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);
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 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
212 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
215 RB_LEN_TIME_EXTEND = 8,
216 RB_LEN_TIME_STAMP = 16,
219 static inline int rb_null_event(struct ring_buffer_event *event)
221 return event->type_len == RINGBUF_TYPE_PADDING
222 && event->time_delta == 0;
225 static inline int rb_discarded_event(struct ring_buffer_event *event)
227 return event->type_len == RINGBUF_TYPE_PADDING && event->time_delta;
230 static void rb_event_set_padding(struct ring_buffer_event *event)
232 event->type_len = RINGBUF_TYPE_PADDING;
233 event->time_delta = 0;
237 rb_event_data_length(struct ring_buffer_event *event)
242 length = event->type_len * RB_ALIGNMENT;
244 length = event->array[0];
245 return length + RB_EVNT_HDR_SIZE;
248 /* inline for ring buffer fast paths */
250 rb_event_length(struct ring_buffer_event *event)
252 switch (event->type_len) {
253 case RINGBUF_TYPE_PADDING:
254 if (rb_null_event(event))
257 return event->array[0] + RB_EVNT_HDR_SIZE;
259 case RINGBUF_TYPE_TIME_EXTEND:
260 return RB_LEN_TIME_EXTEND;
262 case RINGBUF_TYPE_TIME_STAMP:
263 return RB_LEN_TIME_STAMP;
265 case RINGBUF_TYPE_DATA:
266 return rb_event_data_length(event);
275 * ring_buffer_event_length - return the length of the event
276 * @event: the event to get the length of
278 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
280 unsigned length = rb_event_length(event);
281 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
283 length -= RB_EVNT_HDR_SIZE;
284 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
285 length -= sizeof(event->array[0]);
288 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
290 /* inline for ring buffer fast paths */
292 rb_event_data(struct ring_buffer_event *event)
294 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
295 /* If length is in len field, then array[0] has the data */
297 return (void *)&event->array[0];
298 /* Otherwise length is in array[0] and array[1] has the data */
299 return (void *)&event->array[1];
303 * ring_buffer_event_data - return the data of the event
304 * @event: the event to get the data from
306 void *ring_buffer_event_data(struct ring_buffer_event *event)
308 return rb_event_data(event);
310 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
312 #define for_each_buffer_cpu(buffer, cpu) \
313 for_each_cpu(cpu, buffer->cpumask)
316 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
317 #define TS_DELTA_TEST (~TS_MASK)
319 struct buffer_data_page {
320 u64 time_stamp; /* page time stamp */
321 local_t commit; /* write committed index */
322 unsigned char data[]; /* data of buffer page */
326 * Note, the buffer_page list must be first. The buffer pages
327 * are allocated in cache lines, which means that each buffer
328 * page will be at the beginning of a cache line, and thus
329 * the least significant bits will be zero. We use this to
330 * add flags in the list struct pointers, to make the ring buffer
334 struct list_head list; /* list of buffer pages */
335 local_t write; /* index for next write */
336 unsigned read; /* index for next read */
337 local_t entries; /* entries on this page */
338 struct buffer_data_page *page; /* Actual data page */
342 * The buffer page counters, write and entries, must be reset
343 * atomically when crossing page boundaries. To synchronize this
344 * update, two counters are inserted into the number. One is
345 * the actual counter for the write position or count on the page.
347 * The other is a counter of updaters. Before an update happens
348 * the update partition of the counter is incremented. This will
349 * allow the updater to update the counter atomically.
351 * The counter is 20 bits, and the state data is 12.
353 #define RB_WRITE_MASK 0xfffff
354 #define RB_WRITE_INTCNT (1 << 20)
356 static void rb_init_page(struct buffer_data_page *bpage)
358 local_set(&bpage->commit, 0);
362 * ring_buffer_page_len - the size of data on the page.
363 * @page: The page to read
365 * Returns the amount of data on the page, including buffer page header.
367 size_t ring_buffer_page_len(void *page)
369 return local_read(&((struct buffer_data_page *)page)->commit)
374 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
377 static void free_buffer_page(struct buffer_page *bpage)
379 free_page((unsigned long)bpage->page);
384 * We need to fit the time_stamp delta into 27 bits.
386 static inline int test_time_stamp(u64 delta)
388 if (delta & TS_DELTA_TEST)
393 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
395 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
396 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
398 /* Max number of timestamps that can fit on a page */
399 #define RB_TIMESTAMPS_PER_PAGE (BUF_PAGE_SIZE / RB_LEN_TIME_STAMP)
401 int ring_buffer_print_page_header(struct trace_seq *s)
403 struct buffer_data_page field;
406 ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t"
407 "offset:0;\tsize:%u;\n",
408 (unsigned int)sizeof(field.time_stamp));
410 ret = trace_seq_printf(s, "\tfield: local_t commit;\t"
411 "offset:%u;\tsize:%u;\n",
412 (unsigned int)offsetof(typeof(field), commit),
413 (unsigned int)sizeof(field.commit));
415 ret = trace_seq_printf(s, "\tfield: char data;\t"
416 "offset:%u;\tsize:%u;\n",
417 (unsigned int)offsetof(typeof(field), data),
418 (unsigned int)BUF_PAGE_SIZE);
424 * head_page == tail_page && head == tail then buffer is empty.
426 struct ring_buffer_per_cpu {
428 struct ring_buffer *buffer;
429 spinlock_t reader_lock; /* serialize readers */
431 struct lock_class_key lock_key;
432 struct list_head *pages;
433 struct buffer_page *head_page; /* read from head */
434 struct buffer_page *tail_page; /* write to tail */
435 struct buffer_page *commit_page; /* committed pages */
436 struct buffer_page *reader_page;
437 local_t commit_overrun;
445 atomic_t record_disabled;
452 atomic_t record_disabled;
453 cpumask_var_t cpumask;
455 struct lock_class_key *reader_lock_key;
459 struct ring_buffer_per_cpu **buffers;
461 #ifdef CONFIG_HOTPLUG_CPU
462 struct notifier_block cpu_notify;
467 struct ring_buffer_iter {
468 struct ring_buffer_per_cpu *cpu_buffer;
470 struct buffer_page *head_page;
474 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
475 #define RB_WARN_ON(buffer, cond) \
477 int _____ret = unlikely(cond); \
479 atomic_inc(&buffer->record_disabled); \
485 /* Up this if you want to test the TIME_EXTENTS and normalization */
486 #define DEBUG_SHIFT 0
488 static inline u64 rb_time_stamp(struct ring_buffer *buffer, int cpu)
490 /* shift to debug/test normalization and TIME_EXTENTS */
491 return buffer->clock() << DEBUG_SHIFT;
494 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
498 preempt_disable_notrace();
499 time = rb_time_stamp(buffer, cpu);
500 preempt_enable_no_resched_notrace();
504 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
506 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
509 /* Just stupid testing the normalize function and deltas */
512 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
515 * Making the ring buffer lockless makes things tricky.
516 * Although writes only happen on the CPU that they are on,
517 * and they only need to worry about interrupts. Reads can
520 * The reader page is always off the ring buffer, but when the
521 * reader finishes with a page, it needs to swap its page with
522 * a new one from the buffer. The reader needs to take from
523 * the head (writes go to the tail). But if a writer is in overwrite
524 * mode and wraps, it must push the head page forward.
526 * Here lies the problem.
528 * The reader must be careful to replace only the head page, and
529 * not another one. As described at the top of the file in the
530 * ASCII art, the reader sets its old page to point to the next
531 * page after head. It then sets the page after head to point to
532 * the old reader page. But if the writer moves the head page
533 * during this operation, the reader could end up with the tail.
535 * We use cmpxchg to help prevent this race. We also do something
536 * special with the page before head. We set the LSB to 1.
538 * When the writer must push the page forward, it will clear the
539 * bit that points to the head page, move the head, and then set
540 * the bit that points to the new head page.
542 * We also don't want an interrupt coming in and moving the head
543 * page on another writer. Thus we use the second LSB to catch
546 * head->list->prev->next bit 1 bit 0
549 * Points to head page 0 1
552 * Note we can not trust the prev pointer of the head page, because:
554 * +----+ +-----+ +-----+
555 * | |------>| T |---X--->| N |
557 * +----+ +-----+ +-----+
560 * +----------| R |----------+ |
564 * Key: ---X--> HEAD flag set in pointer
569 * (see __rb_reserve_next() to see where this happens)
571 * What the above shows is that the reader just swapped out
572 * the reader page with a page in the buffer, but before it
573 * could make the new header point back to the new page added
574 * it was preempted by a writer. The writer moved forward onto
575 * the new page added by the reader and is about to move forward
578 * You can see, it is legitimate for the previous pointer of
579 * the head (or any page) not to point back to itself. But only
583 #define RB_PAGE_NORMAL 0UL
584 #define RB_PAGE_HEAD 1UL
585 #define RB_PAGE_UPDATE 2UL
588 #define RB_FLAG_MASK 3UL
590 /* PAGE_MOVED is not part of the mask */
591 #define RB_PAGE_MOVED 4UL
594 * rb_list_head - remove any bit
596 static struct list_head *rb_list_head(struct list_head *list)
598 unsigned long val = (unsigned long)list;
600 return (struct list_head *)(val & ~RB_FLAG_MASK);
604 * rb_is_head_page - test if the give page is the head page
606 * Because the reader may move the head_page pointer, we can
607 * not trust what the head page is (it may be pointing to
608 * the reader page). But if the next page is a header page,
609 * its flags will be non zero.
612 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
613 struct buffer_page *page, struct list_head *list)
617 val = (unsigned long)list->next;
619 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
620 return RB_PAGE_MOVED;
622 return val & RB_FLAG_MASK;
628 * The unique thing about the reader page, is that, if the
629 * writer is ever on it, the previous pointer never points
630 * back to the reader page.
632 static int rb_is_reader_page(struct buffer_page *page)
634 struct list_head *list = page->list.prev;
636 return rb_list_head(list->next) != &page->list;
640 * rb_set_list_to_head - set a list_head to be pointing to head.
642 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
643 struct list_head *list)
647 ptr = (unsigned long *)&list->next;
648 *ptr |= RB_PAGE_HEAD;
649 *ptr &= ~RB_PAGE_UPDATE;
653 * rb_head_page_activate - sets up head page
655 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
657 struct buffer_page *head;
659 head = cpu_buffer->head_page;
664 * Set the previous list pointer to have the HEAD flag.
666 rb_set_list_to_head(cpu_buffer, head->list.prev);
669 static void rb_list_head_clear(struct list_head *list)
671 unsigned long *ptr = (unsigned long *)&list->next;
673 *ptr &= ~RB_FLAG_MASK;
677 * rb_head_page_dactivate - clears head page ptr (for free list)
680 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
682 struct list_head *hd;
684 /* Go through the whole list and clear any pointers found. */
685 rb_list_head_clear(cpu_buffer->pages);
687 list_for_each(hd, cpu_buffer->pages)
688 rb_list_head_clear(hd);
691 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
692 struct buffer_page *head,
693 struct buffer_page *prev,
694 int old_flag, int new_flag)
696 struct list_head *list;
697 unsigned long val = (unsigned long)&head->list;
702 val &= ~RB_FLAG_MASK;
704 ret = (unsigned long)cmpxchg(&list->next,
705 val | old_flag, val | new_flag);
707 /* check if the reader took the page */
708 if ((ret & ~RB_FLAG_MASK) != val)
709 return RB_PAGE_MOVED;
711 return ret & RB_FLAG_MASK;
714 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
715 struct buffer_page *head,
716 struct buffer_page *prev,
719 return rb_head_page_set(cpu_buffer, head, prev,
720 old_flag, RB_PAGE_UPDATE);
723 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
724 struct buffer_page *head,
725 struct buffer_page *prev,
728 return rb_head_page_set(cpu_buffer, head, prev,
729 old_flag, RB_PAGE_HEAD);
732 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
733 struct buffer_page *head,
734 struct buffer_page *prev,
737 return rb_head_page_set(cpu_buffer, head, prev,
738 old_flag, RB_PAGE_NORMAL);
741 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
742 struct buffer_page **bpage)
744 struct list_head *p = rb_list_head((*bpage)->list.next);
746 *bpage = list_entry(p, struct buffer_page, list);
749 static struct buffer_page *
750 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
752 struct buffer_page *head;
753 struct buffer_page *page;
754 struct list_head *list;
757 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
761 list = cpu_buffer->pages;
762 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
765 page = head = cpu_buffer->head_page;
767 * It is possible that the writer moves the header behind
768 * where we started, and we miss in one loop.
769 * A second loop should grab the header, but we'll do
770 * three loops just because I'm paranoid.
772 for (i = 0; i < 3; i++) {
774 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
775 cpu_buffer->head_page = page;
778 rb_inc_page(cpu_buffer, &page);
779 } while (page != head);
782 RB_WARN_ON(cpu_buffer, 1);
787 static int rb_head_page_replace(struct buffer_page *old,
788 struct buffer_page *new)
790 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
794 val = *ptr & ~RB_FLAG_MASK;
797 ret = cmpxchg(ptr, val, &new->list);
803 * rb_tail_page_update - move the tail page forward
805 * Returns 1 if moved tail page, 0 if someone else did.
807 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
808 struct buffer_page *tail_page,
809 struct buffer_page *next_page)
811 struct buffer_page *old_tail;
812 unsigned long old_entries;
813 unsigned long old_write;
817 * The tail page now needs to be moved forward.
819 * We need to reset the tail page, but without messing
820 * with possible erasing of data brought in by interrupts
821 * that have moved the tail page and are currently on it.
823 * We add a counter to the write field to denote this.
825 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
826 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
829 * Just make sure we have seen our old_write and synchronize
830 * with any interrupts that come in.
835 * If the tail page is still the same as what we think
836 * it is, then it is up to us to update the tail
839 if (tail_page == cpu_buffer->tail_page) {
840 /* Zero the write counter */
841 unsigned long val = old_write & ~RB_WRITE_MASK;
842 unsigned long eval = old_entries & ~RB_WRITE_MASK;
845 * This will only succeed if an interrupt did
846 * not come in and change it. In which case, we
847 * do not want to modify it.
849 * We add (void) to let the compiler know that we do not care
850 * about the return value of these functions. We use the
851 * cmpxchg to only update if an interrupt did not already
852 * do it for us. If the cmpxchg fails, we don't care.
854 (void)local_cmpxchg(&next_page->write, old_write, val);
855 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
858 * No need to worry about races with clearing out the commit.
859 * it only can increment when a commit takes place. But that
860 * only happens in the outer most nested commit.
