4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 #include <linux/ring_buffer.h>
7 #include <linux/trace_clock.h>
8 #include <linux/spinlock.h>
9 #include <linux/debugfs.h>
10 #include <linux/uaccess.h>
11 #include <linux/hardirq.h>
12 #include <linux/kmemcheck.h>
13 #include <linux/module.h>
14 #include <linux/percpu.h>
15 #include <linux/mutex.h>
16 #include <linux/slab.h>
17 #include <linux/init.h>
18 #include <linux/hash.h>
19 #include <linux/list.h>
20 #include <linux/cpu.h>
23 #include <asm/local.h>
27 * The ring buffer header is special. We must manually up keep it.
29 int ring_buffer_print_entry_header(struct trace_seq *s)
33 ret = trace_seq_printf(s, "# compressed entry header\n");
34 ret = trace_seq_printf(s, "\ttype_len : 5 bits\n");
35 ret = trace_seq_printf(s, "\ttime_delta : 27 bits\n");
36 ret = trace_seq_printf(s, "\tarray : 32 bits\n");
37 ret = trace_seq_printf(s, "\n");
38 ret = trace_seq_printf(s, "\tpadding : type == %d\n",
39 RINGBUF_TYPE_PADDING);
40 ret = trace_seq_printf(s, "\ttime_extend : type == %d\n",
41 RINGBUF_TYPE_TIME_EXTEND);
42 ret = trace_seq_printf(s, "\tdata max type_len == %d\n",
43 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
49 * The ring buffer is made up of a list of pages. A separate list of pages is
50 * allocated for each CPU. A writer may only write to a buffer that is
51 * associated with the CPU it is currently executing on. A reader may read
52 * from any per cpu buffer.
54 * The reader is special. For each per cpu buffer, the reader has its own
55 * reader page. When a reader has read the entire reader page, this reader
56 * page is swapped with another page in the ring buffer.
58 * Now, as long as the writer is off the reader page, the reader can do what
59 * ever it wants with that page. The writer will never write to that page
60 * again (as long as it is out of the ring buffer).
62 * Here's some silly ASCII art.
65 * |reader| RING BUFFER
67 * +------+ +---+ +---+ +---+
76 * |reader| RING BUFFER
77 * |page |------------------v
78 * +------+ +---+ +---+ +---+
87 * |reader| RING BUFFER
88 * |page |------------------v
89 * +------+ +---+ +---+ +---+
94 * +------------------------------+
98 * |buffer| RING BUFFER
99 * |page |------------------v
100 * +------+ +---+ +---+ +---+
102 * | New +---+ +---+ +---+
105 * +------------------------------+
108 * After we make this swap, the reader can hand this page off to the splice
109 * code and be done with it. It can even allocate a new page if it needs to
110 * and swap that into the ring buffer.
112 * We will be using cmpxchg soon to make all this lockless.
117 * A fast way to enable or disable all ring buffers is to
118 * call tracing_on or tracing_off. Turning off the ring buffers
119 * prevents all ring buffers from being recorded to.
120 * Turning this switch on, makes it OK to write to the
121 * ring buffer, if the ring buffer is enabled itself.
123 * There's three layers that must be on in order to write
124 * to the ring buffer.
126 * 1) This global flag must be set.
127 * 2) The ring buffer must be enabled for recording.
128 * 3) The per cpu buffer must be enabled for recording.
130 * In case of an anomaly, this global flag has a bit set that
131 * will permantly disable all ring buffers.
135 * Global flag to disable all recording to ring buffers
136 * This has two bits: ON, DISABLED
140 * 0 0 : ring buffers are off
141 * 1 0 : ring buffers are on
142 * X 1 : ring buffers are permanently disabled
146 RB_BUFFERS_ON_BIT = 0,
147 RB_BUFFERS_DISABLED_BIT = 1,
151 RB_BUFFERS_ON = 1 << RB_BUFFERS_ON_BIT,
152 RB_BUFFERS_DISABLED = 1 << RB_BUFFERS_DISABLED_BIT,
155 static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON;
157 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
160 * tracing_on - enable all tracing buffers
162 * This function enables all tracing buffers that may have been
163 * disabled with tracing_off.
165 void tracing_on(void)
167 set_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
169 EXPORT_SYMBOL_GPL(tracing_on);
172 * tracing_off - turn off all tracing buffers
174 * This function stops all tracing buffers from recording data.
175 * It does not disable any overhead the tracers themselves may
176 * be causing. This function simply causes all recording to
177 * the ring buffers to fail.
179 void tracing_off(void)
181 clear_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
183 EXPORT_SYMBOL_GPL(tracing_off);
186 * tracing_off_permanent - permanently disable ring buffers
188 * This function, once called, will disable all ring buffers
191 void tracing_off_permanent(void)
193 set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags);
197 * tracing_is_on - show state of ring buffers enabled
199 int tracing_is_on(void)
201 return ring_buffer_flags == RB_BUFFERS_ON;
203 EXPORT_SYMBOL_GPL(tracing_is_on);
205 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
206 #define RB_ALIGNMENT 4U
207 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
208 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
210 #if !defined(CONFIG_64BIT) || defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
211 # define RB_FORCE_8BYTE_ALIGNMENT 0
212 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
214 # define RB_FORCE_8BYTE_ALIGNMENT 1
215 # define RB_ARCH_ALIGNMENT 8U
218 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
219 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
222 RB_LEN_TIME_EXTEND = 8,
223 RB_LEN_TIME_STAMP = 16,
226 #define skip_time_extend(event) \
227 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
229 static inline int rb_null_event(struct ring_buffer_event *event)
231 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
234 static void rb_event_set_padding(struct ring_buffer_event *event)
236 /* padding has a NULL time_delta */
237 event->type_len = RINGBUF_TYPE_PADDING;
238 event->time_delta = 0;
242 rb_event_data_length(struct ring_buffer_event *event)
247 length = event->type_len * RB_ALIGNMENT;
249 length = event->array[0];
250 return length + RB_EVNT_HDR_SIZE;
254 * Return the length of the given event. Will return
255 * the length of the time extend if the event is a
258 static inline unsigned
259 rb_event_length(struct ring_buffer_event *event)
261 switch (event->type_len) {
262 case RINGBUF_TYPE_PADDING:
263 if (rb_null_event(event))
266 return event->array[0] + RB_EVNT_HDR_SIZE;
268 case RINGBUF_TYPE_TIME_EXTEND:
269 return RB_LEN_TIME_EXTEND;
271 case RINGBUF_TYPE_TIME_STAMP:
272 return RB_LEN_TIME_STAMP;
274 case RINGBUF_TYPE_DATA:
275 return rb_event_data_length(event);
284 * Return total length of time extend and data,
285 * or just the event length for all other events.
287 static inline unsigned
288 rb_event_ts_length(struct ring_buffer_event *event)
292 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
293 /* time extends include the data event after it */
294 len = RB_LEN_TIME_EXTEND;
295 event = skip_time_extend(event);
297 return len + rb_event_length(event);
301 * ring_buffer_event_length - return the length of the event
302 * @event: the event to get the length of
304 * Returns the size of the data load of a data event.
305 * If the event is something other than a data event, it
306 * returns the size of the event itself. With the exception
307 * of a TIME EXTEND, where it still returns the size of the
308 * data load of the data event after it.
310 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
314 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
315 event = skip_time_extend(event);
317 length = rb_event_length(event);
318 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
320 length -= RB_EVNT_HDR_SIZE;
321 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
322 length -= sizeof(event->array[0]);
325 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
327 /* inline for ring buffer fast paths */
329 rb_event_data(struct ring_buffer_event *event)
331 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
332 event = skip_time_extend(event);
333 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
334 /* If length is in len field, then array[0] has the data */
336 return (void *)&event->array[0];
337 /* Otherwise length is in array[0] and array[1] has the data */
338 return (void *)&event->array[1];
342 * ring_buffer_event_data - return the data of the event
343 * @event: the event to get the data from
345 void *ring_buffer_event_data(struct ring_buffer_event *event)
347 return rb_event_data(event);
349 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
351 #define for_each_buffer_cpu(buffer, cpu) \
352 for_each_cpu(cpu, buffer->cpumask)
355 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
356 #define TS_DELTA_TEST (~TS_MASK)
358 /* Flag when events were overwritten */
359 #define RB_MISSED_EVENTS (1 << 31)
360 /* Missed count stored at end */
361 #define RB_MISSED_STORED (1 << 30)
363 struct buffer_data_page {
364 u64 time_stamp; /* page time stamp */
365 local_t commit; /* write committed index */
366 unsigned char data[]; /* data of buffer page */
370 * Note, the buffer_page list must be first. The buffer pages
371 * are allocated in cache lines, which means that each buffer
372 * page will be at the beginning of a cache line, and thus
373 * the least significant bits will be zero. We use this to
374 * add flags in the list struct pointers, to make the ring buffer
378 struct list_head list; /* list of buffer pages */
379 local_t write; /* index for next write */
380 unsigned read; /* index for next read */
381 local_t entries; /* entries on this page */
382 unsigned long real_end; /* real end of data */
383 struct buffer_data_page *page; /* Actual data page */
387 * The buffer page counters, write and entries, must be reset
388 * atomically when crossing page boundaries. To synchronize this
389 * update, two counters are inserted into the number. One is
390 * the actual counter for the write position or count on the page.
392 * The other is a counter of updaters. Before an update happens
393 * the update partition of the counter is incremented. This will
394 * allow the updater to update the counter atomically.
396 * The counter is 20 bits, and the state data is 12.
398 #define RB_WRITE_MASK 0xfffff
399 #define RB_WRITE_INTCNT (1 << 20)
401 static void rb_init_page(struct buffer_data_page *bpage)
403 local_set(&bpage->commit, 0);
407 * ring_buffer_page_len - the size of data on the page.
408 * @page: The page to read
410 * Returns the amount of data on the page, including buffer page header.
412 size_t ring_buffer_page_len(void *page)
414 return local_read(&((struct buffer_data_page *)page)->commit)
419 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
422 static void free_buffer_page(struct buffer_page *bpage)
424 free_page((unsigned long)bpage->page);
429 * We need to fit the time_stamp delta into 27 bits.
431 static inline int test_time_stamp(u64 delta)
433 if (delta & TS_DELTA_TEST)
438 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
440 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
441 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
443 int ring_buffer_print_page_header(struct trace_seq *s)
445 struct buffer_data_page field;
448 ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t"
449 "offset:0;\tsize:%u;\tsigned:%u;\n",
450 (unsigned int)sizeof(field.time_stamp),
451 (unsigned int)is_signed_type(u64));
453 ret = trace_seq_printf(s, "\tfield: local_t commit;\t"
454 "offset:%u;\tsize:%u;\tsigned:%u;\n",
455 (unsigned int)offsetof(typeof(field), commit),
456 (unsigned int)sizeof(field.commit),
457 (unsigned int)is_signed_type(long));
459 ret = trace_seq_printf(s, "\tfield: int overwrite;\t"
460 "offset:%u;\tsize:%u;\tsigned:%u;\n",
461 (unsigned int)offsetof(typeof(field), commit),
463 (unsigned int)is_signed_type(long));
465 ret = trace_seq_printf(s, "\tfield: char data;\t"
466 "offset:%u;\tsize:%u;\tsigned:%u;\n",
467 (unsigned int)offsetof(typeof(field), data),
468 (unsigned int)BUF_PAGE_SIZE,
469 (unsigned int)is_signed_type(char));
475 * head_page == tail_page && head == tail then buffer is empty.
477 struct ring_buffer_per_cpu {
479 atomic_t record_disabled;
480 struct ring_buffer *buffer;
481 raw_spinlock_t reader_lock; /* serialize readers */
482 arch_spinlock_t lock;
483 struct lock_class_key lock_key;
484 struct list_head *pages;
485 struct buffer_page *head_page; /* read from head */
486 struct buffer_page *tail_page; /* write to tail */
487 struct buffer_page *commit_page; /* committed pages */
488 struct buffer_page *reader_page;
489 unsigned long lost_events;
490 unsigned long last_overrun;
491 local_t entries_bytes;
492 local_t commit_overrun;
498 unsigned long read_bytes;
507 atomic_t record_disabled;
508 cpumask_var_t cpumask;
510 struct lock_class_key *reader_lock_key;
514 struct ring_buffer_per_cpu **buffers;
516 #ifdef CONFIG_HOTPLUG_CPU
517 struct notifier_block cpu_notify;
522 struct ring_buffer_iter {
523 struct ring_buffer_per_cpu *cpu_buffer;
525 struct buffer_page *head_page;
526 struct buffer_page *cache_reader_page;
527 unsigned long cache_read;
531 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
532 #define RB_WARN_ON(b, cond) \
534 int _____ret = unlikely(cond); \
536 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
537 struct ring_buffer_per_cpu *__b = \
539 atomic_inc(&__b->buffer->record_disabled); \
541 atomic_inc(&b->record_disabled); \
547 /* Up this if you want to test the TIME_EXTENTS and normalization */
548 #define DEBUG_SHIFT 0
550 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
552 /* shift to debug/test normalization and TIME_EXTENTS */
553 return buffer->clock() << DEBUG_SHIFT;
556 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
560 preempt_disable_notrace();
561 time = rb_time_stamp(buffer);
562 preempt_enable_no_resched_notrace();
566 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
568 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
571 /* Just stupid testing the normalize function and deltas */
574 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
577 * Making the ring buffer lockless makes things tricky.
578 * Although writes only happen on the CPU that they are on,
579 * and they only need to worry about interrupts. Reads can
582 * The reader page is always off the ring buffer, but when the
583 * reader finishes with a page, it needs to swap its page with
584 * a new one from the buffer. The reader needs to take from
585 * the head (writes go to the tail). But if a writer is in overwrite
586 * mode and wraps, it must push the head page forward.
588 * Here lies the problem.
590 * The reader must be careful to replace only the head page, and
591 * not another one. As described at the top of the file in the
592 * ASCII art, the reader sets its old page to point to the next
593 * page after head. It then sets the page after head to point to
594 * the old reader page. But if the writer moves the head page
595 * during this operation, the reader could end up with the tail.
597 * We use cmpxchg to help prevent this race. We also do something
598 * special with the page before head. We set the LSB to 1.
600 * When the writer must push the page forward, it will clear the
601 * bit that points to the head page, move the head, and then set
602 * the bit that points to the new head page.
604 * We also don't want an interrupt coming in and moving the head
605 * page on another writer. Thus we use the second LSB to catch
608 * head->list->prev->next bit 1 bit 0
611 * Points to head page 0 1
614 * Note we can not trust the prev pointer of the head page, because:
616 * +----+ +-----+ +-----+
617 * | |------>| T |---X--->| N |
619 * +----+ +-----+ +-----+
622 * +----------| R |----------+ |
626 * Key: ---X--> HEAD flag set in pointer
631 * (see __rb_reserve_next() to see where this happens)
633 * What the above shows is that the reader just swapped out
634 * the reader page with a page in the buffer, but before it
635 * could make the new header point back to the new page added
636 * it was preempted by a writer. The writer moved forward onto
637 * the new page added by the reader and is about to move forward
640 * You can see, it is legitimate for the previous pointer of
641 * the head (or any page) not to point back to itself. But only
645 #define RB_PAGE_NORMAL 0UL
646 #define RB_PAGE_HEAD 1UL
647 #define RB_PAGE_UPDATE 2UL
650 #define RB_FLAG_MASK 3UL
652 /* PAGE_MOVED is not part of the mask */
653 #define RB_PAGE_MOVED 4UL
656 * rb_list_head - remove any bit
658 static struct list_head *rb_list_head(struct list_head *list)
660 unsigned long val = (unsigned long)list;
662 return (struct list_head *)(val & ~RB_FLAG_MASK);
666 * rb_is_head_page - test if the given page is the head page
668 * Because the reader may move the head_page pointer, we can
669 * not trust what the head page is (it may be pointing to
670 * the reader page). But if the next page is a header page,
671 * its flags will be non zero.
674 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
675 struct buffer_page *page, struct list_head *list)
679 val = (unsigned long)list->next;
681 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
682 return RB_PAGE_MOVED;
684 return val & RB_FLAG_MASK;
690 * The unique thing about the reader page, is that, if the
691 * writer is ever on it, the previous pointer never points
692 * back to the reader page.
694 static int rb_is_reader_page(struct buffer_page *page)
696 struct list_head *list = page->list.prev;
698 return rb_list_head(list->next) != &page->list;
702 * rb_set_list_to_head - set a list_head to be pointing to head.
704 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
705 struct list_head *list)
709 ptr = (unsigned long *)&list->next;
710 *ptr |= RB_PAGE_HEAD;
711 *ptr &= ~RB_PAGE_UPDATE;
715 * rb_head_page_activate - sets up head page
717 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
719 struct buffer_page *head;
721 head = cpu_buffer->head_page;
726 * Set the previous list pointer to have the HEAD flag.
728 rb_set_list_to_head(cpu_buffer, head->list.prev);
731 static void rb_list_head_clear(struct list_head *list)
733 unsigned long *ptr = (unsigned long *)&list->next;
735 *ptr &= ~RB_FLAG_MASK;
739 * rb_head_page_dactivate - clears head page ptr (for free list)
742 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
744 struct list_head *hd;
746 /* Go through the whole list and clear any pointers found. */
747 rb_list_head_clear(cpu_buffer->pages);
749 list_for_each(hd, cpu_buffer->pages)
750 rb_list_head_clear(hd);
753 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
754 struct buffer_page *head,
755 struct buffer_page *prev,
756 int old_flag, int new_flag)
758 struct list_head *list;
759 unsigned long val = (unsigned long)&head->list;
764 val &= ~RB_FLAG_MASK;
766 ret = cmpxchg((unsigned long *)&list->next,
767 val | old_flag, val | new_flag);
769 /* check if the reader took the page */
770 if ((ret & ~RB_FLAG_MASK) != val)
771 return RB_PAGE_MOVED;
773 return ret & RB_FLAG_MASK;
776 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
777 struct buffer_page *head,
778 struct buffer_page *prev,
781 return rb_head_page_set(cpu_buffer, head, prev,
782 old_flag, RB_PAGE_UPDATE);
785 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
786 struct buffer_page *head,
787 struct buffer_page *prev,
790 return rb_head_page_set(cpu_buffer, head, prev,
791 old_flag, RB_PAGE_HEAD);
794 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
795 struct buffer_page *head,
796 struct buffer_page *prev,
799 return rb_head_page_set(cpu_buffer, head, prev,
800 old_flag, RB_PAGE_NORMAL);
803 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
804 struct buffer_page **bpage)
806 struct list_head *p = rb_list_head((*bpage)->list.next);
808 *bpage = list_entry(p, struct buffer_page, list);
811 static struct buffer_page *
812 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
814 struct buffer_page *head;
815 struct buffer_page *page;
816 struct list_head *list;
819 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
823 list = cpu_buffer->pages;
824 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
827 page = head = cpu_buffer->head_page;
829 * It is possible that the writer moves the header behind
830 * where we started, and we miss in one loop.
831 * A second loop should grab the header, but we'll do
832 * three loops just because I'm paranoid.
834 for (i = 0; i < 3; i++) {
836 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
837 cpu_buffer->head_page = page;
840 rb_inc_page(cpu_buffer, &page);
841 } while (page != head);
844 RB_WARN_ON(cpu_buffer, 1);
849 static int rb_head_page_replace(struct buffer_page *old,
850 struct buffer_page *new)
852 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
856 val = *ptr & ~RB_FLAG_MASK;
859 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
865 * rb_tail_page_update - move the tail page forward
867 * Returns 1 if moved tail page, 0 if someone else did.
869 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
870 struct buffer_page *tail_page,
871 struct buffer_page *next_page)
873 struct buffer_page *old_tail;
874 unsigned long old_entries;
875 unsigned long old_write;
879 * The tail page now needs to be moved forward.
881 * We need to reset the tail page, but without messing
882 * with possible erasing of data brought in by interrupts
883 * that have moved the tail page and are currently on it.
885 * We add a counter to the write field to denote this.
887 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
888 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
891 * Just make sure we have seen our old_write and synchronize
892 * with any interrupts that come in.
897 * If the tail page is still the same as what we think
898 * it is, then it is up to us to update the tail
901 if (tail_page == cpu_buffer->tail_page) {
902 /* Zero the write counter */
903 unsigned long val = old_write & ~RB_WRITE_MASK;
904 unsigned long eval = old_entries & ~RB_WRITE_MASK;
907 * This will only succeed if an interrupt did
908 * not come in and change it. In which case, we
909 * do not want to modify it.
911 * We add (void) to let the compiler know that we do not care
912 * about the return value of these functions. We use the
913 * cmpxchg to only update if an interrupt did not already
914 * do it for us. If the cmpxchg fails, we don't care.
916 (void)local_cmpxchg(&next_page->write, old_write, val);
917 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
920 * No need to worry about races with clearing out the commit.
921 * it only can increment when a commit takes place. But that
922 * only happens in the outer most nested commit.
924 local_set(&next_page->page->commit, 0);
926 old_tail = cmpxchg(&cpu_buffer->tail_page,
927 tail_page, next_page);
929 if (old_tail == tail_page)
936 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
937 struct buffer_page *bpage)
939 unsigned long val = (unsigned long)bpage;
941 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
948 * rb_check_list - make sure a pointer to a list has the last bits zero
950 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
951 struct list_head *list)
953 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
955 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
961 * check_pages - integrity check of buffer pages
962 * @cpu_buffer: CPU buffer with pages to test
964 * As a safety measure we check to make sure the data pages have not
967 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
969 struct list_head *head = cpu_buffer->pages;
970 struct buffer_page *bpage, *tmp;
972 rb_head_page_deactivate(cpu_buffer);
974 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
976 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
979 if (rb_check_list(cpu_buffer, head))
982 list_for_each_entry_safe(bpage, tmp, head, list) {
983 if (RB_WARN_ON(cpu_buffer,
984 bpage->list.next->prev != &bpage->list))
986 if (RB_WARN_ON(cpu_buffer,
987 bpage->list.prev->next != &bpage->list))
989 if (rb_check_list(cpu_buffer, &bpage->list))
993 rb_head_page_activate(cpu_buffer);
998 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1001 struct buffer_page *bpage, *tmp;
1007 for (i = 0; i < nr_pages; i++) {
1010 * __GFP_NORETRY flag makes sure that the allocation fails
1011 * gracefully without invoking oom-killer and the system is
1014 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1015 GFP_KERNEL | __GFP_NORETRY,
1016 cpu_to_node(cpu_buffer->cpu));
1020 rb_check_bpage(cpu_buffer, bpage);
1022 list_add(&bpage->list, &pages);
1024 page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu),
1025 GFP_KERNEL | __GFP_NORETRY, 0);
1028 bpage->page = page_address(page);
1029 rb_init_page(bpage->page);
1033 * The ring buffer page list is a circular list that does not
1034 * start and end with a list head. All page list items point to
1037 cpu_buffer->pages = pages.next;
1040 rb_check_pages(cpu_buffer);
1045 list_for_each_entry_safe(bpage, tmp, &pages, list) {
1046 list_del_init(&bpage->list);
1047 free_buffer_page(bpage);
1052 static struct ring_buffer_per_cpu *
1053 rb_allocate_cpu_buffer(struct ring_buffer *buffer, int cpu)
1055 struct ring_buffer_per_cpu *cpu_buffer;
1056 struct buffer_page *bpage;
1060 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1061 GFP_KERNEL, cpu_to_node(cpu));
1065 cpu_buffer->cpu = cpu;
1066 cpu_buffer->buffer = buffer;
1067 raw_spin_lock_init(&cpu_buffer->reader_lock);
1068 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1069 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1071 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1072 GFP_KERNEL, cpu_to_node(cpu));
1074 goto fail_free_buffer;
1076 rb_check_bpage(cpu_buffer, bpage);
1078 cpu_buffer->reader_page = bpage;
1079 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1081 goto fail_free_reader;
1082 bpage->page = page_address(page);
1083 rb_init_page(bpage->page);
1085 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1087 ret = rb_allocate_pages(cpu_buffer, buffer->pages);
1089 goto fail_free_reader;
1091 cpu_buffer->head_page
1092 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1093 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1095 rb_head_page_activate(cpu_buffer);
1100 free_buffer_page(cpu_buffer->reader_page);
1107 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1109 struct list_head *head = cpu_buffer->pages;
1110 struct buffer_page *bpage, *tmp;
1112 free_buffer_page(cpu_buffer->reader_page);
1114 rb_head_page_deactivate(cpu_buffer);
1117 list_for_each_entry_safe(bpage, tmp, head, list) {
1118 list_del_init(&bpage->list);
1119 free_buffer_page(bpage);
1121 bpage = list_entry(head, struct buffer_page, list);
1122 free_buffer_page(bpage);
1128 #ifdef CONFIG_HOTPLUG_CPU
1129 static int rb_cpu_notify(struct notifier_block *self,
1130 unsigned long action, void *hcpu);
1134 * ring_buffer_alloc - allocate a new ring_buffer
1135 * @size: the size in bytes per cpu that is needed.
1136 * @flags: attributes to set for the ring buffer.
1138 * Currently the only flag that is available is the RB_FL_OVERWRITE
1139 * flag. This flag means that the buffer will overwrite old data
1140 * when the buffer wraps. If this flag is not set, the buffer will
1141 * drop data when the tail hits the head.
1143 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1144 struct lock_class_key *key)
1146 struct ring_buffer *buffer;
1150 /* keep it in its own cache line */
1151 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1156 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1157 goto fail_free_buffer;
1159 buffer->pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1160 buffer->flags = flags;
1161 buffer->clock = trace_clock_local;
1162 buffer->reader_lock_key = key;
1164 /* need at least two pages */
1165 if (buffer->pages < 2)
1169 * In case of non-hotplug cpu, if the ring-buffer is allocated
1170 * in early initcall, it will not be notified of secondary cpus.
1171 * In that off case, we need to allocate for all possible cpus.
1173 #ifdef CONFIG_HOTPLUG_CPU
1175 cpumask_copy(buffer->cpumask, cpu_online_mask);
1177 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1179 buffer->cpus = nr_cpu_ids;
1181 bsize = sizeof(void *) * nr_cpu_ids;
1182 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1184 if (!buffer->buffers)
1185 goto fail_free_cpumask;
1187 for_each_buffer_cpu(buffer, cpu) {
1188 buffer->buffers[cpu] =
1189 rb_allocate_cpu_buffer(buffer, cpu);
1190 if (!buffer->buffers[cpu])
1191 goto fail_free_buffers;
1194 #ifdef CONFIG_HOTPLUG_CPU
1195 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1196 buffer->cpu_notify.priority = 0;
1197 register_cpu_notifier(&buffer->cpu_notify);
1201 mutex_init(&buffer->mutex);
1206 for_each_buffer_cpu(buffer, cpu) {
1207 if (buffer->buffers[cpu])
1208 rb_free_cpu_buffer(buffer->buffers[cpu]);
1210 kfree(buffer->buffers);
1213 free_cpumask_var(buffer->cpumask);
1220 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1223 * ring_buffer_free - free a ring buffer.
1224 * @buffer: the buffer to free.
1227 ring_buffer_free(struct ring_buffer *buffer)
1233 #ifdef CONFIG_HOTPLUG_CPU
1234 unregister_cpu_notifier(&buffer->cpu_notify);
1237 for_each_buffer_cpu(buffer, cpu)
1238 rb_free_cpu_buffer(buffer->buffers[cpu]);
1242 kfree(buffer->buffers);
1243 free_cpumask_var(buffer->cpumask);
1247 EXPORT_SYMBOL_GPL(ring_buffer_free);
1249 void ring_buffer_set_clock(struct ring_buffer *buffer,
1252 buffer->clock = clock;
1255 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1258 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned nr_pages)
1260 struct buffer_page *bpage;
1261 struct list_head *p;
1264 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1265 rb_head_page_deactivate(cpu_buffer);
1267 for (i = 0; i < nr_pages; i++) {
1268 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1270 p = cpu_buffer->pages->next;
1271 bpage = list_entry(p, struct buffer_page, list);
1272 list_del_init(&bpage->list);
1273 free_buffer_page(bpage);
1275 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1278 rb_reset_cpu(cpu_buffer);
1279 rb_check_pages(cpu_buffer);
1282 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1286 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer,
1287 struct list_head *pages, unsigned nr_pages)
1289 struct buffer_page *bpage;
1290 struct list_head *p;
1293 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1294 rb_head_page_deactivate(cpu_buffer);
1296 for (i = 0; i < nr_pages; i++) {
1297 if (RB_WARN_ON(cpu_buffer, list_empty(pages)))
1300 bpage = list_entry(p, struct buffer_page, list);
1301 list_del_init(&bpage->list);
1302 list_add_tail(&bpage->list, cpu_buffer->pages);
1304 rb_reset_cpu(cpu_buffer);
1305 rb_check_pages(cpu_buffer);
1308 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1312 * ring_buffer_resize - resize the ring buffer
1313 * @buffer: the buffer to resize.
1314 * @size: the new size.
1316 * Minimum size is 2 * BUF_PAGE_SIZE.
1318 * Returns -1 on failure.
1320 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size)
1322 struct ring_buffer_per_cpu *cpu_buffer;
1323 unsigned nr_pages, rm_pages, new_pages;
1324 struct buffer_page *bpage, *tmp;
1325 unsigned long buffer_size;
1330 * Always succeed at resizing a non-existent buffer:
1335 size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1336 size *= BUF_PAGE_SIZE;
1337 buffer_size = buffer->pages * BUF_PAGE_SIZE;
1339 /* we need a minimum of two pages */
1340 if (size < BUF_PAGE_SIZE * 2)
1341 size = BUF_PAGE_SIZE * 2;
1343 if (size == buffer_size)
1346 atomic_inc(&buffer->record_disabled);
1348 /* Make sure all writers are done with this buffer. */
1349 synchronize_sched();
1351 mutex_lock(&buffer->mutex);
1354 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1356 if (size < buffer_size) {
1358 /* easy case, just free pages */
1359 if (RB_WARN_ON(buffer, nr_pages >= buffer->pages))
1362 rm_pages = buffer->pages - nr_pages;
1364 for_each_buffer_cpu(buffer, cpu) {
1365 cpu_buffer = buffer->buffers[cpu];
1366 rb_remove_pages(cpu_buffer, rm_pages);
1372 * This is a bit more difficult. We only want to add pages
1373 * when we can allocate enough for all CPUs. We do this
1374 * by allocating all the pages and storing them on a local
1375 * link list. If we succeed in our allocation, then we
1376 * add these pages to the cpu_buffers. Otherwise we just free
1377 * them all and return -ENOMEM;
1379 if (RB_WARN_ON(buffer, nr_pages <= buffer->pages))
1382 new_pages = nr_pages - buffer->pages;
1384 for_each_buffer_cpu(buffer, cpu) {
1385 for (i = 0; i < new_pages; i++) {
1388 * __GFP_NORETRY flag makes sure that the allocation
1389 * fails gracefully without invoking oom-killer and
1390 * the system is not destabilized.
1392 bpage = kzalloc_node(ALIGN(sizeof(*bpage),
1394 GFP_KERNEL | __GFP_NORETRY,
1398 list_add(&bpage->list, &pages);
1399 page = alloc_pages_node(cpu_to_node(cpu),
1400 GFP_KERNEL | __GFP_NORETRY, 0);
1403 bpage->page = page_address(page);
1404 rb_init_page(bpage->page);
1408 for_each_buffer_cpu(buffer, cpu) {
1409 cpu_buffer = buffer->buffers[cpu];
1410 rb_insert_pages(cpu_buffer, &pages, new_pages);
1413 if (RB_WARN_ON(buffer, !list_empty(&pages)))
1417 buffer->pages = nr_pages;
1419 mutex_unlock(&buffer->mutex);
1421 atomic_dec(&buffer->record_disabled);
1426 list_for_each_entry_safe(bpage, tmp, &pages, list) {
1427 list_del_init(&bpage->list);
1428 free_buffer_page(bpage);
1431 mutex_unlock(&buffer->mutex);
1432 atomic_dec(&buffer->record_disabled);
1436 * Something went totally wrong, and we are too paranoid
1437 * to even clean up the mess.
1441 mutex_unlock(&buffer->mutex);
1442 atomic_dec(&buffer->record_disabled);
1445 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1447 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1449 mutex_lock(&buffer->mutex);
1451 buffer->flags |= RB_FL_OVERWRITE;
1453 buffer->flags &= ~RB_FL_OVERWRITE;
1454 mutex_unlock(&buffer->mutex);
1456 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1458 static inline void *
1459 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1461 return bpage->data + index;
1464 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1466 return bpage->page->data + index;
1469 static inline struct ring_buffer_event *
1470 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1472 return __rb_page_index(cpu_buffer->reader_page,
1473 cpu_buffer->reader_page->read);
1476 static inline struct ring_buffer_event *
1477 rb_iter_head_event(struct ring_buffer_iter *iter)
1479 return __rb_page_index(iter->head_page, iter->head);
1482 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1484 return local_read(&bpage->write) & RB_WRITE_MASK;
1487 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1489 return local_read(&bpage->page->commit);
1492 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1494 return local_read(&bpage->entries) & RB_WRITE_MASK;
1497 /* Size is determined by what has been committed */
1498 static inline unsigned rb_page_size(struct buffer_page *bpage)
1500 return rb_page_commit(bpage);
1503 static inline unsigned
1504 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1506 return rb_page_commit(cpu_buffer->commit_page);
1509 static inline unsigned
1510 rb_event_index(struct ring_buffer_event *event)
1512 unsigned long addr = (unsigned long)event;
1514 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1518 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1519 struct ring_buffer_event *event)
1521 unsigned long addr = (unsigned long)event;
1522 unsigned long index;
1524 index = rb_event_index(event);
1527 return cpu_buffer->commit_page->page == (void *)addr &&
1528 rb_commit_index(cpu_buffer) == index;
1532 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1534 unsigned long max_count;
1537 * We only race with interrupts and NMIs on this CPU.
1538 * If we own the commit event, then we can commit
1539 * all others that interrupted us, since the interruptions
1540 * are in stack format (they finish before they come
1541 * back to us). This allows us to do a simple loop to
1542 * assign the commit to the tail.
1545 max_count = cpu_buffer->buffer->pages * 100;
1547 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1548 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1550 if (RB_WARN_ON(cpu_buffer,
1551 rb_is_reader_page(cpu_buffer->tail_page)))
1553 local_set(&cpu_buffer->commit_page->page->commit,
1554 rb_page_write(cpu_buffer->commit_page));
1555 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1556 cpu_buffer->write_stamp =
1557 cpu_buffer->commit_page->page->time_stamp;
1558 /* add barrier to keep gcc from optimizing too much */
1561 while (rb_commit_index(cpu_buffer) !=
1562 rb_page_write(cpu_buffer->commit_page)) {
1564 local_set(&cpu_buffer->commit_page->page->commit,
1565 rb_page_write(cpu_buffer->commit_page));
1566 RB_WARN_ON(cpu_buffer,
1567 local_read(&cpu_buffer->commit_page->page->commit) &
1572 /* again, keep gcc from optimizing */
1576 * If an interrupt came in just after the first while loop
1577 * and pushed the tail page forward, we will be left with
1578 * a dangling commit that will never go forward.
1580 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1584 static void rb_inc_iter(struct ring_buffer_iter *iter)
1586 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1589 * The iterator could be on the reader page (it starts there).
1590 * But the head could have moved, since the reader was
1591 * found. Check for this case and assign the iterator
1592 * to the head page instead of next.
1594 if (iter->head_page == cpu_buffer->reader_page)
1595 iter->head_page = rb_set_head_page(cpu_buffer);
1597 rb_inc_page(cpu_buffer, &iter->head_page);
1599 iter->read_stamp = iter->head_page->page->time_stamp;
1603 /* Slow path, do not inline */
1604 static noinline struct ring_buffer_event *
1605 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
1607 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
1609 /* Not the first event on the page? */
1610 if (rb_event_index(event)) {
1611 event->time_delta = delta & TS_MASK;
1612 event->array[0] = delta >> TS_SHIFT;
1614 /* nope, just zero it */
1615 event->time_delta = 0;
1616 event->array[0] = 0;
1619 return skip_time_extend(event);
1623 * ring_buffer_update_event - update event type and data
1624 * @event: the even to update
1625 * @type: the type of event
1626 * @length: the size of the event field in the ring buffer
1628 * Update the type and data fields of the event. The length
1629 * is the actual size that is written to the ring buffer,
1630 * and with this, we can determine what to place into the
1634 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
1635 struct ring_buffer_event *event, unsigned length,
1636 int add_timestamp, u64 delta)
1638 /* Only a commit updates the timestamp */
1639 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
1643 * If we need to add a timestamp, then we
1644 * add it to the start of the resevered space.
1646 if (unlikely(add_timestamp)) {
1647 event = rb_add_time_stamp(event, delta);
1648 length -= RB_LEN_TIME_EXTEND;
1652 event->time_delta = delta;
1653 length -= RB_EVNT_HDR_SIZE;
1654 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
1655 event->type_len = 0;
1656 event->array[0] = length;
1658 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
1662 * rb_handle_head_page - writer hit the head page
1664 * Returns: +1 to retry page
1669 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1670 struct buffer_page *tail_page,
1671 struct buffer_page *next_page)
1673 struct buffer_page *new_head;
1678 entries = rb_page_entries(next_page);
1681 * The hard part is here. We need to move the head
1682 * forward, and protect against both readers on
1683 * other CPUs and writers coming in via interrupts.
1685 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1689 * type can be one of four:
1690 * NORMAL - an interrupt already moved it for us
1691 * HEAD - we are the first to get here.
1692 * UPDATE - we are the interrupt interrupting
1694 * MOVED - a reader on another CPU moved the next
1695 * pointer to its reader page. Give up
1702 * We changed the head to UPDATE, thus
1703 * it is our responsibility to update
1706 local_add(entries, &cpu_buffer->overrun);
1707 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1710 * The entries will be zeroed out when we move the
1714 /* still more to do */
1717 case RB_PAGE_UPDATE:
1719 * This is an interrupt that interrupt the
1720 * previous update. Still more to do.
1723 case RB_PAGE_NORMAL:
1725 * An interrupt came in before the update
1726 * and processed this for us.
1727 * Nothing left to do.
1732 * The reader is on another CPU and just did
1733 * a swap with our next_page.
1738 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1743 * Now that we are here, the old head pointer is
1744 * set to UPDATE. This will keep the reader from
1745 * swapping the head page with the reader page.
1746 * The reader (on another CPU) will spin till
1749 * We just need to protect against interrupts
1750 * doing the job. We will set the next pointer
1751 * to HEAD. After that, we set the old pointer
1752 * to NORMAL, but only if it was HEAD before.
1753 * otherwise we are an interrupt, and only
1754 * want the outer most commit to reset it.
1756 new_head = next_page;
1757 rb_inc_page(cpu_buffer, &new_head);
1759 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1763 * Valid returns are:
1764 * HEAD - an interrupt came in and already set it.
1765 * NORMAL - One of two things:
1766 * 1) We really set it.
1767 * 2) A bunch of interrupts came in and moved
1768 * the page forward again.
1772 case RB_PAGE_NORMAL:
1776 RB_WARN_ON(cpu_buffer, 1);
1781 * It is possible that an interrupt came in,
1782 * set the head up, then more interrupts came in
1783 * and moved it again. When we get back here,
1784 * the page would have been set to NORMAL but we
1785 * just set it back to HEAD.
1787 * How do you detect this? Well, if that happened
1788 * the tail page would have moved.
1790 if (ret == RB_PAGE_NORMAL) {
1792 * If the tail had moved passed next, then we need
1793 * to reset the pointer.
1795 if (cpu_buffer->tail_page != tail_page &&
1796 cpu_buffer->tail_page != next_page)
1797 rb_head_page_set_normal(cpu_buffer, new_head,
1803 * If this was the outer most commit (the one that
1804 * changed the original pointer from HEAD to UPDATE),
1805 * then it is up to us to reset it to NORMAL.
1807 if (type == RB_PAGE_HEAD) {
1808 ret = rb_head_page_set_normal(cpu_buffer, next_page,
1811 if (RB_WARN_ON(cpu_buffer,
1812 ret != RB_PAGE_UPDATE))
1819 static unsigned rb_calculate_event_length(unsigned length)
1821 struct ring_buffer_event event; /* Used only for sizeof array */
1823 /* zero length can cause confusions */
1827 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
1828 length += sizeof(event.array[0]);
1830 length += RB_EVNT_HDR_SIZE;
1831 length = ALIGN(length, RB_ARCH_ALIGNMENT);
1837 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
1838 struct buffer_page *tail_page,
1839 unsigned long tail, unsigned long length)
1841 struct ring_buffer_event *event;
1844 * Only the event that crossed the page boundary
1845 * must fill the old tail_page with padding.
1847 if (tail >= BUF_PAGE_SIZE) {
1849 * If the page was filled, then we still need
1850 * to update the real_end. Reset it to zero
1851 * and the reader will ignore it.
1853 if (tail == BUF_PAGE_SIZE)
1854 tail_page->real_end = 0;
1856 local_sub(length, &tail_page->write);
1860 event = __rb_page_index(tail_page, tail);
1861 kmemcheck_annotate_bitfield(event, bitfield);
1863 /* account for padding bytes */
1864 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
1867 * Save the original length to the meta data.
1868 * This will be used by the reader to add lost event
1871 tail_page->real_end = tail;
1874 * If this event is bigger than the minimum size, then
1875 * we need to be careful that we don't subtract the
1876 * write counter enough to allow another writer to slip
1878 * We put in a discarded commit instead, to make sure
1879 * that this space is not used again.
1881 * If we are less than the minimum size, we don't need to
1884 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
1885 /* No room for any events */
1887 /* Mark the rest of the page with padding */
1888 rb_event_set_padding(event);
1890 /* Set the write back to the previous setting */
1891 local_sub(length, &tail_page->write);
1895 /* Put in a discarded event */
1896 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
1897 event->type_len = RINGBUF_TYPE_PADDING;
1898 /* time delta must be non zero */
1899 event->time_delta = 1;
1901 /* Set write to end of buffer */
1902 length = (tail + length) - BUF_PAGE_SIZE;
1903 local_sub(length, &tail_page->write);
1907 * This is the slow path, force gcc not to inline it.
1909 static noinline struct ring_buffer_event *
1910 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
1911 unsigned long length, unsigned long tail,
1912 struct buffer_page *tail_page, u64 ts)
1914 struct buffer_page *commit_page = cpu_buffer->commit_page;
1915 struct ring_buffer *buffer = cpu_buffer->buffer;
1916 struct buffer_page *next_page;
1919 next_page = tail_page;
1921 rb_inc_page(cpu_buffer, &next_page);
1924 * If for some reason, we had an interrupt storm that made
1925 * it all the way around the buffer, bail, and warn
1928 if (unlikely(next_page == commit_page)) {
1929 local_inc(&cpu_buffer->commit_overrun);
1934 * This is where the fun begins!
1936 * We are fighting against races between a reader that
1937 * could be on another CPU trying to swap its reader
1938 * page with the buffer head.
1940 * We are also fighting against interrupts coming in and
1941 * moving the head or tail on us as well.
1943 * If the next page is the head page then we have filled
1944 * the buffer, unless the commit page is still on the
1947 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
1950 * If the commit is not on the reader page, then
1951 * move the header page.
1953 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
1955 * If we are not in overwrite mode,
1956 * this is easy, just stop here.
1958 if (!(buffer->flags & RB_FL_OVERWRITE))
1961 ret = rb_handle_head_page(cpu_buffer,
1970 * We need to be careful here too. The
1971 * commit page could still be on the reader
1972 * page. We could have a small buffer, and
1973 * have filled up the buffer with events
1974 * from interrupts and such, and wrapped.
1976 * Note, if the tail page is also the on the
1977 * reader_page, we let it move out.
1979 if (unlikely((cpu_buffer->commit_page !=
1980 cpu_buffer->tail_page) &&
1981 (cpu_buffer->commit_page ==
1982 cpu_buffer->reader_page))) {
1983 local_inc(&cpu_buffer->commit_overrun);
1989 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
1992 * Nested commits always have zero deltas, so
1993 * just reread the time stamp
1995 ts = rb_time_stamp(buffer);
1996 next_page->page->time_stamp = ts;
2001 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2003 /* fail and let the caller try again */
2004 return ERR_PTR(-EAGAIN);
2008 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2013 static struct ring_buffer_event *
2014 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2015 unsigned long length, u64 ts,
2016 u64 delta, int add_timestamp)
2018 struct buffer_page *tail_page;
2019 struct ring_buffer_event *event;
2020 unsigned long tail, write;
2023 * If the time delta since the last event is too big to
2024 * hold in the time field of the event, then we append a
2025 * TIME EXTEND event ahead of the data event.
2027 if (unlikely(add_timestamp))
2028 length += RB_LEN_TIME_EXTEND;
2030 tail_page = cpu_buffer->tail_page;
2031 write = local_add_return(length, &tail_page->write);
2033 /* set write to only the index of the write */
2034 write &= RB_WRITE_MASK;
2035 tail = write - length;
2038 * If this is the first commit on the page, then it has the same
2039 * timestamp as the page itself.
2044 /* See if we shot pass the end of this buffer page */
2045 if (unlikely(write > BUF_PAGE_SIZE))
2046 return rb_move_tail(cpu_buffer, length, tail,
2049 /* We reserved something on the buffer */
2051 event = __rb_page_index(tail_page, tail);
2052 kmemcheck_annotate_bitfield(event, bitfield);
2053 rb_update_event(cpu_buffer, event, length, add_timestamp, delta);
2055 local_inc(&tail_page->entries);
2058 * If this is the first commit on the page, then update
2062 tail_page->page->time_stamp = ts;
2064 /* account for these added bytes */
2065 local_add(length, &cpu_buffer->entries_bytes);
2071 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2072 struct ring_buffer_event *event)
2074 unsigned long new_index, old_index;
2075 struct buffer_page *bpage;
2076 unsigned long index;
2079 new_index = rb_event_index(event);
2080 old_index = new_index + rb_event_ts_length(event);
2081 addr = (unsigned long)event;
2084 bpage = cpu_buffer->tail_page;
2086 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2087 unsigned long write_mask =
2088 local_read(&bpage->write) & ~RB_WRITE_MASK;
2089 unsigned long event_length = rb_event_length(event);
2091 * This is on the tail page. It is possible that
2092 * a write could come in and move the tail page
2093 * and write to the next page. That is fine
2094 * because we just shorten what is on this page.
2096 old_index += write_mask;
2097 new_index += write_mask;
2098 index = local_cmpxchg(&bpage->write, old_index, new_index);
2099 if (index == old_index) {
2100 /* update counters */
2101 local_sub(event_length, &cpu_buffer->entries_bytes);
2106 /* could not discard */
2110 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2112 local_inc(&cpu_buffer->committing);
2113 local_inc(&cpu_buffer->commits);
2116 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2118 unsigned long commits;
2120 if (RB_WARN_ON(cpu_buffer,
2121 !local_read(&cpu_buffer->committing)))
2125 commits = local_read(&cpu_buffer->commits);
2126 /* synchronize with interrupts */
2128 if (local_read(&cpu_buffer->committing) == 1)
2129 rb_set_commit_to_write(cpu_buffer);
2131 local_dec(&cpu_buffer->committing);
2133 /* synchronize with interrupts */
2137 * Need to account for interrupts coming in between the
2138 * updating of the commit page and the clearing of the
2139 * committing counter.
2141 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2142 !local_read(&cpu_buffer->committing)) {
2143 local_inc(&cpu_buffer->committing);
2148 static struct ring_buffer_event *
2149 rb_reserve_next_event(struct ring_buffer *buffer,
2150 struct ring_buffer_per_cpu *cpu_buffer,
2151 unsigned long length)
2153 struct ring_buffer_event *event;
2159 rb_start_commit(cpu_buffer);
2161 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2163 * Due to the ability to swap a cpu buffer from a buffer
2164 * it is possible it was swapped before we committed.
2165 * (committing stops a swap). We check for it here and
2166 * if it happened, we have to fail the write.
2169 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2170 local_dec(&cpu_buffer->committing);
2171 local_dec(&cpu_buffer->commits);
2176 length = rb_calculate_event_length(length);
2182 * We allow for interrupts to reenter here and do a trace.
2183 * If one does, it will cause this original code to loop
2184 * back here. Even with heavy interrupts happening, this
2185 * should only happen a few times in a row. If this happens
2186 * 1000 times in a row, there must be either an interrupt
2187 * storm or we have something buggy.
2190 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2193 ts = rb_time_stamp(cpu_buffer->buffer);
2194 diff = ts - cpu_buffer->write_stamp;
2196 /* make sure this diff is calculated here */
2199 /* Did the write stamp get updated already? */
2200 if (likely(ts >= cpu_buffer->write_stamp)) {
2202 if (unlikely(test_time_stamp(delta))) {
2203 int local_clock_stable = 1;
2204 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2205 local_clock_stable = sched_clock_stable;
2207 WARN_ONCE(delta > (1ULL << 59),
2208 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2209 (unsigned long long)delta,
2210 (unsigned long long)ts,
2211 (unsigned long long)cpu_buffer->write_stamp,
2212 local_clock_stable ? "" :
2213 "If you just came from a suspend/resume,\n"
2214 "please switch to the trace global clock:\n"
2215 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2220 event = __rb_reserve_next(cpu_buffer, length, ts,
2221 delta, add_timestamp);
2222 if (unlikely(PTR_ERR(event) == -EAGAIN))
2231 rb_end_commit(cpu_buffer);
2235 #ifdef CONFIG_TRACING
2237 #define TRACE_RECURSIVE_DEPTH 16
2239 /* Keep this code out of the fast path cache */
2240 static noinline void trace_recursive_fail(void)
2242 /* Disable all tracing before we do anything else */
2243 tracing_off_permanent();
2245 printk_once(KERN_WARNING "Tracing recursion: depth[%ld]:"
2246 "HC[%lu]:SC[%lu]:NMI[%lu]\n",
2247 trace_recursion_buffer(),
2248 hardirq_count() >> HARDIRQ_SHIFT,
2249 softirq_count() >> SOFTIRQ_SHIFT,
2255 static inline int trace_recursive_lock(void)
2257 trace_recursion_inc();
2259 if (likely(trace_recursion_buffer() < TRACE_RECURSIVE_DEPTH))
2262 trace_recursive_fail();
2267 static inline void trace_recursive_unlock(void)
2269 WARN_ON_ONCE(!trace_recursion_buffer());
2271 trace_recursion_dec();
2276 #define trace_recursive_lock() (0)
2277 #define trace_recursive_unlock() do { } while (0)
2282 * ring_buffer_lock_reserve - reserve a part of the buffer
2283 * @buffer: the ring buffer to reserve from
2284 * @length: the length of the data to reserve (excluding event header)
2286 * Returns a reseverd event on the ring buffer to copy directly to.
2287 * The user of this interface will need to get the body to write into
2288 * and can use the ring_buffer_event_data() interface.
2290 * The length is the length of the data needed, not the event length
2291 * which also includes the event header.
2293 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2294 * If NULL is returned, then nothing has been allocated or locked.
2296 struct ring_buffer_event *
2297 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2299 struct ring_buffer_per_cpu *cpu_buffer;
2300 struct ring_buffer_event *event;
2303 if (ring_buffer_flags != RB_BUFFERS_ON)
2306 /* If we are tracing schedule, we don't want to recurse */
2307 preempt_disable_notrace();
2309 if (atomic_read(&buffer->record_disabled))
2312 if (trace_recursive_lock())
2315 cpu = raw_smp_processor_id();
2317 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2320 cpu_buffer = buffer->buffers[cpu];
2322 if (atomic_read(&cpu_buffer->record_disabled))
2325 if (length > BUF_MAX_DATA_SIZE)
2328 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2335 trace_recursive_unlock();
2338 preempt_enable_notrace();
2341 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2344 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2345 struct ring_buffer_event *event)
2350 * The event first in the commit queue updates the
2353 if (rb_event_is_commit(cpu_buffer, event)) {
2355 * A commit event that is first on a page
2356 * updates the write timestamp with the page stamp
2358 if (!rb_event_index(event))
2359 cpu_buffer->write_stamp =
2360 cpu_buffer->commit_page->page->time_stamp;
2361 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2362 delta = event->array[0];
2364 delta += event->time_delta;
2365 cpu_buffer->write_stamp += delta;
2367 cpu_buffer->write_stamp += event->time_delta;
2371 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2372 struct ring_buffer_event *event)
2374 local_inc(&cpu_buffer->entries);
2375 rb_update_write_stamp(cpu_buffer, event);
2376 rb_end_commit(cpu_buffer);
2380 * ring_buffer_unlock_commit - commit a reserved
2381 * @buffer: The buffer to commit to
2382 * @event: The event pointer to commit.
2384 * This commits the data to the ring buffer, and releases any locks held.
2386 * Must be paired with ring_buffer_lock_reserve.
2388 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2389 struct ring_buffer_event *event)
2391 struct ring_buffer_per_cpu *cpu_buffer;
2392 int cpu = raw_smp_processor_id();
2394 cpu_buffer = buffer->buffers[cpu];
2396 rb_commit(cpu_buffer, event);
2398 trace_recursive_unlock();
2400 preempt_enable_notrace();
2404 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2406 static inline void rb_event_discard(struct ring_buffer_event *event)
2408 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2409 event = skip_time_extend(event);
2411 /* array[0] holds the actual length for the discarded event */
2412 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2413 event->type_len = RINGBUF_TYPE_PADDING;
2414 /* time delta must be non zero */
2415 if (!event->time_delta)
2416 event->time_delta = 1;
2420 * Decrement the entries to the page that an event is on.
2421 * The event does not even need to exist, only the pointer
2422 * to the page it is on. This may only be called before the commit
2426 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2427 struct ring_buffer_event *event)
2429 unsigned long addr = (unsigned long)event;
2430 struct buffer_page *bpage = cpu_buffer->commit_page;
2431 struct buffer_page *start;
2435 /* Do the likely case first */
2436 if (likely(bpage->page == (void *)addr)) {
2437 local_dec(&bpage->entries);
2442 * Because the commit page may be on the reader page we
2443 * start with the next page and check the end loop there.
2445 rb_inc_page(cpu_buffer, &bpage);
2448 if (bpage->page == (void *)addr) {
2449 local_dec(&bpage->entries);
2452 rb_inc_page(cpu_buffer, &bpage);
2453 } while (bpage != start);
2455 /* commit not part of this buffer?? */
2456 RB_WARN_ON(cpu_buffer, 1);
2460 * ring_buffer_commit_discard - discard an event that has not been committed
2461 * @buffer: the ring buffer
2462 * @event: non committed event to discard
2464 * Sometimes an event that is in the ring buffer needs to be ignored.
2465 * This function lets the user discard an event in the ring buffer
2466 * and then that event will not be read later.
2468 * This function only works if it is called before the the item has been
2469 * committed. It will try to free the event from the ring buffer
2470 * if another event has not been added behind it.
2472 * If another event has been added behind it, it will set the event
2473 * up as discarded, and perform the commit.
2475 * If this function is called, do not call ring_buffer_unlock_commit on
2478 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2479 struct ring_buffer_event *event)
2481 struct ring_buffer_per_cpu *cpu_buffer;
2484 /* The event is discarded regardless */
2485 rb_event_discard(event);
2487 cpu = smp_processor_id();
2488 cpu_buffer = buffer->buffers[cpu];
2491 * This must only be called if the event has not been
2492 * committed yet. Thus we can assume that preemption
2493 * is still disabled.
2495 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2497 rb_decrement_entry(cpu_buffer, event);
2498 if (rb_try_to_discard(cpu_buffer, event))
2502 * The commit is still visible by the reader, so we
2503 * must still update the timestamp.
2505 rb_update_write_stamp(cpu_buffer, event);
2507 rb_end_commit(cpu_buffer);
2509 trace_recursive_unlock();
2511 preempt_enable_notrace();
2514 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2517 * ring_buffer_write - write data to the buffer without reserving
2518 * @buffer: The ring buffer to write to.
2519 * @length: The length of the data being written (excluding the event header)
2520 * @data: The data to write to the buffer.
2522 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2523 * one function. If you already have the data to write to the buffer, it
2524 * may be easier to simply call this function.
2526 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2527 * and not the length of the event which would hold the header.
2529 int ring_buffer_write(struct ring_buffer *buffer,
2530 unsigned long length,
2533 struct ring_buffer_per_cpu *cpu_buffer;
2534 struct ring_buffer_event *event;
2539 if (ring_buffer_flags != RB_BUFFERS_ON)
2542 preempt_disable_notrace();
2544 if (atomic_read(&buffer->record_disabled))
2547 cpu = raw_smp_processor_id();
2549 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2552 cpu_buffer = buffer->buffers[cpu];
2554 if (atomic_read(&cpu_buffer->record_disabled))
2557 if (length > BUF_MAX_DATA_SIZE)
2560 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2564 body = rb_event_data(event);
2566 memcpy(body, data, length);
2568 rb_commit(cpu_buffer, event);
2572 preempt_enable_notrace();
2576 EXPORT_SYMBOL_GPL(ring_buffer_write);
2578 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
2580 struct buffer_page *reader = cpu_buffer->reader_page;
2581 struct buffer_page *head = rb_set_head_page(cpu_buffer);
2582 struct buffer_page *commit = cpu_buffer->commit_page;
2584 /* In case of error, head will be NULL */
2585 if (unlikely(!head))
2588 return reader->read == rb_page_commit(reader) &&
2589 (commit == reader ||
2591 head->read == rb_page_commit(commit)));
2595 * ring_buffer_record_disable - stop all writes into the buffer
2596 * @buffer: The ring buffer to stop writes to.
2598 * This prevents all writes to the buffer. Any attempt to write
2599 * to the buffer after this will fail and return NULL.
2601 * The caller should call synchronize_sched() after this.
2603 void ring_buffer_record_disable(struct ring_buffer *buffer)
2605 atomic_inc(&buffer->record_disabled);
2607 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
2610 * ring_buffer_record_enable - enable writes to the buffer
2611 * @buffer: The ring buffer to enable writes
2613 * Note, multiple disables will need the same number of enables
2614 * to truly enable the writing (much like preempt_disable).
2616 void ring_buffer_record_enable(struct ring_buffer *buffer)
2618 atomic_dec(&buffer->record_disabled);
2620 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
2623 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
2624 * @buffer: The ring buffer to stop writes to.
2625 * @cpu: The CPU buffer to stop
2627 * This prevents all writes to the buffer. Any attempt to write
2628 * to the buffer after this will fail and return NULL.
2630 * The caller should call synchronize_sched() after this.
2632 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
2634 struct ring_buffer_per_cpu *cpu_buffer;
2636 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2639 cpu_buffer = buffer->buffers[cpu];
2640 atomic_inc(&cpu_buffer->record_disabled);
2642 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
2645 * ring_buffer_record_enable_cpu - enable writes to the buffer
2646 * @buffer: The ring buffer to enable writes
2647 * @cpu: The CPU to enable.
2649 * Note, multiple disables will need the same number of enables
2650 * to truly enable the writing (much like preempt_disable).
2652 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
2654 struct ring_buffer_per_cpu *cpu_buffer;
2656 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2659 cpu_buffer = buffer->buffers[cpu];
2660 atomic_dec(&cpu_buffer->record_disabled);
2662 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
2665 * The total entries in the ring buffer is the running counter
2666 * of entries entered into the ring buffer, minus the sum of
2667 * the entries read from the ring buffer and the number of
2668 * entries that were overwritten.
2670 static inline unsigned long
2671 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
2673 return local_read(&cpu_buffer->entries) -
2674 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
2678 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
2679 * @buffer: The ring buffer
2680 * @cpu: The per CPU buffer to read from.
2682 unsigned long ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
2684 unsigned long flags;
2685 struct ring_buffer_per_cpu *cpu_buffer;
2686 struct buffer_page *bpage;
2687 unsigned long ret = 0;
2689 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2692 cpu_buffer = buffer->buffers[cpu];
2693 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2695 * if the tail is on reader_page, oldest time stamp is on the reader
2698 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
2699 bpage = cpu_buffer->reader_page;
2701 bpage = rb_set_head_page(cpu_buffer);
2703 ret = bpage->page->time_stamp;
2704 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2708 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
2711 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
2712 * @buffer: The ring buffer
2713 * @cpu: The per CPU buffer to read from.
2715 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
2717 struct ring_buffer_per_cpu *cpu_buffer;
2720 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2723 cpu_buffer = buffer->buffers[cpu];
2724 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
2728 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
2731 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
2732 * @buffer: The ring buffer
2733 * @cpu: The per CPU buffer to get the entries from.
2735 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
2737 struct ring_buffer_per_cpu *cpu_buffer;
2739 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2742 cpu_buffer = buffer->buffers[cpu];
2744 return rb_num_of_entries(cpu_buffer);
2746 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
2749 * ring_buffer_overrun_cpu - get the number of overruns in a cpu_buffer
2750 * @buffer: The ring buffer
2751 * @cpu: The per CPU buffer to get the number of overruns from
2753 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
2755 struct ring_buffer_per_cpu *cpu_buffer;
2758 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2761 cpu_buffer = buffer->buffers[cpu];
2762 ret = local_read(&cpu_buffer->overrun);
2766 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
2769 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by commits
2770 * @buffer: The ring buffer
2771 * @cpu: The per CPU buffer to get the number of overruns from
2774 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
2776 struct ring_buffer_per_cpu *cpu_buffer;
2779 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2782 cpu_buffer = buffer->buffers[cpu];
2783 ret = local_read(&cpu_buffer->commit_overrun);
2787 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
2790 * ring_buffer_entries - get the number of entries in a buffer
2791 * @buffer: The ring buffer
2793 * Returns the total number of entries in the ring buffer
2796 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
2798 struct ring_buffer_per_cpu *cpu_buffer;
2799 unsigned long entries = 0;
2802 /* if you care about this being correct, lock the buffer */
2803 for_each_buffer_cpu(buffer, cpu) {
2804 cpu_buffer = buffer->buffers[cpu];
2805 entries += rb_num_of_entries(cpu_buffer);
2810 EXPORT_SYMBOL_GPL(ring_buffer_entries);
2813 * ring_buffer_overruns - get the number of overruns in buffer
2814 * @buffer: The ring buffer
2816 * Returns the total number of overruns in the ring buffer
2819 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
2821 struct ring_buffer_per_cpu *cpu_buffer;
2822 unsigned long overruns = 0;
2825 /* if you care about this being correct, lock the buffer */
2826 for_each_buffer_cpu(buffer, cpu) {
2827 cpu_buffer = buffer->buffers[cpu];
2828 overruns += local_read(&cpu_buffer->overrun);
2833 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
2835 static void rb_iter_reset(struct ring_buffer_iter *iter)
2837 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2839 /* Iterator usage is expected to have record disabled */
2840 iter->head_page = cpu_buffer->reader_page;
2841 iter->head = cpu_buffer->reader_page->read;
2843 iter->cache_reader_page = iter->head_page;
2844 iter->cache_read = cpu_buffer->read;
2847 iter->read_stamp = cpu_buffer->read_stamp;
2849 iter->read_stamp = iter->head_page->page->time_stamp;
2853 * ring_buffer_iter_reset - reset an iterator
2854 * @iter: The iterator to reset
2856 * Resets the iterator, so that it will start from the beginning
2859 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
2861 struct ring_buffer_per_cpu *cpu_buffer;
2862 unsigned long flags;
2867 cpu_buffer = iter->cpu_buffer;
2869 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2870 rb_iter_reset(iter);
2871 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2873 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
2876 * ring_buffer_iter_empty - check if an iterator has no more to read
2877 * @iter: The iterator to check
2879 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
2881 struct ring_buffer_per_cpu *cpu_buffer;
2882 struct buffer_page *reader;
2883 struct buffer_page *head_page;
2884 struct buffer_page *commit_page;
2887 cpu_buffer = iter->cpu_buffer;
2889 /* Remember, trace recording is off when iterator is in use */
2890 reader = cpu_buffer->reader_page;
2891 head_page = cpu_buffer->head_page;
2892 commit_page = cpu_buffer->commit_page;
2893 commit = rb_page_commit(commit_page);
2895 return ((iter->head_page == commit_page && iter->head == commit) ||
2896 (iter->head_page == reader && commit_page == head_page &&
2897 head_page->read == commit &&
2898 iter->head == rb_page_commit(cpu_buffer->reader_page)));
2900 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
2903 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2904 struct ring_buffer_event *event)
2908 switch (event->type_len) {
2909 case RINGBUF_TYPE_PADDING:
2912 case RINGBUF_TYPE_TIME_EXTEND:
2913 delta = event->array[0];
2915 delta += event->time_delta;
2916 cpu_buffer->read_stamp += delta;
2919 case RINGBUF_TYPE_TIME_STAMP:
2920 /* FIXME: not implemented */
2923 case RINGBUF_TYPE_DATA:
2924 cpu_buffer->read_stamp += event->time_delta;
2934 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
2935 struct ring_buffer_event *event)
2939 switch (event->type_len) {
2940 case RINGBUF_TYPE_PADDING:
2943 case RINGBUF_TYPE_TIME_EXTEND:
2944 delta = event->array[0];
2946 delta += event->time_delta;
2947 iter->read_stamp += delta;
2950 case RINGBUF_TYPE_TIME_STAMP:
2951 /* FIXME: not implemented */
2954 case RINGBUF_TYPE_DATA:
2955 iter->read_stamp += event->time_delta;
2964 static struct buffer_page *
2965 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
2967 struct buffer_page *reader = NULL;
2968 unsigned long overwrite;
2969 unsigned long flags;
2973 local_irq_save(flags);
2974 arch_spin_lock(&cpu_buffer->lock);
2978 * This should normally only loop twice. But because the
2979 * start of the reader inserts an empty page, it causes
2980 * a case where we will loop three times. There should be no
2981 * reason to loop four times (that I know of).
2983 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
2988 reader = cpu_buffer->reader_page;
2990 /* If there's more to read, return this page */
2991 if (cpu_buffer->reader_page->read < rb_page_size(reader))
2994 /* Never should we have an index greater than the size */
2995 if (RB_WARN_ON(cpu_buffer,
2996 cpu_buffer->reader_page->read > rb_page_size(reader)))
2999 /* check if we caught up to the tail */
3001 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3005 * Reset the reader page to size zero.
3007 local_set(&cpu_buffer->reader_page->write, 0);
3008 local_set(&cpu_buffer->reader_page->entries, 0);
3009 local_set(&cpu_buffer->reader_page->page->commit, 0);
3010 cpu_buffer->reader_page->real_end = 0;
3014 * Splice the empty reader page into the list around the head.
3016 reader = rb_set_head_page(cpu_buffer);
3019 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3020 cpu_buffer->reader_page->list.prev = reader->list.prev;
3023 * cpu_buffer->pages just needs to point to the buffer, it
3024 * has no specific buffer page to point to. Lets move it out
3025 * of our way so we don't accidentally swap it.
3027 cpu_buffer->pages = reader->list.prev;
3029 /* The reader page will be pointing to the new head */
3030 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3033 * We want to make sure we read the overruns after we set up our
3034 * pointers to the next object. The writer side does a
3035 * cmpxchg to cross pages which acts as the mb on the writer
3036 * side. Note, the reader will constantly fail the swap
3037 * while the writer is updating the pointers, so this
3038 * guarantees that the overwrite recorded here is the one we
3039 * want to compare with the last_overrun.
3042 overwrite = local_read(&(cpu_buffer->overrun));
3045 * Here's the tricky part.
3047 * We need to move the pointer past the header page.
3048 * But we can only do that if a writer is not currently
3049 * moving it. The page before the header page has the
3050 * flag bit '1' set if it is pointing to the page we want.
3051 * but if the writer is in the process of moving it
3052 * than it will be '2' or already moved '0'.
3055 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3058 * If we did not convert it, then we must try again.
3064 * Yeah! We succeeded in replacing the page.
3066 * Now make the new head point back to the reader page.
3068 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3069 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3071 /* Finally update the reader page to the new head */
3072 cpu_buffer->reader_page = reader;
3073 cpu_buffer->reader_page->read = 0;
3075 if (overwrite != cpu_buffer->last_overrun) {
3076 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3077 cpu_buffer->last_overrun = overwrite;
3083 /* Update the read_stamp on the first event */
3084 if (reader && reader->read == 0)
3085 cpu_buffer->read_stamp = reader->page->time_stamp;
3087 arch_spin_unlock(&cpu_buffer->lock);
3088 local_irq_restore(flags);
3093 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3095 struct ring_buffer_event *event;
3096 struct buffer_page *reader;
3099 reader = rb_get_reader_page(cpu_buffer);
3101 /* This function should not be called when buffer is empty */
3102 if (RB_WARN_ON(cpu_buffer, !reader))
3105 event = rb_reader_event(cpu_buffer);
3107 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3110 rb_update_read_stamp(cpu_buffer, event);
3112 length = rb_event_length(event);
3113 cpu_buffer->reader_page->read += length;
3116 static void rb_advance_iter(struct ring_buffer_iter *iter)
3118 struct ring_buffer_per_cpu *cpu_buffer;
3119 struct ring_buffer_event *event;
3122 cpu_buffer = iter->cpu_buffer;
3125 * Check if we are at the end of the buffer.
3127 if (iter->head >= rb_page_size(iter->head_page)) {
3128 /* discarded commits can make the page empty */
3129 if (iter->head_page == cpu_buffer->commit_page)
3135 event = rb_iter_head_event(iter);
3137 length = rb_event_length(event);
3140 * This should not be called to advance the header if we are
3141 * at the tail of the buffer.
3143 if (RB_WARN_ON(cpu_buffer,
3144 (iter->head_page == cpu_buffer->commit_page) &&
3145 (iter->head + length > rb_commit_index(cpu_buffer))))
3148 rb_update_iter_read_stamp(iter, event);
3150 iter->head += length;
3152 /* check for end of page padding */
3153 if ((iter->head >= rb_page_size(iter->head_page)) &&
3154 (iter->head_page != cpu_buffer->commit_page))
3155 rb_advance_iter(iter);
3158 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3160 return cpu_buffer->lost_events;
3163 static struct ring_buffer_event *
3164 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3165 unsigned long *lost_events)
3167 struct ring_buffer_event *event;
3168 struct buffer_page *reader;
3173 * We repeat when a time extend is encountered.
3174 * Since the time extend is always attached to a data event,
3175 * we should never loop more than once.
3176 * (We never hit the following condition more than twice).
3178 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3181 reader = rb_get_reader_page(cpu_buffer);
3185 event = rb_reader_event(cpu_buffer);
3187 switch (event->type_len) {
3188 case RINGBUF_TYPE_PADDING:
3189 if (rb_null_event(event))
3190 RB_WARN_ON(cpu_buffer, 1);
3192 * Because the writer could be discarding every
3193 * event it creates (which would probably be bad)
3194 * if we were to go back to "again" then we may never
3195 * catch up, and will trigger the warn on, or lock
3196 * the box. Return the padding, and we will release
3197 * the current locks, and try again.
3201 case RINGBUF_TYPE_TIME_EXTEND:
3202 /* Internal data, OK to advance */
3203 rb_advance_reader(cpu_buffer);
3206 case RINGBUF_TYPE_TIME_STAMP:
3207 /* FIXME: not implemented */
3208 rb_advance_reader(cpu_buffer);
3211 case RINGBUF_TYPE_DATA:
3213 *ts = cpu_buffer->read_stamp + event->time_delta;
3214 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3215 cpu_buffer->cpu, ts);
3218 *lost_events = rb_lost_events(cpu_buffer);
3227 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3229 static struct ring_buffer_event *
3230 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3232 struct ring_buffer *buffer;
3233 struct ring_buffer_per_cpu *cpu_buffer;
3234 struct ring_buffer_event *event;
3237 cpu_buffer = iter->cpu_buffer;
3238 buffer = cpu_buffer->buffer;
3241 * Check if someone performed a consuming read to
3242 * the buffer. A consuming read invalidates the iterator
3243 * and we need to reset the iterator in this case.
3245 if (unlikely(iter->cache_read != cpu_buffer->read ||
3246 iter->cache_reader_page != cpu_buffer->reader_page))
3247 rb_iter_reset(iter);
3250 if (ring_buffer_iter_empty(iter))
3254 * We repeat when a time extend is encountered or we hit
3255 * the end of the page. Since the time extend is always attached
3256 * to a data event, we should never loop more than three times.
3257 * Once for going to next page, once on time extend, and
3258 * finally once to get the event.
3259 * (We never hit the following condition more than thrice).
3261 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3264 if (rb_per_cpu_empty(cpu_buffer))
3267 if (iter->head >= local_read(&iter->head_page->page->commit)) {
3272 event = rb_iter_head_event(iter);
3274 switch (event->type_len) {
3275 case RINGBUF_TYPE_PADDING:
3276 if (rb_null_event(event)) {
3280 rb_advance_iter(iter);
3283 case RINGBUF_TYPE_TIME_EXTEND:
3284 /* Internal data, OK to advance */
3285 rb_advance_iter(iter);
3288 case RINGBUF_TYPE_TIME_STAMP:
3289 /* FIXME: not implemented */
3290 rb_advance_iter(iter);
3293 case RINGBUF_TYPE_DATA:
3295 *ts = iter->read_stamp + event->time_delta;
3296 ring_buffer_normalize_time_stamp(buffer,
3297 cpu_buffer->cpu, ts);
3307 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3309 static inline int rb_ok_to_lock(void)
3312 * If an NMI die dumps out the content of the ring buffer
3313 * do not grab locks. We also permanently disable the ring
3314 * buffer too. A one time deal is all you get from reading
3315 * the ring buffer from an NMI.
3317 if (likely(!in_nmi()))
3320 tracing_off_permanent();
3325 * ring_buffer_peek - peek at the next event to be read
3326 * @buffer: The ring buffer to read
3327 * @cpu: The cpu to peak at
3328 * @ts: The timestamp counter of this event.
3329 * @lost_events: a variable to store if events were lost (may be NULL)
3331 * This will return the event that will be read next, but does
3332 * not consume the data.
3334 struct ring_buffer_event *
3335 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3336 unsigned long *lost_events)
3338 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3339 struct ring_buffer_event *event;
3340 unsigned long flags;
3343 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3346 dolock = rb_ok_to_lock();
3348 local_irq_save(flags);
3350 raw_spin_lock(&cpu_buffer->reader_lock);
3351 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3352 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3353 rb_advance_reader(cpu_buffer);
3355 raw_spin_unlock(&cpu_buffer->reader_lock);
3356 local_irq_restore(flags);
3358 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3365 * ring_buffer_iter_peek - peek at the next event to be read
3366 * @iter: The ring buffer iterator
3367 * @ts: The timestamp counter of this event.
3369 * This will return the event that will be read next, but does
3370 * not increment the iterator.
3372 struct ring_buffer_event *
3373 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3375 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3376 struct ring_buffer_event *event;
3377 unsigned long flags;
3380 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3381 event = rb_iter_peek(iter, ts);
3382 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3384 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3391 * ring_buffer_consume - return an event and consume it
3392 * @buffer: The ring buffer to get the next event from
3393 * @cpu: the cpu to read the buffer from
3394 * @ts: a variable to store the timestamp (may be NULL)
3395 * @lost_events: a variable to store if events were lost (may be NULL)
3397 * Returns the next event in the ring buffer, and that event is consumed.
3398 * Meaning, that sequential reads will keep returning a different event,
3399 * and eventually empty the ring buffer if the producer is slower.
3401 struct ring_buffer_event *
3402 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3403 unsigned long *lost_events)
3405 struct ring_buffer_per_cpu *cpu_buffer;
3406 struct ring_buffer_event *event = NULL;
3407 unsigned long flags;
3410 dolock = rb_ok_to_lock();
3413 /* might be called in atomic */
3416 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3419 cpu_buffer = buffer->buffers[cpu];
3420 local_irq_save(flags);
3422 raw_spin_lock(&cpu_buffer->reader_lock);
3424 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3426 cpu_buffer->lost_events = 0;
3427 rb_advance_reader(cpu_buffer);
3431 raw_spin_unlock(&cpu_buffer->reader_lock);
3432 local_irq_restore(flags);
3437 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3442 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3445 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3446 * @buffer: The ring buffer to read from
3447 * @cpu: The cpu buffer to iterate over
3449 * This performs the initial preparations necessary to iterate
3450 * through the buffer. Memory is allocated, buffer recording
3451 * is disabled, and the iterator pointer is returned to the caller.
3453 * Disabling buffer recordng prevents the reading from being
3454 * corrupted. This is not a consuming read, so a producer is not
3457 * After a sequence of ring_buffer_read_prepare calls, the user is
3458 * expected to make at least one call to ring_buffer_prepare_sync.
3459 * Afterwards, ring_buffer_read_start is invoked to get things going
3462 * This overall must be paired with ring_buffer_finish.
3464 struct ring_buffer_iter *
3465 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3467 struct ring_buffer_per_cpu *cpu_buffer;
3468 struct ring_buffer_iter *iter;
3470 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3473 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3477 cpu_buffer = buffer->buffers[cpu];
3479 iter->cpu_buffer = cpu_buffer;
3481 atomic_inc(&cpu_buffer->record_disabled);
3485 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
3488 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
3490 * All previously invoked ring_buffer_read_prepare calls to prepare
3491 * iterators will be synchronized. Afterwards, read_buffer_read_start
3492 * calls on those iterators are allowed.
3495 ring_buffer_read_prepare_sync(void)
3497 synchronize_sched();
3499 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
3502 * ring_buffer_read_start - start a non consuming read of the buffer
3503 * @iter: The iterator returned by ring_buffer_read_prepare
3505 * This finalizes the startup of an iteration through the buffer.
3506 * The iterator comes from a call to ring_buffer_read_prepare and
3507 * an intervening ring_buffer_read_prepare_sync must have been
3510 * Must be paired with ring_buffer_finish.
3513 ring_buffer_read_start(struct ring_buffer_iter *iter)
3515 struct ring_buffer_per_cpu *cpu_buffer;
3516 unsigned long flags;
3521 cpu_buffer = iter->cpu_buffer;
3523 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3524 arch_spin_lock(&cpu_buffer->lock);
3525 rb_iter_reset(iter);
3526 arch_spin_unlock(&cpu_buffer->lock);
3527 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3529 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
3532 * ring_buffer_finish - finish reading the iterator of the buffer
3533 * @iter: The iterator retrieved by ring_buffer_start
3535 * This re-enables the recording to the buffer, and frees the
3539 ring_buffer_read_finish(struct ring_buffer_iter *iter)
3541 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3543 atomic_dec(&cpu_buffer->record_disabled);
3546 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
3549 * ring_buffer_read - read the next item in the ring buffer by the iterator
3550 * @iter: The ring buffer iterator
3551 * @ts: The time stamp of the event read.
3553 * This reads the next event in the ring buffer and increments the iterator.
3555 struct ring_buffer_event *
3556 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
3558 struct ring_buffer_event *event;
3559 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3560 unsigned long flags;
3562 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3564 event = rb_iter_peek(iter, ts);
3568 if (event->type_len == RINGBUF_TYPE_PADDING)
3571 rb_advance_iter(iter);
3573 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3577 EXPORT_SYMBOL_GPL(ring_buffer_read);
3580 * ring_buffer_size - return the size of the ring buffer (in bytes)
3581 * @buffer: The ring buffer.
3583 unsigned long ring_buffer_size(struct ring_buffer *buffer)
3585 return BUF_PAGE_SIZE * buffer->pages;
3587 EXPORT_SYMBOL_GPL(ring_buffer_size);
3590 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
3592 rb_head_page_deactivate(cpu_buffer);
3594 cpu_buffer->head_page
3595 = list_entry(cpu_buffer->pages, struct buffer_page, list);
3596 local_set(&cpu_buffer->head_page->write, 0);
3597 local_set(&cpu_buffer->head_page->entries, 0);
3598 local_set(&cpu_buffer->head_page->page->commit, 0);
3600 cpu_buffer->head_page->read = 0;
3602 cpu_buffer->tail_page = cpu_buffer->head_page;
3603 cpu_buffer->commit_page = cpu_buffer->head_page;
3605 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
3606 local_set(&cpu_buffer->reader_page->write, 0);
3607 local_set(&cpu_buffer->reader_page->entries, 0);
3608 local_set(&cpu_buffer->reader_page->page->commit, 0);
3609 cpu_buffer->reader_page->read = 0;
3611 local_set(&cpu_buffer->commit_overrun, 0);
3612 local_set(&cpu_buffer->entries_bytes, 0);
3613 local_set(&cpu_buffer->overrun, 0);
3614 local_set(&cpu_buffer->entries, 0);
3615 local_set(&cpu_buffer->committing, 0);
3616 local_set(&cpu_buffer->commits, 0);
3617 cpu_buffer->read = 0;
3618 cpu_buffer->read_bytes = 0;
3620 cpu_buffer->write_stamp = 0;
3621 cpu_buffer->read_stamp = 0;
3623 cpu_buffer->lost_events = 0;
3624 cpu_buffer->last_overrun = 0;
3626 rb_head_page_activate(cpu_buffer);
3630 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
3631 * @buffer: The ring buffer to reset a per cpu buffer of
3632 * @cpu: The CPU buffer to be reset
3634 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
3636 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3637 unsigned long flags;
3639 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3642 atomic_inc(&cpu_buffer->record_disabled);
3644 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3646 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
3649 arch_spin_lock(&cpu_buffer->lock);
3651 rb_reset_cpu(cpu_buffer);
3653 arch_spin_unlock(&cpu_buffer->lock);
3656 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3658 atomic_dec(&cpu_buffer->record_disabled);
3660 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
3663 * ring_buffer_reset - reset a ring buffer
3664 * @buffer: The ring buffer to reset all cpu buffers
3666 void ring_buffer_reset(struct ring_buffer *buffer)
3670 for_each_buffer_cpu(buffer, cpu)
3671 ring_buffer_reset_cpu(buffer, cpu);
3673 EXPORT_SYMBOL_GPL(ring_buffer_reset);
3676 * rind_buffer_empty - is the ring buffer empty?
3677 * @buffer: The ring buffer to test
3679 int ring_buffer_empty(struct ring_buffer *buffer)
3681 struct ring_buffer_per_cpu *cpu_buffer;
3682 unsigned long flags;
3687 dolock = rb_ok_to_lock();
3689 /* yes this is racy, but if you don't like the race, lock the buffer */
3690 for_each_buffer_cpu(buffer, cpu) {
3691 cpu_buffer = buffer->buffers[cpu];
3692 local_irq_save(flags);
3694 raw_spin_lock(&cpu_buffer->reader_lock);
3695 ret = rb_per_cpu_empty(cpu_buffer);
3697 raw_spin_unlock(&cpu_buffer->reader_lock);
3698 local_irq_restore(flags);
3706 EXPORT_SYMBOL_GPL(ring_buffer_empty);
3709 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
3710 * @buffer: The ring buffer
3711 * @cpu: The CPU buffer to test
3713 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
3715 struct ring_buffer_per_cpu *cpu_buffer;
3716 unsigned long flags;
3720 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3723 dolock = rb_ok_to_lock();
3725 cpu_buffer = buffer->buffers[cpu];
3726 local_irq_save(flags);
3728 raw_spin_lock(&cpu_buffer->reader_lock);
3729 ret = rb_per_cpu_empty(cpu_buffer);
3731 raw_spin_unlock(&cpu_buffer->reader_lock);
3732 local_irq_restore(flags);
3736 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
3738 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3740 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
3741 * @buffer_a: One buffer to swap with
3742 * @buffer_b: The other buffer to swap with
3744 * This function is useful for tracers that want to take a "snapshot"
3745 * of a CPU buffer and has another back up buffer lying around.
3746 * it is expected that the tracer handles the cpu buffer not being
3747 * used at the moment.
3749 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
3750 struct ring_buffer *buffer_b, int cpu)
3752 struct ring_buffer_per_cpu *cpu_buffer_a;
3753 struct ring_buffer_per_cpu *cpu_buffer_b;
3756 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
3757 !cpumask_test_cpu(cpu, buffer_b->cpumask))
3760 /* At least make sure the two buffers are somewhat the same */
3761 if (buffer_a->pages != buffer_b->pages)
3766 if (ring_buffer_flags != RB_BUFFERS_ON)
3769 if (atomic_read(&buffer_a->record_disabled))
3772 if (atomic_read(&buffer_b->record_disabled))
3775 cpu_buffer_a = buffer_a->buffers[cpu];
3776 cpu_buffer_b = buffer_b->buffers[cpu];
3778 if (atomic_read(&cpu_buffer_a->record_disabled))
3781 if (atomic_read(&cpu_buffer_b->record_disabled))
3785 * We can't do a synchronize_sched here because this
3786 * function can be called in atomic context.
3787 * Normally this will be called from the same CPU as cpu.
3788 * If not it's up to the caller to protect this.
3790 atomic_inc(&cpu_buffer_a->record_disabled);
3791 atomic_inc(&cpu_buffer_b->record_disabled);
3794 if (local_read(&cpu_buffer_a->committing))
3796 if (local_read(&cpu_buffer_b->committing))
3799 buffer_a->buffers[cpu] = cpu_buffer_b;
3800 buffer_b->buffers[cpu] = cpu_buffer_a;
3802 cpu_buffer_b->buffer = buffer_a;
3803 cpu_buffer_a->buffer = buffer_b;
3808 atomic_dec(&cpu_buffer_a->record_disabled);
3809 atomic_dec(&cpu_buffer_b->record_disabled);
3813 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
3814 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
3817 * ring_buffer_alloc_read_page - allocate a page to read from buffer
3818 * @buffer: the buffer to allocate for.
3820 * This function is used in conjunction with ring_buffer_read_page.
3821 * When reading a full page from the ring buffer, these functions
3822 * can be used to speed up the process. The calling function should
3823 * allocate a few pages first with this function. Then when it
3824 * needs to get pages from the ring buffer, it passes the result
3825 * of this function into ring_buffer_read_page, which will swap
3826 * the page that was allocated, with the read page of the buffer.
3829 * The page allocated, or NULL on error.
3831 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
3833 struct buffer_data_page *bpage;
3836 page = alloc_pages_node(cpu_to_node(cpu),
3837 GFP_KERNEL | __GFP_NORETRY, 0);
3841 bpage = page_address(page);
3843 rb_init_page(bpage);
3847 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
3850 * ring_buffer_free_read_page - free an allocated read page
3851 * @buffer: the buffer the page was allocate for
3852 * @data: the page to free
3854 * Free a page allocated from ring_buffer_alloc_read_page.
3856 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
3858 free_page((unsigned long)data);
3860 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
3863 * ring_buffer_read_page - extract a page from the ring buffer
3864 * @buffer: buffer to extract from
3865 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
3866 * @len: amount to extract
3867 * @cpu: the cpu of the buffer to extract
3868 * @full: should the extraction only happen when the page is full.
3870 * This function will pull out a page from the ring buffer and consume it.
3871 * @data_page must be the address of the variable that was returned
3872 * from ring_buffer_alloc_read_page. This is because the page might be used
3873 * to swap with a page in the ring buffer.
3876 * rpage = ring_buffer_alloc_read_page(buffer);
3879 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
3881 * process_page(rpage, ret);
3883 * When @full is set, the function will not return true unless
3884 * the writer is off the reader page.
3886 * Note: it is up to the calling functions to handle sleeps and wakeups.
3887 * The ring buffer can be used anywhere in the kernel and can not
3888 * blindly call wake_up. The layer that uses the ring buffer must be
3889 * responsible for that.
3892 * >=0 if data has been transferred, returns the offset of consumed data.
3893 * <0 if no data has been transferred.
3895 int ring_buffer_read_page(struct ring_buffer *buffer,
3896 void **data_page, size_t len, int cpu, int full)
3898 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3899 struct ring_buffer_event *event;
3900 struct buffer_data_page *bpage;
3901 struct buffer_page *reader;
3902 unsigned long missed_events;
3903 unsigned long flags;
3904 unsigned int commit;
3909 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3913 * If len is not big enough to hold the page header, then
3914 * we can not copy anything.
3916 if (len <= BUF_PAGE_HDR_SIZE)
3919 len -= BUF_PAGE_HDR_SIZE;
3928 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3930 reader = rb_get_reader_page(cpu_buffer);
3934 event = rb_reader_event(cpu_buffer);
3936 read = reader->read;
3937 commit = rb_page_commit(reader);
3939 /* Check if any events were dropped */
3940 missed_events = cpu_buffer->lost_events;
3943 * If this page has been partially read or
3944 * if len is not big enough to read the rest of the page or
3945 * a writer is still on the page, then
3946 * we must copy the data from the page to the buffer.
3947 * Otherwise, we can simply swap the page with the one passed in.
3949 if (read || (len < (commit - read)) ||
3950 cpu_buffer->reader_page == cpu_buffer->commit_page) {
3951 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
3952 unsigned int rpos = read;
3953 unsigned int pos = 0;
3959 if (len > (commit - read))
3960 len = (commit - read);
3962 /* Always keep the time extend and data together */
3963 size = rb_event_ts_length(event);
3968 /* save the current timestamp, since the user will need it */
3969 save_timestamp = cpu_buffer->read_stamp;
3971 /* Need to copy one event at a time */
3973 /* We need the size of one event, because
3974 * rb_advance_reader only advances by one event,
3975 * whereas rb_event_ts_length may include the size of
3976 * one or two events.
3977 * We have already ensured there's enough space if this
3978 * is a time extend. */
3979 size = rb_event_length(event);
3980 memcpy(bpage->data + pos, rpage->data + rpos, size);
3984 rb_advance_reader(cpu_buffer);
3985 rpos = reader->read;
3991 event = rb_reader_event(cpu_buffer);
3992 /* Always keep the time extend and data together */
3993 size = rb_event_ts_length(event);
3994 } while (len >= size);
3997 local_set(&bpage->commit, pos);
3998 bpage->time_stamp = save_timestamp;
4000 /* we copied everything to the beginning */
4003 /* update the entry counter */
4004 cpu_buffer->read += rb_page_entries(reader);
4005 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4007 /* swap the pages */
4008 rb_init_page(bpage);
4009 bpage = reader->page;
4010 reader->page = *data_page;
4011 local_set(&reader->write, 0);
4012 local_set(&reader->entries, 0);
4017 * Use the real_end for the data size,
4018 * This gives us a chance to store the lost events
4021 if (reader->real_end)
4022 local_set(&bpage->commit, reader->real_end);
4026 cpu_buffer->lost_events = 0;
4028 commit = local_read(&bpage->commit);
4030 * Set a flag in the commit field if we lost events
4032 if (missed_events) {
4033 /* If there is room at the end of the page to save the
4034 * missed events, then record it there.
4036 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4037 memcpy(&bpage->data[commit], &missed_events,
4038 sizeof(missed_events));
4039 local_add(RB_MISSED_STORED, &bpage->commit);
4040 commit += sizeof(missed_events);
4042 local_add(RB_MISSED_EVENTS, &bpage->commit);
4046 * This page may be off to user land. Zero it out here.
4048 if (commit < BUF_PAGE_SIZE)
4049 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4052 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4057 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4059 #ifdef CONFIG_TRACING
4061 rb_simple_read(struct file *filp, char __user *ubuf,
4062 size_t cnt, loff_t *ppos)
4064 unsigned long *p = filp->private_data;
4068 if (test_bit(RB_BUFFERS_DISABLED_BIT, p))
4069 r = sprintf(buf, "permanently disabled\n");
4071 r = sprintf(buf, "%d\n", test_bit(RB_BUFFERS_ON_BIT, p));
4073 return simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
4077 rb_simple_write(struct file *filp, const char __user *ubuf,
4078 size_t cnt, loff_t *ppos)
4080 unsigned long *p = filp->private_data;
4084 ret = kstrtoul_from_user(ubuf, cnt, 10, &val);
4089 set_bit(RB_BUFFERS_ON_BIT, p);
4091 clear_bit(RB_BUFFERS_ON_BIT, p);
4098 static const struct file_operations rb_simple_fops = {
4099 .open = tracing_open_generic,
4100 .read = rb_simple_read,
4101 .write = rb_simple_write,
4102 .llseek = default_llseek,
4106 static __init int rb_init_debugfs(void)
4108 struct dentry *d_tracer;
4110 d_tracer = tracing_init_dentry();
4112 trace_create_file("tracing_on", 0644, d_tracer,
4113 &ring_buffer_flags, &rb_simple_fops);
4118 fs_initcall(rb_init_debugfs);
4121 #ifdef CONFIG_HOTPLUG_CPU
4122 static int rb_cpu_notify(struct notifier_block *self,
4123 unsigned long action, void *hcpu)
4125 struct ring_buffer *buffer =
4126 container_of(self, struct ring_buffer, cpu_notify);
4127 long cpu = (long)hcpu;
4130 case CPU_UP_PREPARE:
4131 case CPU_UP_PREPARE_FROZEN:
4132 if (cpumask_test_cpu(cpu, buffer->cpumask))
4135 buffer->buffers[cpu] =
4136 rb_allocate_cpu_buffer(buffer, cpu);
4137 if (!buffer->buffers[cpu]) {
4138 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4143 cpumask_set_cpu(cpu, buffer->cpumask);
4145 case CPU_DOWN_PREPARE:
4146 case CPU_DOWN_PREPARE_FROZEN:
4149 * If we were to free the buffer, then the user would
4150 * lose any trace that was in the buffer.