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/sched/clock.h>
10 #include <linux/trace_seq.h>
11 #include <linux/spinlock.h>
12 #include <linux/irq_work.h>
13 #include <linux/uaccess.h>
14 #include <linux/hardirq.h>
15 #include <linux/kthread.h> /* for self test */
16 #include <linux/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 */
248 static __always_inline void *
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.
425 * See trace_recursive_lock() comment below for more details.
437 * head_page == tail_page && head == tail then buffer is empty.
439 struct ring_buffer_per_cpu {
441 atomic_t record_disabled;
442 struct ring_buffer *buffer;
443 raw_spinlock_t reader_lock; /* serialize readers */
444 arch_spinlock_t lock;
445 struct lock_class_key lock_key;
446 struct buffer_data_page *free_page;
447 unsigned long nr_pages;
448 unsigned int current_context;
449 struct list_head *pages;
450 struct buffer_page *head_page; /* read from head */
451 struct buffer_page *tail_page; /* write to tail */
452 struct buffer_page *commit_page; /* committed pages */
453 struct buffer_page *reader_page;
454 unsigned long lost_events;
455 unsigned long last_overrun;
456 local_t entries_bytes;
459 local_t commit_overrun;
460 local_t dropped_events;
464 unsigned long read_bytes;
467 /* pages removed since last reset */
468 unsigned long pages_removed;
469 /* ring buffer pages to update, > 0 to add, < 0 to remove */
470 long nr_pages_to_update;
471 struct list_head new_pages; /* new pages to add */
472 struct work_struct update_pages_work;
473 struct completion update_done;
475 struct rb_irq_work irq_work;
481 atomic_t record_disabled;
482 atomic_t resize_disabled;
483 cpumask_var_t cpumask;
485 struct lock_class_key *reader_lock_key;
489 struct ring_buffer_per_cpu **buffers;
491 struct hlist_node node;
494 struct rb_irq_work irq_work;
497 struct ring_buffer_iter {
498 struct ring_buffer_per_cpu *cpu_buffer;
500 struct buffer_page *head_page;
501 struct buffer_page *cache_reader_page;
502 unsigned long cache_read;
503 unsigned long cache_pages_removed;
508 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
510 * Schedules a delayed work to wake up any task that is blocked on the
511 * ring buffer waiters queue.
513 static void rb_wake_up_waiters(struct irq_work *work)
515 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
517 wake_up_all(&rbwork->waiters);
518 if (rbwork->full_waiters_pending || rbwork->wakeup_full) {
519 rbwork->wakeup_full = false;
520 rbwork->full_waiters_pending = false;
521 wake_up_all(&rbwork->full_waiters);
526 * ring_buffer_wait - wait for input to the ring buffer
527 * @buffer: buffer to wait on
528 * @cpu: the cpu buffer to wait on
529 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
531 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
532 * as data is added to any of the @buffer's cpu buffers. Otherwise
533 * it will wait for data to be added to a specific cpu buffer.
535 int ring_buffer_wait(struct ring_buffer *buffer, int cpu, bool full)
537 struct ring_buffer_per_cpu *cpu_buffer;
539 struct rb_irq_work *work;
543 * Depending on what the caller is waiting for, either any
544 * data in any cpu buffer, or a specific buffer, put the
545 * caller on the appropriate wait queue.
547 if (cpu == RING_BUFFER_ALL_CPUS) {
548 work = &buffer->irq_work;
549 /* Full only makes sense on per cpu reads */
552 if (!cpumask_test_cpu(cpu, buffer->cpumask))
554 cpu_buffer = buffer->buffers[cpu];
555 work = &cpu_buffer->irq_work;
561 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
563 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
566 * The events can happen in critical sections where
567 * checking a work queue can cause deadlocks.
568 * After adding a task to the queue, this flag is set
569 * only to notify events to try to wake up the queue
572 * We don't clear it even if the buffer is no longer
573 * empty. The flag only causes the next event to run
574 * irq_work to do the work queue wake up. The worse
575 * that can happen if we race with !trace_empty() is that
576 * an event will cause an irq_work to try to wake up
579 * There's no reason to protect this flag either, as
580 * the work queue and irq_work logic will do the necessary
581 * synchronization for the wake ups. The only thing
582 * that is necessary is that the wake up happens after
583 * a task has been queued. It's OK for spurious wake ups.
586 work->full_waiters_pending = true;
588 work->waiters_pending = true;
590 if (signal_pending(current)) {
595 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
598 if (cpu != RING_BUFFER_ALL_CPUS &&
599 !ring_buffer_empty_cpu(buffer, cpu)) {
606 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
607 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
608 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
618 finish_wait(&work->full_waiters, &wait);
620 finish_wait(&work->waiters, &wait);
626 * ring_buffer_poll_wait - poll on buffer input
627 * @buffer: buffer to wait on
628 * @cpu: the cpu buffer to wait on
629 * @filp: the file descriptor
630 * @poll_table: The poll descriptor
632 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
633 * as data is added to any of the @buffer's cpu buffers. Otherwise
634 * it will wait for data to be added to a specific cpu buffer.
636 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
639 int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
640 struct file *filp, poll_table *poll_table)
642 struct ring_buffer_per_cpu *cpu_buffer;
643 struct rb_irq_work *work;
645 if (cpu == RING_BUFFER_ALL_CPUS)
646 work = &buffer->irq_work;
648 if (!cpumask_test_cpu(cpu, buffer->cpumask))
651 cpu_buffer = buffer->buffers[cpu];
652 work = &cpu_buffer->irq_work;
655 poll_wait(filp, &work->waiters, poll_table);
656 work->waiters_pending = true;
658 * There's a tight race between setting the waiters_pending and
659 * checking if the ring buffer is empty. Once the waiters_pending bit
660 * is set, the next event will wake the task up, but we can get stuck
661 * if there's only a single event in.
663 * FIXME: Ideally, we need a memory barrier on the writer side as well,
664 * but adding a memory barrier to all events will cause too much of a
665 * performance hit in the fast path. We only need a memory barrier when
666 * the buffer goes from empty to having content. But as this race is
667 * extremely small, and it's not a problem if another event comes in, we
672 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
673 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
674 return POLLIN | POLLRDNORM;
678 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
679 #define RB_WARN_ON(b, cond) \
681 int _____ret = unlikely(cond); \
683 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
684 struct ring_buffer_per_cpu *__b = \
686 atomic_inc(&__b->buffer->record_disabled); \
688 atomic_inc(&b->record_disabled); \
694 /* Up this if you want to test the TIME_EXTENTS and normalization */
695 #define DEBUG_SHIFT 0
697 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
699 /* shift to debug/test normalization and TIME_EXTENTS */
700 return buffer->clock() << DEBUG_SHIFT;
703 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
707 preempt_disable_notrace();
708 time = rb_time_stamp(buffer);
709 preempt_enable_notrace();
713 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
715 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
718 /* Just stupid testing the normalize function and deltas */
721 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
724 * Making the ring buffer lockless makes things tricky.
725 * Although writes only happen on the CPU that they are on,
726 * and they only need to worry about interrupts. Reads can
729 * The reader page is always off the ring buffer, but when the
730 * reader finishes with a page, it needs to swap its page with
731 * a new one from the buffer. The reader needs to take from
732 * the head (writes go to the tail). But if a writer is in overwrite
733 * mode and wraps, it must push the head page forward.
735 * Here lies the problem.
737 * The reader must be careful to replace only the head page, and
738 * not another one. As described at the top of the file in the
739 * ASCII art, the reader sets its old page to point to the next
740 * page after head. It then sets the page after head to point to
741 * the old reader page. But if the writer moves the head page
742 * during this operation, the reader could end up with the tail.
744 * We use cmpxchg to help prevent this race. We also do something
745 * special with the page before head. We set the LSB to 1.
747 * When the writer must push the page forward, it will clear the
748 * bit that points to the head page, move the head, and then set
749 * the bit that points to the new head page.
751 * We also don't want an interrupt coming in and moving the head
752 * page on another writer. Thus we use the second LSB to catch
755 * head->list->prev->next bit 1 bit 0
758 * Points to head page 0 1
761 * Note we can not trust the prev pointer of the head page, because:
763 * +----+ +-----+ +-----+
764 * | |------>| T |---X--->| N |
766 * +----+ +-----+ +-----+
769 * +----------| R |----------+ |
773 * Key: ---X--> HEAD flag set in pointer
778 * (see __rb_reserve_next() to see where this happens)
780 * What the above shows is that the reader just swapped out
781 * the reader page with a page in the buffer, but before it
782 * could make the new header point back to the new page added
783 * it was preempted by a writer. The writer moved forward onto
784 * the new page added by the reader and is about to move forward
787 * You can see, it is legitimate for the previous pointer of
788 * the head (or any page) not to point back to itself. But only
792 #define RB_PAGE_NORMAL 0UL
793 #define RB_PAGE_HEAD 1UL
794 #define RB_PAGE_UPDATE 2UL
797 #define RB_FLAG_MASK 3UL
799 /* PAGE_MOVED is not part of the mask */
800 #define RB_PAGE_MOVED 4UL
803 * rb_list_head - remove any bit
805 static struct list_head *rb_list_head(struct list_head *list)
807 unsigned long val = (unsigned long)list;
809 return (struct list_head *)(val & ~RB_FLAG_MASK);
813 * rb_is_head_page - test if the given page is the head page
815 * Because the reader may move the head_page pointer, we can
816 * not trust what the head page is (it may be pointing to
817 * the reader page). But if the next page is a header page,
818 * its flags will be non zero.
821 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
822 struct buffer_page *page, struct list_head *list)
826 val = (unsigned long)list->next;
828 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
829 return RB_PAGE_MOVED;
831 return val & RB_FLAG_MASK;
837 * The unique thing about the reader page, is that, if the
838 * writer is ever on it, the previous pointer never points
839 * back to the reader page.
841 static bool rb_is_reader_page(struct buffer_page *page)
843 struct list_head *list = page->list.prev;
845 return rb_list_head(list->next) != &page->list;
849 * rb_set_list_to_head - set a list_head to be pointing to head.
851 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
852 struct list_head *list)
856 ptr = (unsigned long *)&list->next;
857 *ptr |= RB_PAGE_HEAD;
858 *ptr &= ~RB_PAGE_UPDATE;
862 * rb_head_page_activate - sets up head page
864 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
866 struct buffer_page *head;
868 head = cpu_buffer->head_page;
873 * Set the previous list pointer to have the HEAD flag.
875 rb_set_list_to_head(cpu_buffer, head->list.prev);
878 static void rb_list_head_clear(struct list_head *list)
880 unsigned long *ptr = (unsigned long *)&list->next;
882 *ptr &= ~RB_FLAG_MASK;
886 * rb_head_page_dactivate - clears head page ptr (for free list)
889 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
891 struct list_head *hd;
893 /* Go through the whole list and clear any pointers found. */
894 rb_list_head_clear(cpu_buffer->pages);
896 list_for_each(hd, cpu_buffer->pages)
897 rb_list_head_clear(hd);
900 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
901 struct buffer_page *head,
902 struct buffer_page *prev,
903 int old_flag, int new_flag)
905 struct list_head *list;
906 unsigned long val = (unsigned long)&head->list;
911 val &= ~RB_FLAG_MASK;
913 ret = cmpxchg((unsigned long *)&list->next,
914 val | old_flag, val | new_flag);
916 /* check if the reader took the page */
917 if ((ret & ~RB_FLAG_MASK) != val)
918 return RB_PAGE_MOVED;
920 return ret & RB_FLAG_MASK;
923 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
924 struct buffer_page *head,
925 struct buffer_page *prev,
928 return rb_head_page_set(cpu_buffer, head, prev,
929 old_flag, RB_PAGE_UPDATE);
932 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
933 struct buffer_page *head,
934 struct buffer_page *prev,
937 return rb_head_page_set(cpu_buffer, head, prev,
938 old_flag, RB_PAGE_HEAD);
941 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
942 struct buffer_page *head,
943 struct buffer_page *prev,
946 return rb_head_page_set(cpu_buffer, head, prev,
947 old_flag, RB_PAGE_NORMAL);
950 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
951 struct buffer_page **bpage)
953 struct list_head *p = rb_list_head((*bpage)->list.next);
955 *bpage = list_entry(p, struct buffer_page, list);
958 static struct buffer_page *
959 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
961 struct buffer_page *head;
962 struct buffer_page *page;
963 struct list_head *list;
966 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
970 list = cpu_buffer->pages;
971 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
974 page = head = cpu_buffer->head_page;
976 * It is possible that the writer moves the header behind
977 * where we started, and we miss in one loop.
978 * A second loop should grab the header, but we'll do
979 * three loops just because I'm paranoid.
981 for (i = 0; i < 3; i++) {
983 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
984 cpu_buffer->head_page = page;
987 rb_inc_page(cpu_buffer, &page);
988 } while (page != head);
991 RB_WARN_ON(cpu_buffer, 1);
996 static int rb_head_page_replace(struct buffer_page *old,
997 struct buffer_page *new)
999 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1003 val = *ptr & ~RB_FLAG_MASK;
1004 val |= RB_PAGE_HEAD;
1006 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1012 * rb_tail_page_update - move the tail page forward
1014 static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1015 struct buffer_page *tail_page,
1016 struct buffer_page *next_page)
1018 unsigned long old_entries;
1019 unsigned long old_write;
1022 * The tail page now needs to be moved forward.
1024 * We need to reset the tail page, but without messing
1025 * with possible erasing of data brought in by interrupts
1026 * that have moved the tail page and are currently on it.
1028 * We add a counter to the write field to denote this.
1030 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1031 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1034 * Just make sure we have seen our old_write and synchronize
1035 * with any interrupts that come in.
1040 * If the tail page is still the same as what we think
1041 * it is, then it is up to us to update the tail
1044 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1045 /* Zero the write counter */
1046 unsigned long val = old_write & ~RB_WRITE_MASK;
1047 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1050 * This will only succeed if an interrupt did
1051 * not come in and change it. In which case, we
1052 * do not want to modify it.
1054 * We add (void) to let the compiler know that we do not care
1055 * about the return value of these functions. We use the
1056 * cmpxchg to only update if an interrupt did not already
1057 * do it for us. If the cmpxchg fails, we don't care.
1059 (void)local_cmpxchg(&next_page->write, old_write, val);
1060 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1063 * No need to worry about races with clearing out the commit.
1064 * it only can increment when a commit takes place. But that
1065 * only happens in the outer most nested commit.
1067 local_set(&next_page->page->commit, 0);
1069 /* Again, either we update tail_page or an interrupt does */
1070 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1074 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1075 struct buffer_page *bpage)
1077 unsigned long val = (unsigned long)bpage;
1079 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1086 * rb_check_list - make sure a pointer to a list has the last bits zero
1088 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1089 struct list_head *list)
1091 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1093 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1099 * rb_check_pages - integrity check of buffer pages
1100 * @cpu_buffer: CPU buffer with pages to test
1102 * As a safety measure we check to make sure the data pages have not
1105 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1107 struct list_head *head = cpu_buffer->pages;
1108 struct buffer_page *bpage, *tmp;
1110 /* Reset the head page if it exists */
1111 if (cpu_buffer->head_page)
1112 rb_set_head_page(cpu_buffer);
1114 rb_head_page_deactivate(cpu_buffer);
1116 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1118 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1121 if (rb_check_list(cpu_buffer, head))
1124 list_for_each_entry_safe(bpage, tmp, head, list) {
1125 if (RB_WARN_ON(cpu_buffer,
1126 bpage->list.next->prev != &bpage->list))
1128 if (RB_WARN_ON(cpu_buffer,
1129 bpage->list.prev->next != &bpage->list))
1131 if (rb_check_list(cpu_buffer, &bpage->list))
1135 rb_head_page_activate(cpu_buffer);
1140 static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1142 struct buffer_page *bpage, *tmp;
1145 /* Check if the available memory is there first */
1146 i = si_mem_available();
1150 for (i = 0; i < nr_pages; i++) {
1153 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1154 * gracefully without invoking oom-killer and the system is not
1157 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1158 GFP_KERNEL | __GFP_RETRY_MAYFAIL,
1163 list_add(&bpage->list, pages);
1165 page = alloc_pages_node(cpu_to_node(cpu),
1166 GFP_KERNEL | __GFP_RETRY_MAYFAIL, 0);
1169 bpage->page = page_address(page);
1170 rb_init_page(bpage->page);
1176 list_for_each_entry_safe(bpage, tmp, pages, list) {
1177 list_del_init(&bpage->list);
1178 free_buffer_page(bpage);
1184 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1185 unsigned long nr_pages)
1191 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1195 * The ring buffer page list is a circular list that does not
1196 * start and end with a list head. All page list items point to
1199 cpu_buffer->pages = pages.next;
1202 cpu_buffer->nr_pages = nr_pages;
1204 rb_check_pages(cpu_buffer);
1209 static struct ring_buffer_per_cpu *
1210 rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1212 struct ring_buffer_per_cpu *cpu_buffer;
1213 struct buffer_page *bpage;
1217 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1218 GFP_KERNEL, cpu_to_node(cpu));
1222 cpu_buffer->cpu = cpu;
1223 cpu_buffer->buffer = buffer;
1224 raw_spin_lock_init(&cpu_buffer->reader_lock);
1225 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1226 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1227 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1228 init_completion(&cpu_buffer->update_done);
1229 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1230 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1231 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1233 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1234 GFP_KERNEL, cpu_to_node(cpu));
1236 goto fail_free_buffer;
1238 rb_check_bpage(cpu_buffer, bpage);
1240 cpu_buffer->reader_page = bpage;
1241 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1243 goto fail_free_reader;
1244 bpage->page = page_address(page);
1245 rb_init_page(bpage->page);
1247 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1248 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1250 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1252 goto fail_free_reader;
1254 cpu_buffer->head_page
1255 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1256 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1258 rb_head_page_activate(cpu_buffer);
1263 free_buffer_page(cpu_buffer->reader_page);
1270 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1272 struct list_head *head = cpu_buffer->pages;
1273 struct buffer_page *bpage, *tmp;
1275 irq_work_sync(&cpu_buffer->irq_work.work);
1277 free_buffer_page(cpu_buffer->reader_page);
1280 rb_head_page_deactivate(cpu_buffer);
1282 list_for_each_entry_safe(bpage, tmp, head, list) {
1283 list_del_init(&bpage->list);
1284 free_buffer_page(bpage);
1286 bpage = list_entry(head, struct buffer_page, list);
1287 free_buffer_page(bpage);
1294 * __ring_buffer_alloc - allocate a new ring_buffer
1295 * @size: the size in bytes per cpu that is needed.
1296 * @flags: attributes to set for the ring buffer.
1298 * Currently the only flag that is available is the RB_FL_OVERWRITE
1299 * flag. This flag means that the buffer will overwrite old data
1300 * when the buffer wraps. If this flag is not set, the buffer will
1301 * drop data when the tail hits the head.
1303 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1304 struct lock_class_key *key)
1306 struct ring_buffer *buffer;
1312 /* keep it in its own cache line */
1313 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1318 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1319 goto fail_free_buffer;
1321 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1322 buffer->flags = flags;
1323 buffer->clock = trace_clock_local;
1324 buffer->reader_lock_key = key;
1326 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1327 init_waitqueue_head(&buffer->irq_work.waiters);
1329 /* need at least two pages */
1333 buffer->cpus = nr_cpu_ids;
1335 bsize = sizeof(void *) * nr_cpu_ids;
1336 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1338 if (!buffer->buffers)
1339 goto fail_free_cpumask;
1341 cpu = raw_smp_processor_id();
1342 cpumask_set_cpu(cpu, buffer->cpumask);
1343 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1344 if (!buffer->buffers[cpu])
1345 goto fail_free_buffers;
1347 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1349 goto fail_free_buffers;
1351 mutex_init(&buffer->mutex);
1356 for_each_buffer_cpu(buffer, cpu) {
1357 if (buffer->buffers[cpu])
1358 rb_free_cpu_buffer(buffer->buffers[cpu]);
1360 kfree(buffer->buffers);
1363 free_cpumask_var(buffer->cpumask);
1369 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1372 * ring_buffer_free - free a ring buffer.
1373 * @buffer: the buffer to free.
1376 ring_buffer_free(struct ring_buffer *buffer)
1380 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1382 irq_work_sync(&buffer->irq_work.work);
1384 for_each_buffer_cpu(buffer, cpu)
1385 rb_free_cpu_buffer(buffer->buffers[cpu]);
1387 kfree(buffer->buffers);
1388 free_cpumask_var(buffer->cpumask);
1392 EXPORT_SYMBOL_GPL(ring_buffer_free);
1394 void ring_buffer_set_clock(struct ring_buffer *buffer,
1397 buffer->clock = clock;
1400 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1402 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1404 return local_read(&bpage->entries) & RB_WRITE_MASK;
1407 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1409 return local_read(&bpage->write) & RB_WRITE_MASK;
1413 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1415 struct list_head *tail_page, *to_remove, *next_page;
1416 struct buffer_page *to_remove_page, *tmp_iter_page;
1417 struct buffer_page *last_page, *first_page;
1418 unsigned long nr_removed;
1419 unsigned long head_bit;
1424 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1425 atomic_inc(&cpu_buffer->record_disabled);
1427 * We don't race with the readers since we have acquired the reader
1428 * lock. We also don't race with writers after disabling recording.
1429 * This makes it easy to figure out the first and the last page to be
1430 * removed from the list. We unlink all the pages in between including
1431 * the first and last pages. This is done in a busy loop so that we
1432 * lose the least number of traces.
1433 * The pages are freed after we restart recording and unlock readers.
1435 tail_page = &cpu_buffer->tail_page->list;
1438 * tail page might be on reader page, we remove the next page
1439 * from the ring buffer
1441 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1442 tail_page = rb_list_head(tail_page->next);
1443 to_remove = tail_page;
1445 /* start of pages to remove */
1446 first_page = list_entry(rb_list_head(to_remove->next),
1447 struct buffer_page, list);
1449 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1450 to_remove = rb_list_head(to_remove)->next;
1451 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1453 /* Read iterators need to reset themselves when some pages removed */
1454 cpu_buffer->pages_removed += nr_removed;
1456 next_page = rb_list_head(to_remove)->next;
1459 * Now we remove all pages between tail_page and next_page.
1460 * Make sure that we have head_bit value preserved for the
1463 tail_page->next = (struct list_head *)((unsigned long)next_page |
1465 next_page = rb_list_head(next_page);
1466 next_page->prev = tail_page;
1468 /* make sure pages points to a valid page in the ring buffer */
1469 cpu_buffer->pages = next_page;
1471 /* update head page */
1473 cpu_buffer->head_page = list_entry(next_page,
1474 struct buffer_page, list);
1476 /* pages are removed, resume tracing and then free the pages */
1477 atomic_dec(&cpu_buffer->record_disabled);
1478 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1480 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1482 /* last buffer page to remove */
1483 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1485 tmp_iter_page = first_page;
1490 to_remove_page = tmp_iter_page;
1491 rb_inc_page(cpu_buffer, &tmp_iter_page);
1493 /* update the counters */
1494 page_entries = rb_page_entries(to_remove_page);
1497 * If something was added to this page, it was full
1498 * since it is not the tail page. So we deduct the
1499 * bytes consumed in ring buffer from here.
1500 * Increment overrun to account for the lost events.
1502 local_add(page_entries, &cpu_buffer->overrun);
1503 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1507 * We have already removed references to this list item, just
1508 * free up the buffer_page and its page
1510 free_buffer_page(to_remove_page);
1513 } while (to_remove_page != last_page);
1515 RB_WARN_ON(cpu_buffer, nr_removed);
1517 return nr_removed == 0;
1521 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1523 struct list_head *pages = &cpu_buffer->new_pages;
1524 int retries, success;
1526 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1528 * We are holding the reader lock, so the reader page won't be swapped
1529 * in the ring buffer. Now we are racing with the writer trying to
1530 * move head page and the tail page.
1531 * We are going to adapt the reader page update process where:
1532 * 1. We first splice the start and end of list of new pages between
1533 * the head page and its previous page.
1534 * 2. We cmpxchg the prev_page->next to point from head page to the
1535 * start of new pages list.
1536 * 3. Finally, we update the head->prev to the end of new list.
1538 * We will try this process 10 times, to make sure that we don't keep
1544 struct list_head *head_page, *prev_page, *r;
1545 struct list_head *last_page, *first_page;
1546 struct list_head *head_page_with_bit;
1548 head_page = &rb_set_head_page(cpu_buffer)->list;
1551 prev_page = head_page->prev;
1553 first_page = pages->next;
1554 last_page = pages->prev;
1556 head_page_with_bit = (struct list_head *)
1557 ((unsigned long)head_page | RB_PAGE_HEAD);
1559 last_page->next = head_page_with_bit;
1560 first_page->prev = prev_page;
1562 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1564 if (r == head_page_with_bit) {
1566 * yay, we replaced the page pointer to our new list,
1567 * now, we just have to update to head page's prev
1568 * pointer to point to end of list
1570 head_page->prev = last_page;
1577 INIT_LIST_HEAD(pages);
1579 * If we weren't successful in adding in new pages, warn and stop
1582 RB_WARN_ON(cpu_buffer, !success);
1583 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1585 /* free pages if they weren't inserted */
1587 struct buffer_page *bpage, *tmp;
1588 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1590 list_del_init(&bpage->list);
1591 free_buffer_page(bpage);
1597 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1601 if (cpu_buffer->nr_pages_to_update > 0)
1602 success = rb_insert_pages(cpu_buffer);
1604 success = rb_remove_pages(cpu_buffer,
1605 -cpu_buffer->nr_pages_to_update);
1608 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1611 static void update_pages_handler(struct work_struct *work)
1613 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1614 struct ring_buffer_per_cpu, update_pages_work);
1615 rb_update_pages(cpu_buffer);
1616 complete(&cpu_buffer->update_done);
1620 * ring_buffer_resize - resize the ring buffer
1621 * @buffer: the buffer to resize.
1622 * @size: the new size.
1623 * @cpu_id: the cpu buffer to resize
1625 * Minimum size is 2 * BUF_PAGE_SIZE.
1627 * Returns 0 on success and < 0 on failure.
1629 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1632 struct ring_buffer_per_cpu *cpu_buffer;
1633 unsigned long nr_pages;
1637 * Always succeed at resizing a non-existent buffer:
1642 /* Make sure the requested buffer exists */
1643 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1644 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1647 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1649 /* we need a minimum of two pages */
1653 size = nr_pages * BUF_PAGE_SIZE;
1656 * Don't succeed if resizing is disabled, as a reader might be
1657 * manipulating the ring buffer and is expecting a sane state while
1660 if (atomic_read(&buffer->resize_disabled))
1663 /* prevent another thread from changing buffer sizes */
1664 mutex_lock(&buffer->mutex);
1666 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1667 /* calculate the pages to update */
1668 for_each_buffer_cpu(buffer, cpu) {
1669 cpu_buffer = buffer->buffers[cpu];
1671 cpu_buffer->nr_pages_to_update = nr_pages -
1672 cpu_buffer->nr_pages;
1674 * nothing more to do for removing pages or no update
1676 if (cpu_buffer->nr_pages_to_update <= 0)
1679 * to add pages, make sure all new pages can be
1680 * allocated without receiving ENOMEM
1682 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1683 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1684 &cpu_buffer->new_pages, cpu)) {
1685 /* not enough memory for new pages */
1695 * Fire off all the required work handlers
1696 * We can't schedule on offline CPUs, but it's not necessary
1697 * since we can change their buffer sizes without any race.
1699 for_each_buffer_cpu(buffer, cpu) {
1700 cpu_buffer = buffer->buffers[cpu];
1701 if (!cpu_buffer->nr_pages_to_update)
1704 /* Can't run something on an offline CPU. */
1705 if (!cpu_online(cpu)) {
1706 rb_update_pages(cpu_buffer);
1707 cpu_buffer->nr_pages_to_update = 0;
1709 schedule_work_on(cpu,
1710 &cpu_buffer->update_pages_work);
1714 /* wait for all the updates to complete */
1715 for_each_buffer_cpu(buffer, cpu) {
1716 cpu_buffer = buffer->buffers[cpu];
1717 if (!cpu_buffer->nr_pages_to_update)
1720 if (cpu_online(cpu))
1721 wait_for_completion(&cpu_buffer->update_done);
1722 cpu_buffer->nr_pages_to_update = 0;
1727 /* Make sure this CPU has been intitialized */
1728 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1731 cpu_buffer = buffer->buffers[cpu_id];
1733 if (nr_pages == cpu_buffer->nr_pages)
1736 cpu_buffer->nr_pages_to_update = nr_pages -
1737 cpu_buffer->nr_pages;
1739 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1740 if (cpu_buffer->nr_pages_to_update > 0 &&
1741 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1742 &cpu_buffer->new_pages, cpu_id)) {
1749 /* Can't run something on an offline CPU. */
1750 if (!cpu_online(cpu_id))
1751 rb_update_pages(cpu_buffer);
1753 schedule_work_on(cpu_id,
1754 &cpu_buffer->update_pages_work);
1755 wait_for_completion(&cpu_buffer->update_done);
1758 cpu_buffer->nr_pages_to_update = 0;
1764 * The ring buffer resize can happen with the ring buffer
1765 * enabled, so that the update disturbs the tracing as little
1766 * as possible. But if the buffer is disabled, we do not need
1767 * to worry about that, and we can take the time to verify
1768 * that the buffer is not corrupt.
1770 if (atomic_read(&buffer->record_disabled)) {
1771 atomic_inc(&buffer->record_disabled);
1773 * Even though the buffer was disabled, we must make sure
1774 * that it is truly disabled before calling rb_check_pages.
1775 * There could have been a race between checking
1776 * record_disable and incrementing it.
1778 synchronize_sched();
1779 for_each_buffer_cpu(buffer, cpu) {
1780 cpu_buffer = buffer->buffers[cpu];
1781 rb_check_pages(cpu_buffer);
1783 atomic_dec(&buffer->record_disabled);
1786 mutex_unlock(&buffer->mutex);
1790 for_each_buffer_cpu(buffer, cpu) {
1791 struct buffer_page *bpage, *tmp;
1793 cpu_buffer = buffer->buffers[cpu];
1794 cpu_buffer->nr_pages_to_update = 0;
1796 if (list_empty(&cpu_buffer->new_pages))
1799 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1801 list_del_init(&bpage->list);
1802 free_buffer_page(bpage);
1805 mutex_unlock(&buffer->mutex);
1808 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1810 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1812 mutex_lock(&buffer->mutex);
1814 buffer->flags |= RB_FL_OVERWRITE;
1816 buffer->flags &= ~RB_FL_OVERWRITE;
1817 mutex_unlock(&buffer->mutex);
1819 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1821 static __always_inline void *
1822 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1824 return bpage->data + index;
1827 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1829 return bpage->page->data + index;
1832 static __always_inline struct ring_buffer_event *
1833 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1835 return __rb_page_index(cpu_buffer->reader_page,
1836 cpu_buffer->reader_page->read);
1839 static __always_inline struct ring_buffer_event *
1840 rb_iter_head_event(struct ring_buffer_iter *iter)
1842 return __rb_page_index(iter->head_page, iter->head);
1845 static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
1847 return local_read(&bpage->page->commit);
1850 /* Size is determined by what has been committed */
1851 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
1853 return rb_page_commit(bpage);
1856 static __always_inline unsigned
1857 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1859 return rb_page_commit(cpu_buffer->commit_page);
1862 static __always_inline unsigned
1863 rb_event_index(struct ring_buffer_event *event)
1865 unsigned long addr = (unsigned long)event;
1867 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1870 static void rb_inc_iter(struct ring_buffer_iter *iter)
1872 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1875 * The iterator could be on the reader page (it starts there).
1876 * But the head could have moved, since the reader was
1877 * found. Check for this case and assign the iterator
1878 * to the head page instead of next.
1880 if (iter->head_page == cpu_buffer->reader_page)
1881 iter->head_page = rb_set_head_page(cpu_buffer);
1883 rb_inc_page(cpu_buffer, &iter->head_page);
1885 iter->read_stamp = iter->head_page->page->time_stamp;
1890 * rb_handle_head_page - writer hit the head page
1892 * Returns: +1 to retry page
1897 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1898 struct buffer_page *tail_page,
1899 struct buffer_page *next_page)
1901 struct buffer_page *new_head;
1906 entries = rb_page_entries(next_page);
1909 * The hard part is here. We need to move the head
1910 * forward, and protect against both readers on
1911 * other CPUs and writers coming in via interrupts.
1913 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1917 * type can be one of four:
1918 * NORMAL - an interrupt already moved it for us
1919 * HEAD - we are the first to get here.
1920 * UPDATE - we are the interrupt interrupting
1922 * MOVED - a reader on another CPU moved the next
1923 * pointer to its reader page. Give up
1930 * We changed the head to UPDATE, thus
1931 * it is our responsibility to update
1934 local_add(entries, &cpu_buffer->overrun);
1935 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1938 * The entries will be zeroed out when we move the
1942 /* still more to do */
1945 case RB_PAGE_UPDATE:
1947 * This is an interrupt that interrupt the
1948 * previous update. Still more to do.
1951 case RB_PAGE_NORMAL:
1953 * An interrupt came in before the update
1954 * and processed this for us.
1955 * Nothing left to do.
1960 * The reader is on another CPU and just did
1961 * a swap with our next_page.
1966 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1971 * Now that we are here, the old head pointer is
1972 * set to UPDATE. This will keep the reader from
1973 * swapping the head page with the reader page.
1974 * The reader (on another CPU) will spin till
1977 * We just need to protect against interrupts
1978 * doing the job. We will set the next pointer
1979 * to HEAD. After that, we set the old pointer
1980 * to NORMAL, but only if it was HEAD before.
1981 * otherwise we are an interrupt, and only
1982 * want the outer most commit to reset it.
1984 new_head = next_page;
1985 rb_inc_page(cpu_buffer, &new_head);
1987 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1991 * Valid returns are:
1992 * HEAD - an interrupt came in and already set it.
1993 * NORMAL - One of two things:
1994 * 1) We really set it.
1995 * 2) A bunch of interrupts came in and moved
1996 * the page forward again.
2000 case RB_PAGE_NORMAL:
2004 RB_WARN_ON(cpu_buffer, 1);
2009 * It is possible that an interrupt came in,
2010 * set the head up, then more interrupts came in
2011 * and moved it again. When we get back here,
2012 * the page would have been set to NORMAL but we
2013 * just set it back to HEAD.
2015 * How do you detect this? Well, if that happened
2016 * the tail page would have moved.
2018 if (ret == RB_PAGE_NORMAL) {
2019 struct buffer_page *buffer_tail_page;
2021 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2023 * If the tail had moved passed next, then we need
2024 * to reset the pointer.
2026 if (buffer_tail_page != tail_page &&
2027 buffer_tail_page != next_page)
2028 rb_head_page_set_normal(cpu_buffer, new_head,
2034 * If this was the outer most commit (the one that
2035 * changed the original pointer from HEAD to UPDATE),
2036 * then it is up to us to reset it to NORMAL.
2038 if (type == RB_PAGE_HEAD) {
2039 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2042 if (RB_WARN_ON(cpu_buffer,
2043 ret != RB_PAGE_UPDATE))
2051 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2052 unsigned long tail, struct rb_event_info *info)
2054 struct buffer_page *tail_page = info->tail_page;
2055 struct ring_buffer_event *event;
2056 unsigned long length = info->length;
2059 * Only the event that crossed the page boundary
2060 * must fill the old tail_page with padding.
2062 if (tail >= BUF_PAGE_SIZE) {
2064 * If the page was filled, then we still need
2065 * to update the real_end. Reset it to zero
2066 * and the reader will ignore it.
2068 if (tail == BUF_PAGE_SIZE)
2069 tail_page->real_end = 0;
2071 local_sub(length, &tail_page->write);
2075 event = __rb_page_index(tail_page, tail);
2077 /* account for padding bytes */
2078 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2081 * Save the original length to the meta data.
2082 * This will be used by the reader to add lost event
2085 tail_page->real_end = tail;
2088 * If this event is bigger than the minimum size, then
2089 * we need to be careful that we don't subtract the
2090 * write counter enough to allow another writer to slip
2092 * We put in a discarded commit instead, to make sure
2093 * that this space is not used again.
2095 * If we are less than the minimum size, we don't need to
2098 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2099 /* No room for any events */
2101 /* Mark the rest of the page with padding */
2102 rb_event_set_padding(event);
2104 /* Make sure the padding is visible before the write update */
2107 /* Set the write back to the previous setting */
2108 local_sub(length, &tail_page->write);
2112 /* Put in a discarded event */
2113 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2114 event->type_len = RINGBUF_TYPE_PADDING;
2115 /* time delta must be non zero */
2116 event->time_delta = 1;
2118 /* Make sure the padding is visible before the tail_page->write update */
2121 /* Set write to end of buffer */
2122 length = (tail + length) - BUF_PAGE_SIZE;
2123 local_sub(length, &tail_page->write);
2126 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2129 * This is the slow path, force gcc not to inline it.
2131 static noinline struct ring_buffer_event *
2132 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2133 unsigned long tail, struct rb_event_info *info)
2135 struct buffer_page *tail_page = info->tail_page;
2136 struct buffer_page *commit_page = cpu_buffer->commit_page;
2137 struct ring_buffer *buffer = cpu_buffer->buffer;
2138 struct buffer_page *next_page;
2141 next_page = tail_page;
2143 rb_inc_page(cpu_buffer, &next_page);
2146 * If for some reason, we had an interrupt storm that made
2147 * it all the way around the buffer, bail, and warn
2150 if (unlikely(next_page == commit_page)) {
2151 local_inc(&cpu_buffer->commit_overrun);
2156 * This is where the fun begins!
2158 * We are fighting against races between a reader that
2159 * could be on another CPU trying to swap its reader
2160 * page with the buffer head.
2162 * We are also fighting against interrupts coming in and
2163 * moving the head or tail on us as well.
2165 * If the next page is the head page then we have filled
2166 * the buffer, unless the commit page is still on the
2169 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2172 * If the commit is not on the reader page, then
2173 * move the header page.
2175 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2177 * If we are not in overwrite mode,
2178 * this is easy, just stop here.
2180 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2181 local_inc(&cpu_buffer->dropped_events);
2185 ret = rb_handle_head_page(cpu_buffer,
2194 * We need to be careful here too. The
2195 * commit page could still be on the reader
2196 * page. We could have a small buffer, and
2197 * have filled up the buffer with events
2198 * from interrupts and such, and wrapped.
2200 * Note, if the tail page is also the on the
2201 * reader_page, we let it move out.
2203 if (unlikely((cpu_buffer->commit_page !=
2204 cpu_buffer->tail_page) &&
2205 (cpu_buffer->commit_page ==
2206 cpu_buffer->reader_page))) {
2207 local_inc(&cpu_buffer->commit_overrun);
2213 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2217 rb_reset_tail(cpu_buffer, tail, info);
2219 /* Commit what we have for now. */
2220 rb_end_commit(cpu_buffer);
2221 /* rb_end_commit() decs committing */
2222 local_inc(&cpu_buffer->committing);
2224 /* fail and let the caller try again */
2225 return ERR_PTR(-EAGAIN);
2229 rb_reset_tail(cpu_buffer, tail, info);
2234 /* Slow path, do not inline */
2235 static noinline struct ring_buffer_event *
2236 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
2238 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2240 /* Not the first event on the page? */
2241 if (rb_event_index(event)) {
2242 event->time_delta = delta & TS_MASK;
2243 event->array[0] = delta >> TS_SHIFT;
2245 /* nope, just zero it */
2246 event->time_delta = 0;
2247 event->array[0] = 0;
2250 return skip_time_extend(event);
2253 static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2254 struct ring_buffer_event *event);
2257 * rb_update_event - update event type and data
2258 * @event: the event to update
2259 * @type: the type of event
2260 * @length: the size of the event field in the ring buffer
2262 * Update the type and data fields of the event. The length
2263 * is the actual size that is written to the ring buffer,
2264 * and with this, we can determine what to place into the
2268 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2269 struct ring_buffer_event *event,
2270 struct rb_event_info *info)
2272 unsigned length = info->length;
2273 u64 delta = info->delta;
2275 /* Only a commit updates the timestamp */
2276 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2280 * If we need to add a timestamp, then we
2281 * add it to the start of the resevered space.
2283 if (unlikely(info->add_timestamp)) {
2284 event = rb_add_time_stamp(event, delta);
2285 length -= RB_LEN_TIME_EXTEND;
2289 event->time_delta = delta;
2290 length -= RB_EVNT_HDR_SIZE;
2291 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2292 event->type_len = 0;
2293 event->array[0] = length;
2295 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2298 static unsigned rb_calculate_event_length(unsigned length)
2300 struct ring_buffer_event event; /* Used only for sizeof array */
2302 /* zero length can cause confusions */
2306 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2307 length += sizeof(event.array[0]);
2309 length += RB_EVNT_HDR_SIZE;
2310 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2313 * In case the time delta is larger than the 27 bits for it
2314 * in the header, we need to add a timestamp. If another
2315 * event comes in when trying to discard this one to increase
2316 * the length, then the timestamp will be added in the allocated
2317 * space of this event. If length is bigger than the size needed
2318 * for the TIME_EXTEND, then padding has to be used. The events
2319 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2320 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2321 * As length is a multiple of 4, we only need to worry if it
2322 * is 12 (RB_LEN_TIME_EXTEND + 4).
2324 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2325 length += RB_ALIGNMENT;
2330 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2331 static inline bool sched_clock_stable(void)
2338 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2339 struct ring_buffer_event *event)
2341 unsigned long new_index, old_index;
2342 struct buffer_page *bpage;
2343 unsigned long index;
2346 new_index = rb_event_index(event);
2347 old_index = new_index + rb_event_ts_length(event);
2348 addr = (unsigned long)event;
2351 bpage = READ_ONCE(cpu_buffer->tail_page);
2353 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2354 unsigned long write_mask =
2355 local_read(&bpage->write) & ~RB_WRITE_MASK;
2356 unsigned long event_length = rb_event_length(event);
2358 * This is on the tail page. It is possible that
2359 * a write could come in and move the tail page
2360 * and write to the next page. That is fine
2361 * because we just shorten what is on this page.
2363 old_index += write_mask;
2364 new_index += write_mask;
2365 index = local_cmpxchg(&bpage->write, old_index, new_index);
2366 if (index == old_index) {
2367 /* update counters */
2368 local_sub(event_length, &cpu_buffer->entries_bytes);
2373 /* could not discard */
2377 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2379 local_inc(&cpu_buffer->committing);
2380 local_inc(&cpu_buffer->commits);
2383 static __always_inline void
2384 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2386 unsigned long max_count;
2389 * We only race with interrupts and NMIs on this CPU.
2390 * If we own the commit event, then we can commit
2391 * all others that interrupted us, since the interruptions
2392 * are in stack format (they finish before they come
2393 * back to us). This allows us to do a simple loop to
2394 * assign the commit to the tail.
2397 max_count = cpu_buffer->nr_pages * 100;
2399 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2400 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2402 if (RB_WARN_ON(cpu_buffer,
2403 rb_is_reader_page(cpu_buffer->tail_page)))
2406 * No need for a memory barrier here, as the update
2407 * of the tail_page did it for this page.
2409 local_set(&cpu_buffer->commit_page->page->commit,
2410 rb_page_write(cpu_buffer->commit_page));
2411 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2412 /* Only update the write stamp if the page has an event */
2413 if (rb_page_write(cpu_buffer->commit_page))
2414 cpu_buffer->write_stamp =
2415 cpu_buffer->commit_page->page->time_stamp;
2416 /* add barrier to keep gcc from optimizing too much */
2419 while (rb_commit_index(cpu_buffer) !=
2420 rb_page_write(cpu_buffer->commit_page)) {
2422 /* Make sure the readers see the content of what is committed. */
2424 local_set(&cpu_buffer->commit_page->page->commit,
2425 rb_page_write(cpu_buffer->commit_page));
2426 RB_WARN_ON(cpu_buffer,
2427 local_read(&cpu_buffer->commit_page->page->commit) &
2432 /* again, keep gcc from optimizing */
2436 * If an interrupt came in just after the first while loop
2437 * and pushed the tail page forward, we will be left with
2438 * a dangling commit that will never go forward.
2440 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2444 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2446 unsigned long commits;
2448 if (RB_WARN_ON(cpu_buffer,
2449 !local_read(&cpu_buffer->committing)))
2453 commits = local_read(&cpu_buffer->commits);
2454 /* synchronize with interrupts */
2456 if (local_read(&cpu_buffer->committing) == 1)
2457 rb_set_commit_to_write(cpu_buffer);
2459 local_dec(&cpu_buffer->committing);
2461 /* synchronize with interrupts */
2465 * Need to account for interrupts coming in between the
2466 * updating of the commit page and the clearing of the
2467 * committing counter.
2469 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2470 !local_read(&cpu_buffer->committing)) {
2471 local_inc(&cpu_buffer->committing);
2476 static inline void rb_event_discard(struct ring_buffer_event *event)
2478 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2479 event = skip_time_extend(event);
2481 /* array[0] holds the actual length for the discarded event */
2482 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2483 event->type_len = RINGBUF_TYPE_PADDING;
2484 /* time delta must be non zero */
2485 if (!event->time_delta)
2486 event->time_delta = 1;
2489 static __always_inline bool
2490 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2491 struct ring_buffer_event *event)
2493 unsigned long addr = (unsigned long)event;
2494 unsigned long index;
2496 index = rb_event_index(event);
2499 return cpu_buffer->commit_page->page == (void *)addr &&
2500 rb_commit_index(cpu_buffer) == index;
2503 static __always_inline void
2504 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2505 struct ring_buffer_event *event)
2510 * The event first in the commit queue updates the
2513 if (rb_event_is_commit(cpu_buffer, event)) {
2515 * A commit event that is first on a page
2516 * updates the write timestamp with the page stamp
2518 if (!rb_event_index(event))
2519 cpu_buffer->write_stamp =
2520 cpu_buffer->commit_page->page->time_stamp;
2521 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2522 delta = event->array[0];
2524 delta += event->time_delta;
2525 cpu_buffer->write_stamp += delta;
2527 cpu_buffer->write_stamp += event->time_delta;
2531 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2532 struct ring_buffer_event *event)
2534 local_inc(&cpu_buffer->entries);
2535 rb_update_write_stamp(cpu_buffer, event);
2536 rb_end_commit(cpu_buffer);
2539 static __always_inline void
2540 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2544 if (buffer->irq_work.waiters_pending) {
2545 buffer->irq_work.waiters_pending = false;
2546 /* irq_work_queue() supplies it's own memory barriers */
2547 irq_work_queue(&buffer->irq_work.work);
2550 if (cpu_buffer->irq_work.waiters_pending) {
2551 cpu_buffer->irq_work.waiters_pending = false;
2552 /* irq_work_queue() supplies it's own memory barriers */
2553 irq_work_queue(&cpu_buffer->irq_work.work);
2556 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
2558 if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2559 cpu_buffer->irq_work.wakeup_full = true;
2560 cpu_buffer->irq_work.full_waiters_pending = false;
2561 /* irq_work_queue() supplies it's own memory barriers */
2562 irq_work_queue(&cpu_buffer->irq_work.work);
2567 * The lock and unlock are done within a preempt disable section.
2568 * The current_context per_cpu variable can only be modified
2569 * by the current task between lock and unlock. But it can
2570 * be modified more than once via an interrupt. To pass this
2571 * information from the lock to the unlock without having to
2572 * access the 'in_interrupt()' functions again (which do show
2573 * a bit of overhead in something as critical as function tracing,
2574 * we use a bitmask trick.
2576 * bit 1 = NMI context
2577 * bit 2 = IRQ context
2578 * bit 3 = SoftIRQ context
2579 * bit 4 = normal context.
2581 * This works because this is the order of contexts that can
2582 * preempt other contexts. A SoftIRQ never preempts an IRQ
2585 * When the context is determined, the corresponding bit is
2586 * checked and set (if it was set, then a recursion of that context
2589 * On unlock, we need to clear this bit. To do so, just subtract
2590 * 1 from the current_context and AND it to itself.
2594 * 101 & 100 = 100 (clearing bit zero)
2597 * 1010 & 1001 = 1000 (clearing bit 1)
2599 * The least significant bit can be cleared this way, and it
2600 * just so happens that it is the same bit corresponding to
2601 * the current context.
2603 * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
2604 * is set when a recursion is detected at the current context, and if
2605 * the TRANSITION bit is already set, it will fail the recursion.
2606 * This is needed because there's a lag between the changing of
2607 * interrupt context and updating the preempt count. In this case,
2608 * a false positive will be found. To handle this, one extra recursion
2609 * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
2610 * bit is already set, then it is considered a recursion and the function
2611 * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
2613 * On the trace_recursive_unlock(), the TRANSITION bit will be the first
2614 * to be cleared. Even if it wasn't the context that set it. That is,
2615 * if an interrupt comes in while NORMAL bit is set and the ring buffer
2616 * is called before preempt_count() is updated, since the check will
2617 * be on the NORMAL bit, the TRANSITION bit will then be set. If an
2618 * NMI then comes in, it will set the NMI bit, but when the NMI code
2619 * does the trace_recursive_unlock() it will clear the TRANSTION bit
2620 * and leave the NMI bit set. But this is fine, because the interrupt
2621 * code that set the TRANSITION bit will then clear the NMI bit when it
2622 * calls trace_recursive_unlock(). If another NMI comes in, it will
2623 * set the TRANSITION bit and continue.
2625 * Note: The TRANSITION bit only handles a single transition between context.
2628 static __always_inline int
2629 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2631 unsigned int val = cpu_buffer->current_context;
2634 if (in_interrupt()) {
2640 bit = RB_CTX_SOFTIRQ;
2642 bit = RB_CTX_NORMAL;
2644 if (unlikely(val & (1 << bit))) {
2646 * It is possible that this was called by transitioning
2647 * between interrupt context, and preempt_count() has not
2648 * been updated yet. In this case, use the TRANSITION bit.
2650 bit = RB_CTX_TRANSITION;
2651 if (val & (1 << bit))
2656 cpu_buffer->current_context = val;
2661 static __always_inline void
2662 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2664 cpu_buffer->current_context &= cpu_buffer->current_context - 1;
2668 * ring_buffer_unlock_commit - commit a reserved
2669 * @buffer: The buffer to commit to
2670 * @event: The event pointer to commit.
2672 * This commits the data to the ring buffer, and releases any locks held.
2674 * Must be paired with ring_buffer_lock_reserve.
2676 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2677 struct ring_buffer_event *event)
2679 struct ring_buffer_per_cpu *cpu_buffer;
2680 int cpu = raw_smp_processor_id();
2682 cpu_buffer = buffer->buffers[cpu];
2684 rb_commit(cpu_buffer, event);
2686 rb_wakeups(buffer, cpu_buffer);
2688 trace_recursive_unlock(cpu_buffer);
2690 preempt_enable_notrace();
2694 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2696 static noinline void
2697 rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2698 struct rb_event_info *info)
2700 WARN_ONCE(info->delta > (1ULL << 59),
2701 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2702 (unsigned long long)info->delta,
2703 (unsigned long long)info->ts,
2704 (unsigned long long)cpu_buffer->write_stamp,
2705 sched_clock_stable() ? "" :
2706 "If you just came from a suspend/resume,\n"
2707 "please switch to the trace global clock:\n"
2708 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2709 info->add_timestamp = 1;
2712 static struct ring_buffer_event *
2713 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2714 struct rb_event_info *info)
2716 struct ring_buffer_event *event;
2717 struct buffer_page *tail_page;
2718 unsigned long tail, write;
2721 * If the time delta since the last event is too big to
2722 * hold in the time field of the event, then we append a
2723 * TIME EXTEND event ahead of the data event.
2725 if (unlikely(info->add_timestamp))
2726 info->length += RB_LEN_TIME_EXTEND;
2728 /* Don't let the compiler play games with cpu_buffer->tail_page */
2729 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2730 write = local_add_return(info->length, &tail_page->write);
2732 /* set write to only the index of the write */
2733 write &= RB_WRITE_MASK;
2734 tail = write - info->length;
2737 * If this is the first commit on the page, then it has the same
2738 * timestamp as the page itself.
2743 /* See if we shot pass the end of this buffer page */
2744 if (unlikely(write > BUF_PAGE_SIZE))
2745 return rb_move_tail(cpu_buffer, tail, info);
2747 /* We reserved something on the buffer */
2749 event = __rb_page_index(tail_page, tail);
2750 rb_update_event(cpu_buffer, event, info);
2752 local_inc(&tail_page->entries);
2755 * If this is the first commit on the page, then update
2759 tail_page->page->time_stamp = info->ts;
2761 /* account for these added bytes */
2762 local_add(info->length, &cpu_buffer->entries_bytes);
2767 static __always_inline struct ring_buffer_event *
2768 rb_reserve_next_event(struct ring_buffer *buffer,
2769 struct ring_buffer_per_cpu *cpu_buffer,
2770 unsigned long length)
2772 struct ring_buffer_event *event;
2773 struct rb_event_info info;
2777 rb_start_commit(cpu_buffer);
2779 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2781 * Due to the ability to swap a cpu buffer from a buffer
2782 * it is possible it was swapped before we committed.
2783 * (committing stops a swap). We check for it here and
2784 * if it happened, we have to fail the write.
2787 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2788 local_dec(&cpu_buffer->committing);
2789 local_dec(&cpu_buffer->commits);
2794 info.length = rb_calculate_event_length(length);
2796 info.add_timestamp = 0;
2800 * We allow for interrupts to reenter here and do a trace.
2801 * If one does, it will cause this original code to loop
2802 * back here. Even with heavy interrupts happening, this
2803 * should only happen a few times in a row. If this happens
2804 * 1000 times in a row, there must be either an interrupt
2805 * storm or we have something buggy.
2808 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2811 info.ts = rb_time_stamp(cpu_buffer->buffer);
2812 diff = info.ts - cpu_buffer->write_stamp;
2814 /* make sure this diff is calculated here */
2817 /* Did the write stamp get updated already? */
2818 if (likely(info.ts >= cpu_buffer->write_stamp)) {
2820 if (unlikely(test_time_stamp(info.delta)))
2821 rb_handle_timestamp(cpu_buffer, &info);
2824 event = __rb_reserve_next(cpu_buffer, &info);
2826 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2827 if (info.add_timestamp)
2828 info.length -= RB_LEN_TIME_EXTEND;
2838 rb_end_commit(cpu_buffer);
2843 * ring_buffer_lock_reserve - reserve a part of the buffer
2844 * @buffer: the ring buffer to reserve from
2845 * @length: the length of the data to reserve (excluding event header)
2847 * Returns a reseverd event on the ring buffer to copy directly to.
2848 * The user of this interface will need to get the body to write into
2849 * and can use the ring_buffer_event_data() interface.
2851 * The length is the length of the data needed, not the event length
2852 * which also includes the event header.
2854 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2855 * If NULL is returned, then nothing has been allocated or locked.
2857 struct ring_buffer_event *
2858 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2860 struct ring_buffer_per_cpu *cpu_buffer;
2861 struct ring_buffer_event *event;
2864 /* If we are tracing schedule, we don't want to recurse */
2865 preempt_disable_notrace();
2867 if (unlikely(atomic_read(&buffer->record_disabled)))
2870 cpu = raw_smp_processor_id();
2872 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2875 cpu_buffer = buffer->buffers[cpu];
2877 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2880 if (unlikely(length > BUF_MAX_DATA_SIZE))
2883 if (unlikely(trace_recursive_lock(cpu_buffer)))
2886 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2893 trace_recursive_unlock(cpu_buffer);
2895 preempt_enable_notrace();
2898 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2901 * Decrement the entries to the page that an event is on.
2902 * The event does not even need to exist, only the pointer
2903 * to the page it is on. This may only be called before the commit
2907 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2908 struct ring_buffer_event *event)
2910 unsigned long addr = (unsigned long)event;
2911 struct buffer_page *bpage = cpu_buffer->commit_page;
2912 struct buffer_page *start;
2916 /* Do the likely case first */
2917 if (likely(bpage->page == (void *)addr)) {
2918 local_dec(&bpage->entries);
2923 * Because the commit page may be on the reader page we
2924 * start with the next page and check the end loop there.
2926 rb_inc_page(cpu_buffer, &bpage);
2929 if (bpage->page == (void *)addr) {
2930 local_dec(&bpage->entries);
2933 rb_inc_page(cpu_buffer, &bpage);
2934 } while (bpage != start);
2936 /* commit not part of this buffer?? */
2937 RB_WARN_ON(cpu_buffer, 1);
2941 * ring_buffer_commit_discard - discard an event that has not been committed
2942 * @buffer: the ring buffer
2943 * @event: non committed event to discard
2945 * Sometimes an event that is in the ring buffer needs to be ignored.
2946 * This function lets the user discard an event in the ring buffer
2947 * and then that event will not be read later.
2949 * This function only works if it is called before the the item has been
2950 * committed. It will try to free the event from the ring buffer
2951 * if another event has not been added behind it.
2953 * If another event has been added behind it, it will set the event
2954 * up as discarded, and perform the commit.
2956 * If this function is called, do not call ring_buffer_unlock_commit on
2959 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2960 struct ring_buffer_event *event)
2962 struct ring_buffer_per_cpu *cpu_buffer;
2965 /* The event is discarded regardless */
2966 rb_event_discard(event);
2968 cpu = smp_processor_id();
2969 cpu_buffer = buffer->buffers[cpu];
2972 * This must only be called if the event has not been
2973 * committed yet. Thus we can assume that preemption
2974 * is still disabled.
2976 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2978 rb_decrement_entry(cpu_buffer, event);
2979 if (rb_try_to_discard(cpu_buffer, event))
2983 * The commit is still visible by the reader, so we
2984 * must still update the timestamp.
2986 rb_update_write_stamp(cpu_buffer, event);
2988 rb_end_commit(cpu_buffer);
2990 trace_recursive_unlock(cpu_buffer);
2992 preempt_enable_notrace();
2995 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2998 * ring_buffer_write - write data to the buffer without reserving
2999 * @buffer: The ring buffer to write to.
3000 * @length: The length of the data being written (excluding the event header)
3001 * @data: The data to write to the buffer.
3003 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3004 * one function. If you already have the data to write to the buffer, it
3005 * may be easier to simply call this function.
3007 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3008 * and not the length of the event which would hold the header.
3010 int ring_buffer_write(struct ring_buffer *buffer,
3011 unsigned long length,
3014 struct ring_buffer_per_cpu *cpu_buffer;
3015 struct ring_buffer_event *event;
3020 preempt_disable_notrace();
3022 if (atomic_read(&buffer->record_disabled))
3025 cpu = raw_smp_processor_id();
3027 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3030 cpu_buffer = buffer->buffers[cpu];
3032 if (atomic_read(&cpu_buffer->record_disabled))
3035 if (length > BUF_MAX_DATA_SIZE)
3038 if (unlikely(trace_recursive_lock(cpu_buffer)))
3041 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3045 body = rb_event_data(event);
3047 memcpy(body, data, length);
3049 rb_commit(cpu_buffer, event);
3051 rb_wakeups(buffer, cpu_buffer);
3056 trace_recursive_unlock(cpu_buffer);
3059 preempt_enable_notrace();
3063 EXPORT_SYMBOL_GPL(ring_buffer_write);
3065 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3067 struct buffer_page *reader = cpu_buffer->reader_page;
3068 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3069 struct buffer_page *commit = cpu_buffer->commit_page;
3071 /* In case of error, head will be NULL */
3072 if (unlikely(!head))
3075 /* Reader should exhaust content in reader page */
3076 if (reader->read != rb_page_commit(reader))
3080 * If writers are committing on the reader page, knowing all
3081 * committed content has been read, the ring buffer is empty.
3083 if (commit == reader)
3087 * If writers are committing on a page other than reader page
3088 * and head page, there should always be content to read.
3094 * Writers are committing on the head page, we just need
3095 * to care about there're committed data, and the reader will
3096 * swap reader page with head page when it is to read data.
3098 return rb_page_commit(commit) == 0;
3102 * ring_buffer_record_disable - stop all writes into the buffer
3103 * @buffer: The ring buffer to stop writes to.
3105 * This prevents all writes to the buffer. Any attempt to write
3106 * to the buffer after this will fail and return NULL.
3108 * The caller should call synchronize_sched() after this.
3110 void ring_buffer_record_disable(struct ring_buffer *buffer)
3112 atomic_inc(&buffer->record_disabled);
3114 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3117 * ring_buffer_record_enable - enable writes to the buffer
3118 * @buffer: The ring buffer to enable writes
3120 * Note, multiple disables will need the same number of enables
3121 * to truly enable the writing (much like preempt_disable).
3123 void ring_buffer_record_enable(struct ring_buffer *buffer)
3125 atomic_dec(&buffer->record_disabled);
3127 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3130 * ring_buffer_record_off - stop all writes into the buffer
3131 * @buffer: The ring buffer to stop writes to.
3133 * This prevents all writes to the buffer. Any attempt to write
3134 * to the buffer after this will fail and return NULL.
3136 * This is different than ring_buffer_record_disable() as
3137 * it works like an on/off switch, where as the disable() version
3138 * must be paired with a enable().
3140 void ring_buffer_record_off(struct ring_buffer *buffer)
3143 unsigned int new_rd;
3146 rd = atomic_read(&buffer->record_disabled);
3147 new_rd = rd | RB_BUFFER_OFF;
3148 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3150 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3153 * ring_buffer_record_on - restart writes into the buffer
3154 * @buffer: The ring buffer to start writes to.
3156 * This enables all writes to the buffer that was disabled by
3157 * ring_buffer_record_off().
3159 * This is different than ring_buffer_record_enable() as
3160 * it works like an on/off switch, where as the enable() version
3161 * must be paired with a disable().
3163 void ring_buffer_record_on(struct ring_buffer *buffer)
3166 unsigned int new_rd;
3169 rd = atomic_read(&buffer->record_disabled);
3170 new_rd = rd & ~RB_BUFFER_OFF;
3171 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3173 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3176 * ring_buffer_record_is_on - return true if the ring buffer can write
3177 * @buffer: The ring buffer to see if write is enabled
3179 * Returns true if the ring buffer is in a state that it accepts writes.
3181 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3183 return !atomic_read(&buffer->record_disabled);
3187 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
3188 * @buffer: The ring buffer to see if write is set enabled
3190 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
3191 * Note that this does NOT mean it is in a writable state.
3193 * It may return true when the ring buffer has been disabled by
3194 * ring_buffer_record_disable(), as that is a temporary disabling of
3197 int ring_buffer_record_is_set_on(struct ring_buffer *buffer)
3199 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
3203 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3204 * @buffer: The ring buffer to stop writes to.
3205 * @cpu: The CPU buffer to stop
3207 * This prevents all writes to the buffer. Any attempt to write
3208 * to the buffer after this will fail and return NULL.
3210 * The caller should call synchronize_sched() after this.
3212 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3214 struct ring_buffer_per_cpu *cpu_buffer;
3216 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3219 cpu_buffer = buffer->buffers[cpu];
3220 atomic_inc(&cpu_buffer->record_disabled);
3222 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3225 * ring_buffer_record_enable_cpu - enable writes to the buffer
3226 * @buffer: The ring buffer to enable writes
3227 * @cpu: The CPU to enable.
3229 * Note, multiple disables will need the same number of enables
3230 * to truly enable the writing (much like preempt_disable).
3232 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3234 struct ring_buffer_per_cpu *cpu_buffer;
3236 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3239 cpu_buffer = buffer->buffers[cpu];
3240 atomic_dec(&cpu_buffer->record_disabled);
3242 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3245 * The total entries in the ring buffer is the running counter
3246 * of entries entered into the ring buffer, minus the sum of
3247 * the entries read from the ring buffer and the number of
3248 * entries that were overwritten.
3250 static inline unsigned long
3251 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3253 return local_read(&cpu_buffer->entries) -
3254 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3258 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3259 * @buffer: The ring buffer
3260 * @cpu: The per CPU buffer to read from.
3262 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3264 unsigned long flags;
3265 struct ring_buffer_per_cpu *cpu_buffer;
3266 struct buffer_page *bpage;
3269 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3272 cpu_buffer = buffer->buffers[cpu];
3273 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3275 * if the tail is on reader_page, oldest time stamp is on the reader
3278 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3279 bpage = cpu_buffer->reader_page;
3281 bpage = rb_set_head_page(cpu_buffer);
3283 ret = bpage->page->time_stamp;
3284 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3288 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3291 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3292 * @buffer: The ring buffer
3293 * @cpu: The per CPU buffer to read from.
3295 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3297 struct ring_buffer_per_cpu *cpu_buffer;
3300 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3303 cpu_buffer = buffer->buffers[cpu];
3304 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3308 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3311 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3312 * @buffer: The ring buffer
3313 * @cpu: The per CPU buffer to get the entries from.
3315 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3317 struct ring_buffer_per_cpu *cpu_buffer;
3319 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3322 cpu_buffer = buffer->buffers[cpu];
3324 return rb_num_of_entries(cpu_buffer);
3326 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3329 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3330 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3331 * @buffer: The ring buffer
3332 * @cpu: The per CPU buffer to get the number of overruns from
3334 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3336 struct ring_buffer_per_cpu *cpu_buffer;
3339 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3342 cpu_buffer = buffer->buffers[cpu];
3343 ret = local_read(&cpu_buffer->overrun);
3347 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3350 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3351 * commits failing due to the buffer wrapping around while there are uncommitted
3352 * events, such as during an interrupt storm.
3353 * @buffer: The ring buffer
3354 * @cpu: The per CPU buffer to get the number of overruns from
3357 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3359 struct ring_buffer_per_cpu *cpu_buffer;
3362 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3365 cpu_buffer = buffer->buffers[cpu];
3366 ret = local_read(&cpu_buffer->commit_overrun);
3370 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3373 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3374 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3375 * @buffer: The ring buffer
3376 * @cpu: The per CPU buffer to get the number of overruns from
3379 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3381 struct ring_buffer_per_cpu *cpu_buffer;
3384 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3387 cpu_buffer = buffer->buffers[cpu];
3388 ret = local_read(&cpu_buffer->dropped_events);
3392 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3395 * ring_buffer_read_events_cpu - get the number of events successfully read
3396 * @buffer: The ring buffer
3397 * @cpu: The per CPU buffer to get the number of events read
3400 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3402 struct ring_buffer_per_cpu *cpu_buffer;
3404 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3407 cpu_buffer = buffer->buffers[cpu];
3408 return cpu_buffer->read;
3410 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3413 * ring_buffer_entries - get the number of entries in a buffer
3414 * @buffer: The ring buffer
3416 * Returns the total number of entries in the ring buffer
3419 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3421 struct ring_buffer_per_cpu *cpu_buffer;
3422 unsigned long entries = 0;
3425 /* if you care about this being correct, lock the buffer */
3426 for_each_buffer_cpu(buffer, cpu) {
3427 cpu_buffer = buffer->buffers[cpu];
3428 entries += rb_num_of_entries(cpu_buffer);
3433 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3436 * ring_buffer_overruns - get the number of overruns in buffer
3437 * @buffer: The ring buffer
3439 * Returns the total number of overruns in the ring buffer
3442 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3444 struct ring_buffer_per_cpu *cpu_buffer;
3445 unsigned long overruns = 0;
3448 /* if you care about this being correct, lock the buffer */
3449 for_each_buffer_cpu(buffer, cpu) {
3450 cpu_buffer = buffer->buffers[cpu];
3451 overruns += local_read(&cpu_buffer->overrun);
3456 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3458 static void rb_iter_reset(struct ring_buffer_iter *iter)
3460 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3462 /* Iterator usage is expected to have record disabled */
3463 iter->head_page = cpu_buffer->reader_page;
3464 iter->head = cpu_buffer->reader_page->read;
3466 iter->cache_reader_page = iter->head_page;
3467 iter->cache_read = cpu_buffer->read;
3468 iter->cache_pages_removed = cpu_buffer->pages_removed;
3471 iter->read_stamp = cpu_buffer->read_stamp;
3473 iter->read_stamp = iter->head_page->page->time_stamp;
3477 * ring_buffer_iter_reset - reset an iterator
3478 * @iter: The iterator to reset
3480 * Resets the iterator, so that it will start from the beginning
3483 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3485 struct ring_buffer_per_cpu *cpu_buffer;
3486 unsigned long flags;
3491 cpu_buffer = iter->cpu_buffer;
3493 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3494 rb_iter_reset(iter);
3495 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3497 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3500 * ring_buffer_iter_empty - check if an iterator has no more to read
3501 * @iter: The iterator to check
3503 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3505 struct ring_buffer_per_cpu *cpu_buffer;
3506 struct buffer_page *reader;
3507 struct buffer_page *head_page;
3508 struct buffer_page *commit_page;
3511 cpu_buffer = iter->cpu_buffer;
3513 /* Remember, trace recording is off when iterator is in use */
3514 reader = cpu_buffer->reader_page;
3515 head_page = cpu_buffer->head_page;
3516 commit_page = cpu_buffer->commit_page;
3517 commit = rb_page_commit(commit_page);
3519 return ((iter->head_page == commit_page && iter->head == commit) ||
3520 (iter->head_page == reader && commit_page == head_page &&
3521 head_page->read == commit &&
3522 iter->head == rb_page_commit(cpu_buffer->reader_page)));
3524 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3527 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3528 struct ring_buffer_event *event)
3532 switch (event->type_len) {
3533 case RINGBUF_TYPE_PADDING:
3536 case RINGBUF_TYPE_TIME_EXTEND:
3537 delta = event->array[0];
3539 delta += event->time_delta;
3540 cpu_buffer->read_stamp += delta;
3543 case RINGBUF_TYPE_TIME_STAMP:
3544 /* FIXME: not implemented */
3547 case RINGBUF_TYPE_DATA:
3548 cpu_buffer->read_stamp += event->time_delta;
3558 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3559 struct ring_buffer_event *event)
3563 switch (event->type_len) {
3564 case RINGBUF_TYPE_PADDING:
3567 case RINGBUF_TYPE_TIME_EXTEND:
3568 delta = event->array[0];
3570 delta += event->time_delta;
3571 iter->read_stamp += delta;
3574 case RINGBUF_TYPE_TIME_STAMP:
3575 /* FIXME: not implemented */
3578 case RINGBUF_TYPE_DATA:
3579 iter->read_stamp += event->time_delta;
3588 static struct buffer_page *
3589 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3591 struct buffer_page *reader = NULL;
3592 unsigned long overwrite;
3593 unsigned long flags;
3597 local_irq_save(flags);
3598 arch_spin_lock(&cpu_buffer->lock);
3602 * This should normally only loop twice. But because the
3603 * start of the reader inserts an empty page, it causes
3604 * a case where we will loop three times. There should be no
3605 * reason to loop four times (that I know of).
3607 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3612 reader = cpu_buffer->reader_page;
3614 /* If there's more to read, return this page */
3615 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3618 /* Never should we have an index greater than the size */
3619 if (RB_WARN_ON(cpu_buffer,
3620 cpu_buffer->reader_page->read > rb_page_size(reader)))
3623 /* check if we caught up to the tail */
3625 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3628 /* Don't bother swapping if the ring buffer is empty */
3629 if (rb_num_of_entries(cpu_buffer) == 0)
3633 * Reset the reader page to size zero.
3635 local_set(&cpu_buffer->reader_page->write, 0);
3636 local_set(&cpu_buffer->reader_page->entries, 0);
3637 local_set(&cpu_buffer->reader_page->page->commit, 0);
3638 cpu_buffer->reader_page->real_end = 0;
3642 * Splice the empty reader page into the list around the head.
3644 reader = rb_set_head_page(cpu_buffer);
3647 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3648 cpu_buffer->reader_page->list.prev = reader->list.prev;
3651 * cpu_buffer->pages just needs to point to the buffer, it
3652 * has no specific buffer page to point to. Lets move it out
3653 * of our way so we don't accidentally swap it.
3655 cpu_buffer->pages = reader->list.prev;
3657 /* The reader page will be pointing to the new head */
3658 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3661 * We want to make sure we read the overruns after we set up our
3662 * pointers to the next object. The writer side does a
3663 * cmpxchg to cross pages which acts as the mb on the writer
3664 * side. Note, the reader will constantly fail the swap
3665 * while the writer is updating the pointers, so this
3666 * guarantees that the overwrite recorded here is the one we
3667 * want to compare with the last_overrun.
3670 overwrite = local_read(&(cpu_buffer->overrun));
3673 * Here's the tricky part.
3675 * We need to move the pointer past the header page.
3676 * But we can only do that if a writer is not currently
3677 * moving it. The page before the header page has the
3678 * flag bit '1' set if it is pointing to the page we want.
3679 * but if the writer is in the process of moving it
3680 * than it will be '2' or already moved '0'.
3683 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3686 * If we did not convert it, then we must try again.
3692 * Yeah! We succeeded in replacing the page.
3694 * Now make the new head point back to the reader page.
3696 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3697 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3699 /* Finally update the reader page to the new head */
3700 cpu_buffer->reader_page = reader;
3701 cpu_buffer->reader_page->read = 0;
3703 if (overwrite != cpu_buffer->last_overrun) {
3704 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3705 cpu_buffer->last_overrun = overwrite;
3711 /* Update the read_stamp on the first event */
3712 if (reader && reader->read == 0)
3713 cpu_buffer->read_stamp = reader->page->time_stamp;
3715 arch_spin_unlock(&cpu_buffer->lock);
3716 local_irq_restore(flags);
3719 * The writer has preempt disable, wait for it. But not forever
3720 * Although, 1 second is pretty much "forever"
3722 #define USECS_WAIT 1000000
3723 for (nr_loops = 0; nr_loops < USECS_WAIT; nr_loops++) {
3724 /* If the write is past the end of page, a writer is still updating it */
3725 if (likely(!reader || rb_page_write(reader) <= BUF_PAGE_SIZE))
3730 /* Get the latest version of the reader write value */
3734 /* The writer is not moving forward? Something is wrong */
3735 if (RB_WARN_ON(cpu_buffer, nr_loops == USECS_WAIT))
3739 * Make sure we see any padding after the write update
3740 * (see rb_reset_tail()).
3742 * In addition, a writer may be writing on the reader page
3743 * if the page has not been fully filled, so the read barrier
3744 * is also needed to make sure we see the content of what is
3745 * committed by the writer (see rb_set_commit_to_write()).
3753 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3755 struct ring_buffer_event *event;
3756 struct buffer_page *reader;
3759 reader = rb_get_reader_page(cpu_buffer);
3761 /* This function should not be called when buffer is empty */
3762 if (RB_WARN_ON(cpu_buffer, !reader))
3765 event = rb_reader_event(cpu_buffer);
3767 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3770 rb_update_read_stamp(cpu_buffer, event);
3772 length = rb_event_length(event);
3773 cpu_buffer->reader_page->read += length;
3776 static void rb_advance_iter(struct ring_buffer_iter *iter)
3778 struct ring_buffer_per_cpu *cpu_buffer;
3779 struct ring_buffer_event *event;
3782 cpu_buffer = iter->cpu_buffer;
3785 * Check if we are at the end of the buffer.
3787 if (iter->head >= rb_page_size(iter->head_page)) {
3788 /* discarded commits can make the page empty */
3789 if (iter->head_page == cpu_buffer->commit_page)
3795 event = rb_iter_head_event(iter);
3797 length = rb_event_length(event);
3800 * This should not be called to advance the header if we are
3801 * at the tail of the buffer.
3803 if (RB_WARN_ON(cpu_buffer,
3804 (iter->head_page == cpu_buffer->commit_page) &&
3805 (iter->head + length > rb_commit_index(cpu_buffer))))
3808 rb_update_iter_read_stamp(iter, event);
3810 iter->head += length;
3812 /* check for end of page padding */
3813 if ((iter->head >= rb_page_size(iter->head_page)) &&
3814 (iter->head_page != cpu_buffer->commit_page))
3818 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3820 return cpu_buffer->lost_events;
3823 static struct ring_buffer_event *
3824 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3825 unsigned long *lost_events)
3827 struct ring_buffer_event *event;
3828 struct buffer_page *reader;
3833 * We repeat when a time extend is encountered.
3834 * Since the time extend is always attached to a data event,
3835 * we should never loop more than once.
3836 * (We never hit the following condition more than twice).
3838 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3841 reader = rb_get_reader_page(cpu_buffer);
3845 event = rb_reader_event(cpu_buffer);
3847 switch (event->type_len) {
3848 case RINGBUF_TYPE_PADDING:
3849 if (rb_null_event(event))
3850 RB_WARN_ON(cpu_buffer, 1);
3852 * Because the writer could be discarding every
3853 * event it creates (which would probably be bad)
3854 * if we were to go back to "again" then we may never
3855 * catch up, and will trigger the warn on, or lock
3856 * the box. Return the padding, and we will release
3857 * the current locks, and try again.
3861 case RINGBUF_TYPE_TIME_EXTEND:
3862 /* Internal data, OK to advance */
3863 rb_advance_reader(cpu_buffer);
3866 case RINGBUF_TYPE_TIME_STAMP:
3867 /* FIXME: not implemented */
3868 rb_advance_reader(cpu_buffer);
3871 case RINGBUF_TYPE_DATA:
3873 *ts = cpu_buffer->read_stamp + event->time_delta;
3874 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3875 cpu_buffer->cpu, ts);
3878 *lost_events = rb_lost_events(cpu_buffer);
3887 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3889 static struct ring_buffer_event *
3890 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3892 struct ring_buffer *buffer;
3893 struct ring_buffer_per_cpu *cpu_buffer;
3894 struct ring_buffer_event *event;
3897 cpu_buffer = iter->cpu_buffer;
3898 buffer = cpu_buffer->buffer;
3901 * Check if someone performed a consuming read to the buffer
3902 * or removed some pages from the buffer. In these cases,
3903 * iterator was invalidated and we need to reset it.
3905 if (unlikely(iter->cache_read != cpu_buffer->read ||
3906 iter->cache_reader_page != cpu_buffer->reader_page ||
3907 iter->cache_pages_removed != cpu_buffer->pages_removed))
3908 rb_iter_reset(iter);
3911 if (ring_buffer_iter_empty(iter))
3915 * We repeat when a time extend is encountered or we hit
3916 * the end of the page. Since the time extend is always attached
3917 * to a data event, we should never loop more than three times.
3918 * Once for going to next page, once on time extend, and
3919 * finally once to get the event.
3920 * (We never hit the following condition more than thrice).
3922 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3925 if (rb_per_cpu_empty(cpu_buffer))
3928 if (iter->head >= rb_page_size(iter->head_page)) {
3933 event = rb_iter_head_event(iter);
3935 switch (event->type_len) {
3936 case RINGBUF_TYPE_PADDING:
3937 if (rb_null_event(event)) {
3941 rb_advance_iter(iter);
3944 case RINGBUF_TYPE_TIME_EXTEND:
3945 /* Internal data, OK to advance */
3946 rb_advance_iter(iter);
3949 case RINGBUF_TYPE_TIME_STAMP:
3950 /* FIXME: not implemented */
3951 rb_advance_iter(iter);
3954 case RINGBUF_TYPE_DATA:
3956 *ts = iter->read_stamp + event->time_delta;
3957 ring_buffer_normalize_time_stamp(buffer,
3958 cpu_buffer->cpu, ts);
3968 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3970 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
3972 if (likely(!in_nmi())) {
3973 raw_spin_lock(&cpu_buffer->reader_lock);
3978 * If an NMI die dumps out the content of the ring buffer
3979 * trylock must be used to prevent a deadlock if the NMI
3980 * preempted a task that holds the ring buffer locks. If
3981 * we get the lock then all is fine, if not, then continue
3982 * to do the read, but this can corrupt the ring buffer,
3983 * so it must be permanently disabled from future writes.
3984 * Reading from NMI is a oneshot deal.
3986 if (raw_spin_trylock(&cpu_buffer->reader_lock))
3989 /* Continue without locking, but disable the ring buffer */
3990 atomic_inc(&cpu_buffer->record_disabled);
3995 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
3998 raw_spin_unlock(&cpu_buffer->reader_lock);
4003 * ring_buffer_peek - peek at the next event to be read
4004 * @buffer: The ring buffer to read
4005 * @cpu: The cpu to peak at
4006 * @ts: The timestamp counter of this event.
4007 * @lost_events: a variable to store if events were lost (may be NULL)
4009 * This will return the event that will be read next, but does
4010 * not consume the data.
4012 struct ring_buffer_event *
4013 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
4014 unsigned long *lost_events)
4016 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4017 struct ring_buffer_event *event;
4018 unsigned long flags;
4021 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4025 local_irq_save(flags);
4026 dolock = rb_reader_lock(cpu_buffer);
4027 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4028 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4029 rb_advance_reader(cpu_buffer);
4030 rb_reader_unlock(cpu_buffer, dolock);
4031 local_irq_restore(flags);
4033 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4040 * ring_buffer_iter_peek - peek at the next event to be read
4041 * @iter: The ring buffer iterator
4042 * @ts: The timestamp counter of this event.
4044 * This will return the event that will be read next, but does
4045 * not increment the iterator.
4047 struct ring_buffer_event *
4048 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4050 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4051 struct ring_buffer_event *event;
4052 unsigned long flags;
4055 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4056 event = rb_iter_peek(iter, ts);
4057 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4059 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4066 * ring_buffer_consume - return an event and consume it
4067 * @buffer: The ring buffer to get the next event from
4068 * @cpu: the cpu to read the buffer from
4069 * @ts: a variable to store the timestamp (may be NULL)
4070 * @lost_events: a variable to store if events were lost (may be NULL)
4072 * Returns the next event in the ring buffer, and that event is consumed.
4073 * Meaning, that sequential reads will keep returning a different event,
4074 * and eventually empty the ring buffer if the producer is slower.
4076 struct ring_buffer_event *
4077 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
4078 unsigned long *lost_events)
4080 struct ring_buffer_per_cpu *cpu_buffer;
4081 struct ring_buffer_event *event = NULL;
4082 unsigned long flags;
4086 /* might be called in atomic */
4089 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4092 cpu_buffer = buffer->buffers[cpu];
4093 local_irq_save(flags);
4094 dolock = rb_reader_lock(cpu_buffer);
4096 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4098 cpu_buffer->lost_events = 0;
4099 rb_advance_reader(cpu_buffer);
4102 rb_reader_unlock(cpu_buffer, dolock);
4103 local_irq_restore(flags);
4108 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4113 EXPORT_SYMBOL_GPL(ring_buffer_consume);
4116 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4117 * @buffer: The ring buffer to read from
4118 * @cpu: The cpu buffer to iterate over
4119 * @flags: gfp flags to use for memory allocation
4121 * This performs the initial preparations necessary to iterate
4122 * through the buffer. Memory is allocated, buffer recording
4123 * is disabled, and the iterator pointer is returned to the caller.
4125 * Disabling buffer recordng prevents the reading from being
4126 * corrupted. This is not a consuming read, so a producer is not
4129 * After a sequence of ring_buffer_read_prepare calls, the user is
4130 * expected to make at least one call to ring_buffer_read_prepare_sync.
4131 * Afterwards, ring_buffer_read_start is invoked to get things going
4134 * This overall must be paired with ring_buffer_read_finish.
4136 struct ring_buffer_iter *
4137 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu, gfp_t flags)
4139 struct ring_buffer_per_cpu *cpu_buffer;
4140 struct ring_buffer_iter *iter;
4142 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4145 iter = kmalloc(sizeof(*iter), flags);
4149 cpu_buffer = buffer->buffers[cpu];
4151 iter->cpu_buffer = cpu_buffer;
4153 atomic_inc(&buffer->resize_disabled);
4154 atomic_inc(&cpu_buffer->record_disabled);
4158 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4161 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4163 * All previously invoked ring_buffer_read_prepare calls to prepare
4164 * iterators will be synchronized. Afterwards, read_buffer_read_start
4165 * calls on those iterators are allowed.
4168 ring_buffer_read_prepare_sync(void)
4170 synchronize_sched();
4172 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4175 * ring_buffer_read_start - start a non consuming read of the buffer
4176 * @iter: The iterator returned by ring_buffer_read_prepare
4178 * This finalizes the startup of an iteration through the buffer.
4179 * The iterator comes from a call to ring_buffer_read_prepare and
4180 * an intervening ring_buffer_read_prepare_sync must have been
4183 * Must be paired with ring_buffer_read_finish.
4186 ring_buffer_read_start(struct ring_buffer_iter *iter)
4188 struct ring_buffer_per_cpu *cpu_buffer;
4189 unsigned long flags;
4194 cpu_buffer = iter->cpu_buffer;
4196 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4197 arch_spin_lock(&cpu_buffer->lock);
4198 rb_iter_reset(iter);
4199 arch_spin_unlock(&cpu_buffer->lock);
4200 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4202 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4205 * ring_buffer_read_finish - finish reading the iterator of the buffer
4206 * @iter: The iterator retrieved by ring_buffer_start
4208 * This re-enables the recording to the buffer, and frees the
4212 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4214 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4215 unsigned long flags;
4218 * Ring buffer is disabled from recording, here's a good place
4219 * to check the integrity of the ring buffer.
4220 * Must prevent readers from trying to read, as the check
4221 * clears the HEAD page and readers require it.
4223 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4224 rb_check_pages(cpu_buffer);
4225 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4227 atomic_dec(&cpu_buffer->record_disabled);
4228 atomic_dec(&cpu_buffer->buffer->resize_disabled);
4231 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4234 * ring_buffer_read - read the next item in the ring buffer by the iterator
4235 * @iter: The ring buffer iterator
4236 * @ts: The time stamp of the event read.
4238 * This reads the next event in the ring buffer and increments the iterator.
4240 struct ring_buffer_event *
4241 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4243 struct ring_buffer_event *event;
4244 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4245 unsigned long flags;
4247 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4249 event = rb_iter_peek(iter, ts);
4253 if (event->type_len == RINGBUF_TYPE_PADDING)
4256 rb_advance_iter(iter);
4258 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4262 EXPORT_SYMBOL_GPL(ring_buffer_read);
4265 * ring_buffer_size - return the size of the ring buffer (in bytes)
4266 * @buffer: The ring buffer.
4268 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4271 * Earlier, this method returned
4272 * BUF_PAGE_SIZE * buffer->nr_pages
4273 * Since the nr_pages field is now removed, we have converted this to
4274 * return the per cpu buffer value.
4276 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4279 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4281 EXPORT_SYMBOL_GPL(ring_buffer_size);
4283 static void rb_clear_buffer_page(struct buffer_page *page)
4285 local_set(&page->write, 0);
4286 local_set(&page->entries, 0);
4287 rb_init_page(page->page);
4292 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4294 struct buffer_page *page;
4296 rb_head_page_deactivate(cpu_buffer);
4298 cpu_buffer->head_page
4299 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4300 rb_clear_buffer_page(cpu_buffer->head_page);
4301 list_for_each_entry(page, cpu_buffer->pages, list) {
4302 rb_clear_buffer_page(page);
4305 cpu_buffer->tail_page = cpu_buffer->head_page;
4306 cpu_buffer->commit_page = cpu_buffer->head_page;
4308 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4309 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4310 rb_clear_buffer_page(cpu_buffer->reader_page);
4312 local_set(&cpu_buffer->entries_bytes, 0);
4313 local_set(&cpu_buffer->overrun, 0);
4314 local_set(&cpu_buffer->commit_overrun, 0);
4315 local_set(&cpu_buffer->dropped_events, 0);
4316 local_set(&cpu_buffer->entries, 0);
4317 local_set(&cpu_buffer->committing, 0);
4318 local_set(&cpu_buffer->commits, 0);
4319 cpu_buffer->read = 0;
4320 cpu_buffer->read_bytes = 0;
4322 cpu_buffer->write_stamp = 0;
4323 cpu_buffer->read_stamp = 0;
4325 cpu_buffer->lost_events = 0;
4326 cpu_buffer->last_overrun = 0;
4328 rb_head_page_activate(cpu_buffer);
4329 cpu_buffer->pages_removed = 0;
4333 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4334 * @buffer: The ring buffer to reset a per cpu buffer of
4335 * @cpu: The CPU buffer to be reset
4337 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4339 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4340 unsigned long flags;
4342 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4344 /* prevent another thread from changing buffer sizes */
4345 mutex_lock(&buffer->mutex);
4347 atomic_inc(&buffer->resize_disabled);
4348 atomic_inc(&cpu_buffer->record_disabled);
4350 /* Make sure all commits have finished */
4351 synchronize_sched();
4353 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4355 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4358 arch_spin_lock(&cpu_buffer->lock);
4360 rb_reset_cpu(cpu_buffer);
4362 arch_spin_unlock(&cpu_buffer->lock);
4365 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4367 atomic_dec(&cpu_buffer->record_disabled);
4368 atomic_dec(&buffer->resize_disabled);
4370 mutex_unlock(&buffer->mutex);
4372 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4375 * ring_buffer_reset - reset a ring buffer
4376 * @buffer: The ring buffer to reset all cpu buffers
4378 void ring_buffer_reset(struct ring_buffer *buffer)
4382 for_each_buffer_cpu(buffer, cpu)
4383 ring_buffer_reset_cpu(buffer, cpu);
4385 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4388 * rind_buffer_empty - is the ring buffer empty?
4389 * @buffer: The ring buffer to test
4391 bool ring_buffer_empty(struct ring_buffer *buffer)
4393 struct ring_buffer_per_cpu *cpu_buffer;
4394 unsigned long flags;
4399 /* yes this is racy, but if you don't like the race, lock the buffer */
4400 for_each_buffer_cpu(buffer, cpu) {
4401 cpu_buffer = buffer->buffers[cpu];
4402 local_irq_save(flags);
4403 dolock = rb_reader_lock(cpu_buffer);
4404 ret = rb_per_cpu_empty(cpu_buffer);
4405 rb_reader_unlock(cpu_buffer, dolock);
4406 local_irq_restore(flags);
4414 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4417 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4418 * @buffer: The ring buffer
4419 * @cpu: The CPU buffer to test
4421 bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4423 struct ring_buffer_per_cpu *cpu_buffer;
4424 unsigned long flags;
4428 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4431 cpu_buffer = buffer->buffers[cpu];
4432 local_irq_save(flags);
4433 dolock = rb_reader_lock(cpu_buffer);
4434 ret = rb_per_cpu_empty(cpu_buffer);
4435 rb_reader_unlock(cpu_buffer, dolock);
4436 local_irq_restore(flags);
4440 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4442 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4444 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4445 * @buffer_a: One buffer to swap with
4446 * @buffer_b: The other buffer to swap with
4448 * This function is useful for tracers that want to take a "snapshot"
4449 * of a CPU buffer and has another back up buffer lying around.
4450 * it is expected that the tracer handles the cpu buffer not being
4451 * used at the moment.
4453 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4454 struct ring_buffer *buffer_b, int cpu)
4456 struct ring_buffer_per_cpu *cpu_buffer_a;
4457 struct ring_buffer_per_cpu *cpu_buffer_b;
4460 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4461 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4464 cpu_buffer_a = buffer_a->buffers[cpu];
4465 cpu_buffer_b = buffer_b->buffers[cpu];
4467 /* At least make sure the two buffers are somewhat the same */
4468 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4473 if (atomic_read(&buffer_a->record_disabled))
4476 if (atomic_read(&buffer_b->record_disabled))
4479 if (atomic_read(&cpu_buffer_a->record_disabled))
4482 if (atomic_read(&cpu_buffer_b->record_disabled))
4486 * We can't do a synchronize_sched here because this
4487 * function can be called in atomic context.
4488 * Normally this will be called from the same CPU as cpu.
4489 * If not it's up to the caller to protect this.
4491 atomic_inc(&cpu_buffer_a->record_disabled);
4492 atomic_inc(&cpu_buffer_b->record_disabled);
4495 if (local_read(&cpu_buffer_a->committing))
4497 if (local_read(&cpu_buffer_b->committing))
4500 buffer_a->buffers[cpu] = cpu_buffer_b;
4501 buffer_b->buffers[cpu] = cpu_buffer_a;
4503 cpu_buffer_b->buffer = buffer_a;
4504 cpu_buffer_a->buffer = buffer_b;
4509 atomic_dec(&cpu_buffer_a->record_disabled);
4510 atomic_dec(&cpu_buffer_b->record_disabled);
4514 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4515 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4518 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4519 * @buffer: the buffer to allocate for.
4520 * @cpu: the cpu buffer to allocate.
4522 * This function is used in conjunction with ring_buffer_read_page.
4523 * When reading a full page from the ring buffer, these functions
4524 * can be used to speed up the process. The calling function should
4525 * allocate a few pages first with this function. Then when it
4526 * needs to get pages from the ring buffer, it passes the result
4527 * of this function into ring_buffer_read_page, which will swap
4528 * the page that was allocated, with the read page of the buffer.
4531 * The page allocated, or ERR_PTR
4533 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4535 struct ring_buffer_per_cpu *cpu_buffer;
4536 struct buffer_data_page *bpage = NULL;
4537 unsigned long flags;
4540 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4541 return ERR_PTR(-ENODEV);
4543 cpu_buffer = buffer->buffers[cpu];
4544 local_irq_save(flags);
4545 arch_spin_lock(&cpu_buffer->lock);
4547 if (cpu_buffer->free_page) {
4548 bpage = cpu_buffer->free_page;
4549 cpu_buffer->free_page = NULL;
4552 arch_spin_unlock(&cpu_buffer->lock);
4553 local_irq_restore(flags);
4558 page = alloc_pages_node(cpu_to_node(cpu),
4559 GFP_KERNEL | __GFP_NORETRY, 0);
4561 return ERR_PTR(-ENOMEM);
4563 bpage = page_address(page);
4566 rb_init_page(bpage);
4570 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4573 * ring_buffer_free_read_page - free an allocated read page
4574 * @buffer: the buffer the page was allocate for
4575 * @cpu: the cpu buffer the page came from
4576 * @data: the page to free
4578 * Free a page allocated from ring_buffer_alloc_read_page.
4580 void ring_buffer_free_read_page(struct ring_buffer *buffer, int cpu, void *data)
4582 struct ring_buffer_per_cpu *cpu_buffer;
4583 struct buffer_data_page *bpage = data;
4584 struct page *page = virt_to_page(bpage);
4585 unsigned long flags;
4587 if (!buffer || !buffer->buffers || !buffer->buffers[cpu])
4590 cpu_buffer = buffer->buffers[cpu];
4592 /* If the page is still in use someplace else, we can't reuse it */
4593 if (page_ref_count(page) > 1)
4596 local_irq_save(flags);
4597 arch_spin_lock(&cpu_buffer->lock);
4599 if (!cpu_buffer->free_page) {
4600 cpu_buffer->free_page = bpage;
4604 arch_spin_unlock(&cpu_buffer->lock);
4605 local_irq_restore(flags);
4608 free_page((unsigned long)bpage);
4610 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4613 * ring_buffer_read_page - extract a page from the ring buffer
4614 * @buffer: buffer to extract from
4615 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4616 * @len: amount to extract
4617 * @cpu: the cpu of the buffer to extract
4618 * @full: should the extraction only happen when the page is full.
4620 * This function will pull out a page from the ring buffer and consume it.
4621 * @data_page must be the address of the variable that was returned
4622 * from ring_buffer_alloc_read_page. This is because the page might be used
4623 * to swap with a page in the ring buffer.
4626 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4627 * if (IS_ERR(rpage))
4628 * return PTR_ERR(rpage);
4629 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4631 * process_page(rpage, ret);
4633 * When @full is set, the function will not return true unless
4634 * the writer is off the reader page.
4636 * Note: it is up to the calling functions to handle sleeps and wakeups.
4637 * The ring buffer can be used anywhere in the kernel and can not
4638 * blindly call wake_up. The layer that uses the ring buffer must be
4639 * responsible for that.
4642 * >=0 if data has been transferred, returns the offset of consumed data.
4643 * <0 if no data has been transferred.
4645 int ring_buffer_read_page(struct ring_buffer *buffer,
4646 void **data_page, size_t len, int cpu, int full)
4648 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4649 struct ring_buffer_event *event;
4650 struct buffer_data_page *bpage;
4651 struct buffer_page *reader;
4652 unsigned long missed_events;
4653 unsigned long flags;
4654 unsigned int commit;
4659 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4663 * If len is not big enough to hold the page header, then
4664 * we can not copy anything.
4666 if (len <= BUF_PAGE_HDR_SIZE)
4669 len -= BUF_PAGE_HDR_SIZE;
4678 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4680 reader = rb_get_reader_page(cpu_buffer);
4684 event = rb_reader_event(cpu_buffer);
4686 read = reader->read;
4687 commit = rb_page_commit(reader);
4689 /* Check if any events were dropped */
4690 missed_events = cpu_buffer->lost_events;
4693 * If this page has been partially read or
4694 * if len is not big enough to read the rest of the page or
4695 * a writer is still on the page, then
4696 * we must copy the data from the page to the buffer.
4697 * Otherwise, we can simply swap the page with the one passed in.
4699 if (read || (len < (commit - read)) ||
4700 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4701 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4702 unsigned int rpos = read;
4703 unsigned int pos = 0;
4707 * If a full page is expected, this can still be returned
4708 * if there's been a previous partial read and the
4709 * rest of the page can be read and the commit page is off
4713 (!read || (len < (commit - read)) ||
4714 cpu_buffer->reader_page == cpu_buffer->commit_page))
4717 if (len > (commit - read))
4718 len = (commit - read);
4720 /* Always keep the time extend and data together */
4721 size = rb_event_ts_length(event);
4726 /* save the current timestamp, since the user will need it */
4727 save_timestamp = cpu_buffer->read_stamp;
4729 /* Need to copy one event at a time */
4731 /* We need the size of one event, because
4732 * rb_advance_reader only advances by one event,
4733 * whereas rb_event_ts_length may include the size of
4734 * one or two events.
4735 * We have already ensured there's enough space if this
4736 * is a time extend. */
4737 size = rb_event_length(event);
4738 memcpy(bpage->data + pos, rpage->data + rpos, size);
4742 rb_advance_reader(cpu_buffer);
4743 rpos = reader->read;
4749 event = rb_reader_event(cpu_buffer);
4750 /* Always keep the time extend and data together */
4751 size = rb_event_ts_length(event);
4752 } while (len >= size);
4755 local_set(&bpage->commit, pos);
4756 bpage->time_stamp = save_timestamp;
4758 /* we copied everything to the beginning */
4761 /* update the entry counter */
4762 cpu_buffer->read += rb_page_entries(reader);
4763 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4765 /* swap the pages */
4766 rb_init_page(bpage);
4767 bpage = reader->page;
4768 reader->page = *data_page;
4769 local_set(&reader->write, 0);
4770 local_set(&reader->entries, 0);
4775 * Use the real_end for the data size,
4776 * This gives us a chance to store the lost events
4779 if (reader->real_end)
4780 local_set(&bpage->commit, reader->real_end);
4784 cpu_buffer->lost_events = 0;
4786 commit = local_read(&bpage->commit);
4788 * Set a flag in the commit field if we lost events
4790 if (missed_events) {
4791 /* If there is room at the end of the page to save the
4792 * missed events, then record it there.
4794 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4795 memcpy(&bpage->data[commit], &missed_events,
4796 sizeof(missed_events));
4797 local_add(RB_MISSED_STORED, &bpage->commit);
4798 commit += sizeof(missed_events);
4800 local_add(RB_MISSED_EVENTS, &bpage->commit);
4804 * This page may be off to user land. Zero it out here.
4806 if (commit < BUF_PAGE_SIZE)
4807 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4810 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4815 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4818 * We only allocate new buffers, never free them if the CPU goes down.
4819 * If we were to free the buffer, then the user would lose any trace that was in
4822 int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
4824 struct ring_buffer *buffer;
4827 unsigned long nr_pages;
4829 buffer = container_of(node, struct ring_buffer, node);
4830 if (cpumask_test_cpu(cpu, buffer->cpumask))
4835 /* check if all cpu sizes are same */
4836 for_each_buffer_cpu(buffer, cpu_i) {
4837 /* fill in the size from first enabled cpu */
4839 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4840 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4845 /* allocate minimum pages, user can later expand it */
4848 buffer->buffers[cpu] =
4849 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4850 if (!buffer->buffers[cpu]) {
4851 WARN(1, "failed to allocate ring buffer on CPU %u\n",
4856 cpumask_set_cpu(cpu, buffer->cpumask);
4860 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4862 * This is a basic integrity check of the ring buffer.
4863 * Late in the boot cycle this test will run when configured in.
4864 * It will kick off a thread per CPU that will go into a loop
4865 * writing to the per cpu ring buffer various sizes of data.
4866 * Some of the data will be large items, some small.
4868 * Another thread is created that goes into a spin, sending out
4869 * IPIs to the other CPUs to also write into the ring buffer.
4870 * this is to test the nesting ability of the buffer.
4872 * Basic stats are recorded and reported. If something in the
4873 * ring buffer should happen that's not expected, a big warning
4874 * is displayed and all ring buffers are disabled.
4876 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4878 struct rb_test_data {
4879 struct ring_buffer *buffer;
4880 unsigned long events;
4881 unsigned long bytes_written;
4882 unsigned long bytes_alloc;
4883 unsigned long bytes_dropped;
4884 unsigned long events_nested;
4885 unsigned long bytes_written_nested;
4886 unsigned long bytes_alloc_nested;
4887 unsigned long bytes_dropped_nested;
4888 int min_size_nested;
4889 int max_size_nested;
4896 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4899 #define RB_TEST_BUFFER_SIZE 1048576
4901 static char rb_string[] __initdata =
4902 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4903 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4904 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4906 static bool rb_test_started __initdata;
4913 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4915 struct ring_buffer_event *event;
4916 struct rb_item *item;
4923 /* Have nested writes different that what is written */
4924 cnt = data->cnt + (nested ? 27 : 0);
4926 /* Multiply cnt by ~e, to make some unique increment */
4927 size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4929 len = size + sizeof(struct rb_item);
4931 started = rb_test_started;
4932 /* read rb_test_started before checking buffer enabled */
4935 event = ring_buffer_lock_reserve(data->buffer, len);
4937 /* Ignore dropped events before test starts. */
4940 data->bytes_dropped += len;
4942 data->bytes_dropped_nested += len;
4947 event_len = ring_buffer_event_length(event);
4949 if (RB_WARN_ON(data->buffer, event_len < len))
4952 item = ring_buffer_event_data(event);
4954 memcpy(item->str, rb_string, size);
4957 data->bytes_alloc_nested += event_len;
4958 data->bytes_written_nested += len;
4959 data->events_nested++;
4960 if (!data->min_size_nested || len < data->min_size_nested)
4961 data->min_size_nested = len;
4962 if (len > data->max_size_nested)
4963 data->max_size_nested = len;
4965 data->bytes_alloc += event_len;
4966 data->bytes_written += len;
4968 if (!data->min_size || len < data->min_size)
4969 data->max_size = len;
4970 if (len > data->max_size)
4971 data->max_size = len;
4975 ring_buffer_unlock_commit(data->buffer, event);
4980 static __init int rb_test(void *arg)
4982 struct rb_test_data *data = arg;
4984 while (!kthread_should_stop()) {
4985 rb_write_something(data, false);
4988 set_current_state(TASK_INTERRUPTIBLE);
4989 /* Now sleep between a min of 100-300us and a max of 1ms */
4990 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4996 static __init void rb_ipi(void *ignore)
4998 struct rb_test_data *data;
4999 int cpu = smp_processor_id();
5001 data = &rb_data[cpu];
5002 rb_write_something(data, true);
5005 static __init int rb_hammer_test(void *arg)
5007 while (!kthread_should_stop()) {
5009 /* Send an IPI to all cpus to write data! */
5010 smp_call_function(rb_ipi, NULL, 1);
5011 /* No sleep, but for non preempt, let others run */
5018 static __init int test_ringbuffer(void)
5020 struct task_struct *rb_hammer;
5021 struct ring_buffer *buffer;
5025 pr_info("Running ring buffer tests...\n");
5027 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
5028 if (WARN_ON(!buffer))
5031 /* Disable buffer so that threads can't write to it yet */
5032 ring_buffer_record_off(buffer);
5034 for_each_online_cpu(cpu) {
5035 rb_data[cpu].buffer = buffer;
5036 rb_data[cpu].cpu = cpu;
5037 rb_data[cpu].cnt = cpu;
5038 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
5039 "rbtester/%d", cpu);
5040 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
5041 pr_cont("FAILED\n");
5042 ret = PTR_ERR(rb_threads[cpu]);
5046 kthread_bind(rb_threads[cpu], cpu);
5047 wake_up_process(rb_threads[cpu]);
5050 /* Now create the rb hammer! */
5051 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
5052 if (WARN_ON(IS_ERR(rb_hammer))) {
5053 pr_cont("FAILED\n");
5054 ret = PTR_ERR(rb_hammer);
5058 ring_buffer_record_on(buffer);
5060 * Show buffer is enabled before setting rb_test_started.
5061 * Yes there's a small race window where events could be
5062 * dropped and the thread wont catch it. But when a ring
5063 * buffer gets enabled, there will always be some kind of
5064 * delay before other CPUs see it. Thus, we don't care about
5065 * those dropped events. We care about events dropped after
5066 * the threads see that the buffer is active.
5069 rb_test_started = true;
5071 set_current_state(TASK_INTERRUPTIBLE);
5072 /* Just run for 10 seconds */;
5073 schedule_timeout(10 * HZ);
5075 kthread_stop(rb_hammer);
5078 for_each_online_cpu(cpu) {
5079 if (!rb_threads[cpu])
5081 kthread_stop(rb_threads[cpu]);
5084 ring_buffer_free(buffer);
5089 pr_info("finished\n");
5090 for_each_online_cpu(cpu) {
5091 struct ring_buffer_event *event;
5092 struct rb_test_data *data = &rb_data[cpu];
5093 struct rb_item *item;
5094 unsigned long total_events;
5095 unsigned long total_dropped;
5096 unsigned long total_written;
5097 unsigned long total_alloc;
5098 unsigned long total_read = 0;
5099 unsigned long total_size = 0;
5100 unsigned long total_len = 0;
5101 unsigned long total_lost = 0;
5104 int small_event_size;
5108 total_events = data->events + data->events_nested;
5109 total_written = data->bytes_written + data->bytes_written_nested;
5110 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
5111 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
5113 big_event_size = data->max_size + data->max_size_nested;
5114 small_event_size = data->min_size + data->min_size_nested;
5116 pr_info("CPU %d:\n", cpu);
5117 pr_info(" events: %ld\n", total_events);
5118 pr_info(" dropped bytes: %ld\n", total_dropped);
5119 pr_info(" alloced bytes: %ld\n", total_alloc);
5120 pr_info(" written bytes: %ld\n", total_written);
5121 pr_info(" biggest event: %d\n", big_event_size);
5122 pr_info(" smallest event: %d\n", small_event_size);
5124 if (RB_WARN_ON(buffer, total_dropped))
5129 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
5131 item = ring_buffer_event_data(event);
5132 total_len += ring_buffer_event_length(event);
5133 total_size += item->size + sizeof(struct rb_item);
5134 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
5135 pr_info("FAILED!\n");
5136 pr_info("buffer had: %.*s\n", item->size, item->str);
5137 pr_info("expected: %.*s\n", item->size, rb_string);
5138 RB_WARN_ON(buffer, 1);
5149 pr_info(" read events: %ld\n", total_read);
5150 pr_info(" lost events: %ld\n", total_lost);
5151 pr_info(" total events: %ld\n", total_lost + total_read);
5152 pr_info(" recorded len bytes: %ld\n", total_len);
5153 pr_info(" recorded size bytes: %ld\n", total_size);
5155 pr_info(" With dropped events, record len and size may not match\n"
5156 " alloced and written from above\n");
5158 if (RB_WARN_ON(buffer, total_len != total_alloc ||
5159 total_size != total_written))
5162 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5168 pr_info("Ring buffer PASSED!\n");
5170 ring_buffer_free(buffer);
5174 late_initcall(test_ringbuffer);
5175 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */