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