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