rcu: Make non-preemptive schedule be Tasks RCU quiescent state
[linux/fpc-iii.git] / kernel / trace / ring_buffer.c
blob96fc3c043ad654e0c2de8a9d65a38695788b46b2
1 /*
2 * Generic ring buffer
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/sched/clock.h>
10 #include <linux/trace_seq.h>
11 #include <linux/spinlock.h>
12 #include <linux/irq_work.h>
13 #include <linux/uaccess.h>
14 #include <linux/hardirq.h>
15 #include <linux/kthread.h> /* for self test */
16 #include <linux/kmemcheck.h>
17 #include <linux/module.h>
18 #include <linux/percpu.h>
19 #include <linux/mutex.h>
20 #include <linux/delay.h>
21 #include <linux/slab.h>
22 #include <linux/init.h>
23 #include <linux/hash.h>
24 #include <linux/list.h>
25 #include <linux/cpu.h>
27 #include <asm/local.h>
29 static void update_pages_handler(struct work_struct *work);
32 * The ring buffer header is special. We must manually up keep it.
34 int ring_buffer_print_entry_header(struct trace_seq *s)
36 trace_seq_puts(s, "# compressed entry header\n");
37 trace_seq_puts(s, "\ttype_len : 5 bits\n");
38 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
39 trace_seq_puts(s, "\tarray : 32 bits\n");
40 trace_seq_putc(s, '\n');
41 trace_seq_printf(s, "\tpadding : type == %d\n",
42 RINGBUF_TYPE_PADDING);
43 trace_seq_printf(s, "\ttime_extend : type == %d\n",
44 RINGBUF_TYPE_TIME_EXTEND);
45 trace_seq_printf(s, "\tdata max type_len == %d\n",
46 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
48 return !trace_seq_has_overflowed(s);
52 * The ring buffer is made up of a list of pages. A separate list of pages is
53 * allocated for each CPU. A writer may only write to a buffer that is
54 * associated with the CPU it is currently executing on. A reader may read
55 * from any per cpu buffer.
57 * The reader is special. For each per cpu buffer, the reader has its own
58 * reader page. When a reader has read the entire reader page, this reader
59 * page is swapped with another page in the ring buffer.
61 * Now, as long as the writer is off the reader page, the reader can do what
62 * ever it wants with that page. The writer will never write to that page
63 * again (as long as it is out of the ring buffer).
65 * Here's some silly ASCII art.
67 * +------+
68 * |reader| RING BUFFER
69 * |page |
70 * +------+ +---+ +---+ +---+
71 * | |-->| |-->| |
72 * +---+ +---+ +---+
73 * ^ |
74 * | |
75 * +---------------+
78 * +------+
79 * |reader| RING BUFFER
80 * |page |------------------v
81 * +------+ +---+ +---+ +---+
82 * | |-->| |-->| |
83 * +---+ +---+ +---+
84 * ^ |
85 * | |
86 * +---------------+
89 * +------+
90 * |reader| RING BUFFER
91 * |page |------------------v
92 * +------+ +---+ +---+ +---+
93 * ^ | |-->| |-->| |
94 * | +---+ +---+ +---+
95 * | |
96 * | |
97 * +------------------------------+
100 * +------+
101 * |buffer| RING BUFFER
102 * |page |------------------v
103 * +------+ +---+ +---+ +---+
104 * ^ | | | |-->| |
105 * | New +---+ +---+ +---+
106 * | Reader------^ |
107 * | page |
108 * +------------------------------+
111 * After we make this swap, the reader can hand this page off to the splice
112 * code and be done with it. It can even allocate a new page if it needs to
113 * and swap that into the ring buffer.
115 * We will be using cmpxchg soon to make all this lockless.
119 /* Used for individual buffers (after the counter) */
120 #define RB_BUFFER_OFF (1 << 20)
122 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
124 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
125 #define RB_ALIGNMENT 4U
126 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
127 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
129 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
130 # define RB_FORCE_8BYTE_ALIGNMENT 0
131 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
132 #else
133 # define RB_FORCE_8BYTE_ALIGNMENT 1
134 # define RB_ARCH_ALIGNMENT 8U
135 #endif
137 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
139 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
140 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
142 enum {
143 RB_LEN_TIME_EXTEND = 8,
144 RB_LEN_TIME_STAMP = 16,
147 #define skip_time_extend(event) \
148 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
150 static inline int rb_null_event(struct ring_buffer_event *event)
152 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
155 static void rb_event_set_padding(struct ring_buffer_event *event)
157 /* padding has a NULL time_delta */
158 event->type_len = RINGBUF_TYPE_PADDING;
159 event->time_delta = 0;
162 static unsigned
163 rb_event_data_length(struct ring_buffer_event *event)
165 unsigned length;
167 if (event->type_len)
168 length = event->type_len * RB_ALIGNMENT;
169 else
170 length = event->array[0];
171 return length + RB_EVNT_HDR_SIZE;
175 * Return the length of the given event. Will return
176 * the length of the time extend if the event is a
177 * time extend.
179 static inline unsigned
180 rb_event_length(struct ring_buffer_event *event)
182 switch (event->type_len) {
183 case RINGBUF_TYPE_PADDING:
184 if (rb_null_event(event))
185 /* undefined */
186 return -1;
187 return event->array[0] + RB_EVNT_HDR_SIZE;
189 case RINGBUF_TYPE_TIME_EXTEND:
190 return RB_LEN_TIME_EXTEND;
192 case RINGBUF_TYPE_TIME_STAMP:
193 return RB_LEN_TIME_STAMP;
195 case RINGBUF_TYPE_DATA:
196 return rb_event_data_length(event);
197 default:
198 BUG();
200 /* not hit */
201 return 0;
205 * Return total length of time extend and data,
206 * or just the event length for all other events.
208 static inline unsigned
209 rb_event_ts_length(struct ring_buffer_event *event)
211 unsigned len = 0;
213 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
214 /* time extends include the data event after it */
215 len = RB_LEN_TIME_EXTEND;
216 event = skip_time_extend(event);
218 return len + rb_event_length(event);
222 * ring_buffer_event_length - return the length of the event
223 * @event: the event to get the length of
225 * Returns the size of the data load of a data event.
226 * If the event is something other than a data event, it
227 * returns the size of the event itself. With the exception
228 * of a TIME EXTEND, where it still returns the size of the
229 * data load of the data event after it.
231 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
233 unsigned length;
235 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
236 event = skip_time_extend(event);
238 length = rb_event_length(event);
239 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
240 return length;
241 length -= RB_EVNT_HDR_SIZE;
242 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
243 length -= sizeof(event->array[0]);
244 return length;
246 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
248 /* inline for ring buffer fast paths */
249 static __always_inline void *
250 rb_event_data(struct ring_buffer_event *event)
252 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
253 event = skip_time_extend(event);
254 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
255 /* If length is in len field, then array[0] has the data */
256 if (event->type_len)
257 return (void *)&event->array[0];
258 /* Otherwise length is in array[0] and array[1] has the data */
259 return (void *)&event->array[1];
263 * ring_buffer_event_data - return the data of the event
264 * @event: the event to get the data from
266 void *ring_buffer_event_data(struct ring_buffer_event *event)
268 return rb_event_data(event);
270 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
272 #define for_each_buffer_cpu(buffer, cpu) \
273 for_each_cpu(cpu, buffer->cpumask)
275 #define TS_SHIFT 27
276 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
277 #define TS_DELTA_TEST (~TS_MASK)
279 /* Flag when events were overwritten */
280 #define RB_MISSED_EVENTS (1 << 31)
281 /* Missed count stored at end */
282 #define RB_MISSED_STORED (1 << 30)
284 struct buffer_data_page {
285 u64 time_stamp; /* page time stamp */
286 local_t commit; /* write committed index */
287 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
291 * Note, the buffer_page list must be first. The buffer pages
292 * are allocated in cache lines, which means that each buffer
293 * page will be at the beginning of a cache line, and thus
294 * the least significant bits will be zero. We use this to
295 * add flags in the list struct pointers, to make the ring buffer
296 * lockless.
298 struct buffer_page {
299 struct list_head list; /* list of buffer pages */
300 local_t write; /* index for next write */
301 unsigned read; /* index for next read */
302 local_t entries; /* entries on this page */
303 unsigned long real_end; /* real end of data */
304 struct buffer_data_page *page; /* Actual data page */
308 * The buffer page counters, write and entries, must be reset
309 * atomically when crossing page boundaries. To synchronize this
310 * update, two counters are inserted into the number. One is
311 * the actual counter for the write position or count on the page.
313 * The other is a counter of updaters. Before an update happens
314 * the update partition of the counter is incremented. This will
315 * allow the updater to update the counter atomically.
317 * The counter is 20 bits, and the state data is 12.
319 #define RB_WRITE_MASK 0xfffff
320 #define RB_WRITE_INTCNT (1 << 20)
322 static void rb_init_page(struct buffer_data_page *bpage)
324 local_set(&bpage->commit, 0);
328 * ring_buffer_page_len - the size of data on the page.
329 * @page: The page to read
331 * Returns the amount of data on the page, including buffer page header.
333 size_t ring_buffer_page_len(void *page)
335 return local_read(&((struct buffer_data_page *)page)->commit)
336 + BUF_PAGE_HDR_SIZE;
340 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
341 * this issue out.
343 static void free_buffer_page(struct buffer_page *bpage)
345 free_page((unsigned long)bpage->page);
346 kfree(bpage);
350 * We need to fit the time_stamp delta into 27 bits.
352 static inline int test_time_stamp(u64 delta)
354 if (delta & TS_DELTA_TEST)
355 return 1;
356 return 0;
359 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
361 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
362 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
364 int ring_buffer_print_page_header(struct trace_seq *s)
366 struct buffer_data_page field;
368 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
369 "offset:0;\tsize:%u;\tsigned:%u;\n",
370 (unsigned int)sizeof(field.time_stamp),
371 (unsigned int)is_signed_type(u64));
373 trace_seq_printf(s, "\tfield: local_t commit;\t"
374 "offset:%u;\tsize:%u;\tsigned:%u;\n",
375 (unsigned int)offsetof(typeof(field), commit),
376 (unsigned int)sizeof(field.commit),
377 (unsigned int)is_signed_type(long));
379 trace_seq_printf(s, "\tfield: int overwrite;\t"
380 "offset:%u;\tsize:%u;\tsigned:%u;\n",
381 (unsigned int)offsetof(typeof(field), commit),
383 (unsigned int)is_signed_type(long));
385 trace_seq_printf(s, "\tfield: char data;\t"
386 "offset:%u;\tsize:%u;\tsigned:%u;\n",
387 (unsigned int)offsetof(typeof(field), data),
388 (unsigned int)BUF_PAGE_SIZE,
389 (unsigned int)is_signed_type(char));
391 return !trace_seq_has_overflowed(s);
394 struct rb_irq_work {
395 struct irq_work work;
396 wait_queue_head_t waiters;
397 wait_queue_head_t full_waiters;
398 bool waiters_pending;
399 bool full_waiters_pending;
400 bool wakeup_full;
404 * Structure to hold event state and handle nested events.
406 struct rb_event_info {
407 u64 ts;
408 u64 delta;
409 unsigned long length;
410 struct buffer_page *tail_page;
411 int add_timestamp;
415 * Used for which event context the event is in.
416 * NMI = 0
417 * IRQ = 1
418 * SOFTIRQ = 2
419 * NORMAL = 3
421 * See trace_recursive_lock() comment below for more details.
423 enum {
424 RB_CTX_NMI,
425 RB_CTX_IRQ,
426 RB_CTX_SOFTIRQ,
427 RB_CTX_NORMAL,
428 RB_CTX_MAX
432 * head_page == tail_page && head == tail then buffer is empty.
434 struct ring_buffer_per_cpu {
435 int cpu;
436 atomic_t record_disabled;
437 struct ring_buffer *buffer;
438 raw_spinlock_t reader_lock; /* serialize readers */
439 arch_spinlock_t lock;
440 struct lock_class_key lock_key;
441 unsigned long nr_pages;
442 unsigned int current_context;
443 struct list_head *pages;
444 struct buffer_page *head_page; /* read from head */
445 struct buffer_page *tail_page; /* write to tail */
446 struct buffer_page *commit_page; /* committed pages */
447 struct buffer_page *reader_page;
448 unsigned long lost_events;
449 unsigned long last_overrun;
450 local_t entries_bytes;
451 local_t entries;
452 local_t overrun;
453 local_t commit_overrun;
454 local_t dropped_events;
455 local_t committing;
456 local_t commits;
457 unsigned long read;
458 unsigned long read_bytes;
459 u64 write_stamp;
460 u64 read_stamp;
461 /* ring buffer pages to update, > 0 to add, < 0 to remove */
462 long nr_pages_to_update;
463 struct list_head new_pages; /* new pages to add */
464 struct work_struct update_pages_work;
465 struct completion update_done;
467 struct rb_irq_work irq_work;
470 struct ring_buffer {
471 unsigned flags;
472 int cpus;
473 atomic_t record_disabled;
474 atomic_t resize_disabled;
475 cpumask_var_t cpumask;
477 struct lock_class_key *reader_lock_key;
479 struct mutex mutex;
481 struct ring_buffer_per_cpu **buffers;
483 struct hlist_node node;
484 u64 (*clock)(void);
486 struct rb_irq_work irq_work;
489 struct ring_buffer_iter {
490 struct ring_buffer_per_cpu *cpu_buffer;
491 unsigned long head;
492 struct buffer_page *head_page;
493 struct buffer_page *cache_reader_page;
494 unsigned long cache_read;
495 u64 read_stamp;
499 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
501 * Schedules a delayed work to wake up any task that is blocked on the
502 * ring buffer waiters queue.
504 static void rb_wake_up_waiters(struct irq_work *work)
506 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
508 wake_up_all(&rbwork->waiters);
509 if (rbwork->wakeup_full) {
510 rbwork->wakeup_full = false;
511 wake_up_all(&rbwork->full_waiters);
516 * ring_buffer_wait - wait for input to the ring buffer
517 * @buffer: buffer to wait on
518 * @cpu: the cpu buffer to wait on
519 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
521 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
522 * as data is added to any of the @buffer's cpu buffers. Otherwise
523 * it will wait for data to be added to a specific cpu buffer.
525 int ring_buffer_wait(struct ring_buffer *buffer, int cpu, bool full)
527 struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
528 DEFINE_WAIT(wait);
529 struct rb_irq_work *work;
530 int ret = 0;
533 * Depending on what the caller is waiting for, either any
534 * data in any cpu buffer, or a specific buffer, put the
535 * caller on the appropriate wait queue.
537 if (cpu == RING_BUFFER_ALL_CPUS) {
538 work = &buffer->irq_work;
539 /* Full only makes sense on per cpu reads */
540 full = false;
541 } else {
542 if (!cpumask_test_cpu(cpu, buffer->cpumask))
543 return -ENODEV;
544 cpu_buffer = buffer->buffers[cpu];
545 work = &cpu_buffer->irq_work;
549 while (true) {
550 if (full)
551 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
552 else
553 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
556 * The events can happen in critical sections where
557 * checking a work queue can cause deadlocks.
558 * After adding a task to the queue, this flag is set
559 * only to notify events to try to wake up the queue
560 * using irq_work.
562 * We don't clear it even if the buffer is no longer
563 * empty. The flag only causes the next event to run
564 * irq_work to do the work queue wake up. The worse
565 * that can happen if we race with !trace_empty() is that
566 * an event will cause an irq_work to try to wake up
567 * an empty queue.
569 * There's no reason to protect this flag either, as
570 * the work queue and irq_work logic will do the necessary
571 * synchronization for the wake ups. The only thing
572 * that is necessary is that the wake up happens after
573 * a task has been queued. It's OK for spurious wake ups.
575 if (full)
576 work->full_waiters_pending = true;
577 else
578 work->waiters_pending = true;
580 if (signal_pending(current)) {
581 ret = -EINTR;
582 break;
585 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
586 break;
588 if (cpu != RING_BUFFER_ALL_CPUS &&
589 !ring_buffer_empty_cpu(buffer, cpu)) {
590 unsigned long flags;
591 bool pagebusy;
593 if (!full)
594 break;
596 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
597 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
598 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
600 if (!pagebusy)
601 break;
604 schedule();
607 if (full)
608 finish_wait(&work->full_waiters, &wait);
609 else
610 finish_wait(&work->waiters, &wait);
612 return ret;
616 * ring_buffer_poll_wait - poll on buffer input
617 * @buffer: buffer to wait on
618 * @cpu: the cpu buffer to wait on
619 * @filp: the file descriptor
620 * @poll_table: The poll descriptor
622 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
623 * as data is added to any of the @buffer's cpu buffers. Otherwise
624 * it will wait for data to be added to a specific cpu buffer.
626 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
627 * zero otherwise.
629 int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
630 struct file *filp, poll_table *poll_table)
632 struct ring_buffer_per_cpu *cpu_buffer;
633 struct rb_irq_work *work;
635 if (cpu == RING_BUFFER_ALL_CPUS)
636 work = &buffer->irq_work;
637 else {
638 if (!cpumask_test_cpu(cpu, buffer->cpumask))
639 return -EINVAL;
641 cpu_buffer = buffer->buffers[cpu];
642 work = &cpu_buffer->irq_work;
645 poll_wait(filp, &work->waiters, poll_table);
646 work->waiters_pending = true;
648 * There's a tight race between setting the waiters_pending and
649 * checking if the ring buffer is empty. Once the waiters_pending bit
650 * is set, the next event will wake the task up, but we can get stuck
651 * if there's only a single event in.
653 * FIXME: Ideally, we need a memory barrier on the writer side as well,
654 * but adding a memory barrier to all events will cause too much of a
655 * performance hit in the fast path. We only need a memory barrier when
656 * the buffer goes from empty to having content. But as this race is
657 * extremely small, and it's not a problem if another event comes in, we
658 * will fix it later.
660 smp_mb();
662 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
663 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
664 return POLLIN | POLLRDNORM;
665 return 0;
668 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
669 #define RB_WARN_ON(b, cond) \
670 ({ \
671 int _____ret = unlikely(cond); \
672 if (_____ret) { \
673 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
674 struct ring_buffer_per_cpu *__b = \
675 (void *)b; \
676 atomic_inc(&__b->buffer->record_disabled); \
677 } else \
678 atomic_inc(&b->record_disabled); \
679 WARN_ON(1); \
681 _____ret; \
684 /* Up this if you want to test the TIME_EXTENTS and normalization */
685 #define DEBUG_SHIFT 0
687 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
689 /* shift to debug/test normalization and TIME_EXTENTS */
690 return buffer->clock() << DEBUG_SHIFT;
693 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
695 u64 time;
697 preempt_disable_notrace();
698 time = rb_time_stamp(buffer);
699 preempt_enable_no_resched_notrace();
701 return time;
703 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
705 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
706 int cpu, u64 *ts)
708 /* Just stupid testing the normalize function and deltas */
709 *ts >>= DEBUG_SHIFT;
711 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
714 * Making the ring buffer lockless makes things tricky.
715 * Although writes only happen on the CPU that they are on,
716 * and they only need to worry about interrupts. Reads can
717 * happen on any CPU.
719 * The reader page is always off the ring buffer, but when the
720 * reader finishes with a page, it needs to swap its page with
721 * a new one from the buffer. The reader needs to take from
722 * the head (writes go to the tail). But if a writer is in overwrite
723 * mode and wraps, it must push the head page forward.
725 * Here lies the problem.
727 * The reader must be careful to replace only the head page, and
728 * not another one. As described at the top of the file in the
729 * ASCII art, the reader sets its old page to point to the next
730 * page after head. It then sets the page after head to point to
731 * the old reader page. But if the writer moves the head page
732 * during this operation, the reader could end up with the tail.
734 * We use cmpxchg to help prevent this race. We also do something
735 * special with the page before head. We set the LSB to 1.
737 * When the writer must push the page forward, it will clear the
738 * bit that points to the head page, move the head, and then set
739 * the bit that points to the new head page.
741 * We also don't want an interrupt coming in and moving the head
742 * page on another writer. Thus we use the second LSB to catch
743 * that too. Thus:
745 * head->list->prev->next bit 1 bit 0
746 * ------- -------
747 * Normal page 0 0
748 * Points to head page 0 1
749 * New head page 1 0
751 * Note we can not trust the prev pointer of the head page, because:
753 * +----+ +-----+ +-----+
754 * | |------>| T |---X--->| N |
755 * | |<------| | | |
756 * +----+ +-----+ +-----+
757 * ^ ^ |
758 * | +-----+ | |
759 * +----------| R |----------+ |
760 * | |<-----------+
761 * +-----+
763 * Key: ---X--> HEAD flag set in pointer
764 * T Tail page
765 * R Reader page
766 * N Next page
768 * (see __rb_reserve_next() to see where this happens)
770 * What the above shows is that the reader just swapped out
771 * the reader page with a page in the buffer, but before it
772 * could make the new header point back to the new page added
773 * it was preempted by a writer. The writer moved forward onto
774 * the new page added by the reader and is about to move forward
775 * again.
777 * You can see, it is legitimate for the previous pointer of
778 * the head (or any page) not to point back to itself. But only
779 * temporarially.
782 #define RB_PAGE_NORMAL 0UL
783 #define RB_PAGE_HEAD 1UL
784 #define RB_PAGE_UPDATE 2UL
787 #define RB_FLAG_MASK 3UL
789 /* PAGE_MOVED is not part of the mask */
790 #define RB_PAGE_MOVED 4UL
793 * rb_list_head - remove any bit
795 static struct list_head *rb_list_head(struct list_head *list)
797 unsigned long val = (unsigned long)list;
799 return (struct list_head *)(val & ~RB_FLAG_MASK);
803 * rb_is_head_page - test if the given page is the head page
805 * Because the reader may move the head_page pointer, we can
806 * not trust what the head page is (it may be pointing to
807 * the reader page). But if the next page is a header page,
808 * its flags will be non zero.
810 static inline int
811 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
812 struct buffer_page *page, struct list_head *list)
814 unsigned long val;
816 val = (unsigned long)list->next;
818 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
819 return RB_PAGE_MOVED;
821 return val & RB_FLAG_MASK;
825 * rb_is_reader_page
827 * The unique thing about the reader page, is that, if the
828 * writer is ever on it, the previous pointer never points
829 * back to the reader page.
831 static bool rb_is_reader_page(struct buffer_page *page)
833 struct list_head *list = page->list.prev;
835 return rb_list_head(list->next) != &page->list;
839 * rb_set_list_to_head - set a list_head to be pointing to head.
841 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
842 struct list_head *list)
844 unsigned long *ptr;
846 ptr = (unsigned long *)&list->next;
847 *ptr |= RB_PAGE_HEAD;
848 *ptr &= ~RB_PAGE_UPDATE;
852 * rb_head_page_activate - sets up head page
854 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
856 struct buffer_page *head;
858 head = cpu_buffer->head_page;
859 if (!head)
860 return;
863 * Set the previous list pointer to have the HEAD flag.
865 rb_set_list_to_head(cpu_buffer, head->list.prev);
868 static void rb_list_head_clear(struct list_head *list)
870 unsigned long *ptr = (unsigned long *)&list->next;
872 *ptr &= ~RB_FLAG_MASK;
876 * rb_head_page_dactivate - clears head page ptr (for free list)
878 static void
879 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
881 struct list_head *hd;
883 /* Go through the whole list and clear any pointers found. */
884 rb_list_head_clear(cpu_buffer->pages);
886 list_for_each(hd, cpu_buffer->pages)
887 rb_list_head_clear(hd);
890 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
891 struct buffer_page *head,
892 struct buffer_page *prev,
893 int old_flag, int new_flag)
895 struct list_head *list;
896 unsigned long val = (unsigned long)&head->list;
897 unsigned long ret;
899 list = &prev->list;
901 val &= ~RB_FLAG_MASK;
903 ret = cmpxchg((unsigned long *)&list->next,
904 val | old_flag, val | new_flag);
906 /* check if the reader took the page */
907 if ((ret & ~RB_FLAG_MASK) != val)
908 return RB_PAGE_MOVED;
910 return ret & RB_FLAG_MASK;
913 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
914 struct buffer_page *head,
915 struct buffer_page *prev,
916 int old_flag)
918 return rb_head_page_set(cpu_buffer, head, prev,
919 old_flag, RB_PAGE_UPDATE);
922 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
923 struct buffer_page *head,
924 struct buffer_page *prev,
925 int old_flag)
927 return rb_head_page_set(cpu_buffer, head, prev,
928 old_flag, RB_PAGE_HEAD);
931 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
932 struct buffer_page *head,
933 struct buffer_page *prev,
934 int old_flag)
936 return rb_head_page_set(cpu_buffer, head, prev,
937 old_flag, RB_PAGE_NORMAL);
940 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
941 struct buffer_page **bpage)
943 struct list_head *p = rb_list_head((*bpage)->list.next);
945 *bpage = list_entry(p, struct buffer_page, list);
948 static struct buffer_page *
949 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
951 struct buffer_page *head;
952 struct buffer_page *page;
953 struct list_head *list;
954 int i;
956 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
957 return NULL;
959 /* sanity check */
960 list = cpu_buffer->pages;
961 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
962 return NULL;
964 page = head = cpu_buffer->head_page;
966 * It is possible that the writer moves the header behind
967 * where we started, and we miss in one loop.
968 * A second loop should grab the header, but we'll do
969 * three loops just because I'm paranoid.
971 for (i = 0; i < 3; i++) {
972 do {
973 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
974 cpu_buffer->head_page = page;
975 return page;
977 rb_inc_page(cpu_buffer, &page);
978 } while (page != head);
981 RB_WARN_ON(cpu_buffer, 1);
983 return NULL;
986 static int rb_head_page_replace(struct buffer_page *old,
987 struct buffer_page *new)
989 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
990 unsigned long val;
991 unsigned long ret;
993 val = *ptr & ~RB_FLAG_MASK;
994 val |= RB_PAGE_HEAD;
996 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
998 return ret == val;
1002 * rb_tail_page_update - move the tail page forward
1004 static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1005 struct buffer_page *tail_page,
1006 struct buffer_page *next_page)
1008 unsigned long old_entries;
1009 unsigned long old_write;
1012 * The tail page now needs to be moved forward.
1014 * We need to reset the tail page, but without messing
1015 * with possible erasing of data brought in by interrupts
1016 * that have moved the tail page and are currently on it.
1018 * We add a counter to the write field to denote this.
1020 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1021 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1024 * Just make sure we have seen our old_write and synchronize
1025 * with any interrupts that come in.
1027 barrier();
1030 * If the tail page is still the same as what we think
1031 * it is, then it is up to us to update the tail
1032 * pointer.
1034 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1035 /* Zero the write counter */
1036 unsigned long val = old_write & ~RB_WRITE_MASK;
1037 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1040 * This will only succeed if an interrupt did
1041 * not come in and change it. In which case, we
1042 * do not want to modify it.
1044 * We add (void) to let the compiler know that we do not care
1045 * about the return value of these functions. We use the
1046 * cmpxchg to only update if an interrupt did not already
1047 * do it for us. If the cmpxchg fails, we don't care.
1049 (void)local_cmpxchg(&next_page->write, old_write, val);
1050 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1053 * No need to worry about races with clearing out the commit.
1054 * it only can increment when a commit takes place. But that
1055 * only happens in the outer most nested commit.
1057 local_set(&next_page->page->commit, 0);
1059 /* Again, either we update tail_page or an interrupt does */
1060 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1064 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1065 struct buffer_page *bpage)
1067 unsigned long val = (unsigned long)bpage;
1069 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1070 return 1;
1072 return 0;
1076 * rb_check_list - make sure a pointer to a list has the last bits zero
1078 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1079 struct list_head *list)
1081 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1082 return 1;
1083 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1084 return 1;
1085 return 0;
1089 * rb_check_pages - integrity check of buffer pages
1090 * @cpu_buffer: CPU buffer with pages to test
1092 * As a safety measure we check to make sure the data pages have not
1093 * been corrupted.
1095 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1097 struct list_head *head = cpu_buffer->pages;
1098 struct buffer_page *bpage, *tmp;
1100 /* Reset the head page if it exists */
1101 if (cpu_buffer->head_page)
1102 rb_set_head_page(cpu_buffer);
1104 rb_head_page_deactivate(cpu_buffer);
1106 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1107 return -1;
1108 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1109 return -1;
1111 if (rb_check_list(cpu_buffer, head))
1112 return -1;
1114 list_for_each_entry_safe(bpage, tmp, head, list) {
1115 if (RB_WARN_ON(cpu_buffer,
1116 bpage->list.next->prev != &bpage->list))
1117 return -1;
1118 if (RB_WARN_ON(cpu_buffer,
1119 bpage->list.prev->next != &bpage->list))
1120 return -1;
1121 if (rb_check_list(cpu_buffer, &bpage->list))
1122 return -1;
1125 rb_head_page_activate(cpu_buffer);
1127 return 0;
1130 static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1132 struct buffer_page *bpage, *tmp;
1133 long i;
1135 for (i = 0; i < nr_pages; i++) {
1136 struct page *page;
1138 * __GFP_NORETRY flag makes sure that the allocation fails
1139 * gracefully without invoking oom-killer and the system is
1140 * not destabilized.
1142 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1143 GFP_KERNEL | __GFP_NORETRY,
1144 cpu_to_node(cpu));
1145 if (!bpage)
1146 goto free_pages;
1148 list_add(&bpage->list, pages);
1150 page = alloc_pages_node(cpu_to_node(cpu),
1151 GFP_KERNEL | __GFP_NORETRY, 0);
1152 if (!page)
1153 goto free_pages;
1154 bpage->page = page_address(page);
1155 rb_init_page(bpage->page);
1158 return 0;
1160 free_pages:
1161 list_for_each_entry_safe(bpage, tmp, pages, list) {
1162 list_del_init(&bpage->list);
1163 free_buffer_page(bpage);
1166 return -ENOMEM;
1169 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1170 unsigned long nr_pages)
1172 LIST_HEAD(pages);
1174 WARN_ON(!nr_pages);
1176 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1177 return -ENOMEM;
1180 * The ring buffer page list is a circular list that does not
1181 * start and end with a list head. All page list items point to
1182 * other pages.
1184 cpu_buffer->pages = pages.next;
1185 list_del(&pages);
1187 cpu_buffer->nr_pages = nr_pages;
1189 rb_check_pages(cpu_buffer);
1191 return 0;
1194 static struct ring_buffer_per_cpu *
1195 rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1197 struct ring_buffer_per_cpu *cpu_buffer;
1198 struct buffer_page *bpage;
1199 struct page *page;
1200 int ret;
1202 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1203 GFP_KERNEL, cpu_to_node(cpu));
1204 if (!cpu_buffer)
1205 return NULL;
1207 cpu_buffer->cpu = cpu;
1208 cpu_buffer->buffer = buffer;
1209 raw_spin_lock_init(&cpu_buffer->reader_lock);
1210 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1211 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1212 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1213 init_completion(&cpu_buffer->update_done);
1214 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1215 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1216 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1218 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1219 GFP_KERNEL, cpu_to_node(cpu));
1220 if (!bpage)
1221 goto fail_free_buffer;
1223 rb_check_bpage(cpu_buffer, bpage);
1225 cpu_buffer->reader_page = bpage;
1226 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1227 if (!page)
1228 goto fail_free_reader;
1229 bpage->page = page_address(page);
1230 rb_init_page(bpage->page);
1232 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1233 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1235 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1236 if (ret < 0)
1237 goto fail_free_reader;
1239 cpu_buffer->head_page
1240 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1241 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1243 rb_head_page_activate(cpu_buffer);
1245 return cpu_buffer;
1247 fail_free_reader:
1248 free_buffer_page(cpu_buffer->reader_page);
1250 fail_free_buffer:
1251 kfree(cpu_buffer);
1252 return NULL;
1255 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1257 struct list_head *head = cpu_buffer->pages;
1258 struct buffer_page *bpage, *tmp;
1260 free_buffer_page(cpu_buffer->reader_page);
1262 rb_head_page_deactivate(cpu_buffer);
1264 if (head) {
1265 list_for_each_entry_safe(bpage, tmp, head, list) {
1266 list_del_init(&bpage->list);
1267 free_buffer_page(bpage);
1269 bpage = list_entry(head, struct buffer_page, list);
1270 free_buffer_page(bpage);
1273 kfree(cpu_buffer);
1277 * __ring_buffer_alloc - allocate a new ring_buffer
1278 * @size: the size in bytes per cpu that is needed.
1279 * @flags: attributes to set for the ring buffer.
1281 * Currently the only flag that is available is the RB_FL_OVERWRITE
1282 * flag. This flag means that the buffer will overwrite old data
1283 * when the buffer wraps. If this flag is not set, the buffer will
1284 * drop data when the tail hits the head.
1286 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1287 struct lock_class_key *key)
1289 struct ring_buffer *buffer;
1290 long nr_pages;
1291 int bsize;
1292 int cpu;
1293 int ret;
1295 /* keep it in its own cache line */
1296 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1297 GFP_KERNEL);
1298 if (!buffer)
1299 return NULL;
1301 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1302 goto fail_free_buffer;
1304 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1305 buffer->flags = flags;
1306 buffer->clock = trace_clock_local;
1307 buffer->reader_lock_key = key;
1309 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1310 init_waitqueue_head(&buffer->irq_work.waiters);
1312 /* need at least two pages */
1313 if (nr_pages < 2)
1314 nr_pages = 2;
1316 buffer->cpus = nr_cpu_ids;
1318 bsize = sizeof(void *) * nr_cpu_ids;
1319 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1320 GFP_KERNEL);
1321 if (!buffer->buffers)
1322 goto fail_free_cpumask;
1324 cpu = raw_smp_processor_id();
1325 cpumask_set_cpu(cpu, buffer->cpumask);
1326 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1327 if (!buffer->buffers[cpu])
1328 goto fail_free_buffers;
1330 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1331 if (ret < 0)
1332 goto fail_free_buffers;
1334 mutex_init(&buffer->mutex);
1336 return buffer;
1338 fail_free_buffers:
1339 for_each_buffer_cpu(buffer, cpu) {
1340 if (buffer->buffers[cpu])
1341 rb_free_cpu_buffer(buffer->buffers[cpu]);
1343 kfree(buffer->buffers);
1345 fail_free_cpumask:
1346 free_cpumask_var(buffer->cpumask);
1348 fail_free_buffer:
1349 kfree(buffer);
1350 return NULL;
1352 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1355 * ring_buffer_free - free a ring buffer.
1356 * @buffer: the buffer to free.
1358 void
1359 ring_buffer_free(struct ring_buffer *buffer)
1361 int cpu;
1363 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1365 for_each_buffer_cpu(buffer, cpu)
1366 rb_free_cpu_buffer(buffer->buffers[cpu]);
1368 kfree(buffer->buffers);
1369 free_cpumask_var(buffer->cpumask);
1371 kfree(buffer);
1373 EXPORT_SYMBOL_GPL(ring_buffer_free);
1375 void ring_buffer_set_clock(struct ring_buffer *buffer,
1376 u64 (*clock)(void))
1378 buffer->clock = clock;
1381 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1383 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1385 return local_read(&bpage->entries) & RB_WRITE_MASK;
1388 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1390 return local_read(&bpage->write) & RB_WRITE_MASK;
1393 static int
1394 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1396 struct list_head *tail_page, *to_remove, *next_page;
1397 struct buffer_page *to_remove_page, *tmp_iter_page;
1398 struct buffer_page *last_page, *first_page;
1399 unsigned long nr_removed;
1400 unsigned long head_bit;
1401 int page_entries;
1403 head_bit = 0;
1405 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1406 atomic_inc(&cpu_buffer->record_disabled);
1408 * We don't race with the readers since we have acquired the reader
1409 * lock. We also don't race with writers after disabling recording.
1410 * This makes it easy to figure out the first and the last page to be
1411 * removed from the list. We unlink all the pages in between including
1412 * the first and last pages. This is done in a busy loop so that we
1413 * lose the least number of traces.
1414 * The pages are freed after we restart recording and unlock readers.
1416 tail_page = &cpu_buffer->tail_page->list;
1419 * tail page might be on reader page, we remove the next page
1420 * from the ring buffer
1422 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1423 tail_page = rb_list_head(tail_page->next);
1424 to_remove = tail_page;
1426 /* start of pages to remove */
1427 first_page = list_entry(rb_list_head(to_remove->next),
1428 struct buffer_page, list);
1430 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1431 to_remove = rb_list_head(to_remove)->next;
1432 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1435 next_page = rb_list_head(to_remove)->next;
1438 * Now we remove all pages between tail_page and next_page.
1439 * Make sure that we have head_bit value preserved for the
1440 * next page
1442 tail_page->next = (struct list_head *)((unsigned long)next_page |
1443 head_bit);
1444 next_page = rb_list_head(next_page);
1445 next_page->prev = tail_page;
1447 /* make sure pages points to a valid page in the ring buffer */
1448 cpu_buffer->pages = next_page;
1450 /* update head page */
1451 if (head_bit)
1452 cpu_buffer->head_page = list_entry(next_page,
1453 struct buffer_page, list);
1456 * change read pointer to make sure any read iterators reset
1457 * themselves
1459 cpu_buffer->read = 0;
1461 /* pages are removed, resume tracing and then free the pages */
1462 atomic_dec(&cpu_buffer->record_disabled);
1463 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1465 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1467 /* last buffer page to remove */
1468 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1469 list);
1470 tmp_iter_page = first_page;
1472 do {
1473 to_remove_page = tmp_iter_page;
1474 rb_inc_page(cpu_buffer, &tmp_iter_page);
1476 /* update the counters */
1477 page_entries = rb_page_entries(to_remove_page);
1478 if (page_entries) {
1480 * If something was added to this page, it was full
1481 * since it is not the tail page. So we deduct the
1482 * bytes consumed in ring buffer from here.
1483 * Increment overrun to account for the lost events.
1485 local_add(page_entries, &cpu_buffer->overrun);
1486 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1490 * We have already removed references to this list item, just
1491 * free up the buffer_page and its page
1493 free_buffer_page(to_remove_page);
1494 nr_removed--;
1496 } while (to_remove_page != last_page);
1498 RB_WARN_ON(cpu_buffer, nr_removed);
1500 return nr_removed == 0;
1503 static int
1504 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1506 struct list_head *pages = &cpu_buffer->new_pages;
1507 int retries, success;
1509 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1511 * We are holding the reader lock, so the reader page won't be swapped
1512 * in the ring buffer. Now we are racing with the writer trying to
1513 * move head page and the tail page.
1514 * We are going to adapt the reader page update process where:
1515 * 1. We first splice the start and end of list of new pages between
1516 * the head page and its previous page.
1517 * 2. We cmpxchg the prev_page->next to point from head page to the
1518 * start of new pages list.
1519 * 3. Finally, we update the head->prev to the end of new list.
1521 * We will try this process 10 times, to make sure that we don't keep
1522 * spinning.
1524 retries = 10;
1525 success = 0;
1526 while (retries--) {
1527 struct list_head *head_page, *prev_page, *r;
1528 struct list_head *last_page, *first_page;
1529 struct list_head *head_page_with_bit;
1531 head_page = &rb_set_head_page(cpu_buffer)->list;
1532 if (!head_page)
1533 break;
1534 prev_page = head_page->prev;
1536 first_page = pages->next;
1537 last_page = pages->prev;
1539 head_page_with_bit = (struct list_head *)
1540 ((unsigned long)head_page | RB_PAGE_HEAD);
1542 last_page->next = head_page_with_bit;
1543 first_page->prev = prev_page;
1545 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1547 if (r == head_page_with_bit) {
1549 * yay, we replaced the page pointer to our new list,
1550 * now, we just have to update to head page's prev
1551 * pointer to point to end of list
1553 head_page->prev = last_page;
1554 success = 1;
1555 break;
1559 if (success)
1560 INIT_LIST_HEAD(pages);
1562 * If we weren't successful in adding in new pages, warn and stop
1563 * tracing
1565 RB_WARN_ON(cpu_buffer, !success);
1566 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1568 /* free pages if they weren't inserted */
1569 if (!success) {
1570 struct buffer_page *bpage, *tmp;
1571 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1572 list) {
1573 list_del_init(&bpage->list);
1574 free_buffer_page(bpage);
1577 return success;
1580 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1582 int success;
1584 if (cpu_buffer->nr_pages_to_update > 0)
1585 success = rb_insert_pages(cpu_buffer);
1586 else
1587 success = rb_remove_pages(cpu_buffer,
1588 -cpu_buffer->nr_pages_to_update);
1590 if (success)
1591 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1594 static void update_pages_handler(struct work_struct *work)
1596 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1597 struct ring_buffer_per_cpu, update_pages_work);
1598 rb_update_pages(cpu_buffer);
1599 complete(&cpu_buffer->update_done);
1603 * ring_buffer_resize - resize the ring buffer
1604 * @buffer: the buffer to resize.
1605 * @size: the new size.
1606 * @cpu_id: the cpu buffer to resize
1608 * Minimum size is 2 * BUF_PAGE_SIZE.
1610 * Returns 0 on success and < 0 on failure.
1612 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1613 int cpu_id)
1615 struct ring_buffer_per_cpu *cpu_buffer;
1616 unsigned long nr_pages;
1617 int cpu, err = 0;
1620 * Always succeed at resizing a non-existent buffer:
1622 if (!buffer)
1623 return size;
1625 /* Make sure the requested buffer exists */
1626 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1627 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1628 return size;
1630 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1632 /* we need a minimum of two pages */
1633 if (nr_pages < 2)
1634 nr_pages = 2;
1636 size = nr_pages * BUF_PAGE_SIZE;
1639 * Don't succeed if resizing is disabled, as a reader might be
1640 * manipulating the ring buffer and is expecting a sane state while
1641 * this is true.
1643 if (atomic_read(&buffer->resize_disabled))
1644 return -EBUSY;
1646 /* prevent another thread from changing buffer sizes */
1647 mutex_lock(&buffer->mutex);
1649 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1650 /* calculate the pages to update */
1651 for_each_buffer_cpu(buffer, cpu) {
1652 cpu_buffer = buffer->buffers[cpu];
1654 cpu_buffer->nr_pages_to_update = nr_pages -
1655 cpu_buffer->nr_pages;
1657 * nothing more to do for removing pages or no update
1659 if (cpu_buffer->nr_pages_to_update <= 0)
1660 continue;
1662 * to add pages, make sure all new pages can be
1663 * allocated without receiving ENOMEM
1665 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1666 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1667 &cpu_buffer->new_pages, cpu)) {
1668 /* not enough memory for new pages */
1669 err = -ENOMEM;
1670 goto out_err;
1674 get_online_cpus();
1676 * Fire off all the required work handlers
1677 * We can't schedule on offline CPUs, but it's not necessary
1678 * since we can change their buffer sizes without any race.
1680 for_each_buffer_cpu(buffer, cpu) {
1681 cpu_buffer = buffer->buffers[cpu];
1682 if (!cpu_buffer->nr_pages_to_update)
1683 continue;
1685 /* Can't run something on an offline CPU. */
1686 if (!cpu_online(cpu)) {
1687 rb_update_pages(cpu_buffer);
1688 cpu_buffer->nr_pages_to_update = 0;
1689 } else {
1690 schedule_work_on(cpu,
1691 &cpu_buffer->update_pages_work);
1695 /* wait for all the updates to complete */
1696 for_each_buffer_cpu(buffer, cpu) {
1697 cpu_buffer = buffer->buffers[cpu];
1698 if (!cpu_buffer->nr_pages_to_update)
1699 continue;
1701 if (cpu_online(cpu))
1702 wait_for_completion(&cpu_buffer->update_done);
1703 cpu_buffer->nr_pages_to_update = 0;
1706 put_online_cpus();
1707 } else {
1708 /* Make sure this CPU has been intitialized */
1709 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1710 goto out;
1712 cpu_buffer = buffer->buffers[cpu_id];
1714 if (nr_pages == cpu_buffer->nr_pages)
1715 goto out;
1717 cpu_buffer->nr_pages_to_update = nr_pages -
1718 cpu_buffer->nr_pages;
1720 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1721 if (cpu_buffer->nr_pages_to_update > 0 &&
1722 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1723 &cpu_buffer->new_pages, cpu_id)) {
1724 err = -ENOMEM;
1725 goto out_err;
1728 get_online_cpus();
1730 /* Can't run something on an offline CPU. */
1731 if (!cpu_online(cpu_id))
1732 rb_update_pages(cpu_buffer);
1733 else {
1734 schedule_work_on(cpu_id,
1735 &cpu_buffer->update_pages_work);
1736 wait_for_completion(&cpu_buffer->update_done);
1739 cpu_buffer->nr_pages_to_update = 0;
1740 put_online_cpus();
1743 out:
1745 * The ring buffer resize can happen with the ring buffer
1746 * enabled, so that the update disturbs the tracing as little
1747 * as possible. But if the buffer is disabled, we do not need
1748 * to worry about that, and we can take the time to verify
1749 * that the buffer is not corrupt.
1751 if (atomic_read(&buffer->record_disabled)) {
1752 atomic_inc(&buffer->record_disabled);
1754 * Even though the buffer was disabled, we must make sure
1755 * that it is truly disabled before calling rb_check_pages.
1756 * There could have been a race between checking
1757 * record_disable and incrementing it.
1759 synchronize_sched();
1760 for_each_buffer_cpu(buffer, cpu) {
1761 cpu_buffer = buffer->buffers[cpu];
1762 rb_check_pages(cpu_buffer);
1764 atomic_dec(&buffer->record_disabled);
1767 mutex_unlock(&buffer->mutex);
1768 return size;
1770 out_err:
1771 for_each_buffer_cpu(buffer, cpu) {
1772 struct buffer_page *bpage, *tmp;
1774 cpu_buffer = buffer->buffers[cpu];
1775 cpu_buffer->nr_pages_to_update = 0;
1777 if (list_empty(&cpu_buffer->new_pages))
1778 continue;
1780 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1781 list) {
1782 list_del_init(&bpage->list);
1783 free_buffer_page(bpage);
1786 mutex_unlock(&buffer->mutex);
1787 return err;
1789 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1791 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1793 mutex_lock(&buffer->mutex);
1794 if (val)
1795 buffer->flags |= RB_FL_OVERWRITE;
1796 else
1797 buffer->flags &= ~RB_FL_OVERWRITE;
1798 mutex_unlock(&buffer->mutex);
1800 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1802 static __always_inline void *
1803 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1805 return bpage->data + index;
1808 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1810 return bpage->page->data + index;
1813 static __always_inline struct ring_buffer_event *
1814 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1816 return __rb_page_index(cpu_buffer->reader_page,
1817 cpu_buffer->reader_page->read);
1820 static __always_inline struct ring_buffer_event *
1821 rb_iter_head_event(struct ring_buffer_iter *iter)
1823 return __rb_page_index(iter->head_page, iter->head);
1826 static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
1828 return local_read(&bpage->page->commit);
1831 /* Size is determined by what has been committed */
1832 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
1834 return rb_page_commit(bpage);
1837 static __always_inline unsigned
1838 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1840 return rb_page_commit(cpu_buffer->commit_page);
1843 static __always_inline unsigned
1844 rb_event_index(struct ring_buffer_event *event)
1846 unsigned long addr = (unsigned long)event;
1848 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1851 static void rb_inc_iter(struct ring_buffer_iter *iter)
1853 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1856 * The iterator could be on the reader page (it starts there).
1857 * But the head could have moved, since the reader was
1858 * found. Check for this case and assign the iterator
1859 * to the head page instead of next.
1861 if (iter->head_page == cpu_buffer->reader_page)
1862 iter->head_page = rb_set_head_page(cpu_buffer);
1863 else
1864 rb_inc_page(cpu_buffer, &iter->head_page);
1866 iter->read_stamp = iter->head_page->page->time_stamp;
1867 iter->head = 0;
1871 * rb_handle_head_page - writer hit the head page
1873 * Returns: +1 to retry page
1874 * 0 to continue
1875 * -1 on error
1877 static int
1878 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1879 struct buffer_page *tail_page,
1880 struct buffer_page *next_page)
1882 struct buffer_page *new_head;
1883 int entries;
1884 int type;
1885 int ret;
1887 entries = rb_page_entries(next_page);
1890 * The hard part is here. We need to move the head
1891 * forward, and protect against both readers on
1892 * other CPUs and writers coming in via interrupts.
1894 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1895 RB_PAGE_HEAD);
1898 * type can be one of four:
1899 * NORMAL - an interrupt already moved it for us
1900 * HEAD - we are the first to get here.
1901 * UPDATE - we are the interrupt interrupting
1902 * a current move.
1903 * MOVED - a reader on another CPU moved the next
1904 * pointer to its reader page. Give up
1905 * and try again.
1908 switch (type) {
1909 case RB_PAGE_HEAD:
1911 * We changed the head to UPDATE, thus
1912 * it is our responsibility to update
1913 * the counters.
1915 local_add(entries, &cpu_buffer->overrun);
1916 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1919 * The entries will be zeroed out when we move the
1920 * tail page.
1923 /* still more to do */
1924 break;
1926 case RB_PAGE_UPDATE:
1928 * This is an interrupt that interrupt the
1929 * previous update. Still more to do.
1931 break;
1932 case RB_PAGE_NORMAL:
1934 * An interrupt came in before the update
1935 * and processed this for us.
1936 * Nothing left to do.
1938 return 1;
1939 case RB_PAGE_MOVED:
1941 * The reader is on another CPU and just did
1942 * a swap with our next_page.
1943 * Try again.
1945 return 1;
1946 default:
1947 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1948 return -1;
1952 * Now that we are here, the old head pointer is
1953 * set to UPDATE. This will keep the reader from
1954 * swapping the head page with the reader page.
1955 * The reader (on another CPU) will spin till
1956 * we are finished.
1958 * We just need to protect against interrupts
1959 * doing the job. We will set the next pointer
1960 * to HEAD. After that, we set the old pointer
1961 * to NORMAL, but only if it was HEAD before.
1962 * otherwise we are an interrupt, and only
1963 * want the outer most commit to reset it.
1965 new_head = next_page;
1966 rb_inc_page(cpu_buffer, &new_head);
1968 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1969 RB_PAGE_NORMAL);
1972 * Valid returns are:
1973 * HEAD - an interrupt came in and already set it.
1974 * NORMAL - One of two things:
1975 * 1) We really set it.
1976 * 2) A bunch of interrupts came in and moved
1977 * the page forward again.
1979 switch (ret) {
1980 case RB_PAGE_HEAD:
1981 case RB_PAGE_NORMAL:
1982 /* OK */
1983 break;
1984 default:
1985 RB_WARN_ON(cpu_buffer, 1);
1986 return -1;
1990 * It is possible that an interrupt came in,
1991 * set the head up, then more interrupts came in
1992 * and moved it again. When we get back here,
1993 * the page would have been set to NORMAL but we
1994 * just set it back to HEAD.
1996 * How do you detect this? Well, if that happened
1997 * the tail page would have moved.
1999 if (ret == RB_PAGE_NORMAL) {
2000 struct buffer_page *buffer_tail_page;
2002 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2004 * If the tail had moved passed next, then we need
2005 * to reset the pointer.
2007 if (buffer_tail_page != tail_page &&
2008 buffer_tail_page != next_page)
2009 rb_head_page_set_normal(cpu_buffer, new_head,
2010 next_page,
2011 RB_PAGE_HEAD);
2015 * If this was the outer most commit (the one that
2016 * changed the original pointer from HEAD to UPDATE),
2017 * then it is up to us to reset it to NORMAL.
2019 if (type == RB_PAGE_HEAD) {
2020 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2021 tail_page,
2022 RB_PAGE_UPDATE);
2023 if (RB_WARN_ON(cpu_buffer,
2024 ret != RB_PAGE_UPDATE))
2025 return -1;
2028 return 0;
2031 static inline void
2032 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2033 unsigned long tail, struct rb_event_info *info)
2035 struct buffer_page *tail_page = info->tail_page;
2036 struct ring_buffer_event *event;
2037 unsigned long length = info->length;
2040 * Only the event that crossed the page boundary
2041 * must fill the old tail_page with padding.
2043 if (tail >= BUF_PAGE_SIZE) {
2045 * If the page was filled, then we still need
2046 * to update the real_end. Reset it to zero
2047 * and the reader will ignore it.
2049 if (tail == BUF_PAGE_SIZE)
2050 tail_page->real_end = 0;
2052 local_sub(length, &tail_page->write);
2053 return;
2056 event = __rb_page_index(tail_page, tail);
2057 kmemcheck_annotate_bitfield(event, bitfield);
2059 /* account for padding bytes */
2060 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2063 * Save the original length to the meta data.
2064 * This will be used by the reader to add lost event
2065 * counter.
2067 tail_page->real_end = tail;
2070 * If this event is bigger than the minimum size, then
2071 * we need to be careful that we don't subtract the
2072 * write counter enough to allow another writer to slip
2073 * in on this page.
2074 * We put in a discarded commit instead, to make sure
2075 * that this space is not used again.
2077 * If we are less than the minimum size, we don't need to
2078 * worry about it.
2080 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2081 /* No room for any events */
2083 /* Mark the rest of the page with padding */
2084 rb_event_set_padding(event);
2086 /* Set the write back to the previous setting */
2087 local_sub(length, &tail_page->write);
2088 return;
2091 /* Put in a discarded event */
2092 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2093 event->type_len = RINGBUF_TYPE_PADDING;
2094 /* time delta must be non zero */
2095 event->time_delta = 1;
2097 /* Set write to end of buffer */
2098 length = (tail + length) - BUF_PAGE_SIZE;
2099 local_sub(length, &tail_page->write);
2102 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2105 * This is the slow path, force gcc not to inline it.
2107 static noinline struct ring_buffer_event *
2108 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2109 unsigned long tail, struct rb_event_info *info)
2111 struct buffer_page *tail_page = info->tail_page;
2112 struct buffer_page *commit_page = cpu_buffer->commit_page;
2113 struct ring_buffer *buffer = cpu_buffer->buffer;
2114 struct buffer_page *next_page;
2115 int ret;
2117 next_page = tail_page;
2119 rb_inc_page(cpu_buffer, &next_page);
2122 * If for some reason, we had an interrupt storm that made
2123 * it all the way around the buffer, bail, and warn
2124 * about it.
2126 if (unlikely(next_page == commit_page)) {
2127 local_inc(&cpu_buffer->commit_overrun);
2128 goto out_reset;
2132 * This is where the fun begins!
2134 * We are fighting against races between a reader that
2135 * could be on another CPU trying to swap its reader
2136 * page with the buffer head.
2138 * We are also fighting against interrupts coming in and
2139 * moving the head or tail on us as well.
2141 * If the next page is the head page then we have filled
2142 * the buffer, unless the commit page is still on the
2143 * reader page.
2145 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2148 * If the commit is not on the reader page, then
2149 * move the header page.
2151 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2153 * If we are not in overwrite mode,
2154 * this is easy, just stop here.
2156 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2157 local_inc(&cpu_buffer->dropped_events);
2158 goto out_reset;
2161 ret = rb_handle_head_page(cpu_buffer,
2162 tail_page,
2163 next_page);
2164 if (ret < 0)
2165 goto out_reset;
2166 if (ret)
2167 goto out_again;
2168 } else {
2170 * We need to be careful here too. The
2171 * commit page could still be on the reader
2172 * page. We could have a small buffer, and
2173 * have filled up the buffer with events
2174 * from interrupts and such, and wrapped.
2176 * Note, if the tail page is also the on the
2177 * reader_page, we let it move out.
2179 if (unlikely((cpu_buffer->commit_page !=
2180 cpu_buffer->tail_page) &&
2181 (cpu_buffer->commit_page ==
2182 cpu_buffer->reader_page))) {
2183 local_inc(&cpu_buffer->commit_overrun);
2184 goto out_reset;
2189 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2191 out_again:
2193 rb_reset_tail(cpu_buffer, tail, info);
2195 /* Commit what we have for now. */
2196 rb_end_commit(cpu_buffer);
2197 /* rb_end_commit() decs committing */
2198 local_inc(&cpu_buffer->committing);
2200 /* fail and let the caller try again */
2201 return ERR_PTR(-EAGAIN);
2203 out_reset:
2204 /* reset write */
2205 rb_reset_tail(cpu_buffer, tail, info);
2207 return NULL;
2210 /* Slow path, do not inline */
2211 static noinline struct ring_buffer_event *
2212 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
2214 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2216 /* Not the first event on the page? */
2217 if (rb_event_index(event)) {
2218 event->time_delta = delta & TS_MASK;
2219 event->array[0] = delta >> TS_SHIFT;
2220 } else {
2221 /* nope, just zero it */
2222 event->time_delta = 0;
2223 event->array[0] = 0;
2226 return skip_time_extend(event);
2229 static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2230 struct ring_buffer_event *event);
2233 * rb_update_event - update event type and data
2234 * @event: the event to update
2235 * @type: the type of event
2236 * @length: the size of the event field in the ring buffer
2238 * Update the type and data fields of the event. The length
2239 * is the actual size that is written to the ring buffer,
2240 * and with this, we can determine what to place into the
2241 * data field.
2243 static void
2244 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2245 struct ring_buffer_event *event,
2246 struct rb_event_info *info)
2248 unsigned length = info->length;
2249 u64 delta = info->delta;
2251 /* Only a commit updates the timestamp */
2252 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2253 delta = 0;
2256 * If we need to add a timestamp, then we
2257 * add it to the start of the resevered space.
2259 if (unlikely(info->add_timestamp)) {
2260 event = rb_add_time_stamp(event, delta);
2261 length -= RB_LEN_TIME_EXTEND;
2262 delta = 0;
2265 event->time_delta = delta;
2266 length -= RB_EVNT_HDR_SIZE;
2267 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2268 event->type_len = 0;
2269 event->array[0] = length;
2270 } else
2271 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2274 static unsigned rb_calculate_event_length(unsigned length)
2276 struct ring_buffer_event event; /* Used only for sizeof array */
2278 /* zero length can cause confusions */
2279 if (!length)
2280 length++;
2282 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2283 length += sizeof(event.array[0]);
2285 length += RB_EVNT_HDR_SIZE;
2286 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2289 * In case the time delta is larger than the 27 bits for it
2290 * in the header, we need to add a timestamp. If another
2291 * event comes in when trying to discard this one to increase
2292 * the length, then the timestamp will be added in the allocated
2293 * space of this event. If length is bigger than the size needed
2294 * for the TIME_EXTEND, then padding has to be used. The events
2295 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2296 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2297 * As length is a multiple of 4, we only need to worry if it
2298 * is 12 (RB_LEN_TIME_EXTEND + 4).
2300 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2301 length += RB_ALIGNMENT;
2303 return length;
2306 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2307 static inline bool sched_clock_stable(void)
2309 return true;
2311 #endif
2313 static inline int
2314 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2315 struct ring_buffer_event *event)
2317 unsigned long new_index, old_index;
2318 struct buffer_page *bpage;
2319 unsigned long index;
2320 unsigned long addr;
2322 new_index = rb_event_index(event);
2323 old_index = new_index + rb_event_ts_length(event);
2324 addr = (unsigned long)event;
2325 addr &= PAGE_MASK;
2327 bpage = READ_ONCE(cpu_buffer->tail_page);
2329 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2330 unsigned long write_mask =
2331 local_read(&bpage->write) & ~RB_WRITE_MASK;
2332 unsigned long event_length = rb_event_length(event);
2334 * This is on the tail page. It is possible that
2335 * a write could come in and move the tail page
2336 * and write to the next page. That is fine
2337 * because we just shorten what is on this page.
2339 old_index += write_mask;
2340 new_index += write_mask;
2341 index = local_cmpxchg(&bpage->write, old_index, new_index);
2342 if (index == old_index) {
2343 /* update counters */
2344 local_sub(event_length, &cpu_buffer->entries_bytes);
2345 return 1;
2349 /* could not discard */
2350 return 0;
2353 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2355 local_inc(&cpu_buffer->committing);
2356 local_inc(&cpu_buffer->commits);
2359 static __always_inline void
2360 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2362 unsigned long max_count;
2365 * We only race with interrupts and NMIs on this CPU.
2366 * If we own the commit event, then we can commit
2367 * all others that interrupted us, since the interruptions
2368 * are in stack format (they finish before they come
2369 * back to us). This allows us to do a simple loop to
2370 * assign the commit to the tail.
2372 again:
2373 max_count = cpu_buffer->nr_pages * 100;
2375 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2376 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2377 return;
2378 if (RB_WARN_ON(cpu_buffer,
2379 rb_is_reader_page(cpu_buffer->tail_page)))
2380 return;
2381 local_set(&cpu_buffer->commit_page->page->commit,
2382 rb_page_write(cpu_buffer->commit_page));
2383 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2384 /* Only update the write stamp if the page has an event */
2385 if (rb_page_write(cpu_buffer->commit_page))
2386 cpu_buffer->write_stamp =
2387 cpu_buffer->commit_page->page->time_stamp;
2388 /* add barrier to keep gcc from optimizing too much */
2389 barrier();
2391 while (rb_commit_index(cpu_buffer) !=
2392 rb_page_write(cpu_buffer->commit_page)) {
2394 local_set(&cpu_buffer->commit_page->page->commit,
2395 rb_page_write(cpu_buffer->commit_page));
2396 RB_WARN_ON(cpu_buffer,
2397 local_read(&cpu_buffer->commit_page->page->commit) &
2398 ~RB_WRITE_MASK);
2399 barrier();
2402 /* again, keep gcc from optimizing */
2403 barrier();
2406 * If an interrupt came in just after the first while loop
2407 * and pushed the tail page forward, we will be left with
2408 * a dangling commit that will never go forward.
2410 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2411 goto again;
2414 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2416 unsigned long commits;
2418 if (RB_WARN_ON(cpu_buffer,
2419 !local_read(&cpu_buffer->committing)))
2420 return;
2422 again:
2423 commits = local_read(&cpu_buffer->commits);
2424 /* synchronize with interrupts */
2425 barrier();
2426 if (local_read(&cpu_buffer->committing) == 1)
2427 rb_set_commit_to_write(cpu_buffer);
2429 local_dec(&cpu_buffer->committing);
2431 /* synchronize with interrupts */
2432 barrier();
2435 * Need to account for interrupts coming in between the
2436 * updating of the commit page and the clearing of the
2437 * committing counter.
2439 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2440 !local_read(&cpu_buffer->committing)) {
2441 local_inc(&cpu_buffer->committing);
2442 goto again;
2446 static inline void rb_event_discard(struct ring_buffer_event *event)
2448 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2449 event = skip_time_extend(event);
2451 /* array[0] holds the actual length for the discarded event */
2452 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2453 event->type_len = RINGBUF_TYPE_PADDING;
2454 /* time delta must be non zero */
2455 if (!event->time_delta)
2456 event->time_delta = 1;
2459 static __always_inline bool
2460 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2461 struct ring_buffer_event *event)
2463 unsigned long addr = (unsigned long)event;
2464 unsigned long index;
2466 index = rb_event_index(event);
2467 addr &= PAGE_MASK;
2469 return cpu_buffer->commit_page->page == (void *)addr &&
2470 rb_commit_index(cpu_buffer) == index;
2473 static __always_inline void
2474 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2475 struct ring_buffer_event *event)
2477 u64 delta;
2480 * The event first in the commit queue updates the
2481 * time stamp.
2483 if (rb_event_is_commit(cpu_buffer, event)) {
2485 * A commit event that is first on a page
2486 * updates the write timestamp with the page stamp
2488 if (!rb_event_index(event))
2489 cpu_buffer->write_stamp =
2490 cpu_buffer->commit_page->page->time_stamp;
2491 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2492 delta = event->array[0];
2493 delta <<= TS_SHIFT;
2494 delta += event->time_delta;
2495 cpu_buffer->write_stamp += delta;
2496 } else
2497 cpu_buffer->write_stamp += event->time_delta;
2501 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2502 struct ring_buffer_event *event)
2504 local_inc(&cpu_buffer->entries);
2505 rb_update_write_stamp(cpu_buffer, event);
2506 rb_end_commit(cpu_buffer);
2509 static __always_inline void
2510 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2512 bool pagebusy;
2514 if (buffer->irq_work.waiters_pending) {
2515 buffer->irq_work.waiters_pending = false;
2516 /* irq_work_queue() supplies it's own memory barriers */
2517 irq_work_queue(&buffer->irq_work.work);
2520 if (cpu_buffer->irq_work.waiters_pending) {
2521 cpu_buffer->irq_work.waiters_pending = false;
2522 /* irq_work_queue() supplies it's own memory barriers */
2523 irq_work_queue(&cpu_buffer->irq_work.work);
2526 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
2528 if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2529 cpu_buffer->irq_work.wakeup_full = true;
2530 cpu_buffer->irq_work.full_waiters_pending = false;
2531 /* irq_work_queue() supplies it's own memory barriers */
2532 irq_work_queue(&cpu_buffer->irq_work.work);
2537 * The lock and unlock are done within a preempt disable section.
2538 * The current_context per_cpu variable can only be modified
2539 * by the current task between lock and unlock. But it can
2540 * be modified more than once via an interrupt. To pass this
2541 * information from the lock to the unlock without having to
2542 * access the 'in_interrupt()' functions again (which do show
2543 * a bit of overhead in something as critical as function tracing,
2544 * we use a bitmask trick.
2546 * bit 0 = NMI context
2547 * bit 1 = IRQ context
2548 * bit 2 = SoftIRQ context
2549 * bit 3 = normal context.
2551 * This works because this is the order of contexts that can
2552 * preempt other contexts. A SoftIRQ never preempts an IRQ
2553 * context.
2555 * When the context is determined, the corresponding bit is
2556 * checked and set (if it was set, then a recursion of that context
2557 * happened).
2559 * On unlock, we need to clear this bit. To do so, just subtract
2560 * 1 from the current_context and AND it to itself.
2562 * (binary)
2563 * 101 - 1 = 100
2564 * 101 & 100 = 100 (clearing bit zero)
2566 * 1010 - 1 = 1001
2567 * 1010 & 1001 = 1000 (clearing bit 1)
2569 * The least significant bit can be cleared this way, and it
2570 * just so happens that it is the same bit corresponding to
2571 * the current context.
2574 static __always_inline int
2575 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2577 unsigned int val = cpu_buffer->current_context;
2578 int bit;
2580 if (in_interrupt()) {
2581 if (in_nmi())
2582 bit = RB_CTX_NMI;
2583 else if (in_irq())
2584 bit = RB_CTX_IRQ;
2585 else
2586 bit = RB_CTX_SOFTIRQ;
2587 } else
2588 bit = RB_CTX_NORMAL;
2590 if (unlikely(val & (1 << bit)))
2591 return 1;
2593 val |= (1 << bit);
2594 cpu_buffer->current_context = val;
2596 return 0;
2599 static __always_inline void
2600 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2602 cpu_buffer->current_context &= cpu_buffer->current_context - 1;
2606 * ring_buffer_unlock_commit - commit a reserved
2607 * @buffer: The buffer to commit to
2608 * @event: The event pointer to commit.
2610 * This commits the data to the ring buffer, and releases any locks held.
2612 * Must be paired with ring_buffer_lock_reserve.
2614 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2615 struct ring_buffer_event *event)
2617 struct ring_buffer_per_cpu *cpu_buffer;
2618 int cpu = raw_smp_processor_id();
2620 cpu_buffer = buffer->buffers[cpu];
2622 rb_commit(cpu_buffer, event);
2624 rb_wakeups(buffer, cpu_buffer);
2626 trace_recursive_unlock(cpu_buffer);
2628 preempt_enable_notrace();
2630 return 0;
2632 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2634 static noinline void
2635 rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2636 struct rb_event_info *info)
2638 WARN_ONCE(info->delta > (1ULL << 59),
2639 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2640 (unsigned long long)info->delta,
2641 (unsigned long long)info->ts,
2642 (unsigned long long)cpu_buffer->write_stamp,
2643 sched_clock_stable() ? "" :
2644 "If you just came from a suspend/resume,\n"
2645 "please switch to the trace global clock:\n"
2646 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2647 info->add_timestamp = 1;
2650 static struct ring_buffer_event *
2651 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2652 struct rb_event_info *info)
2654 struct ring_buffer_event *event;
2655 struct buffer_page *tail_page;
2656 unsigned long tail, write;
2659 * If the time delta since the last event is too big to
2660 * hold in the time field of the event, then we append a
2661 * TIME EXTEND event ahead of the data event.
2663 if (unlikely(info->add_timestamp))
2664 info->length += RB_LEN_TIME_EXTEND;
2666 /* Don't let the compiler play games with cpu_buffer->tail_page */
2667 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2668 write = local_add_return(info->length, &tail_page->write);
2670 /* set write to only the index of the write */
2671 write &= RB_WRITE_MASK;
2672 tail = write - info->length;
2675 * If this is the first commit on the page, then it has the same
2676 * timestamp as the page itself.
2678 if (!tail)
2679 info->delta = 0;
2681 /* See if we shot pass the end of this buffer page */
2682 if (unlikely(write > BUF_PAGE_SIZE))
2683 return rb_move_tail(cpu_buffer, tail, info);
2685 /* We reserved something on the buffer */
2687 event = __rb_page_index(tail_page, tail);
2688 kmemcheck_annotate_bitfield(event, bitfield);
2689 rb_update_event(cpu_buffer, event, info);
2691 local_inc(&tail_page->entries);
2694 * If this is the first commit on the page, then update
2695 * its timestamp.
2697 if (!tail)
2698 tail_page->page->time_stamp = info->ts;
2700 /* account for these added bytes */
2701 local_add(info->length, &cpu_buffer->entries_bytes);
2703 return event;
2706 static __always_inline struct ring_buffer_event *
2707 rb_reserve_next_event(struct ring_buffer *buffer,
2708 struct ring_buffer_per_cpu *cpu_buffer,
2709 unsigned long length)
2711 struct ring_buffer_event *event;
2712 struct rb_event_info info;
2713 int nr_loops = 0;
2714 u64 diff;
2716 rb_start_commit(cpu_buffer);
2718 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2720 * Due to the ability to swap a cpu buffer from a buffer
2721 * it is possible it was swapped before we committed.
2722 * (committing stops a swap). We check for it here and
2723 * if it happened, we have to fail the write.
2725 barrier();
2726 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2727 local_dec(&cpu_buffer->committing);
2728 local_dec(&cpu_buffer->commits);
2729 return NULL;
2731 #endif
2733 info.length = rb_calculate_event_length(length);
2734 again:
2735 info.add_timestamp = 0;
2736 info.delta = 0;
2739 * We allow for interrupts to reenter here and do a trace.
2740 * If one does, it will cause this original code to loop
2741 * back here. Even with heavy interrupts happening, this
2742 * should only happen a few times in a row. If this happens
2743 * 1000 times in a row, there must be either an interrupt
2744 * storm or we have something buggy.
2745 * Bail!
2747 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2748 goto out_fail;
2750 info.ts = rb_time_stamp(cpu_buffer->buffer);
2751 diff = info.ts - cpu_buffer->write_stamp;
2753 /* make sure this diff is calculated here */
2754 barrier();
2756 /* Did the write stamp get updated already? */
2757 if (likely(info.ts >= cpu_buffer->write_stamp)) {
2758 info.delta = diff;
2759 if (unlikely(test_time_stamp(info.delta)))
2760 rb_handle_timestamp(cpu_buffer, &info);
2763 event = __rb_reserve_next(cpu_buffer, &info);
2765 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2766 if (info.add_timestamp)
2767 info.length -= RB_LEN_TIME_EXTEND;
2768 goto again;
2771 if (!event)
2772 goto out_fail;
2774 return event;
2776 out_fail:
2777 rb_end_commit(cpu_buffer);
2778 return NULL;
2782 * ring_buffer_lock_reserve - reserve a part of the buffer
2783 * @buffer: the ring buffer to reserve from
2784 * @length: the length of the data to reserve (excluding event header)
2786 * Returns a reseverd event on the ring buffer to copy directly to.
2787 * The user of this interface will need to get the body to write into
2788 * and can use the ring_buffer_event_data() interface.
2790 * The length is the length of the data needed, not the event length
2791 * which also includes the event header.
2793 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2794 * If NULL is returned, then nothing has been allocated or locked.
2796 struct ring_buffer_event *
2797 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2799 struct ring_buffer_per_cpu *cpu_buffer;
2800 struct ring_buffer_event *event;
2801 int cpu;
2803 /* If we are tracing schedule, we don't want to recurse */
2804 preempt_disable_notrace();
2806 if (unlikely(atomic_read(&buffer->record_disabled)))
2807 goto out;
2809 cpu = raw_smp_processor_id();
2811 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2812 goto out;
2814 cpu_buffer = buffer->buffers[cpu];
2816 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2817 goto out;
2819 if (unlikely(length > BUF_MAX_DATA_SIZE))
2820 goto out;
2822 if (unlikely(trace_recursive_lock(cpu_buffer)))
2823 goto out;
2825 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2826 if (!event)
2827 goto out_unlock;
2829 return event;
2831 out_unlock:
2832 trace_recursive_unlock(cpu_buffer);
2833 out:
2834 preempt_enable_notrace();
2835 return NULL;
2837 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2840 * Decrement the entries to the page that an event is on.
2841 * The event does not even need to exist, only the pointer
2842 * to the page it is on. This may only be called before the commit
2843 * takes place.
2845 static inline void
2846 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2847 struct ring_buffer_event *event)
2849 unsigned long addr = (unsigned long)event;
2850 struct buffer_page *bpage = cpu_buffer->commit_page;
2851 struct buffer_page *start;
2853 addr &= PAGE_MASK;
2855 /* Do the likely case first */
2856 if (likely(bpage->page == (void *)addr)) {
2857 local_dec(&bpage->entries);
2858 return;
2862 * Because the commit page may be on the reader page we
2863 * start with the next page and check the end loop there.
2865 rb_inc_page(cpu_buffer, &bpage);
2866 start = bpage;
2867 do {
2868 if (bpage->page == (void *)addr) {
2869 local_dec(&bpage->entries);
2870 return;
2872 rb_inc_page(cpu_buffer, &bpage);
2873 } while (bpage != start);
2875 /* commit not part of this buffer?? */
2876 RB_WARN_ON(cpu_buffer, 1);
2880 * ring_buffer_commit_discard - discard an event that has not been committed
2881 * @buffer: the ring buffer
2882 * @event: non committed event to discard
2884 * Sometimes an event that is in the ring buffer needs to be ignored.
2885 * This function lets the user discard an event in the ring buffer
2886 * and then that event will not be read later.
2888 * This function only works if it is called before the the item has been
2889 * committed. It will try to free the event from the ring buffer
2890 * if another event has not been added behind it.
2892 * If another event has been added behind it, it will set the event
2893 * up as discarded, and perform the commit.
2895 * If this function is called, do not call ring_buffer_unlock_commit on
2896 * the event.
2898 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2899 struct ring_buffer_event *event)
2901 struct ring_buffer_per_cpu *cpu_buffer;
2902 int cpu;
2904 /* The event is discarded regardless */
2905 rb_event_discard(event);
2907 cpu = smp_processor_id();
2908 cpu_buffer = buffer->buffers[cpu];
2911 * This must only be called if the event has not been
2912 * committed yet. Thus we can assume that preemption
2913 * is still disabled.
2915 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2917 rb_decrement_entry(cpu_buffer, event);
2918 if (rb_try_to_discard(cpu_buffer, event))
2919 goto out;
2922 * The commit is still visible by the reader, so we
2923 * must still update the timestamp.
2925 rb_update_write_stamp(cpu_buffer, event);
2926 out:
2927 rb_end_commit(cpu_buffer);
2929 trace_recursive_unlock(cpu_buffer);
2931 preempt_enable_notrace();
2934 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2937 * ring_buffer_write - write data to the buffer without reserving
2938 * @buffer: The ring buffer to write to.
2939 * @length: The length of the data being written (excluding the event header)
2940 * @data: The data to write to the buffer.
2942 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2943 * one function. If you already have the data to write to the buffer, it
2944 * may be easier to simply call this function.
2946 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2947 * and not the length of the event which would hold the header.
2949 int ring_buffer_write(struct ring_buffer *buffer,
2950 unsigned long length,
2951 void *data)
2953 struct ring_buffer_per_cpu *cpu_buffer;
2954 struct ring_buffer_event *event;
2955 void *body;
2956 int ret = -EBUSY;
2957 int cpu;
2959 preempt_disable_notrace();
2961 if (atomic_read(&buffer->record_disabled))
2962 goto out;
2964 cpu = raw_smp_processor_id();
2966 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2967 goto out;
2969 cpu_buffer = buffer->buffers[cpu];
2971 if (atomic_read(&cpu_buffer->record_disabled))
2972 goto out;
2974 if (length > BUF_MAX_DATA_SIZE)
2975 goto out;
2977 if (unlikely(trace_recursive_lock(cpu_buffer)))
2978 goto out;
2980 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2981 if (!event)
2982 goto out_unlock;
2984 body = rb_event_data(event);
2986 memcpy(body, data, length);
2988 rb_commit(cpu_buffer, event);
2990 rb_wakeups(buffer, cpu_buffer);
2992 ret = 0;
2994 out_unlock:
2995 trace_recursive_unlock(cpu_buffer);
2997 out:
2998 preempt_enable_notrace();
3000 return ret;
3002 EXPORT_SYMBOL_GPL(ring_buffer_write);
3004 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3006 struct buffer_page *reader = cpu_buffer->reader_page;
3007 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3008 struct buffer_page *commit = cpu_buffer->commit_page;
3010 /* In case of error, head will be NULL */
3011 if (unlikely(!head))
3012 return true;
3014 return reader->read == rb_page_commit(reader) &&
3015 (commit == reader ||
3016 (commit == head &&
3017 head->read == rb_page_commit(commit)));
3021 * ring_buffer_record_disable - stop all writes into the buffer
3022 * @buffer: The ring buffer to stop writes to.
3024 * This prevents all writes to the buffer. Any attempt to write
3025 * to the buffer after this will fail and return NULL.
3027 * The caller should call synchronize_sched() after this.
3029 void ring_buffer_record_disable(struct ring_buffer *buffer)
3031 atomic_inc(&buffer->record_disabled);
3033 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3036 * ring_buffer_record_enable - enable writes to the buffer
3037 * @buffer: The ring buffer to enable writes
3039 * Note, multiple disables will need the same number of enables
3040 * to truly enable the writing (much like preempt_disable).
3042 void ring_buffer_record_enable(struct ring_buffer *buffer)
3044 atomic_dec(&buffer->record_disabled);
3046 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3049 * ring_buffer_record_off - stop all writes into the buffer
3050 * @buffer: The ring buffer to stop writes to.
3052 * This prevents all writes to the buffer. Any attempt to write
3053 * to the buffer after this will fail and return NULL.
3055 * This is different than ring_buffer_record_disable() as
3056 * it works like an on/off switch, where as the disable() version
3057 * must be paired with a enable().
3059 void ring_buffer_record_off(struct ring_buffer *buffer)
3061 unsigned int rd;
3062 unsigned int new_rd;
3064 do {
3065 rd = atomic_read(&buffer->record_disabled);
3066 new_rd = rd | RB_BUFFER_OFF;
3067 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3069 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3072 * ring_buffer_record_on - restart writes into the buffer
3073 * @buffer: The ring buffer to start writes to.
3075 * This enables all writes to the buffer that was disabled by
3076 * ring_buffer_record_off().
3078 * This is different than ring_buffer_record_enable() as
3079 * it works like an on/off switch, where as the enable() version
3080 * must be paired with a disable().
3082 void ring_buffer_record_on(struct ring_buffer *buffer)
3084 unsigned int rd;
3085 unsigned int new_rd;
3087 do {
3088 rd = atomic_read(&buffer->record_disabled);
3089 new_rd = rd & ~RB_BUFFER_OFF;
3090 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3092 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3095 * ring_buffer_record_is_on - return true if the ring buffer can write
3096 * @buffer: The ring buffer to see if write is enabled
3098 * Returns true if the ring buffer is in a state that it accepts writes.
3100 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3102 return !atomic_read(&buffer->record_disabled);
3106 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3107 * @buffer: The ring buffer to stop writes to.
3108 * @cpu: The CPU buffer to stop
3110 * This prevents all writes to the buffer. Any attempt to write
3111 * to the buffer after this will fail and return NULL.
3113 * The caller should call synchronize_sched() after this.
3115 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3117 struct ring_buffer_per_cpu *cpu_buffer;
3119 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3120 return;
3122 cpu_buffer = buffer->buffers[cpu];
3123 atomic_inc(&cpu_buffer->record_disabled);
3125 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3128 * ring_buffer_record_enable_cpu - enable writes to the buffer
3129 * @buffer: The ring buffer to enable writes
3130 * @cpu: The CPU to enable.
3132 * Note, multiple disables will need the same number of enables
3133 * to truly enable the writing (much like preempt_disable).
3135 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3137 struct ring_buffer_per_cpu *cpu_buffer;
3139 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3140 return;
3142 cpu_buffer = buffer->buffers[cpu];
3143 atomic_dec(&cpu_buffer->record_disabled);
3145 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3148 * The total entries in the ring buffer is the running counter
3149 * of entries entered into the ring buffer, minus the sum of
3150 * the entries read from the ring buffer and the number of
3151 * entries that were overwritten.
3153 static inline unsigned long
3154 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3156 return local_read(&cpu_buffer->entries) -
3157 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3161 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3162 * @buffer: The ring buffer
3163 * @cpu: The per CPU buffer to read from.
3165 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3167 unsigned long flags;
3168 struct ring_buffer_per_cpu *cpu_buffer;
3169 struct buffer_page *bpage;
3170 u64 ret = 0;
3172 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3173 return 0;
3175 cpu_buffer = buffer->buffers[cpu];
3176 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3178 * if the tail is on reader_page, oldest time stamp is on the reader
3179 * page
3181 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3182 bpage = cpu_buffer->reader_page;
3183 else
3184 bpage = rb_set_head_page(cpu_buffer);
3185 if (bpage)
3186 ret = bpage->page->time_stamp;
3187 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3189 return ret;
3191 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3194 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3195 * @buffer: The ring buffer
3196 * @cpu: The per CPU buffer to read from.
3198 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3200 struct ring_buffer_per_cpu *cpu_buffer;
3201 unsigned long ret;
3203 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3204 return 0;
3206 cpu_buffer = buffer->buffers[cpu];
3207 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3209 return ret;
3211 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3214 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3215 * @buffer: The ring buffer
3216 * @cpu: The per CPU buffer to get the entries from.
3218 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3220 struct ring_buffer_per_cpu *cpu_buffer;
3222 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3223 return 0;
3225 cpu_buffer = buffer->buffers[cpu];
3227 return rb_num_of_entries(cpu_buffer);
3229 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3232 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3233 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3234 * @buffer: The ring buffer
3235 * @cpu: The per CPU buffer to get the number of overruns from
3237 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3239 struct ring_buffer_per_cpu *cpu_buffer;
3240 unsigned long ret;
3242 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3243 return 0;
3245 cpu_buffer = buffer->buffers[cpu];
3246 ret = local_read(&cpu_buffer->overrun);
3248 return ret;
3250 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3253 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3254 * commits failing due to the buffer wrapping around while there are uncommitted
3255 * events, such as during an interrupt storm.
3256 * @buffer: The ring buffer
3257 * @cpu: The per CPU buffer to get the number of overruns from
3259 unsigned long
3260 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3262 struct ring_buffer_per_cpu *cpu_buffer;
3263 unsigned long ret;
3265 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3266 return 0;
3268 cpu_buffer = buffer->buffers[cpu];
3269 ret = local_read(&cpu_buffer->commit_overrun);
3271 return ret;
3273 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3276 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3277 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3278 * @buffer: The ring buffer
3279 * @cpu: The per CPU buffer to get the number of overruns from
3281 unsigned long
3282 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3284 struct ring_buffer_per_cpu *cpu_buffer;
3285 unsigned long ret;
3287 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3288 return 0;
3290 cpu_buffer = buffer->buffers[cpu];
3291 ret = local_read(&cpu_buffer->dropped_events);
3293 return ret;
3295 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3298 * ring_buffer_read_events_cpu - get the number of events successfully read
3299 * @buffer: The ring buffer
3300 * @cpu: The per CPU buffer to get the number of events read
3302 unsigned long
3303 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3305 struct ring_buffer_per_cpu *cpu_buffer;
3307 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3308 return 0;
3310 cpu_buffer = buffer->buffers[cpu];
3311 return cpu_buffer->read;
3313 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3316 * ring_buffer_entries - get the number of entries in a buffer
3317 * @buffer: The ring buffer
3319 * Returns the total number of entries in the ring buffer
3320 * (all CPU entries)
3322 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3324 struct ring_buffer_per_cpu *cpu_buffer;
3325 unsigned long entries = 0;
3326 int cpu;
3328 /* if you care about this being correct, lock the buffer */
3329 for_each_buffer_cpu(buffer, cpu) {
3330 cpu_buffer = buffer->buffers[cpu];
3331 entries += rb_num_of_entries(cpu_buffer);
3334 return entries;
3336 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3339 * ring_buffer_overruns - get the number of overruns in buffer
3340 * @buffer: The ring buffer
3342 * Returns the total number of overruns in the ring buffer
3343 * (all CPU entries)
3345 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3347 struct ring_buffer_per_cpu *cpu_buffer;
3348 unsigned long overruns = 0;
3349 int cpu;
3351 /* if you care about this being correct, lock the buffer */
3352 for_each_buffer_cpu(buffer, cpu) {
3353 cpu_buffer = buffer->buffers[cpu];
3354 overruns += local_read(&cpu_buffer->overrun);
3357 return overruns;
3359 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3361 static void rb_iter_reset(struct ring_buffer_iter *iter)
3363 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3365 /* Iterator usage is expected to have record disabled */
3366 iter->head_page = cpu_buffer->reader_page;
3367 iter->head = cpu_buffer->reader_page->read;
3369 iter->cache_reader_page = iter->head_page;
3370 iter->cache_read = cpu_buffer->read;
3372 if (iter->head)
3373 iter->read_stamp = cpu_buffer->read_stamp;
3374 else
3375 iter->read_stamp = iter->head_page->page->time_stamp;
3379 * ring_buffer_iter_reset - reset an iterator
3380 * @iter: The iterator to reset
3382 * Resets the iterator, so that it will start from the beginning
3383 * again.
3385 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3387 struct ring_buffer_per_cpu *cpu_buffer;
3388 unsigned long flags;
3390 if (!iter)
3391 return;
3393 cpu_buffer = iter->cpu_buffer;
3395 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3396 rb_iter_reset(iter);
3397 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3399 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3402 * ring_buffer_iter_empty - check if an iterator has no more to read
3403 * @iter: The iterator to check
3405 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3407 struct ring_buffer_per_cpu *cpu_buffer;
3409 cpu_buffer = iter->cpu_buffer;
3411 return iter->head_page == cpu_buffer->commit_page &&
3412 iter->head == rb_commit_index(cpu_buffer);
3414 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3416 static void
3417 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3418 struct ring_buffer_event *event)
3420 u64 delta;
3422 switch (event->type_len) {
3423 case RINGBUF_TYPE_PADDING:
3424 return;
3426 case RINGBUF_TYPE_TIME_EXTEND:
3427 delta = event->array[0];
3428 delta <<= TS_SHIFT;
3429 delta += event->time_delta;
3430 cpu_buffer->read_stamp += delta;
3431 return;
3433 case RINGBUF_TYPE_TIME_STAMP:
3434 /* FIXME: not implemented */
3435 return;
3437 case RINGBUF_TYPE_DATA:
3438 cpu_buffer->read_stamp += event->time_delta;
3439 return;
3441 default:
3442 BUG();
3444 return;
3447 static void
3448 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3449 struct ring_buffer_event *event)
3451 u64 delta;
3453 switch (event->type_len) {
3454 case RINGBUF_TYPE_PADDING:
3455 return;
3457 case RINGBUF_TYPE_TIME_EXTEND:
3458 delta = event->array[0];
3459 delta <<= TS_SHIFT;
3460 delta += event->time_delta;
3461 iter->read_stamp += delta;
3462 return;
3464 case RINGBUF_TYPE_TIME_STAMP:
3465 /* FIXME: not implemented */
3466 return;
3468 case RINGBUF_TYPE_DATA:
3469 iter->read_stamp += event->time_delta;
3470 return;
3472 default:
3473 BUG();
3475 return;
3478 static struct buffer_page *
3479 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3481 struct buffer_page *reader = NULL;
3482 unsigned long overwrite;
3483 unsigned long flags;
3484 int nr_loops = 0;
3485 int ret;
3487 local_irq_save(flags);
3488 arch_spin_lock(&cpu_buffer->lock);
3490 again:
3492 * This should normally only loop twice. But because the
3493 * start of the reader inserts an empty page, it causes
3494 * a case where we will loop three times. There should be no
3495 * reason to loop four times (that I know of).
3497 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3498 reader = NULL;
3499 goto out;
3502 reader = cpu_buffer->reader_page;
3504 /* If there's more to read, return this page */
3505 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3506 goto out;
3508 /* Never should we have an index greater than the size */
3509 if (RB_WARN_ON(cpu_buffer,
3510 cpu_buffer->reader_page->read > rb_page_size(reader)))
3511 goto out;
3513 /* check if we caught up to the tail */
3514 reader = NULL;
3515 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3516 goto out;
3518 /* Don't bother swapping if the ring buffer is empty */
3519 if (rb_num_of_entries(cpu_buffer) == 0)
3520 goto out;
3523 * Reset the reader page to size zero.
3525 local_set(&cpu_buffer->reader_page->write, 0);
3526 local_set(&cpu_buffer->reader_page->entries, 0);
3527 local_set(&cpu_buffer->reader_page->page->commit, 0);
3528 cpu_buffer->reader_page->real_end = 0;
3530 spin:
3532 * Splice the empty reader page into the list around the head.
3534 reader = rb_set_head_page(cpu_buffer);
3535 if (!reader)
3536 goto out;
3537 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3538 cpu_buffer->reader_page->list.prev = reader->list.prev;
3541 * cpu_buffer->pages just needs to point to the buffer, it
3542 * has no specific buffer page to point to. Lets move it out
3543 * of our way so we don't accidentally swap it.
3545 cpu_buffer->pages = reader->list.prev;
3547 /* The reader page will be pointing to the new head */
3548 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3551 * We want to make sure we read the overruns after we set up our
3552 * pointers to the next object. The writer side does a
3553 * cmpxchg to cross pages which acts as the mb on the writer
3554 * side. Note, the reader will constantly fail the swap
3555 * while the writer is updating the pointers, so this
3556 * guarantees that the overwrite recorded here is the one we
3557 * want to compare with the last_overrun.
3559 smp_mb();
3560 overwrite = local_read(&(cpu_buffer->overrun));
3563 * Here's the tricky part.
3565 * We need to move the pointer past the header page.
3566 * But we can only do that if a writer is not currently
3567 * moving it. The page before the header page has the
3568 * flag bit '1' set if it is pointing to the page we want.
3569 * but if the writer is in the process of moving it
3570 * than it will be '2' or already moved '0'.
3573 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3576 * If we did not convert it, then we must try again.
3578 if (!ret)
3579 goto spin;
3582 * Yeah! We succeeded in replacing the page.
3584 * Now make the new head point back to the reader page.
3586 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3587 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3589 /* Finally update the reader page to the new head */
3590 cpu_buffer->reader_page = reader;
3591 cpu_buffer->reader_page->read = 0;
3593 if (overwrite != cpu_buffer->last_overrun) {
3594 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3595 cpu_buffer->last_overrun = overwrite;
3598 goto again;
3600 out:
3601 /* Update the read_stamp on the first event */
3602 if (reader && reader->read == 0)
3603 cpu_buffer->read_stamp = reader->page->time_stamp;
3605 arch_spin_unlock(&cpu_buffer->lock);
3606 local_irq_restore(flags);
3608 return reader;
3611 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3613 struct ring_buffer_event *event;
3614 struct buffer_page *reader;
3615 unsigned length;
3617 reader = rb_get_reader_page(cpu_buffer);
3619 /* This function should not be called when buffer is empty */
3620 if (RB_WARN_ON(cpu_buffer, !reader))
3621 return;
3623 event = rb_reader_event(cpu_buffer);
3625 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3626 cpu_buffer->read++;
3628 rb_update_read_stamp(cpu_buffer, event);
3630 length = rb_event_length(event);
3631 cpu_buffer->reader_page->read += length;
3634 static void rb_advance_iter(struct ring_buffer_iter *iter)
3636 struct ring_buffer_per_cpu *cpu_buffer;
3637 struct ring_buffer_event *event;
3638 unsigned length;
3640 cpu_buffer = iter->cpu_buffer;
3643 * Check if we are at the end of the buffer.
3645 if (iter->head >= rb_page_size(iter->head_page)) {
3646 /* discarded commits can make the page empty */
3647 if (iter->head_page == cpu_buffer->commit_page)
3648 return;
3649 rb_inc_iter(iter);
3650 return;
3653 event = rb_iter_head_event(iter);
3655 length = rb_event_length(event);
3658 * This should not be called to advance the header if we are
3659 * at the tail of the buffer.
3661 if (RB_WARN_ON(cpu_buffer,
3662 (iter->head_page == cpu_buffer->commit_page) &&
3663 (iter->head + length > rb_commit_index(cpu_buffer))))
3664 return;
3666 rb_update_iter_read_stamp(iter, event);
3668 iter->head += length;
3670 /* check for end of page padding */
3671 if ((iter->head >= rb_page_size(iter->head_page)) &&
3672 (iter->head_page != cpu_buffer->commit_page))
3673 rb_inc_iter(iter);
3676 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3678 return cpu_buffer->lost_events;
3681 static struct ring_buffer_event *
3682 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3683 unsigned long *lost_events)
3685 struct ring_buffer_event *event;
3686 struct buffer_page *reader;
3687 int nr_loops = 0;
3689 again:
3691 * We repeat when a time extend is encountered.
3692 * Since the time extend is always attached to a data event,
3693 * we should never loop more than once.
3694 * (We never hit the following condition more than twice).
3696 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3697 return NULL;
3699 reader = rb_get_reader_page(cpu_buffer);
3700 if (!reader)
3701 return NULL;
3703 event = rb_reader_event(cpu_buffer);
3705 switch (event->type_len) {
3706 case RINGBUF_TYPE_PADDING:
3707 if (rb_null_event(event))
3708 RB_WARN_ON(cpu_buffer, 1);
3710 * Because the writer could be discarding every
3711 * event it creates (which would probably be bad)
3712 * if we were to go back to "again" then we may never
3713 * catch up, and will trigger the warn on, or lock
3714 * the box. Return the padding, and we will release
3715 * the current locks, and try again.
3717 return event;
3719 case RINGBUF_TYPE_TIME_EXTEND:
3720 /* Internal data, OK to advance */
3721 rb_advance_reader(cpu_buffer);
3722 goto again;
3724 case RINGBUF_TYPE_TIME_STAMP:
3725 /* FIXME: not implemented */
3726 rb_advance_reader(cpu_buffer);
3727 goto again;
3729 case RINGBUF_TYPE_DATA:
3730 if (ts) {
3731 *ts = cpu_buffer->read_stamp + event->time_delta;
3732 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3733 cpu_buffer->cpu, ts);
3735 if (lost_events)
3736 *lost_events = rb_lost_events(cpu_buffer);
3737 return event;
3739 default:
3740 BUG();
3743 return NULL;
3745 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3747 static struct ring_buffer_event *
3748 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3750 struct ring_buffer *buffer;
3751 struct ring_buffer_per_cpu *cpu_buffer;
3752 struct ring_buffer_event *event;
3753 int nr_loops = 0;
3755 cpu_buffer = iter->cpu_buffer;
3756 buffer = cpu_buffer->buffer;
3759 * Check if someone performed a consuming read to
3760 * the buffer. A consuming read invalidates the iterator
3761 * and we need to reset the iterator in this case.
3763 if (unlikely(iter->cache_read != cpu_buffer->read ||
3764 iter->cache_reader_page != cpu_buffer->reader_page))
3765 rb_iter_reset(iter);
3767 again:
3768 if (ring_buffer_iter_empty(iter))
3769 return NULL;
3772 * We repeat when a time extend is encountered or we hit
3773 * the end of the page. Since the time extend is always attached
3774 * to a data event, we should never loop more than three times.
3775 * Once for going to next page, once on time extend, and
3776 * finally once to get the event.
3777 * (We never hit the following condition more than thrice).
3779 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3780 return NULL;
3782 if (rb_per_cpu_empty(cpu_buffer))
3783 return NULL;
3785 if (iter->head >= rb_page_size(iter->head_page)) {
3786 rb_inc_iter(iter);
3787 goto again;
3790 event = rb_iter_head_event(iter);
3792 switch (event->type_len) {
3793 case RINGBUF_TYPE_PADDING:
3794 if (rb_null_event(event)) {
3795 rb_inc_iter(iter);
3796 goto again;
3798 rb_advance_iter(iter);
3799 return event;
3801 case RINGBUF_TYPE_TIME_EXTEND:
3802 /* Internal data, OK to advance */
3803 rb_advance_iter(iter);
3804 goto again;
3806 case RINGBUF_TYPE_TIME_STAMP:
3807 /* FIXME: not implemented */
3808 rb_advance_iter(iter);
3809 goto again;
3811 case RINGBUF_TYPE_DATA:
3812 if (ts) {
3813 *ts = iter->read_stamp + event->time_delta;
3814 ring_buffer_normalize_time_stamp(buffer,
3815 cpu_buffer->cpu, ts);
3817 return event;
3819 default:
3820 BUG();
3823 return NULL;
3825 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3827 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
3829 if (likely(!in_nmi())) {
3830 raw_spin_lock(&cpu_buffer->reader_lock);
3831 return true;
3835 * If an NMI die dumps out the content of the ring buffer
3836 * trylock must be used to prevent a deadlock if the NMI
3837 * preempted a task that holds the ring buffer locks. If
3838 * we get the lock then all is fine, if not, then continue
3839 * to do the read, but this can corrupt the ring buffer,
3840 * so it must be permanently disabled from future writes.
3841 * Reading from NMI is a oneshot deal.
3843 if (raw_spin_trylock(&cpu_buffer->reader_lock))
3844 return true;
3846 /* Continue without locking, but disable the ring buffer */
3847 atomic_inc(&cpu_buffer->record_disabled);
3848 return false;
3851 static inline void
3852 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
3854 if (likely(locked))
3855 raw_spin_unlock(&cpu_buffer->reader_lock);
3856 return;
3860 * ring_buffer_peek - peek at the next event to be read
3861 * @buffer: The ring buffer to read
3862 * @cpu: The cpu to peak at
3863 * @ts: The timestamp counter of this event.
3864 * @lost_events: a variable to store if events were lost (may be NULL)
3866 * This will return the event that will be read next, but does
3867 * not consume the data.
3869 struct ring_buffer_event *
3870 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3871 unsigned long *lost_events)
3873 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3874 struct ring_buffer_event *event;
3875 unsigned long flags;
3876 bool dolock;
3878 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3879 return NULL;
3881 again:
3882 local_irq_save(flags);
3883 dolock = rb_reader_lock(cpu_buffer);
3884 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3885 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3886 rb_advance_reader(cpu_buffer);
3887 rb_reader_unlock(cpu_buffer, dolock);
3888 local_irq_restore(flags);
3890 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3891 goto again;
3893 return event;
3897 * ring_buffer_iter_peek - peek at the next event to be read
3898 * @iter: The ring buffer iterator
3899 * @ts: The timestamp counter of this event.
3901 * This will return the event that will be read next, but does
3902 * not increment the iterator.
3904 struct ring_buffer_event *
3905 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3907 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3908 struct ring_buffer_event *event;
3909 unsigned long flags;
3911 again:
3912 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3913 event = rb_iter_peek(iter, ts);
3914 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3916 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3917 goto again;
3919 return event;
3923 * ring_buffer_consume - return an event and consume it
3924 * @buffer: The ring buffer to get the next event from
3925 * @cpu: the cpu to read the buffer from
3926 * @ts: a variable to store the timestamp (may be NULL)
3927 * @lost_events: a variable to store if events were lost (may be NULL)
3929 * Returns the next event in the ring buffer, and that event is consumed.
3930 * Meaning, that sequential reads will keep returning a different event,
3931 * and eventually empty the ring buffer if the producer is slower.
3933 struct ring_buffer_event *
3934 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3935 unsigned long *lost_events)
3937 struct ring_buffer_per_cpu *cpu_buffer;
3938 struct ring_buffer_event *event = NULL;
3939 unsigned long flags;
3940 bool dolock;
3942 again:
3943 /* might be called in atomic */
3944 preempt_disable();
3946 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3947 goto out;
3949 cpu_buffer = buffer->buffers[cpu];
3950 local_irq_save(flags);
3951 dolock = rb_reader_lock(cpu_buffer);
3953 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3954 if (event) {
3955 cpu_buffer->lost_events = 0;
3956 rb_advance_reader(cpu_buffer);
3959 rb_reader_unlock(cpu_buffer, dolock);
3960 local_irq_restore(flags);
3962 out:
3963 preempt_enable();
3965 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3966 goto again;
3968 return event;
3970 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3973 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3974 * @buffer: The ring buffer to read from
3975 * @cpu: The cpu buffer to iterate over
3977 * This performs the initial preparations necessary to iterate
3978 * through the buffer. Memory is allocated, buffer recording
3979 * is disabled, and the iterator pointer is returned to the caller.
3981 * Disabling buffer recordng prevents the reading from being
3982 * corrupted. This is not a consuming read, so a producer is not
3983 * expected.
3985 * After a sequence of ring_buffer_read_prepare calls, the user is
3986 * expected to make at least one call to ring_buffer_read_prepare_sync.
3987 * Afterwards, ring_buffer_read_start is invoked to get things going
3988 * for real.
3990 * This overall must be paired with ring_buffer_read_finish.
3992 struct ring_buffer_iter *
3993 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3995 struct ring_buffer_per_cpu *cpu_buffer;
3996 struct ring_buffer_iter *iter;
3998 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3999 return NULL;
4001 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
4002 if (!iter)
4003 return NULL;
4005 cpu_buffer = buffer->buffers[cpu];
4007 iter->cpu_buffer = cpu_buffer;
4009 atomic_inc(&buffer->resize_disabled);
4010 atomic_inc(&cpu_buffer->record_disabled);
4012 return iter;
4014 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4017 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4019 * All previously invoked ring_buffer_read_prepare calls to prepare
4020 * iterators will be synchronized. Afterwards, read_buffer_read_start
4021 * calls on those iterators are allowed.
4023 void
4024 ring_buffer_read_prepare_sync(void)
4026 synchronize_sched();
4028 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4031 * ring_buffer_read_start - start a non consuming read of the buffer
4032 * @iter: The iterator returned by ring_buffer_read_prepare
4034 * This finalizes the startup of an iteration through the buffer.
4035 * The iterator comes from a call to ring_buffer_read_prepare and
4036 * an intervening ring_buffer_read_prepare_sync must have been
4037 * performed.
4039 * Must be paired with ring_buffer_read_finish.
4041 void
4042 ring_buffer_read_start(struct ring_buffer_iter *iter)
4044 struct ring_buffer_per_cpu *cpu_buffer;
4045 unsigned long flags;
4047 if (!iter)
4048 return;
4050 cpu_buffer = iter->cpu_buffer;
4052 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4053 arch_spin_lock(&cpu_buffer->lock);
4054 rb_iter_reset(iter);
4055 arch_spin_unlock(&cpu_buffer->lock);
4056 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4058 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4061 * ring_buffer_read_finish - finish reading the iterator of the buffer
4062 * @iter: The iterator retrieved by ring_buffer_start
4064 * This re-enables the recording to the buffer, and frees the
4065 * iterator.
4067 void
4068 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4070 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4071 unsigned long flags;
4074 * Ring buffer is disabled from recording, here's a good place
4075 * to check the integrity of the ring buffer.
4076 * Must prevent readers from trying to read, as the check
4077 * clears the HEAD page and readers require it.
4079 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4080 rb_check_pages(cpu_buffer);
4081 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4083 atomic_dec(&cpu_buffer->record_disabled);
4084 atomic_dec(&cpu_buffer->buffer->resize_disabled);
4085 kfree(iter);
4087 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4090 * ring_buffer_read - read the next item in the ring buffer by the iterator
4091 * @iter: The ring buffer iterator
4092 * @ts: The time stamp of the event read.
4094 * This reads the next event in the ring buffer and increments the iterator.
4096 struct ring_buffer_event *
4097 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4099 struct ring_buffer_event *event;
4100 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4101 unsigned long flags;
4103 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4104 again:
4105 event = rb_iter_peek(iter, ts);
4106 if (!event)
4107 goto out;
4109 if (event->type_len == RINGBUF_TYPE_PADDING)
4110 goto again;
4112 rb_advance_iter(iter);
4113 out:
4114 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4116 return event;
4118 EXPORT_SYMBOL_GPL(ring_buffer_read);
4121 * ring_buffer_size - return the size of the ring buffer (in bytes)
4122 * @buffer: The ring buffer.
4124 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4127 * Earlier, this method returned
4128 * BUF_PAGE_SIZE * buffer->nr_pages
4129 * Since the nr_pages field is now removed, we have converted this to
4130 * return the per cpu buffer value.
4132 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4133 return 0;
4135 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4137 EXPORT_SYMBOL_GPL(ring_buffer_size);
4139 static void
4140 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4142 rb_head_page_deactivate(cpu_buffer);
4144 cpu_buffer->head_page
4145 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4146 local_set(&cpu_buffer->head_page->write, 0);
4147 local_set(&cpu_buffer->head_page->entries, 0);
4148 local_set(&cpu_buffer->head_page->page->commit, 0);
4150 cpu_buffer->head_page->read = 0;
4152 cpu_buffer->tail_page = cpu_buffer->head_page;
4153 cpu_buffer->commit_page = cpu_buffer->head_page;
4155 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4156 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4157 local_set(&cpu_buffer->reader_page->write, 0);
4158 local_set(&cpu_buffer->reader_page->entries, 0);
4159 local_set(&cpu_buffer->reader_page->page->commit, 0);
4160 cpu_buffer->reader_page->read = 0;
4162 local_set(&cpu_buffer->entries_bytes, 0);
4163 local_set(&cpu_buffer->overrun, 0);
4164 local_set(&cpu_buffer->commit_overrun, 0);
4165 local_set(&cpu_buffer->dropped_events, 0);
4166 local_set(&cpu_buffer->entries, 0);
4167 local_set(&cpu_buffer->committing, 0);
4168 local_set(&cpu_buffer->commits, 0);
4169 cpu_buffer->read = 0;
4170 cpu_buffer->read_bytes = 0;
4172 cpu_buffer->write_stamp = 0;
4173 cpu_buffer->read_stamp = 0;
4175 cpu_buffer->lost_events = 0;
4176 cpu_buffer->last_overrun = 0;
4178 rb_head_page_activate(cpu_buffer);
4182 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4183 * @buffer: The ring buffer to reset a per cpu buffer of
4184 * @cpu: The CPU buffer to be reset
4186 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4188 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4189 unsigned long flags;
4191 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4192 return;
4194 atomic_inc(&buffer->resize_disabled);
4195 atomic_inc(&cpu_buffer->record_disabled);
4197 /* Make sure all commits have finished */
4198 synchronize_sched();
4200 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4202 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4203 goto out;
4205 arch_spin_lock(&cpu_buffer->lock);
4207 rb_reset_cpu(cpu_buffer);
4209 arch_spin_unlock(&cpu_buffer->lock);
4211 out:
4212 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4214 atomic_dec(&cpu_buffer->record_disabled);
4215 atomic_dec(&buffer->resize_disabled);
4217 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4220 * ring_buffer_reset - reset a ring buffer
4221 * @buffer: The ring buffer to reset all cpu buffers
4223 void ring_buffer_reset(struct ring_buffer *buffer)
4225 int cpu;
4227 for_each_buffer_cpu(buffer, cpu)
4228 ring_buffer_reset_cpu(buffer, cpu);
4230 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4233 * rind_buffer_empty - is the ring buffer empty?
4234 * @buffer: The ring buffer to test
4236 bool ring_buffer_empty(struct ring_buffer *buffer)
4238 struct ring_buffer_per_cpu *cpu_buffer;
4239 unsigned long flags;
4240 bool dolock;
4241 int cpu;
4242 int ret;
4244 /* yes this is racy, but if you don't like the race, lock the buffer */
4245 for_each_buffer_cpu(buffer, cpu) {
4246 cpu_buffer = buffer->buffers[cpu];
4247 local_irq_save(flags);
4248 dolock = rb_reader_lock(cpu_buffer);
4249 ret = rb_per_cpu_empty(cpu_buffer);
4250 rb_reader_unlock(cpu_buffer, dolock);
4251 local_irq_restore(flags);
4253 if (!ret)
4254 return false;
4257 return true;
4259 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4262 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4263 * @buffer: The ring buffer
4264 * @cpu: The CPU buffer to test
4266 bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4268 struct ring_buffer_per_cpu *cpu_buffer;
4269 unsigned long flags;
4270 bool dolock;
4271 int ret;
4273 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4274 return true;
4276 cpu_buffer = buffer->buffers[cpu];
4277 local_irq_save(flags);
4278 dolock = rb_reader_lock(cpu_buffer);
4279 ret = rb_per_cpu_empty(cpu_buffer);
4280 rb_reader_unlock(cpu_buffer, dolock);
4281 local_irq_restore(flags);
4283 return ret;
4285 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4287 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4289 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4290 * @buffer_a: One buffer to swap with
4291 * @buffer_b: The other buffer to swap with
4293 * This function is useful for tracers that want to take a "snapshot"
4294 * of a CPU buffer and has another back up buffer lying around.
4295 * it is expected that the tracer handles the cpu buffer not being
4296 * used at the moment.
4298 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4299 struct ring_buffer *buffer_b, int cpu)
4301 struct ring_buffer_per_cpu *cpu_buffer_a;
4302 struct ring_buffer_per_cpu *cpu_buffer_b;
4303 int ret = -EINVAL;
4305 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4306 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4307 goto out;
4309 cpu_buffer_a = buffer_a->buffers[cpu];
4310 cpu_buffer_b = buffer_b->buffers[cpu];
4312 /* At least make sure the two buffers are somewhat the same */
4313 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4314 goto out;
4316 ret = -EAGAIN;
4318 if (atomic_read(&buffer_a->record_disabled))
4319 goto out;
4321 if (atomic_read(&buffer_b->record_disabled))
4322 goto out;
4324 if (atomic_read(&cpu_buffer_a->record_disabled))
4325 goto out;
4327 if (atomic_read(&cpu_buffer_b->record_disabled))
4328 goto out;
4331 * We can't do a synchronize_sched here because this
4332 * function can be called in atomic context.
4333 * Normally this will be called from the same CPU as cpu.
4334 * If not it's up to the caller to protect this.
4336 atomic_inc(&cpu_buffer_a->record_disabled);
4337 atomic_inc(&cpu_buffer_b->record_disabled);
4339 ret = -EBUSY;
4340 if (local_read(&cpu_buffer_a->committing))
4341 goto out_dec;
4342 if (local_read(&cpu_buffer_b->committing))
4343 goto out_dec;
4345 buffer_a->buffers[cpu] = cpu_buffer_b;
4346 buffer_b->buffers[cpu] = cpu_buffer_a;
4348 cpu_buffer_b->buffer = buffer_a;
4349 cpu_buffer_a->buffer = buffer_b;
4351 ret = 0;
4353 out_dec:
4354 atomic_dec(&cpu_buffer_a->record_disabled);
4355 atomic_dec(&cpu_buffer_b->record_disabled);
4356 out:
4357 return ret;
4359 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4360 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4363 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4364 * @buffer: the buffer to allocate for.
4365 * @cpu: the cpu buffer to allocate.
4367 * This function is used in conjunction with ring_buffer_read_page.
4368 * When reading a full page from the ring buffer, these functions
4369 * can be used to speed up the process. The calling function should
4370 * allocate a few pages first with this function. Then when it
4371 * needs to get pages from the ring buffer, it passes the result
4372 * of this function into ring_buffer_read_page, which will swap
4373 * the page that was allocated, with the read page of the buffer.
4375 * Returns:
4376 * The page allocated, or NULL on error.
4378 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4380 struct buffer_data_page *bpage;
4381 struct page *page;
4383 page = alloc_pages_node(cpu_to_node(cpu),
4384 GFP_KERNEL | __GFP_NORETRY, 0);
4385 if (!page)
4386 return NULL;
4388 bpage = page_address(page);
4390 rb_init_page(bpage);
4392 return bpage;
4394 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4397 * ring_buffer_free_read_page - free an allocated read page
4398 * @buffer: the buffer the page was allocate for
4399 * @data: the page to free
4401 * Free a page allocated from ring_buffer_alloc_read_page.
4403 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4405 free_page((unsigned long)data);
4407 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4410 * ring_buffer_read_page - extract a page from the ring buffer
4411 * @buffer: buffer to extract from
4412 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4413 * @len: amount to extract
4414 * @cpu: the cpu of the buffer to extract
4415 * @full: should the extraction only happen when the page is full.
4417 * This function will pull out a page from the ring buffer and consume it.
4418 * @data_page must be the address of the variable that was returned
4419 * from ring_buffer_alloc_read_page. This is because the page might be used
4420 * to swap with a page in the ring buffer.
4422 * for example:
4423 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4424 * if (!rpage)
4425 * return error;
4426 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4427 * if (ret >= 0)
4428 * process_page(rpage, ret);
4430 * When @full is set, the function will not return true unless
4431 * the writer is off the reader page.
4433 * Note: it is up to the calling functions to handle sleeps and wakeups.
4434 * The ring buffer can be used anywhere in the kernel and can not
4435 * blindly call wake_up. The layer that uses the ring buffer must be
4436 * responsible for that.
4438 * Returns:
4439 * >=0 if data has been transferred, returns the offset of consumed data.
4440 * <0 if no data has been transferred.
4442 int ring_buffer_read_page(struct ring_buffer *buffer,
4443 void **data_page, size_t len, int cpu, int full)
4445 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4446 struct ring_buffer_event *event;
4447 struct buffer_data_page *bpage;
4448 struct buffer_page *reader;
4449 unsigned long missed_events;
4450 unsigned long flags;
4451 unsigned int commit;
4452 unsigned int read;
4453 u64 save_timestamp;
4454 int ret = -1;
4456 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4457 goto out;
4460 * If len is not big enough to hold the page header, then
4461 * we can not copy anything.
4463 if (len <= BUF_PAGE_HDR_SIZE)
4464 goto out;
4466 len -= BUF_PAGE_HDR_SIZE;
4468 if (!data_page)
4469 goto out;
4471 bpage = *data_page;
4472 if (!bpage)
4473 goto out;
4475 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4477 reader = rb_get_reader_page(cpu_buffer);
4478 if (!reader)
4479 goto out_unlock;
4481 event = rb_reader_event(cpu_buffer);
4483 read = reader->read;
4484 commit = rb_page_commit(reader);
4486 /* Check if any events were dropped */
4487 missed_events = cpu_buffer->lost_events;
4490 * If this page has been partially read or
4491 * if len is not big enough to read the rest of the page or
4492 * a writer is still on the page, then
4493 * we must copy the data from the page to the buffer.
4494 * Otherwise, we can simply swap the page with the one passed in.
4496 if (read || (len < (commit - read)) ||
4497 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4498 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4499 unsigned int rpos = read;
4500 unsigned int pos = 0;
4501 unsigned int size;
4503 if (full)
4504 goto out_unlock;
4506 if (len > (commit - read))
4507 len = (commit - read);
4509 /* Always keep the time extend and data together */
4510 size = rb_event_ts_length(event);
4512 if (len < size)
4513 goto out_unlock;
4515 /* save the current timestamp, since the user will need it */
4516 save_timestamp = cpu_buffer->read_stamp;
4518 /* Need to copy one event at a time */
4519 do {
4520 /* We need the size of one event, because
4521 * rb_advance_reader only advances by one event,
4522 * whereas rb_event_ts_length may include the size of
4523 * one or two events.
4524 * We have already ensured there's enough space if this
4525 * is a time extend. */
4526 size = rb_event_length(event);
4527 memcpy(bpage->data + pos, rpage->data + rpos, size);
4529 len -= size;
4531 rb_advance_reader(cpu_buffer);
4532 rpos = reader->read;
4533 pos += size;
4535 if (rpos >= commit)
4536 break;
4538 event = rb_reader_event(cpu_buffer);
4539 /* Always keep the time extend and data together */
4540 size = rb_event_ts_length(event);
4541 } while (len >= size);
4543 /* update bpage */
4544 local_set(&bpage->commit, pos);
4545 bpage->time_stamp = save_timestamp;
4547 /* we copied everything to the beginning */
4548 read = 0;
4549 } else {
4550 /* update the entry counter */
4551 cpu_buffer->read += rb_page_entries(reader);
4552 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4554 /* swap the pages */
4555 rb_init_page(bpage);
4556 bpage = reader->page;
4557 reader->page = *data_page;
4558 local_set(&reader->write, 0);
4559 local_set(&reader->entries, 0);
4560 reader->read = 0;
4561 *data_page = bpage;
4564 * Use the real_end for the data size,
4565 * This gives us a chance to store the lost events
4566 * on the page.
4568 if (reader->real_end)
4569 local_set(&bpage->commit, reader->real_end);
4571 ret = read;
4573 cpu_buffer->lost_events = 0;
4575 commit = local_read(&bpage->commit);
4577 * Set a flag in the commit field if we lost events
4579 if (missed_events) {
4580 /* If there is room at the end of the page to save the
4581 * missed events, then record it there.
4583 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4584 memcpy(&bpage->data[commit], &missed_events,
4585 sizeof(missed_events));
4586 local_add(RB_MISSED_STORED, &bpage->commit);
4587 commit += sizeof(missed_events);
4589 local_add(RB_MISSED_EVENTS, &bpage->commit);
4593 * This page may be off to user land. Zero it out here.
4595 if (commit < BUF_PAGE_SIZE)
4596 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4598 out_unlock:
4599 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4601 out:
4602 return ret;
4604 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4607 * We only allocate new buffers, never free them if the CPU goes down.
4608 * If we were to free the buffer, then the user would lose any trace that was in
4609 * the buffer.
4611 int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
4613 struct ring_buffer *buffer;
4614 long nr_pages_same;
4615 int cpu_i;
4616 unsigned long nr_pages;
4618 buffer = container_of(node, struct ring_buffer, node);
4619 if (cpumask_test_cpu(cpu, buffer->cpumask))
4620 return 0;
4622 nr_pages = 0;
4623 nr_pages_same = 1;
4624 /* check if all cpu sizes are same */
4625 for_each_buffer_cpu(buffer, cpu_i) {
4626 /* fill in the size from first enabled cpu */
4627 if (nr_pages == 0)
4628 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4629 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4630 nr_pages_same = 0;
4631 break;
4634 /* allocate minimum pages, user can later expand it */
4635 if (!nr_pages_same)
4636 nr_pages = 2;
4637 buffer->buffers[cpu] =
4638 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4639 if (!buffer->buffers[cpu]) {
4640 WARN(1, "failed to allocate ring buffer on CPU %u\n",
4641 cpu);
4642 return -ENOMEM;
4644 smp_wmb();
4645 cpumask_set_cpu(cpu, buffer->cpumask);
4646 return 0;
4649 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4651 * This is a basic integrity check of the ring buffer.
4652 * Late in the boot cycle this test will run when configured in.
4653 * It will kick off a thread per CPU that will go into a loop
4654 * writing to the per cpu ring buffer various sizes of data.
4655 * Some of the data will be large items, some small.
4657 * Another thread is created that goes into a spin, sending out
4658 * IPIs to the other CPUs to also write into the ring buffer.
4659 * this is to test the nesting ability of the buffer.
4661 * Basic stats are recorded and reported. If something in the
4662 * ring buffer should happen that's not expected, a big warning
4663 * is displayed and all ring buffers are disabled.
4665 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4667 struct rb_test_data {
4668 struct ring_buffer *buffer;
4669 unsigned long events;
4670 unsigned long bytes_written;
4671 unsigned long bytes_alloc;
4672 unsigned long bytes_dropped;
4673 unsigned long events_nested;
4674 unsigned long bytes_written_nested;
4675 unsigned long bytes_alloc_nested;
4676 unsigned long bytes_dropped_nested;
4677 int min_size_nested;
4678 int max_size_nested;
4679 int max_size;
4680 int min_size;
4681 int cpu;
4682 int cnt;
4685 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4687 /* 1 meg per cpu */
4688 #define RB_TEST_BUFFER_SIZE 1048576
4690 static char rb_string[] __initdata =
4691 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4692 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4693 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4695 static bool rb_test_started __initdata;
4697 struct rb_item {
4698 int size;
4699 char str[];
4702 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4704 struct ring_buffer_event *event;
4705 struct rb_item *item;
4706 bool started;
4707 int event_len;
4708 int size;
4709 int len;
4710 int cnt;
4712 /* Have nested writes different that what is written */
4713 cnt = data->cnt + (nested ? 27 : 0);
4715 /* Multiply cnt by ~e, to make some unique increment */
4716 size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4718 len = size + sizeof(struct rb_item);
4720 started = rb_test_started;
4721 /* read rb_test_started before checking buffer enabled */
4722 smp_rmb();
4724 event = ring_buffer_lock_reserve(data->buffer, len);
4725 if (!event) {
4726 /* Ignore dropped events before test starts. */
4727 if (started) {
4728 if (nested)
4729 data->bytes_dropped += len;
4730 else
4731 data->bytes_dropped_nested += len;
4733 return len;
4736 event_len = ring_buffer_event_length(event);
4738 if (RB_WARN_ON(data->buffer, event_len < len))
4739 goto out;
4741 item = ring_buffer_event_data(event);
4742 item->size = size;
4743 memcpy(item->str, rb_string, size);
4745 if (nested) {
4746 data->bytes_alloc_nested += event_len;
4747 data->bytes_written_nested += len;
4748 data->events_nested++;
4749 if (!data->min_size_nested || len < data->min_size_nested)
4750 data->min_size_nested = len;
4751 if (len > data->max_size_nested)
4752 data->max_size_nested = len;
4753 } else {
4754 data->bytes_alloc += event_len;
4755 data->bytes_written += len;
4756 data->events++;
4757 if (!data->min_size || len < data->min_size)
4758 data->max_size = len;
4759 if (len > data->max_size)
4760 data->max_size = len;
4763 out:
4764 ring_buffer_unlock_commit(data->buffer, event);
4766 return 0;
4769 static __init int rb_test(void *arg)
4771 struct rb_test_data *data = arg;
4773 while (!kthread_should_stop()) {
4774 rb_write_something(data, false);
4775 data->cnt++;
4777 set_current_state(TASK_INTERRUPTIBLE);
4778 /* Now sleep between a min of 100-300us and a max of 1ms */
4779 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4782 return 0;
4785 static __init void rb_ipi(void *ignore)
4787 struct rb_test_data *data;
4788 int cpu = smp_processor_id();
4790 data = &rb_data[cpu];
4791 rb_write_something(data, true);
4794 static __init int rb_hammer_test(void *arg)
4796 while (!kthread_should_stop()) {
4798 /* Send an IPI to all cpus to write data! */
4799 smp_call_function(rb_ipi, NULL, 1);
4800 /* No sleep, but for non preempt, let others run */
4801 schedule();
4804 return 0;
4807 static __init int test_ringbuffer(void)
4809 struct task_struct *rb_hammer;
4810 struct ring_buffer *buffer;
4811 int cpu;
4812 int ret = 0;
4814 pr_info("Running ring buffer tests...\n");
4816 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4817 if (WARN_ON(!buffer))
4818 return 0;
4820 /* Disable buffer so that threads can't write to it yet */
4821 ring_buffer_record_off(buffer);
4823 for_each_online_cpu(cpu) {
4824 rb_data[cpu].buffer = buffer;
4825 rb_data[cpu].cpu = cpu;
4826 rb_data[cpu].cnt = cpu;
4827 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4828 "rbtester/%d", cpu);
4829 if (WARN_ON(!rb_threads[cpu])) {
4830 pr_cont("FAILED\n");
4831 ret = -1;
4832 goto out_free;
4835 kthread_bind(rb_threads[cpu], cpu);
4836 wake_up_process(rb_threads[cpu]);
4839 /* Now create the rb hammer! */
4840 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4841 if (WARN_ON(!rb_hammer)) {
4842 pr_cont("FAILED\n");
4843 ret = -1;
4844 goto out_free;
4847 ring_buffer_record_on(buffer);
4849 * Show buffer is enabled before setting rb_test_started.
4850 * Yes there's a small race window where events could be
4851 * dropped and the thread wont catch it. But when a ring
4852 * buffer gets enabled, there will always be some kind of
4853 * delay before other CPUs see it. Thus, we don't care about
4854 * those dropped events. We care about events dropped after
4855 * the threads see that the buffer is active.
4857 smp_wmb();
4858 rb_test_started = true;
4860 set_current_state(TASK_INTERRUPTIBLE);
4861 /* Just run for 10 seconds */;
4862 schedule_timeout(10 * HZ);
4864 kthread_stop(rb_hammer);
4866 out_free:
4867 for_each_online_cpu(cpu) {
4868 if (!rb_threads[cpu])
4869 break;
4870 kthread_stop(rb_threads[cpu]);
4872 if (ret) {
4873 ring_buffer_free(buffer);
4874 return ret;
4877 /* Report! */
4878 pr_info("finished\n");
4879 for_each_online_cpu(cpu) {
4880 struct ring_buffer_event *event;
4881 struct rb_test_data *data = &rb_data[cpu];
4882 struct rb_item *item;
4883 unsigned long total_events;
4884 unsigned long total_dropped;
4885 unsigned long total_written;
4886 unsigned long total_alloc;
4887 unsigned long total_read = 0;
4888 unsigned long total_size = 0;
4889 unsigned long total_len = 0;
4890 unsigned long total_lost = 0;
4891 unsigned long lost;
4892 int big_event_size;
4893 int small_event_size;
4895 ret = -1;
4897 total_events = data->events + data->events_nested;
4898 total_written = data->bytes_written + data->bytes_written_nested;
4899 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
4900 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
4902 big_event_size = data->max_size + data->max_size_nested;
4903 small_event_size = data->min_size + data->min_size_nested;
4905 pr_info("CPU %d:\n", cpu);
4906 pr_info(" events: %ld\n", total_events);
4907 pr_info(" dropped bytes: %ld\n", total_dropped);
4908 pr_info(" alloced bytes: %ld\n", total_alloc);
4909 pr_info(" written bytes: %ld\n", total_written);
4910 pr_info(" biggest event: %d\n", big_event_size);
4911 pr_info(" smallest event: %d\n", small_event_size);
4913 if (RB_WARN_ON(buffer, total_dropped))
4914 break;
4916 ret = 0;
4918 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
4919 total_lost += lost;
4920 item = ring_buffer_event_data(event);
4921 total_len += ring_buffer_event_length(event);
4922 total_size += item->size + sizeof(struct rb_item);
4923 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
4924 pr_info("FAILED!\n");
4925 pr_info("buffer had: %.*s\n", item->size, item->str);
4926 pr_info("expected: %.*s\n", item->size, rb_string);
4927 RB_WARN_ON(buffer, 1);
4928 ret = -1;
4929 break;
4931 total_read++;
4933 if (ret)
4934 break;
4936 ret = -1;
4938 pr_info(" read events: %ld\n", total_read);
4939 pr_info(" lost events: %ld\n", total_lost);
4940 pr_info(" total events: %ld\n", total_lost + total_read);
4941 pr_info(" recorded len bytes: %ld\n", total_len);
4942 pr_info(" recorded size bytes: %ld\n", total_size);
4943 if (total_lost)
4944 pr_info(" With dropped events, record len and size may not match\n"
4945 " alloced and written from above\n");
4946 if (!total_lost) {
4947 if (RB_WARN_ON(buffer, total_len != total_alloc ||
4948 total_size != total_written))
4949 break;
4951 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
4952 break;
4954 ret = 0;
4956 if (!ret)
4957 pr_info("Ring buffer PASSED!\n");
4959 ring_buffer_free(buffer);
4960 return 0;
4963 late_initcall(test_ringbuffer);
4964 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */