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 static inline int rb_null_event(struct ring_buffer_event *event)
229 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
232 static void rb_event_set_padding(struct ring_buffer_event *event)
234 /* padding has a NULL time_delta */
235 event->type_len = RINGBUF_TYPE_PADDING;
236 event->time_delta = 0;
240 rb_event_data_length(struct ring_buffer_event *event)
245 length = event->type_len * RB_ALIGNMENT;
247 length = event->array[0];
248 return length + RB_EVNT_HDR_SIZE;
251 /* inline for ring buffer fast paths */
253 rb_event_length(struct ring_buffer_event *event)
255 switch (event->type_len) {
256 case RINGBUF_TYPE_PADDING:
257 if (rb_null_event(event))
260 return event->array[0] + RB_EVNT_HDR_SIZE;
262 case RINGBUF_TYPE_TIME_EXTEND:
263 return RB_LEN_TIME_EXTEND;
265 case RINGBUF_TYPE_TIME_STAMP:
266 return RB_LEN_TIME_STAMP;
268 case RINGBUF_TYPE_DATA:
269 return rb_event_data_length(event);
278 * ring_buffer_event_length - return the length of the event
279 * @event: the event to get the length of
281 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
283 unsigned length = rb_event_length(event);
284 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
286 length -= RB_EVNT_HDR_SIZE;
287 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
288 length -= sizeof(event->array[0]);
291 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
293 /* inline for ring buffer fast paths */
295 rb_event_data(struct ring_buffer_event *event)
297 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
298 /* If length is in len field, then array[0] has the data */
300 return (void *)&event->array[0];
301 /* Otherwise length is in array[0] and array[1] has the data */
302 return (void *)&event->array[1];
306 * ring_buffer_event_data - return the data of the event
307 * @event: the event to get the data from
309 void *ring_buffer_event_data(struct ring_buffer_event *event)
311 return rb_event_data(event);
313 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
315 #define for_each_buffer_cpu(buffer, cpu) \
316 for_each_cpu(cpu, buffer->cpumask)
319 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
320 #define TS_DELTA_TEST (~TS_MASK)
322 struct buffer_data_page {
323 u64 time_stamp; /* page time stamp */
324 local_t commit; /* write committed index */
325 unsigned char data[]; /* data of buffer page */
329 * Note, the buffer_page list must be first. The buffer pages
330 * are allocated in cache lines, which means that each buffer
331 * page will be at the beginning of a cache line, and thus
332 * the least significant bits will be zero. We use this to
333 * add flags in the list struct pointers, to make the ring buffer
337 struct list_head list; /* list of buffer pages */
338 local_t write; /* index for next write */
339 unsigned read; /* index for next read */
340 local_t entries; /* entries on this page */
341 struct buffer_data_page *page; /* Actual data page */
345 * The buffer page counters, write and entries, must be reset
346 * atomically when crossing page boundaries. To synchronize this
347 * update, two counters are inserted into the number. One is
348 * the actual counter for the write position or count on the page.
350 * The other is a counter of updaters. Before an update happens
351 * the update partition of the counter is incremented. This will
352 * allow the updater to update the counter atomically.
354 * The counter is 20 bits, and the state data is 12.
356 #define RB_WRITE_MASK 0xfffff
357 #define RB_WRITE_INTCNT (1 << 20)
359 static void rb_init_page(struct buffer_data_page *bpage)
361 local_set(&bpage->commit, 0);
365 * ring_buffer_page_len - the size of data on the page.
366 * @page: The page to read
368 * Returns the amount of data on the page, including buffer page header.
370 size_t ring_buffer_page_len(void *page)
372 return local_read(&((struct buffer_data_page *)page)->commit)
377 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
380 static void free_buffer_page(struct buffer_page *bpage)
382 free_page((unsigned long)bpage->page);
387 * We need to fit the time_stamp delta into 27 bits.
389 static inline int test_time_stamp(u64 delta)
391 if (delta & TS_DELTA_TEST)
396 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
398 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
399 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
401 /* Max number of timestamps that can fit on a page */
402 #define RB_TIMESTAMPS_PER_PAGE (BUF_PAGE_SIZE / RB_LEN_TIME_STAMP)
404 int ring_buffer_print_page_header(struct trace_seq *s)
406 struct buffer_data_page field;
409 ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t"
410 "offset:0;\tsize:%u;\tsigned:%u;\n",
411 (unsigned int)sizeof(field.time_stamp),
412 (unsigned int)is_signed_type(u64));
414 ret = trace_seq_printf(s, "\tfield: local_t commit;\t"
415 "offset:%u;\tsize:%u;\tsigned:%u;\n",
416 (unsigned int)offsetof(typeof(field), commit),
417 (unsigned int)sizeof(field.commit),
418 (unsigned int)is_signed_type(long));
420 ret = trace_seq_printf(s, "\tfield: char data;\t"
421 "offset:%u;\tsize:%u;\tsigned:%u;\n",
422 (unsigned int)offsetof(typeof(field), data),
423 (unsigned int)BUF_PAGE_SIZE,
424 (unsigned int)is_signed_type(char));
430 * head_page == tail_page && head == tail then buffer is empty.
432 struct ring_buffer_per_cpu {
434 struct ring_buffer *buffer;
435 spinlock_t reader_lock; /* serialize readers */
436 arch_spinlock_t lock;
437 struct lock_class_key lock_key;
438 struct list_head *pages;
439 struct buffer_page *head_page; /* read from head */
440 struct buffer_page *tail_page; /* write to tail */
441 struct buffer_page *commit_page; /* committed pages */
442 struct buffer_page *reader_page;
443 local_t commit_overrun;
451 atomic_t record_disabled;
458 atomic_t record_disabled;
459 cpumask_var_t cpumask;
461 struct lock_class_key *reader_lock_key;
465 struct ring_buffer_per_cpu **buffers;
467 #ifdef CONFIG_HOTPLUG_CPU
468 struct notifier_block cpu_notify;
473 struct ring_buffer_iter {
474 struct ring_buffer_per_cpu *cpu_buffer;
476 struct buffer_page *head_page;
477 struct buffer_page *cache_reader_page;
478 unsigned long cache_read;
482 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
483 #define RB_WARN_ON(b, cond) \
485 int _____ret = unlikely(cond); \
487 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
488 struct ring_buffer_per_cpu *__b = \
490 atomic_inc(&__b->buffer->record_disabled); \
492 atomic_inc(&b->record_disabled); \
498 /* Up this if you want to test the TIME_EXTENTS and normalization */
499 #define DEBUG_SHIFT 0
501 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
503 /* shift to debug/test normalization and TIME_EXTENTS */
504 return buffer->clock() << DEBUG_SHIFT;
507 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
511 preempt_disable_notrace();
512 time = rb_time_stamp(buffer);
513 preempt_enable_no_resched_notrace();
517 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
519 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
522 /* Just stupid testing the normalize function and deltas */
525 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
528 * Making the ring buffer lockless makes things tricky.
529 * Although writes only happen on the CPU that they are on,
530 * and they only need to worry about interrupts. Reads can
533 * The reader page is always off the ring buffer, but when the
534 * reader finishes with a page, it needs to swap its page with
535 * a new one from the buffer. The reader needs to take from
536 * the head (writes go to the tail). But if a writer is in overwrite
537 * mode and wraps, it must push the head page forward.
539 * Here lies the problem.
541 * The reader must be careful to replace only the head page, and
542 * not another one. As described at the top of the file in the
543 * ASCII art, the reader sets its old page to point to the next
544 * page after head. It then sets the page after head to point to
545 * the old reader page. But if the writer moves the head page
546 * during this operation, the reader could end up with the tail.
548 * We use cmpxchg to help prevent this race. We also do something
549 * special with the page before head. We set the LSB to 1.
551 * When the writer must push the page forward, it will clear the
552 * bit that points to the head page, move the head, and then set
553 * the bit that points to the new head page.
555 * We also don't want an interrupt coming in and moving the head
556 * page on another writer. Thus we use the second LSB to catch
559 * head->list->prev->next bit 1 bit 0
562 * Points to head page 0 1
565 * Note we can not trust the prev pointer of the head page, because:
567 * +----+ +-----+ +-----+
568 * | |------>| T |---X--->| N |
570 * +----+ +-----+ +-----+
573 * +----------| R |----------+ |
577 * Key: ---X--> HEAD flag set in pointer
582 * (see __rb_reserve_next() to see where this happens)
584 * What the above shows is that the reader just swapped out
585 * the reader page with a page in the buffer, but before it
586 * could make the new header point back to the new page added
587 * it was preempted by a writer. The writer moved forward onto
588 * the new page added by the reader and is about to move forward
591 * You can see, it is legitimate for the previous pointer of
592 * the head (or any page) not to point back to itself. But only
596 #define RB_PAGE_NORMAL 0UL
597 #define RB_PAGE_HEAD 1UL
598 #define RB_PAGE_UPDATE 2UL
601 #define RB_FLAG_MASK 3UL
603 /* PAGE_MOVED is not part of the mask */
604 #define RB_PAGE_MOVED 4UL
607 * rb_list_head - remove any bit
609 static struct list_head *rb_list_head(struct list_head *list)
611 unsigned long val = (unsigned long)list;
613 return (struct list_head *)(val & ~RB_FLAG_MASK);
617 * rb_is_head_page - test if the given page is the head page
619 * Because the reader may move the head_page pointer, we can
620 * not trust what the head page is (it may be pointing to
621 * the reader page). But if the next page is a header page,
622 * its flags will be non zero.
625 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
626 struct buffer_page *page, struct list_head *list)
630 val = (unsigned long)list->next;
632 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
633 return RB_PAGE_MOVED;
635 return val & RB_FLAG_MASK;
641 * The unique thing about the reader page, is that, if the
642 * writer is ever on it, the previous pointer never points
643 * back to the reader page.
645 static int rb_is_reader_page(struct buffer_page *page)
647 struct list_head *list = page->list.prev;
649 return rb_list_head(list->next) != &page->list;
653 * rb_set_list_to_head - set a list_head to be pointing to head.
655 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
656 struct list_head *list)
660 ptr = (unsigned long *)&list->next;
661 *ptr |= RB_PAGE_HEAD;
662 *ptr &= ~RB_PAGE_UPDATE;
666 * rb_head_page_activate - sets up head page
668 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
670 struct buffer_page *head;
672 head = cpu_buffer->head_page;
677 * Set the previous list pointer to have the HEAD flag.
679 rb_set_list_to_head(cpu_buffer, head->list.prev);
682 static void rb_list_head_clear(struct list_head *list)
684 unsigned long *ptr = (unsigned long *)&list->next;
686 *ptr &= ~RB_FLAG_MASK;
690 * rb_head_page_dactivate - clears head page ptr (for free list)
693 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
695 struct list_head *hd;
697 /* Go through the whole list and clear any pointers found. */
698 rb_list_head_clear(cpu_buffer->pages);
700 list_for_each(hd, cpu_buffer->pages)
701 rb_list_head_clear(hd);
704 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
705 struct buffer_page *head,
706 struct buffer_page *prev,
707 int old_flag, int new_flag)
709 struct list_head *list;
710 unsigned long val = (unsigned long)&head->list;
715 val &= ~RB_FLAG_MASK;
717 ret = cmpxchg((unsigned long *)&list->next,
718 val | old_flag, val | new_flag);
720 /* check if the reader took the page */
721 if ((ret & ~RB_FLAG_MASK) != val)
722 return RB_PAGE_MOVED;
724 return ret & RB_FLAG_MASK;
727 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
728 struct buffer_page *head,
729 struct buffer_page *prev,
732 return rb_head_page_set(cpu_buffer, head, prev,
733 old_flag, RB_PAGE_UPDATE);
736 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
737 struct buffer_page *head,
738 struct buffer_page *prev,
741 return rb_head_page_set(cpu_buffer, head, prev,
742 old_flag, RB_PAGE_HEAD);
745 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
746 struct buffer_page *head,
747 struct buffer_page *prev,
750 return rb_head_page_set(cpu_buffer, head, prev,
751 old_flag, RB_PAGE_NORMAL);
754 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
755 struct buffer_page **bpage)
757 struct list_head *p = rb_list_head((*bpage)->list.next);
759 *bpage = list_entry(p, struct buffer_page, list);
762 static struct buffer_page *
763 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
765 struct buffer_page *head;
766 struct buffer_page *page;
767 struct list_head *list;
770 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
774 list = cpu_buffer->pages;
775 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
778 page = head = cpu_buffer->head_page;
780 * It is possible that the writer moves the header behind
781 * where we started, and we miss in one loop.
782 * A second loop should grab the header, but we'll do
783 * three loops just because I'm paranoid.
785 for (i = 0; i < 3; i++) {
787 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
788 cpu_buffer->head_page = page;
791 rb_inc_page(cpu_buffer, &page);
792 } while (page != head);
795 RB_WARN_ON(cpu_buffer, 1);
800 static int rb_head_page_replace(struct buffer_page *old,
801 struct buffer_page *new)
803 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
807 val = *ptr & ~RB_FLAG_MASK;
810 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
816 * rb_tail_page_update - move the tail page forward
818 * Returns 1 if moved tail page, 0 if someone else did.
820 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
821 struct buffer_page *tail_page,
822 struct buffer_page *next_page)
824 struct buffer_page *old_tail;
825 unsigned long old_entries;
826 unsigned long old_write;
830 * The tail page now needs to be moved forward.
832 * We need to reset the tail page, but without messing
833 * with possible erasing of data brought in by interrupts
834 * that have moved the tail page and are currently on it.
836 * We add a counter to the write field to denote this.
838 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
839 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
842 * Just make sure we have seen our old_write and synchronize
843 * with any interrupts that come in.
848 * If the tail page is still the same as what we think
849 * it is, then it is up to us to update the tail
852 if (tail_page == cpu_buffer->tail_page) {
853 /* Zero the write counter */
854 unsigned long val = old_write & ~RB_WRITE_MASK;
855 unsigned long eval = old_entries & ~RB_WRITE_MASK;
858 * This will only succeed if an interrupt did
859 * not come in and change it. In which case, we
860 * do not want to modify it.
862 * We add (void) to let the compiler know that we do not care
863 * about the return value of these functions. We use the
864 * cmpxchg to only update if an interrupt did not already
865 * do it for us. If the cmpxchg fails, we don't care.
867 (void)local_cmpxchg(&next_page->write, old_write, val);
868 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
871 * No need to worry about races with clearing out the commit.
872 * it only can increment when a commit takes place. But that
873 * only happens in the outer most nested commit.
875 local_set(&next_page->page->commit, 0);
877 old_tail = cmpxchg(&cpu_buffer->tail_page,
878 tail_page, next_page);
880 if (old_tail == tail_page)
887 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
888 struct buffer_page *bpage)
890 unsigned long val = (unsigned long)bpage;
892 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
899 * rb_check_list - make sure a pointer to a list has the last bits zero
901 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
902 struct list_head *list)
904 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
906 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
912 * check_pages - integrity check of buffer pages
913 * @cpu_buffer: CPU buffer with pages to test
915 * As a safety measure we check to make sure the data pages have not
918 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
920 struct list_head *head = cpu_buffer->pages;
921 struct buffer_page *bpage, *tmp;
923 rb_head_page_deactivate(cpu_buffer);
925 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
927 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
930 if (rb_check_list(cpu_buffer, head))
933 list_for_each_entry_safe(bpage, tmp, head, list) {
934 if (RB_WARN_ON(cpu_buffer,
935 bpage->list.next->prev != &bpage->list))
937 if (RB_WARN_ON(cpu_buffer,
938 bpage->list.prev->next != &bpage->list))
940 if (rb_check_list(cpu_buffer, &bpage->list))
944 rb_head_page_activate(cpu_buffer);
949 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
952 struct buffer_page *bpage, *tmp;
959 for (i = 0; i < nr_pages; i++) {
960 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
961 GFP_KERNEL, cpu_to_node(cpu_buffer->cpu));
965 rb_check_bpage(cpu_buffer, bpage);
967 list_add(&bpage->list, &pages);
969 addr = __get_free_page(GFP_KERNEL);
972 bpage->page = (void *)addr;
973 rb_init_page(bpage->page);
977 * The ring buffer page list is a circular list that does not
978 * start and end with a list head. All page list items point to
981 cpu_buffer->pages = pages.next;
984 rb_check_pages(cpu_buffer);
989 list_for_each_entry_safe(bpage, tmp, &pages, list) {
990 list_del_init(&bpage->list);
991 free_buffer_page(bpage);
996 static struct ring_buffer_per_cpu *
997 rb_allocate_cpu_buffer(struct ring_buffer *buffer, int cpu)
999 struct ring_buffer_per_cpu *cpu_buffer;
1000 struct buffer_page *bpage;
1004 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1005 GFP_KERNEL, cpu_to_node(cpu));
1009 cpu_buffer->cpu = cpu;
1010 cpu_buffer->buffer = buffer;
1011 spin_lock_init(&cpu_buffer->reader_lock);
1012 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1013 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1015 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1016 GFP_KERNEL, cpu_to_node(cpu));
1018 goto fail_free_buffer;
1020 rb_check_bpage(cpu_buffer, bpage);
1022 cpu_buffer->reader_page = bpage;
1023 addr = __get_free_page(GFP_KERNEL);
1025 goto fail_free_reader;
1026 bpage->page = (void *)addr;
1027 rb_init_page(bpage->page);
1029 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1031 ret = rb_allocate_pages(cpu_buffer, buffer->pages);
1033 goto fail_free_reader;
1035 cpu_buffer->head_page
1036 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1037 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1039 rb_head_page_activate(cpu_buffer);
1044 free_buffer_page(cpu_buffer->reader_page);
1051 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1053 struct list_head *head = cpu_buffer->pages;
1054 struct buffer_page *bpage, *tmp;
1056 free_buffer_page(cpu_buffer->reader_page);
1058 rb_head_page_deactivate(cpu_buffer);
1061 list_for_each_entry_safe(bpage, tmp, head, list) {
1062 list_del_init(&bpage->list);
1063 free_buffer_page(bpage);
1065 bpage = list_entry(head, struct buffer_page, list);
1066 free_buffer_page(bpage);
1072 #ifdef CONFIG_HOTPLUG_CPU
1073 static int rb_cpu_notify(struct notifier_block *self,
1074 unsigned long action, void *hcpu);
1078 * ring_buffer_alloc - allocate a new ring_buffer
1079 * @size: the size in bytes per cpu that is needed.
1080 * @flags: attributes to set for the ring buffer.
1082 * Currently the only flag that is available is the RB_FL_OVERWRITE
1083 * flag. This flag means that the buffer will overwrite old data
1084 * when the buffer wraps. If this flag is not set, the buffer will
1085 * drop data when the tail hits the head.
1087 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1088 struct lock_class_key *key)
1090 struct ring_buffer *buffer;
1094 /* keep it in its own cache line */
1095 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1100 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1101 goto fail_free_buffer;
1103 buffer->pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1104 buffer->flags = flags;
1105 buffer->clock = trace_clock_local;
1106 buffer->reader_lock_key = key;
1108 /* need at least two pages */
1109 if (buffer->pages < 2)
1113 * In case of non-hotplug cpu, if the ring-buffer is allocated
1114 * in early initcall, it will not be notified of secondary cpus.
1115 * In that off case, we need to allocate for all possible cpus.
1117 #ifdef CONFIG_HOTPLUG_CPU
1119 cpumask_copy(buffer->cpumask, cpu_online_mask);
1121 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1123 buffer->cpus = nr_cpu_ids;
1125 bsize = sizeof(void *) * nr_cpu_ids;
1126 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1128 if (!buffer->buffers)
1129 goto fail_free_cpumask;
1131 for_each_buffer_cpu(buffer, cpu) {
1132 buffer->buffers[cpu] =
1133 rb_allocate_cpu_buffer(buffer, cpu);
1134 if (!buffer->buffers[cpu])
1135 goto fail_free_buffers;
1138 #ifdef CONFIG_HOTPLUG_CPU
1139 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1140 buffer->cpu_notify.priority = 0;
1141 register_cpu_notifier(&buffer->cpu_notify);
1145 mutex_init(&buffer->mutex);
1150 for_each_buffer_cpu(buffer, cpu) {
1151 if (buffer->buffers[cpu])
1152 rb_free_cpu_buffer(buffer->buffers[cpu]);
1154 kfree(buffer->buffers);
1157 free_cpumask_var(buffer->cpumask);
1164 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1167 * ring_buffer_free - free a ring buffer.
1168 * @buffer: the buffer to free.
1171 ring_buffer_free(struct ring_buffer *buffer)
1177 #ifdef CONFIG_HOTPLUG_CPU
1178 unregister_cpu_notifier(&buffer->cpu_notify);
1181 for_each_buffer_cpu(buffer, cpu)
1182 rb_free_cpu_buffer(buffer->buffers[cpu]);
1186 kfree(buffer->buffers);
1187 free_cpumask_var(buffer->cpumask);
1191 EXPORT_SYMBOL_GPL(ring_buffer_free);
1193 void ring_buffer_set_clock(struct ring_buffer *buffer,
1196 buffer->clock = clock;
1199 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1202 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned nr_pages)
1204 struct buffer_page *bpage;
1205 struct list_head *p;
1208 spin_lock_irq(&cpu_buffer->reader_lock);
1209 rb_head_page_deactivate(cpu_buffer);
1211 for (i = 0; i < nr_pages; i++) {
1212 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1214 p = cpu_buffer->pages->next;
1215 bpage = list_entry(p, struct buffer_page, list);
1216 list_del_init(&bpage->list);
1217 free_buffer_page(bpage);
1219 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1222 rb_reset_cpu(cpu_buffer);
1223 rb_check_pages(cpu_buffer);
1225 spin_unlock_irq(&cpu_buffer->reader_lock);
1229 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer,
1230 struct list_head *pages, unsigned nr_pages)
1232 struct buffer_page *bpage;
1233 struct list_head *p;
1236 spin_lock_irq(&cpu_buffer->reader_lock);
1237 rb_head_page_deactivate(cpu_buffer);
1239 for (i = 0; i < nr_pages; i++) {
1240 if (RB_WARN_ON(cpu_buffer, list_empty(pages)))
1243 bpage = list_entry(p, struct buffer_page, list);
1244 list_del_init(&bpage->list);
1245 list_add_tail(&bpage->list, cpu_buffer->pages);
1247 rb_reset_cpu(cpu_buffer);
1248 rb_check_pages(cpu_buffer);
1250 spin_unlock_irq(&cpu_buffer->reader_lock);
1254 * ring_buffer_resize - resize the ring buffer
1255 * @buffer: the buffer to resize.
1256 * @size: the new size.
1258 * Minimum size is 2 * BUF_PAGE_SIZE.
1260 * Returns -1 on failure.
1262 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size)
1264 struct ring_buffer_per_cpu *cpu_buffer;
1265 unsigned nr_pages, rm_pages, new_pages;
1266 struct buffer_page *bpage, *tmp;
1267 unsigned long buffer_size;
1273 * Always succeed at resizing a non-existent buffer:
1278 size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1279 size *= BUF_PAGE_SIZE;
1280 buffer_size = buffer->pages * BUF_PAGE_SIZE;
1282 /* we need a minimum of two pages */
1283 if (size < BUF_PAGE_SIZE * 2)
1284 size = BUF_PAGE_SIZE * 2;
1286 if (size == buffer_size)
1289 atomic_inc(&buffer->record_disabled);
1291 /* Make sure all writers are done with this buffer. */
1292 synchronize_sched();
1294 mutex_lock(&buffer->mutex);
1297 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1299 if (size < buffer_size) {
1301 /* easy case, just free pages */
1302 if (RB_WARN_ON(buffer, nr_pages >= buffer->pages))
1305 rm_pages = buffer->pages - nr_pages;
1307 for_each_buffer_cpu(buffer, cpu) {
1308 cpu_buffer = buffer->buffers[cpu];
1309 rb_remove_pages(cpu_buffer, rm_pages);
1315 * This is a bit more difficult. We only want to add pages
1316 * when we can allocate enough for all CPUs. We do this
1317 * by allocating all the pages and storing them on a local
1318 * link list. If we succeed in our allocation, then we
1319 * add these pages to the cpu_buffers. Otherwise we just free
1320 * them all and return -ENOMEM;
1322 if (RB_WARN_ON(buffer, nr_pages <= buffer->pages))
1325 new_pages = nr_pages - buffer->pages;
1327 for_each_buffer_cpu(buffer, cpu) {
1328 for (i = 0; i < new_pages; i++) {
1329 bpage = kzalloc_node(ALIGN(sizeof(*bpage),
1331 GFP_KERNEL, cpu_to_node(cpu));
1334 list_add(&bpage->list, &pages);
1335 addr = __get_free_page(GFP_KERNEL);
1338 bpage->page = (void *)addr;
1339 rb_init_page(bpage->page);
1343 for_each_buffer_cpu(buffer, cpu) {
1344 cpu_buffer = buffer->buffers[cpu];
1345 rb_insert_pages(cpu_buffer, &pages, new_pages);
1348 if (RB_WARN_ON(buffer, !list_empty(&pages)))
1352 buffer->pages = nr_pages;
1354 mutex_unlock(&buffer->mutex);
1356 atomic_dec(&buffer->record_disabled);
1361 list_for_each_entry_safe(bpage, tmp, &pages, list) {
1362 list_del_init(&bpage->list);
1363 free_buffer_page(bpage);
1366 mutex_unlock(&buffer->mutex);
1367 atomic_dec(&buffer->record_disabled);
1371 * Something went totally wrong, and we are too paranoid
1372 * to even clean up the mess.
1376 mutex_unlock(&buffer->mutex);
1377 atomic_dec(&buffer->record_disabled);
1380 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1382 static inline void *
1383 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1385 return bpage->data + index;
1388 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1390 return bpage->page->data + index;
1393 static inline struct ring_buffer_event *
1394 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1396 return __rb_page_index(cpu_buffer->reader_page,
1397 cpu_buffer->reader_page->read);
1400 static inline struct ring_buffer_event *
1401 rb_iter_head_event(struct ring_buffer_iter *iter)
1403 return __rb_page_index(iter->head_page, iter->head);
1406 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1408 return local_read(&bpage->write) & RB_WRITE_MASK;
1411 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1413 return local_read(&bpage->page->commit);
1416 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1418 return local_read(&bpage->entries) & RB_WRITE_MASK;
1421 /* Size is determined by what has been commited */
1422 static inline unsigned rb_page_size(struct buffer_page *bpage)
1424 return rb_page_commit(bpage);
1427 static inline unsigned
1428 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1430 return rb_page_commit(cpu_buffer->commit_page);
1433 static inline unsigned
1434 rb_event_index(struct ring_buffer_event *event)
1436 unsigned long addr = (unsigned long)event;
1438 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1442 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1443 struct ring_buffer_event *event)
1445 unsigned long addr = (unsigned long)event;
1446 unsigned long index;
1448 index = rb_event_index(event);
1451 return cpu_buffer->commit_page->page == (void *)addr &&
1452 rb_commit_index(cpu_buffer) == index;
1456 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1458 unsigned long max_count;
1461 * We only race with interrupts and NMIs on this CPU.
1462 * If we own the commit event, then we can commit
1463 * all others that interrupted us, since the interruptions
1464 * are in stack format (they finish before they come
1465 * back to us). This allows us to do a simple loop to
1466 * assign the commit to the tail.
1469 max_count = cpu_buffer->buffer->pages * 100;
1471 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1472 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1474 if (RB_WARN_ON(cpu_buffer,
1475 rb_is_reader_page(cpu_buffer->tail_page)))
1477 local_set(&cpu_buffer->commit_page->page->commit,
1478 rb_page_write(cpu_buffer->commit_page));
1479 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1480 cpu_buffer->write_stamp =
1481 cpu_buffer->commit_page->page->time_stamp;
1482 /* add barrier to keep gcc from optimizing too much */
1485 while (rb_commit_index(cpu_buffer) !=
1486 rb_page_write(cpu_buffer->commit_page)) {
1488 local_set(&cpu_buffer->commit_page->page->commit,
1489 rb_page_write(cpu_buffer->commit_page));
1490 RB_WARN_ON(cpu_buffer,
1491 local_read(&cpu_buffer->commit_page->page->commit) &
1496 /* again, keep gcc from optimizing */
1500 * If an interrupt came in just after the first while loop
1501 * and pushed the tail page forward, we will be left with
1502 * a dangling commit that will never go forward.
1504 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1508 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1510 cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1511 cpu_buffer->reader_page->read = 0;
1514 static void rb_inc_iter(struct ring_buffer_iter *iter)
1516 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1519 * The iterator could be on the reader page (it starts there).
1520 * But the head could have moved, since the reader was
1521 * found. Check for this case and assign the iterator
1522 * to the head page instead of next.
1524 if (iter->head_page == cpu_buffer->reader_page)
1525 iter->head_page = rb_set_head_page(cpu_buffer);
1527 rb_inc_page(cpu_buffer, &iter->head_page);
1529 iter->read_stamp = iter->head_page->page->time_stamp;
1534 * ring_buffer_update_event - update event type and data
1535 * @event: the even to update
1536 * @type: the type of event
1537 * @length: the size of the event field in the ring buffer
1539 * Update the type and data fields of the event. The length
1540 * is the actual size that is written to the ring buffer,
1541 * and with this, we can determine what to place into the
1545 rb_update_event(struct ring_buffer_event *event,
1546 unsigned type, unsigned length)
1548 event->type_len = type;
1552 case RINGBUF_TYPE_PADDING:
1553 case RINGBUF_TYPE_TIME_EXTEND:
1554 case RINGBUF_TYPE_TIME_STAMP:
1558 length -= RB_EVNT_HDR_SIZE;
1559 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
1560 event->array[0] = length;
1562 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
1570 * rb_handle_head_page - writer hit the head page
1572 * Returns: +1 to retry page
1577 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1578 struct buffer_page *tail_page,
1579 struct buffer_page *next_page)
1581 struct buffer_page *new_head;
1586 entries = rb_page_entries(next_page);
1589 * The hard part is here. We need to move the head
1590 * forward, and protect against both readers on
1591 * other CPUs and writers coming in via interrupts.
1593 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1597 * type can be one of four:
1598 * NORMAL - an interrupt already moved it for us
1599 * HEAD - we are the first to get here.
1600 * UPDATE - we are the interrupt interrupting
1602 * MOVED - a reader on another CPU moved the next
1603 * pointer to its reader page. Give up
1610 * We changed the head to UPDATE, thus
1611 * it is our responsibility to update
1614 local_add(entries, &cpu_buffer->overrun);
1617 * The entries will be zeroed out when we move the
1621 /* still more to do */
1624 case RB_PAGE_UPDATE:
1626 * This is an interrupt that interrupt the
1627 * previous update. Still more to do.
1630 case RB_PAGE_NORMAL:
1632 * An interrupt came in before the update
1633 * and processed this for us.
1634 * Nothing left to do.
1639 * The reader is on another CPU and just did
1640 * a swap with our next_page.
1645 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1650 * Now that we are here, the old head pointer is
1651 * set to UPDATE. This will keep the reader from
1652 * swapping the head page with the reader page.
1653 * The reader (on another CPU) will spin till
1656 * We just need to protect against interrupts
1657 * doing the job. We will set the next pointer
1658 * to HEAD. After that, we set the old pointer
1659 * to NORMAL, but only if it was HEAD before.
1660 * otherwise we are an interrupt, and only
1661 * want the outer most commit to reset it.
1663 new_head = next_page;
1664 rb_inc_page(cpu_buffer, &new_head);
1666 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1670 * Valid returns are:
1671 * HEAD - an interrupt came in and already set it.
1672 * NORMAL - One of two things:
1673 * 1) We really set it.
1674 * 2) A bunch of interrupts came in and moved
1675 * the page forward again.
1679 case RB_PAGE_NORMAL:
1683 RB_WARN_ON(cpu_buffer, 1);
1688 * It is possible that an interrupt came in,
1689 * set the head up, then more interrupts came in
1690 * and moved it again. When we get back here,
1691 * the page would have been set to NORMAL but we
1692 * just set it back to HEAD.
1694 * How do you detect this? Well, if that happened
1695 * the tail page would have moved.
1697 if (ret == RB_PAGE_NORMAL) {
1699 * If the tail had moved passed next, then we need
1700 * to reset the pointer.
1702 if (cpu_buffer->tail_page != tail_page &&
1703 cpu_buffer->tail_page != next_page)
1704 rb_head_page_set_normal(cpu_buffer, new_head,
1710 * If this was the outer most commit (the one that
1711 * changed the original pointer from HEAD to UPDATE),
1712 * then it is up to us to reset it to NORMAL.
1714 if (type == RB_PAGE_HEAD) {
1715 ret = rb_head_page_set_normal(cpu_buffer, next_page,
1718 if (RB_WARN_ON(cpu_buffer,
1719 ret != RB_PAGE_UPDATE))
1726 static unsigned rb_calculate_event_length(unsigned length)
1728 struct ring_buffer_event event; /* Used only for sizeof array */
1730 /* zero length can cause confusions */
1734 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
1735 length += sizeof(event.array[0]);
1737 length += RB_EVNT_HDR_SIZE;
1738 length = ALIGN(length, RB_ARCH_ALIGNMENT);
1744 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
1745 struct buffer_page *tail_page,
1746 unsigned long tail, unsigned long length)
1748 struct ring_buffer_event *event;
1751 * Only the event that crossed the page boundary
1752 * must fill the old tail_page with padding.
1754 if (tail >= BUF_PAGE_SIZE) {
1755 local_sub(length, &tail_page->write);
1759 event = __rb_page_index(tail_page, tail);
1760 kmemcheck_annotate_bitfield(event, bitfield);
1763 * If this event is bigger than the minimum size, then
1764 * we need to be careful that we don't subtract the
1765 * write counter enough to allow another writer to slip
1767 * We put in a discarded commit instead, to make sure
1768 * that this space is not used again.
1770 * If we are less than the minimum size, we don't need to
1773 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
1774 /* No room for any events */
1776 /* Mark the rest of the page with padding */
1777 rb_event_set_padding(event);
1779 /* Set the write back to the previous setting */
1780 local_sub(length, &tail_page->write);
1784 /* Put in a discarded event */
1785 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
1786 event->type_len = RINGBUF_TYPE_PADDING;
1787 /* time delta must be non zero */
1788 event->time_delta = 1;
1790 /* Set write to end of buffer */
1791 length = (tail + length) - BUF_PAGE_SIZE;
1792 local_sub(length, &tail_page->write);
1795 static struct ring_buffer_event *
1796 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
1797 unsigned long length, unsigned long tail,
1798 struct buffer_page *tail_page, u64 *ts)
1800 struct buffer_page *commit_page = cpu_buffer->commit_page;
1801 struct ring_buffer *buffer = cpu_buffer->buffer;
1802 struct buffer_page *next_page;
1805 next_page = tail_page;
1807 rb_inc_page(cpu_buffer, &next_page);
1810 * If for some reason, we had an interrupt storm that made
1811 * it all the way around the buffer, bail, and warn
1814 if (unlikely(next_page == commit_page)) {
1815 local_inc(&cpu_buffer->commit_overrun);
1820 * This is where the fun begins!
1822 * We are fighting against races between a reader that
1823 * could be on another CPU trying to swap its reader
1824 * page with the buffer head.
1826 * We are also fighting against interrupts coming in and
1827 * moving the head or tail on us as well.
1829 * If the next page is the head page then we have filled
1830 * the buffer, unless the commit page is still on the
1833 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
1836 * If the commit is not on the reader page, then
1837 * move the header page.
1839 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
1841 * If we are not in overwrite mode,
1842 * this is easy, just stop here.
1844 if (!(buffer->flags & RB_FL_OVERWRITE))
1847 ret = rb_handle_head_page(cpu_buffer,
1856 * We need to be careful here too. The
1857 * commit page could still be on the reader
1858 * page. We could have a small buffer, and
1859 * have filled up the buffer with events
1860 * from interrupts and such, and wrapped.
1862 * Note, if the tail page is also the on the
1863 * reader_page, we let it move out.
1865 if (unlikely((cpu_buffer->commit_page !=
1866 cpu_buffer->tail_page) &&
1867 (cpu_buffer->commit_page ==
1868 cpu_buffer->reader_page))) {
1869 local_inc(&cpu_buffer->commit_overrun);
1875 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
1878 * Nested commits always have zero deltas, so
1879 * just reread the time stamp
1881 *ts = rb_time_stamp(buffer);
1882 next_page->page->time_stamp = *ts;
1887 rb_reset_tail(cpu_buffer, tail_page, tail, length);
1889 /* fail and let the caller try again */
1890 return ERR_PTR(-EAGAIN);
1894 rb_reset_tail(cpu_buffer, tail_page, tail, length);
1899 static struct ring_buffer_event *
1900 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
1901 unsigned type, unsigned long length, u64 *ts)
1903 struct buffer_page *tail_page;
1904 struct ring_buffer_event *event;
1905 unsigned long tail, write;
1907 tail_page = cpu_buffer->tail_page;
1908 write = local_add_return(length, &tail_page->write);
1910 /* set write to only the index of the write */
1911 write &= RB_WRITE_MASK;
1912 tail = write - length;
1914 /* See if we shot pass the end of this buffer page */
1915 if (write > BUF_PAGE_SIZE)
1916 return rb_move_tail(cpu_buffer, length, tail,
1919 /* We reserved something on the buffer */
1921 event = __rb_page_index(tail_page, tail);
1922 kmemcheck_annotate_bitfield(event, bitfield);
1923 rb_update_event(event, type, length);
1925 /* The passed in type is zero for DATA */
1927 local_inc(&tail_page->entries);
1930 * If this is the first commit on the page, then update
1934 tail_page->page->time_stamp = *ts;
1940 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
1941 struct ring_buffer_event *event)
1943 unsigned long new_index, old_index;
1944 struct buffer_page *bpage;
1945 unsigned long index;
1948 new_index = rb_event_index(event);
1949 old_index = new_index + rb_event_length(event);
1950 addr = (unsigned long)event;
1953 bpage = cpu_buffer->tail_page;
1955 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
1956 unsigned long write_mask =
1957 local_read(&bpage->write) & ~RB_WRITE_MASK;
1959 * This is on the tail page. It is possible that
1960 * a write could come in and move the tail page
1961 * and write to the next page. That is fine
1962 * because we just shorten what is on this page.
1964 old_index += write_mask;
1965 new_index += write_mask;
1966 index = local_cmpxchg(&bpage->write, old_index, new_index);
1967 if (index == old_index)
1971 /* could not discard */
1976 rb_add_time_stamp(struct ring_buffer_per_cpu *cpu_buffer,
1977 u64 *ts, u64 *delta)
1979 struct ring_buffer_event *event;
1983 if (unlikely(*delta > (1ULL << 59) && !once++)) {
1984 printk(KERN_WARNING "Delta way too big! %llu"
1985 " ts=%llu write stamp = %llu\n",
1986 (unsigned long long)*delta,
1987 (unsigned long long)*ts,
1988 (unsigned long long)cpu_buffer->write_stamp);
1993 * The delta is too big, we to add a
1996 event = __rb_reserve_next(cpu_buffer,
1997 RINGBUF_TYPE_TIME_EXTEND,
2003 if (PTR_ERR(event) == -EAGAIN)
2006 /* Only a commited time event can update the write stamp */
2007 if (rb_event_is_commit(cpu_buffer, event)) {
2009 * If this is the first on the page, then it was
2010 * updated with the page itself. Try to discard it
2011 * and if we can't just make it zero.
2013 if (rb_event_index(event)) {
2014 event->time_delta = *delta & TS_MASK;
2015 event->array[0] = *delta >> TS_SHIFT;
2017 /* try to discard, since we do not need this */
2018 if (!rb_try_to_discard(cpu_buffer, event)) {
2019 /* nope, just zero it */
2020 event->time_delta = 0;
2021 event->array[0] = 0;
2024 cpu_buffer->write_stamp = *ts;
2025 /* let the caller know this was the commit */
2028 /* Try to discard the event */
2029 if (!rb_try_to_discard(cpu_buffer, event)) {
2030 /* Darn, this is just wasted space */
2031 event->time_delta = 0;
2032 event->array[0] = 0;
2042 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2044 local_inc(&cpu_buffer->committing);
2045 local_inc(&cpu_buffer->commits);
2048 static void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2050 unsigned long commits;
2052 if (RB_WARN_ON(cpu_buffer,
2053 !local_read(&cpu_buffer->committing)))
2057 commits = local_read(&cpu_buffer->commits);
2058 /* synchronize with interrupts */
2060 if (local_read(&cpu_buffer->committing) == 1)
2061 rb_set_commit_to_write(cpu_buffer);
2063 local_dec(&cpu_buffer->committing);
2065 /* synchronize with interrupts */
2069 * Need to account for interrupts coming in between the
2070 * updating of the commit page and the clearing of the
2071 * committing counter.
2073 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2074 !local_read(&cpu_buffer->committing)) {
2075 local_inc(&cpu_buffer->committing);
2080 static struct ring_buffer_event *
2081 rb_reserve_next_event(struct ring_buffer *buffer,
2082 struct ring_buffer_per_cpu *cpu_buffer,
2083 unsigned long length)
2085 struct ring_buffer_event *event;
2090 rb_start_commit(cpu_buffer);
2092 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2094 * Due to the ability to swap a cpu buffer from a buffer
2095 * it is possible it was swapped before we committed.
2096 * (committing stops a swap). We check for it here and
2097 * if it happened, we have to fail the write.
2100 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2101 local_dec(&cpu_buffer->committing);
2102 local_dec(&cpu_buffer->commits);
2107 length = rb_calculate_event_length(length);
2110 * We allow for interrupts to reenter here and do a trace.
2111 * If one does, it will cause this original code to loop
2112 * back here. Even with heavy interrupts happening, this
2113 * should only happen a few times in a row. If this happens
2114 * 1000 times in a row, there must be either an interrupt
2115 * storm or we have something buggy.
2118 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2121 ts = rb_time_stamp(cpu_buffer->buffer);
2124 * Only the first commit can update the timestamp.
2125 * Yes there is a race here. If an interrupt comes in
2126 * just after the conditional and it traces too, then it
2127 * will also check the deltas. More than one timestamp may
2128 * also be made. But only the entry that did the actual
2129 * commit will be something other than zero.
2131 if (likely(cpu_buffer->tail_page == cpu_buffer->commit_page &&
2132 rb_page_write(cpu_buffer->tail_page) ==
2133 rb_commit_index(cpu_buffer))) {
2136 diff = ts - cpu_buffer->write_stamp;
2138 /* make sure this diff is calculated here */
2141 /* Did the write stamp get updated already? */
2142 if (unlikely(ts < cpu_buffer->write_stamp))
2146 if (unlikely(test_time_stamp(delta))) {
2148 commit = rb_add_time_stamp(cpu_buffer, &ts, &delta);
2149 if (commit == -EBUSY)
2152 if (commit == -EAGAIN)
2155 RB_WARN_ON(cpu_buffer, commit < 0);
2160 event = __rb_reserve_next(cpu_buffer, 0, length, &ts);
2161 if (unlikely(PTR_ERR(event) == -EAGAIN))
2167 if (!rb_event_is_commit(cpu_buffer, event))
2170 event->time_delta = delta;
2175 rb_end_commit(cpu_buffer);
2179 #ifdef CONFIG_TRACING
2181 #define TRACE_RECURSIVE_DEPTH 16
2183 static int trace_recursive_lock(void)
2185 current->trace_recursion++;
2187 if (likely(current->trace_recursion < TRACE_RECURSIVE_DEPTH))
2190 /* Disable all tracing before we do anything else */
2191 tracing_off_permanent();
2193 printk_once(KERN_WARNING "Tracing recursion: depth[%ld]:"
2194 "HC[%lu]:SC[%lu]:NMI[%lu]\n",
2195 current->trace_recursion,
2196 hardirq_count() >> HARDIRQ_SHIFT,
2197 softirq_count() >> SOFTIRQ_SHIFT,
2204 static void trace_recursive_unlock(void)
2206 WARN_ON_ONCE(!current->trace_recursion);
2208 current->trace_recursion--;
2213 #define trace_recursive_lock() (0)
2214 #define trace_recursive_unlock() do { } while (0)
2218 static DEFINE_PER_CPU(int, rb_need_resched);
2221 * ring_buffer_lock_reserve - reserve a part of the buffer
2222 * @buffer: the ring buffer to reserve from
2223 * @length: the length of the data to reserve (excluding event header)
2225 * Returns a reseverd event on the ring buffer to copy directly to.
2226 * The user of this interface will need to get the body to write into
2227 * and can use the ring_buffer_event_data() interface.
2229 * The length is the length of the data needed, not the event length
2230 * which also includes the event header.
2232 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2233 * If NULL is returned, then nothing has been allocated or locked.
2235 struct ring_buffer_event *
2236 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2238 struct ring_buffer_per_cpu *cpu_buffer;
2239 struct ring_buffer_event *event;
2242 if (ring_buffer_flags != RB_BUFFERS_ON)
2245 /* If we are tracing schedule, we don't want to recurse */
2246 resched = ftrace_preempt_disable();
2248 if (atomic_read(&buffer->record_disabled))
2251 if (trace_recursive_lock())
2254 cpu = raw_smp_processor_id();
2256 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2259 cpu_buffer = buffer->buffers[cpu];
2261 if (atomic_read(&cpu_buffer->record_disabled))
2264 if (length > BUF_MAX_DATA_SIZE)
2267 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2272 * Need to store resched state on this cpu.
2273 * Only the first needs to.
2276 if (preempt_count() == 1)
2277 per_cpu(rb_need_resched, cpu) = resched;
2282 trace_recursive_unlock();
2285 ftrace_preempt_enable(resched);
2288 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2291 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2292 struct ring_buffer_event *event)
2295 * The event first in the commit queue updates the
2298 if (rb_event_is_commit(cpu_buffer, event))
2299 cpu_buffer->write_stamp += event->time_delta;
2302 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2303 struct ring_buffer_event *event)
2305 local_inc(&cpu_buffer->entries);
2306 rb_update_write_stamp(cpu_buffer, event);
2307 rb_end_commit(cpu_buffer);
2311 * ring_buffer_unlock_commit - commit a reserved
2312 * @buffer: The buffer to commit to
2313 * @event: The event pointer to commit.
2315 * This commits the data to the ring buffer, and releases any locks held.
2317 * Must be paired with ring_buffer_lock_reserve.
2319 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2320 struct ring_buffer_event *event)
2322 struct ring_buffer_per_cpu *cpu_buffer;
2323 int cpu = raw_smp_processor_id();
2325 cpu_buffer = buffer->buffers[cpu];
2327 rb_commit(cpu_buffer, event);
2329 trace_recursive_unlock();
2332 * Only the last preempt count needs to restore preemption.
2334 if (preempt_count() == 1)
2335 ftrace_preempt_enable(per_cpu(rb_need_resched, cpu));
2337 preempt_enable_no_resched_notrace();
2341 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2343 static inline void rb_event_discard(struct ring_buffer_event *event)
2345 /* array[0] holds the actual length for the discarded event */
2346 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2347 event->type_len = RINGBUF_TYPE_PADDING;
2348 /* time delta must be non zero */
2349 if (!event->time_delta)
2350 event->time_delta = 1;
2354 * Decrement the entries to the page that an event is on.
2355 * The event does not even need to exist, only the pointer
2356 * to the page it is on. This may only be called before the commit
2360 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2361 struct ring_buffer_event *event)
2363 unsigned long addr = (unsigned long)event;
2364 struct buffer_page *bpage = cpu_buffer->commit_page;
2365 struct buffer_page *start;
2369 /* Do the likely case first */
2370 if (likely(bpage->page == (void *)addr)) {
2371 local_dec(&bpage->entries);
2376 * Because the commit page may be on the reader page we
2377 * start with the next page and check the end loop there.
2379 rb_inc_page(cpu_buffer, &bpage);
2382 if (bpage->page == (void *)addr) {
2383 local_dec(&bpage->entries);
2386 rb_inc_page(cpu_buffer, &bpage);
2387 } while (bpage != start);
2389 /* commit not part of this buffer?? */
2390 RB_WARN_ON(cpu_buffer, 1);
2394 * ring_buffer_commit_discard - discard an event that has not been committed
2395 * @buffer: the ring buffer
2396 * @event: non committed event to discard
2398 * Sometimes an event that is in the ring buffer needs to be ignored.
2399 * This function lets the user discard an event in the ring buffer
2400 * and then that event will not be read later.
2402 * This function only works if it is called before the the item has been
2403 * committed. It will try to free the event from the ring buffer
2404 * if another event has not been added behind it.
2406 * If another event has been added behind it, it will set the event
2407 * up as discarded, and perform the commit.
2409 * If this function is called, do not call ring_buffer_unlock_commit on
2412 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2413 struct ring_buffer_event *event)
2415 struct ring_buffer_per_cpu *cpu_buffer;
2418 /* The event is discarded regardless */
2419 rb_event_discard(event);
2421 cpu = smp_processor_id();
2422 cpu_buffer = buffer->buffers[cpu];
2425 * This must only be called if the event has not been
2426 * committed yet. Thus we can assume that preemption
2427 * is still disabled.
2429 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2431 rb_decrement_entry(cpu_buffer, event);
2432 if (rb_try_to_discard(cpu_buffer, event))
2436 * The commit is still visible by the reader, so we
2437 * must still update the timestamp.
2439 rb_update_write_stamp(cpu_buffer, event);
2441 rb_end_commit(cpu_buffer);
2443 trace_recursive_unlock();
2446 * Only the last preempt count needs to restore preemption.
2448 if (preempt_count() == 1)
2449 ftrace_preempt_enable(per_cpu(rb_need_resched, cpu));
2451 preempt_enable_no_resched_notrace();
2454 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2457 * ring_buffer_write - write data to the buffer without reserving
2458 * @buffer: The ring buffer to write to.
2459 * @length: The length of the data being written (excluding the event header)
2460 * @data: The data to write to the buffer.
2462 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2463 * one function. If you already have the data to write to the buffer, it
2464 * may be easier to simply call this function.
2466 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2467 * and not the length of the event which would hold the header.
2469 int ring_buffer_write(struct ring_buffer *buffer,
2470 unsigned long length,
2473 struct ring_buffer_per_cpu *cpu_buffer;
2474 struct ring_buffer_event *event;
2479 if (ring_buffer_flags != RB_BUFFERS_ON)
2482 resched = ftrace_preempt_disable();
2484 if (atomic_read(&buffer->record_disabled))
2487 cpu = raw_smp_processor_id();
2489 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2492 cpu_buffer = buffer->buffers[cpu];
2494 if (atomic_read(&cpu_buffer->record_disabled))
2497 if (length > BUF_MAX_DATA_SIZE)
2500 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2504 body = rb_event_data(event);
2506 memcpy(body, data, length);
2508 rb_commit(cpu_buffer, event);
2512 ftrace_preempt_enable(resched);
2516 EXPORT_SYMBOL_GPL(ring_buffer_write);
2518 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
2520 struct buffer_page *reader = cpu_buffer->reader_page;
2521 struct buffer_page *head = rb_set_head_page(cpu_buffer);
2522 struct buffer_page *commit = cpu_buffer->commit_page;
2524 /* In case of error, head will be NULL */
2525 if (unlikely(!head))
2528 return reader->read == rb_page_commit(reader) &&
2529 (commit == reader ||
2531 head->read == rb_page_commit(commit)));
2535 * ring_buffer_record_disable - stop all writes into the buffer
2536 * @buffer: The ring buffer to stop writes to.
2538 * This prevents all writes to the buffer. Any attempt to write
2539 * to the buffer after this will fail and return NULL.
2541 * The caller should call synchronize_sched() after this.
2543 void ring_buffer_record_disable(struct ring_buffer *buffer)
2545 atomic_inc(&buffer->record_disabled);
2547 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
2550 * ring_buffer_record_enable - enable writes to the buffer
2551 * @buffer: The ring buffer to enable writes
2553 * Note, multiple disables will need the same number of enables
2554 * to truly enable the writing (much like preempt_disable).
2556 void ring_buffer_record_enable(struct ring_buffer *buffer)
2558 atomic_dec(&buffer->record_disabled);
2560 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
2563 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
2564 * @buffer: The ring buffer to stop writes to.
2565 * @cpu: The CPU buffer to stop
2567 * This prevents all writes to the buffer. Any attempt to write
2568 * to the buffer after this will fail and return NULL.
2570 * The caller should call synchronize_sched() after this.
2572 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
2574 struct ring_buffer_per_cpu *cpu_buffer;
2576 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2579 cpu_buffer = buffer->buffers[cpu];
2580 atomic_inc(&cpu_buffer->record_disabled);
2582 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
2585 * ring_buffer_record_enable_cpu - enable writes to the buffer
2586 * @buffer: The ring buffer to enable writes
2587 * @cpu: The CPU to enable.
2589 * Note, multiple disables will need the same number of enables
2590 * to truly enable the writing (much like preempt_disable).
2592 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
2594 struct ring_buffer_per_cpu *cpu_buffer;
2596 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2599 cpu_buffer = buffer->buffers[cpu];
2600 atomic_dec(&cpu_buffer->record_disabled);
2602 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
2605 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
2606 * @buffer: The ring buffer
2607 * @cpu: The per CPU buffer to get the entries from.
2609 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
2611 struct ring_buffer_per_cpu *cpu_buffer;
2614 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2617 cpu_buffer = buffer->buffers[cpu];
2618 ret = (local_read(&cpu_buffer->entries) - local_read(&cpu_buffer->overrun))
2623 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
2626 * ring_buffer_overrun_cpu - get the number of overruns in a cpu_buffer
2627 * @buffer: The ring buffer
2628 * @cpu: The per CPU buffer to get the number of overruns from
2630 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
2632 struct ring_buffer_per_cpu *cpu_buffer;
2635 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2638 cpu_buffer = buffer->buffers[cpu];
2639 ret = local_read(&cpu_buffer->overrun);
2643 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
2646 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by commits
2647 * @buffer: The ring buffer
2648 * @cpu: The per CPU buffer to get the number of overruns from
2651 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
2653 struct ring_buffer_per_cpu *cpu_buffer;
2656 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2659 cpu_buffer = buffer->buffers[cpu];
2660 ret = local_read(&cpu_buffer->commit_overrun);
2664 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
2667 * ring_buffer_entries - get the number of entries in a buffer
2668 * @buffer: The ring buffer
2670 * Returns the total number of entries in the ring buffer
2673 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
2675 struct ring_buffer_per_cpu *cpu_buffer;
2676 unsigned long entries = 0;
2679 /* if you care about this being correct, lock the buffer */
2680 for_each_buffer_cpu(buffer, cpu) {
2681 cpu_buffer = buffer->buffers[cpu];
2682 entries += (local_read(&cpu_buffer->entries) -
2683 local_read(&cpu_buffer->overrun)) - cpu_buffer->read;
2688 EXPORT_SYMBOL_GPL(ring_buffer_entries);
2691 * ring_buffer_overruns - get the number of overruns in buffer
2692 * @buffer: The ring buffer
2694 * Returns the total number of overruns in the ring buffer
2697 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
2699 struct ring_buffer_per_cpu *cpu_buffer;
2700 unsigned long overruns = 0;
2703 /* if you care about this being correct, lock the buffer */
2704 for_each_buffer_cpu(buffer, cpu) {
2705 cpu_buffer = buffer->buffers[cpu];
2706 overruns += local_read(&cpu_buffer->overrun);
2711 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
2713 static void rb_iter_reset(struct ring_buffer_iter *iter)
2715 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2717 /* Iterator usage is expected to have record disabled */
2718 if (list_empty(&cpu_buffer->reader_page->list)) {
2719 iter->head_page = rb_set_head_page(cpu_buffer);
2720 if (unlikely(!iter->head_page))
2722 iter->head = iter->head_page->read;
2724 iter->head_page = cpu_buffer->reader_page;
2725 iter->head = cpu_buffer->reader_page->read;
2728 iter->read_stamp = cpu_buffer->read_stamp;
2730 iter->read_stamp = iter->head_page->page->time_stamp;
2731 iter->cache_reader_page = cpu_buffer->reader_page;
2732 iter->cache_read = cpu_buffer->read;
2736 * ring_buffer_iter_reset - reset an iterator
2737 * @iter: The iterator to reset
2739 * Resets the iterator, so that it will start from the beginning
2742 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
2744 struct ring_buffer_per_cpu *cpu_buffer;
2745 unsigned long flags;
2750 cpu_buffer = iter->cpu_buffer;
2752 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2753 rb_iter_reset(iter);
2754 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2756 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
2759 * ring_buffer_iter_empty - check if an iterator has no more to read
2760 * @iter: The iterator to check
2762 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
2764 struct ring_buffer_per_cpu *cpu_buffer;
2766 cpu_buffer = iter->cpu_buffer;
2768 return iter->head_page == cpu_buffer->commit_page &&
2769 iter->head == rb_commit_index(cpu_buffer);
2771 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
2774 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2775 struct ring_buffer_event *event)
2779 switch (event->type_len) {
2780 case RINGBUF_TYPE_PADDING:
2783 case RINGBUF_TYPE_TIME_EXTEND:
2784 delta = event->array[0];
2786 delta += event->time_delta;
2787 cpu_buffer->read_stamp += delta;
2790 case RINGBUF_TYPE_TIME_STAMP:
2791 /* FIXME: not implemented */
2794 case RINGBUF_TYPE_DATA:
2795 cpu_buffer->read_stamp += event->time_delta;
2805 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
2806 struct ring_buffer_event *event)
2810 switch (event->type_len) {
2811 case RINGBUF_TYPE_PADDING:
2814 case RINGBUF_TYPE_TIME_EXTEND:
2815 delta = event->array[0];
2817 delta += event->time_delta;
2818 iter->read_stamp += delta;
2821 case RINGBUF_TYPE_TIME_STAMP:
2822 /* FIXME: not implemented */
2825 case RINGBUF_TYPE_DATA:
2826 iter->read_stamp += event->time_delta;
2835 static struct buffer_page *
2836 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
2838 struct buffer_page *reader = NULL;
2839 unsigned long flags;
2843 local_irq_save(flags);
2844 arch_spin_lock(&cpu_buffer->lock);
2848 * This should normally only loop twice. But because the
2849 * start of the reader inserts an empty page, it causes
2850 * a case where we will loop three times. There should be no
2851 * reason to loop four times (that I know of).
2853 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
2858 reader = cpu_buffer->reader_page;
2860 /* If there's more to read, return this page */
2861 if (cpu_buffer->reader_page->read < rb_page_size(reader))
2864 /* Never should we have an index greater than the size */
2865 if (RB_WARN_ON(cpu_buffer,
2866 cpu_buffer->reader_page->read > rb_page_size(reader)))
2869 /* check if we caught up to the tail */
2871 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
2875 * Reset the reader page to size zero.
2877 local_set(&cpu_buffer->reader_page->write, 0);
2878 local_set(&cpu_buffer->reader_page->entries, 0);
2879 local_set(&cpu_buffer->reader_page->page->commit, 0);
2883 * Splice the empty reader page into the list around the head.
2885 reader = rb_set_head_page(cpu_buffer);
2886 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
2887 cpu_buffer->reader_page->list.prev = reader->list.prev;
2890 * cpu_buffer->pages just needs to point to the buffer, it
2891 * has no specific buffer page to point to. Lets move it out
2892 * of our way so we don't accidently swap it.
2894 cpu_buffer->pages = reader->list.prev;
2896 /* The reader page will be pointing to the new head */
2897 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
2900 * Here's the tricky part.
2902 * We need to move the pointer past the header page.
2903 * But we can only do that if a writer is not currently
2904 * moving it. The page before the header page has the
2905 * flag bit '1' set if it is pointing to the page we want.
2906 * but if the writer is in the process of moving it
2907 * than it will be '2' or already moved '0'.
2910 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
2913 * If we did not convert it, then we must try again.
2919 * Yeah! We succeeded in replacing the page.
2921 * Now make the new head point back to the reader page.
2923 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
2924 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
2926 /* Finally update the reader page to the new head */
2927 cpu_buffer->reader_page = reader;
2928 rb_reset_reader_page(cpu_buffer);
2933 arch_spin_unlock(&cpu_buffer->lock);
2934 local_irq_restore(flags);
2939 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
2941 struct ring_buffer_event *event;
2942 struct buffer_page *reader;
2945 reader = rb_get_reader_page(cpu_buffer);
2947 /* This function should not be called when buffer is empty */
2948 if (RB_WARN_ON(cpu_buffer, !reader))
2951 event = rb_reader_event(cpu_buffer);
2953 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
2956 rb_update_read_stamp(cpu_buffer, event);
2958 length = rb_event_length(event);
2959 cpu_buffer->reader_page->read += length;
2962 static void rb_advance_iter(struct ring_buffer_iter *iter)
2964 struct ring_buffer *buffer;
2965 struct ring_buffer_per_cpu *cpu_buffer;
2966 struct ring_buffer_event *event;
2969 cpu_buffer = iter->cpu_buffer;
2970 buffer = cpu_buffer->buffer;
2973 * Check if we are at the end of the buffer.
2975 if (iter->head >= rb_page_size(iter->head_page)) {
2976 /* discarded commits can make the page empty */
2977 if (iter->head_page == cpu_buffer->commit_page)
2983 event = rb_iter_head_event(iter);
2985 length = rb_event_length(event);
2988 * This should not be called to advance the header if we are
2989 * at the tail of the buffer.
2991 if (RB_WARN_ON(cpu_buffer,
2992 (iter->head_page == cpu_buffer->commit_page) &&
2993 (iter->head + length > rb_commit_index(cpu_buffer))))
2996 rb_update_iter_read_stamp(iter, event);
2998 iter->head += length;
3000 /* check for end of page padding */
3001 if ((iter->head >= rb_page_size(iter->head_page)) &&
3002 (iter->head_page != cpu_buffer->commit_page))
3003 rb_advance_iter(iter);
3006 static struct ring_buffer_event *
3007 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts)
3009 struct ring_buffer_event *event;
3010 struct buffer_page *reader;
3015 * We repeat when a timestamp is encountered. It is possible
3016 * to get multiple timestamps from an interrupt entering just
3017 * as one timestamp is about to be written, or from discarded
3018 * commits. The most that we can have is the number on a single page.
3020 if (RB_WARN_ON(cpu_buffer, ++nr_loops > RB_TIMESTAMPS_PER_PAGE))
3023 reader = rb_get_reader_page(cpu_buffer);
3027 event = rb_reader_event(cpu_buffer);
3029 switch (event->type_len) {
3030 case RINGBUF_TYPE_PADDING:
3031 if (rb_null_event(event))
3032 RB_WARN_ON(cpu_buffer, 1);
3034 * Because the writer could be discarding every
3035 * event it creates (which would probably be bad)
3036 * if we were to go back to "again" then we may never
3037 * catch up, and will trigger the warn on, or lock
3038 * the box. Return the padding, and we will release
3039 * the current locks, and try again.
3043 case RINGBUF_TYPE_TIME_EXTEND:
3044 /* Internal data, OK to advance */
3045 rb_advance_reader(cpu_buffer);
3048 case RINGBUF_TYPE_TIME_STAMP:
3049 /* FIXME: not implemented */
3050 rb_advance_reader(cpu_buffer);
3053 case RINGBUF_TYPE_DATA:
3055 *ts = cpu_buffer->read_stamp + event->time_delta;
3056 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3057 cpu_buffer->cpu, ts);
3067 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3069 static struct ring_buffer_event *
3070 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3072 struct ring_buffer *buffer;
3073 struct ring_buffer_per_cpu *cpu_buffer;
3074 struct ring_buffer_event *event;
3077 cpu_buffer = iter->cpu_buffer;
3078 buffer = cpu_buffer->buffer;
3081 * Check if someone performed a consuming read to
3082 * the buffer. A consuming read invalidates the iterator
3083 * and we need to reset the iterator in this case.
3085 if (unlikely(iter->cache_read != cpu_buffer->read ||
3086 iter->cache_reader_page != cpu_buffer->reader_page))
3087 rb_iter_reset(iter);
3090 if (ring_buffer_iter_empty(iter))
3094 * We repeat when a timestamp is encountered.
3095 * We can get multiple timestamps by nested interrupts or also
3096 * if filtering is on (discarding commits). Since discarding
3097 * commits can be frequent we can get a lot of timestamps.
3098 * But we limit them by not adding timestamps if they begin
3099 * at the start of a page.
3101 if (RB_WARN_ON(cpu_buffer, ++nr_loops > RB_TIMESTAMPS_PER_PAGE))
3104 if (rb_per_cpu_empty(cpu_buffer))
3107 if (iter->head >= local_read(&iter->head_page->page->commit)) {
3112 event = rb_iter_head_event(iter);
3114 switch (event->type_len) {
3115 case RINGBUF_TYPE_PADDING:
3116 if (rb_null_event(event)) {
3120 rb_advance_iter(iter);
3123 case RINGBUF_TYPE_TIME_EXTEND:
3124 /* Internal data, OK to advance */
3125 rb_advance_iter(iter);
3128 case RINGBUF_TYPE_TIME_STAMP:
3129 /* FIXME: not implemented */
3130 rb_advance_iter(iter);
3133 case RINGBUF_TYPE_DATA:
3135 *ts = iter->read_stamp + event->time_delta;
3136 ring_buffer_normalize_time_stamp(buffer,
3137 cpu_buffer->cpu, ts);
3147 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3149 static inline int rb_ok_to_lock(void)
3152 * If an NMI die dumps out the content of the ring buffer
3153 * do not grab locks. We also permanently disable the ring
3154 * buffer too. A one time deal is all you get from reading
3155 * the ring buffer from an NMI.
3157 if (likely(!in_nmi()))
3160 tracing_off_permanent();
3165 * ring_buffer_peek - peek at the next event to be read
3166 * @buffer: The ring buffer to read
3167 * @cpu: The cpu to peak at
3168 * @ts: The timestamp counter of this event.
3170 * This will return the event that will be read next, but does
3171 * not consume the data.
3173 struct ring_buffer_event *
3174 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts)
3176 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3177 struct ring_buffer_event *event;
3178 unsigned long flags;
3181 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3184 dolock = rb_ok_to_lock();
3186 local_irq_save(flags);
3188 spin_lock(&cpu_buffer->reader_lock);
3189 event = rb_buffer_peek(cpu_buffer, ts);
3190 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3191 rb_advance_reader(cpu_buffer);
3193 spin_unlock(&cpu_buffer->reader_lock);
3194 local_irq_restore(flags);
3196 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3203 * ring_buffer_iter_peek - peek at the next event to be read
3204 * @iter: The ring buffer iterator
3205 * @ts: The timestamp counter of this event.
3207 * This will return the event that will be read next, but does
3208 * not increment the iterator.
3210 struct ring_buffer_event *
3211 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3213 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3214 struct ring_buffer_event *event;
3215 unsigned long flags;
3218 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3219 event = rb_iter_peek(iter, ts);
3220 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3222 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3229 * ring_buffer_consume - return an event and consume it
3230 * @buffer: The ring buffer to get the next event from
3232 * Returns the next event in the ring buffer, and that event is consumed.
3233 * Meaning, that sequential reads will keep returning a different event,
3234 * and eventually empty the ring buffer if the producer is slower.
3236 struct ring_buffer_event *
3237 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts)
3239 struct ring_buffer_per_cpu *cpu_buffer;
3240 struct ring_buffer_event *event = NULL;
3241 unsigned long flags;
3244 dolock = rb_ok_to_lock();
3247 /* might be called in atomic */
3250 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3253 cpu_buffer = buffer->buffers[cpu];
3254 local_irq_save(flags);
3256 spin_lock(&cpu_buffer->reader_lock);
3258 event = rb_buffer_peek(cpu_buffer, ts);
3260 rb_advance_reader(cpu_buffer);
3263 spin_unlock(&cpu_buffer->reader_lock);
3264 local_irq_restore(flags);
3269 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3274 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3277 * ring_buffer_read_start - start a non consuming read of the buffer
3278 * @buffer: The ring buffer to read from
3279 * @cpu: The cpu buffer to iterate over
3281 * This starts up an iteration through the buffer. It also disables
3282 * the recording to the buffer until the reading is finished.
3283 * This prevents the reading from being corrupted. This is not
3284 * a consuming read, so a producer is not expected.
3286 * Must be paired with ring_buffer_finish.
3288 struct ring_buffer_iter *
3289 ring_buffer_read_start(struct ring_buffer *buffer, int cpu)
3291 struct ring_buffer_per_cpu *cpu_buffer;
3292 struct ring_buffer_iter *iter;
3293 unsigned long flags;
3295 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3298 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3302 cpu_buffer = buffer->buffers[cpu];
3304 iter->cpu_buffer = cpu_buffer;
3306 atomic_inc(&cpu_buffer->record_disabled);
3307 synchronize_sched();
3309 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3310 arch_spin_lock(&cpu_buffer->lock);
3311 rb_iter_reset(iter);
3312 arch_spin_unlock(&cpu_buffer->lock);
3313 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3317 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
3320 * ring_buffer_finish - finish reading the iterator of the buffer
3321 * @iter: The iterator retrieved by ring_buffer_start
3323 * This re-enables the recording to the buffer, and frees the
3327 ring_buffer_read_finish(struct ring_buffer_iter *iter)
3329 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3331 atomic_dec(&cpu_buffer->record_disabled);
3334 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
3337 * ring_buffer_read - read the next item in the ring buffer by the iterator
3338 * @iter: The ring buffer iterator
3339 * @ts: The time stamp of the event read.
3341 * This reads the next event in the ring buffer and increments the iterator.
3343 struct ring_buffer_event *
3344 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
3346 struct ring_buffer_event *event;
3347 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3348 unsigned long flags;
3350 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3352 event = rb_iter_peek(iter, ts);
3356 if (event->type_len == RINGBUF_TYPE_PADDING)
3359 rb_advance_iter(iter);
3361 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3365 EXPORT_SYMBOL_GPL(ring_buffer_read);
3368 * ring_buffer_size - return the size of the ring buffer (in bytes)
3369 * @buffer: The ring buffer.
3371 unsigned long ring_buffer_size(struct ring_buffer *buffer)
3373 return BUF_PAGE_SIZE * buffer->pages;
3375 EXPORT_SYMBOL_GPL(ring_buffer_size);
3378 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
3380 rb_head_page_deactivate(cpu_buffer);
3382 cpu_buffer->head_page
3383 = list_entry(cpu_buffer->pages, struct buffer_page, list);
3384 local_set(&cpu_buffer->head_page->write, 0);
3385 local_set(&cpu_buffer->head_page->entries, 0);
3386 local_set(&cpu_buffer->head_page->page->commit, 0);
3388 cpu_buffer->head_page->read = 0;
3390 cpu_buffer->tail_page = cpu_buffer->head_page;
3391 cpu_buffer->commit_page = cpu_buffer->head_page;
3393 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
3394 local_set(&cpu_buffer->reader_page->write, 0);
3395 local_set(&cpu_buffer->reader_page->entries, 0);
3396 local_set(&cpu_buffer->reader_page->page->commit, 0);
3397 cpu_buffer->reader_page->read = 0;
3399 local_set(&cpu_buffer->commit_overrun, 0);
3400 local_set(&cpu_buffer->overrun, 0);
3401 local_set(&cpu_buffer->entries, 0);
3402 local_set(&cpu_buffer->committing, 0);
3403 local_set(&cpu_buffer->commits, 0);
3404 cpu_buffer->read = 0;
3406 cpu_buffer->write_stamp = 0;
3407 cpu_buffer->read_stamp = 0;
3409 rb_head_page_activate(cpu_buffer);
3413 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
3414 * @buffer: The ring buffer to reset a per cpu buffer of
3415 * @cpu: The CPU buffer to be reset
3417 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
3419 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3420 unsigned long flags;
3422 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3425 atomic_inc(&cpu_buffer->record_disabled);
3427 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3429 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
3432 arch_spin_lock(&cpu_buffer->lock);
3434 rb_reset_cpu(cpu_buffer);
3436 arch_spin_unlock(&cpu_buffer->lock);
3439 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3441 atomic_dec(&cpu_buffer->record_disabled);
3443 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
3446 * ring_buffer_reset - reset a ring buffer
3447 * @buffer: The ring buffer to reset all cpu buffers
3449 void ring_buffer_reset(struct ring_buffer *buffer)
3453 for_each_buffer_cpu(buffer, cpu)
3454 ring_buffer_reset_cpu(buffer, cpu);
3456 EXPORT_SYMBOL_GPL(ring_buffer_reset);
3459 * rind_buffer_empty - is the ring buffer empty?
3460 * @buffer: The ring buffer to test
3462 int ring_buffer_empty(struct ring_buffer *buffer)
3464 struct ring_buffer_per_cpu *cpu_buffer;
3465 unsigned long flags;
3470 dolock = rb_ok_to_lock();
3472 /* yes this is racy, but if you don't like the race, lock the buffer */
3473 for_each_buffer_cpu(buffer, cpu) {
3474 cpu_buffer = buffer->buffers[cpu];
3475 local_irq_save(flags);
3477 spin_lock(&cpu_buffer->reader_lock);
3478 ret = rb_per_cpu_empty(cpu_buffer);
3480 spin_unlock(&cpu_buffer->reader_lock);
3481 local_irq_restore(flags);
3489 EXPORT_SYMBOL_GPL(ring_buffer_empty);
3492 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
3493 * @buffer: The ring buffer
3494 * @cpu: The CPU buffer to test
3496 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
3498 struct ring_buffer_per_cpu *cpu_buffer;
3499 unsigned long flags;
3503 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3506 dolock = rb_ok_to_lock();
3508 cpu_buffer = buffer->buffers[cpu];
3509 local_irq_save(flags);
3511 spin_lock(&cpu_buffer->reader_lock);
3512 ret = rb_per_cpu_empty(cpu_buffer);
3514 spin_unlock(&cpu_buffer->reader_lock);
3515 local_irq_restore(flags);
3519 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
3521 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3523 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
3524 * @buffer_a: One buffer to swap with
3525 * @buffer_b: The other buffer to swap with
3527 * This function is useful for tracers that want to take a "snapshot"
3528 * of a CPU buffer and has another back up buffer lying around.
3529 * it is expected that the tracer handles the cpu buffer not being
3530 * used at the moment.
3532 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
3533 struct ring_buffer *buffer_b, int cpu)
3535 struct ring_buffer_per_cpu *cpu_buffer_a;
3536 struct ring_buffer_per_cpu *cpu_buffer_b;
3539 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
3540 !cpumask_test_cpu(cpu, buffer_b->cpumask))
3543 /* At least make sure the two buffers are somewhat the same */
3544 if (buffer_a->pages != buffer_b->pages)
3549 if (ring_buffer_flags != RB_BUFFERS_ON)
3552 if (atomic_read(&buffer_a->record_disabled))
3555 if (atomic_read(&buffer_b->record_disabled))
3558 cpu_buffer_a = buffer_a->buffers[cpu];
3559 cpu_buffer_b = buffer_b->buffers[cpu];
3561 if (atomic_read(&cpu_buffer_a->record_disabled))
3564 if (atomic_read(&cpu_buffer_b->record_disabled))
3568 * We can't do a synchronize_sched here because this
3569 * function can be called in atomic context.
3570 * Normally this will be called from the same CPU as cpu.
3571 * If not it's up to the caller to protect this.
3573 atomic_inc(&cpu_buffer_a->record_disabled);
3574 atomic_inc(&cpu_buffer_b->record_disabled);
3577 if (local_read(&cpu_buffer_a->committing))
3579 if (local_read(&cpu_buffer_b->committing))
3582 buffer_a->buffers[cpu] = cpu_buffer_b;
3583 buffer_b->buffers[cpu] = cpu_buffer_a;
3585 cpu_buffer_b->buffer = buffer_a;
3586 cpu_buffer_a->buffer = buffer_b;
3591 atomic_dec(&cpu_buffer_a->record_disabled);
3592 atomic_dec(&cpu_buffer_b->record_disabled);
3596 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
3597 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
3600 * ring_buffer_alloc_read_page - allocate a page to read from buffer
3601 * @buffer: the buffer to allocate for.
3603 * This function is used in conjunction with ring_buffer_read_page.
3604 * When reading a full page from the ring buffer, these functions
3605 * can be used to speed up the process. The calling function should
3606 * allocate a few pages first with this function. Then when it
3607 * needs to get pages from the ring buffer, it passes the result
3608 * of this function into ring_buffer_read_page, which will swap
3609 * the page that was allocated, with the read page of the buffer.
3612 * The page allocated, or NULL on error.
3614 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer)
3616 struct buffer_data_page *bpage;
3619 addr = __get_free_page(GFP_KERNEL);
3623 bpage = (void *)addr;
3625 rb_init_page(bpage);
3629 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
3632 * ring_buffer_free_read_page - free an allocated read page
3633 * @buffer: the buffer the page was allocate for
3634 * @data: the page to free
3636 * Free a page allocated from ring_buffer_alloc_read_page.
3638 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
3640 free_page((unsigned long)data);
3642 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
3645 * ring_buffer_read_page - extract a page from the ring buffer
3646 * @buffer: buffer to extract from
3647 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
3648 * @len: amount to extract
3649 * @cpu: the cpu of the buffer to extract
3650 * @full: should the extraction only happen when the page is full.
3652 * This function will pull out a page from the ring buffer and consume it.
3653 * @data_page must be the address of the variable that was returned
3654 * from ring_buffer_alloc_read_page. This is because the page might be used
3655 * to swap with a page in the ring buffer.
3658 * rpage = ring_buffer_alloc_read_page(buffer);
3661 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
3663 * process_page(rpage, ret);
3665 * When @full is set, the function will not return true unless
3666 * the writer is off the reader page.
3668 * Note: it is up to the calling functions to handle sleeps and wakeups.
3669 * The ring buffer can be used anywhere in the kernel and can not
3670 * blindly call wake_up. The layer that uses the ring buffer must be
3671 * responsible for that.
3674 * >=0 if data has been transferred, returns the offset of consumed data.
3675 * <0 if no data has been transferred.
3677 int ring_buffer_read_page(struct ring_buffer *buffer,
3678 void **data_page, size_t len, int cpu, int full)
3680 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3681 struct ring_buffer_event *event;
3682 struct buffer_data_page *bpage;
3683 struct buffer_page *reader;
3684 unsigned long flags;
3685 unsigned int commit;
3690 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3694 * If len is not big enough to hold the page header, then
3695 * we can not copy anything.
3697 if (len <= BUF_PAGE_HDR_SIZE)
3700 len -= BUF_PAGE_HDR_SIZE;
3709 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3711 reader = rb_get_reader_page(cpu_buffer);
3715 event = rb_reader_event(cpu_buffer);
3717 read = reader->read;
3718 commit = rb_page_commit(reader);
3721 * If this page has been partially read or
3722 * if len is not big enough to read the rest of the page or
3723 * a writer is still on the page, then
3724 * we must copy the data from the page to the buffer.
3725 * Otherwise, we can simply swap the page with the one passed in.
3727 if (read || (len < (commit - read)) ||
3728 cpu_buffer->reader_page == cpu_buffer->commit_page) {
3729 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
3730 unsigned int rpos = read;
3731 unsigned int pos = 0;
3737 if (len > (commit - read))
3738 len = (commit - read);
3740 size = rb_event_length(event);
3745 /* save the current timestamp, since the user will need it */
3746 save_timestamp = cpu_buffer->read_stamp;
3748 /* Need to copy one event at a time */
3750 memcpy(bpage->data + pos, rpage->data + rpos, size);
3754 rb_advance_reader(cpu_buffer);
3755 rpos = reader->read;
3758 event = rb_reader_event(cpu_buffer);
3759 size = rb_event_length(event);
3760 } while (len > size);
3763 local_set(&bpage->commit, pos);
3764 bpage->time_stamp = save_timestamp;
3766 /* we copied everything to the beginning */
3769 /* update the entry counter */
3770 cpu_buffer->read += rb_page_entries(reader);
3772 /* swap the pages */
3773 rb_init_page(bpage);
3774 bpage = reader->page;
3775 reader->page = *data_page;
3776 local_set(&reader->write, 0);
3777 local_set(&reader->entries, 0);
3784 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3789 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
3791 #ifdef CONFIG_TRACING
3793 rb_simple_read(struct file *filp, char __user *ubuf,
3794 size_t cnt, loff_t *ppos)
3796 unsigned long *p = filp->private_data;
3800 if (test_bit(RB_BUFFERS_DISABLED_BIT, p))
3801 r = sprintf(buf, "permanently disabled\n");
3803 r = sprintf(buf, "%d\n", test_bit(RB_BUFFERS_ON_BIT, p));
3805 return simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
3809 rb_simple_write(struct file *filp, const char __user *ubuf,
3810 size_t cnt, loff_t *ppos)
3812 unsigned long *p = filp->private_data;
3817 if (cnt >= sizeof(buf))
3820 if (copy_from_user(&buf, ubuf, cnt))
3825 ret = strict_strtoul(buf, 10, &val);
3830 set_bit(RB_BUFFERS_ON_BIT, p);
3832 clear_bit(RB_BUFFERS_ON_BIT, p);
3839 static const struct file_operations rb_simple_fops = {
3840 .open = tracing_open_generic,
3841 .read = rb_simple_read,
3842 .write = rb_simple_write,
3846 static __init int rb_init_debugfs(void)
3848 struct dentry *d_tracer;
3850 d_tracer = tracing_init_dentry();
3852 trace_create_file("tracing_on", 0644, d_tracer,
3853 &ring_buffer_flags, &rb_simple_fops);
3858 fs_initcall(rb_init_debugfs);
3861 #ifdef CONFIG_HOTPLUG_CPU
3862 static int rb_cpu_notify(struct notifier_block *self,
3863 unsigned long action, void *hcpu)
3865 struct ring_buffer *buffer =
3866 container_of(self, struct ring_buffer, cpu_notify);
3867 long cpu = (long)hcpu;
3870 case CPU_UP_PREPARE:
3871 case CPU_UP_PREPARE_FROZEN:
3872 if (cpumask_test_cpu(cpu, buffer->cpumask))
3875 buffer->buffers[cpu] =
3876 rb_allocate_cpu_buffer(buffer, cpu);
3877 if (!buffer->buffers[cpu]) {
3878 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
3883 cpumask_set_cpu(cpu, buffer->cpumask);
3885 case CPU_DOWN_PREPARE:
3886 case CPU_DOWN_PREPARE_FROZEN:
3889 * If we were to free the buffer, then the user would
3890 * lose any trace that was in the buffer.