4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 #include <linux/trace_events.h>
7 #include <linux/ring_buffer.h>
8 #include <linux/trace_clock.h>
9 #include <linux/trace_seq.h>
10 #include <linux/spinlock.h>
11 #include <linux/irq_work.h>
12 #include <linux/uaccess.h>
13 #include <linux/hardirq.h>
14 #include <linux/kthread.h> /* for self test */
15 #include <linux/kmemcheck.h>
16 #include <linux/module.h>
17 #include <linux/percpu.h>
18 #include <linux/mutex.h>
19 #include <linux/delay.h>
20 #include <linux/slab.h>
21 #include <linux/init.h>
22 #include <linux/hash.h>
23 #include <linux/list.h>
24 #include <linux/cpu.h>
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.
118 /* Used for individual buffers (after the counter) */
119 #define RB_BUFFER_OFF (1 << 20)
121 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
123 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
124 #define RB_ALIGNMENT 4U
125 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
126 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
128 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
129 # define RB_FORCE_8BYTE_ALIGNMENT 0
130 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
132 # define RB_FORCE_8BYTE_ALIGNMENT 1
133 # define RB_ARCH_ALIGNMENT 8U
136 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
138 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
139 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
142 RB_LEN_TIME_EXTEND = 8,
143 RB_LEN_TIME_STAMP = 16,
146 #define skip_time_extend(event) \
147 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
149 static inline int rb_null_event(struct ring_buffer_event *event)
151 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
154 static void rb_event_set_padding(struct ring_buffer_event *event)
156 /* padding has a NULL time_delta */
157 event->type_len = RINGBUF_TYPE_PADDING;
158 event->time_delta = 0;
162 rb_event_data_length(struct ring_buffer_event *event)
167 length = event->type_len * RB_ALIGNMENT;
169 length = event->array[0];
170 return length + RB_EVNT_HDR_SIZE;
174 * Return the length of the given event. Will return
175 * the length of the time extend if the event is a
178 static inline unsigned
179 rb_event_length(struct ring_buffer_event *event)
181 switch (event->type_len) {
182 case RINGBUF_TYPE_PADDING:
183 if (rb_null_event(event))
186 return event->array[0] + RB_EVNT_HDR_SIZE;
188 case RINGBUF_TYPE_TIME_EXTEND:
189 return RB_LEN_TIME_EXTEND;
191 case RINGBUF_TYPE_TIME_STAMP:
192 return RB_LEN_TIME_STAMP;
194 case RINGBUF_TYPE_DATA:
195 return rb_event_data_length(event);
204 * Return total length of time extend and data,
205 * or just the event length for all other events.
207 static inline unsigned
208 rb_event_ts_length(struct ring_buffer_event *event)
212 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
213 /* time extends include the data event after it */
214 len = RB_LEN_TIME_EXTEND;
215 event = skip_time_extend(event);
217 return len + rb_event_length(event);
221 * ring_buffer_event_length - return the length of the event
222 * @event: the event to get the length of
224 * Returns the size of the data load of a data event.
225 * If the event is something other than a data event, it
226 * returns the size of the event itself. With the exception
227 * of a TIME EXTEND, where it still returns the size of the
228 * data load of the data event after it.
230 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
234 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
235 event = skip_time_extend(event);
237 length = rb_event_length(event);
238 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
240 length -= RB_EVNT_HDR_SIZE;
241 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
242 length -= sizeof(event->array[0]);
245 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
247 /* inline for ring buffer fast paths */
249 rb_event_data(struct ring_buffer_event *event)
251 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
252 event = skip_time_extend(event);
253 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
254 /* If length is in len field, then array[0] has the data */
256 return (void *)&event->array[0];
257 /* Otherwise length is in array[0] and array[1] has the data */
258 return (void *)&event->array[1];
262 * ring_buffer_event_data - return the data of the event
263 * @event: the event to get the data from
265 void *ring_buffer_event_data(struct ring_buffer_event *event)
267 return rb_event_data(event);
269 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
271 #define for_each_buffer_cpu(buffer, cpu) \
272 for_each_cpu(cpu, buffer->cpumask)
275 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
276 #define TS_DELTA_TEST (~TS_MASK)
278 /* Flag when events were overwritten */
279 #define RB_MISSED_EVENTS (1 << 31)
280 /* Missed count stored at end */
281 #define RB_MISSED_STORED (1 << 30)
283 #define RB_MISSED_FLAGS (RB_MISSED_EVENTS|RB_MISSED_STORED)
285 struct buffer_data_page {
286 u64 time_stamp; /* page time stamp */
287 local_t commit; /* write committed index */
288 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
292 * Note, the buffer_page list must be first. The buffer pages
293 * are allocated in cache lines, which means that each buffer
294 * page will be at the beginning of a cache line, and thus
295 * the least significant bits will be zero. We use this to
296 * add flags in the list struct pointers, to make the ring buffer
300 struct list_head list; /* list of buffer pages */
301 local_t write; /* index for next write */
302 unsigned read; /* index for next read */
303 local_t entries; /* entries on this page */
304 unsigned long real_end; /* real end of data */
305 struct buffer_data_page *page; /* Actual data page */
309 * The buffer page counters, write and entries, must be reset
310 * atomically when crossing page boundaries. To synchronize this
311 * update, two counters are inserted into the number. One is
312 * the actual counter for the write position or count on the page.
314 * The other is a counter of updaters. Before an update happens
315 * the update partition of the counter is incremented. This will
316 * allow the updater to update the counter atomically.
318 * The counter is 20 bits, and the state data is 12.
320 #define RB_WRITE_MASK 0xfffff
321 #define RB_WRITE_INTCNT (1 << 20)
323 static void rb_init_page(struct buffer_data_page *bpage)
325 local_set(&bpage->commit, 0);
329 * ring_buffer_page_len - the size of data on the page.
330 * @page: The page to read
332 * Returns the amount of data on the page, including buffer page header.
334 size_t ring_buffer_page_len(void *page)
336 struct buffer_data_page *bpage = page;
338 return (local_read(&bpage->commit) & ~RB_MISSED_FLAGS)
343 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
346 static void free_buffer_page(struct buffer_page *bpage)
348 free_page((unsigned long)bpage->page);
353 * We need to fit the time_stamp delta into 27 bits.
355 static inline int test_time_stamp(u64 delta)
357 if (delta & TS_DELTA_TEST)
362 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
364 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
365 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
367 int ring_buffer_print_page_header(struct trace_seq *s)
369 struct buffer_data_page field;
371 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
372 "offset:0;\tsize:%u;\tsigned:%u;\n",
373 (unsigned int)sizeof(field.time_stamp),
374 (unsigned int)is_signed_type(u64));
376 trace_seq_printf(s, "\tfield: local_t commit;\t"
377 "offset:%u;\tsize:%u;\tsigned:%u;\n",
378 (unsigned int)offsetof(typeof(field), commit),
379 (unsigned int)sizeof(field.commit),
380 (unsigned int)is_signed_type(long));
382 trace_seq_printf(s, "\tfield: int overwrite;\t"
383 "offset:%u;\tsize:%u;\tsigned:%u;\n",
384 (unsigned int)offsetof(typeof(field), commit),
386 (unsigned int)is_signed_type(long));
388 trace_seq_printf(s, "\tfield: char data;\t"
389 "offset:%u;\tsize:%u;\tsigned:%u;\n",
390 (unsigned int)offsetof(typeof(field), data),
391 (unsigned int)BUF_PAGE_SIZE,
392 (unsigned int)is_signed_type(char));
394 return !trace_seq_has_overflowed(s);
398 struct irq_work work;
399 wait_queue_head_t waiters;
400 wait_queue_head_t full_waiters;
401 bool waiters_pending;
402 bool full_waiters_pending;
407 * Structure to hold event state and handle nested events.
409 struct rb_event_info {
412 unsigned long length;
413 struct buffer_page *tail_page;
418 * Used for which event context the event is in.
424 * See trace_recursive_lock() comment below for more details.
435 * head_page == tail_page && head == tail then buffer is empty.
437 struct ring_buffer_per_cpu {
439 atomic_t record_disabled;
440 struct ring_buffer *buffer;
441 raw_spinlock_t reader_lock; /* serialize readers */
442 arch_spinlock_t lock;
443 struct lock_class_key lock_key;
444 unsigned long nr_pages;
445 unsigned int current_context;
446 struct list_head *pages;
447 struct buffer_page *head_page; /* read from head */
448 struct buffer_page *tail_page; /* write to tail */
449 struct buffer_page *commit_page; /* committed pages */
450 struct buffer_page *reader_page;
451 unsigned long lost_events;
452 unsigned long last_overrun;
453 local_t entries_bytes;
456 local_t commit_overrun;
457 local_t dropped_events;
461 unsigned long read_bytes;
464 /* ring buffer pages to update, > 0 to add, < 0 to remove */
465 long nr_pages_to_update;
466 struct list_head new_pages; /* new pages to add */
467 struct work_struct update_pages_work;
468 struct completion update_done;
470 struct rb_irq_work irq_work;
476 atomic_t record_disabled;
477 atomic_t resize_disabled;
478 cpumask_var_t cpumask;
480 struct lock_class_key *reader_lock_key;
484 struct ring_buffer_per_cpu **buffers;
486 #ifdef CONFIG_HOTPLUG_CPU
487 struct notifier_block cpu_notify;
491 struct rb_irq_work irq_work;
494 struct ring_buffer_iter {
495 struct ring_buffer_per_cpu *cpu_buffer;
497 struct buffer_page *head_page;
498 struct buffer_page *cache_reader_page;
499 unsigned long cache_read;
504 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
506 * Schedules a delayed work to wake up any task that is blocked on the
507 * ring buffer waiters queue.
509 static void rb_wake_up_waiters(struct irq_work *work)
511 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
513 wake_up_all(&rbwork->waiters);
514 if (rbwork->wakeup_full) {
515 rbwork->wakeup_full = false;
516 wake_up_all(&rbwork->full_waiters);
521 * ring_buffer_wait - wait for input to the ring buffer
522 * @buffer: buffer to wait on
523 * @cpu: the cpu buffer to wait on
524 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
526 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
527 * as data is added to any of the @buffer's cpu buffers. Otherwise
528 * it will wait for data to be added to a specific cpu buffer.
530 int ring_buffer_wait(struct ring_buffer *buffer, int cpu, bool full)
532 struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
534 struct rb_irq_work *work;
538 * Depending on what the caller is waiting for, either any
539 * data in any cpu buffer, or a specific buffer, put the
540 * caller on the appropriate wait queue.
542 if (cpu == RING_BUFFER_ALL_CPUS) {
543 work = &buffer->irq_work;
544 /* Full only makes sense on per cpu reads */
547 if (!cpumask_test_cpu(cpu, buffer->cpumask))
549 cpu_buffer = buffer->buffers[cpu];
550 work = &cpu_buffer->irq_work;
556 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
558 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
561 * The events can happen in critical sections where
562 * checking a work queue can cause deadlocks.
563 * After adding a task to the queue, this flag is set
564 * only to notify events to try to wake up the queue
567 * We don't clear it even if the buffer is no longer
568 * empty. The flag only causes the next event to run
569 * irq_work to do the work queue wake up. The worse
570 * that can happen if we race with !trace_empty() is that
571 * an event will cause an irq_work to try to wake up
574 * There's no reason to protect this flag either, as
575 * the work queue and irq_work logic will do the necessary
576 * synchronization for the wake ups. The only thing
577 * that is necessary is that the wake up happens after
578 * a task has been queued. It's OK for spurious wake ups.
581 work->full_waiters_pending = true;
583 work->waiters_pending = true;
585 if (signal_pending(current)) {
590 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
593 if (cpu != RING_BUFFER_ALL_CPUS &&
594 !ring_buffer_empty_cpu(buffer, cpu)) {
601 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
602 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
603 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
613 finish_wait(&work->full_waiters, &wait);
615 finish_wait(&work->waiters, &wait);
621 * ring_buffer_poll_wait - poll on buffer input
622 * @buffer: buffer to wait on
623 * @cpu: the cpu buffer to wait on
624 * @filp: the file descriptor
625 * @poll_table: The poll descriptor
627 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
628 * as data is added to any of the @buffer's cpu buffers. Otherwise
629 * it will wait for data to be added to a specific cpu buffer.
631 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
634 int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
635 struct file *filp, poll_table *poll_table)
637 struct ring_buffer_per_cpu *cpu_buffer;
638 struct rb_irq_work *work;
640 if (cpu == RING_BUFFER_ALL_CPUS)
641 work = &buffer->irq_work;
643 if (!cpumask_test_cpu(cpu, buffer->cpumask))
646 cpu_buffer = buffer->buffers[cpu];
647 work = &cpu_buffer->irq_work;
650 poll_wait(filp, &work->waiters, poll_table);
651 work->waiters_pending = true;
653 * There's a tight race between setting the waiters_pending and
654 * checking if the ring buffer is empty. Once the waiters_pending bit
655 * is set, the next event will wake the task up, but we can get stuck
656 * if there's only a single event in.
658 * FIXME: Ideally, we need a memory barrier on the writer side as well,
659 * but adding a memory barrier to all events will cause too much of a
660 * performance hit in the fast path. We only need a memory barrier when
661 * the buffer goes from empty to having content. But as this race is
662 * extremely small, and it's not a problem if another event comes in, we
667 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
668 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
669 return POLLIN | POLLRDNORM;
673 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
674 #define RB_WARN_ON(b, cond) \
676 int _____ret = unlikely(cond); \
678 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
679 struct ring_buffer_per_cpu *__b = \
681 atomic_inc(&__b->buffer->record_disabled); \
683 atomic_inc(&b->record_disabled); \
689 /* Up this if you want to test the TIME_EXTENTS and normalization */
690 #define DEBUG_SHIFT 0
692 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
694 /* shift to debug/test normalization and TIME_EXTENTS */
695 return buffer->clock() << DEBUG_SHIFT;
698 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
702 preempt_disable_notrace();
703 time = rb_time_stamp(buffer);
704 preempt_enable_no_resched_notrace();
708 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
710 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
713 /* Just stupid testing the normalize function and deltas */
716 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
719 * Making the ring buffer lockless makes things tricky.
720 * Although writes only happen on the CPU that they are on,
721 * and they only need to worry about interrupts. Reads can
724 * The reader page is always off the ring buffer, but when the
725 * reader finishes with a page, it needs to swap its page with
726 * a new one from the buffer. The reader needs to take from
727 * the head (writes go to the tail). But if a writer is in overwrite
728 * mode and wraps, it must push the head page forward.
730 * Here lies the problem.
732 * The reader must be careful to replace only the head page, and
733 * not another one. As described at the top of the file in the
734 * ASCII art, the reader sets its old page to point to the next
735 * page after head. It then sets the page after head to point to
736 * the old reader page. But if the writer moves the head page
737 * during this operation, the reader could end up with the tail.
739 * We use cmpxchg to help prevent this race. We also do something
740 * special with the page before head. We set the LSB to 1.
742 * When the writer must push the page forward, it will clear the
743 * bit that points to the head page, move the head, and then set
744 * the bit that points to the new head page.
746 * We also don't want an interrupt coming in and moving the head
747 * page on another writer. Thus we use the second LSB to catch
750 * head->list->prev->next bit 1 bit 0
753 * Points to head page 0 1
756 * Note we can not trust the prev pointer of the head page, because:
758 * +----+ +-----+ +-----+
759 * | |------>| T |---X--->| N |
761 * +----+ +-----+ +-----+
764 * +----------| R |----------+ |
768 * Key: ---X--> HEAD flag set in pointer
773 * (see __rb_reserve_next() to see where this happens)
775 * What the above shows is that the reader just swapped out
776 * the reader page with a page in the buffer, but before it
777 * could make the new header point back to the new page added
778 * it was preempted by a writer. The writer moved forward onto
779 * the new page added by the reader and is about to move forward
782 * You can see, it is legitimate for the previous pointer of
783 * the head (or any page) not to point back to itself. But only
787 #define RB_PAGE_NORMAL 0UL
788 #define RB_PAGE_HEAD 1UL
789 #define RB_PAGE_UPDATE 2UL
792 #define RB_FLAG_MASK 3UL
794 /* PAGE_MOVED is not part of the mask */
795 #define RB_PAGE_MOVED 4UL
798 * rb_list_head - remove any bit
800 static struct list_head *rb_list_head(struct list_head *list)
802 unsigned long val = (unsigned long)list;
804 return (struct list_head *)(val & ~RB_FLAG_MASK);
808 * rb_is_head_page - test if the given page is the head page
810 * Because the reader may move the head_page pointer, we can
811 * not trust what the head page is (it may be pointing to
812 * the reader page). But if the next page is a header page,
813 * its flags will be non zero.
816 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
817 struct buffer_page *page, struct list_head *list)
821 val = (unsigned long)list->next;
823 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
824 return RB_PAGE_MOVED;
826 return val & RB_FLAG_MASK;
832 * The unique thing about the reader page, is that, if the
833 * writer is ever on it, the previous pointer never points
834 * back to the reader page.
836 static bool rb_is_reader_page(struct buffer_page *page)
838 struct list_head *list = page->list.prev;
840 return rb_list_head(list->next) != &page->list;
844 * rb_set_list_to_head - set a list_head to be pointing to head.
846 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
847 struct list_head *list)
851 ptr = (unsigned long *)&list->next;
852 *ptr |= RB_PAGE_HEAD;
853 *ptr &= ~RB_PAGE_UPDATE;
857 * rb_head_page_activate - sets up head page
859 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
861 struct buffer_page *head;
863 head = cpu_buffer->head_page;
868 * Set the previous list pointer to have the HEAD flag.
870 rb_set_list_to_head(cpu_buffer, head->list.prev);
873 static void rb_list_head_clear(struct list_head *list)
875 unsigned long *ptr = (unsigned long *)&list->next;
877 *ptr &= ~RB_FLAG_MASK;
881 * rb_head_page_dactivate - clears head page ptr (for free list)
884 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
886 struct list_head *hd;
888 /* Go through the whole list and clear any pointers found. */
889 rb_list_head_clear(cpu_buffer->pages);
891 list_for_each(hd, cpu_buffer->pages)
892 rb_list_head_clear(hd);
895 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
896 struct buffer_page *head,
897 struct buffer_page *prev,
898 int old_flag, int new_flag)
900 struct list_head *list;
901 unsigned long val = (unsigned long)&head->list;
906 val &= ~RB_FLAG_MASK;
908 ret = cmpxchg((unsigned long *)&list->next,
909 val | old_flag, val | new_flag);
911 /* check if the reader took the page */
912 if ((ret & ~RB_FLAG_MASK) != val)
913 return RB_PAGE_MOVED;
915 return ret & RB_FLAG_MASK;
918 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
919 struct buffer_page *head,
920 struct buffer_page *prev,
923 return rb_head_page_set(cpu_buffer, head, prev,
924 old_flag, RB_PAGE_UPDATE);
927 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
928 struct buffer_page *head,
929 struct buffer_page *prev,
932 return rb_head_page_set(cpu_buffer, head, prev,
933 old_flag, RB_PAGE_HEAD);
936 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
937 struct buffer_page *head,
938 struct buffer_page *prev,
941 return rb_head_page_set(cpu_buffer, head, prev,
942 old_flag, RB_PAGE_NORMAL);
945 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
946 struct buffer_page **bpage)
948 struct list_head *p = rb_list_head((*bpage)->list.next);
950 *bpage = list_entry(p, struct buffer_page, list);
953 static struct buffer_page *
954 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
956 struct buffer_page *head;
957 struct buffer_page *page;
958 struct list_head *list;
961 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
965 list = cpu_buffer->pages;
966 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
969 page = head = cpu_buffer->head_page;
971 * It is possible that the writer moves the header behind
972 * where we started, and we miss in one loop.
973 * A second loop should grab the header, but we'll do
974 * three loops just because I'm paranoid.
976 for (i = 0; i < 3; i++) {
978 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
979 cpu_buffer->head_page = page;
982 rb_inc_page(cpu_buffer, &page);
983 } while (page != head);
986 RB_WARN_ON(cpu_buffer, 1);
991 static int rb_head_page_replace(struct buffer_page *old,
992 struct buffer_page *new)
994 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
998 val = *ptr & ~RB_FLAG_MASK;
1001 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1007 * rb_tail_page_update - move the tail page forward
1009 * Returns 1 if moved tail page, 0 if someone else did.
1011 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1012 struct buffer_page *tail_page,
1013 struct buffer_page *next_page)
1015 struct buffer_page *old_tail;
1016 unsigned long old_entries;
1017 unsigned long old_write;
1021 * The tail page now needs to be moved forward.
1023 * We need to reset the tail page, but without messing
1024 * with possible erasing of data brought in by interrupts
1025 * that have moved the tail page and are currently on it.
1027 * We add a counter to the write field to denote this.
1029 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1030 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1033 * Just make sure we have seen our old_write and synchronize
1034 * with any interrupts that come in.
1039 * If the tail page is still the same as what we think
1040 * it is, then it is up to us to update the tail
1043 if (tail_page == cpu_buffer->tail_page) {
1044 /* Zero the write counter */
1045 unsigned long val = old_write & ~RB_WRITE_MASK;
1046 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1049 * This will only succeed if an interrupt did
1050 * not come in and change it. In which case, we
1051 * do not want to modify it.
1053 * We add (void) to let the compiler know that we do not care
1054 * about the return value of these functions. We use the
1055 * cmpxchg to only update if an interrupt did not already
1056 * do it for us. If the cmpxchg fails, we don't care.
1058 (void)local_cmpxchg(&next_page->write, old_write, val);
1059 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1062 * No need to worry about races with clearing out the commit.
1063 * it only can increment when a commit takes place. But that
1064 * only happens in the outer most nested commit.
1066 local_set(&next_page->page->commit, 0);
1068 old_tail = cmpxchg(&cpu_buffer->tail_page,
1069 tail_page, next_page);
1071 if (old_tail == tail_page)
1078 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1079 struct buffer_page *bpage)
1081 unsigned long val = (unsigned long)bpage;
1083 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1090 * rb_check_list - make sure a pointer to a list has the last bits zero
1092 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1093 struct list_head *list)
1095 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1097 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1103 * rb_check_pages - integrity check of buffer pages
1104 * @cpu_buffer: CPU buffer with pages to test
1106 * As a safety measure we check to make sure the data pages have not
1109 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1111 struct list_head *head = cpu_buffer->pages;
1112 struct buffer_page *bpage, *tmp;
1114 /* Reset the head page if it exists */
1115 if (cpu_buffer->head_page)
1116 rb_set_head_page(cpu_buffer);
1118 rb_head_page_deactivate(cpu_buffer);
1120 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1122 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1125 if (rb_check_list(cpu_buffer, head))
1128 list_for_each_entry_safe(bpage, tmp, head, list) {
1129 if (RB_WARN_ON(cpu_buffer,
1130 bpage->list.next->prev != &bpage->list))
1132 if (RB_WARN_ON(cpu_buffer,
1133 bpage->list.prev->next != &bpage->list))
1135 if (rb_check_list(cpu_buffer, &bpage->list))
1139 rb_head_page_activate(cpu_buffer);
1144 static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1146 struct buffer_page *bpage, *tmp;
1149 for (i = 0; i < nr_pages; i++) {
1152 * __GFP_NORETRY flag makes sure that the allocation fails
1153 * gracefully without invoking oom-killer and the system is
1156 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1157 GFP_KERNEL | __GFP_NORETRY,
1162 list_add(&bpage->list, pages);
1164 page = alloc_pages_node(cpu_to_node(cpu),
1165 GFP_KERNEL | __GFP_NORETRY, 0);
1168 bpage->page = page_address(page);
1169 rb_init_page(bpage->page);
1175 list_for_each_entry_safe(bpage, tmp, pages, list) {
1176 list_del_init(&bpage->list);
1177 free_buffer_page(bpage);
1183 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1184 unsigned long nr_pages)
1190 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1194 * The ring buffer page list is a circular list that does not
1195 * start and end with a list head. All page list items point to
1198 cpu_buffer->pages = pages.next;
1201 cpu_buffer->nr_pages = nr_pages;
1203 rb_check_pages(cpu_buffer);
1208 static struct ring_buffer_per_cpu *
1209 rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1211 struct ring_buffer_per_cpu *cpu_buffer;
1212 struct buffer_page *bpage;
1216 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1217 GFP_KERNEL, cpu_to_node(cpu));
1221 cpu_buffer->cpu = cpu;
1222 cpu_buffer->buffer = buffer;
1223 raw_spin_lock_init(&cpu_buffer->reader_lock);
1224 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1225 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1226 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1227 init_completion(&cpu_buffer->update_done);
1228 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1229 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1230 init_waitqueue_head(&cpu_buffer->irq_work.full_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;
1313 /* keep it in its own cache line */
1314 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1319 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1320 goto fail_free_buffer;
1322 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1323 buffer->flags = flags;
1324 buffer->clock = trace_clock_local;
1325 buffer->reader_lock_key = key;
1327 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1328 init_waitqueue_head(&buffer->irq_work.waiters);
1330 /* need at least two pages */
1335 * In case of non-hotplug cpu, if the ring-buffer is allocated
1336 * in early initcall, it will not be notified of secondary cpus.
1337 * In that off case, we need to allocate for all possible cpus.
1339 #ifdef CONFIG_HOTPLUG_CPU
1340 cpu_notifier_register_begin();
1341 cpumask_copy(buffer->cpumask, cpu_online_mask);
1343 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1345 buffer->cpus = nr_cpu_ids;
1347 bsize = sizeof(void *) * nr_cpu_ids;
1348 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1350 if (!buffer->buffers)
1351 goto fail_free_cpumask;
1353 for_each_buffer_cpu(buffer, cpu) {
1354 buffer->buffers[cpu] =
1355 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1356 if (!buffer->buffers[cpu])
1357 goto fail_free_buffers;
1360 #ifdef CONFIG_HOTPLUG_CPU
1361 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1362 buffer->cpu_notify.priority = 0;
1363 __register_cpu_notifier(&buffer->cpu_notify);
1364 cpu_notifier_register_done();
1367 mutex_init(&buffer->mutex);
1372 for_each_buffer_cpu(buffer, cpu) {
1373 if (buffer->buffers[cpu])
1374 rb_free_cpu_buffer(buffer->buffers[cpu]);
1376 kfree(buffer->buffers);
1379 free_cpumask_var(buffer->cpumask);
1380 #ifdef CONFIG_HOTPLUG_CPU
1381 cpu_notifier_register_done();
1388 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1391 * ring_buffer_free - free a ring buffer.
1392 * @buffer: the buffer to free.
1395 ring_buffer_free(struct ring_buffer *buffer)
1399 #ifdef CONFIG_HOTPLUG_CPU
1400 cpu_notifier_register_begin();
1401 __unregister_cpu_notifier(&buffer->cpu_notify);
1404 for_each_buffer_cpu(buffer, cpu)
1405 rb_free_cpu_buffer(buffer->buffers[cpu]);
1407 #ifdef CONFIG_HOTPLUG_CPU
1408 cpu_notifier_register_done();
1411 kfree(buffer->buffers);
1412 free_cpumask_var(buffer->cpumask);
1416 EXPORT_SYMBOL_GPL(ring_buffer_free);
1418 void ring_buffer_set_clock(struct ring_buffer *buffer,
1421 buffer->clock = clock;
1424 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1426 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1428 return local_read(&bpage->entries) & RB_WRITE_MASK;
1431 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1433 return local_read(&bpage->write) & RB_WRITE_MASK;
1437 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1439 struct list_head *tail_page, *to_remove, *next_page;
1440 struct buffer_page *to_remove_page, *tmp_iter_page;
1441 struct buffer_page *last_page, *first_page;
1442 unsigned long nr_removed;
1443 unsigned long head_bit;
1448 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1449 atomic_inc(&cpu_buffer->record_disabled);
1451 * We don't race with the readers since we have acquired the reader
1452 * lock. We also don't race with writers after disabling recording.
1453 * This makes it easy to figure out the first and the last page to be
1454 * removed from the list. We unlink all the pages in between including
1455 * the first and last pages. This is done in a busy loop so that we
1456 * lose the least number of traces.
1457 * The pages are freed after we restart recording and unlock readers.
1459 tail_page = &cpu_buffer->tail_page->list;
1462 * tail page might be on reader page, we remove the next page
1463 * from the ring buffer
1465 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1466 tail_page = rb_list_head(tail_page->next);
1467 to_remove = tail_page;
1469 /* start of pages to remove */
1470 first_page = list_entry(rb_list_head(to_remove->next),
1471 struct buffer_page, list);
1473 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1474 to_remove = rb_list_head(to_remove)->next;
1475 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1478 next_page = rb_list_head(to_remove)->next;
1481 * Now we remove all pages between tail_page and next_page.
1482 * Make sure that we have head_bit value preserved for the
1485 tail_page->next = (struct list_head *)((unsigned long)next_page |
1487 next_page = rb_list_head(next_page);
1488 next_page->prev = tail_page;
1490 /* make sure pages points to a valid page in the ring buffer */
1491 cpu_buffer->pages = next_page;
1493 /* update head page */
1495 cpu_buffer->head_page = list_entry(next_page,
1496 struct buffer_page, list);
1499 * change read pointer to make sure any read iterators reset
1502 cpu_buffer->read = 0;
1504 /* pages are removed, resume tracing and then free the pages */
1505 atomic_dec(&cpu_buffer->record_disabled);
1506 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1508 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1510 /* last buffer page to remove */
1511 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1513 tmp_iter_page = first_page;
1516 to_remove_page = tmp_iter_page;
1517 rb_inc_page(cpu_buffer, &tmp_iter_page);
1519 /* update the counters */
1520 page_entries = rb_page_entries(to_remove_page);
1523 * If something was added to this page, it was full
1524 * since it is not the tail page. So we deduct the
1525 * bytes consumed in ring buffer from here.
1526 * Increment overrun to account for the lost events.
1528 local_add(page_entries, &cpu_buffer->overrun);
1529 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1533 * We have already removed references to this list item, just
1534 * free up the buffer_page and its page
1536 free_buffer_page(to_remove_page);
1539 } while (to_remove_page != last_page);
1541 RB_WARN_ON(cpu_buffer, nr_removed);
1543 return nr_removed == 0;
1547 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1549 struct list_head *pages = &cpu_buffer->new_pages;
1550 int retries, success;
1552 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1554 * We are holding the reader lock, so the reader page won't be swapped
1555 * in the ring buffer. Now we are racing with the writer trying to
1556 * move head page and the tail page.
1557 * We are going to adapt the reader page update process where:
1558 * 1. We first splice the start and end of list of new pages between
1559 * the head page and its previous page.
1560 * 2. We cmpxchg the prev_page->next to point from head page to the
1561 * start of new pages list.
1562 * 3. Finally, we update the head->prev to the end of new list.
1564 * We will try this process 10 times, to make sure that we don't keep
1570 struct list_head *head_page, *prev_page, *r;
1571 struct list_head *last_page, *first_page;
1572 struct list_head *head_page_with_bit;
1574 head_page = &rb_set_head_page(cpu_buffer)->list;
1577 prev_page = head_page->prev;
1579 first_page = pages->next;
1580 last_page = pages->prev;
1582 head_page_with_bit = (struct list_head *)
1583 ((unsigned long)head_page | RB_PAGE_HEAD);
1585 last_page->next = head_page_with_bit;
1586 first_page->prev = prev_page;
1588 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1590 if (r == head_page_with_bit) {
1592 * yay, we replaced the page pointer to our new list,
1593 * now, we just have to update to head page's prev
1594 * pointer to point to end of list
1596 head_page->prev = last_page;
1603 INIT_LIST_HEAD(pages);
1605 * If we weren't successful in adding in new pages, warn and stop
1608 RB_WARN_ON(cpu_buffer, !success);
1609 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1611 /* free pages if they weren't inserted */
1613 struct buffer_page *bpage, *tmp;
1614 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1616 list_del_init(&bpage->list);
1617 free_buffer_page(bpage);
1623 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1627 if (cpu_buffer->nr_pages_to_update > 0)
1628 success = rb_insert_pages(cpu_buffer);
1630 success = rb_remove_pages(cpu_buffer,
1631 -cpu_buffer->nr_pages_to_update);
1634 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1637 static void update_pages_handler(struct work_struct *work)
1639 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1640 struct ring_buffer_per_cpu, update_pages_work);
1641 rb_update_pages(cpu_buffer);
1642 complete(&cpu_buffer->update_done);
1646 * ring_buffer_resize - resize the ring buffer
1647 * @buffer: the buffer to resize.
1648 * @size: the new size.
1649 * @cpu_id: the cpu buffer to resize
1651 * Minimum size is 2 * BUF_PAGE_SIZE.
1653 * Returns 0 on success and < 0 on failure.
1655 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1658 struct ring_buffer_per_cpu *cpu_buffer;
1659 unsigned long nr_pages;
1663 * Always succeed at resizing a non-existent buffer:
1668 /* Make sure the requested buffer exists */
1669 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1670 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1673 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1675 /* we need a minimum of two pages */
1679 size = nr_pages * 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;
1894 static void rb_inc_iter(struct ring_buffer_iter *iter)
1896 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1899 * The iterator could be on the reader page (it starts there).
1900 * But the head could have moved, since the reader was
1901 * found. Check for this case and assign the iterator
1902 * to the head page instead of next.
1904 if (iter->head_page == cpu_buffer->reader_page)
1905 iter->head_page = rb_set_head_page(cpu_buffer);
1907 rb_inc_page(cpu_buffer, &iter->head_page);
1909 iter->read_stamp = iter->head_page->page->time_stamp;
1914 * rb_handle_head_page - writer hit the head page
1916 * Returns: +1 to retry page
1921 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1922 struct buffer_page *tail_page,
1923 struct buffer_page *next_page)
1925 struct buffer_page *new_head;
1930 entries = rb_page_entries(next_page);
1933 * The hard part is here. We need to move the head
1934 * forward, and protect against both readers on
1935 * other CPUs and writers coming in via interrupts.
1937 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1941 * type can be one of four:
1942 * NORMAL - an interrupt already moved it for us
1943 * HEAD - we are the first to get here.
1944 * UPDATE - we are the interrupt interrupting
1946 * MOVED - a reader on another CPU moved the next
1947 * pointer to its reader page. Give up
1954 * We changed the head to UPDATE, thus
1955 * it is our responsibility to update
1958 local_add(entries, &cpu_buffer->overrun);
1959 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1962 * The entries will be zeroed out when we move the
1966 /* still more to do */
1969 case RB_PAGE_UPDATE:
1971 * This is an interrupt that interrupt the
1972 * previous update. Still more to do.
1975 case RB_PAGE_NORMAL:
1977 * An interrupt came in before the update
1978 * and processed this for us.
1979 * Nothing left to do.
1984 * The reader is on another CPU and just did
1985 * a swap with our next_page.
1990 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1995 * Now that we are here, the old head pointer is
1996 * set to UPDATE. This will keep the reader from
1997 * swapping the head page with the reader page.
1998 * The reader (on another CPU) will spin till
2001 * We just need to protect against interrupts
2002 * doing the job. We will set the next pointer
2003 * to HEAD. After that, we set the old pointer
2004 * to NORMAL, but only if it was HEAD before.
2005 * otherwise we are an interrupt, and only
2006 * want the outer most commit to reset it.
2008 new_head = next_page;
2009 rb_inc_page(cpu_buffer, &new_head);
2011 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2015 * Valid returns are:
2016 * HEAD - an interrupt came in and already set it.
2017 * NORMAL - One of two things:
2018 * 1) We really set it.
2019 * 2) A bunch of interrupts came in and moved
2020 * the page forward again.
2024 case RB_PAGE_NORMAL:
2028 RB_WARN_ON(cpu_buffer, 1);
2033 * It is possible that an interrupt came in,
2034 * set the head up, then more interrupts came in
2035 * and moved it again. When we get back here,
2036 * the page would have been set to NORMAL but we
2037 * just set it back to HEAD.
2039 * How do you detect this? Well, if that happened
2040 * the tail page would have moved.
2042 if (ret == RB_PAGE_NORMAL) {
2044 * If the tail had moved passed next, then we need
2045 * to reset the pointer.
2047 if (cpu_buffer->tail_page != tail_page &&
2048 cpu_buffer->tail_page != next_page)
2049 rb_head_page_set_normal(cpu_buffer, new_head,
2055 * If this was the outer most commit (the one that
2056 * changed the original pointer from HEAD to UPDATE),
2057 * then it is up to us to reset it to NORMAL.
2059 if (type == RB_PAGE_HEAD) {
2060 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2063 if (RB_WARN_ON(cpu_buffer,
2064 ret != RB_PAGE_UPDATE))
2072 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2073 unsigned long tail, struct rb_event_info *info)
2075 struct buffer_page *tail_page = info->tail_page;
2076 struct ring_buffer_event *event;
2077 unsigned long length = info->length;
2080 * Only the event that crossed the page boundary
2081 * must fill the old tail_page with padding.
2083 if (tail >= BUF_PAGE_SIZE) {
2085 * If the page was filled, then we still need
2086 * to update the real_end. Reset it to zero
2087 * and the reader will ignore it.
2089 if (tail == BUF_PAGE_SIZE)
2090 tail_page->real_end = 0;
2092 local_sub(length, &tail_page->write);
2096 event = __rb_page_index(tail_page, tail);
2097 kmemcheck_annotate_bitfield(event, bitfield);
2099 /* account for padding bytes */
2100 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2103 * Save the original length to the meta data.
2104 * This will be used by the reader to add lost event
2107 tail_page->real_end = tail;
2110 * If this event is bigger than the minimum size, then
2111 * we need to be careful that we don't subtract the
2112 * write counter enough to allow another writer to slip
2114 * We put in a discarded commit instead, to make sure
2115 * that this space is not used again.
2117 * If we are less than the minimum size, we don't need to
2120 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2121 /* No room for any events */
2123 /* Mark the rest of the page with padding */
2124 rb_event_set_padding(event);
2126 /* Set the write back to the previous setting */
2127 local_sub(length, &tail_page->write);
2131 /* Put in a discarded event */
2132 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2133 event->type_len = RINGBUF_TYPE_PADDING;
2134 /* time delta must be non zero */
2135 event->time_delta = 1;
2137 /* Set write to end of buffer */
2138 length = (tail + length) - BUF_PAGE_SIZE;
2139 local_sub(length, &tail_page->write);
2143 * This is the slow path, force gcc not to inline it.
2145 static noinline struct ring_buffer_event *
2146 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2147 unsigned long tail, struct rb_event_info *info)
2149 struct buffer_page *tail_page = info->tail_page;
2150 struct buffer_page *commit_page = cpu_buffer->commit_page;
2151 struct ring_buffer *buffer = cpu_buffer->buffer;
2152 struct buffer_page *next_page;
2156 next_page = tail_page;
2158 rb_inc_page(cpu_buffer, &next_page);
2161 * If for some reason, we had an interrupt storm that made
2162 * it all the way around the buffer, bail, and warn
2165 if (unlikely(next_page == commit_page)) {
2166 local_inc(&cpu_buffer->commit_overrun);
2171 * This is where the fun begins!
2173 * We are fighting against races between a reader that
2174 * could be on another CPU trying to swap its reader
2175 * page with the buffer head.
2177 * We are also fighting against interrupts coming in and
2178 * moving the head or tail on us as well.
2180 * If the next page is the head page then we have filled
2181 * the buffer, unless the commit page is still on the
2184 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2187 * If the commit is not on the reader page, then
2188 * move the header page.
2190 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2192 * If we are not in overwrite mode,
2193 * this is easy, just stop here.
2195 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2196 local_inc(&cpu_buffer->dropped_events);
2200 ret = rb_handle_head_page(cpu_buffer,
2209 * We need to be careful here too. The
2210 * commit page could still be on the reader
2211 * page. We could have a small buffer, and
2212 * have filled up the buffer with events
2213 * from interrupts and such, and wrapped.
2215 * Note, if the tail page is also the on the
2216 * reader_page, we let it move out.
2218 if (unlikely((cpu_buffer->commit_page !=
2219 cpu_buffer->tail_page) &&
2220 (cpu_buffer->commit_page ==
2221 cpu_buffer->reader_page))) {
2222 local_inc(&cpu_buffer->commit_overrun);
2228 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
2231 * Nested commits always have zero deltas, so
2232 * just reread the time stamp
2234 ts = rb_time_stamp(buffer);
2235 next_page->page->time_stamp = ts;
2240 rb_reset_tail(cpu_buffer, tail, info);
2242 /* fail and let the caller try again */
2243 return ERR_PTR(-EAGAIN);
2247 rb_reset_tail(cpu_buffer, tail, info);
2252 /* Slow path, do not inline */
2253 static noinline struct ring_buffer_event *
2254 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
2256 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2258 /* Not the first event on the page? */
2259 if (rb_event_index(event)) {
2260 event->time_delta = delta & TS_MASK;
2261 event->array[0] = delta >> TS_SHIFT;
2263 /* nope, just zero it */
2264 event->time_delta = 0;
2265 event->array[0] = 0;
2268 return skip_time_extend(event);
2271 static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2272 struct ring_buffer_event *event);
2275 * rb_update_event - update event type and data
2276 * @event: the event to update
2277 * @type: the type of event
2278 * @length: the size of the event field in the ring buffer
2280 * Update the type and data fields of the event. The length
2281 * is the actual size that is written to the ring buffer,
2282 * and with this, we can determine what to place into the
2286 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2287 struct ring_buffer_event *event,
2288 struct rb_event_info *info)
2290 unsigned length = info->length;
2291 u64 delta = info->delta;
2293 /* Only a commit updates the timestamp */
2294 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2298 * If we need to add a timestamp, then we
2299 * add it to the start of the resevered space.
2301 if (unlikely(info->add_timestamp)) {
2302 event = rb_add_time_stamp(event, delta);
2303 length -= RB_LEN_TIME_EXTEND;
2307 event->time_delta = delta;
2308 length -= RB_EVNT_HDR_SIZE;
2309 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2310 event->type_len = 0;
2311 event->array[0] = length;
2313 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2316 static unsigned rb_calculate_event_length(unsigned length)
2318 struct ring_buffer_event event; /* Used only for sizeof array */
2320 /* zero length can cause confusions */
2324 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2325 length += sizeof(event.array[0]);
2327 length += RB_EVNT_HDR_SIZE;
2328 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2331 * In case the time delta is larger than the 27 bits for it
2332 * in the header, we need to add a timestamp. If another
2333 * event comes in when trying to discard this one to increase
2334 * the length, then the timestamp will be added in the allocated
2335 * space of this event. If length is bigger than the size needed
2336 * for the TIME_EXTEND, then padding has to be used. The events
2337 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2338 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2339 * As length is a multiple of 4, we only need to worry if it
2340 * is 12 (RB_LEN_TIME_EXTEND + 4).
2342 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2343 length += RB_ALIGNMENT;
2348 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2349 static inline bool sched_clock_stable(void)
2356 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2357 struct ring_buffer_event *event)
2359 unsigned long new_index, old_index;
2360 struct buffer_page *bpage;
2361 unsigned long index;
2364 new_index = rb_event_index(event);
2365 old_index = new_index + rb_event_ts_length(event);
2366 addr = (unsigned long)event;
2369 bpage = cpu_buffer->tail_page;
2371 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2372 unsigned long write_mask =
2373 local_read(&bpage->write) & ~RB_WRITE_MASK;
2374 unsigned long event_length = rb_event_length(event);
2376 * This is on the tail page. It is possible that
2377 * a write could come in and move the tail page
2378 * and write to the next page. That is fine
2379 * because we just shorten what is on this page.
2381 old_index += write_mask;
2382 new_index += write_mask;
2383 index = local_cmpxchg(&bpage->write, old_index, new_index);
2384 if (index == old_index) {
2385 /* update counters */
2386 local_sub(event_length, &cpu_buffer->entries_bytes);
2391 /* could not discard */
2395 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2397 local_inc(&cpu_buffer->committing);
2398 local_inc(&cpu_buffer->commits);
2402 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2404 unsigned long max_count;
2407 * We only race with interrupts and NMIs on this CPU.
2408 * If we own the commit event, then we can commit
2409 * all others that interrupted us, since the interruptions
2410 * are in stack format (they finish before they come
2411 * back to us). This allows us to do a simple loop to
2412 * assign the commit to the tail.
2415 max_count = cpu_buffer->nr_pages * 100;
2417 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
2418 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2420 if (RB_WARN_ON(cpu_buffer,
2421 rb_is_reader_page(cpu_buffer->tail_page)))
2423 local_set(&cpu_buffer->commit_page->page->commit,
2424 rb_page_write(cpu_buffer->commit_page));
2425 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2426 cpu_buffer->write_stamp =
2427 cpu_buffer->commit_page->page->time_stamp;
2428 /* add barrier to keep gcc from optimizing too much */
2431 while (rb_commit_index(cpu_buffer) !=
2432 rb_page_write(cpu_buffer->commit_page)) {
2434 local_set(&cpu_buffer->commit_page->page->commit,
2435 rb_page_write(cpu_buffer->commit_page));
2436 RB_WARN_ON(cpu_buffer,
2437 local_read(&cpu_buffer->commit_page->page->commit) &
2442 /* again, keep gcc from optimizing */
2446 * If an interrupt came in just after the first while loop
2447 * and pushed the tail page forward, we will be left with
2448 * a dangling commit that will never go forward.
2450 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
2454 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2456 unsigned long commits;
2458 if (RB_WARN_ON(cpu_buffer,
2459 !local_read(&cpu_buffer->committing)))
2463 commits = local_read(&cpu_buffer->commits);
2464 /* synchronize with interrupts */
2466 if (local_read(&cpu_buffer->committing) == 1)
2467 rb_set_commit_to_write(cpu_buffer);
2469 local_dec(&cpu_buffer->committing);
2471 /* synchronize with interrupts */
2475 * Need to account for interrupts coming in between the
2476 * updating of the commit page and the clearing of the
2477 * committing counter.
2479 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2480 !local_read(&cpu_buffer->committing)) {
2481 local_inc(&cpu_buffer->committing);
2486 static inline void rb_event_discard(struct ring_buffer_event *event)
2488 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2489 event = skip_time_extend(event);
2491 /* array[0] holds the actual length for the discarded event */
2492 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2493 event->type_len = RINGBUF_TYPE_PADDING;
2494 /* time delta must be non zero */
2495 if (!event->time_delta)
2496 event->time_delta = 1;
2500 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2501 struct ring_buffer_event *event)
2503 unsigned long addr = (unsigned long)event;
2504 unsigned long index;
2506 index = rb_event_index(event);
2509 return cpu_buffer->commit_page->page == (void *)addr &&
2510 rb_commit_index(cpu_buffer) == index;
2514 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2515 struct ring_buffer_event *event)
2520 * The event first in the commit queue updates the
2523 if (rb_event_is_commit(cpu_buffer, event)) {
2525 * A commit event that is first on a page
2526 * updates the write timestamp with the page stamp
2528 if (!rb_event_index(event))
2529 cpu_buffer->write_stamp =
2530 cpu_buffer->commit_page->page->time_stamp;
2531 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2532 delta = event->array[0];
2534 delta += event->time_delta;
2535 cpu_buffer->write_stamp += delta;
2537 cpu_buffer->write_stamp += event->time_delta;
2541 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2542 struct ring_buffer_event *event)
2544 local_inc(&cpu_buffer->entries);
2545 rb_update_write_stamp(cpu_buffer, event);
2546 rb_end_commit(cpu_buffer);
2549 static __always_inline void
2550 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2554 if (buffer->irq_work.waiters_pending) {
2555 buffer->irq_work.waiters_pending = false;
2556 /* irq_work_queue() supplies it's own memory barriers */
2557 irq_work_queue(&buffer->irq_work.work);
2560 if (cpu_buffer->irq_work.waiters_pending) {
2561 cpu_buffer->irq_work.waiters_pending = false;
2562 /* irq_work_queue() supplies it's own memory barriers */
2563 irq_work_queue(&cpu_buffer->irq_work.work);
2566 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
2568 if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2569 cpu_buffer->irq_work.wakeup_full = true;
2570 cpu_buffer->irq_work.full_waiters_pending = false;
2571 /* irq_work_queue() supplies it's own memory barriers */
2572 irq_work_queue(&cpu_buffer->irq_work.work);
2577 * The lock and unlock are done within a preempt disable section.
2578 * The current_context per_cpu variable can only be modified
2579 * by the current task between lock and unlock. But it can
2580 * be modified more than once via an interrupt. To pass this
2581 * information from the lock to the unlock without having to
2582 * access the 'in_interrupt()' functions again (which do show
2583 * a bit of overhead in something as critical as function tracing,
2584 * we use a bitmask trick.
2586 * bit 0 = NMI context
2587 * bit 1 = IRQ context
2588 * bit 2 = SoftIRQ context
2589 * bit 3 = normal context.
2591 * This works because this is the order of contexts that can
2592 * preempt other contexts. A SoftIRQ never preempts an IRQ
2595 * When the context is determined, the corresponding bit is
2596 * checked and set (if it was set, then a recursion of that context
2599 * On unlock, we need to clear this bit. To do so, just subtract
2600 * 1 from the current_context and AND it to itself.
2604 * 101 & 100 = 100 (clearing bit zero)
2607 * 1010 & 1001 = 1000 (clearing bit 1)
2609 * The least significant bit can be cleared this way, and it
2610 * just so happens that it is the same bit corresponding to
2611 * the current context.
2614 static __always_inline int
2615 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2617 unsigned int val = cpu_buffer->current_context;
2620 if (in_interrupt()) {
2626 bit = RB_CTX_SOFTIRQ;
2628 bit = RB_CTX_NORMAL;
2630 if (unlikely(val & (1 << bit)))
2634 cpu_buffer->current_context = val;
2639 static __always_inline void
2640 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2642 cpu_buffer->current_context &= cpu_buffer->current_context - 1;
2646 * ring_buffer_unlock_commit - commit a reserved
2647 * @buffer: The buffer to commit to
2648 * @event: The event pointer to commit.
2650 * This commits the data to the ring buffer, and releases any locks held.
2652 * Must be paired with ring_buffer_lock_reserve.
2654 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2655 struct ring_buffer_event *event)
2657 struct ring_buffer_per_cpu *cpu_buffer;
2658 int cpu = raw_smp_processor_id();
2660 cpu_buffer = buffer->buffers[cpu];
2662 rb_commit(cpu_buffer, event);
2664 rb_wakeups(buffer, cpu_buffer);
2666 trace_recursive_unlock(cpu_buffer);
2668 preempt_enable_notrace();
2672 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2674 static noinline void
2675 rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2676 struct rb_event_info *info)
2678 WARN_ONCE(info->delta > (1ULL << 59),
2679 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2680 (unsigned long long)info->delta,
2681 (unsigned long long)info->ts,
2682 (unsigned long long)cpu_buffer->write_stamp,
2683 sched_clock_stable() ? "" :
2684 "If you just came from a suspend/resume,\n"
2685 "please switch to the trace global clock:\n"
2686 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2687 info->add_timestamp = 1;
2690 static struct ring_buffer_event *
2691 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2692 struct rb_event_info *info)
2694 struct ring_buffer_event *event;
2695 struct buffer_page *tail_page;
2696 unsigned long tail, write;
2699 * If the time delta since the last event is too big to
2700 * hold in the time field of the event, then we append a
2701 * TIME EXTEND event ahead of the data event.
2703 if (unlikely(info->add_timestamp))
2704 info->length += RB_LEN_TIME_EXTEND;
2706 tail_page = info->tail_page = cpu_buffer->tail_page;
2707 write = local_add_return(info->length, &tail_page->write);
2709 /* set write to only the index of the write */
2710 write &= RB_WRITE_MASK;
2711 tail = write - info->length;
2714 * If this is the first commit on the page, then it has the same
2715 * timestamp as the page itself.
2720 /* See if we shot pass the end of this buffer page */
2721 if (unlikely(write > BUF_PAGE_SIZE))
2722 return rb_move_tail(cpu_buffer, tail, info);
2724 /* We reserved something on the buffer */
2726 event = __rb_page_index(tail_page, tail);
2727 kmemcheck_annotate_bitfield(event, bitfield);
2728 rb_update_event(cpu_buffer, event, info);
2730 local_inc(&tail_page->entries);
2733 * If this is the first commit on the page, then update
2737 tail_page->page->time_stamp = info->ts;
2739 /* account for these added bytes */
2740 local_add(info->length, &cpu_buffer->entries_bytes);
2745 static struct ring_buffer_event *
2746 rb_reserve_next_event(struct ring_buffer *buffer,
2747 struct ring_buffer_per_cpu *cpu_buffer,
2748 unsigned long length)
2750 struct ring_buffer_event *event;
2751 struct rb_event_info info;
2755 rb_start_commit(cpu_buffer);
2757 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2759 * Due to the ability to swap a cpu buffer from a buffer
2760 * it is possible it was swapped before we committed.
2761 * (committing stops a swap). We check for it here and
2762 * if it happened, we have to fail the write.
2765 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2766 local_dec(&cpu_buffer->committing);
2767 local_dec(&cpu_buffer->commits);
2772 info.length = rb_calculate_event_length(length);
2774 info.add_timestamp = 0;
2778 * We allow for interrupts to reenter here and do a trace.
2779 * If one does, it will cause this original code to loop
2780 * back here. Even with heavy interrupts happening, this
2781 * should only happen a few times in a row. If this happens
2782 * 1000 times in a row, there must be either an interrupt
2783 * storm or we have something buggy.
2786 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2789 info.ts = rb_time_stamp(cpu_buffer->buffer);
2790 diff = info.ts - cpu_buffer->write_stamp;
2792 /* make sure this diff is calculated here */
2795 /* Did the write stamp get updated already? */
2796 if (likely(info.ts >= cpu_buffer->write_stamp)) {
2798 if (unlikely(test_time_stamp(info.delta)))
2799 rb_handle_timestamp(cpu_buffer, &info);
2802 event = __rb_reserve_next(cpu_buffer, &info);
2804 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2805 if (info.add_timestamp)
2806 info.length -= RB_LEN_TIME_EXTEND;
2816 rb_end_commit(cpu_buffer);
2821 * ring_buffer_lock_reserve - reserve a part of the buffer
2822 * @buffer: the ring buffer to reserve from
2823 * @length: the length of the data to reserve (excluding event header)
2825 * Returns a reseverd event on the ring buffer to copy directly to.
2826 * The user of this interface will need to get the body to write into
2827 * and can use the ring_buffer_event_data() interface.
2829 * The length is the length of the data needed, not the event length
2830 * which also includes the event header.
2832 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2833 * If NULL is returned, then nothing has been allocated or locked.
2835 struct ring_buffer_event *
2836 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2838 struct ring_buffer_per_cpu *cpu_buffer;
2839 struct ring_buffer_event *event;
2842 /* If we are tracing schedule, we don't want to recurse */
2843 preempt_disable_notrace();
2845 if (unlikely(atomic_read(&buffer->record_disabled)))
2848 cpu = raw_smp_processor_id();
2850 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2853 cpu_buffer = buffer->buffers[cpu];
2855 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2858 if (unlikely(length > BUF_MAX_DATA_SIZE))
2861 if (unlikely(trace_recursive_lock(cpu_buffer)))
2864 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2871 trace_recursive_unlock(cpu_buffer);
2873 preempt_enable_notrace();
2876 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2879 * Decrement the entries to the page that an event is on.
2880 * The event does not even need to exist, only the pointer
2881 * to the page it is on. This may only be called before the commit
2885 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2886 struct ring_buffer_event *event)
2888 unsigned long addr = (unsigned long)event;
2889 struct buffer_page *bpage = cpu_buffer->commit_page;
2890 struct buffer_page *start;
2894 /* Do the likely case first */
2895 if (likely(bpage->page == (void *)addr)) {
2896 local_dec(&bpage->entries);
2901 * Because the commit page may be on the reader page we
2902 * start with the next page and check the end loop there.
2904 rb_inc_page(cpu_buffer, &bpage);
2907 if (bpage->page == (void *)addr) {
2908 local_dec(&bpage->entries);
2911 rb_inc_page(cpu_buffer, &bpage);
2912 } while (bpage != start);
2914 /* commit not part of this buffer?? */
2915 RB_WARN_ON(cpu_buffer, 1);
2919 * ring_buffer_commit_discard - discard an event that has not been committed
2920 * @buffer: the ring buffer
2921 * @event: non committed event to discard
2923 * Sometimes an event that is in the ring buffer needs to be ignored.
2924 * This function lets the user discard an event in the ring buffer
2925 * and then that event will not be read later.
2927 * This function only works if it is called before the the item has been
2928 * committed. It will try to free the event from the ring buffer
2929 * if another event has not been added behind it.
2931 * If another event has been added behind it, it will set the event
2932 * up as discarded, and perform the commit.
2934 * If this function is called, do not call ring_buffer_unlock_commit on
2937 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2938 struct ring_buffer_event *event)
2940 struct ring_buffer_per_cpu *cpu_buffer;
2943 /* The event is discarded regardless */
2944 rb_event_discard(event);
2946 cpu = smp_processor_id();
2947 cpu_buffer = buffer->buffers[cpu];
2950 * This must only be called if the event has not been
2951 * committed yet. Thus we can assume that preemption
2952 * is still disabled.
2954 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2956 rb_decrement_entry(cpu_buffer, event);
2957 if (rb_try_to_discard(cpu_buffer, event))
2961 * The commit is still visible by the reader, so we
2962 * must still update the timestamp.
2964 rb_update_write_stamp(cpu_buffer, event);
2966 rb_end_commit(cpu_buffer);
2968 trace_recursive_unlock(cpu_buffer);
2970 preempt_enable_notrace();
2973 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2976 * ring_buffer_write - write data to the buffer without reserving
2977 * @buffer: The ring buffer to write to.
2978 * @length: The length of the data being written (excluding the event header)
2979 * @data: The data to write to the buffer.
2981 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2982 * one function. If you already have the data to write to the buffer, it
2983 * may be easier to simply call this function.
2985 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2986 * and not the length of the event which would hold the header.
2988 int ring_buffer_write(struct ring_buffer *buffer,
2989 unsigned long length,
2992 struct ring_buffer_per_cpu *cpu_buffer;
2993 struct ring_buffer_event *event;
2998 preempt_disable_notrace();
3000 if (atomic_read(&buffer->record_disabled))
3003 cpu = raw_smp_processor_id();
3005 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3008 cpu_buffer = buffer->buffers[cpu];
3010 if (atomic_read(&cpu_buffer->record_disabled))
3013 if (length > BUF_MAX_DATA_SIZE)
3016 if (unlikely(trace_recursive_lock(cpu_buffer)))
3019 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3023 body = rb_event_data(event);
3025 memcpy(body, data, length);
3027 rb_commit(cpu_buffer, event);
3029 rb_wakeups(buffer, cpu_buffer);
3034 trace_recursive_unlock(cpu_buffer);
3037 preempt_enable_notrace();
3041 EXPORT_SYMBOL_GPL(ring_buffer_write);
3043 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3045 struct buffer_page *reader = cpu_buffer->reader_page;
3046 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3047 struct buffer_page *commit = cpu_buffer->commit_page;
3049 /* In case of error, head will be NULL */
3050 if (unlikely(!head))
3053 return reader->read == rb_page_commit(reader) &&
3054 (commit == reader ||
3056 head->read == rb_page_commit(commit)));
3060 * ring_buffer_record_disable - stop all writes into the buffer
3061 * @buffer: The ring buffer to stop writes to.
3063 * This prevents all writes to the buffer. Any attempt to write
3064 * to the buffer after this will fail and return NULL.
3066 * The caller should call synchronize_sched() after this.
3068 void ring_buffer_record_disable(struct ring_buffer *buffer)
3070 atomic_inc(&buffer->record_disabled);
3072 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3075 * ring_buffer_record_enable - enable writes to the buffer
3076 * @buffer: The ring buffer to enable writes
3078 * Note, multiple disables will need the same number of enables
3079 * to truly enable the writing (much like preempt_disable).
3081 void ring_buffer_record_enable(struct ring_buffer *buffer)
3083 atomic_dec(&buffer->record_disabled);
3085 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3088 * ring_buffer_record_off - stop all writes into the buffer
3089 * @buffer: The ring buffer to stop writes to.
3091 * This prevents all writes to the buffer. Any attempt to write
3092 * to the buffer after this will fail and return NULL.
3094 * This is different than ring_buffer_record_disable() as
3095 * it works like an on/off switch, where as the disable() version
3096 * must be paired with a enable().
3098 void ring_buffer_record_off(struct ring_buffer *buffer)
3101 unsigned int new_rd;
3104 rd = atomic_read(&buffer->record_disabled);
3105 new_rd = rd | RB_BUFFER_OFF;
3106 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3108 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3111 * ring_buffer_record_on - restart writes into the buffer
3112 * @buffer: The ring buffer to start writes to.
3114 * This enables all writes to the buffer that was disabled by
3115 * ring_buffer_record_off().
3117 * This is different than ring_buffer_record_enable() as
3118 * it works like an on/off switch, where as the enable() version
3119 * must be paired with a disable().
3121 void ring_buffer_record_on(struct ring_buffer *buffer)
3124 unsigned int new_rd;
3127 rd = atomic_read(&buffer->record_disabled);
3128 new_rd = rd & ~RB_BUFFER_OFF;
3129 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3131 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3134 * ring_buffer_record_is_on - return true if the ring buffer can write
3135 * @buffer: The ring buffer to see if write is enabled
3137 * Returns true if the ring buffer is in a state that it accepts writes.
3139 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3141 return !atomic_read(&buffer->record_disabled);
3145 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
3146 * @buffer: The ring buffer to see if write is set enabled
3148 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
3149 * Note that this does NOT mean it is in a writable state.
3151 * It may return true when the ring buffer has been disabled by
3152 * ring_buffer_record_disable(), as that is a temporary disabling of
3155 int ring_buffer_record_is_set_on(struct ring_buffer *buffer)
3157 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
3161 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3162 * @buffer: The ring buffer to stop writes to.
3163 * @cpu: The CPU buffer to stop
3165 * This prevents all writes to the buffer. Any attempt to write
3166 * to the buffer after this will fail and return NULL.
3168 * The caller should call synchronize_sched() after this.
3170 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3172 struct ring_buffer_per_cpu *cpu_buffer;
3174 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3177 cpu_buffer = buffer->buffers[cpu];
3178 atomic_inc(&cpu_buffer->record_disabled);
3180 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3183 * ring_buffer_record_enable_cpu - enable writes to the buffer
3184 * @buffer: The ring buffer to enable writes
3185 * @cpu: The CPU to enable.
3187 * Note, multiple disables will need the same number of enables
3188 * to truly enable the writing (much like preempt_disable).
3190 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3192 struct ring_buffer_per_cpu *cpu_buffer;
3194 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3197 cpu_buffer = buffer->buffers[cpu];
3198 atomic_dec(&cpu_buffer->record_disabled);
3200 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3203 * The total entries in the ring buffer is the running counter
3204 * of entries entered into the ring buffer, minus the sum of
3205 * the entries read from the ring buffer and the number of
3206 * entries that were overwritten.
3208 static inline unsigned long
3209 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3211 return local_read(&cpu_buffer->entries) -
3212 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3216 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3217 * @buffer: The ring buffer
3218 * @cpu: The per CPU buffer to read from.
3220 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3222 unsigned long flags;
3223 struct ring_buffer_per_cpu *cpu_buffer;
3224 struct buffer_page *bpage;
3227 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3230 cpu_buffer = buffer->buffers[cpu];
3231 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3233 * if the tail is on reader_page, oldest time stamp is on the reader
3236 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3237 bpage = cpu_buffer->reader_page;
3239 bpage = rb_set_head_page(cpu_buffer);
3241 ret = bpage->page->time_stamp;
3242 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3246 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3249 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3250 * @buffer: The ring buffer
3251 * @cpu: The per CPU buffer to read from.
3253 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3255 struct ring_buffer_per_cpu *cpu_buffer;
3258 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3261 cpu_buffer = buffer->buffers[cpu];
3262 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3266 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3269 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3270 * @buffer: The ring buffer
3271 * @cpu: The per CPU buffer to get the entries from.
3273 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3275 struct ring_buffer_per_cpu *cpu_buffer;
3277 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3280 cpu_buffer = buffer->buffers[cpu];
3282 return rb_num_of_entries(cpu_buffer);
3284 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3287 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3288 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3289 * @buffer: The ring buffer
3290 * @cpu: The per CPU buffer to get the number of overruns from
3292 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3294 struct ring_buffer_per_cpu *cpu_buffer;
3297 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3300 cpu_buffer = buffer->buffers[cpu];
3301 ret = local_read(&cpu_buffer->overrun);
3305 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3308 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3309 * commits failing due to the buffer wrapping around while there are uncommitted
3310 * events, such as during an interrupt storm.
3311 * @buffer: The ring buffer
3312 * @cpu: The per CPU buffer to get the number of overruns from
3315 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3317 struct ring_buffer_per_cpu *cpu_buffer;
3320 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3323 cpu_buffer = buffer->buffers[cpu];
3324 ret = local_read(&cpu_buffer->commit_overrun);
3328 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3331 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3332 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3333 * @buffer: The ring buffer
3334 * @cpu: The per CPU buffer to get the number of overruns from
3337 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3339 struct ring_buffer_per_cpu *cpu_buffer;
3342 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3345 cpu_buffer = buffer->buffers[cpu];
3346 ret = local_read(&cpu_buffer->dropped_events);
3350 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3353 * ring_buffer_read_events_cpu - get the number of events successfully read
3354 * @buffer: The ring buffer
3355 * @cpu: The per CPU buffer to get the number of events read
3358 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3360 struct ring_buffer_per_cpu *cpu_buffer;
3362 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3365 cpu_buffer = buffer->buffers[cpu];
3366 return cpu_buffer->read;
3368 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3371 * ring_buffer_entries - get the number of entries in a buffer
3372 * @buffer: The ring buffer
3374 * Returns the total number of entries in the ring buffer
3377 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3379 struct ring_buffer_per_cpu *cpu_buffer;
3380 unsigned long entries = 0;
3383 /* if you care about this being correct, lock the buffer */
3384 for_each_buffer_cpu(buffer, cpu) {
3385 cpu_buffer = buffer->buffers[cpu];
3386 entries += rb_num_of_entries(cpu_buffer);
3391 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3394 * ring_buffer_overruns - get the number of overruns in buffer
3395 * @buffer: The ring buffer
3397 * Returns the total number of overruns in the ring buffer
3400 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3402 struct ring_buffer_per_cpu *cpu_buffer;
3403 unsigned long overruns = 0;
3406 /* if you care about this being correct, lock the buffer */
3407 for_each_buffer_cpu(buffer, cpu) {
3408 cpu_buffer = buffer->buffers[cpu];
3409 overruns += local_read(&cpu_buffer->overrun);
3414 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3416 static void rb_iter_reset(struct ring_buffer_iter *iter)
3418 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3420 /* Iterator usage is expected to have record disabled */
3421 iter->head_page = cpu_buffer->reader_page;
3422 iter->head = cpu_buffer->reader_page->read;
3424 iter->cache_reader_page = iter->head_page;
3425 iter->cache_read = cpu_buffer->read;
3428 iter->read_stamp = cpu_buffer->read_stamp;
3430 iter->read_stamp = iter->head_page->page->time_stamp;
3434 * ring_buffer_iter_reset - reset an iterator
3435 * @iter: The iterator to reset
3437 * Resets the iterator, so that it will start from the beginning
3440 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3442 struct ring_buffer_per_cpu *cpu_buffer;
3443 unsigned long flags;
3448 cpu_buffer = iter->cpu_buffer;
3450 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3451 rb_iter_reset(iter);
3452 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3454 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3457 * ring_buffer_iter_empty - check if an iterator has no more to read
3458 * @iter: The iterator to check
3460 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3462 struct ring_buffer_per_cpu *cpu_buffer;
3463 struct buffer_page *reader;
3464 struct buffer_page *head_page;
3465 struct buffer_page *commit_page;
3468 cpu_buffer = iter->cpu_buffer;
3470 /* Remember, trace recording is off when iterator is in use */
3471 reader = cpu_buffer->reader_page;
3472 head_page = cpu_buffer->head_page;
3473 commit_page = cpu_buffer->commit_page;
3474 commit = rb_page_commit(commit_page);
3476 return ((iter->head_page == commit_page && iter->head == commit) ||
3477 (iter->head_page == reader && commit_page == head_page &&
3478 head_page->read == commit &&
3479 iter->head == rb_page_commit(cpu_buffer->reader_page)));
3481 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3484 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3485 struct ring_buffer_event *event)
3489 switch (event->type_len) {
3490 case RINGBUF_TYPE_PADDING:
3493 case RINGBUF_TYPE_TIME_EXTEND:
3494 delta = event->array[0];
3496 delta += event->time_delta;
3497 cpu_buffer->read_stamp += delta;
3500 case RINGBUF_TYPE_TIME_STAMP:
3501 /* FIXME: not implemented */
3504 case RINGBUF_TYPE_DATA:
3505 cpu_buffer->read_stamp += event->time_delta;
3515 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3516 struct ring_buffer_event *event)
3520 switch (event->type_len) {
3521 case RINGBUF_TYPE_PADDING:
3524 case RINGBUF_TYPE_TIME_EXTEND:
3525 delta = event->array[0];
3527 delta += event->time_delta;
3528 iter->read_stamp += delta;
3531 case RINGBUF_TYPE_TIME_STAMP:
3532 /* FIXME: not implemented */
3535 case RINGBUF_TYPE_DATA:
3536 iter->read_stamp += event->time_delta;
3545 static struct buffer_page *
3546 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3548 struct buffer_page *reader = NULL;
3549 unsigned long overwrite;
3550 unsigned long flags;
3554 local_irq_save(flags);
3555 arch_spin_lock(&cpu_buffer->lock);
3559 * This should normally only loop twice. But because the
3560 * start of the reader inserts an empty page, it causes
3561 * a case where we will loop three times. There should be no
3562 * reason to loop four times (that I know of).
3564 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3569 reader = cpu_buffer->reader_page;
3571 /* If there's more to read, return this page */
3572 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3575 /* Never should we have an index greater than the size */
3576 if (RB_WARN_ON(cpu_buffer,
3577 cpu_buffer->reader_page->read > rb_page_size(reader)))
3580 /* check if we caught up to the tail */
3582 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3585 /* Don't bother swapping if the ring buffer is empty */
3586 if (rb_num_of_entries(cpu_buffer) == 0)
3590 * Reset the reader page to size zero.
3592 local_set(&cpu_buffer->reader_page->write, 0);
3593 local_set(&cpu_buffer->reader_page->entries, 0);
3594 local_set(&cpu_buffer->reader_page->page->commit, 0);
3595 cpu_buffer->reader_page->real_end = 0;
3599 * Splice the empty reader page into the list around the head.
3601 reader = rb_set_head_page(cpu_buffer);
3604 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3605 cpu_buffer->reader_page->list.prev = reader->list.prev;
3608 * cpu_buffer->pages just needs to point to the buffer, it
3609 * has no specific buffer page to point to. Lets move it out
3610 * of our way so we don't accidentally swap it.
3612 cpu_buffer->pages = reader->list.prev;
3614 /* The reader page will be pointing to the new head */
3615 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3618 * We want to make sure we read the overruns after we set up our
3619 * pointers to the next object. The writer side does a
3620 * cmpxchg to cross pages which acts as the mb on the writer
3621 * side. Note, the reader will constantly fail the swap
3622 * while the writer is updating the pointers, so this
3623 * guarantees that the overwrite recorded here is the one we
3624 * want to compare with the last_overrun.
3627 overwrite = local_read(&(cpu_buffer->overrun));
3630 * Here's the tricky part.
3632 * We need to move the pointer past the header page.
3633 * But we can only do that if a writer is not currently
3634 * moving it. The page before the header page has the
3635 * flag bit '1' set if it is pointing to the page we want.
3636 * but if the writer is in the process of moving it
3637 * than it will be '2' or already moved '0'.
3640 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3643 * If we did not convert it, then we must try again.
3649 * Yeah! We succeeded in replacing the page.
3651 * Now make the new head point back to the reader page.
3653 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3654 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3656 /* Finally update the reader page to the new head */
3657 cpu_buffer->reader_page = reader;
3658 cpu_buffer->reader_page->read = 0;
3660 if (overwrite != cpu_buffer->last_overrun) {
3661 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3662 cpu_buffer->last_overrun = overwrite;
3668 /* Update the read_stamp on the first event */
3669 if (reader && reader->read == 0)
3670 cpu_buffer->read_stamp = reader->page->time_stamp;
3672 arch_spin_unlock(&cpu_buffer->lock);
3673 local_irq_restore(flags);
3678 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3680 struct ring_buffer_event *event;
3681 struct buffer_page *reader;
3684 reader = rb_get_reader_page(cpu_buffer);
3686 /* This function should not be called when buffer is empty */
3687 if (RB_WARN_ON(cpu_buffer, !reader))
3690 event = rb_reader_event(cpu_buffer);
3692 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3695 rb_update_read_stamp(cpu_buffer, event);
3697 length = rb_event_length(event);
3698 cpu_buffer->reader_page->read += length;
3701 static void rb_advance_iter(struct ring_buffer_iter *iter)
3703 struct ring_buffer_per_cpu *cpu_buffer;
3704 struct ring_buffer_event *event;
3707 cpu_buffer = iter->cpu_buffer;
3710 * Check if we are at the end of the buffer.
3712 if (iter->head >= rb_page_size(iter->head_page)) {
3713 /* discarded commits can make the page empty */
3714 if (iter->head_page == cpu_buffer->commit_page)
3720 event = rb_iter_head_event(iter);
3722 length = rb_event_length(event);
3725 * This should not be called to advance the header if we are
3726 * at the tail of the buffer.
3728 if (RB_WARN_ON(cpu_buffer,
3729 (iter->head_page == cpu_buffer->commit_page) &&
3730 (iter->head + length > rb_commit_index(cpu_buffer))))
3733 rb_update_iter_read_stamp(iter, event);
3735 iter->head += length;
3737 /* check for end of page padding */
3738 if ((iter->head >= rb_page_size(iter->head_page)) &&
3739 (iter->head_page != cpu_buffer->commit_page))
3743 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3745 return cpu_buffer->lost_events;
3748 static struct ring_buffer_event *
3749 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3750 unsigned long *lost_events)
3752 struct ring_buffer_event *event;
3753 struct buffer_page *reader;
3758 * We repeat when a time extend is encountered.
3759 * Since the time extend is always attached to a data event,
3760 * we should never loop more than once.
3761 * (We never hit the following condition more than twice).
3763 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3766 reader = rb_get_reader_page(cpu_buffer);
3770 event = rb_reader_event(cpu_buffer);
3772 switch (event->type_len) {
3773 case RINGBUF_TYPE_PADDING:
3774 if (rb_null_event(event))
3775 RB_WARN_ON(cpu_buffer, 1);
3777 * Because the writer could be discarding every
3778 * event it creates (which would probably be bad)
3779 * if we were to go back to "again" then we may never
3780 * catch up, and will trigger the warn on, or lock
3781 * the box. Return the padding, and we will release
3782 * the current locks, and try again.
3786 case RINGBUF_TYPE_TIME_EXTEND:
3787 /* Internal data, OK to advance */
3788 rb_advance_reader(cpu_buffer);
3791 case RINGBUF_TYPE_TIME_STAMP:
3792 /* FIXME: not implemented */
3793 rb_advance_reader(cpu_buffer);
3796 case RINGBUF_TYPE_DATA:
3798 *ts = cpu_buffer->read_stamp + event->time_delta;
3799 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3800 cpu_buffer->cpu, ts);
3803 *lost_events = rb_lost_events(cpu_buffer);
3812 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3814 static struct ring_buffer_event *
3815 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3817 struct ring_buffer *buffer;
3818 struct ring_buffer_per_cpu *cpu_buffer;
3819 struct ring_buffer_event *event;
3822 cpu_buffer = iter->cpu_buffer;
3823 buffer = cpu_buffer->buffer;
3826 * Check if someone performed a consuming read to
3827 * the buffer. A consuming read invalidates the iterator
3828 * and we need to reset the iterator in this case.
3830 if (unlikely(iter->cache_read != cpu_buffer->read ||
3831 iter->cache_reader_page != cpu_buffer->reader_page))
3832 rb_iter_reset(iter);
3835 if (ring_buffer_iter_empty(iter))
3839 * We repeat when a time extend is encountered or we hit
3840 * the end of the page. Since the time extend is always attached
3841 * to a data event, we should never loop more than three times.
3842 * Once for going to next page, once on time extend, and
3843 * finally once to get the event.
3844 * (We never hit the following condition more than thrice).
3846 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3849 if (rb_per_cpu_empty(cpu_buffer))
3852 if (iter->head >= rb_page_size(iter->head_page)) {
3857 event = rb_iter_head_event(iter);
3859 switch (event->type_len) {
3860 case RINGBUF_TYPE_PADDING:
3861 if (rb_null_event(event)) {
3865 rb_advance_iter(iter);
3868 case RINGBUF_TYPE_TIME_EXTEND:
3869 /* Internal data, OK to advance */
3870 rb_advance_iter(iter);
3873 case RINGBUF_TYPE_TIME_STAMP:
3874 /* FIXME: not implemented */
3875 rb_advance_iter(iter);
3878 case RINGBUF_TYPE_DATA:
3880 *ts = iter->read_stamp + event->time_delta;
3881 ring_buffer_normalize_time_stamp(buffer,
3882 cpu_buffer->cpu, ts);
3892 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3894 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
3896 if (likely(!in_nmi())) {
3897 raw_spin_lock(&cpu_buffer->reader_lock);
3902 * If an NMI die dumps out the content of the ring buffer
3903 * trylock must be used to prevent a deadlock if the NMI
3904 * preempted a task that holds the ring buffer locks. If
3905 * we get the lock then all is fine, if not, then continue
3906 * to do the read, but this can corrupt the ring buffer,
3907 * so it must be permanently disabled from future writes.
3908 * Reading from NMI is a oneshot deal.
3910 if (raw_spin_trylock(&cpu_buffer->reader_lock))
3913 /* Continue without locking, but disable the ring buffer */
3914 atomic_inc(&cpu_buffer->record_disabled);
3919 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
3922 raw_spin_unlock(&cpu_buffer->reader_lock);
3927 * ring_buffer_peek - peek at the next event to be read
3928 * @buffer: The ring buffer to read
3929 * @cpu: The cpu to peak at
3930 * @ts: The timestamp counter of this event.
3931 * @lost_events: a variable to store if events were lost (may be NULL)
3933 * This will return the event that will be read next, but does
3934 * not consume the data.
3936 struct ring_buffer_event *
3937 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3938 unsigned long *lost_events)
3940 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3941 struct ring_buffer_event *event;
3942 unsigned long flags;
3945 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3949 local_irq_save(flags);
3950 dolock = rb_reader_lock(cpu_buffer);
3951 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3952 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3953 rb_advance_reader(cpu_buffer);
3954 rb_reader_unlock(cpu_buffer, dolock);
3955 local_irq_restore(flags);
3957 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3964 * ring_buffer_iter_peek - peek at the next event to be read
3965 * @iter: The ring buffer iterator
3966 * @ts: The timestamp counter of this event.
3968 * This will return the event that will be read next, but does
3969 * not increment the iterator.
3971 struct ring_buffer_event *
3972 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3974 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3975 struct ring_buffer_event *event;
3976 unsigned long flags;
3979 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3980 event = rb_iter_peek(iter, ts);
3981 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3983 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3990 * ring_buffer_consume - return an event and consume it
3991 * @buffer: The ring buffer to get the next event from
3992 * @cpu: the cpu to read the buffer from
3993 * @ts: a variable to store the timestamp (may be NULL)
3994 * @lost_events: a variable to store if events were lost (may be NULL)
3996 * Returns the next event in the ring buffer, and that event is consumed.
3997 * Meaning, that sequential reads will keep returning a different event,
3998 * and eventually empty the ring buffer if the producer is slower.
4000 struct ring_buffer_event *
4001 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
4002 unsigned long *lost_events)
4004 struct ring_buffer_per_cpu *cpu_buffer;
4005 struct ring_buffer_event *event = NULL;
4006 unsigned long flags;
4010 /* might be called in atomic */
4013 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4016 cpu_buffer = buffer->buffers[cpu];
4017 local_irq_save(flags);
4018 dolock = rb_reader_lock(cpu_buffer);
4020 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4022 cpu_buffer->lost_events = 0;
4023 rb_advance_reader(cpu_buffer);
4026 rb_reader_unlock(cpu_buffer, dolock);
4027 local_irq_restore(flags);
4032 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4037 EXPORT_SYMBOL_GPL(ring_buffer_consume);
4040 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4041 * @buffer: The ring buffer to read from
4042 * @cpu: The cpu buffer to iterate over
4044 * This performs the initial preparations necessary to iterate
4045 * through the buffer. Memory is allocated, buffer recording
4046 * is disabled, and the iterator pointer is returned to the caller.
4048 * Disabling buffer recordng prevents the reading from being
4049 * corrupted. This is not a consuming read, so a producer is not
4052 * After a sequence of ring_buffer_read_prepare calls, the user is
4053 * expected to make at least one call to ring_buffer_read_prepare_sync.
4054 * Afterwards, ring_buffer_read_start is invoked to get things going
4057 * This overall must be paired with ring_buffer_read_finish.
4059 struct ring_buffer_iter *
4060 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
4062 struct ring_buffer_per_cpu *cpu_buffer;
4063 struct ring_buffer_iter *iter;
4065 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4068 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
4072 cpu_buffer = buffer->buffers[cpu];
4074 iter->cpu_buffer = cpu_buffer;
4076 atomic_inc(&buffer->resize_disabled);
4077 atomic_inc(&cpu_buffer->record_disabled);
4081 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4084 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4086 * All previously invoked ring_buffer_read_prepare calls to prepare
4087 * iterators will be synchronized. Afterwards, read_buffer_read_start
4088 * calls on those iterators are allowed.
4091 ring_buffer_read_prepare_sync(void)
4093 synchronize_sched();
4095 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4098 * ring_buffer_read_start - start a non consuming read of the buffer
4099 * @iter: The iterator returned by ring_buffer_read_prepare
4101 * This finalizes the startup of an iteration through the buffer.
4102 * The iterator comes from a call to ring_buffer_read_prepare and
4103 * an intervening ring_buffer_read_prepare_sync must have been
4106 * Must be paired with ring_buffer_read_finish.
4109 ring_buffer_read_start(struct ring_buffer_iter *iter)
4111 struct ring_buffer_per_cpu *cpu_buffer;
4112 unsigned long flags;
4117 cpu_buffer = iter->cpu_buffer;
4119 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4120 arch_spin_lock(&cpu_buffer->lock);
4121 rb_iter_reset(iter);
4122 arch_spin_unlock(&cpu_buffer->lock);
4123 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4125 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4128 * ring_buffer_read_finish - finish reading the iterator of the buffer
4129 * @iter: The iterator retrieved by ring_buffer_start
4131 * This re-enables the recording to the buffer, and frees the
4135 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4137 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4138 unsigned long flags;
4141 * Ring buffer is disabled from recording, here's a good place
4142 * to check the integrity of the ring buffer.
4143 * Must prevent readers from trying to read, as the check
4144 * clears the HEAD page and readers require it.
4146 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4147 rb_check_pages(cpu_buffer);
4148 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4150 atomic_dec(&cpu_buffer->record_disabled);
4151 atomic_dec(&cpu_buffer->buffer->resize_disabled);
4154 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4157 * ring_buffer_read - read the next item in the ring buffer by the iterator
4158 * @iter: The ring buffer iterator
4159 * @ts: The time stamp of the event read.
4161 * This reads the next event in the ring buffer and increments the iterator.
4163 struct ring_buffer_event *
4164 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4166 struct ring_buffer_event *event;
4167 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4168 unsigned long flags;
4170 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4172 event = rb_iter_peek(iter, ts);
4176 if (event->type_len == RINGBUF_TYPE_PADDING)
4179 rb_advance_iter(iter);
4181 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4185 EXPORT_SYMBOL_GPL(ring_buffer_read);
4188 * ring_buffer_size - return the size of the ring buffer (in bytes)
4189 * @buffer: The ring buffer.
4191 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4194 * Earlier, this method returned
4195 * BUF_PAGE_SIZE * buffer->nr_pages
4196 * Since the nr_pages field is now removed, we have converted this to
4197 * return the per cpu buffer value.
4199 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4202 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4204 EXPORT_SYMBOL_GPL(ring_buffer_size);
4207 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4209 rb_head_page_deactivate(cpu_buffer);
4211 cpu_buffer->head_page
4212 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4213 local_set(&cpu_buffer->head_page->write, 0);
4214 local_set(&cpu_buffer->head_page->entries, 0);
4215 local_set(&cpu_buffer->head_page->page->commit, 0);
4217 cpu_buffer->head_page->read = 0;
4219 cpu_buffer->tail_page = cpu_buffer->head_page;
4220 cpu_buffer->commit_page = cpu_buffer->head_page;
4222 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4223 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4224 local_set(&cpu_buffer->reader_page->write, 0);
4225 local_set(&cpu_buffer->reader_page->entries, 0);
4226 local_set(&cpu_buffer->reader_page->page->commit, 0);
4227 cpu_buffer->reader_page->read = 0;
4229 local_set(&cpu_buffer->entries_bytes, 0);
4230 local_set(&cpu_buffer->overrun, 0);
4231 local_set(&cpu_buffer->commit_overrun, 0);
4232 local_set(&cpu_buffer->dropped_events, 0);
4233 local_set(&cpu_buffer->entries, 0);
4234 local_set(&cpu_buffer->committing, 0);
4235 local_set(&cpu_buffer->commits, 0);
4236 cpu_buffer->read = 0;
4237 cpu_buffer->read_bytes = 0;
4239 cpu_buffer->write_stamp = 0;
4240 cpu_buffer->read_stamp = 0;
4242 cpu_buffer->lost_events = 0;
4243 cpu_buffer->last_overrun = 0;
4245 rb_head_page_activate(cpu_buffer);
4249 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4250 * @buffer: The ring buffer to reset a per cpu buffer of
4251 * @cpu: The CPU buffer to be reset
4253 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4255 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4256 unsigned long flags;
4258 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4261 atomic_inc(&buffer->resize_disabled);
4262 atomic_inc(&cpu_buffer->record_disabled);
4264 /* Make sure all commits have finished */
4265 synchronize_sched();
4267 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4269 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4272 arch_spin_lock(&cpu_buffer->lock);
4274 rb_reset_cpu(cpu_buffer);
4276 arch_spin_unlock(&cpu_buffer->lock);
4279 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4281 atomic_dec(&cpu_buffer->record_disabled);
4282 atomic_dec(&buffer->resize_disabled);
4284 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4287 * ring_buffer_reset - reset a ring buffer
4288 * @buffer: The ring buffer to reset all cpu buffers
4290 void ring_buffer_reset(struct ring_buffer *buffer)
4294 for_each_buffer_cpu(buffer, cpu)
4295 ring_buffer_reset_cpu(buffer, cpu);
4297 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4300 * rind_buffer_empty - is the ring buffer empty?
4301 * @buffer: The ring buffer to test
4303 bool ring_buffer_empty(struct ring_buffer *buffer)
4305 struct ring_buffer_per_cpu *cpu_buffer;
4306 unsigned long flags;
4311 /* yes this is racy, but if you don't like the race, lock the buffer */
4312 for_each_buffer_cpu(buffer, cpu) {
4313 cpu_buffer = buffer->buffers[cpu];
4314 local_irq_save(flags);
4315 dolock = rb_reader_lock(cpu_buffer);
4316 ret = rb_per_cpu_empty(cpu_buffer);
4317 rb_reader_unlock(cpu_buffer, dolock);
4318 local_irq_restore(flags);
4326 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4329 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4330 * @buffer: The ring buffer
4331 * @cpu: The CPU buffer to test
4333 bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4335 struct ring_buffer_per_cpu *cpu_buffer;
4336 unsigned long flags;
4340 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4343 cpu_buffer = buffer->buffers[cpu];
4344 local_irq_save(flags);
4345 dolock = rb_reader_lock(cpu_buffer);
4346 ret = rb_per_cpu_empty(cpu_buffer);
4347 rb_reader_unlock(cpu_buffer, dolock);
4348 local_irq_restore(flags);
4352 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4354 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4356 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4357 * @buffer_a: One buffer to swap with
4358 * @buffer_b: The other buffer to swap with
4360 * This function is useful for tracers that want to take a "snapshot"
4361 * of a CPU buffer and has another back up buffer lying around.
4362 * it is expected that the tracer handles the cpu buffer not being
4363 * used at the moment.
4365 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4366 struct ring_buffer *buffer_b, int cpu)
4368 struct ring_buffer_per_cpu *cpu_buffer_a;
4369 struct ring_buffer_per_cpu *cpu_buffer_b;
4372 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4373 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4376 cpu_buffer_a = buffer_a->buffers[cpu];
4377 cpu_buffer_b = buffer_b->buffers[cpu];
4379 /* At least make sure the two buffers are somewhat the same */
4380 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4385 if (atomic_read(&buffer_a->record_disabled))
4388 if (atomic_read(&buffer_b->record_disabled))
4391 if (atomic_read(&cpu_buffer_a->record_disabled))
4394 if (atomic_read(&cpu_buffer_b->record_disabled))
4398 * We can't do a synchronize_sched here because this
4399 * function can be called in atomic context.
4400 * Normally this will be called from the same CPU as cpu.
4401 * If not it's up to the caller to protect this.
4403 atomic_inc(&cpu_buffer_a->record_disabled);
4404 atomic_inc(&cpu_buffer_b->record_disabled);
4407 if (local_read(&cpu_buffer_a->committing))
4409 if (local_read(&cpu_buffer_b->committing))
4412 buffer_a->buffers[cpu] = cpu_buffer_b;
4413 buffer_b->buffers[cpu] = cpu_buffer_a;
4415 cpu_buffer_b->buffer = buffer_a;
4416 cpu_buffer_a->buffer = buffer_b;
4421 atomic_dec(&cpu_buffer_a->record_disabled);
4422 atomic_dec(&cpu_buffer_b->record_disabled);
4426 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4427 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4430 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4431 * @buffer: the buffer to allocate for.
4432 * @cpu: the cpu buffer to allocate.
4434 * This function is used in conjunction with ring_buffer_read_page.
4435 * When reading a full page from the ring buffer, these functions
4436 * can be used to speed up the process. The calling function should
4437 * allocate a few pages first with this function. Then when it
4438 * needs to get pages from the ring buffer, it passes the result
4439 * of this function into ring_buffer_read_page, which will swap
4440 * the page that was allocated, with the read page of the buffer.
4443 * The page allocated, or NULL on error.
4445 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4447 struct buffer_data_page *bpage;
4450 page = alloc_pages_node(cpu_to_node(cpu),
4451 GFP_KERNEL | __GFP_NORETRY, 0);
4455 bpage = page_address(page);
4457 rb_init_page(bpage);
4461 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4464 * ring_buffer_free_read_page - free an allocated read page
4465 * @buffer: the buffer the page was allocate for
4466 * @data: the page to free
4468 * Free a page allocated from ring_buffer_alloc_read_page.
4470 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4472 free_page((unsigned long)data);
4474 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4477 * ring_buffer_read_page - extract a page from the ring buffer
4478 * @buffer: buffer to extract from
4479 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4480 * @len: amount to extract
4481 * @cpu: the cpu of the buffer to extract
4482 * @full: should the extraction only happen when the page is full.
4484 * This function will pull out a page from the ring buffer and consume it.
4485 * @data_page must be the address of the variable that was returned
4486 * from ring_buffer_alloc_read_page. This is because the page might be used
4487 * to swap with a page in the ring buffer.
4490 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4493 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4495 * process_page(rpage, ret);
4497 * When @full is set, the function will not return true unless
4498 * the writer is off the reader page.
4500 * Note: it is up to the calling functions to handle sleeps and wakeups.
4501 * The ring buffer can be used anywhere in the kernel and can not
4502 * blindly call wake_up. The layer that uses the ring buffer must be
4503 * responsible for that.
4506 * >=0 if data has been transferred, returns the offset of consumed data.
4507 * <0 if no data has been transferred.
4509 int ring_buffer_read_page(struct ring_buffer *buffer,
4510 void **data_page, size_t len, int cpu, int full)
4512 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4513 struct ring_buffer_event *event;
4514 struct buffer_data_page *bpage;
4515 struct buffer_page *reader;
4516 unsigned long missed_events;
4517 unsigned long flags;
4518 unsigned int commit;
4523 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4527 * If len is not big enough to hold the page header, then
4528 * we can not copy anything.
4530 if (len <= BUF_PAGE_HDR_SIZE)
4533 len -= BUF_PAGE_HDR_SIZE;
4542 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4544 reader = rb_get_reader_page(cpu_buffer);
4548 event = rb_reader_event(cpu_buffer);
4550 read = reader->read;
4551 commit = rb_page_commit(reader);
4553 /* Check if any events were dropped */
4554 missed_events = cpu_buffer->lost_events;
4557 * If this page has been partially read or
4558 * if len is not big enough to read the rest of the page or
4559 * a writer is still on the page, then
4560 * we must copy the data from the page to the buffer.
4561 * Otherwise, we can simply swap the page with the one passed in.
4563 if (read || (len < (commit - read)) ||
4564 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4565 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4566 unsigned int rpos = read;
4567 unsigned int pos = 0;
4573 if (len > (commit - read))
4574 len = (commit - read);
4576 /* Always keep the time extend and data together */
4577 size = rb_event_ts_length(event);
4582 /* save the current timestamp, since the user will need it */
4583 save_timestamp = cpu_buffer->read_stamp;
4585 /* Need to copy one event at a time */
4587 /* We need the size of one event, because
4588 * rb_advance_reader only advances by one event,
4589 * whereas rb_event_ts_length may include the size of
4590 * one or two events.
4591 * We have already ensured there's enough space if this
4592 * is a time extend. */
4593 size = rb_event_length(event);
4594 memcpy(bpage->data + pos, rpage->data + rpos, size);
4598 rb_advance_reader(cpu_buffer);
4599 rpos = reader->read;
4605 event = rb_reader_event(cpu_buffer);
4606 /* Always keep the time extend and data together */
4607 size = rb_event_ts_length(event);
4608 } while (len >= size);
4611 local_set(&bpage->commit, pos);
4612 bpage->time_stamp = save_timestamp;
4614 /* we copied everything to the beginning */
4617 /* update the entry counter */
4618 cpu_buffer->read += rb_page_entries(reader);
4619 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4621 /* swap the pages */
4622 rb_init_page(bpage);
4623 bpage = reader->page;
4624 reader->page = *data_page;
4625 local_set(&reader->write, 0);
4626 local_set(&reader->entries, 0);
4631 * Use the real_end for the data size,
4632 * This gives us a chance to store the lost events
4635 if (reader->real_end)
4636 local_set(&bpage->commit, reader->real_end);
4640 cpu_buffer->lost_events = 0;
4642 commit = local_read(&bpage->commit);
4644 * Set a flag in the commit field if we lost events
4646 if (missed_events) {
4647 /* If there is room at the end of the page to save the
4648 * missed events, then record it there.
4650 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4651 memcpy(&bpage->data[commit], &missed_events,
4652 sizeof(missed_events));
4653 local_add(RB_MISSED_STORED, &bpage->commit);
4654 commit += sizeof(missed_events);
4656 local_add(RB_MISSED_EVENTS, &bpage->commit);
4660 * This page may be off to user land. Zero it out here.
4662 if (commit < BUF_PAGE_SIZE)
4663 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4666 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4671 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4673 #ifdef CONFIG_HOTPLUG_CPU
4674 static int rb_cpu_notify(struct notifier_block *self,
4675 unsigned long action, void *hcpu)
4677 struct ring_buffer *buffer =
4678 container_of(self, struct ring_buffer, cpu_notify);
4679 long cpu = (long)hcpu;
4682 unsigned long nr_pages;
4685 case CPU_UP_PREPARE:
4686 case CPU_UP_PREPARE_FROZEN:
4687 if (cpumask_test_cpu(cpu, buffer->cpumask))
4692 /* check if all cpu sizes are same */
4693 for_each_buffer_cpu(buffer, cpu_i) {
4694 /* fill in the size from first enabled cpu */
4696 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4697 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4702 /* allocate minimum pages, user can later expand it */
4705 buffer->buffers[cpu] =
4706 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4707 if (!buffer->buffers[cpu]) {
4708 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4713 cpumask_set_cpu(cpu, buffer->cpumask);
4715 case CPU_DOWN_PREPARE:
4716 case CPU_DOWN_PREPARE_FROZEN:
4719 * If we were to free the buffer, then the user would
4720 * lose any trace that was in the buffer.
4730 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4732 * This is a basic integrity check of the ring buffer.
4733 * Late in the boot cycle this test will run when configured in.
4734 * It will kick off a thread per CPU that will go into a loop
4735 * writing to the per cpu ring buffer various sizes of data.
4736 * Some of the data will be large items, some small.
4738 * Another thread is created that goes into a spin, sending out
4739 * IPIs to the other CPUs to also write into the ring buffer.
4740 * this is to test the nesting ability of the buffer.
4742 * Basic stats are recorded and reported. If something in the
4743 * ring buffer should happen that's not expected, a big warning
4744 * is displayed and all ring buffers are disabled.
4746 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4748 struct rb_test_data {
4749 struct ring_buffer *buffer;
4750 unsigned long events;
4751 unsigned long bytes_written;
4752 unsigned long bytes_alloc;
4753 unsigned long bytes_dropped;
4754 unsigned long events_nested;
4755 unsigned long bytes_written_nested;
4756 unsigned long bytes_alloc_nested;
4757 unsigned long bytes_dropped_nested;
4758 int min_size_nested;
4759 int max_size_nested;
4766 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4769 #define RB_TEST_BUFFER_SIZE 1048576
4771 static char rb_string[] __initdata =
4772 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4773 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4774 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4776 static bool rb_test_started __initdata;
4783 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4785 struct ring_buffer_event *event;
4786 struct rb_item *item;
4793 /* Have nested writes different that what is written */
4794 cnt = data->cnt + (nested ? 27 : 0);
4796 /* Multiply cnt by ~e, to make some unique increment */
4797 size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4799 len = size + sizeof(struct rb_item);
4801 started = rb_test_started;
4802 /* read rb_test_started before checking buffer enabled */
4805 event = ring_buffer_lock_reserve(data->buffer, len);
4807 /* Ignore dropped events before test starts. */
4810 data->bytes_dropped += len;
4812 data->bytes_dropped_nested += len;
4817 event_len = ring_buffer_event_length(event);
4819 if (RB_WARN_ON(data->buffer, event_len < len))
4822 item = ring_buffer_event_data(event);
4824 memcpy(item->str, rb_string, size);
4827 data->bytes_alloc_nested += event_len;
4828 data->bytes_written_nested += len;
4829 data->events_nested++;
4830 if (!data->min_size_nested || len < data->min_size_nested)
4831 data->min_size_nested = len;
4832 if (len > data->max_size_nested)
4833 data->max_size_nested = len;
4835 data->bytes_alloc += event_len;
4836 data->bytes_written += len;
4838 if (!data->min_size || len < data->min_size)
4839 data->max_size = len;
4840 if (len > data->max_size)
4841 data->max_size = len;
4845 ring_buffer_unlock_commit(data->buffer, event);
4850 static __init int rb_test(void *arg)
4852 struct rb_test_data *data = arg;
4854 while (!kthread_should_stop()) {
4855 rb_write_something(data, false);
4858 set_current_state(TASK_INTERRUPTIBLE);
4859 /* Now sleep between a min of 100-300us and a max of 1ms */
4860 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4866 static __init void rb_ipi(void *ignore)
4868 struct rb_test_data *data;
4869 int cpu = smp_processor_id();
4871 data = &rb_data[cpu];
4872 rb_write_something(data, true);
4875 static __init int rb_hammer_test(void *arg)
4877 while (!kthread_should_stop()) {
4879 /* Send an IPI to all cpus to write data! */
4880 smp_call_function(rb_ipi, NULL, 1);
4881 /* No sleep, but for non preempt, let others run */
4888 static __init int test_ringbuffer(void)
4890 struct task_struct *rb_hammer;
4891 struct ring_buffer *buffer;
4895 pr_info("Running ring buffer tests...\n");
4897 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4898 if (WARN_ON(!buffer))
4901 /* Disable buffer so that threads can't write to it yet */
4902 ring_buffer_record_off(buffer);
4904 for_each_online_cpu(cpu) {
4905 rb_data[cpu].buffer = buffer;
4906 rb_data[cpu].cpu = cpu;
4907 rb_data[cpu].cnt = cpu;
4908 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4909 "rbtester/%d", cpu);
4910 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
4911 pr_cont("FAILED\n");
4912 ret = PTR_ERR(rb_threads[cpu]);
4916 kthread_bind(rb_threads[cpu], cpu);
4917 wake_up_process(rb_threads[cpu]);
4920 /* Now create the rb hammer! */
4921 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4922 if (WARN_ON(IS_ERR(rb_hammer))) {
4923 pr_cont("FAILED\n");
4924 ret = PTR_ERR(rb_hammer);
4928 ring_buffer_record_on(buffer);
4930 * Show buffer is enabled before setting rb_test_started.
4931 * Yes there's a small race window where events could be
4932 * dropped and the thread wont catch it. But when a ring
4933 * buffer gets enabled, there will always be some kind of
4934 * delay before other CPUs see it. Thus, we don't care about
4935 * those dropped events. We care about events dropped after
4936 * the threads see that the buffer is active.
4939 rb_test_started = true;
4941 set_current_state(TASK_INTERRUPTIBLE);
4942 /* Just run for 10 seconds */;
4943 schedule_timeout(10 * HZ);
4945 kthread_stop(rb_hammer);
4948 for_each_online_cpu(cpu) {
4949 if (!rb_threads[cpu])
4951 kthread_stop(rb_threads[cpu]);
4954 ring_buffer_free(buffer);
4959 pr_info("finished\n");
4960 for_each_online_cpu(cpu) {
4961 struct ring_buffer_event *event;
4962 struct rb_test_data *data = &rb_data[cpu];
4963 struct rb_item *item;
4964 unsigned long total_events;
4965 unsigned long total_dropped;
4966 unsigned long total_written;
4967 unsigned long total_alloc;
4968 unsigned long total_read = 0;
4969 unsigned long total_size = 0;
4970 unsigned long total_len = 0;
4971 unsigned long total_lost = 0;
4974 int small_event_size;
4978 total_events = data->events + data->events_nested;
4979 total_written = data->bytes_written + data->bytes_written_nested;
4980 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
4981 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
4983 big_event_size = data->max_size + data->max_size_nested;
4984 small_event_size = data->min_size + data->min_size_nested;
4986 pr_info("CPU %d:\n", cpu);
4987 pr_info(" events: %ld\n", total_events);
4988 pr_info(" dropped bytes: %ld\n", total_dropped);
4989 pr_info(" alloced bytes: %ld\n", total_alloc);
4990 pr_info(" written bytes: %ld\n", total_written);
4991 pr_info(" biggest event: %d\n", big_event_size);
4992 pr_info(" smallest event: %d\n", small_event_size);
4994 if (RB_WARN_ON(buffer, total_dropped))
4999 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
5001 item = ring_buffer_event_data(event);
5002 total_len += ring_buffer_event_length(event);
5003 total_size += item->size + sizeof(struct rb_item);
5004 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
5005 pr_info("FAILED!\n");
5006 pr_info("buffer had: %.*s\n", item->size, item->str);
5007 pr_info("expected: %.*s\n", item->size, rb_string);
5008 RB_WARN_ON(buffer, 1);
5019 pr_info(" read events: %ld\n", total_read);
5020 pr_info(" lost events: %ld\n", total_lost);
5021 pr_info(" total events: %ld\n", total_lost + total_read);
5022 pr_info(" recorded len bytes: %ld\n", total_len);
5023 pr_info(" recorded size bytes: %ld\n", total_size);
5025 pr_info(" With dropped events, record len and size may not match\n"
5026 " alloced and written from above\n");
5028 if (RB_WARN_ON(buffer, total_len != total_alloc ||
5029 total_size != total_written))
5032 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5038 pr_info("Ring buffer PASSED!\n");
5040 ring_buffer_free(buffer);
5044 late_initcall(test_ringbuffer);
5045 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */