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/init.h>
18 #include <linux/hash.h>
19 #include <linux/list.h>
20 #include <linux/cpu.h>
26 * The ring buffer header is special. We must manually up keep it.
28 int ring_buffer_print_entry_header(struct trace_seq *s)
32 ret = trace_seq_printf(s, "# compressed entry header\n");
33 ret = trace_seq_printf(s, "\ttype_len : 5 bits\n");
34 ret = trace_seq_printf(s, "\ttime_delta : 27 bits\n");
35 ret = trace_seq_printf(s, "\tarray : 32 bits\n");
36 ret = trace_seq_printf(s, "\n");
37 ret = trace_seq_printf(s, "\tpadding : type == %d\n",
38 RINGBUF_TYPE_PADDING);
39 ret = trace_seq_printf(s, "\ttime_extend : type == %d\n",
40 RINGBUF_TYPE_TIME_EXTEND);
41 ret = trace_seq_printf(s, "\tdata max type_len == %d\n",
42 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
48 * The ring buffer is made up of a list of pages. A separate list of pages is
49 * allocated for each CPU. A writer may only write to a buffer that is
50 * associated with the CPU it is currently executing on. A reader may read
51 * from any per cpu buffer.
53 * The reader is special. For each per cpu buffer, the reader has its own
54 * reader page. When a reader has read the entire reader page, this reader
55 * page is swapped with another page in the ring buffer.
57 * Now, as long as the writer is off the reader page, the reader can do what
58 * ever it wants with that page. The writer will never write to that page
59 * again (as long as it is out of the ring buffer).
61 * Here's some silly ASCII art.
64 * |reader| RING BUFFER
66 * +------+ +---+ +---+ +---+
75 * |reader| RING BUFFER
76 * |page |------------------v
77 * +------+ +---+ +---+ +---+
86 * |reader| RING BUFFER
87 * |page |------------------v
88 * +------+ +---+ +---+ +---+
93 * +------------------------------+
97 * |buffer| RING BUFFER
98 * |page |------------------v
99 * +------+ +---+ +---+ +---+
101 * | New +---+ +---+ +---+
104 * +------------------------------+
107 * After we make this swap, the reader can hand this page off to the splice
108 * code and be done with it. It can even allocate a new page if it needs to
109 * and swap that into the ring buffer.
111 * We will be using cmpxchg soon to make all this lockless.
116 * A fast way to enable or disable all ring buffers is to
117 * call tracing_on or tracing_off. Turning off the ring buffers
118 * prevents all ring buffers from being recorded to.
119 * Turning this switch on, makes it OK to write to the
120 * ring buffer, if the ring buffer is enabled itself.
122 * There's three layers that must be on in order to write
123 * to the ring buffer.
125 * 1) This global flag must be set.
126 * 2) The ring buffer must be enabled for recording.
127 * 3) The per cpu buffer must be enabled for recording.
129 * In case of an anomaly, this global flag has a bit set that
130 * will permantly disable all ring buffers.
134 * Global flag to disable all recording to ring buffers
135 * This has two bits: ON, DISABLED
139 * 0 0 : ring buffers are off
140 * 1 0 : ring buffers are on
141 * X 1 : ring buffers are permanently disabled
145 RB_BUFFERS_ON_BIT = 0,
146 RB_BUFFERS_DISABLED_BIT = 1,
150 RB_BUFFERS_ON = 1 << RB_BUFFERS_ON_BIT,
151 RB_BUFFERS_DISABLED = 1 << RB_BUFFERS_DISABLED_BIT,
154 static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON;
156 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
159 * tracing_on - enable all tracing buffers
161 * This function enables all tracing buffers that may have been
162 * disabled with tracing_off.
164 void tracing_on(void)
166 set_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
168 EXPORT_SYMBOL_GPL(tracing_on);
171 * tracing_off - turn off all tracing buffers
173 * This function stops all tracing buffers from recording data.
174 * It does not disable any overhead the tracers themselves may
175 * be causing. This function simply causes all recording to
176 * the ring buffers to fail.
178 void tracing_off(void)
180 clear_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
182 EXPORT_SYMBOL_GPL(tracing_off);
185 * tracing_off_permanent - permanently disable ring buffers
187 * This function, once called, will disable all ring buffers
190 void tracing_off_permanent(void)
192 set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags);
196 * tracing_is_on - show state of ring buffers enabled
198 int tracing_is_on(void)
200 return ring_buffer_flags == RB_BUFFERS_ON;
202 EXPORT_SYMBOL_GPL(tracing_is_on);
204 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
205 #define RB_ALIGNMENT 4U
206 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
207 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
209 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
210 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
213 RB_LEN_TIME_EXTEND = 8,
214 RB_LEN_TIME_STAMP = 16,
217 static inline int rb_null_event(struct ring_buffer_event *event)
219 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
222 static void rb_event_set_padding(struct ring_buffer_event *event)
224 /* padding has a NULL time_delta */
225 event->type_len = RINGBUF_TYPE_PADDING;
226 event->time_delta = 0;
230 rb_event_data_length(struct ring_buffer_event *event)
235 length = event->type_len * RB_ALIGNMENT;
237 length = event->array[0];
238 return length + RB_EVNT_HDR_SIZE;
241 /* inline for ring buffer fast paths */
243 rb_event_length(struct ring_buffer_event *event)
245 switch (event->type_len) {
246 case RINGBUF_TYPE_PADDING:
247 if (rb_null_event(event))
250 return event->array[0] + RB_EVNT_HDR_SIZE;
252 case RINGBUF_TYPE_TIME_EXTEND:
253 return RB_LEN_TIME_EXTEND;
255 case RINGBUF_TYPE_TIME_STAMP:
256 return RB_LEN_TIME_STAMP;
258 case RINGBUF_TYPE_DATA:
259 return rb_event_data_length(event);
268 * ring_buffer_event_length - return the length of the event
269 * @event: the event to get the length of
271 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
273 unsigned length = rb_event_length(event);
274 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
276 length -= RB_EVNT_HDR_SIZE;
277 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
278 length -= sizeof(event->array[0]);
281 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
283 /* inline for ring buffer fast paths */
285 rb_event_data(struct ring_buffer_event *event)
287 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
288 /* If length is in len field, then array[0] has the data */
290 return (void *)&event->array[0];
291 /* Otherwise length is in array[0] and array[1] has the data */
292 return (void *)&event->array[1];
296 * ring_buffer_event_data - return the data of the event
297 * @event: the event to get the data from
299 void *ring_buffer_event_data(struct ring_buffer_event *event)
301 return rb_event_data(event);
303 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
305 #define for_each_buffer_cpu(buffer, cpu) \
306 for_each_cpu(cpu, buffer->cpumask)
309 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
310 #define TS_DELTA_TEST (~TS_MASK)
312 struct buffer_data_page {
313 u64 time_stamp; /* page time stamp */
314 local_t commit; /* write committed index */
315 unsigned char data[]; /* data of buffer page */
319 * Note, the buffer_page list must be first. The buffer pages
320 * are allocated in cache lines, which means that each buffer
321 * page will be at the beginning of a cache line, and thus
322 * the least significant bits will be zero. We use this to
323 * add flags in the list struct pointers, to make the ring buffer
327 struct list_head list; /* list of buffer pages */
328 local_t write; /* index for next write */
329 unsigned read; /* index for next read */
330 local_t entries; /* entries on this page */
331 struct buffer_data_page *page; /* Actual data page */
335 * The buffer page counters, write and entries, must be reset
336 * atomically when crossing page boundaries. To synchronize this
337 * update, two counters are inserted into the number. One is
338 * the actual counter for the write position or count on the page.
340 * The other is a counter of updaters. Before an update happens
341 * the update partition of the counter is incremented. This will
342 * allow the updater to update the counter atomically.
344 * The counter is 20 bits, and the state data is 12.
346 #define RB_WRITE_MASK 0xfffff
347 #define RB_WRITE_INTCNT (1 << 20)
349 static void rb_init_page(struct buffer_data_page *bpage)
351 local_set(&bpage->commit, 0);
355 * ring_buffer_page_len - the size of data on the page.
356 * @page: The page to read
358 * Returns the amount of data on the page, including buffer page header.
360 size_t ring_buffer_page_len(void *page)
362 return local_read(&((struct buffer_data_page *)page)->commit)
367 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
370 static void free_buffer_page(struct buffer_page *bpage)
372 free_page((unsigned long)bpage->page);
377 * We need to fit the time_stamp delta into 27 bits.
379 static inline int test_time_stamp(u64 delta)
381 if (delta & TS_DELTA_TEST)
386 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
388 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
389 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
391 /* Max number of timestamps that can fit on a page */
392 #define RB_TIMESTAMPS_PER_PAGE (BUF_PAGE_SIZE / RB_LEN_TIME_STAMP)
394 int ring_buffer_print_page_header(struct trace_seq *s)
396 struct buffer_data_page field;
399 ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t"
400 "offset:0;\tsize:%u;\tsigned:%u;\n",
401 (unsigned int)sizeof(field.time_stamp),
402 (unsigned int)is_signed_type(u64));
404 ret = trace_seq_printf(s, "\tfield: local_t commit;\t"
405 "offset:%u;\tsize:%u;\tsigned:%u;\n",
406 (unsigned int)offsetof(typeof(field), commit),
407 (unsigned int)sizeof(field.commit),
408 (unsigned int)is_signed_type(long));
410 ret = trace_seq_printf(s, "\tfield: char data;\t"
411 "offset:%u;\tsize:%u;\tsigned:%u;\n",
412 (unsigned int)offsetof(typeof(field), data),
413 (unsigned int)BUF_PAGE_SIZE,
414 (unsigned int)is_signed_type(char));
420 * head_page == tail_page && head == tail then buffer is empty.
422 struct ring_buffer_per_cpu {
424 struct ring_buffer *buffer;
425 spinlock_t reader_lock; /* serialize readers */
427 struct lock_class_key lock_key;
428 struct list_head *pages;
429 struct buffer_page *head_page; /* read from head */
430 struct buffer_page *tail_page; /* write to tail */
431 struct buffer_page *commit_page; /* committed pages */
432 struct buffer_page *reader_page;
433 local_t commit_overrun;
441 atomic_t record_disabled;
448 atomic_t record_disabled;
449 cpumask_var_t cpumask;
451 struct lock_class_key *reader_lock_key;
455 struct ring_buffer_per_cpu **buffers;
457 #ifdef CONFIG_HOTPLUG_CPU
458 struct notifier_block cpu_notify;
463 struct ring_buffer_iter {
464 struct ring_buffer_per_cpu *cpu_buffer;
466 struct buffer_page *head_page;
470 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
471 #define RB_WARN_ON(b, cond) \
473 int _____ret = unlikely(cond); \
475 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
476 struct ring_buffer_per_cpu *__b = \
478 atomic_inc(&__b->buffer->record_disabled); \
480 atomic_inc(&b->record_disabled); \
486 /* Up this if you want to test the TIME_EXTENTS and normalization */
487 #define DEBUG_SHIFT 0
489 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
491 /* shift to debug/test normalization and TIME_EXTENTS */
492 return buffer->clock() << DEBUG_SHIFT;
495 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
499 preempt_disable_notrace();
500 time = rb_time_stamp(buffer);
501 preempt_enable_no_resched_notrace();
505 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
507 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
510 /* Just stupid testing the normalize function and deltas */
513 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
516 * Making the ring buffer lockless makes things tricky.
517 * Although writes only happen on the CPU that they are on,
518 * and they only need to worry about interrupts. Reads can
521 * The reader page is always off the ring buffer, but when the
522 * reader finishes with a page, it needs to swap its page with
523 * a new one from the buffer. The reader needs to take from
524 * the head (writes go to the tail). But if a writer is in overwrite
525 * mode and wraps, it must push the head page forward.
527 * Here lies the problem.
529 * The reader must be careful to replace only the head page, and
530 * not another one. As described at the top of the file in the
531 * ASCII art, the reader sets its old page to point to the next
532 * page after head. It then sets the page after head to point to
533 * the old reader page. But if the writer moves the head page
534 * during this operation, the reader could end up with the tail.
536 * We use cmpxchg to help prevent this race. We also do something
537 * special with the page before head. We set the LSB to 1.
539 * When the writer must push the page forward, it will clear the
540 * bit that points to the head page, move the head, and then set
541 * the bit that points to the new head page.
543 * We also don't want an interrupt coming in and moving the head
544 * page on another writer. Thus we use the second LSB to catch
547 * head->list->prev->next bit 1 bit 0
550 * Points to head page 0 1
553 * Note we can not trust the prev pointer of the head page, because:
555 * +----+ +-----+ +-----+
556 * | |------>| T |---X--->| N |
558 * +----+ +-----+ +-----+
561 * +----------| R |----------+ |
565 * Key: ---X--> HEAD flag set in pointer
570 * (see __rb_reserve_next() to see where this happens)
572 * What the above shows is that the reader just swapped out
573 * the reader page with a page in the buffer, but before it
574 * could make the new header point back to the new page added
575 * it was preempted by a writer. The writer moved forward onto
576 * the new page added by the reader and is about to move forward
579 * You can see, it is legitimate for the previous pointer of
580 * the head (or any page) not to point back to itself. But only
584 #define RB_PAGE_NORMAL 0UL
585 #define RB_PAGE_HEAD 1UL
586 #define RB_PAGE_UPDATE 2UL
589 #define RB_FLAG_MASK 3UL
591 /* PAGE_MOVED is not part of the mask */
592 #define RB_PAGE_MOVED 4UL
595 * rb_list_head - remove any bit
597 static struct list_head *rb_list_head(struct list_head *list)
599 unsigned long val = (unsigned long)list;
601 return (struct list_head *)(val & ~RB_FLAG_MASK);
605 * rb_is_head_page - test if the given page is the head page
607 * Because the reader may move the head_page pointer, we can
608 * not trust what the head page is (it may be pointing to
609 * the reader page). But if the next page is a header page,
610 * its flags will be non zero.
613 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
614 struct buffer_page *page, struct list_head *list)
618 val = (unsigned long)list->next;
620 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
621 return RB_PAGE_MOVED;
623 return val & RB_FLAG_MASK;
629 * The unique thing about the reader page, is that, if the
630 * writer is ever on it, the previous pointer never points
631 * back to the reader page.
633 static int rb_is_reader_page(struct buffer_page *page)
635 struct list_head *list = page->list.prev;
637 return rb_list_head(list->next) != &page->list;
641 * rb_set_list_to_head - set a list_head to be pointing to head.
643 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
644 struct list_head *list)
648 ptr = (unsigned long *)&list->next;
649 *ptr |= RB_PAGE_HEAD;
650 *ptr &= ~RB_PAGE_UPDATE;
654 * rb_head_page_activate - sets up head page
656 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
658 struct buffer_page *head;
660 head = cpu_buffer->head_page;
665 * Set the previous list pointer to have the HEAD flag.
667 rb_set_list_to_head(cpu_buffer, head->list.prev);
670 static void rb_list_head_clear(struct list_head *list)
672 unsigned long *ptr = (unsigned long *)&list->next;
674 *ptr &= ~RB_FLAG_MASK;
678 * rb_head_page_dactivate - clears head page ptr (for free list)
681 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
683 struct list_head *hd;
685 /* Go through the whole list and clear any pointers found. */
686 rb_list_head_clear(cpu_buffer->pages);
688 list_for_each(hd, cpu_buffer->pages)
689 rb_list_head_clear(hd);
692 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
693 struct buffer_page *head,
694 struct buffer_page *prev,
695 int old_flag, int new_flag)
697 struct list_head *list;
698 unsigned long val = (unsigned long)&head->list;
703 val &= ~RB_FLAG_MASK;
705 ret = cmpxchg((unsigned long *)&list->next,
706 val | old_flag, val | new_flag);
708 /* check if the reader took the page */
709 if ((ret & ~RB_FLAG_MASK) != val)
710 return RB_PAGE_MOVED;
712 return ret & RB_FLAG_MASK;
715 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
716 struct buffer_page *head,
717 struct buffer_page *prev,
720 return rb_head_page_set(cpu_buffer, head, prev,
721 old_flag, RB_PAGE_UPDATE);
724 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
725 struct buffer_page *head,
726 struct buffer_page *prev,
729 return rb_head_page_set(cpu_buffer, head, prev,
730 old_flag, RB_PAGE_HEAD);
733 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
734 struct buffer_page *head,
735 struct buffer_page *prev,
738 return rb_head_page_set(cpu_buffer, head, prev,
739 old_flag, RB_PAGE_NORMAL);
742 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
743 struct buffer_page **bpage)
745 struct list_head *p = rb_list_head((*bpage)->list.next);
747 *bpage = list_entry(p, struct buffer_page, list);
750 static struct buffer_page *
751 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
753 struct buffer_page *head;
754 struct buffer_page *page;
755 struct list_head *list;
758 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
762 list = cpu_buffer->pages;
763 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
766 page = head = cpu_buffer->head_page;
768 * It is possible that the writer moves the header behind
769 * where we started, and we miss in one loop.
770 * A second loop should grab the header, but we'll do
771 * three loops just because I'm paranoid.
773 for (i = 0; i < 3; i++) {
775 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
776 cpu_buffer->head_page = page;
779 rb_inc_page(cpu_buffer, &page);
780 } while (page != head);
783 RB_WARN_ON(cpu_buffer, 1);
788 static int rb_head_page_replace(struct buffer_page *old,
789 struct buffer_page *new)
791 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
795 val = *ptr & ~RB_FLAG_MASK;
798 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
804 * rb_tail_page_update - move the tail page forward
806 * Returns 1 if moved tail page, 0 if someone else did.
808 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
809 struct buffer_page *tail_page,
810 struct buffer_page *next_page)
812 struct buffer_page *old_tail;
813 unsigned long old_entries;
814 unsigned long old_write;
818 * The tail page now needs to be moved forward.
820 * We need to reset the tail page, but without messing
821 * with possible erasing of data brought in by interrupts
822 * that have moved the tail page and are currently on it.
824 * We add a counter to the write field to denote this.
826 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
827 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
830 * Just make sure we have seen our old_write and synchronize
831 * with any interrupts that come in.
836 * If the tail page is still the same as what we think
837 * it is, then it is up to us to update the tail
840 if (tail_page == cpu_buffer->tail_page) {
841 /* Zero the write counter */
842 unsigned long val = old_write & ~RB_WRITE_MASK;
843 unsigned long eval = old_entries & ~RB_WRITE_MASK;
846 * This will only succeed if an interrupt did
847 * not come in and change it. In which case, we
848 * do not want to modify it.
850 * We add (void) to let the compiler know that we do not care
851 * about the return value of these functions. We use the
852 * cmpxchg to only update if an interrupt did not already
853 * do it for us. If the cmpxchg fails, we don't care.
855 (void)local_cmpxchg(&next_page->write, old_write, val);
856 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
859 * No need to worry about races with clearing out the commit.
860 * it only can increment when a commit takes place. But that
861 * only happens in the outer most nested commit.
863 local_set(&next_page->page->commit, 0);
865 old_tail = cmpxchg(&cpu_buffer->tail_page,
866 tail_page, next_page);
868 if (old_tail == tail_page)
875 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
876 struct buffer_page *bpage)
878 unsigned long val = (unsigned long)bpage;
880 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
887 * rb_check_list - make sure a pointer to a list has the last bits zero
889 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
890 struct list_head *list)
892 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
894 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
900 * check_pages - integrity check of buffer pages
901 * @cpu_buffer: CPU buffer with pages to test
903 * As a safety measure we check to make sure the data pages have not
906 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
908 struct list_head *head = cpu_buffer->pages;
909 struct buffer_page *bpage, *tmp;
911 rb_head_page_deactivate(cpu_buffer);
913 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
915 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
918 if (rb_check_list(cpu_buffer, head))
921 list_for_each_entry_safe(bpage, tmp, head, list) {
922 if (RB_WARN_ON(cpu_buffer,
923 bpage->list.next->prev != &bpage->list))
925 if (RB_WARN_ON(cpu_buffer,
926 bpage->list.prev->next != &bpage->list))
928 if (rb_check_list(cpu_buffer, &bpage->list))
932 rb_head_page_activate(cpu_buffer);
937 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
940 struct buffer_page *bpage, *tmp;
947 for (i = 0; i < nr_pages; i++) {
948 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
949 GFP_KERNEL, cpu_to_node(cpu_buffer->cpu));
953 rb_check_bpage(cpu_buffer, bpage);
955 list_add(&bpage->list, &pages);
957 addr = __get_free_page(GFP_KERNEL);
960 bpage->page = (void *)addr;
961 rb_init_page(bpage->page);
965 * The ring buffer page list is a circular list that does not
966 * start and end with a list head. All page list items point to
969 cpu_buffer->pages = pages.next;
972 rb_check_pages(cpu_buffer);
977 list_for_each_entry_safe(bpage, tmp, &pages, list) {
978 list_del_init(&bpage->list);
979 free_buffer_page(bpage);
984 static struct ring_buffer_per_cpu *
985 rb_allocate_cpu_buffer(struct ring_buffer *buffer, int cpu)
987 struct ring_buffer_per_cpu *cpu_buffer;
988 struct buffer_page *bpage;
992 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
993 GFP_KERNEL, cpu_to_node(cpu));
997 cpu_buffer->cpu = cpu;
998 cpu_buffer->buffer = buffer;
999 spin_lock_init(&cpu_buffer->reader_lock);
1000 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1001 cpu_buffer->lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
1003 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1004 GFP_KERNEL, cpu_to_node(cpu));
1006 goto fail_free_buffer;
1008 rb_check_bpage(cpu_buffer, bpage);
1010 cpu_buffer->reader_page = bpage;
1011 addr = __get_free_page(GFP_KERNEL);
1013 goto fail_free_reader;
1014 bpage->page = (void *)addr;
1015 rb_init_page(bpage->page);
1017 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1019 ret = rb_allocate_pages(cpu_buffer, buffer->pages);
1021 goto fail_free_reader;
1023 cpu_buffer->head_page
1024 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1025 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1027 rb_head_page_activate(cpu_buffer);
1032 free_buffer_page(cpu_buffer->reader_page);
1039 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1041 struct list_head *head = cpu_buffer->pages;
1042 struct buffer_page *bpage, *tmp;
1044 free_buffer_page(cpu_buffer->reader_page);
1046 rb_head_page_deactivate(cpu_buffer);
1049 list_for_each_entry_safe(bpage, tmp, head, list) {
1050 list_del_init(&bpage->list);
1051 free_buffer_page(bpage);
1053 bpage = list_entry(head, struct buffer_page, list);
1054 free_buffer_page(bpage);
1060 #ifdef CONFIG_HOTPLUG_CPU
1061 static int rb_cpu_notify(struct notifier_block *self,
1062 unsigned long action, void *hcpu);
1066 * ring_buffer_alloc - allocate a new ring_buffer
1067 * @size: the size in bytes per cpu that is needed.
1068 * @flags: attributes to set for the ring buffer.
1070 * Currently the only flag that is available is the RB_FL_OVERWRITE
1071 * flag. This flag means that the buffer will overwrite old data
1072 * when the buffer wraps. If this flag is not set, the buffer will
1073 * drop data when the tail hits the head.
1075 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1076 struct lock_class_key *key)
1078 struct ring_buffer *buffer;
1082 /* keep it in its own cache line */
1083 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1088 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1089 goto fail_free_buffer;
1091 buffer->pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1092 buffer->flags = flags;
1093 buffer->clock = trace_clock_local;
1094 buffer->reader_lock_key = key;
1096 /* need at least two pages */
1097 if (buffer->pages < 2)
1101 * In case of non-hotplug cpu, if the ring-buffer is allocated
1102 * in early initcall, it will not be notified of secondary cpus.
1103 * In that off case, we need to allocate for all possible cpus.
1105 #ifdef CONFIG_HOTPLUG_CPU
1107 cpumask_copy(buffer->cpumask, cpu_online_mask);
1109 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1111 buffer->cpus = nr_cpu_ids;
1113 bsize = sizeof(void *) * nr_cpu_ids;
1114 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1116 if (!buffer->buffers)
1117 goto fail_free_cpumask;
1119 for_each_buffer_cpu(buffer, cpu) {
1120 buffer->buffers[cpu] =
1121 rb_allocate_cpu_buffer(buffer, cpu);
1122 if (!buffer->buffers[cpu])
1123 goto fail_free_buffers;
1126 #ifdef CONFIG_HOTPLUG_CPU
1127 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1128 buffer->cpu_notify.priority = 0;
1129 register_cpu_notifier(&buffer->cpu_notify);
1133 mutex_init(&buffer->mutex);
1138 for_each_buffer_cpu(buffer, cpu) {
1139 if (buffer->buffers[cpu])
1140 rb_free_cpu_buffer(buffer->buffers[cpu]);
1142 kfree(buffer->buffers);
1145 free_cpumask_var(buffer->cpumask);
1152 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1155 * ring_buffer_free - free a ring buffer.
1156 * @buffer: the buffer to free.
1159 ring_buffer_free(struct ring_buffer *buffer)
1165 #ifdef CONFIG_HOTPLUG_CPU
1166 unregister_cpu_notifier(&buffer->cpu_notify);
1169 for_each_buffer_cpu(buffer, cpu)
1170 rb_free_cpu_buffer(buffer->buffers[cpu]);
1174 kfree(buffer->buffers);
1175 free_cpumask_var(buffer->cpumask);
1179 EXPORT_SYMBOL_GPL(ring_buffer_free);
1181 void ring_buffer_set_clock(struct ring_buffer *buffer,
1184 buffer->clock = clock;
1187 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1190 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned nr_pages)
1192 struct buffer_page *bpage;
1193 struct list_head *p;
1196 atomic_inc(&cpu_buffer->record_disabled);
1197 synchronize_sched();
1199 spin_lock_irq(&cpu_buffer->reader_lock);
1200 rb_head_page_deactivate(cpu_buffer);
1202 for (i = 0; i < nr_pages; i++) {
1203 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1205 p = cpu_buffer->pages->next;
1206 bpage = list_entry(p, struct buffer_page, list);
1207 list_del_init(&bpage->list);
1208 free_buffer_page(bpage);
1210 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1213 rb_reset_cpu(cpu_buffer);
1214 spin_unlock_irq(&cpu_buffer->reader_lock);
1216 rb_check_pages(cpu_buffer);
1218 atomic_dec(&cpu_buffer->record_disabled);
1223 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer,
1224 struct list_head *pages, unsigned nr_pages)
1226 struct buffer_page *bpage;
1227 struct list_head *p;
1230 atomic_inc(&cpu_buffer->record_disabled);
1231 synchronize_sched();
1233 spin_lock_irq(&cpu_buffer->reader_lock);
1234 rb_head_page_deactivate(cpu_buffer);
1236 for (i = 0; i < nr_pages; i++) {
1237 if (RB_WARN_ON(cpu_buffer, list_empty(pages)))
1240 bpage = list_entry(p, struct buffer_page, list);
1241 list_del_init(&bpage->list);
1242 list_add_tail(&bpage->list, cpu_buffer->pages);
1244 rb_reset_cpu(cpu_buffer);
1245 spin_unlock_irq(&cpu_buffer->reader_lock);
1247 rb_check_pages(cpu_buffer);
1249 atomic_dec(&cpu_buffer->record_disabled);
1253 * ring_buffer_resize - resize the ring buffer
1254 * @buffer: the buffer to resize.
1255 * @size: the new size.
1257 * The tracer is responsible for making sure that the buffer is
1258 * not being used while changing the size.
1259 * Note: We may be able to change the above requirement by using
1260 * RCU synchronizations.
1262 * Minimum size is 2 * BUF_PAGE_SIZE.
1264 * Returns -1 on failure.
1266 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size)
1268 struct ring_buffer_per_cpu *cpu_buffer;
1269 unsigned nr_pages, rm_pages, new_pages;
1270 struct buffer_page *bpage, *tmp;
1271 unsigned long buffer_size;
1277 * Always succeed at resizing a non-existent buffer:
1282 size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1283 size *= BUF_PAGE_SIZE;
1284 buffer_size = buffer->pages * BUF_PAGE_SIZE;
1286 /* we need a minimum of two pages */
1287 if (size < BUF_PAGE_SIZE * 2)
1288 size = BUF_PAGE_SIZE * 2;
1290 if (size == buffer_size)
1293 mutex_lock(&buffer->mutex);
1296 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1298 if (size < buffer_size) {
1300 /* easy case, just free pages */
1301 if (RB_WARN_ON(buffer, nr_pages >= buffer->pages))
1304 rm_pages = buffer->pages - nr_pages;
1306 for_each_buffer_cpu(buffer, cpu) {
1307 cpu_buffer = buffer->buffers[cpu];
1308 rb_remove_pages(cpu_buffer, rm_pages);
1314 * This is a bit more difficult. We only want to add pages
1315 * when we can allocate enough for all CPUs. We do this
1316 * by allocating all the pages and storing them on a local
1317 * link list. If we succeed in our allocation, then we
1318 * add these pages to the cpu_buffers. Otherwise we just free
1319 * them all and return -ENOMEM;
1321 if (RB_WARN_ON(buffer, nr_pages <= buffer->pages))
1324 new_pages = nr_pages - buffer->pages;
1326 for_each_buffer_cpu(buffer, cpu) {
1327 for (i = 0; i < new_pages; i++) {
1328 bpage = kzalloc_node(ALIGN(sizeof(*bpage),
1330 GFP_KERNEL, cpu_to_node(cpu));
1333 list_add(&bpage->list, &pages);
1334 addr = __get_free_page(GFP_KERNEL);
1337 bpage->page = (void *)addr;
1338 rb_init_page(bpage->page);
1342 for_each_buffer_cpu(buffer, cpu) {
1343 cpu_buffer = buffer->buffers[cpu];
1344 rb_insert_pages(cpu_buffer, &pages, new_pages);
1347 if (RB_WARN_ON(buffer, !list_empty(&pages)))
1351 buffer->pages = nr_pages;
1353 mutex_unlock(&buffer->mutex);
1358 list_for_each_entry_safe(bpage, tmp, &pages, list) {
1359 list_del_init(&bpage->list);
1360 free_buffer_page(bpage);
1363 mutex_unlock(&buffer->mutex);
1367 * Something went totally wrong, and we are too paranoid
1368 * to even clean up the mess.
1372 mutex_unlock(&buffer->mutex);
1375 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1377 static inline void *
1378 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1380 return bpage->data + index;
1383 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1385 return bpage->page->data + index;
1388 static inline struct ring_buffer_event *
1389 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1391 return __rb_page_index(cpu_buffer->reader_page,
1392 cpu_buffer->reader_page->read);
1395 static inline struct ring_buffer_event *
1396 rb_iter_head_event(struct ring_buffer_iter *iter)
1398 return __rb_page_index(iter->head_page, iter->head);
1401 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1403 return local_read(&bpage->write) & RB_WRITE_MASK;
1406 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1408 return local_read(&bpage->page->commit);
1411 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1413 return local_read(&bpage->entries) & RB_WRITE_MASK;
1416 /* Size is determined by what has been commited */
1417 static inline unsigned rb_page_size(struct buffer_page *bpage)
1419 return rb_page_commit(bpage);
1422 static inline unsigned
1423 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1425 return rb_page_commit(cpu_buffer->commit_page);
1428 static inline unsigned
1429 rb_event_index(struct ring_buffer_event *event)
1431 unsigned long addr = (unsigned long)event;
1433 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1437 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1438 struct ring_buffer_event *event)
1440 unsigned long addr = (unsigned long)event;
1441 unsigned long index;
1443 index = rb_event_index(event);
1446 return cpu_buffer->commit_page->page == (void *)addr &&
1447 rb_commit_index(cpu_buffer) == index;
1451 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1453 unsigned long max_count;
1456 * We only race with interrupts and NMIs on this CPU.
1457 * If we own the commit event, then we can commit
1458 * all others that interrupted us, since the interruptions
1459 * are in stack format (they finish before they come
1460 * back to us). This allows us to do a simple loop to
1461 * assign the commit to the tail.
1464 max_count = cpu_buffer->buffer->pages * 100;
1466 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1467 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1469 if (RB_WARN_ON(cpu_buffer,
1470 rb_is_reader_page(cpu_buffer->tail_page)))
1472 local_set(&cpu_buffer->commit_page->page->commit,
1473 rb_page_write(cpu_buffer->commit_page));
1474 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1475 cpu_buffer->write_stamp =
1476 cpu_buffer->commit_page->page->time_stamp;
1477 /* add barrier to keep gcc from optimizing too much */
1480 while (rb_commit_index(cpu_buffer) !=
1481 rb_page_write(cpu_buffer->commit_page)) {
1483 local_set(&cpu_buffer->commit_page->page->commit,
1484 rb_page_write(cpu_buffer->commit_page));
1485 RB_WARN_ON(cpu_buffer,
1486 local_read(&cpu_buffer->commit_page->page->commit) &
1491 /* again, keep gcc from optimizing */
1495 * If an interrupt came in just after the first while loop
1496 * and pushed the tail page forward, we will be left with
1497 * a dangling commit that will never go forward.
1499 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1503 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1505 cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1506 cpu_buffer->reader_page->read = 0;
1509 static void rb_inc_iter(struct ring_buffer_iter *iter)
1511 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1514 * The iterator could be on the reader page (it starts there).
1515 * But the head could have moved, since the reader was
1516 * found. Check for this case and assign the iterator
1517 * to the head page instead of next.
1519 if (iter->head_page == cpu_buffer->reader_page)
1520 iter->head_page = rb_set_head_page(cpu_buffer);
1522 rb_inc_page(cpu_buffer, &iter->head_page);
1524 iter->read_stamp = iter->head_page->page->time_stamp;
1529 * ring_buffer_update_event - update event type and data
1530 * @event: the even to update
1531 * @type: the type of event
1532 * @length: the size of the event field in the ring buffer
1534 * Update the type and data fields of the event. The length
1535 * is the actual size that is written to the ring buffer,
1536 * and with this, we can determine what to place into the
1540 rb_update_event(struct ring_buffer_event *event,
1541 unsigned type, unsigned length)
1543 event->type_len = type;
1547 case RINGBUF_TYPE_PADDING:
1548 case RINGBUF_TYPE_TIME_EXTEND:
1549 case RINGBUF_TYPE_TIME_STAMP:
1553 length -= RB_EVNT_HDR_SIZE;
1554 if (length > RB_MAX_SMALL_DATA)
1555 event->array[0] = length;
1557 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
1565 * rb_handle_head_page - writer hit the head page
1567 * Returns: +1 to retry page
1572 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1573 struct buffer_page *tail_page,
1574 struct buffer_page *next_page)
1576 struct buffer_page *new_head;
1581 entries = rb_page_entries(next_page);
1584 * The hard part is here. We need to move the head
1585 * forward, and protect against both readers on
1586 * other CPUs and writers coming in via interrupts.
1588 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1592 * type can be one of four:
1593 * NORMAL - an interrupt already moved it for us
1594 * HEAD - we are the first to get here.
1595 * UPDATE - we are the interrupt interrupting
1597 * MOVED - a reader on another CPU moved the next
1598 * pointer to its reader page. Give up
1605 * We changed the head to UPDATE, thus
1606 * it is our responsibility to update
1609 local_add(entries, &cpu_buffer->overrun);
1612 * The entries will be zeroed out when we move the
1616 /* still more to do */
1619 case RB_PAGE_UPDATE:
1621 * This is an interrupt that interrupt the
1622 * previous update. Still more to do.
1625 case RB_PAGE_NORMAL:
1627 * An interrupt came in before the update
1628 * and processed this for us.
1629 * Nothing left to do.
1634 * The reader is on another CPU and just did
1635 * a swap with our next_page.
1640 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1645 * Now that we are here, the old head pointer is
1646 * set to UPDATE. This will keep the reader from
1647 * swapping the head page with the reader page.
1648 * The reader (on another CPU) will spin till
1651 * We just need to protect against interrupts
1652 * doing the job. We will set the next pointer
1653 * to HEAD. After that, we set the old pointer
1654 * to NORMAL, but only if it was HEAD before.
1655 * otherwise we are an interrupt, and only
1656 * want the outer most commit to reset it.
1658 new_head = next_page;
1659 rb_inc_page(cpu_buffer, &new_head);
1661 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1665 * Valid returns are:
1666 * HEAD - an interrupt came in and already set it.
1667 * NORMAL - One of two things:
1668 * 1) We really set it.
1669 * 2) A bunch of interrupts came in and moved
1670 * the page forward again.
1674 case RB_PAGE_NORMAL:
1678 RB_WARN_ON(cpu_buffer, 1);
1683 * It is possible that an interrupt came in,
1684 * set the head up, then more interrupts came in
1685 * and moved it again. When we get back here,
1686 * the page would have been set to NORMAL but we
1687 * just set it back to HEAD.
1689 * How do you detect this? Well, if that happened
1690 * the tail page would have moved.
1692 if (ret == RB_PAGE_NORMAL) {
1694 * If the tail had moved passed next, then we need
1695 * to reset the pointer.
1697 if (cpu_buffer->tail_page != tail_page &&
1698 cpu_buffer->tail_page != next_page)
1699 rb_head_page_set_normal(cpu_buffer, new_head,
1705 * If this was the outer most commit (the one that
1706 * changed the original pointer from HEAD to UPDATE),
1707 * then it is up to us to reset it to NORMAL.
1709 if (type == RB_PAGE_HEAD) {
1710 ret = rb_head_page_set_normal(cpu_buffer, next_page,
1713 if (RB_WARN_ON(cpu_buffer,
1714 ret != RB_PAGE_UPDATE))
1721 static unsigned rb_calculate_event_length(unsigned length)
1723 struct ring_buffer_event event; /* Used only for sizeof array */
1725 /* zero length can cause confusions */
1729 if (length > RB_MAX_SMALL_DATA)
1730 length += sizeof(event.array[0]);
1732 length += RB_EVNT_HDR_SIZE;
1733 length = ALIGN(length, RB_ALIGNMENT);
1739 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
1740 struct buffer_page *tail_page,
1741 unsigned long tail, unsigned long length)
1743 struct ring_buffer_event *event;
1746 * Only the event that crossed the page boundary
1747 * must fill the old tail_page with padding.
1749 if (tail >= BUF_PAGE_SIZE) {
1750 local_sub(length, &tail_page->write);
1754 event = __rb_page_index(tail_page, tail);
1755 kmemcheck_annotate_bitfield(event, bitfield);
1758 * If this event is bigger than the minimum size, then
1759 * we need to be careful that we don't subtract the
1760 * write counter enough to allow another writer to slip
1762 * We put in a discarded commit instead, to make sure
1763 * that this space is not used again.
1765 * If we are less than the minimum size, we don't need to
1768 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
1769 /* No room for any events */
1771 /* Mark the rest of the page with padding */
1772 rb_event_set_padding(event);
1774 /* Set the write back to the previous setting */
1775 local_sub(length, &tail_page->write);
1779 /* Put in a discarded event */
1780 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
1781 event->type_len = RINGBUF_TYPE_PADDING;
1782 /* time delta must be non zero */
1783 event->time_delta = 1;
1785 /* Set write to end of buffer */
1786 length = (tail + length) - BUF_PAGE_SIZE;
1787 local_sub(length, &tail_page->write);
1790 static struct ring_buffer_event *
1791 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
1792 unsigned long length, unsigned long tail,
1793 struct buffer_page *commit_page,
1794 struct buffer_page *tail_page, u64 *ts)
1796 struct ring_buffer *buffer = cpu_buffer->buffer;
1797 struct buffer_page *next_page;
1800 next_page = tail_page;
1802 rb_inc_page(cpu_buffer, &next_page);
1805 * If for some reason, we had an interrupt storm that made
1806 * it all the way around the buffer, bail, and warn
1809 if (unlikely(next_page == commit_page)) {
1810 local_inc(&cpu_buffer->commit_overrun);
1815 * This is where the fun begins!
1817 * We are fighting against races between a reader that
1818 * could be on another CPU trying to swap its reader
1819 * page with the buffer head.
1821 * We are also fighting against interrupts coming in and
1822 * moving the head or tail on us as well.
1824 * If the next page is the head page then we have filled
1825 * the buffer, unless the commit page is still on the
1828 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
1831 * If the commit is not on the reader page, then
1832 * move the header page.
1834 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
1836 * If we are not in overwrite mode,
1837 * this is easy, just stop here.
1839 if (!(buffer->flags & RB_FL_OVERWRITE))
1842 ret = rb_handle_head_page(cpu_buffer,
1851 * We need to be careful here too. The
1852 * commit page could still be on the reader
1853 * page. We could have a small buffer, and
1854 * have filled up the buffer with events
1855 * from interrupts and such, and wrapped.
1857 * Note, if the tail page is also the on the
1858 * reader_page, we let it move out.
1860 if (unlikely((cpu_buffer->commit_page !=
1861 cpu_buffer->tail_page) &&
1862 (cpu_buffer->commit_page ==
1863 cpu_buffer->reader_page))) {
1864 local_inc(&cpu_buffer->commit_overrun);
1870 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
1873 * Nested commits always have zero deltas, so
1874 * just reread the time stamp
1876 *ts = rb_time_stamp(buffer);
1877 next_page->page->time_stamp = *ts;
1882 rb_reset_tail(cpu_buffer, tail_page, tail, length);
1884 /* fail and let the caller try again */
1885 return ERR_PTR(-EAGAIN);
1889 rb_reset_tail(cpu_buffer, tail_page, tail, length);
1894 static struct ring_buffer_event *
1895 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
1896 unsigned type, unsigned long length, u64 *ts)
1898 struct buffer_page *tail_page, *commit_page;
1899 struct ring_buffer_event *event;
1900 unsigned long tail, write;
1902 commit_page = cpu_buffer->commit_page;
1903 /* we just need to protect against interrupts */
1905 tail_page = cpu_buffer->tail_page;
1906 write = local_add_return(length, &tail_page->write);
1908 /* set write to only the index of the write */
1909 write &= RB_WRITE_MASK;
1910 tail = write - length;
1912 /* See if we shot pass the end of this buffer page */
1913 if (write > BUF_PAGE_SIZE)
1914 return rb_move_tail(cpu_buffer, length, tail,
1915 commit_page, tail_page, ts);
1917 /* We reserved something on the buffer */
1919 event = __rb_page_index(tail_page, tail);
1920 kmemcheck_annotate_bitfield(event, bitfield);
1921 rb_update_event(event, type, length);
1923 /* The passed in type is zero for DATA */
1925 local_inc(&tail_page->entries);
1928 * If this is the first commit on the page, then update
1932 tail_page->page->time_stamp = *ts;
1938 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
1939 struct ring_buffer_event *event)
1941 unsigned long new_index, old_index;
1942 struct buffer_page *bpage;
1943 unsigned long index;
1946 new_index = rb_event_index(event);
1947 old_index = new_index + rb_event_length(event);
1948 addr = (unsigned long)event;
1951 bpage = cpu_buffer->tail_page;
1953 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
1954 unsigned long write_mask =
1955 local_read(&bpage->write) & ~RB_WRITE_MASK;
1957 * This is on the tail page. It is possible that
1958 * a write could come in and move the tail page
1959 * and write to the next page. That is fine
1960 * because we just shorten what is on this page.
1962 old_index += write_mask;
1963 new_index += write_mask;
1964 index = local_cmpxchg(&bpage->write, old_index, new_index);
1965 if (index == old_index)
1969 /* could not discard */
1974 rb_add_time_stamp(struct ring_buffer_per_cpu *cpu_buffer,
1975 u64 *ts, u64 *delta)
1977 struct ring_buffer_event *event;
1981 if (unlikely(*delta > (1ULL << 59) && !once++)) {
1982 printk(KERN_WARNING "Delta way too big! %llu"
1983 " ts=%llu write stamp = %llu\n",
1984 (unsigned long long)*delta,
1985 (unsigned long long)*ts,
1986 (unsigned long long)cpu_buffer->write_stamp);
1991 * The delta is too big, we to add a
1994 event = __rb_reserve_next(cpu_buffer,
1995 RINGBUF_TYPE_TIME_EXTEND,
2001 if (PTR_ERR(event) == -EAGAIN)
2004 /* Only a commited time event can update the write stamp */
2005 if (rb_event_is_commit(cpu_buffer, event)) {
2007 * If this is the first on the page, then it was
2008 * updated with the page itself. Try to discard it
2009 * and if we can't just make it zero.
2011 if (rb_event_index(event)) {
2012 event->time_delta = *delta & TS_MASK;
2013 event->array[0] = *delta >> TS_SHIFT;
2015 /* try to discard, since we do not need this */
2016 if (!rb_try_to_discard(cpu_buffer, event)) {
2017 /* nope, just zero it */
2018 event->time_delta = 0;
2019 event->array[0] = 0;
2022 cpu_buffer->write_stamp = *ts;
2023 /* let the caller know this was the commit */
2026 /* Try to discard the event */
2027 if (!rb_try_to_discard(cpu_buffer, event)) {
2028 /* Darn, this is just wasted space */
2029 event->time_delta = 0;
2030 event->array[0] = 0;
2040 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2042 local_inc(&cpu_buffer->committing);
2043 local_inc(&cpu_buffer->commits);
2046 static void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2048 unsigned long commits;
2050 if (RB_WARN_ON(cpu_buffer,
2051 !local_read(&cpu_buffer->committing)))
2055 commits = local_read(&cpu_buffer->commits);
2056 /* synchronize with interrupts */
2058 if (local_read(&cpu_buffer->committing) == 1)
2059 rb_set_commit_to_write(cpu_buffer);
2061 local_dec(&cpu_buffer->committing);
2063 /* synchronize with interrupts */
2067 * Need to account for interrupts coming in between the
2068 * updating of the commit page and the clearing of the
2069 * committing counter.
2071 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2072 !local_read(&cpu_buffer->committing)) {
2073 local_inc(&cpu_buffer->committing);
2078 static struct ring_buffer_event *
2079 rb_reserve_next_event(struct ring_buffer *buffer,
2080 struct ring_buffer_per_cpu *cpu_buffer,
2081 unsigned long length)
2083 struct ring_buffer_event *event;
2088 rb_start_commit(cpu_buffer);
2090 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2092 * Due to the ability to swap a cpu buffer from a buffer
2093 * it is possible it was swapped before we committed.
2094 * (committing stops a swap). We check for it here and
2095 * if it happened, we have to fail the write.
2098 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2099 local_dec(&cpu_buffer->committing);
2100 local_dec(&cpu_buffer->commits);
2105 length = rb_calculate_event_length(length);
2108 * We allow for interrupts to reenter here and do a trace.
2109 * If one does, it will cause this original code to loop
2110 * back here. Even with heavy interrupts happening, this
2111 * should only happen a few times in a row. If this happens
2112 * 1000 times in a row, there must be either an interrupt
2113 * storm or we have something buggy.
2116 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2119 ts = rb_time_stamp(cpu_buffer->buffer);
2122 * Only the first commit can update the timestamp.
2123 * Yes there is a race here. If an interrupt comes in
2124 * just after the conditional and it traces too, then it
2125 * will also check the deltas. More than one timestamp may
2126 * also be made. But only the entry that did the actual
2127 * commit will be something other than zero.
2129 if (likely(cpu_buffer->tail_page == cpu_buffer->commit_page &&
2130 rb_page_write(cpu_buffer->tail_page) ==
2131 rb_commit_index(cpu_buffer))) {
2134 diff = ts - cpu_buffer->write_stamp;
2136 /* make sure this diff is calculated here */
2139 /* Did the write stamp get updated already? */
2140 if (unlikely(ts < cpu_buffer->write_stamp))
2144 if (unlikely(test_time_stamp(delta))) {
2146 commit = rb_add_time_stamp(cpu_buffer, &ts, &delta);
2147 if (commit == -EBUSY)
2150 if (commit == -EAGAIN)
2153 RB_WARN_ON(cpu_buffer, commit < 0);
2158 event = __rb_reserve_next(cpu_buffer, 0, length, &ts);
2159 if (unlikely(PTR_ERR(event) == -EAGAIN))
2165 if (!rb_event_is_commit(cpu_buffer, event))
2168 event->time_delta = delta;
2173 rb_end_commit(cpu_buffer);
2177 #ifdef CONFIG_TRACING
2179 #define TRACE_RECURSIVE_DEPTH 16
2181 static int trace_recursive_lock(void)
2183 current->trace_recursion++;
2185 if (likely(current->trace_recursion < TRACE_RECURSIVE_DEPTH))
2188 /* Disable all tracing before we do anything else */
2189 tracing_off_permanent();
2191 printk_once(KERN_WARNING "Tracing recursion: depth[%ld]:"
2192 "HC[%lu]:SC[%lu]:NMI[%lu]\n",
2193 current->trace_recursion,
2194 hardirq_count() >> HARDIRQ_SHIFT,
2195 softirq_count() >> SOFTIRQ_SHIFT,
2202 static void trace_recursive_unlock(void)
2204 WARN_ON_ONCE(!current->trace_recursion);
2206 current->trace_recursion--;
2211 #define trace_recursive_lock() (0)
2212 #define trace_recursive_unlock() do { } while (0)
2216 static DEFINE_PER_CPU(int, rb_need_resched);
2219 * ring_buffer_lock_reserve - reserve a part of the buffer
2220 * @buffer: the ring buffer to reserve from
2221 * @length: the length of the data to reserve (excluding event header)
2223 * Returns a reseverd event on the ring buffer to copy directly to.
2224 * The user of this interface will need to get the body to write into
2225 * and can use the ring_buffer_event_data() interface.
2227 * The length is the length of the data needed, not the event length
2228 * which also includes the event header.
2230 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2231 * If NULL is returned, then nothing has been allocated or locked.
2233 struct ring_buffer_event *
2234 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2236 struct ring_buffer_per_cpu *cpu_buffer;
2237 struct ring_buffer_event *event;
2240 if (ring_buffer_flags != RB_BUFFERS_ON)
2243 if (atomic_read(&buffer->record_disabled))
2246 /* If we are tracing schedule, we don't want to recurse */
2247 resched = ftrace_preempt_disable();
2249 if (trace_recursive_lock())
2252 cpu = raw_smp_processor_id();
2254 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2257 cpu_buffer = buffer->buffers[cpu];
2259 if (atomic_read(&cpu_buffer->record_disabled))
2262 if (length > BUF_MAX_DATA_SIZE)
2265 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2270 * Need to store resched state on this cpu.
2271 * Only the first needs to.
2274 if (preempt_count() == 1)
2275 per_cpu(rb_need_resched, cpu) = resched;
2280 trace_recursive_unlock();
2283 ftrace_preempt_enable(resched);
2286 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2289 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2290 struct ring_buffer_event *event)
2293 * The event first in the commit queue updates the
2296 if (rb_event_is_commit(cpu_buffer, event))
2297 cpu_buffer->write_stamp += event->time_delta;
2300 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2301 struct ring_buffer_event *event)
2303 local_inc(&cpu_buffer->entries);
2304 rb_update_write_stamp(cpu_buffer, event);
2305 rb_end_commit(cpu_buffer);
2309 * ring_buffer_unlock_commit - commit a reserved
2310 * @buffer: The buffer to commit to
2311 * @event: The event pointer to commit.
2313 * This commits the data to the ring buffer, and releases any locks held.
2315 * Must be paired with ring_buffer_lock_reserve.
2317 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2318 struct ring_buffer_event *event)
2320 struct ring_buffer_per_cpu *cpu_buffer;
2321 int cpu = raw_smp_processor_id();
2323 cpu_buffer = buffer->buffers[cpu];
2325 rb_commit(cpu_buffer, event);
2327 trace_recursive_unlock();
2330 * Only the last preempt count needs to restore preemption.
2332 if (preempt_count() == 1)
2333 ftrace_preempt_enable(per_cpu(rb_need_resched, cpu));
2335 preempt_enable_no_resched_notrace();
2339 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2341 static inline void rb_event_discard(struct ring_buffer_event *event)
2343 /* array[0] holds the actual length for the discarded event */
2344 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2345 event->type_len = RINGBUF_TYPE_PADDING;
2346 /* time delta must be non zero */
2347 if (!event->time_delta)
2348 event->time_delta = 1;
2352 * Decrement the entries to the page that an event is on.
2353 * The event does not even need to exist, only the pointer
2354 * to the page it is on. This may only be called before the commit
2358 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2359 struct ring_buffer_event *event)
2361 unsigned long addr = (unsigned long)event;
2362 struct buffer_page *bpage = cpu_buffer->commit_page;
2363 struct buffer_page *start;
2367 /* Do the likely case first */
2368 if (likely(bpage->page == (void *)addr)) {
2369 local_dec(&bpage->entries);
2374 * Because the commit page may be on the reader page we
2375 * start with the next page and check the end loop there.
2377 rb_inc_page(cpu_buffer, &bpage);
2380 if (bpage->page == (void *)addr) {
2381 local_dec(&bpage->entries);
2384 rb_inc_page(cpu_buffer, &bpage);
2385 } while (bpage != start);
2387 /* commit not part of this buffer?? */
2388 RB_WARN_ON(cpu_buffer, 1);
2392 * ring_buffer_commit_discard - discard an event that has not been committed
2393 * @buffer: the ring buffer
2394 * @event: non committed event to discard
2396 * Sometimes an event that is in the ring buffer needs to be ignored.
2397 * This function lets the user discard an event in the ring buffer
2398 * and then that event will not be read later.
2400 * This function only works if it is called before the the item has been
2401 * committed. It will try to free the event from the ring buffer
2402 * if another event has not been added behind it.
2404 * If another event has been added behind it, it will set the event
2405 * up as discarded, and perform the commit.
2407 * If this function is called, do not call ring_buffer_unlock_commit on
2410 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2411 struct ring_buffer_event *event)
2413 struct ring_buffer_per_cpu *cpu_buffer;
2416 /* The event is discarded regardless */
2417 rb_event_discard(event);
2419 cpu = smp_processor_id();
2420 cpu_buffer = buffer->buffers[cpu];
2423 * This must only be called if the event has not been
2424 * committed yet. Thus we can assume that preemption
2425 * is still disabled.
2427 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2429 rb_decrement_entry(cpu_buffer, event);
2430 if (rb_try_to_discard(cpu_buffer, event))
2434 * The commit is still visible by the reader, so we
2435 * must still update the timestamp.
2437 rb_update_write_stamp(cpu_buffer, event);
2439 rb_end_commit(cpu_buffer);
2441 trace_recursive_unlock();
2444 * Only the last preempt count needs to restore preemption.
2446 if (preempt_count() == 1)
2447 ftrace_preempt_enable(per_cpu(rb_need_resched, cpu));
2449 preempt_enable_no_resched_notrace();
2452 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2455 * ring_buffer_write - write data to the buffer without reserving
2456 * @buffer: The ring buffer to write to.
2457 * @length: The length of the data being written (excluding the event header)
2458 * @data: The data to write to the buffer.
2460 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2461 * one function. If you already have the data to write to the buffer, it
2462 * may be easier to simply call this function.
2464 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2465 * and not the length of the event which would hold the header.
2467 int ring_buffer_write(struct ring_buffer *buffer,
2468 unsigned long length,
2471 struct ring_buffer_per_cpu *cpu_buffer;
2472 struct ring_buffer_event *event;
2477 if (ring_buffer_flags != RB_BUFFERS_ON)
2480 if (atomic_read(&buffer->record_disabled))
2483 resched = ftrace_preempt_disable();
2485 cpu = raw_smp_processor_id();
2487 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2490 cpu_buffer = buffer->buffers[cpu];
2492 if (atomic_read(&cpu_buffer->record_disabled))
2495 if (length > BUF_MAX_DATA_SIZE)
2498 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2502 body = rb_event_data(event);
2504 memcpy(body, data, length);
2506 rb_commit(cpu_buffer, event);
2510 ftrace_preempt_enable(resched);
2514 EXPORT_SYMBOL_GPL(ring_buffer_write);
2516 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
2518 struct buffer_page *reader = cpu_buffer->reader_page;
2519 struct buffer_page *head = rb_set_head_page(cpu_buffer);
2520 struct buffer_page *commit = cpu_buffer->commit_page;
2522 /* In case of error, head will be NULL */
2523 if (unlikely(!head))
2526 return reader->read == rb_page_commit(reader) &&
2527 (commit == reader ||
2529 head->read == rb_page_commit(commit)));
2533 * ring_buffer_record_disable - stop all writes into the buffer
2534 * @buffer: The ring buffer to stop writes to.
2536 * This prevents all writes to the buffer. Any attempt to write
2537 * to the buffer after this will fail and return NULL.
2539 * The caller should call synchronize_sched() after this.
2541 void ring_buffer_record_disable(struct ring_buffer *buffer)
2543 atomic_inc(&buffer->record_disabled);
2545 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
2548 * ring_buffer_record_enable - enable writes to the buffer
2549 * @buffer: The ring buffer to enable writes
2551 * Note, multiple disables will need the same number of enables
2552 * to truely enable the writing (much like preempt_disable).
2554 void ring_buffer_record_enable(struct ring_buffer *buffer)
2556 atomic_dec(&buffer->record_disabled);
2558 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
2561 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
2562 * @buffer: The ring buffer to stop writes to.
2563 * @cpu: The CPU buffer to stop
2565 * This prevents all writes to the buffer. Any attempt to write
2566 * to the buffer after this will fail and return NULL.
2568 * The caller should call synchronize_sched() after this.
2570 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
2572 struct ring_buffer_per_cpu *cpu_buffer;
2574 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2577 cpu_buffer = buffer->buffers[cpu];
2578 atomic_inc(&cpu_buffer->record_disabled);
2580 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
2583 * ring_buffer_record_enable_cpu - enable writes to the buffer
2584 * @buffer: The ring buffer to enable writes
2585 * @cpu: The CPU to enable.
2587 * Note, multiple disables will need the same number of enables
2588 * to truely enable the writing (much like preempt_disable).
2590 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
2592 struct ring_buffer_per_cpu *cpu_buffer;
2594 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2597 cpu_buffer = buffer->buffers[cpu];
2598 atomic_dec(&cpu_buffer->record_disabled);
2600 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
2603 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
2604 * @buffer: The ring buffer
2605 * @cpu: The per CPU buffer to get the entries from.
2607 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
2609 struct ring_buffer_per_cpu *cpu_buffer;
2612 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2615 cpu_buffer = buffer->buffers[cpu];
2616 ret = (local_read(&cpu_buffer->entries) - local_read(&cpu_buffer->overrun))
2621 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
2624 * ring_buffer_overrun_cpu - get the number of overruns in a cpu_buffer
2625 * @buffer: The ring buffer
2626 * @cpu: The per CPU buffer to get the number of overruns from
2628 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
2630 struct ring_buffer_per_cpu *cpu_buffer;
2633 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2636 cpu_buffer = buffer->buffers[cpu];
2637 ret = local_read(&cpu_buffer->overrun);
2641 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
2644 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by commits
2645 * @buffer: The ring buffer
2646 * @cpu: The per CPU buffer to get the number of overruns from
2649 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
2651 struct ring_buffer_per_cpu *cpu_buffer;
2654 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2657 cpu_buffer = buffer->buffers[cpu];
2658 ret = local_read(&cpu_buffer->commit_overrun);
2662 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
2665 * ring_buffer_entries - get the number of entries in a buffer
2666 * @buffer: The ring buffer
2668 * Returns the total number of entries in the ring buffer
2671 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
2673 struct ring_buffer_per_cpu *cpu_buffer;
2674 unsigned long entries = 0;
2677 /* if you care about this being correct, lock the buffer */
2678 for_each_buffer_cpu(buffer, cpu) {
2679 cpu_buffer = buffer->buffers[cpu];
2680 entries += (local_read(&cpu_buffer->entries) -
2681 local_read(&cpu_buffer->overrun)) - cpu_buffer->read;
2686 EXPORT_SYMBOL_GPL(ring_buffer_entries);
2689 * ring_buffer_overruns - get the number of overruns in buffer
2690 * @buffer: The ring buffer
2692 * Returns the total number of overruns in the ring buffer
2695 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
2697 struct ring_buffer_per_cpu *cpu_buffer;
2698 unsigned long overruns = 0;
2701 /* if you care about this being correct, lock the buffer */
2702 for_each_buffer_cpu(buffer, cpu) {
2703 cpu_buffer = buffer->buffers[cpu];
2704 overruns += local_read(&cpu_buffer->overrun);
2709 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
2711 static void rb_iter_reset(struct ring_buffer_iter *iter)
2713 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2715 /* Iterator usage is expected to have record disabled */
2716 if (list_empty(&cpu_buffer->reader_page->list)) {
2717 iter->head_page = rb_set_head_page(cpu_buffer);
2718 if (unlikely(!iter->head_page))
2720 iter->head = iter->head_page->read;
2722 iter->head_page = cpu_buffer->reader_page;
2723 iter->head = cpu_buffer->reader_page->read;
2726 iter->read_stamp = cpu_buffer->read_stamp;
2728 iter->read_stamp = iter->head_page->page->time_stamp;
2732 * ring_buffer_iter_reset - reset an iterator
2733 * @iter: The iterator to reset
2735 * Resets the iterator, so that it will start from the beginning
2738 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
2740 struct ring_buffer_per_cpu *cpu_buffer;
2741 unsigned long flags;
2746 cpu_buffer = iter->cpu_buffer;
2748 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2749 rb_iter_reset(iter);
2750 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2752 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
2755 * ring_buffer_iter_empty - check if an iterator has no more to read
2756 * @iter: The iterator to check
2758 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
2760 struct ring_buffer_per_cpu *cpu_buffer;
2762 cpu_buffer = iter->cpu_buffer;
2764 return iter->head_page == cpu_buffer->commit_page &&
2765 iter->head == rb_commit_index(cpu_buffer);
2767 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
2770 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2771 struct ring_buffer_event *event)
2775 switch (event->type_len) {
2776 case RINGBUF_TYPE_PADDING:
2779 case RINGBUF_TYPE_TIME_EXTEND:
2780 delta = event->array[0];
2782 delta += event->time_delta;
2783 cpu_buffer->read_stamp += delta;
2786 case RINGBUF_TYPE_TIME_STAMP:
2787 /* FIXME: not implemented */
2790 case RINGBUF_TYPE_DATA:
2791 cpu_buffer->read_stamp += event->time_delta;
2801 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
2802 struct ring_buffer_event *event)
2806 switch (event->type_len) {
2807 case RINGBUF_TYPE_PADDING:
2810 case RINGBUF_TYPE_TIME_EXTEND:
2811 delta = event->array[0];
2813 delta += event->time_delta;
2814 iter->read_stamp += delta;
2817 case RINGBUF_TYPE_TIME_STAMP:
2818 /* FIXME: not implemented */
2821 case RINGBUF_TYPE_DATA:
2822 iter->read_stamp += event->time_delta;
2831 static struct buffer_page *
2832 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
2834 struct buffer_page *reader = NULL;
2835 unsigned long flags;
2839 local_irq_save(flags);
2840 __raw_spin_lock(&cpu_buffer->lock);
2844 * This should normally only loop twice. But because the
2845 * start of the reader inserts an empty page, it causes
2846 * a case where we will loop three times. There should be no
2847 * reason to loop four times (that I know of).
2849 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
2854 reader = cpu_buffer->reader_page;
2856 /* If there's more to read, return this page */
2857 if (cpu_buffer->reader_page->read < rb_page_size(reader))
2860 /* Never should we have an index greater than the size */
2861 if (RB_WARN_ON(cpu_buffer,
2862 cpu_buffer->reader_page->read > rb_page_size(reader)))
2865 /* check if we caught up to the tail */
2867 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
2871 * Reset the reader page to size zero.
2873 local_set(&cpu_buffer->reader_page->write, 0);
2874 local_set(&cpu_buffer->reader_page->entries, 0);
2875 local_set(&cpu_buffer->reader_page->page->commit, 0);
2879 * Splice the empty reader page into the list around the head.
2881 reader = rb_set_head_page(cpu_buffer);
2882 cpu_buffer->reader_page->list.next = reader->list.next;
2883 cpu_buffer->reader_page->list.prev = reader->list.prev;
2886 * cpu_buffer->pages just needs to point to the buffer, it
2887 * has no specific buffer page to point to. Lets move it out
2888 * of our way so we don't accidently swap it.
2890 cpu_buffer->pages = reader->list.prev;
2892 /* The reader page will be pointing to the new head */
2893 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
2896 * Here's the tricky part.
2898 * We need to move the pointer past the header page.
2899 * But we can only do that if a writer is not currently
2900 * moving it. The page before the header page has the
2901 * flag bit '1' set if it is pointing to the page we want.
2902 * but if the writer is in the process of moving it
2903 * than it will be '2' or already moved '0'.
2906 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
2909 * If we did not convert it, then we must try again.
2915 * Yeah! We succeeded in replacing the page.
2917 * Now make the new head point back to the reader page.
2919 reader->list.next->prev = &cpu_buffer->reader_page->list;
2920 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
2922 /* Finally update the reader page to the new head */
2923 cpu_buffer->reader_page = reader;
2924 rb_reset_reader_page(cpu_buffer);
2929 __raw_spin_unlock(&cpu_buffer->lock);
2930 local_irq_restore(flags);
2935 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
2937 struct ring_buffer_event *event;
2938 struct buffer_page *reader;
2941 reader = rb_get_reader_page(cpu_buffer);
2943 /* This function should not be called when buffer is empty */
2944 if (RB_WARN_ON(cpu_buffer, !reader))
2947 event = rb_reader_event(cpu_buffer);
2949 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
2952 rb_update_read_stamp(cpu_buffer, event);
2954 length = rb_event_length(event);
2955 cpu_buffer->reader_page->read += length;
2958 static void rb_advance_iter(struct ring_buffer_iter *iter)
2960 struct ring_buffer *buffer;
2961 struct ring_buffer_per_cpu *cpu_buffer;
2962 struct ring_buffer_event *event;
2965 cpu_buffer = iter->cpu_buffer;
2966 buffer = cpu_buffer->buffer;
2969 * Check if we are at the end of the buffer.
2971 if (iter->head >= rb_page_size(iter->head_page)) {
2972 /* discarded commits can make the page empty */
2973 if (iter->head_page == cpu_buffer->commit_page)
2979 event = rb_iter_head_event(iter);
2981 length = rb_event_length(event);
2984 * This should not be called to advance the header if we are
2985 * at the tail of the buffer.
2987 if (RB_WARN_ON(cpu_buffer,
2988 (iter->head_page == cpu_buffer->commit_page) &&
2989 (iter->head + length > rb_commit_index(cpu_buffer))))
2992 rb_update_iter_read_stamp(iter, event);
2994 iter->head += length;
2996 /* check for end of page padding */
2997 if ((iter->head >= rb_page_size(iter->head_page)) &&
2998 (iter->head_page != cpu_buffer->commit_page))
2999 rb_advance_iter(iter);
3002 static struct ring_buffer_event *
3003 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts)
3005 struct ring_buffer_event *event;
3006 struct buffer_page *reader;
3011 * We repeat when a timestamp is encountered. It is possible
3012 * to get multiple timestamps from an interrupt entering just
3013 * as one timestamp is about to be written, or from discarded
3014 * commits. The most that we can have is the number on a single page.
3016 if (RB_WARN_ON(cpu_buffer, ++nr_loops > RB_TIMESTAMPS_PER_PAGE))
3019 reader = rb_get_reader_page(cpu_buffer);
3023 event = rb_reader_event(cpu_buffer);
3025 switch (event->type_len) {
3026 case RINGBUF_TYPE_PADDING:
3027 if (rb_null_event(event))
3028 RB_WARN_ON(cpu_buffer, 1);
3030 * Because the writer could be discarding every
3031 * event it creates (which would probably be bad)
3032 * if we were to go back to "again" then we may never
3033 * catch up, and will trigger the warn on, or lock
3034 * the box. Return the padding, and we will release
3035 * the current locks, and try again.
3039 case RINGBUF_TYPE_TIME_EXTEND:
3040 /* Internal data, OK to advance */
3041 rb_advance_reader(cpu_buffer);
3044 case RINGBUF_TYPE_TIME_STAMP:
3045 /* FIXME: not implemented */
3046 rb_advance_reader(cpu_buffer);
3049 case RINGBUF_TYPE_DATA:
3051 *ts = cpu_buffer->read_stamp + event->time_delta;
3052 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3053 cpu_buffer->cpu, ts);
3063 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3065 static struct ring_buffer_event *
3066 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3068 struct ring_buffer *buffer;
3069 struct ring_buffer_per_cpu *cpu_buffer;
3070 struct ring_buffer_event *event;
3073 if (ring_buffer_iter_empty(iter))
3076 cpu_buffer = iter->cpu_buffer;
3077 buffer = cpu_buffer->buffer;
3081 * We repeat when a timestamp is encountered.
3082 * We can get multiple timestamps by nested interrupts or also
3083 * if filtering is on (discarding commits). Since discarding
3084 * commits can be frequent we can get a lot of timestamps.
3085 * But we limit them by not adding timestamps if they begin
3086 * at the start of a page.
3088 if (RB_WARN_ON(cpu_buffer, ++nr_loops > RB_TIMESTAMPS_PER_PAGE))
3091 if (rb_per_cpu_empty(cpu_buffer))
3094 event = rb_iter_head_event(iter);
3096 switch (event->type_len) {
3097 case RINGBUF_TYPE_PADDING:
3098 if (rb_null_event(event)) {
3102 rb_advance_iter(iter);
3105 case RINGBUF_TYPE_TIME_EXTEND:
3106 /* Internal data, OK to advance */
3107 rb_advance_iter(iter);
3110 case RINGBUF_TYPE_TIME_STAMP:
3111 /* FIXME: not implemented */
3112 rb_advance_iter(iter);
3115 case RINGBUF_TYPE_DATA:
3117 *ts = iter->read_stamp + event->time_delta;
3118 ring_buffer_normalize_time_stamp(buffer,
3119 cpu_buffer->cpu, ts);
3129 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3131 static inline int rb_ok_to_lock(void)
3134 * If an NMI die dumps out the content of the ring buffer
3135 * do not grab locks. We also permanently disable the ring
3136 * buffer too. A one time deal is all you get from reading
3137 * the ring buffer from an NMI.
3139 if (likely(!in_nmi()))
3142 tracing_off_permanent();
3147 * ring_buffer_peek - peek at the next event to be read
3148 * @buffer: The ring buffer to read
3149 * @cpu: The cpu to peak at
3150 * @ts: The timestamp counter of this event.
3152 * This will return the event that will be read next, but does
3153 * not consume the data.
3155 struct ring_buffer_event *
3156 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts)
3158 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3159 struct ring_buffer_event *event;
3160 unsigned long flags;
3163 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3166 dolock = rb_ok_to_lock();
3168 local_irq_save(flags);
3170 spin_lock(&cpu_buffer->reader_lock);
3171 event = rb_buffer_peek(cpu_buffer, ts);
3172 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3173 rb_advance_reader(cpu_buffer);
3175 spin_unlock(&cpu_buffer->reader_lock);
3176 local_irq_restore(flags);
3178 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3185 * ring_buffer_iter_peek - peek at the next event to be read
3186 * @iter: The ring buffer iterator
3187 * @ts: The timestamp counter of this event.
3189 * This will return the event that will be read next, but does
3190 * not increment the iterator.
3192 struct ring_buffer_event *
3193 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3195 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3196 struct ring_buffer_event *event;
3197 unsigned long flags;
3200 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3201 event = rb_iter_peek(iter, ts);
3202 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3204 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3211 * ring_buffer_consume - return an event and consume it
3212 * @buffer: The ring buffer to get the next event from
3214 * Returns the next event in the ring buffer, and that event is consumed.
3215 * Meaning, that sequential reads will keep returning a different event,
3216 * and eventually empty the ring buffer if the producer is slower.
3218 struct ring_buffer_event *
3219 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts)
3221 struct ring_buffer_per_cpu *cpu_buffer;
3222 struct ring_buffer_event *event = NULL;
3223 unsigned long flags;
3226 dolock = rb_ok_to_lock();
3229 /* might be called in atomic */
3232 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3235 cpu_buffer = buffer->buffers[cpu];
3236 local_irq_save(flags);
3238 spin_lock(&cpu_buffer->reader_lock);
3240 event = rb_buffer_peek(cpu_buffer, ts);
3242 rb_advance_reader(cpu_buffer);
3245 spin_unlock(&cpu_buffer->reader_lock);
3246 local_irq_restore(flags);
3251 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3256 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3259 * ring_buffer_read_start - start a non consuming read of the buffer
3260 * @buffer: The ring buffer to read from
3261 * @cpu: The cpu buffer to iterate over
3263 * This starts up an iteration through the buffer. It also disables
3264 * the recording to the buffer until the reading is finished.
3265 * This prevents the reading from being corrupted. This is not
3266 * a consuming read, so a producer is not expected.
3268 * Must be paired with ring_buffer_finish.
3270 struct ring_buffer_iter *
3271 ring_buffer_read_start(struct ring_buffer *buffer, int cpu)
3273 struct ring_buffer_per_cpu *cpu_buffer;
3274 struct ring_buffer_iter *iter;
3275 unsigned long flags;
3277 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3280 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3284 cpu_buffer = buffer->buffers[cpu];
3286 iter->cpu_buffer = cpu_buffer;
3288 atomic_inc(&cpu_buffer->record_disabled);
3289 synchronize_sched();
3291 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3292 __raw_spin_lock(&cpu_buffer->lock);
3293 rb_iter_reset(iter);
3294 __raw_spin_unlock(&cpu_buffer->lock);
3295 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3299 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
3302 * ring_buffer_finish - finish reading the iterator of the buffer
3303 * @iter: The iterator retrieved by ring_buffer_start
3305 * This re-enables the recording to the buffer, and frees the
3309 ring_buffer_read_finish(struct ring_buffer_iter *iter)
3311 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3313 atomic_dec(&cpu_buffer->record_disabled);
3316 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
3319 * ring_buffer_read - read the next item in the ring buffer by the iterator
3320 * @iter: The ring buffer iterator
3321 * @ts: The time stamp of the event read.
3323 * This reads the next event in the ring buffer and increments the iterator.
3325 struct ring_buffer_event *
3326 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
3328 struct ring_buffer_event *event;
3329 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3330 unsigned long flags;
3332 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3334 event = rb_iter_peek(iter, ts);
3338 if (event->type_len == RINGBUF_TYPE_PADDING)
3341 rb_advance_iter(iter);
3343 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3347 EXPORT_SYMBOL_GPL(ring_buffer_read);
3350 * ring_buffer_size - return the size of the ring buffer (in bytes)
3351 * @buffer: The ring buffer.
3353 unsigned long ring_buffer_size(struct ring_buffer *buffer)
3355 return BUF_PAGE_SIZE * buffer->pages;
3357 EXPORT_SYMBOL_GPL(ring_buffer_size);
3360 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
3362 rb_head_page_deactivate(cpu_buffer);
3364 cpu_buffer->head_page
3365 = list_entry(cpu_buffer->pages, struct buffer_page, list);
3366 local_set(&cpu_buffer->head_page->write, 0);
3367 local_set(&cpu_buffer->head_page->entries, 0);
3368 local_set(&cpu_buffer->head_page->page->commit, 0);
3370 cpu_buffer->head_page->read = 0;
3372 cpu_buffer->tail_page = cpu_buffer->head_page;
3373 cpu_buffer->commit_page = cpu_buffer->head_page;
3375 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
3376 local_set(&cpu_buffer->reader_page->write, 0);
3377 local_set(&cpu_buffer->reader_page->entries, 0);
3378 local_set(&cpu_buffer->reader_page->page->commit, 0);
3379 cpu_buffer->reader_page->read = 0;
3381 local_set(&cpu_buffer->commit_overrun, 0);
3382 local_set(&cpu_buffer->overrun, 0);
3383 local_set(&cpu_buffer->entries, 0);
3384 local_set(&cpu_buffer->committing, 0);
3385 local_set(&cpu_buffer->commits, 0);
3386 cpu_buffer->read = 0;
3388 cpu_buffer->write_stamp = 0;
3389 cpu_buffer->read_stamp = 0;
3391 rb_head_page_activate(cpu_buffer);
3395 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
3396 * @buffer: The ring buffer to reset a per cpu buffer of
3397 * @cpu: The CPU buffer to be reset
3399 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
3401 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3402 unsigned long flags;
3404 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3407 atomic_inc(&cpu_buffer->record_disabled);
3409 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3411 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
3414 __raw_spin_lock(&cpu_buffer->lock);
3416 rb_reset_cpu(cpu_buffer);
3418 __raw_spin_unlock(&cpu_buffer->lock);
3421 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3423 atomic_dec(&cpu_buffer->record_disabled);
3425 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
3428 * ring_buffer_reset - reset a ring buffer
3429 * @buffer: The ring buffer to reset all cpu buffers
3431 void ring_buffer_reset(struct ring_buffer *buffer)
3435 for_each_buffer_cpu(buffer, cpu)
3436 ring_buffer_reset_cpu(buffer, cpu);
3438 EXPORT_SYMBOL_GPL(ring_buffer_reset);
3441 * rind_buffer_empty - is the ring buffer empty?
3442 * @buffer: The ring buffer to test
3444 int ring_buffer_empty(struct ring_buffer *buffer)
3446 struct ring_buffer_per_cpu *cpu_buffer;
3447 unsigned long flags;
3452 dolock = rb_ok_to_lock();
3454 /* yes this is racy, but if you don't like the race, lock the buffer */
3455 for_each_buffer_cpu(buffer, cpu) {
3456 cpu_buffer = buffer->buffers[cpu];
3457 local_irq_save(flags);
3459 spin_lock(&cpu_buffer->reader_lock);
3460 ret = rb_per_cpu_empty(cpu_buffer);
3462 spin_unlock(&cpu_buffer->reader_lock);
3463 local_irq_restore(flags);
3471 EXPORT_SYMBOL_GPL(ring_buffer_empty);
3474 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
3475 * @buffer: The ring buffer
3476 * @cpu: The CPU buffer to test
3478 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
3480 struct ring_buffer_per_cpu *cpu_buffer;
3481 unsigned long flags;
3485 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3488 dolock = rb_ok_to_lock();
3490 cpu_buffer = buffer->buffers[cpu];
3491 local_irq_save(flags);
3493 spin_lock(&cpu_buffer->reader_lock);
3494 ret = rb_per_cpu_empty(cpu_buffer);
3496 spin_unlock(&cpu_buffer->reader_lock);
3497 local_irq_restore(flags);
3501 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
3503 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3505 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
3506 * @buffer_a: One buffer to swap with
3507 * @buffer_b: The other buffer to swap with
3509 * This function is useful for tracers that want to take a "snapshot"
3510 * of a CPU buffer and has another back up buffer lying around.
3511 * it is expected that the tracer handles the cpu buffer not being
3512 * used at the moment.
3514 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
3515 struct ring_buffer *buffer_b, int cpu)
3517 struct ring_buffer_per_cpu *cpu_buffer_a;
3518 struct ring_buffer_per_cpu *cpu_buffer_b;
3521 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
3522 !cpumask_test_cpu(cpu, buffer_b->cpumask))
3525 /* At least make sure the two buffers are somewhat the same */
3526 if (buffer_a->pages != buffer_b->pages)
3531 if (ring_buffer_flags != RB_BUFFERS_ON)
3534 if (atomic_read(&buffer_a->record_disabled))
3537 if (atomic_read(&buffer_b->record_disabled))
3540 cpu_buffer_a = buffer_a->buffers[cpu];
3541 cpu_buffer_b = buffer_b->buffers[cpu];
3543 if (atomic_read(&cpu_buffer_a->record_disabled))
3546 if (atomic_read(&cpu_buffer_b->record_disabled))
3550 * We can't do a synchronize_sched here because this
3551 * function can be called in atomic context.
3552 * Normally this will be called from the same CPU as cpu.
3553 * If not it's up to the caller to protect this.
3555 atomic_inc(&cpu_buffer_a->record_disabled);
3556 atomic_inc(&cpu_buffer_b->record_disabled);
3559 if (local_read(&cpu_buffer_a->committing))
3561 if (local_read(&cpu_buffer_b->committing))
3564 buffer_a->buffers[cpu] = cpu_buffer_b;
3565 buffer_b->buffers[cpu] = cpu_buffer_a;
3567 cpu_buffer_b->buffer = buffer_a;
3568 cpu_buffer_a->buffer = buffer_b;
3573 atomic_dec(&cpu_buffer_a->record_disabled);
3574 atomic_dec(&cpu_buffer_b->record_disabled);
3578 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
3579 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
3582 * ring_buffer_alloc_read_page - allocate a page to read from buffer
3583 * @buffer: the buffer to allocate for.
3585 * This function is used in conjunction with ring_buffer_read_page.
3586 * When reading a full page from the ring buffer, these functions
3587 * can be used to speed up the process. The calling function should
3588 * allocate a few pages first with this function. Then when it
3589 * needs to get pages from the ring buffer, it passes the result
3590 * of this function into ring_buffer_read_page, which will swap
3591 * the page that was allocated, with the read page of the buffer.
3594 * The page allocated, or NULL on error.
3596 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer)
3598 struct buffer_data_page *bpage;
3601 addr = __get_free_page(GFP_KERNEL);
3605 bpage = (void *)addr;
3607 rb_init_page(bpage);
3611 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
3614 * ring_buffer_free_read_page - free an allocated read page
3615 * @buffer: the buffer the page was allocate for
3616 * @data: the page to free
3618 * Free a page allocated from ring_buffer_alloc_read_page.
3620 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
3622 free_page((unsigned long)data);
3624 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
3627 * ring_buffer_read_page - extract a page from the ring buffer
3628 * @buffer: buffer to extract from
3629 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
3630 * @len: amount to extract
3631 * @cpu: the cpu of the buffer to extract
3632 * @full: should the extraction only happen when the page is full.
3634 * This function will pull out a page from the ring buffer and consume it.
3635 * @data_page must be the address of the variable that was returned
3636 * from ring_buffer_alloc_read_page. This is because the page might be used
3637 * to swap with a page in the ring buffer.
3640 * rpage = ring_buffer_alloc_read_page(buffer);
3643 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
3645 * process_page(rpage, ret);
3647 * When @full is set, the function will not return true unless
3648 * the writer is off the reader page.
3650 * Note: it is up to the calling functions to handle sleeps and wakeups.
3651 * The ring buffer can be used anywhere in the kernel and can not
3652 * blindly call wake_up. The layer that uses the ring buffer must be
3653 * responsible for that.
3656 * >=0 if data has been transferred, returns the offset of consumed data.
3657 * <0 if no data has been transferred.
3659 int ring_buffer_read_page(struct ring_buffer *buffer,
3660 void **data_page, size_t len, int cpu, int full)
3662 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3663 struct ring_buffer_event *event;
3664 struct buffer_data_page *bpage;
3665 struct buffer_page *reader;
3666 unsigned long flags;
3667 unsigned int commit;
3672 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3676 * If len is not big enough to hold the page header, then
3677 * we can not copy anything.
3679 if (len <= BUF_PAGE_HDR_SIZE)
3682 len -= BUF_PAGE_HDR_SIZE;
3691 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3693 reader = rb_get_reader_page(cpu_buffer);
3697 event = rb_reader_event(cpu_buffer);
3699 read = reader->read;
3700 commit = rb_page_commit(reader);
3703 * If this page has been partially read or
3704 * if len is not big enough to read the rest of the page or
3705 * a writer is still on the page, then
3706 * we must copy the data from the page to the buffer.
3707 * Otherwise, we can simply swap the page with the one passed in.
3709 if (read || (len < (commit - read)) ||
3710 cpu_buffer->reader_page == cpu_buffer->commit_page) {
3711 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
3712 unsigned int rpos = read;
3713 unsigned int pos = 0;
3719 if (len > (commit - read))
3720 len = (commit - read);
3722 size = rb_event_length(event);
3727 /* save the current timestamp, since the user will need it */
3728 save_timestamp = cpu_buffer->read_stamp;
3730 /* Need to copy one event at a time */
3732 memcpy(bpage->data + pos, rpage->data + rpos, size);
3736 rb_advance_reader(cpu_buffer);
3737 rpos = reader->read;
3740 event = rb_reader_event(cpu_buffer);
3741 size = rb_event_length(event);
3742 } while (len > size);
3745 local_set(&bpage->commit, pos);
3746 bpage->time_stamp = save_timestamp;
3748 /* we copied everything to the beginning */
3751 /* update the entry counter */
3752 cpu_buffer->read += rb_page_entries(reader);
3754 /* swap the pages */
3755 rb_init_page(bpage);
3756 bpage = reader->page;
3757 reader->page = *data_page;
3758 local_set(&reader->write, 0);
3759 local_set(&reader->entries, 0);
3766 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3771 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
3773 #ifdef CONFIG_TRACING
3775 rb_simple_read(struct file *filp, char __user *ubuf,
3776 size_t cnt, loff_t *ppos)
3778 unsigned long *p = filp->private_data;
3782 if (test_bit(RB_BUFFERS_DISABLED_BIT, p))
3783 r = sprintf(buf, "permanently disabled\n");
3785 r = sprintf(buf, "%d\n", test_bit(RB_BUFFERS_ON_BIT, p));
3787 return simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
3791 rb_simple_write(struct file *filp, const char __user *ubuf,
3792 size_t cnt, loff_t *ppos)
3794 unsigned long *p = filp->private_data;
3799 if (cnt >= sizeof(buf))
3802 if (copy_from_user(&buf, ubuf, cnt))
3807 ret = strict_strtoul(buf, 10, &val);
3812 set_bit(RB_BUFFERS_ON_BIT, p);
3814 clear_bit(RB_BUFFERS_ON_BIT, p);
3821 static const struct file_operations rb_simple_fops = {
3822 .open = tracing_open_generic,
3823 .read = rb_simple_read,
3824 .write = rb_simple_write,
3828 static __init int rb_init_debugfs(void)
3830 struct dentry *d_tracer;
3832 d_tracer = tracing_init_dentry();
3834 trace_create_file("tracing_on", 0644, d_tracer,
3835 &ring_buffer_flags, &rb_simple_fops);
3840 fs_initcall(rb_init_debugfs);
3843 #ifdef CONFIG_HOTPLUG_CPU
3844 static int rb_cpu_notify(struct notifier_block *self,
3845 unsigned long action, void *hcpu)
3847 struct ring_buffer *buffer =
3848 container_of(self, struct ring_buffer, cpu_notify);
3849 long cpu = (long)hcpu;
3852 case CPU_UP_PREPARE:
3853 case CPU_UP_PREPARE_FROZEN:
3854 if (cpumask_test_cpu(cpu, buffer->cpumask))
3857 buffer->buffers[cpu] =
3858 rb_allocate_cpu_buffer(buffer, cpu);
3859 if (!buffer->buffers[cpu]) {
3860 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
3865 cpumask_set_cpu(cpu, buffer->cpumask);
3867 case CPU_DOWN_PREPARE:
3868 case CPU_DOWN_PREPARE_FROZEN:
3871 * If we were to free the buffer, then the user would
3872 * lose any trace that was in the buffer.