4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 #include <linux/ftrace_event.h>
7 #include <linux/ring_buffer.h>
8 #include <linux/trace_clock.h>
9 #include <linux/trace_seq.h>
10 #include <linux/spinlock.h>
11 #include <linux/irq_work.h>
12 #include <linux/uaccess.h>
13 #include <linux/hardirq.h>
14 #include <linux/kthread.h> /* for self test */
15 #include <linux/kmemcheck.h>
16 #include <linux/module.h>
17 #include <linux/percpu.h>
18 #include <linux/mutex.h>
19 #include <linux/delay.h>
20 #include <linux/slab.h>
21 #include <linux/init.h>
22 #include <linux/hash.h>
23 #include <linux/list.h>
24 #include <linux/cpu.h>
26 #include <asm/local.h>
28 static void update_pages_handler(struct work_struct *work);
31 * The ring buffer header is special. We must manually up keep it.
33 int ring_buffer_print_entry_header(struct trace_seq *s)
35 trace_seq_puts(s, "# compressed entry header\n");
36 trace_seq_puts(s, "\ttype_len : 5 bits\n");
37 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
38 trace_seq_puts(s, "\tarray : 32 bits\n");
39 trace_seq_putc(s, '\n');
40 trace_seq_printf(s, "\tpadding : type == %d\n",
41 RINGBUF_TYPE_PADDING);
42 trace_seq_printf(s, "\ttime_extend : type == %d\n",
43 RINGBUF_TYPE_TIME_EXTEND);
44 trace_seq_printf(s, "\tdata max type_len == %d\n",
45 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
47 return !trace_seq_has_overflowed(s);
51 * The ring buffer is made up of a list of pages. A separate list of pages is
52 * allocated for each CPU. A writer may only write to a buffer that is
53 * associated with the CPU it is currently executing on. A reader may read
54 * from any per cpu buffer.
56 * The reader is special. For each per cpu buffer, the reader has its own
57 * reader page. When a reader has read the entire reader page, this reader
58 * page is swapped with another page in the ring buffer.
60 * Now, as long as the writer is off the reader page, the reader can do what
61 * ever it wants with that page. The writer will never write to that page
62 * again (as long as it is out of the ring buffer).
64 * Here's some silly ASCII art.
67 * |reader| RING BUFFER
69 * +------+ +---+ +---+ +---+
78 * |reader| RING BUFFER
79 * |page |------------------v
80 * +------+ +---+ +---+ +---+
89 * |reader| RING BUFFER
90 * |page |------------------v
91 * +------+ +---+ +---+ +---+
96 * +------------------------------+
100 * |buffer| RING BUFFER
101 * |page |------------------v
102 * +------+ +---+ +---+ +---+
104 * | New +---+ +---+ +---+
107 * +------------------------------+
110 * After we make this swap, the reader can hand this page off to the splice
111 * code and be done with it. It can even allocate a new page if it needs to
112 * and swap that into the ring buffer.
114 * We will be using cmpxchg soon to make all this lockless.
119 * A fast way to enable or disable all ring buffers is to
120 * call tracing_on or tracing_off. Turning off the ring buffers
121 * prevents all ring buffers from being recorded to.
122 * Turning this switch on, makes it OK to write to the
123 * ring buffer, if the ring buffer is enabled itself.
125 * There's three layers that must be on in order to write
126 * to the ring buffer.
128 * 1) This global flag must be set.
129 * 2) The ring buffer must be enabled for recording.
130 * 3) The per cpu buffer must be enabled for recording.
132 * In case of an anomaly, this global flag has a bit set that
133 * will permantly disable all ring buffers.
137 * Global flag to disable all recording to ring buffers
138 * This has two bits: ON, DISABLED
142 * 0 0 : ring buffers are off
143 * 1 0 : ring buffers are on
144 * X 1 : ring buffers are permanently disabled
148 RB_BUFFERS_ON_BIT = 0,
149 RB_BUFFERS_DISABLED_BIT = 1,
153 RB_BUFFERS_ON = 1 << RB_BUFFERS_ON_BIT,
154 RB_BUFFERS_DISABLED = 1 << RB_BUFFERS_DISABLED_BIT,
157 static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON;
159 /* Used for individual buffers (after the counter) */
160 #define RB_BUFFER_OFF (1 << 20)
162 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
165 * tracing_off_permanent - permanently disable ring buffers
167 * This function, once called, will disable all ring buffers
170 void tracing_off_permanent(void)
172 set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags);
175 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
176 #define RB_ALIGNMENT 4U
177 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
178 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
180 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
181 # define RB_FORCE_8BYTE_ALIGNMENT 0
182 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
184 # define RB_FORCE_8BYTE_ALIGNMENT 1
185 # define RB_ARCH_ALIGNMENT 8U
188 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
190 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
191 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
194 RB_LEN_TIME_EXTEND = 8,
195 RB_LEN_TIME_STAMP = 16,
198 #define skip_time_extend(event) \
199 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
201 static inline int rb_null_event(struct ring_buffer_event *event)
203 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
206 static void rb_event_set_padding(struct ring_buffer_event *event)
208 /* padding has a NULL time_delta */
209 event->type_len = RINGBUF_TYPE_PADDING;
210 event->time_delta = 0;
214 rb_event_data_length(struct ring_buffer_event *event)
219 length = event->type_len * RB_ALIGNMENT;
221 length = event->array[0];
222 return length + RB_EVNT_HDR_SIZE;
226 * Return the length of the given event. Will return
227 * the length of the time extend if the event is a
230 static inline unsigned
231 rb_event_length(struct ring_buffer_event *event)
233 switch (event->type_len) {
234 case RINGBUF_TYPE_PADDING:
235 if (rb_null_event(event))
238 return event->array[0] + RB_EVNT_HDR_SIZE;
240 case RINGBUF_TYPE_TIME_EXTEND:
241 return RB_LEN_TIME_EXTEND;
243 case RINGBUF_TYPE_TIME_STAMP:
244 return RB_LEN_TIME_STAMP;
246 case RINGBUF_TYPE_DATA:
247 return rb_event_data_length(event);
256 * Return total length of time extend and data,
257 * or just the event length for all other events.
259 static inline unsigned
260 rb_event_ts_length(struct ring_buffer_event *event)
264 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
265 /* time extends include the data event after it */
266 len = RB_LEN_TIME_EXTEND;
267 event = skip_time_extend(event);
269 return len + rb_event_length(event);
273 * ring_buffer_event_length - return the length of the event
274 * @event: the event to get the length of
276 * Returns the size of the data load of a data event.
277 * If the event is something other than a data event, it
278 * returns the size of the event itself. With the exception
279 * of a TIME EXTEND, where it still returns the size of the
280 * data load of the data event after it.
282 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
286 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
287 event = skip_time_extend(event);
289 length = rb_event_length(event);
290 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
292 length -= RB_EVNT_HDR_SIZE;
293 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
294 length -= sizeof(event->array[0]);
297 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
299 /* inline for ring buffer fast paths */
301 rb_event_data(struct ring_buffer_event *event)
303 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
304 event = skip_time_extend(event);
305 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
306 /* If length is in len field, then array[0] has the data */
308 return (void *)&event->array[0];
309 /* Otherwise length is in array[0] and array[1] has the data */
310 return (void *)&event->array[1];
314 * ring_buffer_event_data - return the data of the event
315 * @event: the event to get the data from
317 void *ring_buffer_event_data(struct ring_buffer_event *event)
319 return rb_event_data(event);
321 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
323 #define for_each_buffer_cpu(buffer, cpu) \
324 for_each_cpu(cpu, buffer->cpumask)
327 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
328 #define TS_DELTA_TEST (~TS_MASK)
330 /* Flag when events were overwritten */
331 #define RB_MISSED_EVENTS (1 << 31)
332 /* Missed count stored at end */
333 #define RB_MISSED_STORED (1 << 30)
335 struct buffer_data_page {
336 u64 time_stamp; /* page time stamp */
337 local_t commit; /* write committed index */
338 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
342 * Note, the buffer_page list must be first. The buffer pages
343 * are allocated in cache lines, which means that each buffer
344 * page will be at the beginning of a cache line, and thus
345 * the least significant bits will be zero. We use this to
346 * add flags in the list struct pointers, to make the ring buffer
350 struct list_head list; /* list of buffer pages */
351 local_t write; /* index for next write */
352 unsigned read; /* index for next read */
353 local_t entries; /* entries on this page */
354 unsigned long real_end; /* real end of data */
355 struct buffer_data_page *page; /* Actual data page */
359 * The buffer page counters, write and entries, must be reset
360 * atomically when crossing page boundaries. To synchronize this
361 * update, two counters are inserted into the number. One is
362 * the actual counter for the write position or count on the page.
364 * The other is a counter of updaters. Before an update happens
365 * the update partition of the counter is incremented. This will
366 * allow the updater to update the counter atomically.
368 * The counter is 20 bits, and the state data is 12.
370 #define RB_WRITE_MASK 0xfffff
371 #define RB_WRITE_INTCNT (1 << 20)
373 static void rb_init_page(struct buffer_data_page *bpage)
375 local_set(&bpage->commit, 0);
379 * ring_buffer_page_len - the size of data on the page.
380 * @page: The page to read
382 * Returns the amount of data on the page, including buffer page header.
384 size_t ring_buffer_page_len(void *page)
386 return local_read(&((struct buffer_data_page *)page)->commit)
391 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
394 static void free_buffer_page(struct buffer_page *bpage)
396 free_page((unsigned long)bpage->page);
401 * We need to fit the time_stamp delta into 27 bits.
403 static inline int test_time_stamp(u64 delta)
405 if (delta & TS_DELTA_TEST)
410 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
412 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
413 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
415 int ring_buffer_print_page_header(struct trace_seq *s)
417 struct buffer_data_page field;
419 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
420 "offset:0;\tsize:%u;\tsigned:%u;\n",
421 (unsigned int)sizeof(field.time_stamp),
422 (unsigned int)is_signed_type(u64));
424 trace_seq_printf(s, "\tfield: local_t commit;\t"
425 "offset:%u;\tsize:%u;\tsigned:%u;\n",
426 (unsigned int)offsetof(typeof(field), commit),
427 (unsigned int)sizeof(field.commit),
428 (unsigned int)is_signed_type(long));
430 trace_seq_printf(s, "\tfield: int overwrite;\t"
431 "offset:%u;\tsize:%u;\tsigned:%u;\n",
432 (unsigned int)offsetof(typeof(field), commit),
434 (unsigned int)is_signed_type(long));
436 trace_seq_printf(s, "\tfield: char data;\t"
437 "offset:%u;\tsize:%u;\tsigned:%u;\n",
438 (unsigned int)offsetof(typeof(field), data),
439 (unsigned int)BUF_PAGE_SIZE,
440 (unsigned int)is_signed_type(char));
442 return !trace_seq_has_overflowed(s);
446 struct irq_work work;
447 wait_queue_head_t waiters;
448 bool waiters_pending;
452 * head_page == tail_page && head == tail then buffer is empty.
454 struct ring_buffer_per_cpu {
456 atomic_t record_disabled;
457 struct ring_buffer *buffer;
458 raw_spinlock_t reader_lock; /* serialize readers */
459 arch_spinlock_t lock;
460 struct lock_class_key lock_key;
461 unsigned int nr_pages;
462 struct list_head *pages;
463 struct buffer_page *head_page; /* read from head */
464 struct buffer_page *tail_page; /* write to tail */
465 struct buffer_page *commit_page; /* committed pages */
466 struct buffer_page *reader_page;
467 unsigned long lost_events;
468 unsigned long last_overrun;
469 local_t entries_bytes;
472 local_t commit_overrun;
473 local_t dropped_events;
477 unsigned long read_bytes;
480 /* ring buffer pages to update, > 0 to add, < 0 to remove */
481 int nr_pages_to_update;
482 struct list_head new_pages; /* new pages to add */
483 struct work_struct update_pages_work;
484 struct completion update_done;
486 struct rb_irq_work irq_work;
492 atomic_t record_disabled;
493 atomic_t resize_disabled;
494 cpumask_var_t cpumask;
496 struct lock_class_key *reader_lock_key;
500 struct ring_buffer_per_cpu **buffers;
502 #ifdef CONFIG_HOTPLUG_CPU
503 struct notifier_block cpu_notify;
507 struct rb_irq_work irq_work;
510 struct ring_buffer_iter {
511 struct ring_buffer_per_cpu *cpu_buffer;
513 struct buffer_page *head_page;
514 struct buffer_page *cache_reader_page;
515 unsigned long cache_read;
520 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
522 * Schedules a delayed work to wake up any task that is blocked on the
523 * ring buffer waiters queue.
525 static void rb_wake_up_waiters(struct irq_work *work)
527 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
529 wake_up_all(&rbwork->waiters);
533 * ring_buffer_wait - wait for input to the ring buffer
534 * @buffer: buffer to wait on
535 * @cpu: the cpu buffer to wait on
536 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
538 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
539 * as data is added to any of the @buffer's cpu buffers. Otherwise
540 * it will wait for data to be added to a specific cpu buffer.
542 int ring_buffer_wait(struct ring_buffer *buffer, int cpu, bool full)
544 struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
546 struct rb_irq_work *work;
550 * Depending on what the caller is waiting for, either any
551 * data in any cpu buffer, or a specific buffer, put the
552 * caller on the appropriate wait queue.
554 if (cpu == RING_BUFFER_ALL_CPUS)
555 work = &buffer->irq_work;
557 if (!cpumask_test_cpu(cpu, buffer->cpumask))
559 cpu_buffer = buffer->buffers[cpu];
560 work = &cpu_buffer->irq_work;
565 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
568 * The events can happen in critical sections where
569 * checking a work queue can cause deadlocks.
570 * After adding a task to the queue, this flag is set
571 * only to notify events to try to wake up the queue
574 * We don't clear it even if the buffer is no longer
575 * empty. The flag only causes the next event to run
576 * irq_work to do the work queue wake up. The worse
577 * that can happen if we race with !trace_empty() is that
578 * an event will cause an irq_work to try to wake up
581 * There's no reason to protect this flag either, as
582 * the work queue and irq_work logic will do the necessary
583 * synchronization for the wake ups. The only thing
584 * that is necessary is that the wake up happens after
585 * a task has been queued. It's OK for spurious wake ups.
587 work->waiters_pending = true;
589 if (signal_pending(current)) {
594 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
597 if (cpu != RING_BUFFER_ALL_CPUS &&
598 !ring_buffer_empty_cpu(buffer, cpu)) {
605 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
606 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
607 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
616 finish_wait(&work->waiters, &wait);
622 * ring_buffer_poll_wait - poll on buffer input
623 * @buffer: buffer to wait on
624 * @cpu: the cpu buffer to wait on
625 * @filp: the file descriptor
626 * @poll_table: The poll descriptor
628 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
629 * as data is added to any of the @buffer's cpu buffers. Otherwise
630 * it will wait for data to be added to a specific cpu buffer.
632 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
635 int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
636 struct file *filp, poll_table *poll_table)
638 struct ring_buffer_per_cpu *cpu_buffer;
639 struct rb_irq_work *work;
641 if (cpu == RING_BUFFER_ALL_CPUS)
642 work = &buffer->irq_work;
644 if (!cpumask_test_cpu(cpu, buffer->cpumask))
647 cpu_buffer = buffer->buffers[cpu];
648 work = &cpu_buffer->irq_work;
651 poll_wait(filp, &work->waiters, poll_table);
652 work->waiters_pending = true;
654 * There's a tight race between setting the waiters_pending and
655 * checking if the ring buffer is empty. Once the waiters_pending bit
656 * is set, the next event will wake the task up, but we can get stuck
657 * if there's only a single event in.
659 * FIXME: Ideally, we need a memory barrier on the writer side as well,
660 * but adding a memory barrier to all events will cause too much of a
661 * performance hit in the fast path. We only need a memory barrier when
662 * the buffer goes from empty to having content. But as this race is
663 * extremely small, and it's not a problem if another event comes in, we
668 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
669 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
670 return POLLIN | POLLRDNORM;
674 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
675 #define RB_WARN_ON(b, cond) \
677 int _____ret = unlikely(cond); \
679 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
680 struct ring_buffer_per_cpu *__b = \
682 atomic_inc(&__b->buffer->record_disabled); \
684 atomic_inc(&b->record_disabled); \
690 /* Up this if you want to test the TIME_EXTENTS and normalization */
691 #define DEBUG_SHIFT 0
693 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
695 /* shift to debug/test normalization and TIME_EXTENTS */
696 return buffer->clock() << DEBUG_SHIFT;
699 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
703 preempt_disable_notrace();
704 time = rb_time_stamp(buffer);
705 preempt_enable_no_resched_notrace();
709 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
711 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
714 /* Just stupid testing the normalize function and deltas */
717 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
720 * Making the ring buffer lockless makes things tricky.
721 * Although writes only happen on the CPU that they are on,
722 * and they only need to worry about interrupts. Reads can
725 * The reader page is always off the ring buffer, but when the
726 * reader finishes with a page, it needs to swap its page with
727 * a new one from the buffer. The reader needs to take from
728 * the head (writes go to the tail). But if a writer is in overwrite
729 * mode and wraps, it must push the head page forward.
731 * Here lies the problem.
733 * The reader must be careful to replace only the head page, and
734 * not another one. As described at the top of the file in the
735 * ASCII art, the reader sets its old page to point to the next
736 * page after head. It then sets the page after head to point to
737 * the old reader page. But if the writer moves the head page
738 * during this operation, the reader could end up with the tail.
740 * We use cmpxchg to help prevent this race. We also do something
741 * special with the page before head. We set the LSB to 1.
743 * When the writer must push the page forward, it will clear the
744 * bit that points to the head page, move the head, and then set
745 * the bit that points to the new head page.
747 * We also don't want an interrupt coming in and moving the head
748 * page on another writer. Thus we use the second LSB to catch
751 * head->list->prev->next bit 1 bit 0
754 * Points to head page 0 1
757 * Note we can not trust the prev pointer of the head page, because:
759 * +----+ +-----+ +-----+
760 * | |------>| T |---X--->| N |
762 * +----+ +-----+ +-----+
765 * +----------| R |----------+ |
769 * Key: ---X--> HEAD flag set in pointer
774 * (see __rb_reserve_next() to see where this happens)
776 * What the above shows is that the reader just swapped out
777 * the reader page with a page in the buffer, but before it
778 * could make the new header point back to the new page added
779 * it was preempted by a writer. The writer moved forward onto
780 * the new page added by the reader and is about to move forward
783 * You can see, it is legitimate for the previous pointer of
784 * the head (or any page) not to point back to itself. But only
788 #define RB_PAGE_NORMAL 0UL
789 #define RB_PAGE_HEAD 1UL
790 #define RB_PAGE_UPDATE 2UL
793 #define RB_FLAG_MASK 3UL
795 /* PAGE_MOVED is not part of the mask */
796 #define RB_PAGE_MOVED 4UL
799 * rb_list_head - remove any bit
801 static struct list_head *rb_list_head(struct list_head *list)
803 unsigned long val = (unsigned long)list;
805 return (struct list_head *)(val & ~RB_FLAG_MASK);
809 * rb_is_head_page - test if the given page is the head page
811 * Because the reader may move the head_page pointer, we can
812 * not trust what the head page is (it may be pointing to
813 * the reader page). But if the next page is a header page,
814 * its flags will be non zero.
817 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
818 struct buffer_page *page, struct list_head *list)
822 val = (unsigned long)list->next;
824 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
825 return RB_PAGE_MOVED;
827 return val & RB_FLAG_MASK;
833 * The unique thing about the reader page, is that, if the
834 * writer is ever on it, the previous pointer never points
835 * back to the reader page.
837 static int rb_is_reader_page(struct buffer_page *page)
839 struct list_head *list = page->list.prev;
841 return rb_list_head(list->next) != &page->list;
845 * rb_set_list_to_head - set a list_head to be pointing to head.
847 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
848 struct list_head *list)
852 ptr = (unsigned long *)&list->next;
853 *ptr |= RB_PAGE_HEAD;
854 *ptr &= ~RB_PAGE_UPDATE;
858 * rb_head_page_activate - sets up head page
860 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
862 struct buffer_page *head;
864 head = cpu_buffer->head_page;
869 * Set the previous list pointer to have the HEAD flag.
871 rb_set_list_to_head(cpu_buffer, head->list.prev);
874 static void rb_list_head_clear(struct list_head *list)
876 unsigned long *ptr = (unsigned long *)&list->next;
878 *ptr &= ~RB_FLAG_MASK;
882 * rb_head_page_dactivate - clears head page ptr (for free list)
885 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
887 struct list_head *hd;
889 /* Go through the whole list and clear any pointers found. */
890 rb_list_head_clear(cpu_buffer->pages);
892 list_for_each(hd, cpu_buffer->pages)
893 rb_list_head_clear(hd);
896 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
897 struct buffer_page *head,
898 struct buffer_page *prev,
899 int old_flag, int new_flag)
901 struct list_head *list;
902 unsigned long val = (unsigned long)&head->list;
907 val &= ~RB_FLAG_MASK;
909 ret = cmpxchg((unsigned long *)&list->next,
910 val | old_flag, val | new_flag);
912 /* check if the reader took the page */
913 if ((ret & ~RB_FLAG_MASK) != val)
914 return RB_PAGE_MOVED;
916 return ret & RB_FLAG_MASK;
919 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
920 struct buffer_page *head,
921 struct buffer_page *prev,
924 return rb_head_page_set(cpu_buffer, head, prev,
925 old_flag, RB_PAGE_UPDATE);
928 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
929 struct buffer_page *head,
930 struct buffer_page *prev,
933 return rb_head_page_set(cpu_buffer, head, prev,
934 old_flag, RB_PAGE_HEAD);
937 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
938 struct buffer_page *head,
939 struct buffer_page *prev,
942 return rb_head_page_set(cpu_buffer, head, prev,
943 old_flag, RB_PAGE_NORMAL);
946 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
947 struct buffer_page **bpage)
949 struct list_head *p = rb_list_head((*bpage)->list.next);
951 *bpage = list_entry(p, struct buffer_page, list);
954 static struct buffer_page *
955 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
957 struct buffer_page *head;
958 struct buffer_page *page;
959 struct list_head *list;
962 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
966 list = cpu_buffer->pages;
967 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
970 page = head = cpu_buffer->head_page;
972 * It is possible that the writer moves the header behind
973 * where we started, and we miss in one loop.
974 * A second loop should grab the header, but we'll do
975 * three loops just because I'm paranoid.
977 for (i = 0; i < 3; i++) {
979 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
980 cpu_buffer->head_page = page;
983 rb_inc_page(cpu_buffer, &page);
984 } while (page != head);
987 RB_WARN_ON(cpu_buffer, 1);
992 static int rb_head_page_replace(struct buffer_page *old,
993 struct buffer_page *new)
995 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
999 val = *ptr & ~RB_FLAG_MASK;
1000 val |= RB_PAGE_HEAD;
1002 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1008 * rb_tail_page_update - move the tail page forward
1010 * Returns 1 if moved tail page, 0 if someone else did.
1012 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1013 struct buffer_page *tail_page,
1014 struct buffer_page *next_page)
1016 struct buffer_page *old_tail;
1017 unsigned long old_entries;
1018 unsigned long old_write;
1022 * The tail page now needs to be moved forward.
1024 * We need to reset the tail page, but without messing
1025 * with possible erasing of data brought in by interrupts
1026 * that have moved the tail page and are currently on it.
1028 * We add a counter to the write field to denote this.
1030 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1031 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1034 * Just make sure we have seen our old_write and synchronize
1035 * with any interrupts that come in.
1040 * If the tail page is still the same as what we think
1041 * it is, then it is up to us to update the tail
1044 if (tail_page == cpu_buffer->tail_page) {
1045 /* Zero the write counter */
1046 unsigned long val = old_write & ~RB_WRITE_MASK;
1047 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1050 * This will only succeed if an interrupt did
1051 * not come in and change it. In which case, we
1052 * do not want to modify it.
1054 * We add (void) to let the compiler know that we do not care
1055 * about the return value of these functions. We use the
1056 * cmpxchg to only update if an interrupt did not already
1057 * do it for us. If the cmpxchg fails, we don't care.
1059 (void)local_cmpxchg(&next_page->write, old_write, val);
1060 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1063 * No need to worry about races with clearing out the commit.
1064 * it only can increment when a commit takes place. But that
1065 * only happens in the outer most nested commit.
1067 local_set(&next_page->page->commit, 0);
1069 old_tail = cmpxchg(&cpu_buffer->tail_page,
1070 tail_page, next_page);
1072 if (old_tail == tail_page)
1079 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1080 struct buffer_page *bpage)
1082 unsigned long val = (unsigned long)bpage;
1084 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1091 * rb_check_list - make sure a pointer to a list has the last bits zero
1093 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1094 struct list_head *list)
1096 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1098 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1104 * rb_check_pages - integrity check of buffer pages
1105 * @cpu_buffer: CPU buffer with pages to test
1107 * As a safety measure we check to make sure the data pages have not
1110 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1112 struct list_head *head = cpu_buffer->pages;
1113 struct buffer_page *bpage, *tmp;
1115 /* Reset the head page if it exists */
1116 if (cpu_buffer->head_page)
1117 rb_set_head_page(cpu_buffer);
1119 rb_head_page_deactivate(cpu_buffer);
1121 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1123 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1126 if (rb_check_list(cpu_buffer, head))
1129 list_for_each_entry_safe(bpage, tmp, head, list) {
1130 if (RB_WARN_ON(cpu_buffer,
1131 bpage->list.next->prev != &bpage->list))
1133 if (RB_WARN_ON(cpu_buffer,
1134 bpage->list.prev->next != &bpage->list))
1136 if (rb_check_list(cpu_buffer, &bpage->list))
1140 rb_head_page_activate(cpu_buffer);
1145 static int __rb_allocate_pages(int nr_pages, struct list_head *pages, int cpu)
1148 struct buffer_page *bpage, *tmp;
1150 for (i = 0; i < nr_pages; i++) {
1153 * __GFP_NORETRY flag makes sure that the allocation fails
1154 * gracefully without invoking oom-killer and the system is
1157 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1158 GFP_KERNEL | __GFP_NORETRY,
1163 list_add(&bpage->list, pages);
1165 page = alloc_pages_node(cpu_to_node(cpu),
1166 GFP_KERNEL | __GFP_NORETRY, 0);
1169 bpage->page = page_address(page);
1170 rb_init_page(bpage->page);
1176 list_for_each_entry_safe(bpage, tmp, pages, list) {
1177 list_del_init(&bpage->list);
1178 free_buffer_page(bpage);
1184 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1191 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1195 * The ring buffer page list is a circular list that does not
1196 * start and end with a list head. All page list items point to
1199 cpu_buffer->pages = pages.next;
1202 cpu_buffer->nr_pages = nr_pages;
1204 rb_check_pages(cpu_buffer);
1209 static struct ring_buffer_per_cpu *
1210 rb_allocate_cpu_buffer(struct ring_buffer *buffer, int nr_pages, int cpu)
1212 struct ring_buffer_per_cpu *cpu_buffer;
1213 struct buffer_page *bpage;
1217 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1218 GFP_KERNEL, cpu_to_node(cpu));
1222 cpu_buffer->cpu = cpu;
1223 cpu_buffer->buffer = buffer;
1224 raw_spin_lock_init(&cpu_buffer->reader_lock);
1225 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1226 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1227 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1228 init_completion(&cpu_buffer->update_done);
1229 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1230 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1232 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1233 GFP_KERNEL, cpu_to_node(cpu));
1235 goto fail_free_buffer;
1237 rb_check_bpage(cpu_buffer, bpage);
1239 cpu_buffer->reader_page = bpage;
1240 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1242 goto fail_free_reader;
1243 bpage->page = page_address(page);
1244 rb_init_page(bpage->page);
1246 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1247 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1249 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1251 goto fail_free_reader;
1253 cpu_buffer->head_page
1254 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1255 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1257 rb_head_page_activate(cpu_buffer);
1262 free_buffer_page(cpu_buffer->reader_page);
1269 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1271 struct list_head *head = cpu_buffer->pages;
1272 struct buffer_page *bpage, *tmp;
1274 free_buffer_page(cpu_buffer->reader_page);
1276 rb_head_page_deactivate(cpu_buffer);
1279 list_for_each_entry_safe(bpage, tmp, head, list) {
1280 list_del_init(&bpage->list);
1281 free_buffer_page(bpage);
1283 bpage = list_entry(head, struct buffer_page, list);
1284 free_buffer_page(bpage);
1290 #ifdef CONFIG_HOTPLUG_CPU
1291 static int rb_cpu_notify(struct notifier_block *self,
1292 unsigned long action, void *hcpu);
1296 * __ring_buffer_alloc - allocate a new ring_buffer
1297 * @size: the size in bytes per cpu that is needed.
1298 * @flags: attributes to set for the ring buffer.
1300 * Currently the only flag that is available is the RB_FL_OVERWRITE
1301 * flag. This flag means that the buffer will overwrite old data
1302 * when the buffer wraps. If this flag is not set, the buffer will
1303 * drop data when the tail hits the head.
1305 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1306 struct lock_class_key *key)
1308 struct ring_buffer *buffer;
1312 /* keep it in its own cache line */
1313 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1318 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1319 goto fail_free_buffer;
1321 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1322 buffer->flags = flags;
1323 buffer->clock = trace_clock_local;
1324 buffer->reader_lock_key = key;
1326 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1327 init_waitqueue_head(&buffer->irq_work.waiters);
1329 /* need at least two pages */
1334 * In case of non-hotplug cpu, if the ring-buffer is allocated
1335 * in early initcall, it will not be notified of secondary cpus.
1336 * In that off case, we need to allocate for all possible cpus.
1338 #ifdef CONFIG_HOTPLUG_CPU
1339 cpu_notifier_register_begin();
1340 cpumask_copy(buffer->cpumask, cpu_online_mask);
1342 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1344 buffer->cpus = nr_cpu_ids;
1346 bsize = sizeof(void *) * nr_cpu_ids;
1347 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1349 if (!buffer->buffers)
1350 goto fail_free_cpumask;
1352 for_each_buffer_cpu(buffer, cpu) {
1353 buffer->buffers[cpu] =
1354 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1355 if (!buffer->buffers[cpu])
1356 goto fail_free_buffers;
1359 #ifdef CONFIG_HOTPLUG_CPU
1360 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1361 buffer->cpu_notify.priority = 0;
1362 __register_cpu_notifier(&buffer->cpu_notify);
1363 cpu_notifier_register_done();
1366 mutex_init(&buffer->mutex);
1371 for_each_buffer_cpu(buffer, cpu) {
1372 if (buffer->buffers[cpu])
1373 rb_free_cpu_buffer(buffer->buffers[cpu]);
1375 kfree(buffer->buffers);
1378 free_cpumask_var(buffer->cpumask);
1379 #ifdef CONFIG_HOTPLUG_CPU
1380 cpu_notifier_register_done();
1387 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1390 * ring_buffer_free - free a ring buffer.
1391 * @buffer: the buffer to free.
1394 ring_buffer_free(struct ring_buffer *buffer)
1398 #ifdef CONFIG_HOTPLUG_CPU
1399 cpu_notifier_register_begin();
1400 __unregister_cpu_notifier(&buffer->cpu_notify);
1403 for_each_buffer_cpu(buffer, cpu)
1404 rb_free_cpu_buffer(buffer->buffers[cpu]);
1406 #ifdef CONFIG_HOTPLUG_CPU
1407 cpu_notifier_register_done();
1410 kfree(buffer->buffers);
1411 free_cpumask_var(buffer->cpumask);
1415 EXPORT_SYMBOL_GPL(ring_buffer_free);
1417 void ring_buffer_set_clock(struct ring_buffer *buffer,
1420 buffer->clock = clock;
1423 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1425 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1427 return local_read(&bpage->entries) & RB_WRITE_MASK;
1430 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1432 return local_read(&bpage->write) & RB_WRITE_MASK;
1436 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned int nr_pages)
1438 struct list_head *tail_page, *to_remove, *next_page;
1439 struct buffer_page *to_remove_page, *tmp_iter_page;
1440 struct buffer_page *last_page, *first_page;
1441 unsigned int nr_removed;
1442 unsigned long head_bit;
1447 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1448 atomic_inc(&cpu_buffer->record_disabled);
1450 * We don't race with the readers since we have acquired the reader
1451 * lock. We also don't race with writers after disabling recording.
1452 * This makes it easy to figure out the first and the last page to be
1453 * removed from the list. We unlink all the pages in between including
1454 * the first and last pages. This is done in a busy loop so that we
1455 * lose the least number of traces.
1456 * The pages are freed after we restart recording and unlock readers.
1458 tail_page = &cpu_buffer->tail_page->list;
1461 * tail page might be on reader page, we remove the next page
1462 * from the ring buffer
1464 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1465 tail_page = rb_list_head(tail_page->next);
1466 to_remove = tail_page;
1468 /* start of pages to remove */
1469 first_page = list_entry(rb_list_head(to_remove->next),
1470 struct buffer_page, list);
1472 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1473 to_remove = rb_list_head(to_remove)->next;
1474 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1477 next_page = rb_list_head(to_remove)->next;
1480 * Now we remove all pages between tail_page and next_page.
1481 * Make sure that we have head_bit value preserved for the
1484 tail_page->next = (struct list_head *)((unsigned long)next_page |
1486 next_page = rb_list_head(next_page);
1487 next_page->prev = tail_page;
1489 /* make sure pages points to a valid page in the ring buffer */
1490 cpu_buffer->pages = next_page;
1492 /* update head page */
1494 cpu_buffer->head_page = list_entry(next_page,
1495 struct buffer_page, list);
1498 * change read pointer to make sure any read iterators reset
1501 cpu_buffer->read = 0;
1503 /* pages are removed, resume tracing and then free the pages */
1504 atomic_dec(&cpu_buffer->record_disabled);
1505 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1507 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1509 /* last buffer page to remove */
1510 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1512 tmp_iter_page = first_page;
1515 to_remove_page = tmp_iter_page;
1516 rb_inc_page(cpu_buffer, &tmp_iter_page);
1518 /* update the counters */
1519 page_entries = rb_page_entries(to_remove_page);
1522 * If something was added to this page, it was full
1523 * since it is not the tail page. So we deduct the
1524 * bytes consumed in ring buffer from here.
1525 * Increment overrun to account for the lost events.
1527 local_add(page_entries, &cpu_buffer->overrun);
1528 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1532 * We have already removed references to this list item, just
1533 * free up the buffer_page and its page
1535 free_buffer_page(to_remove_page);
1538 } while (to_remove_page != last_page);
1540 RB_WARN_ON(cpu_buffer, nr_removed);
1542 return nr_removed == 0;
1546 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1548 struct list_head *pages = &cpu_buffer->new_pages;
1549 int retries, success;
1551 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1553 * We are holding the reader lock, so the reader page won't be swapped
1554 * in the ring buffer. Now we are racing with the writer trying to
1555 * move head page and the tail page.
1556 * We are going to adapt the reader page update process where:
1557 * 1. We first splice the start and end of list of new pages between
1558 * the head page and its previous page.
1559 * 2. We cmpxchg the prev_page->next to point from head page to the
1560 * start of new pages list.
1561 * 3. Finally, we update the head->prev to the end of new list.
1563 * We will try this process 10 times, to make sure that we don't keep
1569 struct list_head *head_page, *prev_page, *r;
1570 struct list_head *last_page, *first_page;
1571 struct list_head *head_page_with_bit;
1573 head_page = &rb_set_head_page(cpu_buffer)->list;
1576 prev_page = head_page->prev;
1578 first_page = pages->next;
1579 last_page = pages->prev;
1581 head_page_with_bit = (struct list_head *)
1582 ((unsigned long)head_page | RB_PAGE_HEAD);
1584 last_page->next = head_page_with_bit;
1585 first_page->prev = prev_page;
1587 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1589 if (r == head_page_with_bit) {
1591 * yay, we replaced the page pointer to our new list,
1592 * now, we just have to update to head page's prev
1593 * pointer to point to end of list
1595 head_page->prev = last_page;
1602 INIT_LIST_HEAD(pages);
1604 * If we weren't successful in adding in new pages, warn and stop
1607 RB_WARN_ON(cpu_buffer, !success);
1608 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1610 /* free pages if they weren't inserted */
1612 struct buffer_page *bpage, *tmp;
1613 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1615 list_del_init(&bpage->list);
1616 free_buffer_page(bpage);
1622 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1626 if (cpu_buffer->nr_pages_to_update > 0)
1627 success = rb_insert_pages(cpu_buffer);
1629 success = rb_remove_pages(cpu_buffer,
1630 -cpu_buffer->nr_pages_to_update);
1633 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1636 static void update_pages_handler(struct work_struct *work)
1638 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1639 struct ring_buffer_per_cpu, update_pages_work);
1640 rb_update_pages(cpu_buffer);
1641 complete(&cpu_buffer->update_done);
1645 * ring_buffer_resize - resize the ring buffer
1646 * @buffer: the buffer to resize.
1647 * @size: the new size.
1648 * @cpu_id: the cpu buffer to resize
1650 * Minimum size is 2 * BUF_PAGE_SIZE.
1652 * Returns 0 on success and < 0 on failure.
1654 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1657 struct ring_buffer_per_cpu *cpu_buffer;
1662 * Always succeed at resizing a non-existent buffer:
1667 /* Make sure the requested buffer exists */
1668 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1669 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1672 size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1673 size *= BUF_PAGE_SIZE;
1675 /* we need a minimum of two pages */
1676 if (size < BUF_PAGE_SIZE * 2)
1677 size = BUF_PAGE_SIZE * 2;
1679 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1682 * Don't succeed if resizing is disabled, as a reader might be
1683 * manipulating the ring buffer and is expecting a sane state while
1686 if (atomic_read(&buffer->resize_disabled))
1689 /* prevent another thread from changing buffer sizes */
1690 mutex_lock(&buffer->mutex);
1692 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1693 /* calculate the pages to update */
1694 for_each_buffer_cpu(buffer, cpu) {
1695 cpu_buffer = buffer->buffers[cpu];
1697 cpu_buffer->nr_pages_to_update = nr_pages -
1698 cpu_buffer->nr_pages;
1700 * nothing more to do for removing pages or no update
1702 if (cpu_buffer->nr_pages_to_update <= 0)
1705 * to add pages, make sure all new pages can be
1706 * allocated without receiving ENOMEM
1708 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1709 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1710 &cpu_buffer->new_pages, cpu)) {
1711 /* not enough memory for new pages */
1719 * Fire off all the required work handlers
1720 * We can't schedule on offline CPUs, but it's not necessary
1721 * since we can change their buffer sizes without any race.
1723 for_each_buffer_cpu(buffer, cpu) {
1724 cpu_buffer = buffer->buffers[cpu];
1725 if (!cpu_buffer->nr_pages_to_update)
1728 /* Can't run something on an offline CPU. */
1729 if (!cpu_online(cpu)) {
1730 rb_update_pages(cpu_buffer);
1731 cpu_buffer->nr_pages_to_update = 0;
1733 schedule_work_on(cpu,
1734 &cpu_buffer->update_pages_work);
1738 /* wait for all the updates to complete */
1739 for_each_buffer_cpu(buffer, cpu) {
1740 cpu_buffer = buffer->buffers[cpu];
1741 if (!cpu_buffer->nr_pages_to_update)
1744 if (cpu_online(cpu))
1745 wait_for_completion(&cpu_buffer->update_done);
1746 cpu_buffer->nr_pages_to_update = 0;
1751 /* Make sure this CPU has been intitialized */
1752 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1755 cpu_buffer = buffer->buffers[cpu_id];
1757 if (nr_pages == cpu_buffer->nr_pages)
1760 cpu_buffer->nr_pages_to_update = nr_pages -
1761 cpu_buffer->nr_pages;
1763 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1764 if (cpu_buffer->nr_pages_to_update > 0 &&
1765 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1766 &cpu_buffer->new_pages, cpu_id)) {
1773 /* Can't run something on an offline CPU. */
1774 if (!cpu_online(cpu_id))
1775 rb_update_pages(cpu_buffer);
1777 schedule_work_on(cpu_id,
1778 &cpu_buffer->update_pages_work);
1779 wait_for_completion(&cpu_buffer->update_done);
1782 cpu_buffer->nr_pages_to_update = 0;
1788 * The ring buffer resize can happen with the ring buffer
1789 * enabled, so that the update disturbs the tracing as little
1790 * as possible. But if the buffer is disabled, we do not need
1791 * to worry about that, and we can take the time to verify
1792 * that the buffer is not corrupt.
1794 if (atomic_read(&buffer->record_disabled)) {
1795 atomic_inc(&buffer->record_disabled);
1797 * Even though the buffer was disabled, we must make sure
1798 * that it is truly disabled before calling rb_check_pages.
1799 * There could have been a race between checking
1800 * record_disable and incrementing it.
1802 synchronize_sched();
1803 for_each_buffer_cpu(buffer, cpu) {
1804 cpu_buffer = buffer->buffers[cpu];
1805 rb_check_pages(cpu_buffer);
1807 atomic_dec(&buffer->record_disabled);
1810 mutex_unlock(&buffer->mutex);
1814 for_each_buffer_cpu(buffer, cpu) {
1815 struct buffer_page *bpage, *tmp;
1817 cpu_buffer = buffer->buffers[cpu];
1818 cpu_buffer->nr_pages_to_update = 0;
1820 if (list_empty(&cpu_buffer->new_pages))
1823 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1825 list_del_init(&bpage->list);
1826 free_buffer_page(bpage);
1829 mutex_unlock(&buffer->mutex);
1832 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1834 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1836 mutex_lock(&buffer->mutex);
1838 buffer->flags |= RB_FL_OVERWRITE;
1840 buffer->flags &= ~RB_FL_OVERWRITE;
1841 mutex_unlock(&buffer->mutex);
1843 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1845 static inline void *
1846 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1848 return bpage->data + index;
1851 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1853 return bpage->page->data + index;
1856 static inline struct ring_buffer_event *
1857 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1859 return __rb_page_index(cpu_buffer->reader_page,
1860 cpu_buffer->reader_page->read);
1863 static inline struct ring_buffer_event *
1864 rb_iter_head_event(struct ring_buffer_iter *iter)
1866 return __rb_page_index(iter->head_page, iter->head);
1869 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1871 return local_read(&bpage->page->commit);
1874 /* Size is determined by what has been committed */
1875 static inline unsigned rb_page_size(struct buffer_page *bpage)
1877 return rb_page_commit(bpage);
1880 static inline unsigned
1881 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1883 return rb_page_commit(cpu_buffer->commit_page);
1886 static inline unsigned
1887 rb_event_index(struct ring_buffer_event *event)
1889 unsigned long addr = (unsigned long)event;
1891 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1895 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1896 struct ring_buffer_event *event)
1898 unsigned long addr = (unsigned long)event;
1899 unsigned long index;
1901 index = rb_event_index(event);
1904 return cpu_buffer->commit_page->page == (void *)addr &&
1905 rb_commit_index(cpu_buffer) == index;
1909 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1911 unsigned long max_count;
1914 * We only race with interrupts and NMIs on this CPU.
1915 * If we own the commit event, then we can commit
1916 * all others that interrupted us, since the interruptions
1917 * are in stack format (they finish before they come
1918 * back to us). This allows us to do a simple loop to
1919 * assign the commit to the tail.
1922 max_count = cpu_buffer->nr_pages * 100;
1924 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1925 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1927 if (RB_WARN_ON(cpu_buffer,
1928 rb_is_reader_page(cpu_buffer->tail_page)))
1930 local_set(&cpu_buffer->commit_page->page->commit,
1931 rb_page_write(cpu_buffer->commit_page));
1932 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1933 cpu_buffer->write_stamp =
1934 cpu_buffer->commit_page->page->time_stamp;
1935 /* add barrier to keep gcc from optimizing too much */
1938 while (rb_commit_index(cpu_buffer) !=
1939 rb_page_write(cpu_buffer->commit_page)) {
1941 local_set(&cpu_buffer->commit_page->page->commit,
1942 rb_page_write(cpu_buffer->commit_page));
1943 RB_WARN_ON(cpu_buffer,
1944 local_read(&cpu_buffer->commit_page->page->commit) &
1949 /* again, keep gcc from optimizing */
1953 * If an interrupt came in just after the first while loop
1954 * and pushed the tail page forward, we will be left with
1955 * a dangling commit that will never go forward.
1957 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1961 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1963 cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1964 cpu_buffer->reader_page->read = 0;
1967 static void rb_inc_iter(struct ring_buffer_iter *iter)
1969 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1972 * The iterator could be on the reader page (it starts there).
1973 * But the head could have moved, since the reader was
1974 * found. Check for this case and assign the iterator
1975 * to the head page instead of next.
1977 if (iter->head_page == cpu_buffer->reader_page)
1978 iter->head_page = rb_set_head_page(cpu_buffer);
1980 rb_inc_page(cpu_buffer, &iter->head_page);
1982 iter->read_stamp = iter->head_page->page->time_stamp;
1986 /* Slow path, do not inline */
1987 static noinline struct ring_buffer_event *
1988 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
1990 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
1992 /* Not the first event on the page? */
1993 if (rb_event_index(event)) {
1994 event->time_delta = delta & TS_MASK;
1995 event->array[0] = delta >> TS_SHIFT;
1997 /* nope, just zero it */
1998 event->time_delta = 0;
1999 event->array[0] = 0;
2002 return skip_time_extend(event);
2006 * rb_update_event - update event type and data
2007 * @event: the event to update
2008 * @type: the type of event
2009 * @length: the size of the event field in the ring buffer
2011 * Update the type and data fields of the event. The length
2012 * is the actual size that is written to the ring buffer,
2013 * and with this, we can determine what to place into the
2017 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2018 struct ring_buffer_event *event, unsigned length,
2019 int add_timestamp, u64 delta)
2021 /* Only a commit updates the timestamp */
2022 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2026 * If we need to add a timestamp, then we
2027 * add it to the start of the resevered space.
2029 if (unlikely(add_timestamp)) {
2030 event = rb_add_time_stamp(event, delta);
2031 length -= RB_LEN_TIME_EXTEND;
2035 event->time_delta = delta;
2036 length -= RB_EVNT_HDR_SIZE;
2037 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2038 event->type_len = 0;
2039 event->array[0] = length;
2041 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2045 * rb_handle_head_page - writer hit the head page
2047 * Returns: +1 to retry page
2052 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2053 struct buffer_page *tail_page,
2054 struct buffer_page *next_page)
2056 struct buffer_page *new_head;
2061 entries = rb_page_entries(next_page);
2064 * The hard part is here. We need to move the head
2065 * forward, and protect against both readers on
2066 * other CPUs and writers coming in via interrupts.
2068 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2072 * type can be one of four:
2073 * NORMAL - an interrupt already moved it for us
2074 * HEAD - we are the first to get here.
2075 * UPDATE - we are the interrupt interrupting
2077 * MOVED - a reader on another CPU moved the next
2078 * pointer to its reader page. Give up
2085 * We changed the head to UPDATE, thus
2086 * it is our responsibility to update
2089 local_add(entries, &cpu_buffer->overrun);
2090 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2093 * The entries will be zeroed out when we move the
2097 /* still more to do */
2100 case RB_PAGE_UPDATE:
2102 * This is an interrupt that interrupt the
2103 * previous update. Still more to do.
2106 case RB_PAGE_NORMAL:
2108 * An interrupt came in before the update
2109 * and processed this for us.
2110 * Nothing left to do.
2115 * The reader is on another CPU and just did
2116 * a swap with our next_page.
2121 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2126 * Now that we are here, the old head pointer is
2127 * set to UPDATE. This will keep the reader from
2128 * swapping the head page with the reader page.
2129 * The reader (on another CPU) will spin till
2132 * We just need to protect against interrupts
2133 * doing the job. We will set the next pointer
2134 * to HEAD. After that, we set the old pointer
2135 * to NORMAL, but only if it was HEAD before.
2136 * otherwise we are an interrupt, and only
2137 * want the outer most commit to reset it.
2139 new_head = next_page;
2140 rb_inc_page(cpu_buffer, &new_head);
2142 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2146 * Valid returns are:
2147 * HEAD - an interrupt came in and already set it.
2148 * NORMAL - One of two things:
2149 * 1) We really set it.
2150 * 2) A bunch of interrupts came in and moved
2151 * the page forward again.
2155 case RB_PAGE_NORMAL:
2159 RB_WARN_ON(cpu_buffer, 1);
2164 * It is possible that an interrupt came in,
2165 * set the head up, then more interrupts came in
2166 * and moved it again. When we get back here,
2167 * the page would have been set to NORMAL but we
2168 * just set it back to HEAD.
2170 * How do you detect this? Well, if that happened
2171 * the tail page would have moved.
2173 if (ret == RB_PAGE_NORMAL) {
2175 * If the tail had moved passed next, then we need
2176 * to reset the pointer.
2178 if (cpu_buffer->tail_page != tail_page &&
2179 cpu_buffer->tail_page != next_page)
2180 rb_head_page_set_normal(cpu_buffer, new_head,
2186 * If this was the outer most commit (the one that
2187 * changed the original pointer from HEAD to UPDATE),
2188 * then it is up to us to reset it to NORMAL.
2190 if (type == RB_PAGE_HEAD) {
2191 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2194 if (RB_WARN_ON(cpu_buffer,
2195 ret != RB_PAGE_UPDATE))
2202 static unsigned rb_calculate_event_length(unsigned length)
2204 struct ring_buffer_event event; /* Used only for sizeof array */
2206 /* zero length can cause confusions */
2210 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2211 length += sizeof(event.array[0]);
2213 length += RB_EVNT_HDR_SIZE;
2214 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2220 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2221 struct buffer_page *tail_page,
2222 unsigned long tail, unsigned long length)
2224 struct ring_buffer_event *event;
2227 * Only the event that crossed the page boundary
2228 * must fill the old tail_page with padding.
2230 if (tail >= BUF_PAGE_SIZE) {
2232 * If the page was filled, then we still need
2233 * to update the real_end. Reset it to zero
2234 * and the reader will ignore it.
2236 if (tail == BUF_PAGE_SIZE)
2237 tail_page->real_end = 0;
2239 local_sub(length, &tail_page->write);
2243 event = __rb_page_index(tail_page, tail);
2244 kmemcheck_annotate_bitfield(event, bitfield);
2246 /* account for padding bytes */
2247 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2250 * Save the original length to the meta data.
2251 * This will be used by the reader to add lost event
2254 tail_page->real_end = tail;
2257 * If this event is bigger than the minimum size, then
2258 * we need to be careful that we don't subtract the
2259 * write counter enough to allow another writer to slip
2261 * We put in a discarded commit instead, to make sure
2262 * that this space is not used again.
2264 * If we are less than the minimum size, we don't need to
2267 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2268 /* No room for any events */
2270 /* Mark the rest of the page with padding */
2271 rb_event_set_padding(event);
2273 /* Set the write back to the previous setting */
2274 local_sub(length, &tail_page->write);
2278 /* Put in a discarded event */
2279 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2280 event->type_len = RINGBUF_TYPE_PADDING;
2281 /* time delta must be non zero */
2282 event->time_delta = 1;
2284 /* Set write to end of buffer */
2285 length = (tail + length) - BUF_PAGE_SIZE;
2286 local_sub(length, &tail_page->write);
2290 * This is the slow path, force gcc not to inline it.
2292 static noinline struct ring_buffer_event *
2293 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2294 unsigned long length, unsigned long tail,
2295 struct buffer_page *tail_page, u64 ts)
2297 struct buffer_page *commit_page = cpu_buffer->commit_page;
2298 struct ring_buffer *buffer = cpu_buffer->buffer;
2299 struct buffer_page *next_page;
2302 next_page = tail_page;
2304 rb_inc_page(cpu_buffer, &next_page);
2307 * If for some reason, we had an interrupt storm that made
2308 * it all the way around the buffer, bail, and warn
2311 if (unlikely(next_page == commit_page)) {
2312 local_inc(&cpu_buffer->commit_overrun);
2317 * This is where the fun begins!
2319 * We are fighting against races between a reader that
2320 * could be on another CPU trying to swap its reader
2321 * page with the buffer head.
2323 * We are also fighting against interrupts coming in and
2324 * moving the head or tail on us as well.
2326 * If the next page is the head page then we have filled
2327 * the buffer, unless the commit page is still on the
2330 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2333 * If the commit is not on the reader page, then
2334 * move the header page.
2336 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2338 * If we are not in overwrite mode,
2339 * this is easy, just stop here.
2341 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2342 local_inc(&cpu_buffer->dropped_events);
2346 ret = rb_handle_head_page(cpu_buffer,
2355 * We need to be careful here too. The
2356 * commit page could still be on the reader
2357 * page. We could have a small buffer, and
2358 * have filled up the buffer with events
2359 * from interrupts and such, and wrapped.
2361 * Note, if the tail page is also the on the
2362 * reader_page, we let it move out.
2364 if (unlikely((cpu_buffer->commit_page !=
2365 cpu_buffer->tail_page) &&
2366 (cpu_buffer->commit_page ==
2367 cpu_buffer->reader_page))) {
2368 local_inc(&cpu_buffer->commit_overrun);
2374 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
2377 * Nested commits always have zero deltas, so
2378 * just reread the time stamp
2380 ts = rb_time_stamp(buffer);
2381 next_page->page->time_stamp = ts;
2386 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2388 /* fail and let the caller try again */
2389 return ERR_PTR(-EAGAIN);
2393 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2398 static struct ring_buffer_event *
2399 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2400 unsigned long length, u64 ts,
2401 u64 delta, int add_timestamp)
2403 struct buffer_page *tail_page;
2404 struct ring_buffer_event *event;
2405 unsigned long tail, write;
2408 * If the time delta since the last event is too big to
2409 * hold in the time field of the event, then we append a
2410 * TIME EXTEND event ahead of the data event.
2412 if (unlikely(add_timestamp))
2413 length += RB_LEN_TIME_EXTEND;
2415 tail_page = cpu_buffer->tail_page;
2416 write = local_add_return(length, &tail_page->write);
2418 /* set write to only the index of the write */
2419 write &= RB_WRITE_MASK;
2420 tail = write - length;
2423 * If this is the first commit on the page, then it has the same
2424 * timestamp as the page itself.
2429 /* See if we shot pass the end of this buffer page */
2430 if (unlikely(write > BUF_PAGE_SIZE))
2431 return rb_move_tail(cpu_buffer, length, tail,
2434 /* We reserved something on the buffer */
2436 event = __rb_page_index(tail_page, tail);
2437 kmemcheck_annotate_bitfield(event, bitfield);
2438 rb_update_event(cpu_buffer, event, length, add_timestamp, delta);
2440 local_inc(&tail_page->entries);
2443 * If this is the first commit on the page, then update
2447 tail_page->page->time_stamp = ts;
2449 /* account for these added bytes */
2450 local_add(length, &cpu_buffer->entries_bytes);
2456 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2457 struct ring_buffer_event *event)
2459 unsigned long new_index, old_index;
2460 struct buffer_page *bpage;
2461 unsigned long index;
2464 new_index = rb_event_index(event);
2465 old_index = new_index + rb_event_ts_length(event);
2466 addr = (unsigned long)event;
2469 bpage = cpu_buffer->tail_page;
2471 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2472 unsigned long write_mask =
2473 local_read(&bpage->write) & ~RB_WRITE_MASK;
2474 unsigned long event_length = rb_event_length(event);
2476 * This is on the tail page. It is possible that
2477 * a write could come in and move the tail page
2478 * and write to the next page. That is fine
2479 * because we just shorten what is on this page.
2481 old_index += write_mask;
2482 new_index += write_mask;
2483 index = local_cmpxchg(&bpage->write, old_index, new_index);
2484 if (index == old_index) {
2485 /* update counters */
2486 local_sub(event_length, &cpu_buffer->entries_bytes);
2491 /* could not discard */
2495 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2497 local_inc(&cpu_buffer->committing);
2498 local_inc(&cpu_buffer->commits);
2501 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2503 unsigned long commits;
2505 if (RB_WARN_ON(cpu_buffer,
2506 !local_read(&cpu_buffer->committing)))
2510 commits = local_read(&cpu_buffer->commits);
2511 /* synchronize with interrupts */
2513 if (local_read(&cpu_buffer->committing) == 1)
2514 rb_set_commit_to_write(cpu_buffer);
2516 local_dec(&cpu_buffer->committing);
2518 /* synchronize with interrupts */
2522 * Need to account for interrupts coming in between the
2523 * updating of the commit page and the clearing of the
2524 * committing counter.
2526 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2527 !local_read(&cpu_buffer->committing)) {
2528 local_inc(&cpu_buffer->committing);
2533 static struct ring_buffer_event *
2534 rb_reserve_next_event(struct ring_buffer *buffer,
2535 struct ring_buffer_per_cpu *cpu_buffer,
2536 unsigned long length)
2538 struct ring_buffer_event *event;
2544 rb_start_commit(cpu_buffer);
2546 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2548 * Due to the ability to swap a cpu buffer from a buffer
2549 * it is possible it was swapped before we committed.
2550 * (committing stops a swap). We check for it here and
2551 * if it happened, we have to fail the write.
2554 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2555 local_dec(&cpu_buffer->committing);
2556 local_dec(&cpu_buffer->commits);
2561 length = rb_calculate_event_length(length);
2567 * We allow for interrupts to reenter here and do a trace.
2568 * If one does, it will cause this original code to loop
2569 * back here. Even with heavy interrupts happening, this
2570 * should only happen a few times in a row. If this happens
2571 * 1000 times in a row, there must be either an interrupt
2572 * storm or we have something buggy.
2575 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2578 ts = rb_time_stamp(cpu_buffer->buffer);
2579 diff = ts - cpu_buffer->write_stamp;
2581 /* make sure this diff is calculated here */
2584 /* Did the write stamp get updated already? */
2585 if (likely(ts >= cpu_buffer->write_stamp)) {
2587 if (unlikely(test_time_stamp(delta))) {
2588 int local_clock_stable = 1;
2589 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2590 local_clock_stable = sched_clock_stable();
2592 WARN_ONCE(delta > (1ULL << 59),
2593 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2594 (unsigned long long)delta,
2595 (unsigned long long)ts,
2596 (unsigned long long)cpu_buffer->write_stamp,
2597 local_clock_stable ? "" :
2598 "If you just came from a suspend/resume,\n"
2599 "please switch to the trace global clock:\n"
2600 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2605 event = __rb_reserve_next(cpu_buffer, length, ts,
2606 delta, add_timestamp);
2607 if (unlikely(PTR_ERR(event) == -EAGAIN))
2616 rb_end_commit(cpu_buffer);
2620 #ifdef CONFIG_TRACING
2623 * The lock and unlock are done within a preempt disable section.
2624 * The current_context per_cpu variable can only be modified
2625 * by the current task between lock and unlock. But it can
2626 * be modified more than once via an interrupt. To pass this
2627 * information from the lock to the unlock without having to
2628 * access the 'in_interrupt()' functions again (which do show
2629 * a bit of overhead in something as critical as function tracing,
2630 * we use a bitmask trick.
2632 * bit 0 = NMI context
2633 * bit 1 = IRQ context
2634 * bit 2 = SoftIRQ context
2635 * bit 3 = normal context.
2637 * This works because this is the order of contexts that can
2638 * preempt other contexts. A SoftIRQ never preempts an IRQ
2641 * When the context is determined, the corresponding bit is
2642 * checked and set (if it was set, then a recursion of that context
2645 * On unlock, we need to clear this bit. To do so, just subtract
2646 * 1 from the current_context and AND it to itself.
2650 * 101 & 100 = 100 (clearing bit zero)
2653 * 1010 & 1001 = 1000 (clearing bit 1)
2655 * The least significant bit can be cleared this way, and it
2656 * just so happens that it is the same bit corresponding to
2657 * the current context.
2659 static DEFINE_PER_CPU(unsigned int, current_context);
2661 static __always_inline int trace_recursive_lock(void)
2663 unsigned int val = this_cpu_read(current_context);
2666 if (in_interrupt()) {
2676 if (unlikely(val & (1 << bit)))
2680 this_cpu_write(current_context, val);
2685 static __always_inline void trace_recursive_unlock(void)
2687 unsigned int val = this_cpu_read(current_context);
2690 val &= this_cpu_read(current_context);
2691 this_cpu_write(current_context, val);
2696 #define trace_recursive_lock() (0)
2697 #define trace_recursive_unlock() do { } while (0)
2702 * ring_buffer_lock_reserve - reserve a part of the buffer
2703 * @buffer: the ring buffer to reserve from
2704 * @length: the length of the data to reserve (excluding event header)
2706 * Returns a reseverd event on the ring buffer to copy directly to.
2707 * The user of this interface will need to get the body to write into
2708 * and can use the ring_buffer_event_data() interface.
2710 * The length is the length of the data needed, not the event length
2711 * which also includes the event header.
2713 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2714 * If NULL is returned, then nothing has been allocated or locked.
2716 struct ring_buffer_event *
2717 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2719 struct ring_buffer_per_cpu *cpu_buffer;
2720 struct ring_buffer_event *event;
2723 if (ring_buffer_flags != RB_BUFFERS_ON)
2726 /* If we are tracing schedule, we don't want to recurse */
2727 preempt_disable_notrace();
2729 if (atomic_read(&buffer->record_disabled))
2732 if (trace_recursive_lock())
2735 cpu = raw_smp_processor_id();
2737 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2740 cpu_buffer = buffer->buffers[cpu];
2742 if (atomic_read(&cpu_buffer->record_disabled))
2745 if (length > BUF_MAX_DATA_SIZE)
2748 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2755 trace_recursive_unlock();
2758 preempt_enable_notrace();
2761 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2764 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2765 struct ring_buffer_event *event)
2770 * The event first in the commit queue updates the
2773 if (rb_event_is_commit(cpu_buffer, event)) {
2775 * A commit event that is first on a page
2776 * updates the write timestamp with the page stamp
2778 if (!rb_event_index(event))
2779 cpu_buffer->write_stamp =
2780 cpu_buffer->commit_page->page->time_stamp;
2781 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2782 delta = event->array[0];
2784 delta += event->time_delta;
2785 cpu_buffer->write_stamp += delta;
2787 cpu_buffer->write_stamp += event->time_delta;
2791 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2792 struct ring_buffer_event *event)
2794 local_inc(&cpu_buffer->entries);
2795 rb_update_write_stamp(cpu_buffer, event);
2796 rb_end_commit(cpu_buffer);
2799 static __always_inline void
2800 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2802 if (buffer->irq_work.waiters_pending) {
2803 buffer->irq_work.waiters_pending = false;
2804 /* irq_work_queue() supplies it's own memory barriers */
2805 irq_work_queue(&buffer->irq_work.work);
2808 if (cpu_buffer->irq_work.waiters_pending) {
2809 cpu_buffer->irq_work.waiters_pending = false;
2810 /* irq_work_queue() supplies it's own memory barriers */
2811 irq_work_queue(&cpu_buffer->irq_work.work);
2816 * ring_buffer_unlock_commit - commit a reserved
2817 * @buffer: The buffer to commit to
2818 * @event: The event pointer to commit.
2820 * This commits the data to the ring buffer, and releases any locks held.
2822 * Must be paired with ring_buffer_lock_reserve.
2824 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2825 struct ring_buffer_event *event)
2827 struct ring_buffer_per_cpu *cpu_buffer;
2828 int cpu = raw_smp_processor_id();
2830 cpu_buffer = buffer->buffers[cpu];
2832 rb_commit(cpu_buffer, event);
2834 rb_wakeups(buffer, cpu_buffer);
2836 trace_recursive_unlock();
2838 preempt_enable_notrace();
2842 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2844 static inline void rb_event_discard(struct ring_buffer_event *event)
2846 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2847 event = skip_time_extend(event);
2849 /* array[0] holds the actual length for the discarded event */
2850 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2851 event->type_len = RINGBUF_TYPE_PADDING;
2852 /* time delta must be non zero */
2853 if (!event->time_delta)
2854 event->time_delta = 1;
2858 * Decrement the entries to the page that an event is on.
2859 * The event does not even need to exist, only the pointer
2860 * to the page it is on. This may only be called before the commit
2864 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2865 struct ring_buffer_event *event)
2867 unsigned long addr = (unsigned long)event;
2868 struct buffer_page *bpage = cpu_buffer->commit_page;
2869 struct buffer_page *start;
2873 /* Do the likely case first */
2874 if (likely(bpage->page == (void *)addr)) {
2875 local_dec(&bpage->entries);
2880 * Because the commit page may be on the reader page we
2881 * start with the next page and check the end loop there.
2883 rb_inc_page(cpu_buffer, &bpage);
2886 if (bpage->page == (void *)addr) {
2887 local_dec(&bpage->entries);
2890 rb_inc_page(cpu_buffer, &bpage);
2891 } while (bpage != start);
2893 /* commit not part of this buffer?? */
2894 RB_WARN_ON(cpu_buffer, 1);
2898 * ring_buffer_commit_discard - discard an event that has not been committed
2899 * @buffer: the ring buffer
2900 * @event: non committed event to discard
2902 * Sometimes an event that is in the ring buffer needs to be ignored.
2903 * This function lets the user discard an event in the ring buffer
2904 * and then that event will not be read later.
2906 * This function only works if it is called before the the item has been
2907 * committed. It will try to free the event from the ring buffer
2908 * if another event has not been added behind it.
2910 * If another event has been added behind it, it will set the event
2911 * up as discarded, and perform the commit.
2913 * If this function is called, do not call ring_buffer_unlock_commit on
2916 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2917 struct ring_buffer_event *event)
2919 struct ring_buffer_per_cpu *cpu_buffer;
2922 /* The event is discarded regardless */
2923 rb_event_discard(event);
2925 cpu = smp_processor_id();
2926 cpu_buffer = buffer->buffers[cpu];
2929 * This must only be called if the event has not been
2930 * committed yet. Thus we can assume that preemption
2931 * is still disabled.
2933 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2935 rb_decrement_entry(cpu_buffer, event);
2936 if (rb_try_to_discard(cpu_buffer, event))
2940 * The commit is still visible by the reader, so we
2941 * must still update the timestamp.
2943 rb_update_write_stamp(cpu_buffer, event);
2945 rb_end_commit(cpu_buffer);
2947 trace_recursive_unlock();
2949 preempt_enable_notrace();
2952 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2955 * ring_buffer_write - write data to the buffer without reserving
2956 * @buffer: The ring buffer to write to.
2957 * @length: The length of the data being written (excluding the event header)
2958 * @data: The data to write to the buffer.
2960 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2961 * one function. If you already have the data to write to the buffer, it
2962 * may be easier to simply call this function.
2964 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2965 * and not the length of the event which would hold the header.
2967 int ring_buffer_write(struct ring_buffer *buffer,
2968 unsigned long length,
2971 struct ring_buffer_per_cpu *cpu_buffer;
2972 struct ring_buffer_event *event;
2977 if (ring_buffer_flags != RB_BUFFERS_ON)
2980 preempt_disable_notrace();
2982 if (atomic_read(&buffer->record_disabled))
2985 cpu = raw_smp_processor_id();
2987 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2990 cpu_buffer = buffer->buffers[cpu];
2992 if (atomic_read(&cpu_buffer->record_disabled))
2995 if (length > BUF_MAX_DATA_SIZE)
2998 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3002 body = rb_event_data(event);
3004 memcpy(body, data, length);
3006 rb_commit(cpu_buffer, event);
3008 rb_wakeups(buffer, cpu_buffer);
3012 preempt_enable_notrace();
3016 EXPORT_SYMBOL_GPL(ring_buffer_write);
3018 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3020 struct buffer_page *reader = cpu_buffer->reader_page;
3021 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3022 struct buffer_page *commit = cpu_buffer->commit_page;
3024 /* In case of error, head will be NULL */
3025 if (unlikely(!head))
3028 return reader->read == rb_page_commit(reader) &&
3029 (commit == reader ||
3031 head->read == rb_page_commit(commit)));
3035 * ring_buffer_record_disable - stop all writes into the buffer
3036 * @buffer: The ring buffer to stop writes to.
3038 * This prevents all writes to the buffer. Any attempt to write
3039 * to the buffer after this will fail and return NULL.
3041 * The caller should call synchronize_sched() after this.
3043 void ring_buffer_record_disable(struct ring_buffer *buffer)
3045 atomic_inc(&buffer->record_disabled);
3047 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3050 * ring_buffer_record_enable - enable writes to the buffer
3051 * @buffer: The ring buffer to enable writes
3053 * Note, multiple disables will need the same number of enables
3054 * to truly enable the writing (much like preempt_disable).
3056 void ring_buffer_record_enable(struct ring_buffer *buffer)
3058 atomic_dec(&buffer->record_disabled);
3060 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3063 * ring_buffer_record_off - stop all writes into the buffer
3064 * @buffer: The ring buffer to stop writes to.
3066 * This prevents all writes to the buffer. Any attempt to write
3067 * to the buffer after this will fail and return NULL.
3069 * This is different than ring_buffer_record_disable() as
3070 * it works like an on/off switch, where as the disable() version
3071 * must be paired with a enable().
3073 void ring_buffer_record_off(struct ring_buffer *buffer)
3076 unsigned int new_rd;
3079 rd = atomic_read(&buffer->record_disabled);
3080 new_rd = rd | RB_BUFFER_OFF;
3081 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3083 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3086 * ring_buffer_record_on - restart writes into the buffer
3087 * @buffer: The ring buffer to start writes to.
3089 * This enables all writes to the buffer that was disabled by
3090 * ring_buffer_record_off().
3092 * This is different than ring_buffer_record_enable() as
3093 * it works like an on/off switch, where as the enable() version
3094 * must be paired with a disable().
3096 void ring_buffer_record_on(struct ring_buffer *buffer)
3099 unsigned int new_rd;
3102 rd = atomic_read(&buffer->record_disabled);
3103 new_rd = rd & ~RB_BUFFER_OFF;
3104 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3106 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3109 * ring_buffer_record_is_on - return true if the ring buffer can write
3110 * @buffer: The ring buffer to see if write is enabled
3112 * Returns true if the ring buffer is in a state that it accepts writes.
3114 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3116 return !atomic_read(&buffer->record_disabled);
3120 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3121 * @buffer: The ring buffer to stop writes to.
3122 * @cpu: The CPU buffer to stop
3124 * This prevents all writes to the buffer. Any attempt to write
3125 * to the buffer after this will fail and return NULL.
3127 * The caller should call synchronize_sched() after this.
3129 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3131 struct ring_buffer_per_cpu *cpu_buffer;
3133 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3136 cpu_buffer = buffer->buffers[cpu];
3137 atomic_inc(&cpu_buffer->record_disabled);
3139 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3142 * ring_buffer_record_enable_cpu - enable writes to the buffer
3143 * @buffer: The ring buffer to enable writes
3144 * @cpu: The CPU to enable.
3146 * Note, multiple disables will need the same number of enables
3147 * to truly enable the writing (much like preempt_disable).
3149 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3151 struct ring_buffer_per_cpu *cpu_buffer;
3153 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3156 cpu_buffer = buffer->buffers[cpu];
3157 atomic_dec(&cpu_buffer->record_disabled);
3159 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3162 * The total entries in the ring buffer is the running counter
3163 * of entries entered into the ring buffer, minus the sum of
3164 * the entries read from the ring buffer and the number of
3165 * entries that were overwritten.
3167 static inline unsigned long
3168 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3170 return local_read(&cpu_buffer->entries) -
3171 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3175 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3176 * @buffer: The ring buffer
3177 * @cpu: The per CPU buffer to read from.
3179 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3181 unsigned long flags;
3182 struct ring_buffer_per_cpu *cpu_buffer;
3183 struct buffer_page *bpage;
3186 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3189 cpu_buffer = buffer->buffers[cpu];
3190 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3192 * if the tail is on reader_page, oldest time stamp is on the reader
3195 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3196 bpage = cpu_buffer->reader_page;
3198 bpage = rb_set_head_page(cpu_buffer);
3200 ret = bpage->page->time_stamp;
3201 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3205 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3208 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3209 * @buffer: The ring buffer
3210 * @cpu: The per CPU buffer to read from.
3212 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3214 struct ring_buffer_per_cpu *cpu_buffer;
3217 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3220 cpu_buffer = buffer->buffers[cpu];
3221 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3225 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3228 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3229 * @buffer: The ring buffer
3230 * @cpu: The per CPU buffer to get the entries from.
3232 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3234 struct ring_buffer_per_cpu *cpu_buffer;
3236 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3239 cpu_buffer = buffer->buffers[cpu];
3241 return rb_num_of_entries(cpu_buffer);
3243 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3246 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3247 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3248 * @buffer: The ring buffer
3249 * @cpu: The per CPU buffer to get the number of overruns from
3251 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3253 struct ring_buffer_per_cpu *cpu_buffer;
3256 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3259 cpu_buffer = buffer->buffers[cpu];
3260 ret = local_read(&cpu_buffer->overrun);
3264 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3267 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3268 * commits failing due to the buffer wrapping around while there are uncommitted
3269 * events, such as during an interrupt storm.
3270 * @buffer: The ring buffer
3271 * @cpu: The per CPU buffer to get the number of overruns from
3274 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3276 struct ring_buffer_per_cpu *cpu_buffer;
3279 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3282 cpu_buffer = buffer->buffers[cpu];
3283 ret = local_read(&cpu_buffer->commit_overrun);
3287 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3290 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3291 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3292 * @buffer: The ring buffer
3293 * @cpu: The per CPU buffer to get the number of overruns from
3296 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3298 struct ring_buffer_per_cpu *cpu_buffer;
3301 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3304 cpu_buffer = buffer->buffers[cpu];
3305 ret = local_read(&cpu_buffer->dropped_events);
3309 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3312 * ring_buffer_read_events_cpu - get the number of events successfully read
3313 * @buffer: The ring buffer
3314 * @cpu: The per CPU buffer to get the number of events read
3317 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3319 struct ring_buffer_per_cpu *cpu_buffer;
3321 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3324 cpu_buffer = buffer->buffers[cpu];
3325 return cpu_buffer->read;
3327 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3330 * ring_buffer_entries - get the number of entries in a buffer
3331 * @buffer: The ring buffer
3333 * Returns the total number of entries in the ring buffer
3336 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3338 struct ring_buffer_per_cpu *cpu_buffer;
3339 unsigned long entries = 0;
3342 /* if you care about this being correct, lock the buffer */
3343 for_each_buffer_cpu(buffer, cpu) {
3344 cpu_buffer = buffer->buffers[cpu];
3345 entries += rb_num_of_entries(cpu_buffer);
3350 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3353 * ring_buffer_overruns - get the number of overruns in buffer
3354 * @buffer: The ring buffer
3356 * Returns the total number of overruns in the ring buffer
3359 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3361 struct ring_buffer_per_cpu *cpu_buffer;
3362 unsigned long overruns = 0;
3365 /* if you care about this being correct, lock the buffer */
3366 for_each_buffer_cpu(buffer, cpu) {
3367 cpu_buffer = buffer->buffers[cpu];
3368 overruns += local_read(&cpu_buffer->overrun);
3373 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3375 static void rb_iter_reset(struct ring_buffer_iter *iter)
3377 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3379 /* Iterator usage is expected to have record disabled */
3380 iter->head_page = cpu_buffer->reader_page;
3381 iter->head = cpu_buffer->reader_page->read;
3383 iter->cache_reader_page = iter->head_page;
3384 iter->cache_read = cpu_buffer->read;
3387 iter->read_stamp = cpu_buffer->read_stamp;
3389 iter->read_stamp = iter->head_page->page->time_stamp;
3393 * ring_buffer_iter_reset - reset an iterator
3394 * @iter: The iterator to reset
3396 * Resets the iterator, so that it will start from the beginning
3399 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3401 struct ring_buffer_per_cpu *cpu_buffer;
3402 unsigned long flags;
3407 cpu_buffer = iter->cpu_buffer;
3409 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3410 rb_iter_reset(iter);
3411 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3413 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3416 * ring_buffer_iter_empty - check if an iterator has no more to read
3417 * @iter: The iterator to check
3419 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3421 struct ring_buffer_per_cpu *cpu_buffer;
3423 cpu_buffer = iter->cpu_buffer;
3425 return iter->head_page == cpu_buffer->commit_page &&
3426 iter->head == rb_commit_index(cpu_buffer);
3428 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3431 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3432 struct ring_buffer_event *event)
3436 switch (event->type_len) {
3437 case RINGBUF_TYPE_PADDING:
3440 case RINGBUF_TYPE_TIME_EXTEND:
3441 delta = event->array[0];
3443 delta += event->time_delta;
3444 cpu_buffer->read_stamp += delta;
3447 case RINGBUF_TYPE_TIME_STAMP:
3448 /* FIXME: not implemented */
3451 case RINGBUF_TYPE_DATA:
3452 cpu_buffer->read_stamp += event->time_delta;
3462 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3463 struct ring_buffer_event *event)
3467 switch (event->type_len) {
3468 case RINGBUF_TYPE_PADDING:
3471 case RINGBUF_TYPE_TIME_EXTEND:
3472 delta = event->array[0];
3474 delta += event->time_delta;
3475 iter->read_stamp += delta;
3478 case RINGBUF_TYPE_TIME_STAMP:
3479 /* FIXME: not implemented */
3482 case RINGBUF_TYPE_DATA:
3483 iter->read_stamp += event->time_delta;
3492 static struct buffer_page *
3493 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3495 struct buffer_page *reader = NULL;
3496 unsigned long overwrite;
3497 unsigned long flags;
3501 local_irq_save(flags);
3502 arch_spin_lock(&cpu_buffer->lock);
3506 * This should normally only loop twice. But because the
3507 * start of the reader inserts an empty page, it causes
3508 * a case where we will loop three times. There should be no
3509 * reason to loop four times (that I know of).
3511 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3516 reader = cpu_buffer->reader_page;
3518 /* If there's more to read, return this page */
3519 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3522 /* Never should we have an index greater than the size */
3523 if (RB_WARN_ON(cpu_buffer,
3524 cpu_buffer->reader_page->read > rb_page_size(reader)))
3527 /* check if we caught up to the tail */
3529 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3532 /* Don't bother swapping if the ring buffer is empty */
3533 if (rb_num_of_entries(cpu_buffer) == 0)
3537 * Reset the reader page to size zero.
3539 local_set(&cpu_buffer->reader_page->write, 0);
3540 local_set(&cpu_buffer->reader_page->entries, 0);
3541 local_set(&cpu_buffer->reader_page->page->commit, 0);
3542 cpu_buffer->reader_page->real_end = 0;
3546 * Splice the empty reader page into the list around the head.
3548 reader = rb_set_head_page(cpu_buffer);
3551 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3552 cpu_buffer->reader_page->list.prev = reader->list.prev;
3555 * cpu_buffer->pages just needs to point to the buffer, it
3556 * has no specific buffer page to point to. Lets move it out
3557 * of our way so we don't accidentally swap it.
3559 cpu_buffer->pages = reader->list.prev;
3561 /* The reader page will be pointing to the new head */
3562 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3565 * We want to make sure we read the overruns after we set up our
3566 * pointers to the next object. The writer side does a
3567 * cmpxchg to cross pages which acts as the mb on the writer
3568 * side. Note, the reader will constantly fail the swap
3569 * while the writer is updating the pointers, so this
3570 * guarantees that the overwrite recorded here is the one we
3571 * want to compare with the last_overrun.
3574 overwrite = local_read(&(cpu_buffer->overrun));
3577 * Here's the tricky part.
3579 * We need to move the pointer past the header page.
3580 * But we can only do that if a writer is not currently
3581 * moving it. The page before the header page has the
3582 * flag bit '1' set if it is pointing to the page we want.
3583 * but if the writer is in the process of moving it
3584 * than it will be '2' or already moved '0'.
3587 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3590 * If we did not convert it, then we must try again.
3596 * Yeah! We succeeded in replacing the page.
3598 * Now make the new head point back to the reader page.
3600 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3601 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3603 /* Finally update the reader page to the new head */
3604 cpu_buffer->reader_page = reader;
3605 rb_reset_reader_page(cpu_buffer);
3607 if (overwrite != cpu_buffer->last_overrun) {
3608 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3609 cpu_buffer->last_overrun = overwrite;
3615 arch_spin_unlock(&cpu_buffer->lock);
3616 local_irq_restore(flags);
3621 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3623 struct ring_buffer_event *event;
3624 struct buffer_page *reader;
3627 reader = rb_get_reader_page(cpu_buffer);
3629 /* This function should not be called when buffer is empty */
3630 if (RB_WARN_ON(cpu_buffer, !reader))
3633 event = rb_reader_event(cpu_buffer);
3635 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3638 rb_update_read_stamp(cpu_buffer, event);
3640 length = rb_event_length(event);
3641 cpu_buffer->reader_page->read += length;
3644 static void rb_advance_iter(struct ring_buffer_iter *iter)
3646 struct ring_buffer_per_cpu *cpu_buffer;
3647 struct ring_buffer_event *event;
3650 cpu_buffer = iter->cpu_buffer;
3653 * Check if we are at the end of the buffer.
3655 if (iter->head >= rb_page_size(iter->head_page)) {
3656 /* discarded commits can make the page empty */
3657 if (iter->head_page == cpu_buffer->commit_page)
3663 event = rb_iter_head_event(iter);
3665 length = rb_event_length(event);
3668 * This should not be called to advance the header if we are
3669 * at the tail of the buffer.
3671 if (RB_WARN_ON(cpu_buffer,
3672 (iter->head_page == cpu_buffer->commit_page) &&
3673 (iter->head + length > rb_commit_index(cpu_buffer))))
3676 rb_update_iter_read_stamp(iter, event);
3678 iter->head += length;
3680 /* check for end of page padding */
3681 if ((iter->head >= rb_page_size(iter->head_page)) &&
3682 (iter->head_page != cpu_buffer->commit_page))
3686 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3688 return cpu_buffer->lost_events;
3691 static struct ring_buffer_event *
3692 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3693 unsigned long *lost_events)
3695 struct ring_buffer_event *event;
3696 struct buffer_page *reader;
3701 * We repeat when a time extend is encountered.
3702 * Since the time extend is always attached to a data event,
3703 * we should never loop more than once.
3704 * (We never hit the following condition more than twice).
3706 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3709 reader = rb_get_reader_page(cpu_buffer);
3713 event = rb_reader_event(cpu_buffer);
3715 switch (event->type_len) {
3716 case RINGBUF_TYPE_PADDING:
3717 if (rb_null_event(event))
3718 RB_WARN_ON(cpu_buffer, 1);
3720 * Because the writer could be discarding every
3721 * event it creates (which would probably be bad)
3722 * if we were to go back to "again" then we may never
3723 * catch up, and will trigger the warn on, or lock
3724 * the box. Return the padding, and we will release
3725 * the current locks, and try again.
3729 case RINGBUF_TYPE_TIME_EXTEND:
3730 /* Internal data, OK to advance */
3731 rb_advance_reader(cpu_buffer);
3734 case RINGBUF_TYPE_TIME_STAMP:
3735 /* FIXME: not implemented */
3736 rb_advance_reader(cpu_buffer);
3739 case RINGBUF_TYPE_DATA:
3741 *ts = cpu_buffer->read_stamp + event->time_delta;
3742 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3743 cpu_buffer->cpu, ts);
3746 *lost_events = rb_lost_events(cpu_buffer);
3755 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3757 static struct ring_buffer_event *
3758 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3760 struct ring_buffer *buffer;
3761 struct ring_buffer_per_cpu *cpu_buffer;
3762 struct ring_buffer_event *event;
3765 cpu_buffer = iter->cpu_buffer;
3766 buffer = cpu_buffer->buffer;
3769 * Check if someone performed a consuming read to
3770 * the buffer. A consuming read invalidates the iterator
3771 * and we need to reset the iterator in this case.
3773 if (unlikely(iter->cache_read != cpu_buffer->read ||
3774 iter->cache_reader_page != cpu_buffer->reader_page))
3775 rb_iter_reset(iter);
3778 if (ring_buffer_iter_empty(iter))
3782 * We repeat when a time extend is encountered or we hit
3783 * the end of the page. Since the time extend is always attached
3784 * to a data event, we should never loop more than three times.
3785 * Once for going to next page, once on time extend, and
3786 * finally once to get the event.
3787 * (We never hit the following condition more than thrice).
3789 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3792 if (rb_per_cpu_empty(cpu_buffer))
3795 if (iter->head >= rb_page_size(iter->head_page)) {
3800 event = rb_iter_head_event(iter);
3802 switch (event->type_len) {
3803 case RINGBUF_TYPE_PADDING:
3804 if (rb_null_event(event)) {
3808 rb_advance_iter(iter);
3811 case RINGBUF_TYPE_TIME_EXTEND:
3812 /* Internal data, OK to advance */
3813 rb_advance_iter(iter);
3816 case RINGBUF_TYPE_TIME_STAMP:
3817 /* FIXME: not implemented */
3818 rb_advance_iter(iter);
3821 case RINGBUF_TYPE_DATA:
3823 *ts = iter->read_stamp + event->time_delta;
3824 ring_buffer_normalize_time_stamp(buffer,
3825 cpu_buffer->cpu, ts);
3835 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3837 static inline int rb_ok_to_lock(void)
3840 * If an NMI die dumps out the content of the ring buffer
3841 * do not grab locks. We also permanently disable the ring
3842 * buffer too. A one time deal is all you get from reading
3843 * the ring buffer from an NMI.
3845 if (likely(!in_nmi()))
3848 tracing_off_permanent();
3853 * ring_buffer_peek - peek at the next event to be read
3854 * @buffer: The ring buffer to read
3855 * @cpu: The cpu to peak at
3856 * @ts: The timestamp counter of this event.
3857 * @lost_events: a variable to store if events were lost (may be NULL)
3859 * This will return the event that will be read next, but does
3860 * not consume the data.
3862 struct ring_buffer_event *
3863 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3864 unsigned long *lost_events)
3866 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3867 struct ring_buffer_event *event;
3868 unsigned long flags;
3871 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3874 dolock = rb_ok_to_lock();
3876 local_irq_save(flags);
3878 raw_spin_lock(&cpu_buffer->reader_lock);
3879 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3880 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3881 rb_advance_reader(cpu_buffer);
3883 raw_spin_unlock(&cpu_buffer->reader_lock);
3884 local_irq_restore(flags);
3886 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3893 * ring_buffer_iter_peek - peek at the next event to be read
3894 * @iter: The ring buffer iterator
3895 * @ts: The timestamp counter of this event.
3897 * This will return the event that will be read next, but does
3898 * not increment the iterator.
3900 struct ring_buffer_event *
3901 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3903 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3904 struct ring_buffer_event *event;
3905 unsigned long flags;
3908 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3909 event = rb_iter_peek(iter, ts);
3910 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3912 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3919 * ring_buffer_consume - return an event and consume it
3920 * @buffer: The ring buffer to get the next event from
3921 * @cpu: the cpu to read the buffer from
3922 * @ts: a variable to store the timestamp (may be NULL)
3923 * @lost_events: a variable to store if events were lost (may be NULL)
3925 * Returns the next event in the ring buffer, and that event is consumed.
3926 * Meaning, that sequential reads will keep returning a different event,
3927 * and eventually empty the ring buffer if the producer is slower.
3929 struct ring_buffer_event *
3930 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3931 unsigned long *lost_events)
3933 struct ring_buffer_per_cpu *cpu_buffer;
3934 struct ring_buffer_event *event = NULL;
3935 unsigned long flags;
3938 dolock = rb_ok_to_lock();
3941 /* might be called in atomic */
3944 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3947 cpu_buffer = buffer->buffers[cpu];
3948 local_irq_save(flags);
3950 raw_spin_lock(&cpu_buffer->reader_lock);
3952 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3954 cpu_buffer->lost_events = 0;
3955 rb_advance_reader(cpu_buffer);
3959 raw_spin_unlock(&cpu_buffer->reader_lock);
3960 local_irq_restore(flags);
3965 if (event && event->type_len == RINGBUF_TYPE_PADDING)
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
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
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))
4001 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
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);
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.
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
4039 * Must be paired with ring_buffer_read_finish.
4042 ring_buffer_read_start(struct ring_buffer_iter *iter)
4044 struct ring_buffer_per_cpu *cpu_buffer;
4045 unsigned long flags;
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
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);
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);
4105 event = rb_iter_peek(iter, ts);
4109 if (event->type_len == RINGBUF_TYPE_PADDING)
4112 rb_advance_iter(iter);
4114 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
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))
4135 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4137 EXPORT_SYMBOL_GPL(ring_buffer_size);
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))
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)))
4205 arch_spin_lock(&cpu_buffer->lock);
4207 rb_reset_cpu(cpu_buffer);
4209 arch_spin_unlock(&cpu_buffer->lock);
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)
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 int ring_buffer_empty(struct ring_buffer *buffer)
4238 struct ring_buffer_per_cpu *cpu_buffer;
4239 unsigned long flags;
4244 dolock = rb_ok_to_lock();
4246 /* yes this is racy, but if you don't like the race, lock the buffer */
4247 for_each_buffer_cpu(buffer, cpu) {
4248 cpu_buffer = buffer->buffers[cpu];
4249 local_irq_save(flags);
4251 raw_spin_lock(&cpu_buffer->reader_lock);
4252 ret = rb_per_cpu_empty(cpu_buffer);
4254 raw_spin_unlock(&cpu_buffer->reader_lock);
4255 local_irq_restore(flags);
4263 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4266 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4267 * @buffer: The ring buffer
4268 * @cpu: The CPU buffer to test
4270 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4272 struct ring_buffer_per_cpu *cpu_buffer;
4273 unsigned long flags;
4277 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4280 dolock = rb_ok_to_lock();
4282 cpu_buffer = buffer->buffers[cpu];
4283 local_irq_save(flags);
4285 raw_spin_lock(&cpu_buffer->reader_lock);
4286 ret = rb_per_cpu_empty(cpu_buffer);
4288 raw_spin_unlock(&cpu_buffer->reader_lock);
4289 local_irq_restore(flags);
4293 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4295 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4297 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4298 * @buffer_a: One buffer to swap with
4299 * @buffer_b: The other buffer to swap with
4301 * This function is useful for tracers that want to take a "snapshot"
4302 * of a CPU buffer and has another back up buffer lying around.
4303 * it is expected that the tracer handles the cpu buffer not being
4304 * used at the moment.
4306 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4307 struct ring_buffer *buffer_b, int cpu)
4309 struct ring_buffer_per_cpu *cpu_buffer_a;
4310 struct ring_buffer_per_cpu *cpu_buffer_b;
4313 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4314 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4317 cpu_buffer_a = buffer_a->buffers[cpu];
4318 cpu_buffer_b = buffer_b->buffers[cpu];
4320 /* At least make sure the two buffers are somewhat the same */
4321 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4326 if (ring_buffer_flags != RB_BUFFERS_ON)
4329 if (atomic_read(&buffer_a->record_disabled))
4332 if (atomic_read(&buffer_b->record_disabled))
4335 if (atomic_read(&cpu_buffer_a->record_disabled))
4338 if (atomic_read(&cpu_buffer_b->record_disabled))
4342 * We can't do a synchronize_sched here because this
4343 * function can be called in atomic context.
4344 * Normally this will be called from the same CPU as cpu.
4345 * If not it's up to the caller to protect this.
4347 atomic_inc(&cpu_buffer_a->record_disabled);
4348 atomic_inc(&cpu_buffer_b->record_disabled);
4351 if (local_read(&cpu_buffer_a->committing))
4353 if (local_read(&cpu_buffer_b->committing))
4356 buffer_a->buffers[cpu] = cpu_buffer_b;
4357 buffer_b->buffers[cpu] = cpu_buffer_a;
4359 cpu_buffer_b->buffer = buffer_a;
4360 cpu_buffer_a->buffer = buffer_b;
4365 atomic_dec(&cpu_buffer_a->record_disabled);
4366 atomic_dec(&cpu_buffer_b->record_disabled);
4370 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4371 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4374 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4375 * @buffer: the buffer to allocate for.
4376 * @cpu: the cpu buffer to allocate.
4378 * This function is used in conjunction with ring_buffer_read_page.
4379 * When reading a full page from the ring buffer, these functions
4380 * can be used to speed up the process. The calling function should
4381 * allocate a few pages first with this function. Then when it
4382 * needs to get pages from the ring buffer, it passes the result
4383 * of this function into ring_buffer_read_page, which will swap
4384 * the page that was allocated, with the read page of the buffer.
4387 * The page allocated, or NULL on error.
4389 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4391 struct buffer_data_page *bpage;
4394 page = alloc_pages_node(cpu_to_node(cpu),
4395 GFP_KERNEL | __GFP_NORETRY, 0);
4399 bpage = page_address(page);
4401 rb_init_page(bpage);
4405 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4408 * ring_buffer_free_read_page - free an allocated read page
4409 * @buffer: the buffer the page was allocate for
4410 * @data: the page to free
4412 * Free a page allocated from ring_buffer_alloc_read_page.
4414 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4416 free_page((unsigned long)data);
4418 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4421 * ring_buffer_read_page - extract a page from the ring buffer
4422 * @buffer: buffer to extract from
4423 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4424 * @len: amount to extract
4425 * @cpu: the cpu of the buffer to extract
4426 * @full: should the extraction only happen when the page is full.
4428 * This function will pull out a page from the ring buffer and consume it.
4429 * @data_page must be the address of the variable that was returned
4430 * from ring_buffer_alloc_read_page. This is because the page might be used
4431 * to swap with a page in the ring buffer.
4434 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4437 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4439 * process_page(rpage, ret);
4441 * When @full is set, the function will not return true unless
4442 * the writer is off the reader page.
4444 * Note: it is up to the calling functions to handle sleeps and wakeups.
4445 * The ring buffer can be used anywhere in the kernel and can not
4446 * blindly call wake_up. The layer that uses the ring buffer must be
4447 * responsible for that.
4450 * >=0 if data has been transferred, returns the offset of consumed data.
4451 * <0 if no data has been transferred.
4453 int ring_buffer_read_page(struct ring_buffer *buffer,
4454 void **data_page, size_t len, int cpu, int full)
4456 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4457 struct ring_buffer_event *event;
4458 struct buffer_data_page *bpage;
4459 struct buffer_page *reader;
4460 unsigned long missed_events;
4461 unsigned long flags;
4462 unsigned int commit;
4467 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4471 * If len is not big enough to hold the page header, then
4472 * we can not copy anything.
4474 if (len <= BUF_PAGE_HDR_SIZE)
4477 len -= BUF_PAGE_HDR_SIZE;
4486 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4488 reader = rb_get_reader_page(cpu_buffer);
4492 event = rb_reader_event(cpu_buffer);
4494 read = reader->read;
4495 commit = rb_page_commit(reader);
4497 /* Check if any events were dropped */
4498 missed_events = cpu_buffer->lost_events;
4501 * If this page has been partially read or
4502 * if len is not big enough to read the rest of the page or
4503 * a writer is still on the page, then
4504 * we must copy the data from the page to the buffer.
4505 * Otherwise, we can simply swap the page with the one passed in.
4507 if (read || (len < (commit - read)) ||
4508 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4509 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4510 unsigned int rpos = read;
4511 unsigned int pos = 0;
4517 if (len > (commit - read))
4518 len = (commit - read);
4520 /* Always keep the time extend and data together */
4521 size = rb_event_ts_length(event);
4526 /* save the current timestamp, since the user will need it */
4527 save_timestamp = cpu_buffer->read_stamp;
4529 /* Need to copy one event at a time */
4531 /* We need the size of one event, because
4532 * rb_advance_reader only advances by one event,
4533 * whereas rb_event_ts_length may include the size of
4534 * one or two events.
4535 * We have already ensured there's enough space if this
4536 * is a time extend. */
4537 size = rb_event_length(event);
4538 memcpy(bpage->data + pos, rpage->data + rpos, size);
4542 rb_advance_reader(cpu_buffer);
4543 rpos = reader->read;
4549 event = rb_reader_event(cpu_buffer);
4550 /* Always keep the time extend and data together */
4551 size = rb_event_ts_length(event);
4552 } while (len >= size);
4555 local_set(&bpage->commit, pos);
4556 bpage->time_stamp = save_timestamp;
4558 /* we copied everything to the beginning */
4561 /* update the entry counter */
4562 cpu_buffer->read += rb_page_entries(reader);
4563 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4565 /* swap the pages */
4566 rb_init_page(bpage);
4567 bpage = reader->page;
4568 reader->page = *data_page;
4569 local_set(&reader->write, 0);
4570 local_set(&reader->entries, 0);
4575 * Use the real_end for the data size,
4576 * This gives us a chance to store the lost events
4579 if (reader->real_end)
4580 local_set(&bpage->commit, reader->real_end);
4584 cpu_buffer->lost_events = 0;
4586 commit = local_read(&bpage->commit);
4588 * Set a flag in the commit field if we lost events
4590 if (missed_events) {
4591 /* If there is room at the end of the page to save the
4592 * missed events, then record it there.
4594 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4595 memcpy(&bpage->data[commit], &missed_events,
4596 sizeof(missed_events));
4597 local_add(RB_MISSED_STORED, &bpage->commit);
4598 commit += sizeof(missed_events);
4600 local_add(RB_MISSED_EVENTS, &bpage->commit);
4604 * This page may be off to user land. Zero it out here.
4606 if (commit < BUF_PAGE_SIZE)
4607 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4610 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4615 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4617 #ifdef CONFIG_HOTPLUG_CPU
4618 static int rb_cpu_notify(struct notifier_block *self,
4619 unsigned long action, void *hcpu)
4621 struct ring_buffer *buffer =
4622 container_of(self, struct ring_buffer, cpu_notify);
4623 long cpu = (long)hcpu;
4624 int cpu_i, nr_pages_same;
4625 unsigned int nr_pages;
4628 case CPU_UP_PREPARE:
4629 case CPU_UP_PREPARE_FROZEN:
4630 if (cpumask_test_cpu(cpu, buffer->cpumask))
4635 /* check if all cpu sizes are same */
4636 for_each_buffer_cpu(buffer, cpu_i) {
4637 /* fill in the size from first enabled cpu */
4639 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4640 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4645 /* allocate minimum pages, user can later expand it */
4648 buffer->buffers[cpu] =
4649 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4650 if (!buffer->buffers[cpu]) {
4651 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4656 cpumask_set_cpu(cpu, buffer->cpumask);
4658 case CPU_DOWN_PREPARE:
4659 case CPU_DOWN_PREPARE_FROZEN:
4662 * If we were to free the buffer, then the user would
4663 * lose any trace that was in the buffer.
4673 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4675 * This is a basic integrity check of the ring buffer.
4676 * Late in the boot cycle this test will run when configured in.
4677 * It will kick off a thread per CPU that will go into a loop
4678 * writing to the per cpu ring buffer various sizes of data.
4679 * Some of the data will be large items, some small.
4681 * Another thread is created that goes into a spin, sending out
4682 * IPIs to the other CPUs to also write into the ring buffer.
4683 * this is to test the nesting ability of the buffer.
4685 * Basic stats are recorded and reported. If something in the
4686 * ring buffer should happen that's not expected, a big warning
4687 * is displayed and all ring buffers are disabled.
4689 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4691 struct rb_test_data {
4692 struct ring_buffer *buffer;
4693 unsigned long events;
4694 unsigned long bytes_written;
4695 unsigned long bytes_alloc;
4696 unsigned long bytes_dropped;
4697 unsigned long events_nested;
4698 unsigned long bytes_written_nested;
4699 unsigned long bytes_alloc_nested;
4700 unsigned long bytes_dropped_nested;
4701 int min_size_nested;
4702 int max_size_nested;
4709 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4712 #define RB_TEST_BUFFER_SIZE 1048576
4714 static char rb_string[] __initdata =
4715 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4716 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4717 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4719 static bool rb_test_started __initdata;
4726 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4728 struct ring_buffer_event *event;
4729 struct rb_item *item;
4736 /* Have nested writes different that what is written */
4737 cnt = data->cnt + (nested ? 27 : 0);
4739 /* Multiply cnt by ~e, to make some unique increment */
4740 size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4742 len = size + sizeof(struct rb_item);
4744 started = rb_test_started;
4745 /* read rb_test_started before checking buffer enabled */
4748 event = ring_buffer_lock_reserve(data->buffer, len);
4750 /* Ignore dropped events before test starts. */
4753 data->bytes_dropped += len;
4755 data->bytes_dropped_nested += len;
4760 event_len = ring_buffer_event_length(event);
4762 if (RB_WARN_ON(data->buffer, event_len < len))
4765 item = ring_buffer_event_data(event);
4767 memcpy(item->str, rb_string, size);
4770 data->bytes_alloc_nested += event_len;
4771 data->bytes_written_nested += len;
4772 data->events_nested++;
4773 if (!data->min_size_nested || len < data->min_size_nested)
4774 data->min_size_nested = len;
4775 if (len > data->max_size_nested)
4776 data->max_size_nested = len;
4778 data->bytes_alloc += event_len;
4779 data->bytes_written += len;
4781 if (!data->min_size || len < data->min_size)
4782 data->max_size = len;
4783 if (len > data->max_size)
4784 data->max_size = len;
4788 ring_buffer_unlock_commit(data->buffer, event);
4793 static __init int rb_test(void *arg)
4795 struct rb_test_data *data = arg;
4797 while (!kthread_should_stop()) {
4798 rb_write_something(data, false);
4801 set_current_state(TASK_INTERRUPTIBLE);
4802 /* Now sleep between a min of 100-300us and a max of 1ms */
4803 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4809 static __init void rb_ipi(void *ignore)
4811 struct rb_test_data *data;
4812 int cpu = smp_processor_id();
4814 data = &rb_data[cpu];
4815 rb_write_something(data, true);
4818 static __init int rb_hammer_test(void *arg)
4820 while (!kthread_should_stop()) {
4822 /* Send an IPI to all cpus to write data! */
4823 smp_call_function(rb_ipi, NULL, 1);
4824 /* No sleep, but for non preempt, let others run */
4831 static __init int test_ringbuffer(void)
4833 struct task_struct *rb_hammer;
4834 struct ring_buffer *buffer;
4838 pr_info("Running ring buffer tests...\n");
4840 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4841 if (WARN_ON(!buffer))
4844 /* Disable buffer so that threads can't write to it yet */
4845 ring_buffer_record_off(buffer);
4847 for_each_online_cpu(cpu) {
4848 rb_data[cpu].buffer = buffer;
4849 rb_data[cpu].cpu = cpu;
4850 rb_data[cpu].cnt = cpu;
4851 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4852 "rbtester/%d", cpu);
4853 if (WARN_ON(!rb_threads[cpu])) {
4854 pr_cont("FAILED\n");
4859 kthread_bind(rb_threads[cpu], cpu);
4860 wake_up_process(rb_threads[cpu]);
4863 /* Now create the rb hammer! */
4864 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4865 if (WARN_ON(!rb_hammer)) {
4866 pr_cont("FAILED\n");
4871 ring_buffer_record_on(buffer);
4873 * Show buffer is enabled before setting rb_test_started.
4874 * Yes there's a small race window where events could be
4875 * dropped and the thread wont catch it. But when a ring
4876 * buffer gets enabled, there will always be some kind of
4877 * delay before other CPUs see it. Thus, we don't care about
4878 * those dropped events. We care about events dropped after
4879 * the threads see that the buffer is active.
4882 rb_test_started = true;
4884 set_current_state(TASK_INTERRUPTIBLE);
4885 /* Just run for 10 seconds */;
4886 schedule_timeout(10 * HZ);
4888 kthread_stop(rb_hammer);
4891 for_each_online_cpu(cpu) {
4892 if (!rb_threads[cpu])
4894 kthread_stop(rb_threads[cpu]);
4897 ring_buffer_free(buffer);
4902 pr_info("finished\n");
4903 for_each_online_cpu(cpu) {
4904 struct ring_buffer_event *event;
4905 struct rb_test_data *data = &rb_data[cpu];
4906 struct rb_item *item;
4907 unsigned long total_events;
4908 unsigned long total_dropped;
4909 unsigned long total_written;
4910 unsigned long total_alloc;
4911 unsigned long total_read = 0;
4912 unsigned long total_size = 0;
4913 unsigned long total_len = 0;
4914 unsigned long total_lost = 0;
4917 int small_event_size;
4921 total_events = data->events + data->events_nested;
4922 total_written = data->bytes_written + data->bytes_written_nested;
4923 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
4924 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
4926 big_event_size = data->max_size + data->max_size_nested;
4927 small_event_size = data->min_size + data->min_size_nested;
4929 pr_info("CPU %d:\n", cpu);
4930 pr_info(" events: %ld\n", total_events);
4931 pr_info(" dropped bytes: %ld\n", total_dropped);
4932 pr_info(" alloced bytes: %ld\n", total_alloc);
4933 pr_info(" written bytes: %ld\n", total_written);
4934 pr_info(" biggest event: %d\n", big_event_size);
4935 pr_info(" smallest event: %d\n", small_event_size);
4937 if (RB_WARN_ON(buffer, total_dropped))
4942 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
4944 item = ring_buffer_event_data(event);
4945 total_len += ring_buffer_event_length(event);
4946 total_size += item->size + sizeof(struct rb_item);
4947 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
4948 pr_info("FAILED!\n");
4949 pr_info("buffer had: %.*s\n", item->size, item->str);
4950 pr_info("expected: %.*s\n", item->size, rb_string);
4951 RB_WARN_ON(buffer, 1);
4962 pr_info(" read events: %ld\n", total_read);
4963 pr_info(" lost events: %ld\n", total_lost);
4964 pr_info(" total events: %ld\n", total_lost + total_read);
4965 pr_info(" recorded len bytes: %ld\n", total_len);
4966 pr_info(" recorded size bytes: %ld\n", total_size);
4968 pr_info(" With dropped events, record len and size may not match\n"
4969 " alloced and written from above\n");
4971 if (RB_WARN_ON(buffer, total_len != total_alloc ||
4972 total_size != total_written))
4975 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
4981 pr_info("Ring buffer PASSED!\n");
4983 ring_buffer_free(buffer);
4987 late_initcall(test_ringbuffer);
4988 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */