4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 #include <linux/ring_buffer.h>
7 #include <linux/trace_clock.h>
8 #include <linux/ftrace_irq.h>
9 #include <linux/spinlock.h>
10 #include <linux/debugfs.h>
11 #include <linux/uaccess.h>
12 #include <linux/hardirq.h>
13 #include <linux/kmemcheck.h>
14 #include <linux/module.h>
15 #include <linux/percpu.h>
16 #include <linux/mutex.h>
17 #include <linux/slab.h>
18 #include <linux/init.h>
19 #include <linux/hash.h>
20 #include <linux/list.h>
21 #include <linux/cpu.h>
24 #include <asm/local.h>
28 * The ring buffer header is special. We must manually up keep it.
30 int ring_buffer_print_entry_header(struct trace_seq *s)
34 ret = trace_seq_printf(s, "# compressed entry header\n");
35 ret = trace_seq_printf(s, "\ttype_len : 5 bits\n");
36 ret = trace_seq_printf(s, "\ttime_delta : 27 bits\n");
37 ret = trace_seq_printf(s, "\tarray : 32 bits\n");
38 ret = trace_seq_printf(s, "\n");
39 ret = trace_seq_printf(s, "\tpadding : type == %d\n",
40 RINGBUF_TYPE_PADDING);
41 ret = trace_seq_printf(s, "\ttime_extend : type == %d\n",
42 RINGBUF_TYPE_TIME_EXTEND);
43 ret = trace_seq_printf(s, "\tdata max type_len == %d\n",
44 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
50 * The ring buffer is made up of a list of pages. A separate list of pages is
51 * allocated for each CPU. A writer may only write to a buffer that is
52 * associated with the CPU it is currently executing on. A reader may read
53 * from any per cpu buffer.
55 * The reader is special. For each per cpu buffer, the reader has its own
56 * reader page. When a reader has read the entire reader page, this reader
57 * page is swapped with another page in the ring buffer.
59 * Now, as long as the writer is off the reader page, the reader can do what
60 * ever it wants with that page. The writer will never write to that page
61 * again (as long as it is out of the ring buffer).
63 * Here's some silly ASCII art.
66 * |reader| RING BUFFER
68 * +------+ +---+ +---+ +---+
77 * |reader| RING BUFFER
78 * |page |------------------v
79 * +------+ +---+ +---+ +---+
88 * |reader| RING BUFFER
89 * |page |------------------v
90 * +------+ +---+ +---+ +---+
95 * +------------------------------+
99 * |buffer| RING BUFFER
100 * |page |------------------v
101 * +------+ +---+ +---+ +---+
103 * | New +---+ +---+ +---+
106 * +------------------------------+
109 * After we make this swap, the reader can hand this page off to the splice
110 * code and be done with it. It can even allocate a new page if it needs to
111 * and swap that into the ring buffer.
113 * We will be using cmpxchg soon to make all this lockless.
118 * A fast way to enable or disable all ring buffers is to
119 * call tracing_on or tracing_off. Turning off the ring buffers
120 * prevents all ring buffers from being recorded to.
121 * Turning this switch on, makes it OK to write to the
122 * ring buffer, if the ring buffer is enabled itself.
124 * There's three layers that must be on in order to write
125 * to the ring buffer.
127 * 1) This global flag must be set.
128 * 2) The ring buffer must be enabled for recording.
129 * 3) The per cpu buffer must be enabled for recording.
131 * In case of an anomaly, this global flag has a bit set that
132 * will permantly disable all ring buffers.
136 * Global flag to disable all recording to ring buffers
137 * This has two bits: ON, DISABLED
141 * 0 0 : ring buffers are off
142 * 1 0 : ring buffers are on
143 * X 1 : ring buffers are permanently disabled
147 RB_BUFFERS_ON_BIT = 0,
148 RB_BUFFERS_DISABLED_BIT = 1,
152 RB_BUFFERS_ON = 1 << RB_BUFFERS_ON_BIT,
153 RB_BUFFERS_DISABLED = 1 << RB_BUFFERS_DISABLED_BIT,
156 static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON;
158 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
161 * tracing_on - enable all tracing buffers
163 * This function enables all tracing buffers that may have been
164 * disabled with tracing_off.
166 void tracing_on(void)
168 set_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
170 EXPORT_SYMBOL_GPL(tracing_on);
173 * tracing_off - turn off all tracing buffers
175 * This function stops all tracing buffers from recording data.
176 * It does not disable any overhead the tracers themselves may
177 * be causing. This function simply causes all recording to
178 * the ring buffers to fail.
180 void tracing_off(void)
182 clear_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
184 EXPORT_SYMBOL_GPL(tracing_off);
187 * tracing_off_permanent - permanently disable ring buffers
189 * This function, once called, will disable all ring buffers
192 void tracing_off_permanent(void)
194 set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags);
198 * tracing_is_on - show state of ring buffers enabled
200 int tracing_is_on(void)
202 return ring_buffer_flags == RB_BUFFERS_ON;
204 EXPORT_SYMBOL_GPL(tracing_is_on);
206 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
207 #define RB_ALIGNMENT 4U
208 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
209 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
211 #if !defined(CONFIG_64BIT) || defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
212 # define RB_FORCE_8BYTE_ALIGNMENT 0
213 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
215 # define RB_FORCE_8BYTE_ALIGNMENT 1
216 # define RB_ARCH_ALIGNMENT 8U
219 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
220 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
223 RB_LEN_TIME_EXTEND = 8,
224 RB_LEN_TIME_STAMP = 16,
227 #define skip_time_extend(event) \
228 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
230 static inline int rb_null_event(struct ring_buffer_event *event)
232 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
235 static void rb_event_set_padding(struct ring_buffer_event *event)
237 /* padding has a NULL time_delta */
238 event->type_len = RINGBUF_TYPE_PADDING;
239 event->time_delta = 0;
243 rb_event_data_length(struct ring_buffer_event *event)
248 length = event->type_len * RB_ALIGNMENT;
250 length = event->array[0];
251 return length + RB_EVNT_HDR_SIZE;
255 * Return the length of the given event. Will return
256 * the length of the time extend if the event is a
259 static inline unsigned
260 rb_event_length(struct ring_buffer_event *event)
262 switch (event->type_len) {
263 case RINGBUF_TYPE_PADDING:
264 if (rb_null_event(event))
267 return event->array[0] + RB_EVNT_HDR_SIZE;
269 case RINGBUF_TYPE_TIME_EXTEND:
270 return RB_LEN_TIME_EXTEND;
272 case RINGBUF_TYPE_TIME_STAMP:
273 return RB_LEN_TIME_STAMP;
275 case RINGBUF_TYPE_DATA:
276 return rb_event_data_length(event);
285 * Return total length of time extend and data,
286 * or just the event length for all other events.
288 static inline unsigned
289 rb_event_ts_length(struct ring_buffer_event *event)
293 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
294 /* time extends include the data event after it */
295 len = RB_LEN_TIME_EXTEND;
296 event = skip_time_extend(event);
298 return len + rb_event_length(event);
302 * ring_buffer_event_length - return the length of the event
303 * @event: the event to get the length of
305 * Returns the size of the data load of a data event.
306 * If the event is something other than a data event, it
307 * returns the size of the event itself. With the exception
308 * of a TIME EXTEND, where it still returns the size of the
309 * data load of the data event after it.
311 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
315 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
316 event = skip_time_extend(event);
318 length = rb_event_length(event);
319 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
321 length -= RB_EVNT_HDR_SIZE;
322 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
323 length -= sizeof(event->array[0]);
326 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
328 /* inline for ring buffer fast paths */
330 rb_event_data(struct ring_buffer_event *event)
332 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
333 event = skip_time_extend(event);
334 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
335 /* If length is in len field, then array[0] has the data */
337 return (void *)&event->array[0];
338 /* Otherwise length is in array[0] and array[1] has the data */
339 return (void *)&event->array[1];
343 * ring_buffer_event_data - return the data of the event
344 * @event: the event to get the data from
346 void *ring_buffer_event_data(struct ring_buffer_event *event)
348 return rb_event_data(event);
350 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
352 #define for_each_buffer_cpu(buffer, cpu) \
353 for_each_cpu(cpu, buffer->cpumask)
356 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
357 #define TS_DELTA_TEST (~TS_MASK)
359 /* Flag when events were overwritten */
360 #define RB_MISSED_EVENTS (1 << 31)
361 /* Missed count stored at end */
362 #define RB_MISSED_STORED (1 << 30)
364 struct buffer_data_page {
365 u64 time_stamp; /* page time stamp */
366 local_t commit; /* write committed index */
367 unsigned char data[]; /* data of buffer page */
371 * Note, the buffer_page list must be first. The buffer pages
372 * are allocated in cache lines, which means that each buffer
373 * page will be at the beginning of a cache line, and thus
374 * the least significant bits will be zero. We use this to
375 * add flags in the list struct pointers, to make the ring buffer
379 struct list_head list; /* list of buffer pages */
380 local_t write; /* index for next write */
381 unsigned read; /* index for next read */
382 local_t entries; /* entries on this page */
383 unsigned long real_end; /* real end of data */
384 struct buffer_data_page *page; /* Actual data page */
388 * The buffer page counters, write and entries, must be reset
389 * atomically when crossing page boundaries. To synchronize this
390 * update, two counters are inserted into the number. One is
391 * the actual counter for the write position or count on the page.
393 * The other is a counter of updaters. Before an update happens
394 * the update partition of the counter is incremented. This will
395 * allow the updater to update the counter atomically.
397 * The counter is 20 bits, and the state data is 12.
399 #define RB_WRITE_MASK 0xfffff
400 #define RB_WRITE_INTCNT (1 << 20)
402 static void rb_init_page(struct buffer_data_page *bpage)
404 local_set(&bpage->commit, 0);
408 * ring_buffer_page_len - the size of data on the page.
409 * @page: The page to read
411 * Returns the amount of data on the page, including buffer page header.
413 size_t ring_buffer_page_len(void *page)
415 return local_read(&((struct buffer_data_page *)page)->commit)
420 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
423 static void free_buffer_page(struct buffer_page *bpage)
425 free_page((unsigned long)bpage->page);
430 * We need to fit the time_stamp delta into 27 bits.
432 static inline int test_time_stamp(u64 delta)
434 if (delta & TS_DELTA_TEST)
439 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
441 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
442 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
444 int ring_buffer_print_page_header(struct trace_seq *s)
446 struct buffer_data_page field;
449 ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t"
450 "offset:0;\tsize:%u;\tsigned:%u;\n",
451 (unsigned int)sizeof(field.time_stamp),
452 (unsigned int)is_signed_type(u64));
454 ret = trace_seq_printf(s, "\tfield: local_t commit;\t"
455 "offset:%u;\tsize:%u;\tsigned:%u;\n",
456 (unsigned int)offsetof(typeof(field), commit),
457 (unsigned int)sizeof(field.commit),
458 (unsigned int)is_signed_type(long));
460 ret = trace_seq_printf(s, "\tfield: int overwrite;\t"
461 "offset:%u;\tsize:%u;\tsigned:%u;\n",
462 (unsigned int)offsetof(typeof(field), commit),
464 (unsigned int)is_signed_type(long));
466 ret = trace_seq_printf(s, "\tfield: char data;\t"
467 "offset:%u;\tsize:%u;\tsigned:%u;\n",
468 (unsigned int)offsetof(typeof(field), data),
469 (unsigned int)BUF_PAGE_SIZE,
470 (unsigned int)is_signed_type(char));
476 * head_page == tail_page && head == tail then buffer is empty.
478 struct ring_buffer_per_cpu {
480 atomic_t record_disabled;
481 struct ring_buffer *buffer;
482 spinlock_t reader_lock; /* serialize readers */
483 arch_spinlock_t lock;
484 struct lock_class_key lock_key;
485 struct list_head *pages;
486 struct buffer_page *head_page; /* read from head */
487 struct buffer_page *tail_page; /* write to tail */
488 struct buffer_page *commit_page; /* committed pages */
489 struct buffer_page *reader_page;
490 unsigned long lost_events;
491 unsigned long last_overrun;
492 local_t commit_overrun;
506 atomic_t record_disabled;
507 cpumask_var_t cpumask;
509 struct lock_class_key *reader_lock_key;
513 struct ring_buffer_per_cpu **buffers;
515 #ifdef CONFIG_HOTPLUG_CPU
516 struct notifier_block cpu_notify;
521 struct ring_buffer_iter {
522 struct ring_buffer_per_cpu *cpu_buffer;
524 struct buffer_page *head_page;
525 struct buffer_page *cache_reader_page;
526 unsigned long cache_read;
530 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
531 #define RB_WARN_ON(b, cond) \
533 int _____ret = unlikely(cond); \
535 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
536 struct ring_buffer_per_cpu *__b = \
538 atomic_inc(&__b->buffer->record_disabled); \
540 atomic_inc(&b->record_disabled); \
546 /* Up this if you want to test the TIME_EXTENTS and normalization */
547 #define DEBUG_SHIFT 0
549 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
551 /* shift to debug/test normalization and TIME_EXTENTS */
552 return buffer->clock() << DEBUG_SHIFT;
555 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
559 preempt_disable_notrace();
560 time = rb_time_stamp(buffer);
561 preempt_enable_no_resched_notrace();
565 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
567 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
570 /* Just stupid testing the normalize function and deltas */
573 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
576 * Making the ring buffer lockless makes things tricky.
577 * Although writes only happen on the CPU that they are on,
578 * and they only need to worry about interrupts. Reads can
581 * The reader page is always off the ring buffer, but when the
582 * reader finishes with a page, it needs to swap its page with
583 * a new one from the buffer. The reader needs to take from
584 * the head (writes go to the tail). But if a writer is in overwrite
585 * mode and wraps, it must push the head page forward.
587 * Here lies the problem.
589 * The reader must be careful to replace only the head page, and
590 * not another one. As described at the top of the file in the
591 * ASCII art, the reader sets its old page to point to the next
592 * page after head. It then sets the page after head to point to
593 * the old reader page. But if the writer moves the head page
594 * during this operation, the reader could end up with the tail.
596 * We use cmpxchg to help prevent this race. We also do something
597 * special with the page before head. We set the LSB to 1.
599 * When the writer must push the page forward, it will clear the
600 * bit that points to the head page, move the head, and then set
601 * the bit that points to the new head page.
603 * We also don't want an interrupt coming in and moving the head
604 * page on another writer. Thus we use the second LSB to catch
607 * head->list->prev->next bit 1 bit 0
610 * Points to head page 0 1
613 * Note we can not trust the prev pointer of the head page, because:
615 * +----+ +-----+ +-----+
616 * | |------>| T |---X--->| N |
618 * +----+ +-----+ +-----+
621 * +----------| R |----------+ |
625 * Key: ---X--> HEAD flag set in pointer
630 * (see __rb_reserve_next() to see where this happens)
632 * What the above shows is that the reader just swapped out
633 * the reader page with a page in the buffer, but before it
634 * could make the new header point back to the new page added
635 * it was preempted by a writer. The writer moved forward onto
636 * the new page added by the reader and is about to move forward
639 * You can see, it is legitimate for the previous pointer of
640 * the head (or any page) not to point back to itself. But only
644 #define RB_PAGE_NORMAL 0UL
645 #define RB_PAGE_HEAD 1UL
646 #define RB_PAGE_UPDATE 2UL
649 #define RB_FLAG_MASK 3UL
651 /* PAGE_MOVED is not part of the mask */
652 #define RB_PAGE_MOVED 4UL
655 * rb_list_head - remove any bit
657 static struct list_head *rb_list_head(struct list_head *list)
659 unsigned long val = (unsigned long)list;
661 return (struct list_head *)(val & ~RB_FLAG_MASK);
665 * rb_is_head_page - test if the given page is the head page
667 * Because the reader may move the head_page pointer, we can
668 * not trust what the head page is (it may be pointing to
669 * the reader page). But if the next page is a header page,
670 * its flags will be non zero.
673 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
674 struct buffer_page *page, struct list_head *list)
678 val = (unsigned long)list->next;
680 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
681 return RB_PAGE_MOVED;
683 return val & RB_FLAG_MASK;
689 * The unique thing about the reader page, is that, if the
690 * writer is ever on it, the previous pointer never points
691 * back to the reader page.
693 static int rb_is_reader_page(struct buffer_page *page)
695 struct list_head *list = page->list.prev;
697 return rb_list_head(list->next) != &page->list;
701 * rb_set_list_to_head - set a list_head to be pointing to head.
703 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
704 struct list_head *list)
708 ptr = (unsigned long *)&list->next;
709 *ptr |= RB_PAGE_HEAD;
710 *ptr &= ~RB_PAGE_UPDATE;
714 * rb_head_page_activate - sets up head page
716 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
718 struct buffer_page *head;
720 head = cpu_buffer->head_page;
725 * Set the previous list pointer to have the HEAD flag.
727 rb_set_list_to_head(cpu_buffer, head->list.prev);
730 static void rb_list_head_clear(struct list_head *list)
732 unsigned long *ptr = (unsigned long *)&list->next;
734 *ptr &= ~RB_FLAG_MASK;
738 * rb_head_page_dactivate - clears head page ptr (for free list)
741 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
743 struct list_head *hd;
745 /* Go through the whole list and clear any pointers found. */
746 rb_list_head_clear(cpu_buffer->pages);
748 list_for_each(hd, cpu_buffer->pages)
749 rb_list_head_clear(hd);
752 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
753 struct buffer_page *head,
754 struct buffer_page *prev,
755 int old_flag, int new_flag)
757 struct list_head *list;
758 unsigned long val = (unsigned long)&head->list;
763 val &= ~RB_FLAG_MASK;
765 ret = cmpxchg((unsigned long *)&list->next,
766 val | old_flag, val | new_flag);
768 /* check if the reader took the page */
769 if ((ret & ~RB_FLAG_MASK) != val)
770 return RB_PAGE_MOVED;
772 return ret & RB_FLAG_MASK;
775 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
776 struct buffer_page *head,
777 struct buffer_page *prev,
780 return rb_head_page_set(cpu_buffer, head, prev,
781 old_flag, RB_PAGE_UPDATE);
784 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
785 struct buffer_page *head,
786 struct buffer_page *prev,
789 return rb_head_page_set(cpu_buffer, head, prev,
790 old_flag, RB_PAGE_HEAD);
793 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
794 struct buffer_page *head,
795 struct buffer_page *prev,
798 return rb_head_page_set(cpu_buffer, head, prev,
799 old_flag, RB_PAGE_NORMAL);
802 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
803 struct buffer_page **bpage)
805 struct list_head *p = rb_list_head((*bpage)->list.next);
807 *bpage = list_entry(p, struct buffer_page, list);
810 static struct buffer_page *
811 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
813 struct buffer_page *head;
814 struct buffer_page *page;
815 struct list_head *list;
818 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
822 list = cpu_buffer->pages;
823 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
826 page = head = cpu_buffer->head_page;
828 * It is possible that the writer moves the header behind
829 * where we started, and we miss in one loop.
830 * A second loop should grab the header, but we'll do
831 * three loops just because I'm paranoid.
833 for (i = 0; i < 3; i++) {
835 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
836 cpu_buffer->head_page = page;
839 rb_inc_page(cpu_buffer, &page);
840 } while (page != head);
843 RB_WARN_ON(cpu_buffer, 1);
848 static int rb_head_page_replace(struct buffer_page *old,
849 struct buffer_page *new)
851 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
855 val = *ptr & ~RB_FLAG_MASK;
858 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
864 * rb_tail_page_update - move the tail page forward
866 * Returns 1 if moved tail page, 0 if someone else did.
868 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
869 struct buffer_page *tail_page,
870 struct buffer_page *next_page)
872 struct buffer_page *old_tail;
873 unsigned long old_entries;
874 unsigned long old_write;
878 * The tail page now needs to be moved forward.
880 * We need to reset the tail page, but without messing
881 * with possible erasing of data brought in by interrupts
882 * that have moved the tail page and are currently on it.
884 * We add a counter to the write field to denote this.
886 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
887 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
890 * Just make sure we have seen our old_write and synchronize
891 * with any interrupts that come in.
896 * If the tail page is still the same as what we think
897 * it is, then it is up to us to update the tail
900 if (tail_page == cpu_buffer->tail_page) {
901 /* Zero the write counter */
902 unsigned long val = old_write & ~RB_WRITE_MASK;
903 unsigned long eval = old_entries & ~RB_WRITE_MASK;
906 * This will only succeed if an interrupt did
907 * not come in and change it. In which case, we
908 * do not want to modify it.
910 * We add (void) to let the compiler know that we do not care
911 * about the return value of these functions. We use the
912 * cmpxchg to only update if an interrupt did not already
913 * do it for us. If the cmpxchg fails, we don't care.
915 (void)local_cmpxchg(&next_page->write, old_write, val);
916 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
919 * No need to worry about races with clearing out the commit.
920 * it only can increment when a commit takes place. But that
921 * only happens in the outer most nested commit.
923 local_set(&next_page->page->commit, 0);
925 old_tail = cmpxchg(&cpu_buffer->tail_page,
926 tail_page, next_page);
928 if (old_tail == tail_page)
935 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
936 struct buffer_page *bpage)
938 unsigned long val = (unsigned long)bpage;
940 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
947 * rb_check_list - make sure a pointer to a list has the last bits zero
949 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
950 struct list_head *list)
952 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
954 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
960 * check_pages - integrity check of buffer pages
961 * @cpu_buffer: CPU buffer with pages to test
963 * As a safety measure we check to make sure the data pages have not
966 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
968 struct list_head *head = cpu_buffer->pages;
969 struct buffer_page *bpage, *tmp;
971 rb_head_page_deactivate(cpu_buffer);
973 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
975 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
978 if (rb_check_list(cpu_buffer, head))
981 list_for_each_entry_safe(bpage, tmp, head, list) {
982 if (RB_WARN_ON(cpu_buffer,
983 bpage->list.next->prev != &bpage->list))
985 if (RB_WARN_ON(cpu_buffer,
986 bpage->list.prev->next != &bpage->list))
988 if (rb_check_list(cpu_buffer, &bpage->list))
992 rb_head_page_activate(cpu_buffer);
997 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1000 struct buffer_page *bpage, *tmp;
1007 for (i = 0; i < nr_pages; i++) {
1008 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1009 GFP_KERNEL, cpu_to_node(cpu_buffer->cpu));
1013 rb_check_bpage(cpu_buffer, bpage);
1015 list_add(&bpage->list, &pages);
1017 addr = __get_free_page(GFP_KERNEL);
1020 bpage->page = (void *)addr;
1021 rb_init_page(bpage->page);
1025 * The ring buffer page list is a circular list that does not
1026 * start and end with a list head. All page list items point to
1029 cpu_buffer->pages = pages.next;
1032 rb_check_pages(cpu_buffer);
1037 list_for_each_entry_safe(bpage, tmp, &pages, list) {
1038 list_del_init(&bpage->list);
1039 free_buffer_page(bpage);
1044 static struct ring_buffer_per_cpu *
1045 rb_allocate_cpu_buffer(struct ring_buffer *buffer, int cpu)
1047 struct ring_buffer_per_cpu *cpu_buffer;
1048 struct buffer_page *bpage;
1052 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1053 GFP_KERNEL, cpu_to_node(cpu));
1057 cpu_buffer->cpu = cpu;
1058 cpu_buffer->buffer = buffer;
1059 spin_lock_init(&cpu_buffer->reader_lock);
1060 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1061 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1063 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1064 GFP_KERNEL, cpu_to_node(cpu));
1066 goto fail_free_buffer;
1068 rb_check_bpage(cpu_buffer, bpage);
1070 cpu_buffer->reader_page = bpage;
1071 addr = __get_free_page(GFP_KERNEL);
1073 goto fail_free_reader;
1074 bpage->page = (void *)addr;
1075 rb_init_page(bpage->page);
1077 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1079 ret = rb_allocate_pages(cpu_buffer, buffer->pages);
1081 goto fail_free_reader;
1083 cpu_buffer->head_page
1084 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1085 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1087 rb_head_page_activate(cpu_buffer);
1092 free_buffer_page(cpu_buffer->reader_page);
1099 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1101 struct list_head *head = cpu_buffer->pages;
1102 struct buffer_page *bpage, *tmp;
1104 free_buffer_page(cpu_buffer->reader_page);
1106 rb_head_page_deactivate(cpu_buffer);
1109 list_for_each_entry_safe(bpage, tmp, head, list) {
1110 list_del_init(&bpage->list);
1111 free_buffer_page(bpage);
1113 bpage = list_entry(head, struct buffer_page, list);
1114 free_buffer_page(bpage);
1120 #ifdef CONFIG_HOTPLUG_CPU
1121 static int rb_cpu_notify(struct notifier_block *self,
1122 unsigned long action, void *hcpu);
1126 * ring_buffer_alloc - allocate a new ring_buffer
1127 * @size: the size in bytes per cpu that is needed.
1128 * @flags: attributes to set for the ring buffer.
1130 * Currently the only flag that is available is the RB_FL_OVERWRITE
1131 * flag. This flag means that the buffer will overwrite old data
1132 * when the buffer wraps. If this flag is not set, the buffer will
1133 * drop data when the tail hits the head.
1135 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1136 struct lock_class_key *key)
1138 struct ring_buffer *buffer;
1142 /* keep it in its own cache line */
1143 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1148 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1149 goto fail_free_buffer;
1151 buffer->pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1152 buffer->flags = flags;
1153 buffer->clock = trace_clock_local;
1154 buffer->reader_lock_key = key;
1156 /* need at least two pages */
1157 if (buffer->pages < 2)
1161 * In case of non-hotplug cpu, if the ring-buffer is allocated
1162 * in early initcall, it will not be notified of secondary cpus.
1163 * In that off case, we need to allocate for all possible cpus.
1165 #ifdef CONFIG_HOTPLUG_CPU
1167 cpumask_copy(buffer->cpumask, cpu_online_mask);
1169 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1171 buffer->cpus = nr_cpu_ids;
1173 bsize = sizeof(void *) * nr_cpu_ids;
1174 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1176 if (!buffer->buffers)
1177 goto fail_free_cpumask;
1179 for_each_buffer_cpu(buffer, cpu) {
1180 buffer->buffers[cpu] =
1181 rb_allocate_cpu_buffer(buffer, cpu);
1182 if (!buffer->buffers[cpu])
1183 goto fail_free_buffers;
1186 #ifdef CONFIG_HOTPLUG_CPU
1187 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1188 buffer->cpu_notify.priority = 0;
1189 register_cpu_notifier(&buffer->cpu_notify);
1193 mutex_init(&buffer->mutex);
1198 for_each_buffer_cpu(buffer, cpu) {
1199 if (buffer->buffers[cpu])
1200 rb_free_cpu_buffer(buffer->buffers[cpu]);
1202 kfree(buffer->buffers);
1205 free_cpumask_var(buffer->cpumask);
1212 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1215 * ring_buffer_free - free a ring buffer.
1216 * @buffer: the buffer to free.
1219 ring_buffer_free(struct ring_buffer *buffer)
1225 #ifdef CONFIG_HOTPLUG_CPU
1226 unregister_cpu_notifier(&buffer->cpu_notify);
1229 for_each_buffer_cpu(buffer, cpu)
1230 rb_free_cpu_buffer(buffer->buffers[cpu]);
1234 kfree(buffer->buffers);
1235 free_cpumask_var(buffer->cpumask);
1239 EXPORT_SYMBOL_GPL(ring_buffer_free);
1241 void ring_buffer_set_clock(struct ring_buffer *buffer,
1244 buffer->clock = clock;
1247 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1250 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned nr_pages)
1252 struct buffer_page *bpage;
1253 struct list_head *p;
1256 spin_lock_irq(&cpu_buffer->reader_lock);
1257 rb_head_page_deactivate(cpu_buffer);
1259 for (i = 0; i < nr_pages; i++) {
1260 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1262 p = cpu_buffer->pages->next;
1263 bpage = list_entry(p, struct buffer_page, list);
1264 list_del_init(&bpage->list);
1265 free_buffer_page(bpage);
1267 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1270 rb_reset_cpu(cpu_buffer);
1271 rb_check_pages(cpu_buffer);
1274 spin_unlock_irq(&cpu_buffer->reader_lock);
1278 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer,
1279 struct list_head *pages, unsigned nr_pages)
1281 struct buffer_page *bpage;
1282 struct list_head *p;
1285 spin_lock_irq(&cpu_buffer->reader_lock);
1286 rb_head_page_deactivate(cpu_buffer);
1288 for (i = 0; i < nr_pages; i++) {
1289 if (RB_WARN_ON(cpu_buffer, list_empty(pages)))
1292 bpage = list_entry(p, struct buffer_page, list);
1293 list_del_init(&bpage->list);
1294 list_add_tail(&bpage->list, cpu_buffer->pages);
1296 rb_reset_cpu(cpu_buffer);
1297 rb_check_pages(cpu_buffer);
1300 spin_unlock_irq(&cpu_buffer->reader_lock);
1304 * ring_buffer_resize - resize the ring buffer
1305 * @buffer: the buffer to resize.
1306 * @size: the new size.
1308 * Minimum size is 2 * BUF_PAGE_SIZE.
1310 * Returns -1 on failure.
1312 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size)
1314 struct ring_buffer_per_cpu *cpu_buffer;
1315 unsigned nr_pages, rm_pages, new_pages;
1316 struct buffer_page *bpage, *tmp;
1317 unsigned long buffer_size;
1323 * Always succeed at resizing a non-existent buffer:
1328 size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1329 size *= BUF_PAGE_SIZE;
1330 buffer_size = buffer->pages * BUF_PAGE_SIZE;
1332 /* we need a minimum of two pages */
1333 if (size < BUF_PAGE_SIZE * 2)
1334 size = BUF_PAGE_SIZE * 2;
1336 if (size == buffer_size)
1339 atomic_inc(&buffer->record_disabled);
1341 /* Make sure all writers are done with this buffer. */
1342 synchronize_sched();
1344 mutex_lock(&buffer->mutex);
1347 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1349 if (size < buffer_size) {
1351 /* easy case, just free pages */
1352 if (RB_WARN_ON(buffer, nr_pages >= buffer->pages))
1355 rm_pages = buffer->pages - nr_pages;
1357 for_each_buffer_cpu(buffer, cpu) {
1358 cpu_buffer = buffer->buffers[cpu];
1359 rb_remove_pages(cpu_buffer, rm_pages);
1365 * This is a bit more difficult. We only want to add pages
1366 * when we can allocate enough for all CPUs. We do this
1367 * by allocating all the pages and storing them on a local
1368 * link list. If we succeed in our allocation, then we
1369 * add these pages to the cpu_buffers. Otherwise we just free
1370 * them all and return -ENOMEM;
1372 if (RB_WARN_ON(buffer, nr_pages <= buffer->pages))
1375 new_pages = nr_pages - buffer->pages;
1377 for_each_buffer_cpu(buffer, cpu) {
1378 for (i = 0; i < new_pages; i++) {
1379 bpage = kzalloc_node(ALIGN(sizeof(*bpage),
1381 GFP_KERNEL, cpu_to_node(cpu));
1384 list_add(&bpage->list, &pages);
1385 addr = __get_free_page(GFP_KERNEL);
1388 bpage->page = (void *)addr;
1389 rb_init_page(bpage->page);
1393 for_each_buffer_cpu(buffer, cpu) {
1394 cpu_buffer = buffer->buffers[cpu];
1395 rb_insert_pages(cpu_buffer, &pages, new_pages);
1398 if (RB_WARN_ON(buffer, !list_empty(&pages)))
1402 buffer->pages = nr_pages;
1404 mutex_unlock(&buffer->mutex);
1406 atomic_dec(&buffer->record_disabled);
1411 list_for_each_entry_safe(bpage, tmp, &pages, list) {
1412 list_del_init(&bpage->list);
1413 free_buffer_page(bpage);
1416 mutex_unlock(&buffer->mutex);
1417 atomic_dec(&buffer->record_disabled);
1421 * Something went totally wrong, and we are too paranoid
1422 * to even clean up the mess.
1426 mutex_unlock(&buffer->mutex);
1427 atomic_dec(&buffer->record_disabled);
1430 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1432 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1434 mutex_lock(&buffer->mutex);
1436 buffer->flags |= RB_FL_OVERWRITE;
1438 buffer->flags &= ~RB_FL_OVERWRITE;
1439 mutex_unlock(&buffer->mutex);
1441 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1443 static inline void *
1444 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1446 return bpage->data + index;
1449 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1451 return bpage->page->data + index;
1454 static inline struct ring_buffer_event *
1455 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1457 return __rb_page_index(cpu_buffer->reader_page,
1458 cpu_buffer->reader_page->read);
1461 static inline struct ring_buffer_event *
1462 rb_iter_head_event(struct ring_buffer_iter *iter)
1464 return __rb_page_index(iter->head_page, iter->head);
1467 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1469 return local_read(&bpage->write) & RB_WRITE_MASK;
1472 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1474 return local_read(&bpage->page->commit);
1477 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1479 return local_read(&bpage->entries) & RB_WRITE_MASK;
1482 /* Size is determined by what has been commited */
1483 static inline unsigned rb_page_size(struct buffer_page *bpage)
1485 return rb_page_commit(bpage);
1488 static inline unsigned
1489 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1491 return rb_page_commit(cpu_buffer->commit_page);
1494 static inline unsigned
1495 rb_event_index(struct ring_buffer_event *event)
1497 unsigned long addr = (unsigned long)event;
1499 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1503 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1504 struct ring_buffer_event *event)
1506 unsigned long addr = (unsigned long)event;
1507 unsigned long index;
1509 index = rb_event_index(event);
1512 return cpu_buffer->commit_page->page == (void *)addr &&
1513 rb_commit_index(cpu_buffer) == index;
1517 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1519 unsigned long max_count;
1522 * We only race with interrupts and NMIs on this CPU.
1523 * If we own the commit event, then we can commit
1524 * all others that interrupted us, since the interruptions
1525 * are in stack format (they finish before they come
1526 * back to us). This allows us to do a simple loop to
1527 * assign the commit to the tail.
1530 max_count = cpu_buffer->buffer->pages * 100;
1532 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1533 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1535 if (RB_WARN_ON(cpu_buffer,
1536 rb_is_reader_page(cpu_buffer->tail_page)))
1538 local_set(&cpu_buffer->commit_page->page->commit,
1539 rb_page_write(cpu_buffer->commit_page));
1540 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1541 cpu_buffer->write_stamp =
1542 cpu_buffer->commit_page->page->time_stamp;
1543 /* add barrier to keep gcc from optimizing too much */
1546 while (rb_commit_index(cpu_buffer) !=
1547 rb_page_write(cpu_buffer->commit_page)) {
1549 local_set(&cpu_buffer->commit_page->page->commit,
1550 rb_page_write(cpu_buffer->commit_page));
1551 RB_WARN_ON(cpu_buffer,
1552 local_read(&cpu_buffer->commit_page->page->commit) &
1557 /* again, keep gcc from optimizing */
1561 * If an interrupt came in just after the first while loop
1562 * and pushed the tail page forward, we will be left with
1563 * a dangling commit that will never go forward.
1565 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1569 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1571 cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1572 cpu_buffer->reader_page->read = 0;
1575 static void rb_inc_iter(struct ring_buffer_iter *iter)
1577 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1580 * The iterator could be on the reader page (it starts there).
1581 * But the head could have moved, since the reader was
1582 * found. Check for this case and assign the iterator
1583 * to the head page instead of next.
1585 if (iter->head_page == cpu_buffer->reader_page)
1586 iter->head_page = rb_set_head_page(cpu_buffer);
1588 rb_inc_page(cpu_buffer, &iter->head_page);
1590 iter->read_stamp = iter->head_page->page->time_stamp;
1594 /* Slow path, do not inline */
1595 static noinline struct ring_buffer_event *
1596 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
1598 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
1600 /* Not the first event on the page? */
1601 if (rb_event_index(event)) {
1602 event->time_delta = delta & TS_MASK;
1603 event->array[0] = delta >> TS_SHIFT;
1605 /* nope, just zero it */
1606 event->time_delta = 0;
1607 event->array[0] = 0;
1610 return skip_time_extend(event);
1614 * ring_buffer_update_event - update event type and data
1615 * @event: the even to update
1616 * @type: the type of event
1617 * @length: the size of the event field in the ring buffer
1619 * Update the type and data fields of the event. The length
1620 * is the actual size that is written to the ring buffer,
1621 * and with this, we can determine what to place into the
1625 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
1626 struct ring_buffer_event *event, unsigned length,
1627 int add_timestamp, u64 delta)
1629 /* Only a commit updates the timestamp */
1630 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
1634 * If we need to add a timestamp, then we
1635 * add it to the start of the resevered space.
1637 if (unlikely(add_timestamp)) {
1638 event = rb_add_time_stamp(event, delta);
1639 length -= RB_LEN_TIME_EXTEND;
1643 event->time_delta = delta;
1644 length -= RB_EVNT_HDR_SIZE;
1645 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
1646 event->type_len = 0;
1647 event->array[0] = length;
1649 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
1653 * rb_handle_head_page - writer hit the head page
1655 * Returns: +1 to retry page
1660 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1661 struct buffer_page *tail_page,
1662 struct buffer_page *next_page)
1664 struct buffer_page *new_head;
1669 entries = rb_page_entries(next_page);
1672 * The hard part is here. We need to move the head
1673 * forward, and protect against both readers on
1674 * other CPUs and writers coming in via interrupts.
1676 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1680 * type can be one of four:
1681 * NORMAL - an interrupt already moved it for us
1682 * HEAD - we are the first to get here.
1683 * UPDATE - we are the interrupt interrupting
1685 * MOVED - a reader on another CPU moved the next
1686 * pointer to its reader page. Give up
1693 * We changed the head to UPDATE, thus
1694 * it is our responsibility to update
1697 local_add(entries, &cpu_buffer->overrun);
1700 * The entries will be zeroed out when we move the
1704 /* still more to do */
1707 case RB_PAGE_UPDATE:
1709 * This is an interrupt that interrupt the
1710 * previous update. Still more to do.
1713 case RB_PAGE_NORMAL:
1715 * An interrupt came in before the update
1716 * and processed this for us.
1717 * Nothing left to do.
1722 * The reader is on another CPU and just did
1723 * a swap with our next_page.
1728 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1733 * Now that we are here, the old head pointer is
1734 * set to UPDATE. This will keep the reader from
1735 * swapping the head page with the reader page.
1736 * The reader (on another CPU) will spin till
1739 * We just need to protect against interrupts
1740 * doing the job. We will set the next pointer
1741 * to HEAD. After that, we set the old pointer
1742 * to NORMAL, but only if it was HEAD before.
1743 * otherwise we are an interrupt, and only
1744 * want the outer most commit to reset it.
1746 new_head = next_page;
1747 rb_inc_page(cpu_buffer, &new_head);
1749 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1753 * Valid returns are:
1754 * HEAD - an interrupt came in and already set it.
1755 * NORMAL - One of two things:
1756 * 1) We really set it.
1757 * 2) A bunch of interrupts came in and moved
1758 * the page forward again.
1762 case RB_PAGE_NORMAL:
1766 RB_WARN_ON(cpu_buffer, 1);
1771 * It is possible that an interrupt came in,
1772 * set the head up, then more interrupts came in
1773 * and moved it again. When we get back here,
1774 * the page would have been set to NORMAL but we
1775 * just set it back to HEAD.
1777 * How do you detect this? Well, if that happened
1778 * the tail page would have moved.
1780 if (ret == RB_PAGE_NORMAL) {
1782 * If the tail had moved passed next, then we need
1783 * to reset the pointer.
1785 if (cpu_buffer->tail_page != tail_page &&
1786 cpu_buffer->tail_page != next_page)
1787 rb_head_page_set_normal(cpu_buffer, new_head,
1793 * If this was the outer most commit (the one that
1794 * changed the original pointer from HEAD to UPDATE),
1795 * then it is up to us to reset it to NORMAL.
1797 if (type == RB_PAGE_HEAD) {
1798 ret = rb_head_page_set_normal(cpu_buffer, next_page,
1801 if (RB_WARN_ON(cpu_buffer,
1802 ret != RB_PAGE_UPDATE))
1809 static unsigned rb_calculate_event_length(unsigned length)
1811 struct ring_buffer_event event; /* Used only for sizeof array */
1813 /* zero length can cause confusions */
1817 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
1818 length += sizeof(event.array[0]);
1820 length += RB_EVNT_HDR_SIZE;
1821 length = ALIGN(length, RB_ARCH_ALIGNMENT);
1827 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
1828 struct buffer_page *tail_page,
1829 unsigned long tail, unsigned long length)
1831 struct ring_buffer_event *event;
1834 * Only the event that crossed the page boundary
1835 * must fill the old tail_page with padding.
1837 if (tail >= BUF_PAGE_SIZE) {
1839 * If the page was filled, then we still need
1840 * to update the real_end. Reset it to zero
1841 * and the reader will ignore it.
1843 if (tail == BUF_PAGE_SIZE)
1844 tail_page->real_end = 0;
1846 local_sub(length, &tail_page->write);
1850 event = __rb_page_index(tail_page, tail);
1851 kmemcheck_annotate_bitfield(event, bitfield);
1854 * Save the original length to the meta data.
1855 * This will be used by the reader to add lost event
1858 tail_page->real_end = tail;
1861 * If this event is bigger than the minimum size, then
1862 * we need to be careful that we don't subtract the
1863 * write counter enough to allow another writer to slip
1865 * We put in a discarded commit instead, to make sure
1866 * that this space is not used again.
1868 * If we are less than the minimum size, we don't need to
1871 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
1872 /* No room for any events */
1874 /* Mark the rest of the page with padding */
1875 rb_event_set_padding(event);
1877 /* Set the write back to the previous setting */
1878 local_sub(length, &tail_page->write);
1882 /* Put in a discarded event */
1883 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
1884 event->type_len = RINGBUF_TYPE_PADDING;
1885 /* time delta must be non zero */
1886 event->time_delta = 1;
1888 /* Set write to end of buffer */
1889 length = (tail + length) - BUF_PAGE_SIZE;
1890 local_sub(length, &tail_page->write);
1894 * This is the slow path, force gcc not to inline it.
1896 static noinline struct ring_buffer_event *
1897 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
1898 unsigned long length, unsigned long tail,
1899 struct buffer_page *tail_page, u64 ts)
1901 struct buffer_page *commit_page = cpu_buffer->commit_page;
1902 struct ring_buffer *buffer = cpu_buffer->buffer;
1903 struct buffer_page *next_page;
1906 next_page = tail_page;
1908 rb_inc_page(cpu_buffer, &next_page);
1911 * If for some reason, we had an interrupt storm that made
1912 * it all the way around the buffer, bail, and warn
1915 if (unlikely(next_page == commit_page)) {
1916 local_inc(&cpu_buffer->commit_overrun);
1921 * This is where the fun begins!
1923 * We are fighting against races between a reader that
1924 * could be on another CPU trying to swap its reader
1925 * page with the buffer head.
1927 * We are also fighting against interrupts coming in and
1928 * moving the head or tail on us as well.
1930 * If the next page is the head page then we have filled
1931 * the buffer, unless the commit page is still on the
1934 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
1937 * If the commit is not on the reader page, then
1938 * move the header page.
1940 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
1942 * If we are not in overwrite mode,
1943 * this is easy, just stop here.
1945 if (!(buffer->flags & RB_FL_OVERWRITE))
1948 ret = rb_handle_head_page(cpu_buffer,
1957 * We need to be careful here too. The
1958 * commit page could still be on the reader
1959 * page. We could have a small buffer, and
1960 * have filled up the buffer with events
1961 * from interrupts and such, and wrapped.
1963 * Note, if the tail page is also the on the
1964 * reader_page, we let it move out.
1966 if (unlikely((cpu_buffer->commit_page !=
1967 cpu_buffer->tail_page) &&
1968 (cpu_buffer->commit_page ==
1969 cpu_buffer->reader_page))) {
1970 local_inc(&cpu_buffer->commit_overrun);
1976 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
1979 * Nested commits always have zero deltas, so
1980 * just reread the time stamp
1982 ts = rb_time_stamp(buffer);
1983 next_page->page->time_stamp = ts;
1988 rb_reset_tail(cpu_buffer, tail_page, tail, length);
1990 /* fail and let the caller try again */
1991 return ERR_PTR(-EAGAIN);
1995 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2000 static struct ring_buffer_event *
2001 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2002 unsigned long length, u64 ts,
2003 u64 delta, int add_timestamp)
2005 struct buffer_page *tail_page;
2006 struct ring_buffer_event *event;
2007 unsigned long tail, write;
2010 * If the time delta since the last event is too big to
2011 * hold in the time field of the event, then we append a
2012 * TIME EXTEND event ahead of the data event.
2014 if (unlikely(add_timestamp))
2015 length += RB_LEN_TIME_EXTEND;
2017 tail_page = cpu_buffer->tail_page;
2018 write = local_add_return(length, &tail_page->write);
2020 /* set write to only the index of the write */
2021 write &= RB_WRITE_MASK;
2022 tail = write - length;
2024 /* See if we shot pass the end of this buffer page */
2025 if (unlikely(write > BUF_PAGE_SIZE))
2026 return rb_move_tail(cpu_buffer, length, tail,
2029 /* We reserved something on the buffer */
2031 event = __rb_page_index(tail_page, tail);
2032 kmemcheck_annotate_bitfield(event, bitfield);
2033 rb_update_event(cpu_buffer, event, length, add_timestamp, delta);
2035 local_inc(&tail_page->entries);
2038 * If this is the first commit on the page, then update
2042 tail_page->page->time_stamp = ts;
2048 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2049 struct ring_buffer_event *event)
2051 unsigned long new_index, old_index;
2052 struct buffer_page *bpage;
2053 unsigned long index;
2056 new_index = rb_event_index(event);
2057 old_index = new_index + rb_event_ts_length(event);
2058 addr = (unsigned long)event;
2061 bpage = cpu_buffer->tail_page;
2063 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2064 unsigned long write_mask =
2065 local_read(&bpage->write) & ~RB_WRITE_MASK;
2067 * This is on the tail page. It is possible that
2068 * a write could come in and move the tail page
2069 * and write to the next page. That is fine
2070 * because we just shorten what is on this page.
2072 old_index += write_mask;
2073 new_index += write_mask;
2074 index = local_cmpxchg(&bpage->write, old_index, new_index);
2075 if (index == old_index)
2079 /* could not discard */
2083 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2085 local_inc(&cpu_buffer->committing);
2086 local_inc(&cpu_buffer->commits);
2089 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2091 unsigned long commits;
2093 if (RB_WARN_ON(cpu_buffer,
2094 !local_read(&cpu_buffer->committing)))
2098 commits = local_read(&cpu_buffer->commits);
2099 /* synchronize with interrupts */
2101 if (local_read(&cpu_buffer->committing) == 1)
2102 rb_set_commit_to_write(cpu_buffer);
2104 local_dec(&cpu_buffer->committing);
2106 /* synchronize with interrupts */
2110 * Need to account for interrupts coming in between the
2111 * updating of the commit page and the clearing of the
2112 * committing counter.
2114 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2115 !local_read(&cpu_buffer->committing)) {
2116 local_inc(&cpu_buffer->committing);
2121 static struct ring_buffer_event *
2122 rb_reserve_next_event(struct ring_buffer *buffer,
2123 struct ring_buffer_per_cpu *cpu_buffer,
2124 unsigned long length)
2126 struct ring_buffer_event *event;
2132 rb_start_commit(cpu_buffer);
2134 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2136 * Due to the ability to swap a cpu buffer from a buffer
2137 * it is possible it was swapped before we committed.
2138 * (committing stops a swap). We check for it here and
2139 * if it happened, we have to fail the write.
2142 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2143 local_dec(&cpu_buffer->committing);
2144 local_dec(&cpu_buffer->commits);
2149 length = rb_calculate_event_length(length);
2155 * We allow for interrupts to reenter here and do a trace.
2156 * If one does, it will cause this original code to loop
2157 * back here. Even with heavy interrupts happening, this
2158 * should only happen a few times in a row. If this happens
2159 * 1000 times in a row, there must be either an interrupt
2160 * storm or we have something buggy.
2163 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2166 ts = rb_time_stamp(cpu_buffer->buffer);
2167 diff = ts - cpu_buffer->write_stamp;
2169 /* make sure this diff is calculated here */
2172 /* Did the write stamp get updated already? */
2173 if (likely(ts >= cpu_buffer->write_stamp)) {
2175 if (unlikely(test_time_stamp(delta))) {
2176 WARN_ONCE(delta > (1ULL << 59),
2177 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n",
2178 (unsigned long long)delta,
2179 (unsigned long long)ts,
2180 (unsigned long long)cpu_buffer->write_stamp);
2185 event = __rb_reserve_next(cpu_buffer, length, ts,
2186 delta, add_timestamp);
2187 if (unlikely(PTR_ERR(event) == -EAGAIN))
2196 rb_end_commit(cpu_buffer);
2200 #ifdef CONFIG_TRACING
2202 #define TRACE_RECURSIVE_DEPTH 16
2204 /* Keep this code out of the fast path cache */
2205 static noinline void trace_recursive_fail(void)
2207 /* Disable all tracing before we do anything else */
2208 tracing_off_permanent();
2210 printk_once(KERN_WARNING "Tracing recursion: depth[%ld]:"
2211 "HC[%lu]:SC[%lu]:NMI[%lu]\n",
2212 current->trace_recursion,
2213 hardirq_count() >> HARDIRQ_SHIFT,
2214 softirq_count() >> SOFTIRQ_SHIFT,
2220 static inline int trace_recursive_lock(void)
2222 current->trace_recursion++;
2224 if (likely(current->trace_recursion < TRACE_RECURSIVE_DEPTH))
2227 trace_recursive_fail();
2232 static inline void trace_recursive_unlock(void)
2234 WARN_ON_ONCE(!current->trace_recursion);
2236 current->trace_recursion--;
2241 #define trace_recursive_lock() (0)
2242 #define trace_recursive_unlock() do { } while (0)
2247 * ring_buffer_lock_reserve - reserve a part of the buffer
2248 * @buffer: the ring buffer to reserve from
2249 * @length: the length of the data to reserve (excluding event header)
2251 * Returns a reseverd event on the ring buffer to copy directly to.
2252 * The user of this interface will need to get the body to write into
2253 * and can use the ring_buffer_event_data() interface.
2255 * The length is the length of the data needed, not the event length
2256 * which also includes the event header.
2258 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2259 * If NULL is returned, then nothing has been allocated or locked.
2261 struct ring_buffer_event *
2262 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2264 struct ring_buffer_per_cpu *cpu_buffer;
2265 struct ring_buffer_event *event;
2268 if (ring_buffer_flags != RB_BUFFERS_ON)
2271 /* If we are tracing schedule, we don't want to recurse */
2272 preempt_disable_notrace();
2274 if (atomic_read(&buffer->record_disabled))
2277 if (trace_recursive_lock())
2280 cpu = raw_smp_processor_id();
2282 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2285 cpu_buffer = buffer->buffers[cpu];
2287 if (atomic_read(&cpu_buffer->record_disabled))
2290 if (length > BUF_MAX_DATA_SIZE)
2293 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2300 trace_recursive_unlock();
2303 preempt_enable_notrace();
2306 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2309 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2310 struct ring_buffer_event *event)
2315 * The event first in the commit queue updates the
2318 if (rb_event_is_commit(cpu_buffer, event)) {
2320 * A commit event that is first on a page
2321 * updates the write timestamp with the page stamp
2323 if (!rb_event_index(event))
2324 cpu_buffer->write_stamp =
2325 cpu_buffer->commit_page->page->time_stamp;
2326 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2327 delta = event->array[0];
2329 delta += event->time_delta;
2330 cpu_buffer->write_stamp += delta;
2332 cpu_buffer->write_stamp += event->time_delta;
2336 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2337 struct ring_buffer_event *event)
2339 local_inc(&cpu_buffer->entries);
2340 rb_update_write_stamp(cpu_buffer, event);
2341 rb_end_commit(cpu_buffer);
2345 * ring_buffer_unlock_commit - commit a reserved
2346 * @buffer: The buffer to commit to
2347 * @event: The event pointer to commit.
2349 * This commits the data to the ring buffer, and releases any locks held.
2351 * Must be paired with ring_buffer_lock_reserve.
2353 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2354 struct ring_buffer_event *event)
2356 struct ring_buffer_per_cpu *cpu_buffer;
2357 int cpu = raw_smp_processor_id();
2359 cpu_buffer = buffer->buffers[cpu];
2361 rb_commit(cpu_buffer, event);
2363 trace_recursive_unlock();
2365 preempt_enable_notrace();
2369 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2371 static inline void rb_event_discard(struct ring_buffer_event *event)
2373 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2374 event = skip_time_extend(event);
2376 /* array[0] holds the actual length for the discarded event */
2377 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2378 event->type_len = RINGBUF_TYPE_PADDING;
2379 /* time delta must be non zero */
2380 if (!event->time_delta)
2381 event->time_delta = 1;
2385 * Decrement the entries to the page that an event is on.
2386 * The event does not even need to exist, only the pointer
2387 * to the page it is on. This may only be called before the commit
2391 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2392 struct ring_buffer_event *event)
2394 unsigned long addr = (unsigned long)event;
2395 struct buffer_page *bpage = cpu_buffer->commit_page;
2396 struct buffer_page *start;
2400 /* Do the likely case first */
2401 if (likely(bpage->page == (void *)addr)) {
2402 local_dec(&bpage->entries);
2407 * Because the commit page may be on the reader page we
2408 * start with the next page and check the end loop there.
2410 rb_inc_page(cpu_buffer, &bpage);
2413 if (bpage->page == (void *)addr) {
2414 local_dec(&bpage->entries);
2417 rb_inc_page(cpu_buffer, &bpage);
2418 } while (bpage != start);
2420 /* commit not part of this buffer?? */
2421 RB_WARN_ON(cpu_buffer, 1);
2425 * ring_buffer_commit_discard - discard an event that has not been committed
2426 * @buffer: the ring buffer
2427 * @event: non committed event to discard
2429 * Sometimes an event that is in the ring buffer needs to be ignored.
2430 * This function lets the user discard an event in the ring buffer
2431 * and then that event will not be read later.
2433 * This function only works if it is called before the the item has been
2434 * committed. It will try to free the event from the ring buffer
2435 * if another event has not been added behind it.
2437 * If another event has been added behind it, it will set the event
2438 * up as discarded, and perform the commit.
2440 * If this function is called, do not call ring_buffer_unlock_commit on
2443 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2444 struct ring_buffer_event *event)
2446 struct ring_buffer_per_cpu *cpu_buffer;
2449 /* The event is discarded regardless */
2450 rb_event_discard(event);
2452 cpu = smp_processor_id();
2453 cpu_buffer = buffer->buffers[cpu];
2456 * This must only be called if the event has not been
2457 * committed yet. Thus we can assume that preemption
2458 * is still disabled.
2460 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2462 rb_decrement_entry(cpu_buffer, event);
2463 if (rb_try_to_discard(cpu_buffer, event))
2467 * The commit is still visible by the reader, so we
2468 * must still update the timestamp.
2470 rb_update_write_stamp(cpu_buffer, event);
2472 rb_end_commit(cpu_buffer);
2474 trace_recursive_unlock();
2476 preempt_enable_notrace();
2479 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2482 * ring_buffer_write - write data to the buffer without reserving
2483 * @buffer: The ring buffer to write to.
2484 * @length: The length of the data being written (excluding the event header)
2485 * @data: The data to write to the buffer.
2487 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2488 * one function. If you already have the data to write to the buffer, it
2489 * may be easier to simply call this function.
2491 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2492 * and not the length of the event which would hold the header.
2494 int ring_buffer_write(struct ring_buffer *buffer,
2495 unsigned long length,
2498 struct ring_buffer_per_cpu *cpu_buffer;
2499 struct ring_buffer_event *event;
2504 if (ring_buffer_flags != RB_BUFFERS_ON)
2507 preempt_disable_notrace();
2509 if (atomic_read(&buffer->record_disabled))
2512 cpu = raw_smp_processor_id();
2514 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2517 cpu_buffer = buffer->buffers[cpu];
2519 if (atomic_read(&cpu_buffer->record_disabled))
2522 if (length > BUF_MAX_DATA_SIZE)
2525 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2529 body = rb_event_data(event);
2531 memcpy(body, data, length);
2533 rb_commit(cpu_buffer, event);
2537 preempt_enable_notrace();
2541 EXPORT_SYMBOL_GPL(ring_buffer_write);
2543 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
2545 struct buffer_page *reader = cpu_buffer->reader_page;
2546 struct buffer_page *head = rb_set_head_page(cpu_buffer);
2547 struct buffer_page *commit = cpu_buffer->commit_page;
2549 /* In case of error, head will be NULL */
2550 if (unlikely(!head))
2553 return reader->read == rb_page_commit(reader) &&
2554 (commit == reader ||
2556 head->read == rb_page_commit(commit)));
2560 * ring_buffer_record_disable - stop all writes into the buffer
2561 * @buffer: The ring buffer to stop writes to.
2563 * This prevents all writes to the buffer. Any attempt to write
2564 * to the buffer after this will fail and return NULL.
2566 * The caller should call synchronize_sched() after this.
2568 void ring_buffer_record_disable(struct ring_buffer *buffer)
2570 atomic_inc(&buffer->record_disabled);
2572 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
2575 * ring_buffer_record_enable - enable writes to the buffer
2576 * @buffer: The ring buffer to enable writes
2578 * Note, multiple disables will need the same number of enables
2579 * to truly enable the writing (much like preempt_disable).
2581 void ring_buffer_record_enable(struct ring_buffer *buffer)
2583 atomic_dec(&buffer->record_disabled);
2585 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
2588 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
2589 * @buffer: The ring buffer to stop writes to.
2590 * @cpu: The CPU buffer to stop
2592 * This prevents all writes to the buffer. Any attempt to write
2593 * to the buffer after this will fail and return NULL.
2595 * The caller should call synchronize_sched() after this.
2597 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
2599 struct ring_buffer_per_cpu *cpu_buffer;
2601 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2604 cpu_buffer = buffer->buffers[cpu];
2605 atomic_inc(&cpu_buffer->record_disabled);
2607 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
2610 * ring_buffer_record_enable_cpu - enable writes to the buffer
2611 * @buffer: The ring buffer to enable writes
2612 * @cpu: The CPU to enable.
2614 * Note, multiple disables will need the same number of enables
2615 * to truly enable the writing (much like preempt_disable).
2617 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
2619 struct ring_buffer_per_cpu *cpu_buffer;
2621 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2624 cpu_buffer = buffer->buffers[cpu];
2625 atomic_dec(&cpu_buffer->record_disabled);
2627 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
2630 * The total entries in the ring buffer is the running counter
2631 * of entries entered into the ring buffer, minus the sum of
2632 * the entries read from the ring buffer and the number of
2633 * entries that were overwritten.
2635 static inline unsigned long
2636 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
2638 return local_read(&cpu_buffer->entries) -
2639 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
2643 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
2644 * @buffer: The ring buffer
2645 * @cpu: The per CPU buffer to get the entries from.
2647 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
2649 struct ring_buffer_per_cpu *cpu_buffer;
2651 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2654 cpu_buffer = buffer->buffers[cpu];
2656 return rb_num_of_entries(cpu_buffer);
2658 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
2661 * ring_buffer_overrun_cpu - get the number of overruns in a cpu_buffer
2662 * @buffer: The ring buffer
2663 * @cpu: The per CPU buffer to get the number of overruns from
2665 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
2667 struct ring_buffer_per_cpu *cpu_buffer;
2670 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2673 cpu_buffer = buffer->buffers[cpu];
2674 ret = local_read(&cpu_buffer->overrun);
2678 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
2681 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by commits
2682 * @buffer: The ring buffer
2683 * @cpu: The per CPU buffer to get the number of overruns from
2686 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
2688 struct ring_buffer_per_cpu *cpu_buffer;
2691 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2694 cpu_buffer = buffer->buffers[cpu];
2695 ret = local_read(&cpu_buffer->commit_overrun);
2699 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
2702 * ring_buffer_entries - get the number of entries in a buffer
2703 * @buffer: The ring buffer
2705 * Returns the total number of entries in the ring buffer
2708 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
2710 struct ring_buffer_per_cpu *cpu_buffer;
2711 unsigned long entries = 0;
2714 /* if you care about this being correct, lock the buffer */
2715 for_each_buffer_cpu(buffer, cpu) {
2716 cpu_buffer = buffer->buffers[cpu];
2717 entries += rb_num_of_entries(cpu_buffer);
2722 EXPORT_SYMBOL_GPL(ring_buffer_entries);
2725 * ring_buffer_overruns - get the number of overruns in buffer
2726 * @buffer: The ring buffer
2728 * Returns the total number of overruns in the ring buffer
2731 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
2733 struct ring_buffer_per_cpu *cpu_buffer;
2734 unsigned long overruns = 0;
2737 /* if you care about this being correct, lock the buffer */
2738 for_each_buffer_cpu(buffer, cpu) {
2739 cpu_buffer = buffer->buffers[cpu];
2740 overruns += local_read(&cpu_buffer->overrun);
2745 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
2747 static void rb_iter_reset(struct ring_buffer_iter *iter)
2749 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2751 /* Iterator usage is expected to have record disabled */
2752 if (list_empty(&cpu_buffer->reader_page->list)) {
2753 iter->head_page = rb_set_head_page(cpu_buffer);
2754 if (unlikely(!iter->head_page))
2756 iter->head = iter->head_page->read;
2758 iter->head_page = cpu_buffer->reader_page;
2759 iter->head = cpu_buffer->reader_page->read;
2762 iter->read_stamp = cpu_buffer->read_stamp;
2764 iter->read_stamp = iter->head_page->page->time_stamp;
2765 iter->cache_reader_page = cpu_buffer->reader_page;
2766 iter->cache_read = cpu_buffer->read;
2770 * ring_buffer_iter_reset - reset an iterator
2771 * @iter: The iterator to reset
2773 * Resets the iterator, so that it will start from the beginning
2776 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
2778 struct ring_buffer_per_cpu *cpu_buffer;
2779 unsigned long flags;
2784 cpu_buffer = iter->cpu_buffer;
2786 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2787 rb_iter_reset(iter);
2788 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2790 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
2793 * ring_buffer_iter_empty - check if an iterator has no more to read
2794 * @iter: The iterator to check
2796 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
2798 struct ring_buffer_per_cpu *cpu_buffer;
2800 cpu_buffer = iter->cpu_buffer;
2802 return iter->head_page == cpu_buffer->commit_page &&
2803 iter->head == rb_commit_index(cpu_buffer);
2805 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
2808 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2809 struct ring_buffer_event *event)
2813 switch (event->type_len) {
2814 case RINGBUF_TYPE_PADDING:
2817 case RINGBUF_TYPE_TIME_EXTEND:
2818 delta = event->array[0];
2820 delta += event->time_delta;
2821 cpu_buffer->read_stamp += delta;
2824 case RINGBUF_TYPE_TIME_STAMP:
2825 /* FIXME: not implemented */
2828 case RINGBUF_TYPE_DATA:
2829 cpu_buffer->read_stamp += event->time_delta;
2839 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
2840 struct ring_buffer_event *event)
2844 switch (event->type_len) {
2845 case RINGBUF_TYPE_PADDING:
2848 case RINGBUF_TYPE_TIME_EXTEND:
2849 delta = event->array[0];
2851 delta += event->time_delta;
2852 iter->read_stamp += delta;
2855 case RINGBUF_TYPE_TIME_STAMP:
2856 /* FIXME: not implemented */
2859 case RINGBUF_TYPE_DATA:
2860 iter->read_stamp += event->time_delta;
2869 static struct buffer_page *
2870 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
2872 struct buffer_page *reader = NULL;
2873 unsigned long overwrite;
2874 unsigned long flags;
2878 local_irq_save(flags);
2879 arch_spin_lock(&cpu_buffer->lock);
2883 * This should normally only loop twice. But because the
2884 * start of the reader inserts an empty page, it causes
2885 * a case where we will loop three times. There should be no
2886 * reason to loop four times (that I know of).
2888 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
2893 reader = cpu_buffer->reader_page;
2895 /* If there's more to read, return this page */
2896 if (cpu_buffer->reader_page->read < rb_page_size(reader))
2899 /* Never should we have an index greater than the size */
2900 if (RB_WARN_ON(cpu_buffer,
2901 cpu_buffer->reader_page->read > rb_page_size(reader)))
2904 /* check if we caught up to the tail */
2906 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
2910 * Reset the reader page to size zero.
2912 local_set(&cpu_buffer->reader_page->write, 0);
2913 local_set(&cpu_buffer->reader_page->entries, 0);
2914 local_set(&cpu_buffer->reader_page->page->commit, 0);
2915 cpu_buffer->reader_page->real_end = 0;
2919 * Splice the empty reader page into the list around the head.
2921 reader = rb_set_head_page(cpu_buffer);
2922 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
2923 cpu_buffer->reader_page->list.prev = reader->list.prev;
2926 * cpu_buffer->pages just needs to point to the buffer, it
2927 * has no specific buffer page to point to. Lets move it out
2928 * of our way so we don't accidently swap it.
2930 cpu_buffer->pages = reader->list.prev;
2932 /* The reader page will be pointing to the new head */
2933 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
2936 * We want to make sure we read the overruns after we set up our
2937 * pointers to the next object. The writer side does a
2938 * cmpxchg to cross pages which acts as the mb on the writer
2939 * side. Note, the reader will constantly fail the swap
2940 * while the writer is updating the pointers, so this
2941 * guarantees that the overwrite recorded here is the one we
2942 * want to compare with the last_overrun.
2945 overwrite = local_read(&(cpu_buffer->overrun));
2948 * Here's the tricky part.
2950 * We need to move the pointer past the header page.
2951 * But we can only do that if a writer is not currently
2952 * moving it. The page before the header page has the
2953 * flag bit '1' set if it is pointing to the page we want.
2954 * but if the writer is in the process of moving it
2955 * than it will be '2' or already moved '0'.
2958 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
2961 * If we did not convert it, then we must try again.
2967 * Yeah! We succeeded in replacing the page.
2969 * Now make the new head point back to the reader page.
2971 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
2972 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
2974 /* Finally update the reader page to the new head */
2975 cpu_buffer->reader_page = reader;
2976 rb_reset_reader_page(cpu_buffer);
2978 if (overwrite != cpu_buffer->last_overrun) {
2979 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
2980 cpu_buffer->last_overrun = overwrite;
2986 arch_spin_unlock(&cpu_buffer->lock);
2987 local_irq_restore(flags);
2992 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
2994 struct ring_buffer_event *event;
2995 struct buffer_page *reader;
2998 reader = rb_get_reader_page(cpu_buffer);
3000 /* This function should not be called when buffer is empty */
3001 if (RB_WARN_ON(cpu_buffer, !reader))
3004 event = rb_reader_event(cpu_buffer);
3006 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3009 rb_update_read_stamp(cpu_buffer, event);
3011 length = rb_event_length(event);
3012 cpu_buffer->reader_page->read += length;
3015 static void rb_advance_iter(struct ring_buffer_iter *iter)
3017 struct ring_buffer_per_cpu *cpu_buffer;
3018 struct ring_buffer_event *event;
3021 cpu_buffer = iter->cpu_buffer;
3024 * Check if we are at the end of the buffer.
3026 if (iter->head >= rb_page_size(iter->head_page)) {
3027 /* discarded commits can make the page empty */
3028 if (iter->head_page == cpu_buffer->commit_page)
3034 event = rb_iter_head_event(iter);
3036 length = rb_event_length(event);
3039 * This should not be called to advance the header if we are
3040 * at the tail of the buffer.
3042 if (RB_WARN_ON(cpu_buffer,
3043 (iter->head_page == cpu_buffer->commit_page) &&
3044 (iter->head + length > rb_commit_index(cpu_buffer))))
3047 rb_update_iter_read_stamp(iter, event);
3049 iter->head += length;
3051 /* check for end of page padding */
3052 if ((iter->head >= rb_page_size(iter->head_page)) &&
3053 (iter->head_page != cpu_buffer->commit_page))
3054 rb_advance_iter(iter);
3057 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3059 return cpu_buffer->lost_events;
3062 static struct ring_buffer_event *
3063 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3064 unsigned long *lost_events)
3066 struct ring_buffer_event *event;
3067 struct buffer_page *reader;
3072 * We repeat when a time extend is encountered.
3073 * Since the time extend is always attached to a data event,
3074 * we should never loop more than once.
3075 * (We never hit the following condition more than twice).
3077 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3080 reader = rb_get_reader_page(cpu_buffer);
3084 event = rb_reader_event(cpu_buffer);
3086 switch (event->type_len) {
3087 case RINGBUF_TYPE_PADDING:
3088 if (rb_null_event(event))
3089 RB_WARN_ON(cpu_buffer, 1);
3091 * Because the writer could be discarding every
3092 * event it creates (which would probably be bad)
3093 * if we were to go back to "again" then we may never
3094 * catch up, and will trigger the warn on, or lock
3095 * the box. Return the padding, and we will release
3096 * the current locks, and try again.
3100 case RINGBUF_TYPE_TIME_EXTEND:
3101 /* Internal data, OK to advance */
3102 rb_advance_reader(cpu_buffer);
3105 case RINGBUF_TYPE_TIME_STAMP:
3106 /* FIXME: not implemented */
3107 rb_advance_reader(cpu_buffer);
3110 case RINGBUF_TYPE_DATA:
3112 *ts = cpu_buffer->read_stamp + event->time_delta;
3113 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3114 cpu_buffer->cpu, ts);
3117 *lost_events = rb_lost_events(cpu_buffer);
3126 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3128 static struct ring_buffer_event *
3129 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3131 struct ring_buffer *buffer;
3132 struct ring_buffer_per_cpu *cpu_buffer;
3133 struct ring_buffer_event *event;
3136 cpu_buffer = iter->cpu_buffer;
3137 buffer = cpu_buffer->buffer;
3140 * Check if someone performed a consuming read to
3141 * the buffer. A consuming read invalidates the iterator
3142 * and we need to reset the iterator in this case.
3144 if (unlikely(iter->cache_read != cpu_buffer->read ||
3145 iter->cache_reader_page != cpu_buffer->reader_page))
3146 rb_iter_reset(iter);
3149 if (ring_buffer_iter_empty(iter))
3153 * We repeat when a time extend is encountered.
3154 * Since the time extend is always attached to a data event,
3155 * we should never loop more than once.
3156 * (We never hit the following condition more than twice).
3158 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3161 if (rb_per_cpu_empty(cpu_buffer))
3164 if (iter->head >= local_read(&iter->head_page->page->commit)) {
3169 event = rb_iter_head_event(iter);
3171 switch (event->type_len) {
3172 case RINGBUF_TYPE_PADDING:
3173 if (rb_null_event(event)) {
3177 rb_advance_iter(iter);
3180 case RINGBUF_TYPE_TIME_EXTEND:
3181 /* Internal data, OK to advance */
3182 rb_advance_iter(iter);
3185 case RINGBUF_TYPE_TIME_STAMP:
3186 /* FIXME: not implemented */
3187 rb_advance_iter(iter);
3190 case RINGBUF_TYPE_DATA:
3192 *ts = iter->read_stamp + event->time_delta;
3193 ring_buffer_normalize_time_stamp(buffer,
3194 cpu_buffer->cpu, ts);
3204 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3206 static inline int rb_ok_to_lock(void)
3209 * If an NMI die dumps out the content of the ring buffer
3210 * do not grab locks. We also permanently disable the ring
3211 * buffer too. A one time deal is all you get from reading
3212 * the ring buffer from an NMI.
3214 if (likely(!in_nmi()))
3217 tracing_off_permanent();
3222 * ring_buffer_peek - peek at the next event to be read
3223 * @buffer: The ring buffer to read
3224 * @cpu: The cpu to peak at
3225 * @ts: The timestamp counter of this event.
3226 * @lost_events: a variable to store if events were lost (may be NULL)
3228 * This will return the event that will be read next, but does
3229 * not consume the data.
3231 struct ring_buffer_event *
3232 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3233 unsigned long *lost_events)
3235 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3236 struct ring_buffer_event *event;
3237 unsigned long flags;
3240 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3243 dolock = rb_ok_to_lock();
3245 local_irq_save(flags);
3247 spin_lock(&cpu_buffer->reader_lock);
3248 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3249 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3250 rb_advance_reader(cpu_buffer);
3252 spin_unlock(&cpu_buffer->reader_lock);
3253 local_irq_restore(flags);
3255 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3262 * ring_buffer_iter_peek - peek at the next event to be read
3263 * @iter: The ring buffer iterator
3264 * @ts: The timestamp counter of this event.
3266 * This will return the event that will be read next, but does
3267 * not increment the iterator.
3269 struct ring_buffer_event *
3270 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3272 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3273 struct ring_buffer_event *event;
3274 unsigned long flags;
3277 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3278 event = rb_iter_peek(iter, ts);
3279 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3281 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3288 * ring_buffer_consume - return an event and consume it
3289 * @buffer: The ring buffer to get the next event from
3290 * @cpu: the cpu to read the buffer from
3291 * @ts: a variable to store the timestamp (may be NULL)
3292 * @lost_events: a variable to store if events were lost (may be NULL)
3294 * Returns the next event in the ring buffer, and that event is consumed.
3295 * Meaning, that sequential reads will keep returning a different event,
3296 * and eventually empty the ring buffer if the producer is slower.
3298 struct ring_buffer_event *
3299 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3300 unsigned long *lost_events)
3302 struct ring_buffer_per_cpu *cpu_buffer;
3303 struct ring_buffer_event *event = NULL;
3304 unsigned long flags;
3307 dolock = rb_ok_to_lock();
3310 /* might be called in atomic */
3313 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3316 cpu_buffer = buffer->buffers[cpu];
3317 local_irq_save(flags);
3319 spin_lock(&cpu_buffer->reader_lock);
3321 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3323 cpu_buffer->lost_events = 0;
3324 rb_advance_reader(cpu_buffer);
3328 spin_unlock(&cpu_buffer->reader_lock);
3329 local_irq_restore(flags);
3334 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3339 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3342 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3343 * @buffer: The ring buffer to read from
3344 * @cpu: The cpu buffer to iterate over
3346 * This performs the initial preparations necessary to iterate
3347 * through the buffer. Memory is allocated, buffer recording
3348 * is disabled, and the iterator pointer is returned to the caller.
3350 * Disabling buffer recordng prevents the reading from being
3351 * corrupted. This is not a consuming read, so a producer is not
3354 * After a sequence of ring_buffer_read_prepare calls, the user is
3355 * expected to make at least one call to ring_buffer_prepare_sync.
3356 * Afterwards, ring_buffer_read_start is invoked to get things going
3359 * This overall must be paired with ring_buffer_finish.
3361 struct ring_buffer_iter *
3362 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3364 struct ring_buffer_per_cpu *cpu_buffer;
3365 struct ring_buffer_iter *iter;
3367 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3370 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3374 cpu_buffer = buffer->buffers[cpu];
3376 iter->cpu_buffer = cpu_buffer;
3378 atomic_inc(&cpu_buffer->record_disabled);
3382 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
3385 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
3387 * All previously invoked ring_buffer_read_prepare calls to prepare
3388 * iterators will be synchronized. Afterwards, read_buffer_read_start
3389 * calls on those iterators are allowed.
3392 ring_buffer_read_prepare_sync(void)
3394 synchronize_sched();
3396 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
3399 * ring_buffer_read_start - start a non consuming read of the buffer
3400 * @iter: The iterator returned by ring_buffer_read_prepare
3402 * This finalizes the startup of an iteration through the buffer.
3403 * The iterator comes from a call to ring_buffer_read_prepare and
3404 * an intervening ring_buffer_read_prepare_sync must have been
3407 * Must be paired with ring_buffer_finish.
3410 ring_buffer_read_start(struct ring_buffer_iter *iter)
3412 struct ring_buffer_per_cpu *cpu_buffer;
3413 unsigned long flags;
3418 cpu_buffer = iter->cpu_buffer;
3420 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3421 arch_spin_lock(&cpu_buffer->lock);
3422 rb_iter_reset(iter);
3423 arch_spin_unlock(&cpu_buffer->lock);
3424 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3426 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
3429 * ring_buffer_finish - finish reading the iterator of the buffer
3430 * @iter: The iterator retrieved by ring_buffer_start
3432 * This re-enables the recording to the buffer, and frees the
3436 ring_buffer_read_finish(struct ring_buffer_iter *iter)
3438 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3440 atomic_dec(&cpu_buffer->record_disabled);
3443 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
3446 * ring_buffer_read - read the next item in the ring buffer by the iterator
3447 * @iter: The ring buffer iterator
3448 * @ts: The time stamp of the event read.
3450 * This reads the next event in the ring buffer and increments the iterator.
3452 struct ring_buffer_event *
3453 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
3455 struct ring_buffer_event *event;
3456 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3457 unsigned long flags;
3459 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3461 event = rb_iter_peek(iter, ts);
3465 if (event->type_len == RINGBUF_TYPE_PADDING)
3468 rb_advance_iter(iter);
3470 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3474 EXPORT_SYMBOL_GPL(ring_buffer_read);
3477 * ring_buffer_size - return the size of the ring buffer (in bytes)
3478 * @buffer: The ring buffer.
3480 unsigned long ring_buffer_size(struct ring_buffer *buffer)
3482 return BUF_PAGE_SIZE * buffer->pages;
3484 EXPORT_SYMBOL_GPL(ring_buffer_size);
3487 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
3489 rb_head_page_deactivate(cpu_buffer);
3491 cpu_buffer->head_page
3492 = list_entry(cpu_buffer->pages, struct buffer_page, list);
3493 local_set(&cpu_buffer->head_page->write, 0);
3494 local_set(&cpu_buffer->head_page->entries, 0);
3495 local_set(&cpu_buffer->head_page->page->commit, 0);
3497 cpu_buffer->head_page->read = 0;
3499 cpu_buffer->tail_page = cpu_buffer->head_page;
3500 cpu_buffer->commit_page = cpu_buffer->head_page;
3502 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
3503 local_set(&cpu_buffer->reader_page->write, 0);
3504 local_set(&cpu_buffer->reader_page->entries, 0);
3505 local_set(&cpu_buffer->reader_page->page->commit, 0);
3506 cpu_buffer->reader_page->read = 0;
3508 local_set(&cpu_buffer->commit_overrun, 0);
3509 local_set(&cpu_buffer->overrun, 0);
3510 local_set(&cpu_buffer->entries, 0);
3511 local_set(&cpu_buffer->committing, 0);
3512 local_set(&cpu_buffer->commits, 0);
3513 cpu_buffer->read = 0;
3515 cpu_buffer->write_stamp = 0;
3516 cpu_buffer->read_stamp = 0;
3518 cpu_buffer->lost_events = 0;
3519 cpu_buffer->last_overrun = 0;
3521 rb_head_page_activate(cpu_buffer);
3525 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
3526 * @buffer: The ring buffer to reset a per cpu buffer of
3527 * @cpu: The CPU buffer to be reset
3529 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
3531 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3532 unsigned long flags;
3534 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3537 atomic_inc(&cpu_buffer->record_disabled);
3539 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3541 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
3544 arch_spin_lock(&cpu_buffer->lock);
3546 rb_reset_cpu(cpu_buffer);
3548 arch_spin_unlock(&cpu_buffer->lock);
3551 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3553 atomic_dec(&cpu_buffer->record_disabled);
3555 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
3558 * ring_buffer_reset - reset a ring buffer
3559 * @buffer: The ring buffer to reset all cpu buffers
3561 void ring_buffer_reset(struct ring_buffer *buffer)
3565 for_each_buffer_cpu(buffer, cpu)
3566 ring_buffer_reset_cpu(buffer, cpu);
3568 EXPORT_SYMBOL_GPL(ring_buffer_reset);
3571 * rind_buffer_empty - is the ring buffer empty?
3572 * @buffer: The ring buffer to test
3574 int ring_buffer_empty(struct ring_buffer *buffer)
3576 struct ring_buffer_per_cpu *cpu_buffer;
3577 unsigned long flags;
3582 dolock = rb_ok_to_lock();
3584 /* yes this is racy, but if you don't like the race, lock the buffer */
3585 for_each_buffer_cpu(buffer, cpu) {
3586 cpu_buffer = buffer->buffers[cpu];
3587 local_irq_save(flags);
3589 spin_lock(&cpu_buffer->reader_lock);
3590 ret = rb_per_cpu_empty(cpu_buffer);
3592 spin_unlock(&cpu_buffer->reader_lock);
3593 local_irq_restore(flags);
3601 EXPORT_SYMBOL_GPL(ring_buffer_empty);
3604 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
3605 * @buffer: The ring buffer
3606 * @cpu: The CPU buffer to test
3608 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
3610 struct ring_buffer_per_cpu *cpu_buffer;
3611 unsigned long flags;
3615 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3618 dolock = rb_ok_to_lock();
3620 cpu_buffer = buffer->buffers[cpu];
3621 local_irq_save(flags);
3623 spin_lock(&cpu_buffer->reader_lock);
3624 ret = rb_per_cpu_empty(cpu_buffer);
3626 spin_unlock(&cpu_buffer->reader_lock);
3627 local_irq_restore(flags);
3631 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
3633 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3635 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
3636 * @buffer_a: One buffer to swap with
3637 * @buffer_b: The other buffer to swap with
3639 * This function is useful for tracers that want to take a "snapshot"
3640 * of a CPU buffer and has another back up buffer lying around.
3641 * it is expected that the tracer handles the cpu buffer not being
3642 * used at the moment.
3644 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
3645 struct ring_buffer *buffer_b, int cpu)
3647 struct ring_buffer_per_cpu *cpu_buffer_a;
3648 struct ring_buffer_per_cpu *cpu_buffer_b;
3651 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
3652 !cpumask_test_cpu(cpu, buffer_b->cpumask))
3655 /* At least make sure the two buffers are somewhat the same */
3656 if (buffer_a->pages != buffer_b->pages)
3661 if (ring_buffer_flags != RB_BUFFERS_ON)
3664 if (atomic_read(&buffer_a->record_disabled))
3667 if (atomic_read(&buffer_b->record_disabled))
3670 cpu_buffer_a = buffer_a->buffers[cpu];
3671 cpu_buffer_b = buffer_b->buffers[cpu];
3673 if (atomic_read(&cpu_buffer_a->record_disabled))
3676 if (atomic_read(&cpu_buffer_b->record_disabled))
3680 * We can't do a synchronize_sched here because this
3681 * function can be called in atomic context.
3682 * Normally this will be called from the same CPU as cpu.
3683 * If not it's up to the caller to protect this.
3685 atomic_inc(&cpu_buffer_a->record_disabled);
3686 atomic_inc(&cpu_buffer_b->record_disabled);
3689 if (local_read(&cpu_buffer_a->committing))
3691 if (local_read(&cpu_buffer_b->committing))
3694 buffer_a->buffers[cpu] = cpu_buffer_b;
3695 buffer_b->buffers[cpu] = cpu_buffer_a;
3697 cpu_buffer_b->buffer = buffer_a;
3698 cpu_buffer_a->buffer = buffer_b;
3703 atomic_dec(&cpu_buffer_a->record_disabled);
3704 atomic_dec(&cpu_buffer_b->record_disabled);
3708 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
3709 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
3712 * ring_buffer_alloc_read_page - allocate a page to read from buffer
3713 * @buffer: the buffer to allocate for.
3715 * This function is used in conjunction with ring_buffer_read_page.
3716 * When reading a full page from the ring buffer, these functions
3717 * can be used to speed up the process. The calling function should
3718 * allocate a few pages first with this function. Then when it
3719 * needs to get pages from the ring buffer, it passes the result
3720 * of this function into ring_buffer_read_page, which will swap
3721 * the page that was allocated, with the read page of the buffer.
3724 * The page allocated, or NULL on error.
3726 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer)
3728 struct buffer_data_page *bpage;
3731 addr = __get_free_page(GFP_KERNEL);
3735 bpage = (void *)addr;
3737 rb_init_page(bpage);
3741 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
3744 * ring_buffer_free_read_page - free an allocated read page
3745 * @buffer: the buffer the page was allocate for
3746 * @data: the page to free
3748 * Free a page allocated from ring_buffer_alloc_read_page.
3750 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
3752 free_page((unsigned long)data);
3754 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
3757 * ring_buffer_read_page - extract a page from the ring buffer
3758 * @buffer: buffer to extract from
3759 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
3760 * @len: amount to extract
3761 * @cpu: the cpu of the buffer to extract
3762 * @full: should the extraction only happen when the page is full.
3764 * This function will pull out a page from the ring buffer and consume it.
3765 * @data_page must be the address of the variable that was returned
3766 * from ring_buffer_alloc_read_page. This is because the page might be used
3767 * to swap with a page in the ring buffer.
3770 * rpage = ring_buffer_alloc_read_page(buffer);
3773 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
3775 * process_page(rpage, ret);
3777 * When @full is set, the function will not return true unless
3778 * the writer is off the reader page.
3780 * Note: it is up to the calling functions to handle sleeps and wakeups.
3781 * The ring buffer can be used anywhere in the kernel and can not
3782 * blindly call wake_up. The layer that uses the ring buffer must be
3783 * responsible for that.
3786 * >=0 if data has been transferred, returns the offset of consumed data.
3787 * <0 if no data has been transferred.
3789 int ring_buffer_read_page(struct ring_buffer *buffer,
3790 void **data_page, size_t len, int cpu, int full)
3792 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3793 struct ring_buffer_event *event;
3794 struct buffer_data_page *bpage;
3795 struct buffer_page *reader;
3796 unsigned long missed_events;
3797 unsigned long flags;
3798 unsigned int commit;
3803 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3807 * If len is not big enough to hold the page header, then
3808 * we can not copy anything.
3810 if (len <= BUF_PAGE_HDR_SIZE)
3813 len -= BUF_PAGE_HDR_SIZE;
3822 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3824 reader = rb_get_reader_page(cpu_buffer);
3828 event = rb_reader_event(cpu_buffer);
3830 read = reader->read;
3831 commit = rb_page_commit(reader);
3833 /* Check if any events were dropped */
3834 missed_events = cpu_buffer->lost_events;
3837 * If this page has been partially read or
3838 * if len is not big enough to read the rest of the page or
3839 * a writer is still on the page, then
3840 * we must copy the data from the page to the buffer.
3841 * Otherwise, we can simply swap the page with the one passed in.
3843 if (read || (len < (commit - read)) ||
3844 cpu_buffer->reader_page == cpu_buffer->commit_page) {
3845 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
3846 unsigned int rpos = read;
3847 unsigned int pos = 0;
3853 if (len > (commit - read))
3854 len = (commit - read);
3856 /* Always keep the time extend and data together */
3857 size = rb_event_ts_length(event);
3862 /* save the current timestamp, since the user will need it */
3863 save_timestamp = cpu_buffer->read_stamp;
3865 /* Need to copy one event at a time */
3867 /* We need the size of one event, because
3868 * rb_advance_reader only advances by one event,
3869 * whereas rb_event_ts_length may include the size of
3870 * one or two events.
3871 * We have already ensured there's enough space if this
3872 * is a time extend. */
3873 size = rb_event_length(event);
3874 memcpy(bpage->data + pos, rpage->data + rpos, size);
3878 rb_advance_reader(cpu_buffer);
3879 rpos = reader->read;
3885 event = rb_reader_event(cpu_buffer);
3886 /* Always keep the time extend and data together */
3887 size = rb_event_ts_length(event);
3888 } while (len >= size);
3891 local_set(&bpage->commit, pos);
3892 bpage->time_stamp = save_timestamp;
3894 /* we copied everything to the beginning */
3897 /* update the entry counter */
3898 cpu_buffer->read += rb_page_entries(reader);
3900 /* swap the pages */
3901 rb_init_page(bpage);
3902 bpage = reader->page;
3903 reader->page = *data_page;
3904 local_set(&reader->write, 0);
3905 local_set(&reader->entries, 0);
3910 * Use the real_end for the data size,
3911 * This gives us a chance to store the lost events
3914 if (reader->real_end)
3915 local_set(&bpage->commit, reader->real_end);
3919 cpu_buffer->lost_events = 0;
3921 commit = local_read(&bpage->commit);
3923 * Set a flag in the commit field if we lost events
3925 if (missed_events) {
3926 /* If there is room at the end of the page to save the
3927 * missed events, then record it there.
3929 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
3930 memcpy(&bpage->data[commit], &missed_events,
3931 sizeof(missed_events));
3932 local_add(RB_MISSED_STORED, &bpage->commit);
3933 commit += sizeof(missed_events);
3935 local_add(RB_MISSED_EVENTS, &bpage->commit);
3939 * This page may be off to user land. Zero it out here.
3941 if (commit < BUF_PAGE_SIZE)
3942 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
3945 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3950 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
3952 #ifdef CONFIG_TRACING
3954 rb_simple_read(struct file *filp, char __user *ubuf,
3955 size_t cnt, loff_t *ppos)
3957 unsigned long *p = filp->private_data;
3961 if (test_bit(RB_BUFFERS_DISABLED_BIT, p))
3962 r = sprintf(buf, "permanently disabled\n");
3964 r = sprintf(buf, "%d\n", test_bit(RB_BUFFERS_ON_BIT, p));
3966 return simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
3970 rb_simple_write(struct file *filp, const char __user *ubuf,
3971 size_t cnt, loff_t *ppos)
3973 unsigned long *p = filp->private_data;
3978 if (cnt >= sizeof(buf))
3981 if (copy_from_user(&buf, ubuf, cnt))
3986 ret = strict_strtoul(buf, 10, &val);
3991 set_bit(RB_BUFFERS_ON_BIT, p);
3993 clear_bit(RB_BUFFERS_ON_BIT, p);
4000 static const struct file_operations rb_simple_fops = {
4001 .open = tracing_open_generic,
4002 .read = rb_simple_read,
4003 .write = rb_simple_write,
4004 .llseek = default_llseek,
4008 static __init int rb_init_debugfs(void)
4010 struct dentry *d_tracer;
4012 d_tracer = tracing_init_dentry();
4014 trace_create_file("tracing_on", 0644, d_tracer,
4015 &ring_buffer_flags, &rb_simple_fops);
4020 fs_initcall(rb_init_debugfs);
4023 #ifdef CONFIG_HOTPLUG_CPU
4024 static int rb_cpu_notify(struct notifier_block *self,
4025 unsigned long action, void *hcpu)
4027 struct ring_buffer *buffer =
4028 container_of(self, struct ring_buffer, cpu_notify);
4029 long cpu = (long)hcpu;
4032 case CPU_UP_PREPARE:
4033 case CPU_UP_PREPARE_FROZEN:
4034 if (cpumask_test_cpu(cpu, buffer->cpumask))
4037 buffer->buffers[cpu] =
4038 rb_allocate_cpu_buffer(buffer, cpu);
4039 if (!buffer->buffers[cpu]) {
4040 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4045 cpumask_set_cpu(cpu, buffer->cpumask);
4047 case CPU_DOWN_PREPARE:
4048 case CPU_DOWN_PREPARE_FROZEN:
4051 * If we were to free the buffer, then the user would
4052 * lose any trace that was in the buffer.