commit() decs committing */
2197 local_inc(&cpu_buffer->committing);
2198
2199 /* fail and let the caller try again */
2200 return ERR_PTR(-EAGAIN);
2201
2202 out_reset:
2203 /* reset write */
2204 rb_reset_tail(cpu_buffer, tail, info);
2205
2206 return NULL;
2207}
2208
2209/* Slow path, do not inline */
2210static noinline struct ring_buffer_event *
2211rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
2212{
2213 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2214
2215 /* Not the first event on the page? */
2216 if (rb_event_index(event)) {
2217 event->time_delta = delta & TS_MASK;
2218 event->array[0] = delta >> TS_SHIFT;
2219 } else {
2220 /* nope, just zero it */
2221 event->time_delta = 0;
2222 event->array[0] = 0;
2223 }
2224
2225 return skip_time_extend(event);
2226}
2227
2228static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2229 struct ring_buffer_event *event);
2230
2231/**
2232 * rb_update_event - update event type and data
2233 * @event: the event to update
2234 * @type: the type of event
2235 * @length: the size of the event field in the ring buffer
2236 *
2237 * Update the type and data fields of the event. The length
2238 * is the actual size that is written to the ring buffer,
2239 * and with this, we can determine what to place into the
2240 * data field.
2241 */
2242static void
2243rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2244 struct ring_buffer_event *event,
2245 struct rb_event_info *info)
2246{
2247 unsigned length = info->length;
2248 u64 delta = info->delta;
2249
2250 /* Only a commit updates the timestamp */
2251 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2252 delta = 0;
2253
2254 /*
2255 * If we need to add a timestamp, then we
2256 * add it to the start of the resevered space.
2257 */
2258 if (unlikely(info->add_timestamp)) {
2259 event = rb_add_time_stamp(event, delta);
2260 length -= RB_LEN_TIME_EXTEND;
2261 delta = 0;
2262 }
2263
2264 event->time_delta = delta;
2265 length -= RB_EVNT_HDR_SIZE;
2266 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2267 event->type_len = 0;
2268 event->array[0] = length;
2269 } else
2270 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2271}
2272
2273static unsigned rb_calculate_event_length(unsigned length)
2274{
2275 struct ring_buffer_event event; /* Used only for sizeof array */
2276
2277 /* zero length can cause confusions */
2278 if (!length)
2279 length++;
2280
2281 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2282 length += sizeof(event.array[0]);
2283
2284 length += RB_EVNT_HDR_SIZE;
2285 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2286
2287 /*
2288 * In case the time delta is larger than the 27 bits for it
2289 * in the header, we need to add a timestamp. If another
2290 * event comes in when trying to discard this one to increase
2291 * the length, then the timestamp will be added in the allocated
2292 * space of this event. If length is bigger than the size needed
2293 * for the TIME_EXTEND, then padding has to be used. The events
2294 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2295 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2296 * As length is a multiple of 4, we only need to worry if it
2297 * is 12 (RB_LEN_TIME_EXTEND + 4).
2298 */
2299 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2300 length += RB_ALIGNMENT;
2301
2302 return length;
2303}
2304
2305#ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2306static inline bool sched_clock_stable(void)
2307{
2308 return true;
2309}
2310#endif
2311
2312static inline int
2313rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2314 struct ring_buffer_event *event)
2315{
2316 unsigned long new_index, old_index;
2317 struct buffer_page *bpage;
2318 unsigned long index;
2319 unsigned long addr;
2320
2321 new_index = rb_event_index(event);
2322 old_index = new_index + rb_event_ts_length(event);
2323 addr = (unsigned long)event;
2324 addr &= PAGE_MASK;
2325
2326 bpage = READ_ONCE(cpu_buffer->tail_page);
2327
2328 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2329 unsigned long write_mask =
2330 local_read(&bpage->write) & ~RB_WRITE_MASK;
2331 unsigned long event_length = rb_event_length(event);
2332 /*
2333 * This is on the tail page. It is possible that
2334 * a write could come in and move the tail page
2335 * and write to the next page. That is fine
2336 * because we just shorten what is on this page.
2337 */
2338 old_index += write_mask;
2339 new_index += write_mask;
2340 index = local_cmpxchg(&bpage->write, old_index, new_index);
2341 if (index == old_index) {
2342 /* update counters */
2343 local_sub(event_length, &cpu_buffer->entries_bytes);
2344 return 1;
2345 }
2346 }
2347
2348 /* could not discard */
2349 return 0;
2350}
2351
2352static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2353{
2354 local_inc(&cpu_buffer->committing);
2355 local_inc(&cpu_buffer->commits);
2356}
2357
2358static __always_inline void
2359rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2360{
2361 unsigned long max_count;
2362
2363 /*
2364 * We only race with interrupts and NMIs on this CPU.
2365 * If we own the commit event, then we can commit
2366 * all others that interrupted us, since the interruptions
2367 * are in stack format (they finish before they come
2368 * back to us). This allows us to do a simple loop to
2369 * assign the commit to the tail.
2370 */
2371 again:
2372 max_count = cpu_buffer->nr_pages * 100;
2373
2374 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2375 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2376 return;
2377 if (RB_WARN_ON(cpu_buffer,
2378 rb_is_reader_page(cpu_buffer->tail_page)))
2379 return;
2380 local_set(&cpu_buffer->commit_page->page->commit,
2381 rb_page_write(cpu_buffer->commit_page));
2382 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2383 /* Only update the write stamp if the page has an event */
2384 if (rb_page_write(cpu_buffer->commit_page))
2385 cpu_buffer->write_stamp =
2386 cpu_buffer->commit_page->page->time_stamp;
2387 /* add barrier to keep gcc from optimizing too much */
2388 barrier();
2389 }
2390 while (rb_commit_index(cpu_buffer) !=
2391 rb_page_write(cpu_buffer->commit_page)) {
2392
2393 local_set(&cpu_buffer->commit_page->page->commit,
2394 rb_page_write(cpu_buffer->commit_page));
2395 RB_WARN_ON(cpu_buffer,
2396 local_read(&cpu_buffer->commit_page->page->commit) &
2397 ~RB_WRITE_MASK);
2398 barrier();
2399 }
2400
2401 /* again, keep gcc from optimizing */
2402 barrier();
2403
2404 /*
2405 * If an interrupt came in just after the first while loop
2406 * and pushed the tail page forward, we will be left with
2407 * a dangling commit that will never go forward.
2408 */
2409 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2410 goto again;
2411}
2412
2413static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2414{
2415 unsigned long commits;
2416
2417 if (RB_WARN_ON(cpu_buffer,
2418 !local_read(&cpu_buffer->committing)))
2419 return;
2420
2421 again:
2422 commits = local_read(&cpu_buffer->commits);
2423 /* synchronize with interrupts */
2424 barrier();
2425 if (local_read(&cpu_buffer->committing) == 1)
2426 rb_set_commit_to_write(cpu_buffer);
2427
2428 local_dec(&cpu_buffer->committing);
2429
2430 /* synchronize with interrupts */
2431 barrier();
2432
2433 /*
2434 * Need to account for interrupts coming in between the
2435 * updating of the commit page and the clearing of the
2436 * committing counter.
2437 */
2438 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2439 !local_read(&cpu_buffer->committing)) {
2440 local_inc(&cpu_buffer->committing);
2441 goto again;
2442 }
2443}
2444
2445static inline void rb_event_discard(struct ring_buffer_event *event)
2446{
2447 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2448 event = skip_time_extend(event);
2449
2450 /* array[0] holds the actual length for the discarded event */
2451 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2452 event->type_len = RINGBUF_TYPE_PADDING;
2453 /* time delta must be non zero */
2454 if (!event->time_delta)
2455 event->time_delta = 1;
2456}
2457
2458static __always_inline bool
2459rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2460 struct ring_buffer_event *event)
2461{
2462 unsigned long addr = (unsigned long)event;
2463 unsigned long index;
2464
2465 index = rb_event_index(event);
2466 addr &= PAGE_MASK;
2467
2468 return cpu_buffer->commit_page->page == (void *)addr &&
2469 rb_commit_index(cpu_buffer) == index;
2470}
2471
2472static __always_inline void
2473rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2474 struct ring_buffer_event *event)
2475{
2476 u64 delta;
2477
2478 /*
2479 * The event first in the commit queue updates the
2480 * time stamp.
2481 */
2482 if (rb_event_is_commit(cpu_buffer, event)) {
2483 /*
2484 * A commit event that is first on a page
2485 * updates the write timestamp with the page stamp
2486 */
2487 if (!rb_event_index(event))
2488 cpu_buffer->write_stamp =
2489 cpu_buffer->commit_page->page->time_stamp;
2490 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2491 delta = event->array[0];
2492 delta <<= TS_SHIFT;
2493 delta += event->time_delta;
2494 cpu_buffer->write_stamp += delta;
2495 } else
2496 cpu_buffer->write_stamp += event->time_delta;
2497 }
2498}
2499
2500static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2501 struct ring_buffer_event *event)
2502{
2503 local_inc(&cpu_buffer->entries);
2504 rb_update_write_stamp(cpu_buffer, event);
2505 rb_end_commit(cpu_buffer);
2506}
2507
2508static __always_inline void
2509rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2510{
2511 bool pagebusy;
2512
2513 if (buffer->irq_work.waiters_pending) {
2514 buffer->irq_work.waiters_pending = false;
2515 /* irq_work_queue() supplies it's own memory barriers */
2516 irq_work_queue(&buffer->irq_work.work);
2517 }
2518
2519 if (cpu_buffer->irq_work.waiters_pending) {
2520 cpu_buffer->irq_work.waiters_pending = false;
2521 /* irq_work_queue() supplies it's own memory barriers */
2522 irq_work_queue(&cpu_buffer->irq_work.work);
2523 }
2524
2525 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
2526
2527 if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2528 cpu_buffer->irq_work.wakeup_full = true;
2529 cpu_buffer->irq_work.full_waiters_pending = false;
2530 /* irq_work_queue() supplies it's own memory barriers */
2531 irq_work_queue(&cpu_buffer->irq_work.work);
2532 }
2533}
2534
2535/*
2536 * The lock and unlock are done within a preempt disable section.
2537 * The current_context per_cpu variable can only be modified
2538 * by the current task between lock and unlock. But it can
2539 * be modified more than once via an interrupt. To pass this
2540 * information from the lock to the unlock without having to
2541 * access the 'in_interrupt()' functions again (which do show
2542 * a bit of overhead in something as critical as function tracing,
2543 * we use a bitmask trick.
2544 *
2545 * bit 0 = NMI context
2546 * bit 1 = IRQ context
2547 * bit 2 = SoftIRQ context
2548 * bit 3 = normal context.
2549 *
2550 * This works because this is the order of contexts that can
2551 * preempt other contexts. A SoftIRQ never preempts an IRQ
2552 * context.
2553 *
2554 * When the context is determined, the corresponding bit is
2555 * checked and set (if it was set, then a recursion of that context
2556 * happened).
2557 *
2558 * On unlock, we need to clear this bit. To do so, just subtract
2559 * 1 from the current_context and AND it to itself.
2560 *
2561 * (binary)
2562 * 101 - 1 = 100
2563 * 101 & 100 = 100 (clearing bit zero)
2564 *
2565 * 1010 - 1 = 1001
2566 * 1010 & 1001 = 1000 (clearing bit 1)
2567 *
2568 * The least significant bit can be cleared this way, and it
2569 * just so happens that it is the same bit corresponding to
2570 * the current context.
2571 */
2572
2573static __always_inline int
2574trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2575{
2576 unsigned int val = cpu_buffer->current_context;
2577 int bit;
2578
2579 if (in_interrupt()) {
2580 if (in_nmi())
2581 bit = RB_CTX_NMI;
2582 else if (in_irq())
2583 bit = RB_CTX_IRQ;
2584 else
2585 bit = RB_CTX_SOFTIRQ;
2586 } else
2587 bit = RB_CTX_NORMAL;
2588
2589 if (unlikely(val & (1 << bit)))
2590 return 1;
2591
2592 val |= (1 << bit);
2593 cpu_buffer->current_context = val;
2594
2595 return 0;
2596}
2597
2598static __always_inline void
2599trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2600{
2601 cpu_buffer->current_context &= cpu_buffer->current_context - 1;
2602}
2603
2604/**
2605 * ring_buffer_unlock_commit - commit a reserved
2606 * @buffer: The buffer to commit to
2607 * @event: The event pointer to commit.
2608 *
2609 * This commits the data to the ring buffer, and releases any locks held.
2610 *
2611 * Must be paired with ring_buffer_lock_reserve.
2612 */
2613int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2614 struct ring_buffer_event *event)
2615{
2616 struct ring_buffer_per_cpu *cpu_buffer;
2617 int cpu = raw_smp_processor_id();
2618
2619 cpu_buffer = buffer->buffers[cpu];
2620
2621 rb_commit(cpu_buffer, event);
2622
2623 rb_wakeups(buffer, cpu_buffer);
2624
2625 trace_recursive_unlock(cpu_buffer);
2626
2627 preempt_enable_notrace();
2628
2629 return 0;
2630}
2631EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2632
2633static noinline void
2634rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2635 struct rb_event_info *info)
2636{
2637 WARN_ONCE(info->delta > (1ULL << 59),
2638 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2639 (unsigned long long)info->delta,
2640 (unsigned long long)info->ts,
2641 (unsigned long long)cpu_buffer->write_stamp,
2642 sched_clock_stable() ? "" :
2643 "If you just came from a suspend/resume,\n"
2644 "please switch to the trace global clock:\n"
2645 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2646 info->add_timestamp = 1;
2647}
2648
2649static struct ring_buffer_event *
2650__rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2651 struct rb_event_info *info)
2652{
2653 struct ring_buffer_event *event;
2654 struct buffer_page *tail_page;
2655 unsigned long tail, write;
2656
2657 /*
2658 * If the time delta since the last event is too big to
2659 * hold in the time field of the event, then we append a
2660 * TIME EXTEND event ahead of the data event.
2661 */
2662 if (unlikely(info->add_timestamp))
2663 info->length += RB_LEN_TIME_EXTEND;
2664
2665 /* Don't let the compiler play games with cpu_buffer->tail_page */
2666 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2667 write = local_add_return(info->length, &tail_page->write);
2668
2669 /* set write to only the index of the write */
2670 write &= RB_WRITE_MASK;
2671 tail = write - info->length;
2672
2673 /*
2674 * If this is the first commit on the page, then it has the same
2675 * timestamp as the page itself.
2676 */
2677 if (!tail)
2678 info->delta = 0;
2679
2680 /* See if we shot pass the end of this buffer page */
2681 if (unlikely(write > BUF_PAGE_SIZE))
2682 return rb_move_tail(cpu_buffer, tail, info);
2683
2684 /* We reserved something on the buffer */
2685
2686 event = __rb_page_index(tail_page, tail);
2687 kmemcheck_annotate_bitfield(event, bitfield);
2688 rb_update_event(cpu_buffer, event, info);
2689
2690 local_inc(&tail_page->entries);
2691
2692 /*
2693 * If this is the first commit on the page, then update
2694 * its timestamp.
2695 */
2696 if (!tail)
2697 tail_page->page->time_stamp = info->ts;
2698
2699 /* account for these added bytes */
2700 local_add(info->length, &cpu_buffer->entries_bytes);
2701
2702 return event;
2703}
2704
2705static __always_inline struct ring_buffer_event *
2706rb_reserve_next_event(struct ring_buffer *buffer,
2707 struct ring_buffer_per_cpu *cpu_buffer,
2708 unsigned long length)
2709{
2710 struct ring_buffer_event *event;
2711 struct rb_event_info info;
2712 int nr_loops = 0;
2713 u64 diff;
2714
2715 rb_start_commit(cpu_buffer);
2716
2717#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2718 /*
2719 * Due to the ability to swap a cpu buffer from a buffer
2720 * it is possible it was swapped before we committed.
2721 * (committing stops a swap). We check for it here and
2722 * if it happened, we have to fail the write.
2723 */
2724 barrier();
2725 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2726 local_dec(&cpu_buffer->committing);
2727 local_dec(&cpu_buffer->commits);
2728 return NULL;
2729 }
2730#endif
2731
2732 info.length = rb_calculate_event_length(length);
2733 again:
2734 info.add_timestamp = 0;
2735 info.delta = 0;
2736
2737 /*
2738 * We allow for interrupts to reenter here and do a trace.
2739 * If one does, it will cause this original code to loop
2740 * back here. Even with heavy interrupts happening, this
2741 * should only happen a few times in a row. If this happens
2742 * 1000 times in a row, there must be either an interrupt
2743 * storm or we have something buggy.
2744 * Bail!
2745 */
2746 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2747 goto out_fail;
2748
2749 info.ts = rb_time_stamp(cpu_buffer->buffer);
2750 diff = info.ts - cpu_buffer->write_stamp;
2751
2752 /* make sure this diff is calculated here */
2753 barrier();
2754
2755 /* Did the write stamp get updated already? */
2756 if (likely(info.ts >= cpu_buffer->write_stamp)) {
2757 info.delta = diff;
2758 if (unlikely(test_time_stamp(info.delta)))
2759 rb_handle_timestamp(cpu_buffer, &info);
2760 }
2761
2762 event = __rb_reserve_next(cpu_buffer, &info);
2763
2764 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2765 if (info.add_timestamp)
2766 info.length -= RB_LEN_TIME_EXTEND;
2767 goto again;
2768 }
2769
2770 if (!event)
2771 goto out_fail;
2772
2773 return event;
2774
2775 out_fail:
2776 rb_end_commit(cpu_buffer);
2777 return NULL;
2778}
2779
2780/**
2781 * ring_buffer_lock_reserve - reserve a part of the buffer
2782 * @buffer: the ring buffer to reserve from
2783 * @length: the length of the data to reserve (excluding event header)
2784 *
2785 * Returns a reseverd event on the ring buffer to copy directly to.
2786 * The user of this interface will need to get the body to write into
2787 * and can use the ring_buffer_event_data() interface.
2788 *
2789 * The length is the length of the data needed, not the event length
2790 * which also includes the event header.
2791 *
2792 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2793 * If NULL is returned, then nothing has been allocated or locked.
2794 */
2795struct ring_buffer_event *
2796ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2797{
2798 struct ring_buffer_per_cpu *cpu_buffer;
2799 struct ring_buffer_event *event;
2800 int cpu;
2801
2802 /* If we are tracing schedule, we don't want to recurse */
2803 preempt_disable_notrace();
2804
2805 if (unlikely(atomic_read(&buffer->record_disabled)))
2806 goto out;
2807
2808 cpu = raw_smp_processor_id();
2809
2810 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2811 goto out;
2812
2813 cpu_buffer = buffer->buffers[cpu];
2814
2815 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2816 goto out;
2817
2818 if (unlikely(length > BUF_MAX_DATA_SIZE))
2819 goto out;
2820
2821 if (unlikely(trace_recursive_lock(cpu_buffer)))
2822 goto out;
2823
2824 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2825 if (!event)
2826 goto out_unlock;
2827
2828 return event;
2829
2830 out_unlock:
2831 trace_recursive_unlock(cpu_buffer);
2832 out:
2833 preempt_enable_notrace();
2834 return NULL;
2835}
2836EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2837
2838/*
2839 * Decrement the entries to the page that an event is on.
2840 * The event does not even need to exist, only the pointer
2841 * to the page it is on. This may only be called before the commit
2842 * takes place.
2843 */
2844static inline void
2845rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2846 struct ring_buffer_event *event)
2847{
2848 unsigned long addr = (unsigned long)event;
2849 struct buffer_page *bpage = cpu_buffer->commit_page;
2850 struct buffer_page *start;
2851
2852 addr &= PAGE_MASK;
2853
2854 /* Do the likely case first */
2855 if (likely(bpage->page == (void *)addr)) {
2856 local_dec(&bpage->entries);
2857 return;
2858 }
2859
2860 /*
2861 * Because the commit page may be on the reader page we
2862 * start with the next page and check the end loop there.
2863 */
2864 rb_inc_page(cpu_buffer, &bpage);
2865 start = bpage;
2866 do {
2867 if (bpage->page == (void *)addr) {
2868 local_dec(&bpage->entries);
2869 return;
2870 }
2871 rb_inc_page(cpu_buffer, &bpage);
2872 } while (bpage != start);
2873
2874 /* commit not part of this buffer?? */
2875 RB_WARN_ON(cpu_buffer, 1);
2876}
2877
2878/**
2879 * ring_buffer_commit_discard - discard an event that has not been committed
2880 * @buffer: the ring buffer
2881 * @event: non committed event to discard
2882 *
2883 * Sometimes an event that is in the ring buffer needs to be ignored.
2884 * This function lets the user discard an event in the ring buffer
2885 * and then that event will not be read later.
2886 *
2887 * This function only works if it is called before the the item has been
2888 * committed. It will try to free the event from the ring buffer
2889 * if another event has not been added behind it.
2890 *
2891 * If another event has been added behind it, it will set the event
2892 * up as discarded, and perform the commit.
2893 *
2894 * If this function is called, do not call ring_buffer_unlock_commit on
2895 * the event.
2896 */
2897void ring_buffer_discard_commit(struct ring_buffer *buffer,
2898 struct ring_buffer_event *event)
2899{
2900 struct ring_buffer_per_cpu *cpu_buffer;
2901 int cpu;
2902
2903 /* The event is discarded regardless */
2904 rb_event_discard(event);
2905
2906 cpu = smp_processor_id();
2907 cpu_buffer = buffer->buffers[cpu];
2908
2909 /*
2910 * This must only be called if the event has not been
2911 * committed yet. Thus we can assume that preemption
2912 * is still disabled.
2913 */
2914 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2915
2916 rb_decrement_entry(cpu_buffer, event);
2917 if (rb_try_to_discard(cpu_buffer, event))
2918 goto out;
2919
2920 /*
2921 * The commit is still visible by the reader, so we
2922 * must still update the timestamp.
2923 */
2924 rb_update_write_stamp(cpu_buffer, event);
2925 out:
2926 rb_end_commit(cpu_buffer);
2927
2928 trace_recursive_unlock(cpu_buffer);
2929
2930 preempt_enable_notrace();
2931
2932}
2933EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2934
2935/**
2936 * ring_buffer_write - write data to the buffer without reserving
2937 * @buffer: The ring buffer to write to.
2938 * @length: The length of the data being written (excluding the event header)
2939 * @data: The data to write to the buffer.
2940 *
2941 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2942 * one function. If you already have the data to write to the buffer, it
2943 * may be easier to simply call this function.
2944 *
2945 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2946 * and not the length of the event which would hold the header.
2947 */
2948int ring_buffer_write(struct ring_buffer *buffer,
2949 unsigned long length,
2950 void *data)
2951{
2952 struct ring_buffer_per_cpu *cpu_buffer;
2953 struct ring_buffer_event *event;
2954 void *body;
2955 int ret = -EBUSY;
2956 int cpu;
2957
2958 preempt_disable_notrace();
2959
2960 if (atomic_read(&buffer->record_disabled))
2961 goto out;
2962
2963 cpu = raw_smp_processor_id();
2964
2965 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2966 goto out;
2967
2968 cpu_buffer = buffer->buffers[cpu];
2969
2970 if (atomic_read(&cpu_buffer->record_disabled))
2971 goto out;
2972
2973 if (length > BUF_MAX_DATA_SIZE)
2974 goto out;
2975
2976 if (unlikely(trace_recursive_lock(cpu_buffer)))
2977 goto out;
2978
2979 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2980 if (!event)
2981 goto out_unlock;
2982
2983 body = rb_event_data(event);
2984
2985 memcpy(body, data, length);
2986
2987 rb_commit(cpu_buffer, event);
2988
2989 rb_wakeups(buffer, cpu_buffer);
2990
2991 ret = 0;
2992
2993 out_unlock:
2994 trace_recursive_unlock(cpu_buffer);
2995
2996 out:
2997 preempt_enable_notrace();
2998
2999 return ret;
3000}
3001EXPORT_SYMBOL_GPL(ring_buffer_write);
3002
3003static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3004{
3005 struct buffer_page *reader = cpu_buffer->reader_page;
3006 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3007 struct buffer_page *commit = cpu_buffer->commit_page;
3008
3009 /* In case of error, head will be NULL */
3010 if (unlikely(!head))
3011 return true;
3012
3013 return reader->read == rb_page_commit(reader) &&
3014 (commit == reader ||
3015 (commit == head &&
3016 head->read == rb_page_commit(commit)));
3017}
3018
3019/**
3020 * ring_buffer_record_disable - stop all writes into the buffer
3021 * @buffer: The ring buffer to stop writes to.
3022 *
3023 * This prevents all writes to the buffer. Any attempt to write
3024 * to the buffer after this will fail and return NULL.
3025 *
3026 * The caller should call synchronize_sched() after this.
3027 */
3028void ring_buffer_record_disable(struct ring_buffer *buffer)
3029{
3030 atomic_inc(&buffer->record_disabled);
3031}
3032EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3033
3034/**
3035 * ring_buffer_record_enable - enable writes to the buffer
3036 * @buffer: The ring buffer to enable writes
3037 *
3038 * Note, multiple disables will need the same number of enables
3039 * to truly enable the writing (much like preempt_disable).
3040 */
3041void ring_buffer_record_enable(struct ring_buffer *buffer)
3042{
3043 atomic_dec(&buffer->record_disabled);
3044}
3045EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3046
3047/**
3048 * ring_buffer_record_off - stop all writes into the buffer
3049 * @buffer: The ring buffer to stop writes to.
3050 *
3051 * This prevents all writes to the buffer. Any attempt to write
3052 * to the buffer after this will fail and return NULL.
3053 *
3054 * This is different than ring_buffer_record_disable() as
3055 * it works like an on/off switch, where as the disable() version
3056 * must be paired with a enable().
3057 */
3058void ring_buffer_record_off(struct ring_buffer *buffer)
3059{
3060 unsigned int rd;
3061 unsigned int new_rd;
3062
3063 do {
3064 rd = atomic_read(&buffer->record_disabled);
3065 new_rd = rd | RB_BUFFER_OFF;
3066 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3067}
3068EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3069
3070/**
3071 * ring_buffer_record_on - restart writes into the buffer
3072 * @buffer: The ring buffer to start writes to.
3073 *
3074 * This enables all writes to the buffer that was disabled by
3075 * ring_buffer_record_off().
3076 *
3077 * This is different than ring_buffer_record_enable() as
3078 * it works like an on/off switch, where as the enable() version
3079 * must be paired with a disable().
3080 */
3081void ring_buffer_record_on(struct ring_buffer *buffer)
3082{
3083 unsigned int rd;
3084 unsigned int new_rd;
3085
3086 do {
3087 rd = atomic_read(&buffer->record_disabled);
3088 new_rd = rd & ~RB_BUFFER_OFF;
3089 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3090}
3091EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3092
3093/**
3094 * ring_buffer_record_is_on - return true if the ring buffer can write
3095 * @buffer: The ring buffer to see if write is enabled
3096 *
3097 * Returns true if the ring buffer is in a state that it accepts writes.
3098 */
3099int ring_buffer_record_is_on(struct ring_buffer *buffer)
3100{
3101 return !atomic_read(&buffer->record_disabled);
3102}
3103
3104/**
3105 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3106 * @buffer: The ring buffer to stop writes to.
3107 * @cpu: The CPU buffer to stop
3108 *
3109 * This prevents all writes to the buffer. Any attempt to write
3110 * to the buffer after this will fail and return NULL.
3111 *
3112 * The caller should call synchronize_sched() after this.
3113 */
3114void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3115{
3116 struct ring_buffer_per_cpu *cpu_buffer;
3117
3118 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3119 return;
3120
3121 cpu_buffer = buffer->buffers[cpu];
3122 atomic_inc(&cpu_buffer->record_disabled);
3123}
3124EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3125
3126/**
3127 * ring_buffer_record_enable_cpu - enable writes to the buffer
3128 * @buffer: The ring buffer to enable writes
3129 * @cpu: The CPU to enable.
3130 *
3131 * Note, multiple disables will need the same number of enables
3132 * to truly enable the writing (much like preempt_disable).
3133 */
3134void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3135{
3136 struct ring_buffer_per_cpu *cpu_buffer;
3137
3138 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3139 return;
3140
3141 cpu_buffer = buffer->buffers[cpu];
3142 atomic_dec(&cpu_buffer->record_disabled);
3143}
3144EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3145
3146/*
3147 * The total entries in the ring buffer is the running counter
3148 * of entries entered into the ring buffer, minus the sum of
3149 * the entries read from the ring buffer and the number of
3150 * entries that were overwritten.
3151 */
3152static inline unsigned long
3153rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3154{
3155 return local_read(&cpu_buffer->entries) -
3156 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3157}
3158
3159/**
3160 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3161 * @buffer: The ring buffer
3162 * @cpu: The per CPU buffer to read from.
3163 */
3164u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3165{
3166 unsigned long flags;
3167 struct ring_buffer_per_cpu *cpu_buffer;
3168 struct buffer_page *bpage;
3169 u64 ret = 0;
3170
3171 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3172 return 0;
3173
3174 cpu_buffer = buffer->buffers[cpu];
3175 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3176 /*
3177 * if the tail is on reader_page, oldest time stamp is on the reader
3178 * page
3179 */
3180 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3181 bpage = cpu_buffer->reader_page;
3182 else
3183 bpage = rb_set_head_page(cpu_buffer);
3184 if (bpage)
3185 ret = bpage->page->time_stamp;
3186 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3187
3188 return ret;
3189}
3190EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3191
3192/**
3193 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3194 * @buffer: The ring buffer
3195 * @cpu: The per CPU buffer to read from.
3196 */
3197unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3198{
3199 struct ring_buffer_per_cpu *cpu_buffer;
3200 unsigned long ret;
3201
3202 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3203 return 0;
3204
3205 cpu_buffer = buffer->buffers[cpu];
3206 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3207
3208 return ret;
3209}
3210EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3211
3212/**
3213 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3214 * @buffer: The ring buffer
3215 * @cpu: The per CPU buffer to get the entries from.
3216 */
3217unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3218{
3219 struct ring_buffer_per_cpu *cpu_buffer;
3220
3221 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3222 return 0;
3223
3224 cpu_buffer = buffer->buffers[cpu];
3225
3226 return rb_num_of_entries(cpu_buffer);
3227}
3228EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3229
3230/**
3231 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3232 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3233 * @buffer: The ring buffer
3234 * @cpu: The per CPU buffer to get the number of overruns from
3235 */
3236unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3237{
3238 struct ring_buffer_per_cpu *cpu_buffer;
3239 unsigned long ret;
3240
3241 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3242 return 0;
3243
3244 cpu_buffer = buffer->buffers[cpu];
3245 ret = local_read(&cpu_buffer->overrun);
3246
3247 return ret;
3248}
3249EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3250
3251/**
3252 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3253 * commits failing due to the buffer wrapping around while there are uncommitted
3254 * events, such as during an interrupt storm.
3255 * @buffer: The ring buffer
3256 * @cpu: The per CPU buffer to get the number of overruns from
3257 */
3258unsigned long
3259ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3260{
3261 struct ring_buffer_per_cpu *cpu_buffer;
3262 unsigned long ret;
3263
3264 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3265 return 0;
3266
3267 cpu_buffer = buffer->buffers[cpu];
3268 ret = local_read(&cpu_buffer->commit_overrun);
3269
3270 return ret;
3271}
3272EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3273
3274/**
3275 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3276 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3277 * @buffer: The ring buffer
3278 * @cpu: The per CPU buffer to get the number of overruns from
3279 */
3280unsigned long
3281ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3282{
3283 struct ring_buffer_per_cpu *cpu_buffer;
3284 unsigned long ret;
3285
3286 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3287 return 0;
3288
3289 cpu_buffer = buffer->buffers[cpu];
3290 ret = local_read(&cpu_buffer->dropped_events);
3291
3292 return ret;
3293}
3294EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3295
3296/**
3297 * ring_buffer_read_events_cpu - get the number of events successfully read
3298 * @buffer: The ring buffer
3299 * @cpu: The per CPU buffer to get the number of events read
3300 */
3301unsigned long
3302ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3303{
3304 struct ring_buffer_per_cpu *cpu_buffer;
3305
3306 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3307 return 0;
3308
3309 cpu_buffer = buffer->buffers[cpu];
3310 return cpu_buffer->read;
3311}
3312EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3313
3314/**
3315 * ring_buffer_entries - get the number of entries in a buffer
3316 * @buffer: The ring buffer
3317 *
3318 * Returns the total number of entries in the ring buffer
3319 * (all CPU entries)
3320 */
3321unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3322{
3323 struct ring_buffer_per_cpu *cpu_buffer;
3324 unsigned long entries = 0;
3325 int cpu;
3326
3327 /* if you care about this being correct, lock the buffer */
3328 for_each_buffer_cpu(buffer, cpu) {
3329 cpu_buffer = buffer->buffers[cpu];
3330 entries += rb_num_of_entries(cpu_buffer);
3331 }
3332
3333 return entries;
3334}
3335EXPORT_SYMBOL_GPL(ring_buffer_entries);
3336
3337/**
3338 * ring_buffer_overruns - get the number of overruns in buffer
3339 * @buffer: The ring buffer
3340 *
3341 * Returns the total number of overruns in the ring buffer
3342 * (all CPU entries)
3343 */
3344unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3345{
3346 struct ring_buffer_per_cpu *cpu_buffer;
3347 unsigned long overruns = 0;
3348 int cpu;
3349
3350 /* if you care about this being correct, lock the buffer */
3351 for_each_buffer_cpu(buffer, cpu) {
3352 cpu_buffer = buffer->buffers[cpu];
3353 overruns += local_read(&cpu_buffer->overrun);
3354 }
3355
3356 return overruns;
3357}
3358EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3359
3360static void rb_iter_reset(struct ring_buffer_iter *iter)
3361{
3362 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3363
3364 /* Iterator usage is expected to have record disabled */
3365 iter->head_page = cpu_buffer->reader_page;
3366 iter->head = cpu_buffer->reader_page->read;
3367
3368 iter->cache_reader_page = iter->head_page;
3369 iter->cache_read = cpu_buffer->read;
3370
3371 if (iter->head)
3372 iter->read_stamp = cpu_buffer->read_stamp;
3373 else
3374 iter->read_stamp = iter->head_page->page->time_stamp;
3375}
3376
3377/**
3378 * ring_buffer_iter_reset - reset an iterator
3379 * @iter: The iterator to reset
3380 *
3381 * Resets the iterator, so that it will start from the beginning
3382 * again.
3383 */
3384void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3385{
3386 struct ring_buffer_per_cpu *cpu_buffer;
3387 unsigned long flags;
3388
3389 if (!iter)
3390 return;
3391
3392 cpu_buffer = iter->cpu_buffer;
3393
3394 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3395 rb_iter_reset(iter);
3396 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3397}
3398EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3399
3400/**
3401 * ring_buffer_iter_empty - check if an iterator has no more to read
3402 * @iter: The iterator to check
3403 */
3404int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3405{
3406 struct ring_buffer_per_cpu *cpu_buffer;
3407
3408 cpu_buffer = iter->cpu_buffer;
3409
3410 return iter->head_page == cpu_buffer->commit_page &&
3411 iter->head == rb_commit_index(cpu_buffer);
3412}
3413EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3414
3415static void
3416rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3417 struct ring_buffer_event *event)
3418{
3419 u64 delta;
3420
3421 switch (event->type_len) {
3422 case RINGBUF_TYPE_PADDING:
3423 return;
3424
3425 case RINGBUF_TYPE_TIME_EXTEND:
3426 delta = event->array[0];
3427 delta <<= TS_SHIFT;
3428 delta += event->time_delta;
3429 cpu_buffer->read_stamp += delta;
3430 return;
3431
3432 case RINGBUF_TYPE_TIME_STAMP:
3433 /* FIXME: not implemented */
3434 return;
3435
3436 case RINGBUF_TYPE_DATA:
3437 cpu_buffer->read_stamp += event->time_delta;
3438 return;
3439
3440 default:
3441 BUG();
3442 }
3443 return;
3444}
3445
3446static void
3447rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3448 struct ring_buffer_event *event)
3449{
3450 u64 delta;
3451
3452 switch (event->type_len) {
3453 case RINGBUF_TYPE_PADDING:
3454 return;
3455
3456 case RINGBUF_TYPE_TIME_EXTEND:
3457 delta = event->array[0];
3458 delta <<= TS_SHIFT;
3459 delta += event->time_delta;
3460 iter->read_stamp += delta;
3461 return;
3462
3463 case RINGBUF_TYPE_TIME_STAMP:
3464 /* FIXME: not implemented */
3465 return;
3466
3467 case RINGBUF_TYPE_DATA:
3468 iter->read_stamp += event->time_delta;
3469 return;
3470
3471 default:
3472 BUG();
3473 }
3474 return;
3475}
3476
3477static struct buffer_page *
3478rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3479{
3480 struct buffer_page *reader = NULL;
3481 unsigned long overwrite;
3482 unsigned long flags;
3483 int nr_loops = 0;
3484 int ret;
3485
3486 local_irq_save(flags);
3487 arch_spin_lock(&cpu_buffer->lock);
3488
3489 again:
3490 /*
3491 * This should normally only loop twice. But because the
3492 * start of the reader inserts an empty page, it causes
3493 * a case where we will loop three times. There should be no
3494 * reason to loop four times (that I know of).
3495 */
3496 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3497 reader = NULL;
3498 goto out;
3499 }
3500
3501 reader = cpu_buffer->reader_page;
3502
3503 /* If there's more to read, return this page */
3504 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3505 goto out;
3506
3507 /* Never should we have an index greater than the size */
3508 if (RB_WARN_ON(cpu_buffer,
3509 cpu_buffer->reader_page->read > rb_page_size(reader)))
3510 goto out;
3511
3512 /* check if we caught up to the tail */
3513 reader = NULL;
3514 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3515 goto out;
3516
3517 /* Don't bother swapping if the ring buffer is empty */
3518 if (rb_num_of_entries(cpu_buffer) == 0)
3519 goto out;
3520
3521 /*
3522 * Reset the reader page to size zero.
3523 */
3524 local_set(&cpu_buffer->reader_page->write, 0);
3525 local_set(&cpu_buffer->reader_page->entries, 0);
3526 local_set(&cpu_buffer->reader_page->page->commit, 0);
3527 cpu_buffer->reader_page->real_end = 0;
3528
3529 spin:
3530 /*
3531 * Splice the empty reader page into the list around the head.
3532 */
3533 reader = rb_set_head_page(cpu_buffer);
3534 if (!reader)
3535 goto out;
3536 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3537 cpu_buffer->reader_page->list.prev = reader->list.prev;
3538
3539 /*
3540 * cpu_buffer->pages just needs to point to the buffer, it
3541 * has no specific buffer page to point to. Lets move it out
3542 * of our way so we don't accidentally swap it.
3543 */
3544 cpu_buffer->pages = reader->list.prev;
3545
3546 /* The reader page will be pointing to the new head */
3547 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3548
3549 /*
3550 * We want to make sure we read the overruns after we set up our
3551 * pointers to the next object. The writer side does a
3552 * cmpxchg to cross pages which acts as the mb on the writer
3553 * side. Note, the reader will constantly fail the swap
3554 * while the writer is updating the pointers, so this
3555 * guarantees that the overwrite recorded here is the one we
3556 * want to compare with the last_overrun.
3557 */
3558 smp_mb();
3559 overwrite = local_read(&(cpu_buffer->overrun));
3560
3561 /*
3562 * Here's the tricky part.
3563 *
3564 * We need to move the pointer past the header page.
3565 * But we can only do that if a writer is not currently
3566 * moving it. The page before the header page has the
3567 * flag bit '1' set if it is pointing to the page we want.
3568 * but if the writer is in the process of moving it
3569 * than it will be '2' or already moved '0'.
3570 */
3571
3572 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3573
3574 /*
3575 * If we did not convert it, then we must try again.
3576 */
3577 if (!ret)
3578 goto spin;
3579
3580 /*
3581 * Yeah! We succeeded in replacing the page.
3582 *
3583 * Now make the new head point back to the reader page.
3584 */
3585 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3586 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3587
3588 /* Finally update the reader page to the new head */
3589 cpu_buffer->reader_page = reader;
3590 cpu_buffer->reader_page->read = 0;
3591
3592 if (overwrite != cpu_buffer->last_overrun) {
3593 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3594 cpu_buffer->last_overrun = overwrite;
3595 }
3596
3597 goto again;
3598
3599 out:
3600 /* Update the read_stamp on the first event */
3601 if (reader && reader->read == 0)
3602 cpu_buffer->read_stamp = reader->page->time_stamp;
3603
3604 arch_spin_unlock(&cpu_buffer->lock);
3605 local_irq_restore(flags);
3606
3607 return reader;
3608}
3609
3610static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3611{
3612 struct ring_buffer_event *event;
3613 struct buffer_page *reader;
3614 unsigned length;
3615
3616 reader = rb_get_reader_page(cpu_buffer);
3617
3618 /* This function should not be called when buffer is empty */
3619 if (RB_WARN_ON(cpu_buffer, !reader))
3620 return;
3621
3622 event = rb_reader_event(cpu_buffer);
3623
3624 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3625 cpu_buffer->read++;
3626
3627 rb_update_read_stamp(cpu_buffer, event);
3628
3629 length = rb_event_length(event);
3630 cpu_buffer->reader_page->read += length;
3631}
3632
3633static void rb_advance_iter(struct ring_buffer_iter *iter)
3634{
3635 struct ring_buffer_per_cpu *cpu_buffer;
3636 struct ring_buffer_event *event;
3637 unsigned length;
3638
3639 cpu_buffer = iter->cpu_buffer;
3640
3641 /*
3642 * Check if we are at the end of the buffer.
3643 */
3644 if (iter->head >= rb_page_size(iter->head_page)) {
3645 /* discarded commits can make the page empty */
3646 if (iter->head_page == cpu_buffer->commit_page)
3647 return;
3648 rb_inc_iter(iter);
3649 return;
3650 }
3651
3652 event = rb_iter_head_event(iter);
3653
3654 length = rb_event_length(event);
3655
3656 /*
3657 * This should not be called to advance the header if we are
3658 * at the tail of the buffer.
3659 */
3660 if (RB_WARN_ON(cpu_buffer,
3661 (iter->head_page == cpu_buffer->commit_page) &&
3662 (iter->head + length > rb_commit_index(cpu_buffer))))
3663 return;
3664
3665 rb_update_iter_read_stamp(iter, event);
3666
3667 iter->head += length;
3668
3669 /* check for end of page padding */
3670 if ((iter->head >= rb_page_size(iter->head_page)) &&
3671 (iter->head_page != cpu_buffer->commit_page))
3672 rb_inc_iter(iter);
3673}
3674
3675static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3676{
3677 return cpu_buffer->lost_events;
3678}
3679
3680static struct ring_buffer_event *
3681rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3682 unsigned long *lost_events)
3683{
3684 struct ring_buffer_event *event;
3685 struct buffer_page *reader;
3686 int nr_loops = 0;
3687
3688 again:
3689 /*
3690 * We repeat when a time extend is encountered.
3691 * Since the time extend is always attached to a data event,
3692 * we should never loop more than once.
3693 * (We never hit the following condition more than twice).
3694 */
3695 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3696 return NULL;
3697
3698 reader = rb_get_reader_page(cpu_buffer);
3699 if (!reader)
3700 return NULL;
3701
3702 event = rb_reader_event(cpu_buffer);
3703
3704 switch (event->type_len) {
3705 case RINGBUF_TYPE_PADDING:
3706 if (rb_null_event(event))
3707 RB_WARN_ON(cpu_buffer, 1);
3708 /*
3709 * Because the writer could be discarding every
3710 * event it creates (which would probably be bad)
3711 * if we were to go back to "again" then we may never
3712 * catch up, and will trigger the warn on, or lock
3713 * the box. Return the padding, and we will release
3714 * the current locks, and try again.
3715 */
3716 return event;
3717
3718 case RINGBUF_TYPE_TIME_EXTEND:
3719 /* Internal data, OK to advance */
3720 rb_advance_reader(cpu_buffer);
3721 goto again;
3722
3723 case RINGBUF_TYPE_TIME_STAMP:
3724 /* FIXME: not implemented */
3725 rb_advance_reader(cpu_buffer);
3726 goto again;
3727
3728 case RINGBUF_TYPE_DATA:
3729 if (ts) {
3730 *ts = cpu_buffer->read_stamp + event->time_delta;
3731 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3732 cpu_buffer->cpu, ts);
3733 }
3734 if (lost_events)
3735 *lost_events = rb_lost_events(cpu_buffer);
3736 return event;
3737
3738 default:
3739 BUG();
3740 }
3741
3742 return NULL;
3743}
3744EXPORT_SYMBOL_GPL(ring_buffer_peek);
3745
3746static struct ring_buffer_event *
3747rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3748{
3749 struct ring_buffer *buffer;
3750 struct ring_buffer_per_cpu *cpu_buffer;
3751 struct ring_buffer_event *event;
3752 int nr_loops = 0;
3753
3754 cpu_buffer = iter->cpu_buffer;
3755 buffer = cpu_buffer->buffer;
3756
3757 /*
3758 * Check if someone performed a consuming read to
3759 * the buffer. A consuming read invalidates the iterator
3760 * and we need to reset the iterator in this case.
3761 */
3762 if (unlikely(iter->cache_read != cpu_buffer->read ||
3763 iter->cache_reader_page != cpu_buffer->reader_page))
3764 rb_iter_reset(iter);
3765
3766 again:
3767 if (ring_buffer_iter_empty(iter))
3768 return NULL;
3769
3770 /*
3771 * We repeat when a time extend is encountered or we hit
3772 * the end of the page. Since the time extend is always attached
3773 * to a data event, we should never loop more than three times.
3774 * Once for going to next page, once on time extend, and
3775 * finally once to get the event.
3776 * (We never hit the following condition more than thrice).
3777 */
3778 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3779 return NULL;
3780
3781 if (rb_per_cpu_empty(cpu_buffer))
3782 return NULL;
3783
3784 if (iter->head >= rb_page_size(iter->head_page)) {
3785 rb_inc_iter(iter);
3786 goto again;
3787 }
3788
3789 event = rb_iter_head_event(iter);
3790
3791 switch (event->type_len) {
3792 case RINGBUF_TYPE_PADDING:
3793 if (rb_null_event(event)) {
3794 rb_inc_iter(iter);
3795 goto again;
3796 }
3797 rb_advance_iter(iter);
3798 return event;
3799
3800 case RINGBUF_TYPE_TIME_EXTEND:
3801 /* Internal data, OK to advance */
3802 rb_advance_iter(iter);
3803 goto again;
3804
3805 case RINGBUF_TYPE_TIME_STAMP:
3806 /* FIXME: not implemented */
3807 rb_advance_iter(iter);
3808 goto again;
3809
3810 case RINGBUF_TYPE_DATA:
3811 if (ts) {
3812 *ts = iter->read_stamp + event->time_delta;
3813 ring_buffer_normalize_time_stamp(buffer,
3814 cpu_buffer->cpu, ts);
3815 }
3816 return event;
3817
3818 default:
3819 BUG();
3820 }
3821
3822 return NULL;
3823}
3824EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3825
3826static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
3827{
3828 if (likely(!in_nmi())) {
3829 raw_spin_lock(&cpu_buffer->reader_lock);
3830 return true;
3831 }
3832
3833 /*
3834 * If an NMI die dumps out the content of the ring buffer
3835 * trylock must be used to prevent a deadlock if the NMI
3836 * preempted a task that holds the ring buffer locks. If
3837 * we get the lock then all is fine, if not, then continue
3838 * to do the read, but this can corrupt the ring buffer,
3839 * so it must be permanently disabled from future writes.
3840 * Reading from NMI is a oneshot deal.
3841 */
3842 if (raw_spin_trylock(&cpu_buffer->reader_lock))
3843 return true;
3844
3845 /* Continue without locking, but disable the ring buffer */
3846 atomic_inc(&cpu_buffer->record_disabled);
3847 return false;
3848}
3849
3850static inline void
3851rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
3852{
3853 if (likely(locked))
3854 raw_spin_unlock(&cpu_buffer->reader_lock);
3855 return;
3856}
3857
3858/**
3859 * ring_buffer_peek - peek at the next event to be read
3860 * @buffer: The ring buffer to read
3861 * @cpu: The cpu to peak at
3862 * @ts: The timestamp counter of this event.
3863 * @lost_events: a variable to store if events were lost (may be NULL)
3864 *
3865 * This will return the event that will be read next, but does
3866 * not consume the data.
3867 */
3868struct ring_buffer_event *
3869ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3870 unsigned long *lost_events)
3871{
3872 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3873 struct ring_buffer_event *event;
3874 unsigned long flags;
3875 bool dolock;
3876
3877 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3878 return NULL;
3879
3880 again:
3881 local_irq_save(flags);
3882 dolock = rb_reader_lock(cpu_buffer);
3883 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3884 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3885 rb_advance_reader(cpu_buffer);
3886 rb_reader_unlock(cpu_buffer, dolock);
3887 local_irq_restore(flags);
3888
3889 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3890 goto again;
3891
3892 return event;
3893}
3894
3895/**
3896 * ring_buffer_iter_peek - peek at the next event to be read
3897 * @iter: The ring buffer iterator
3898 * @ts: The timestamp counter of this event.
3899 *
3900 * This will return the event that will be read next, but does
3901 * not increment the iterator.
3902 */
3903struct ring_buffer_event *
3904ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3905{
3906 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3907 struct ring_buffer_event *event;
3908 unsigned long flags;
3909
3910 again:
3911 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3912 event = rb_iter_peek(iter, ts);
3913 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3914
3915 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3916 goto again;
3917
3918 return event;
3919}
3920
3921/**
3922 * ring_buffer_consume - return an event and consume it
3923 * @buffer: The ring buffer to get the next event from
3924 * @cpu: the cpu to read the buffer from
3925 * @ts: a variable to store the timestamp (may be NULL)
3926 * @lost_events: a variable to store if events were lost (may be NULL)
3927 *
3928 * Returns the next event in the ring buffer, and that event is consumed.
3929 * Meaning, that sequential reads will keep returning a different event,
3930 * and eventually empty the ring buffer if the producer is slower.
3931 */
3932struct ring_buffer_event *
3933ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3934 unsigned long *lost_events)
3935{
3936 struct ring_buffer_per_cpu *cpu_buffer;
3937 struct ring_buffer_event *event = NULL;
3938 unsigned long flags;
3939 bool dolock;
3940
3941 again:
3942 /* might be called in atomic */
3943 preempt_disable();
3944
3945 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3946 goto out;
3947
3948 cpu_buffer = buffer->buffers[cpu];
3949 local_irq_save(flags);
3950 dolock = rb_reader_lock(cpu_buffer);
3951
3952 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3953 if (event) {
3954 cpu_buffer->lost_events = 0;
3955 rb_advance_reader(cpu_buffer);
3956 }
3957
3958 rb_reader_unlock(cpu_buffer, dolock);
3959 local_irq_restore(flags);
3960
3961 out:
3962 preempt_enable();
3963
3964 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3965 goto again;
3966
3967 return event;
3968}
3969EXPORT_SYMBOL_GPL(ring_buffer_consume);
3970
3971/**
3972 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3973 * @buffer: The ring buffer to read from
3974 * @cpu: The cpu buffer to iterate over
3975 *
3976 * This performs the initial preparations necessary to iterate
3977 * through the buffer. Memory is allocated, buffer recording
3978 * is disabled, and the iterator pointer is returned to the caller.
3979 *
3980 * Disabling buffer recordng prevents the reading from being
3981 * corrupted. This is not a consuming read, so a producer is not
3982 * expected.
3983 *
3984 * After a sequence of ring_buffer_read_prepare calls, the user is
3985 * expected to make at least one call to ring_buffer_read_prepare_sync.
3986 * Afterwards, ring_buffer_read_start is invoked to get things going
3987 * for real.
3988 *
3989 * This overall must be paired with ring_buffer_read_finish.
3990 */
3991struct ring_buffer_iter *
3992ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3993{
3994 struct ring_buffer_per_cpu *cpu_buffer;
3995 struct ring_buffer_iter *iter;
3996
3997 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3998 return NULL;
3999
4000 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
4001 if (!iter)
4002 return NULL;
4003
4004 cpu_buffer = buffer->buffers[cpu];
4005
4006 iter->cpu_buffer = cpu_buffer;
4007
4008 atomic_inc(&buffer->resize_disabled);
4009 atomic_inc(&cpu_buffer->record_disabled);
4010
4011 return iter;
4012}
4013EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4014
4015/**
4016 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4017 *
4018 * All previously invoked ring_buffer_read_prepare calls to prepare
4019 * iterators will be synchronized. Afterwards, read_buffer_read_start
4020 * calls on those iterators are allowed.
4021 */
4022void
4023ring_buffer_read_prepare_sync(void)
4024{
4025 synchronize_sched();
4026}
4027EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4028
4029/**
4030 * ring_buffer_read_start - start a non consuming read of the buffer
4031 * @iter: The iterator returned by ring_buffer_read_prepare
4032 *
4033 * This finalizes the startup of an iteration through the buffer.
4034 * The iterator comes from a call to ring_buffer_read_prepare and
4035 * an intervening ring_buffer_read_prepare_sync must have been
4036 * performed.
4037 *
4038 * Must be paired with ring_buffer_read_finish.
4039 */
4040void
4041ring_buffer_read_start(struct ring_buffer_iter *iter)
4042{
4043 struct ring_buffer_per_cpu *cpu_buffer;
4044 unsigned long flags;
4045
4046 if (!iter)
4047 return;
4048
4049 cpu_buffer = iter->cpu_buffer;
4050
4051 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4052 arch_spin_lock(&cpu_buffer->lock);
4053 rb_iter_reset(iter);
4054 arch_spin_unlock(&cpu_buffer->lock);
4055 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4056}
4057EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4058
4059/**
4060 * ring_buffer_read_finish - finish reading the iterator of the buffer
4061 * @iter: The iterator retrieved by ring_buffer_start
4062 *
4063 * This re-enables the recording to the buffer, and frees the
4064 * iterator.
4065 */
4066void
4067ring_buffer_read_finish(struct ring_buffer_iter *iter)
4068{
4069 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4070 unsigned long flags;
4071
4072 /*
4073 * Ring buffer is disabled from recording, here's a good place
4074 * to check the integrity of the ring buffer.
4075 * Must prevent readers from trying to read, as the check
4076 * clears the HEAD page and readers require it.
4077 */
4078 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4079 rb_check_pages(cpu_buffer);
4080 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4081
4082 atomic_dec(&cpu_buffer->record_disabled);
4083 atomic_dec(&cpu_buffer->buffer->resize_disabled);
4084 kfree(iter);
4085}
4086EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4087
4088/**
4089 * ring_buffer_read - read the next item in the ring buffer by the iterator
4090 * @iter: The ring buffer iterator
4091 * @ts: The time stamp of the event read.
4092 *
4093 * This reads the next event in the ring buffer and increments the iterator.
4094 */
4095struct ring_buffer_event *
4096ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4097{
4098 struct ring_buffer_event *event;
4099 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4100 unsigned long flags;
4101
4102 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4103 again:
4104 event = rb_iter_peek(iter, ts);
4105 if (!event)
4106 goto out;
4107
4108 if (event->type_len == RINGBUF_TYPE_PADDING)
4109 goto again;
4110
4111 rb_advance_iter(iter);
4112 out:
4113 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4114
4115 return event;
4116}
4117EXPORT_SYMBOL_GPL(ring_buffer_read);
4118
4119/**
4120 * ring_buffer_size - return the size of the ring buffer (in bytes)
4121 * @buffer: The ring buffer.
4122 */
4123unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4124{
4125 /*
4126 * Earlier, this method returned
4127 * BUF_PAGE_SIZE * buffer->nr_pages
4128 * Since the nr_pages field is now removed, we have converted this to
4129 * return the per cpu buffer value.
4130 */
4131 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4132 return 0;
4133
4134 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4135}
4136EXPORT_SYMBOL_GPL(ring_buffer_size);
4137
4138static void
4139rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4140{
4141 rb_head_page_deactivate(cpu_buffer);
4142
4143 cpu_buffer->head_page
4144 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4145 local_set(&cpu_buffer->head_page->write, 0);
4146 local_set(&cpu_buffer->head_page->entries, 0);
4147 local_set(&cpu_buffer->head_page->page->commit, 0);
4148
4149 cpu_buffer->head_page->read = 0;
4150
4151 cpu_buffer->tail_page = cpu_buffer->head_page;
4152 cpu_buffer->commit_page = cpu_buffer->head_page;
4153
4154 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4155 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4156 local_set(&cpu_buffer->reader_page->write, 0);
4157 local_set(&cpu_buffer->reader_page->entries, 0);
4158 local_set(&cpu_buffer->reader_page->page->commit, 0);
4159 cpu_buffer->reader_page->read = 0;
4160
4161 local_set(&cpu_buffer->entries_bytes, 0);
4162 local_set(&cpu_buffer->overrun, 0);
4163 local_set(&cpu_buffer->commit_overrun, 0);
4164 local_set(&cpu_buffer->dropped_events, 0);
4165 local_set(&cpu_buffer->entries, 0);
4166 local_set(&cpu_buffer->committing, 0);
4167 local_set(&cpu_buffer->commits, 0);
4168 cpu_buffer->read = 0;
4169 cpu_buffer->read_bytes = 0;
4170
4171 cpu_buffer->write_stamp = 0;
4172 cpu_buffer->read_stamp = 0;
4173
4174 cpu_buffer->lost_events = 0;
4175 cpu_buffer->last_overrun = 0;
4176
4177 rb_head_page_activate(cpu_buffer);
4178}
4179
4180/**
4181 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4182 * @buffer: The ring buffer to reset a per cpu buffer of
4183 * @cpu: The CPU buffer to be reset
4184 */
4185void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4186{
4187 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4188 unsigned long flags;
4189
4190 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4191 return;
4192
4193 atomic_inc(&buffer->resize_disabled);
4194 atomic_inc(&cpu_buffer->record_disabled);
4195
4196 /* Make sure all commits have finished */
4197 synchronize_sched();
4198
4199 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4200
4201 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4202 goto out;
4203
4204 arch_spin_lock(&cpu_buffer->lock);
4205
4206 rb_reset_cpu(cpu_buffer);
4207
4208 arch_spin_unlock(&cpu_buffer->lock);
4209
4210 out:
4211 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4212
4213 atomic_dec(&cpu_buffer->record_disabled);
4214 atomic_dec(&buffer->resize_disabled);
4215}
4216EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4217
4218/**
4219 * ring_buffer_reset - reset a ring buffer
4220 * @buffer: The ring buffer to reset all cpu buffers
4221 */
4222void ring_buffer_reset(struct ring_buffer *buffer)
4223{
4224 int cpu;
4225
4226 for_each_buffer_cpu(buffer, cpu)
4227 ring_buffer_reset_cpu(buffer, cpu);
4228}
4229EXPORT_SYMBOL_GPL(ring_buffer_reset);
4230
4231/**
4232 * rind_buffer_empty - is the ring buffer empty?
4233 * @buffer: The ring buffer to test
4234 */
4235bool ring_buffer_empty(struct ring_buffer *buffer)
4236{
4237 struct ring_buffer_per_cpu *cpu_buffer;
4238 unsigned long flags;
4239 bool dolock;
4240 int cpu;
4241 int ret;
4242
4243 /* yes this is racy, but if you don't like the race, lock the buffer */
4244 for_each_buffer_cpu(buffer, cpu) {
4245 cpu_buffer = buffer->buffers[cpu];
4246 local_irq_save(flags);
4247 dolock = rb_reader_lock(cpu_buffer);
4248 ret = rb_per_cpu_empty(cpu_buffer);
4249 rb_reader_unlock(cpu_buffer, dolock);
4250 local_irq_restore(flags);
4251
4252 if (!ret)
4253 return false;
4254 }
4255
4256 return true;
4257}
4258EXPORT_SYMBOL_GPL(ring_buffer_empty);
4259
4260/**
4261 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4262 * @buffer: The ring buffer
4263 * @cpu: The CPU buffer to test
4264 */
4265bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4266{
4267 struct ring_buffer_per_cpu *cpu_buffer;
4268 unsigned long flags;
4269 bool dolock;
4270 int ret;
4271
4272 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4273 return true;
4274
4275 cpu_buffer = buffer->buffers[cpu];
4276 local_irq_save(flags);
4277 dolock = rb_reader_lock(cpu_buffer);
4278 ret = rb_per_cpu_empty(cpu_buffer);
4279 rb_reader_unlock(cpu_buffer, dolock);
4280 local_irq_restore(flags);
4281
4282 return ret;
4283}
4284EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4285
4286#ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4287/**
4288 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4289 * @buffer_a: One buffer to swap with
4290 * @buffer_b: The other buffer to swap with
4291 *
4292 * This function is useful for tracers that want to take a "snapshot"
4293 * of a CPU buffer and has another back up buffer lying around.
4294 * it is expected that the tracer handles the cpu buffer not being
4295 * used at the moment.
4296 */
4297int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4298 struct ring_buffer *buffer_b, int cpu)
4299{
4300 struct ring_buffer_per_cpu *cpu_buffer_a;
4301 struct ring_buffer_per_cpu *cpu_buffer_b;
4302 int ret = -EINVAL;
4303
4304 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4305 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4306 goto out;
4307
4308 cpu_buffer_a = buffer_a->buffers[cpu];
4309 cpu_buffer_b = buffer_b->buffers[cpu];
4310
4311 /* At least make sure the two buffers are somewhat the same */
4312 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4313 goto out;
4314
4315 ret = -EAGAIN;
4316
4317 if (atomic_read(&buffer_a->record_disabled))
4318 goto out;
4319
4320 if (atomic_read(&buffer_b->record_disabled))
4321 goto out;
4322
4323 if (atomic_read(&cpu_buffer_a->record_disabled))
4324 goto out;
4325
4326 if (atomic_read(&cpu_buffer_b->record_disabled))
4327 goto out;
4328
4329 /*
4330 * We can't do a synchronize_sched here because this
4331 * function can be called in atomic context.
4332 * Normally this will be called from the same CPU as cpu.
4333 * If not it's up to the caller to protect this.
4334 */
4335 atomic_inc(&cpu_buffer_a->record_disabled);
4336 atomic_inc(&cpu_buffer_b->record_disabled);
4337
4338 ret = -EBUSY;
4339 if (local_read(&cpu_buffer_a->committing))
4340 goto out_dec;
4341 if (local_read(&cpu_buffer_b->committing))
4342 goto out_dec;
4343
4344 buffer_a->buffers[cpu] = cpu_buffer_b;
4345 buffer_b->buffers[cpu] = cpu_buffer_a;
4346
4347 cpu_buffer_b->buffer = buffer_a;
4348 cpu_buffer_a->buffer = buffer_b;
4349
4350 ret = 0;
4351
4352out_dec:
4353 atomic_dec(&cpu_buffer_a->record_disabled);
4354 atomic_dec(&cpu_buffer_b->record_disabled);
4355out:
4356 return ret;
4357}
4358EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4359#endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4360
4361/**
4362 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4363 * @buffer: the buffer to allocate for.
4364 * @cpu: the cpu buffer to allocate.
4365 *
4366 * This function is used in conjunction with ring_buffer_read_page.
4367 * When reading a full page from the ring buffer, these functions
4368 * can be used to speed up the process. The calling function should
4369 * allocate a few pages first with this function. Then when it
4370 * needs to get pages from the ring buffer, it passes the result
4371 * of this function into ring_buffer_read_page, which will swap
4372 * the page that was allocated, with the read page of the buffer.
4373 *
4374 * Returns:
4375 * The page allocated, or NULL on error.
4376 */
4377void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4378{
4379 struct buffer_data_page *bpage;
4380 struct page *page;
4381
4382 page = alloc_pages_node(cpu_to_node(cpu),
4383 GFP_KERNEL | __GFP_NORETRY, 0);
4384 if (!page)
4385 return NULL;
4386
4387 bpage = page_address(page);
4388
4389 rb_init_page(bpage);
4390
4391 return bpage;
4392}
4393EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4394
4395/**
4396 * ring_buffer_free_read_page - free an allocated read page
4397 * @buffer: the buffer the page was allocate for
4398 * @data: the page to free
4399 *
4400 * Free a page allocated from ring_buffer_alloc_read_page.
4401 */
4402void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4403{
4404 free_page((unsigned long)data);
4405}
4406EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4407
4408/**
4409 * ring_buffer_read_page - extract a page from the ring buffer
4410 * @buffer: buffer to extract from
4411 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4412 * @len: amount to extract
4413 * @cpu: the cpu of the buffer to extract
4414 * @full: should the extraction only happen when the page is full.
4415 *
4416 * This function will pull out a page from the ring buffer and consume it.
4417 * @data_page must be the address of the variable that was returned
4418 * from ring_buffer_alloc_read_page. This is because the page might be used
4419 * to swap with a page in the ring buffer.
4420 *
4421 * for example:
4422 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4423 * if (!rpage)
4424 * return error;
4425 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4426 * if (ret >= 0)
4427 * process_page(rpage, ret);
4428 *
4429 * When @full is set, the function will not return true unless
4430 * the writer is off the reader page.
4431 *
4432 * Note: it is up to the calling functions to handle sleeps and wakeups.
4433 * The ring buffer can be used anywhere in the kernel and can not
4434 * blindly call wake_up. The layer that uses the ring buffer must be
4435 * responsible for that.
4436 *
4437 * Returns:
4438 * >=0 if data has been transferred, returns the offset of consumed data.
4439 * <0 if no data has been transferred.
4440 */
4441int ring_buffer_read_page(struct ring_buffer *buffer,
4442 void **data_page, size_t len, int cpu, int full)
4443{
4444 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4445 struct ring_buffer_event *event;
4446 struct buffer_data_page *bpage;
4447 struct buffer_page *reader;
4448 unsigned long missed_events;
4449 unsigned long flags;
4450 unsigned int commit;
4451 unsigned int read;
4452 u64 save_timestamp;
4453 int ret = -1;
4454
4455 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4456 goto out;
4457
4458 /*
4459 * If len is not big enough to hold the page header, then
4460 * we can not copy anything.
4461 */
4462 if (len <= BUF_PAGE_HDR_SIZE)
4463 goto out;
4464
4465 len -= BUF_PAGE_HDR_SIZE;
4466
4467 if (!data_page)
4468 goto out;
4469
4470 bpage = *data_page;
4471 if (!bpage)
4472 goto out;
4473
4474 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4475
4476 reader = rb_get_reader_page(cpu_buffer);
4477 if (!reader)
4478 goto out_unlock;
4479
4480 event = rb_reader_event(cpu_buffer);
4481
4482 read = reader->read;
4483 commit = rb_page_commit(reader);
4484
4485 /* Check if any events were dropped */
4486 missed_events = cpu_buffer->lost_events;
4487
4488 /*
4489 * If this page has been partially read or
4490 * if len is not big enough to read the rest of the page or
4491 * a writer is still on the page, then
4492 * we must copy the data from the page to the buffer.
4493 * Otherwise, we can simply swap the page with the one passed in.
4494 */
4495 if (read || (len < (commit - read)) ||
4496 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4497 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4498 unsigned int rpos = read;
4499 unsigned int pos = 0;
4500 unsigned int size;
4501
4502 if (full)
4503 goto out_unlock;
4504
4505 if (len > (commit - read))
4506 len = (commit - read);
4507
4508 /* Always keep the time extend and data together */
4509 size = rb_event_ts_length(event);
4510
4511 if (len < size)
4512 goto out_unlock;
4513
4514 /* save the current timestamp, since the user will need it */
4515 save_timestamp = cpu_buffer->read_stamp;
4516
4517 /* Need to copy one event at a time */
4518 do {
4519 /* We need the size of one event, because
4520 * rb_advance_reader only advances by one event,
4521 * whereas rb_event_ts_length may include the size of
4522 * one or two events.
4523 * We have already ensured there's enough space if this
4524 * is a time extend. */
4525 size = rb_event_length(event);
4526 memcpy(bpage->data + pos, rpage->data + rpos, size);
4527
4528 len -= size;
4529
4530 rb_advance_reader(cpu_buffer);
4531 rpos = reader->read;
4532 pos += size;
4533
4534 if (rpos >= commit)
4535 break;
4536
4537 event = rb_reader_event(cpu_buffer);
4538 /* Always keep the time extend and data together */
4539 size = rb_event_ts_length(event);
4540 } while (len >= size);
4541
4542 /* update bpage */
4543 local_set(&bpage->commit, pos);
4544 bpage->time_stamp = save_timestamp;
4545
4546 /* we copied everything to the beginning */
4547 read = 0;
4548 } else {
4549 /* update the entry counter */
4550 cpu_buffer->read += rb_page_entries(reader);
4551 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4552
4553 /* swap the pages */
4554 rb_init_page(bpage);
4555 bpage = reader->page;
4556 reader->page = *data_page;
4557 local_set(&reader->write, 0);
4558 local_set(&reader->entries, 0);
4559 reader->read = 0;
4560 *data_page = bpage;
4561
4562 /*
4563 * Use the real_end for the data size,
4564 * This gives us a chance to store the lost events
4565 * on the page.
4566 */
4567 if (reader->real_end)
4568 local_set(&bpage->commit, reader->real_end);
4569 }
4570 ret = read;
4571
4572 cpu_buffer->lost_events = 0;
4573
4574 commit = local_read(&bpage->commit);
4575 /*
4576 * Set a flag in the commit field if we lost events
4577 */
4578 if (missed_events) {
4579 /* If there is room at the end of the page to save the
4580 * missed events, then record it there.
4581 */
4582 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4583 memcpy(&bpage->data[commit], &missed_events,
4584 sizeof(missed_events));
4585 local_add(RB_MISSED_STORED, &bpage->commit);
4586 commit += sizeof(missed_events);
4587 }
4588 local_add(RB_MISSED_EVENTS, &bpage->commit);
4589 }
4590
4591 /*
4592 * This page may be off to user land. Zero it out here.
4593 */
4594 if (commit < BUF_PAGE_SIZE)
4595 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4596
4597 out_unlock:
4598 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4599
4600 out:
4601 return ret;
4602}
4603EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4604
4605/*
4606 * We only allocate new buffers, never free them if the CPU goes down.
4607 * If we were to free the buffer, then the user would lose any trace that was in
4608 * the buffer.
4609 */
4610int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
4611{
4612 struct ring_buffer *buffer;
4613 long nr_pages_same;
4614 int cpu_i;
4615 unsigned long nr_pages;
4616
4617 buffer = container_of(node, struct ring_buffer, node);
4618 if (cpumask_test_cpu(cpu, buffer->cpumask))
4619 return 0;
4620
4621 nr_pages = 0;
4622 nr_pages_same = 1;
4623 /* check if all cpu sizes are same */
4624 for_each_buffer_cpu(buffer, cpu_i) {
4625 /* fill in the size from first enabled cpu */
4626 if (nr_pages == 0)
4627 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4628 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4629 nr_pages_same = 0;
4630 break;
4631 }
4632 }
4633 /* allocate minimum pages, user can later expand it */
4634 if (!nr_pages_same)
4635 nr_pages = 2;
4636 buffer->buffers[cpu] =
4637 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4638 if (!buffer->buffers[cpu]) {
4639 WARN(1, "failed to allocate ring buffer on CPU %u\n",
4640 cpu);
4641 return -ENOMEM;
4642 }
4643 smp_wmb();
4644 cpumask_set_cpu(cpu, buffer->cpumask);
4645 return 0;
4646}
4647
4648#ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4649/*
4650 * This is a basic integrity check of the ring buffer.
4651 * Late in the boot cycle this test will run when configured in.
4652 * It will kick off a thread per CPU that will go into a loop
4653 * writing to the per cpu ring buffer various sizes of data.
4654 * Some of the data will be large items, some small.
4655 *
4656 * Another thread is created that goes into a spin, sending out
4657 * IPIs to the other CPUs to also write into the ring buffer.
4658 * this is to test the nesting ability of the buffer.
4659 *
4660 * Basic stats are recorded and reported. If something in the
4661 * ring buffer should happen that's not expected, a big warning
4662 * is displayed and all ring buffers are disabled.
4663 */
4664static struct task_struct *rb_threads[NR_CPUS] __initdata;
4665
4666struct rb_test_data {
4667 struct ring_buffer *buffer;
4668 unsigned long events;
4669 unsigned long bytes_written;
4670 unsigned long bytes_alloc;
4671 unsigned long bytes_dropped;
4672 unsigned long events_nested;
4673 unsigned long bytes_written_nested;
4674 unsigned long bytes_alloc_nested;
4675 unsigned long bytes_dropped_nested;
4676 int min_size_nested;
4677 int max_size_nested;
4678 int max_size;
4679 int min_size;
4680 int cpu;
4681 int cnt;
4682};
4683
4684static struct rb_test_data rb_data[NR_CPUS] __initdata;
4685
4686/* 1 meg per cpu */
4687#define RB_TEST_BUFFER_SIZE 1048576
4688
4689static char rb_string[] __initdata =
4690 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4691 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4692 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4693
4694static bool rb_test_started __initdata;
4695
4696struct rb_item {
4697 int size;
4698 char str[];
4699};
4700
4701static __init int rb_write_something(struct rb_test_data *data, bool nested)
4702{
4703 struct ring_buffer_event *event;
4704 struct rb_item *item;
4705 bool started;
4706 int event_len;
4707 int size;
4708 int len;
4709 int cnt;
4710
4711 /* Have nested writes different that what is written */
4712 cnt = data->cnt + (nested ? 27 : 0);
4713
4714 /* Multiply cnt by ~e, to make some unique increment */
4715 size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4716
4717 len = size + sizeof(struct rb_item);
4718
4719 started = rb_test_started;
4720 /* read rb_test_started before checking buffer enabled */
4721 smp_rmb();
4722
4723 event = ring_buffer_lock_reserve(data->buffer, len);
4724 if (!event) {
4725 /* Ignore dropped events before test starts. */
4726 if (started) {
4727 if (nested)
4728 data->bytes_dropped += len;
4729 else
4730 data->bytes_dropped_nested += len;
4731 }
4732 return len;
4733 }
4734
4735 event_len = ring_buffer_event_length(event);
4736
4737 if (RB_WARN_ON(data->buffer, event_len < len))
4738 goto out;
4739
4740 item = ring_buffer_event_data(event);
4741 item->size = size;
4742 memcpy(item->str, rb_string, size);
4743
4744 if (nested) {
4745 data->bytes_alloc_nested += event_len;
4746 data->bytes_written_nested += len;
4747 data->events_nested++;
4748 if (!data->min_size_nested || len < data->min_size_nested)
4749 data->min_size_nested = len;
4750 if (len > data->max_size_nested)
4751 data->max_size_nested = len;
4752 } else {
4753 data->bytes_alloc += event_len;
4754 data->bytes_written += len;
4755 data->events++;
4756 if (!data->min_size || len < data->min_size)
4757 data->max_size = len;
4758 if (len > data->max_size)
4759 data->max_size = len;
4760 }
4761
4762 out:
4763 ring_buffer_unlock_commit(data->buffer, event);
4764
4765 return 0;
4766}
4767
4768static __init int rb_test(void *arg)
4769{
4770 struct rb_test_data *data = arg;
4771
4772 while (!kthread_should_stop()) {
4773 rb_write_something(data, false);
4774 data->cnt++;
4775
4776 set_current_state(TASK_INTERRUPTIBLE);
4777 /* Now sleep between a min of 100-300us and a max of 1ms */
4778 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4779 }
4780
4781 return 0;
4782}
4783
4784static __init void rb_ipi(void *ignore)
4785{
4786 struct rb_test_data *data;
4787 int cpu = smp_processor_id();
4788
4789 data = &rb_data[cpu];
4790 rb_write_something(data, true);
4791}
4792
4793static __init int rb_hammer_test(void *arg)
4794{
4795 while (!kthread_should_stop()) {
4796
4797 /* Send an IPI to all cpus to write data! */
4798 smp_call_function(rb_ipi, NULL, 1);
4799 /* No sleep, but for non preempt, let others run */
4800 schedule();
4801 }
4802
4803 return 0;
4804}
4805
4806static __init int test_ringbuffer(void)
4807{
4808 struct task_struct *rb_hammer;
4809 struct ring_buffer *buffer;
4810 int cpu;
4811 int ret = 0;
4812
4813 pr_info("Running ring buffer tests...\n");
4814
4815 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4816 if (WARN_ON(!buffer))
4817 return 0;
4818
4819 /* Disable buffer so that threads can't write to it yet */
4820 ring_buffer_record_off(buffer);
4821
4822 for_each_online_cpu(cpu) {
4823 rb_data[cpu].buffer = buffer;
4824 rb_data[cpu].cpu = cpu;
4825 rb_data[cpu].cnt = cpu;
4826 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4827 "rbtester/%d", cpu);
4828 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
4829 pr_cont("FAILED\n");
4830 ret = PTR_ERR(rb_threads[cpu]);
4831 goto out_free;
4832 }
4833
4834 kthread_bind(rb_threads[cpu], cpu);
4835 wake_up_process(rb_threads[cpu]);
4836 }
4837
4838 /* Now create the rb hammer! */
4839 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4840 if (WARN_ON(IS_ERR(rb_hammer))) {
4841 pr_cont("FAILED\n");
4842 ret = PTR_ERR(rb_hammer);
4843 goto out_free;
4844 }
4845
4846 ring_buffer_record_on(buffer);
4847 /*
4848 * Show buffer is enabled before setting rb_test_started.
4849 * Yes there's a small race window where events could be
4850 * dropped and the thread wont catch it. But when a ring
4851 * buffer gets enabled, there will always be some kind of
4852 * delay before other CPUs see it. Thus, we don't care about
4853 * those dropped events. We care about events dropped after
4854 * the threads see that the buffer is active.
4855 */
4856 smp_wmb();
4857 rb_test_started = true;
4858
4859 set_current_state(TASK_INTERRUPTIBLE);
4860 /* Just run for 10 seconds */;
4861 schedule_timeout(10 * HZ);
4862
4863 kthread_stop(rb_hammer);
4864
4865 out_free:
4866 for_each_online_cpu(cpu) {
4867 if (!rb_threads[cpu])
4868 break;
4869 kthread_stop(rb_threads[cpu]);
4870 }
4871 if (ret) {
4872 ring_buffer_free(buffer);
4873 return ret;
4874 }
4875
4876 /* Report! */
4877 pr_info("finished\n");
4878 for_each_online_cpu(cpu) {
4879 struct ring_buffer_event *event;
4880 struct rb_test_data *data = &rb_data[cpu];
4881 struct rb_item *item;
4882 unsigned long total_events;
4883 unsigned long total_dropped;
4884 unsigned long total_written;
4885 unsigned long total_alloc;
4886 unsigned long total_read = 0;
4887 unsigned long total_size = 0;
4888 unsigned long total_len = 0;
4889 unsigned long total_lost = 0;
4890 unsigned long lost;
4891 int big_event_size;
4892 int small_event_size;
4893
4894 ret = -1;
4895
4896 total_events = data->events + data->events_nested;
4897 total_written = data->bytes_written + data->bytes_written_nested;
4898 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
4899 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
4900
4901 big_event_size = data->max_size + data->max_size_nested;
4902 small_event_size = data->min_size + data->min_size_nested;
4903
4904 pr_info("CPU %d:\n", cpu);
4905 pr_info(" events: %ld\n", total_events);
4906 pr_info(" dropped bytes: %ld\n", total_dropped);
4907 pr_info(" alloced bytes: %ld\n", total_alloc);
4908 pr_info(" written bytes: %ld\n", total_written);
4909 pr_info(" biggest event: %d\n", big_event_size);
4910 pr_info(" smallest event: %d\n", small_event_size);
4911
4912 if (RB_WARN_ON(buffer, total_dropped))
4913 break;
4914
4915 ret = 0;
4916
4917 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
4918 total_lost += lost;
4919 item = ring_buffer_event_data(event);
4920 total_len += ring_buffer_event_length(event);
4921 total_size += item->size + sizeof(struct rb_item);
4922 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
4923 pr_info("FAILED!\n");
4924 pr_info("buffer had: %.*s\n", item->size, item->str);
4925 pr_info("expected: %.*s\n", item->size, rb_string);
4926 RB_WARN_ON(buffer, 1);
4927 ret = -1;
4928 break;
4929 }
4930 total_read++;
4931 }
4932 if (ret)
4933 break;
4934
4935 ret = -1;
4936
4937 pr_info(" read events: %ld\n", total_read);
4938 pr_info(" lost events: %ld\n", total_lost);
4939 pr_info(" total events: %ld\n", total_lost + total_read);
4940 pr_info(" recorded len bytes: %ld\n", total_len);
4941 pr_info(" recorded size bytes: %ld\n", total_size);
4942 if (total_lost)
4943 pr_info(" With dropped events, record len and size may not match\n"
4944 " alloced and written from above\n");
4945 if (!total_lost) {
4946 if (RB_WARN_ON(buffer, total_len != total_alloc ||
4947 total_size != total_written))
4948 break;
4949 }
4950 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
4951 break;
4952
4953 ret = 0;
4954 }
4955 if (!ret)
4956 pr_info("Ring buffer PASSED!\n");
4957
4958 ring_buffer_free(buffer);
4959 return 0;
4960}
4961
4962late_initcall(test_ringbuffer);
4963#endif /* CONFIG_RING_BUFFER_STARTUP_TEST */