862 local_set(&next_page->page->commit, 0);
864 old_tail = cmpxchg(&cpu_buffer->tail_page,
865 tail_page, next_page);
867 if (old_tail == tail_page)
874 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
875 struct buffer_page *bpage)
877 unsigned long val = (unsigned long)bpage;
879 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
886 * rb_check_list - make sure a pointer to a list has the last bits zero
888 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
889 struct list_head *list)
891 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
893 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
899 * check_pages - integrity check of buffer pages
900 * @cpu_buffer: CPU buffer with pages to test
902 * As a safety measure we check to make sure the data pages have not
905 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
907 struct list_head *head = cpu_buffer->pages;
908 struct buffer_page *bpage, *tmp;
910 rb_head_page_deactivate(cpu_buffer);
912 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
914 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
917 if (rb_check_list(cpu_buffer, head))
920 list_for_each_entry_safe(bpage, tmp, head, list) {
921 if (RB_WARN_ON(cpu_buffer,
922 bpage->list.next->prev != &bpage->list))
924 if (RB_WARN_ON(cpu_buffer,
925 bpage->list.prev->next != &bpage->list))
927 if (rb_check_list(cpu_buffer, &bpage->list))
931 rb_head_page_activate(cpu_buffer);
936 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
939 struct buffer_page *bpage, *tmp;
946 for (i = 0; i < nr_pages; i++) {
947 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
948 GFP_KERNEL, cpu_to_node(cpu_buffer->cpu));
952 rb_check_bpage(cpu_buffer, bpage);
954 list_add(&bpage->list, &pages);
956 addr = __get_free_page(GFP_KERNEL);
959 bpage->page = (void *)addr;
960 rb_init_page(bpage->page);
964 * The ring buffer page list is a circular list that does not
965 * start and end with a list head. All page list items point to
968 cpu_buffer->pages = pages.next;
971 rb_check_pages(cpu_buffer);
976 list_for_each_entry_safe(bpage, tmp, &pages, list) {
977 list_del_init(&bpage->list);
978 free_buffer_page(bpage);
983 static struct ring_buffer_per_cpu *
984 rb_allocate_cpu_buffer(struct ring_buffer *buffer, int cpu)
986 struct ring_buffer_per_cpu *cpu_buffer;
987 struct buffer_page *bpage;
991 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
992 GFP_KERNEL, cpu_to_node(cpu));
996 cpu_buffer->cpu = cpu;
997 cpu_buffer->buffer = buffer;
998 spin_lock_init(&cpu_buffer->reader_lock);
999 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1000 cpu_buffer->lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
1002 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1003 GFP_KERNEL, cpu_to_node(cpu));
1005 goto fail_free_buffer;
1007 rb_check_bpage(cpu_buffer, bpage);
1009 cpu_buffer->reader_page = bpage;
1010 addr = __get_free_page(GFP_KERNEL);
1012 goto fail_free_reader;
1013 bpage->page = (void *)addr;
1014 rb_init_page(bpage->page);
1016 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1018 ret = rb_allocate_pages(cpu_buffer, buffer->pages);
1020 goto fail_free_reader;
1022 cpu_buffer->head_page
1023 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1024 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1026 rb_head_page_activate(cpu_buffer);
1031 free_buffer_page(cpu_buffer->reader_page);
1038 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1040 struct list_head *head = cpu_buffer->pages;
1041 struct buffer_page *bpage, *tmp;
1043 free_buffer_page(cpu_buffer->reader_page);
1045 rb_head_page_deactivate(cpu_buffer);
1048 list_for_each_entry_safe(bpage, tmp, head, list) {
1049 list_del_init(&bpage->list);
1050 free_buffer_page(bpage);
1052 bpage = list_entry(head, struct buffer_page, list);
1053 free_buffer_page(bpage);
1059 #ifdef CONFIG_HOTPLUG_CPU
1060 static int rb_cpu_notify(struct notifier_block *self,
1061 unsigned long action, void *hcpu);
1065 * ring_buffer_alloc - allocate a new ring_buffer
1066 * @size: the size in bytes per cpu that is needed.
1067 * @flags: attributes to set for the ring buffer.
1069 * Currently the only flag that is available is the RB_FL_OVERWRITE
1070 * flag. This flag means that the buffer will overwrite old data
1071 * when the buffer wraps. If this flag is not set, the buffer will
1072 * drop data when the tail hits the head.
1074 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1075 struct lock_class_key *key)
1077 struct ring_buffer *buffer;
1081 /* keep it in its own cache line */
1082 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1087 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1088 goto fail_free_buffer;
1090 buffer->pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1091 buffer->flags = flags;
1092 buffer->clock = trace_clock_local;
1093 buffer->reader_lock_key = key;
1095 /* need at least two pages */
1096 if (buffer->pages < 2)
1100 * In case of non-hotplug cpu, if the ring-buffer is allocated
1101 * in early initcall, it will not be notified of secondary cpus.
1102 * In that off case, we need to allocate for all possible cpus.
1104 #ifdef CONFIG_HOTPLUG_CPU
1106 cpumask_copy(buffer->cpumask, cpu_online_mask);
1108 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1110 buffer->cpus = nr_cpu_ids;
1112 bsize = sizeof(void *) * nr_cpu_ids;
1113 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1115 if (!buffer->buffers)
1116 goto fail_free_cpumask;
1118 for_each_buffer_cpu(buffer, cpu) {
1119 buffer->buffers[cpu] =
1120 rb_allocate_cpu_buffer(buffer, cpu);
1121 if (!buffer->buffers[cpu])
1122 goto fail_free_buffers;
1125 #ifdef CONFIG_HOTPLUG_CPU
1126 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1127 buffer->cpu_notify.priority = 0;
1128 register_cpu_notifier(&buffer->cpu_notify);
1132 mutex_init(&buffer->mutex);
1137 for_each_buffer_cpu(buffer, cpu) {
1138 if (buffer->buffers[cpu])
1139 rb_free_cpu_buffer(buffer->buffers[cpu]);
1141 kfree(buffer->buffers);
1144 free_cpumask_var(buffer->cpumask);
1151 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1154 * ring_buffer_free - free a ring buffer.
1155 * @buffer: the buffer to free.
1158 ring_buffer_free(struct ring_buffer *buffer)
1164 #ifdef CONFIG_HOTPLUG_CPU
1165 unregister_cpu_notifier(&buffer->cpu_notify);
1168 for_each_buffer_cpu(buffer, cpu)
1169 rb_free_cpu_buffer(buffer->buffers[cpu]);
1173 free_cpumask_var(buffer->cpumask);
1177 EXPORT_SYMBOL_GPL(ring_buffer_free);
1179 void ring_buffer_set_clock(struct ring_buffer *buffer,
1182 buffer->clock = clock;
1185 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1188 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned nr_pages)
1190 struct buffer_page *bpage;
1191 struct list_head *p;
1194 atomic_inc(&cpu_buffer->record_disabled);
1195 synchronize_sched();
1197 rb_head_page_deactivate(cpu_buffer);
1199 for (i = 0; i < nr_pages; i++) {
1200 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1202 p = cpu_buffer->pages->next;
1203 bpage = list_entry(p, struct buffer_page, list);
1204 list_del_init(&bpage->list);
1205 free_buffer_page(bpage);
1207 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1210 rb_reset_cpu(cpu_buffer);
1212 rb_check_pages(cpu_buffer);
1214 atomic_dec(&cpu_buffer->record_disabled);
1219 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer,
1220 struct list_head *pages, unsigned nr_pages)
1222 struct buffer_page *bpage;
1223 struct list_head *p;
1226 atomic_inc(&cpu_buffer->record_disabled);
1227 synchronize_sched();
1229 spin_lock_irq(&cpu_buffer->reader_lock);
1230 rb_head_page_deactivate(cpu_buffer);
1232 for (i = 0; i < nr_pages; i++) {
1233 if (RB_WARN_ON(cpu_buffer, list_empty(pages)))
1236 bpage = list_entry(p, struct buffer_page, list);
1237 list_del_init(&bpage->list);
1238 list_add_tail(&bpage->list, cpu_buffer->pages);
1240 rb_reset_cpu(cpu_buffer);
1241 spin_unlock_irq(&cpu_buffer->reader_lock);
1243 rb_check_pages(cpu_buffer);
1245 atomic_dec(&cpu_buffer->record_disabled);
1249 * ring_buffer_resize - resize the ring buffer
1250 * @buffer: the buffer to resize.
1251 * @size: the new size.
1253 * The tracer is responsible for making sure that the buffer is
1254 * not being used while changing the size.
1255 * Note: We may be able to change the above requirement by using
1256 * RCU synchronizations.
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 mutex_lock(&buffer->mutex);
1292 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1294 if (size < buffer_size) {
1296 /* easy case, just free pages */
1297 if (RB_WARN_ON(buffer, nr_pages >= buffer->pages))
1300 rm_pages = buffer->pages - nr_pages;
1302 for_each_buffer_cpu(buffer, cpu) {
1303 cpu_buffer = buffer->buffers[cpu];
1304 rb_remove_pages(cpu_buffer, rm_pages);
1310 * This is a bit more difficult. We only want to add pages
1311 * when we can allocate enough for all CPUs. We do this
1312 * by allocating all the pages and storing them on a local
1313 * link list. If we succeed in our allocation, then we
1314 * add these pages to the cpu_buffers. Otherwise we just free
1315 * them all and return -ENOMEM;
1317 if (RB_WARN_ON(buffer, nr_pages <= buffer->pages))
1320 new_pages = nr_pages - buffer->pages;
1322 for_each_buffer_cpu(buffer, cpu) {
1323 for (i = 0; i < new_pages; i++) {
1324 bpage = kzalloc_node(ALIGN(sizeof(*bpage),
1326 GFP_KERNEL, cpu_to_node(cpu));
1329 list_add(&bpage->list, &pages);
1330 addr = __get_free_page(GFP_KERNEL);
1333 bpage->page = (void *)addr;
1334 rb_init_page(bpage->page);
1338 for_each_buffer_cpu(buffer, cpu) {
1339 cpu_buffer = buffer->buffers[cpu];
1340 rb_insert_pages(cpu_buffer, &pages, new_pages);
1343 if (RB_WARN_ON(buffer, !list_empty(&pages)))
1347 buffer->pages = nr_pages;
1349 mutex_unlock(&buffer->mutex);
1354 list_for_each_entry_safe(bpage, tmp, &pages, list) {
1355 list_del_init(&bpage->list);
1356 free_buffer_page(bpage);
1359 mutex_unlock(&buffer->mutex);
1363 * Something went totally wrong, and we are too paranoid
1364 * to even clean up the mess.
1368 mutex_unlock(&buffer->mutex);
1371 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1373 static inline void *
1374 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1376 return bpage->data + index;
1379 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1381 return bpage->page->data + index;
1384 static inline struct ring_buffer_event *
1385 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1387 return __rb_page_index(cpu_buffer->reader_page,
1388 cpu_buffer->reader_page->read);
1391 static inline struct ring_buffer_event *
1392 rb_iter_head_event(struct ring_buffer_iter *iter)
1394 return __rb_page_index(iter->head_page, iter->head);
1397 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1399 return local_read(&bpage->write) & RB_WRITE_MASK;
1402 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1404 return local_read(&bpage->page->commit);
1407 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1409 return local_read(&bpage->entries) & RB_WRITE_MASK;
1412 /* Size is determined by what has been commited */
1413 static inline unsigned rb_page_size(struct buffer_page *bpage)
1415 return rb_page_commit(bpage);
1418 static inline unsigned
1419 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1421 return rb_page_commit(cpu_buffer->commit_page);
1424 static inline unsigned
1425 rb_event_index(struct ring_buffer_event *event)
1427 unsigned long addr = (unsigned long)event;
1429 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1433 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1434 struct ring_buffer_event *event)
1436 unsigned long addr = (unsigned long)event;
1437 unsigned long index;
1439 index = rb_event_index(event);
1442 return cpu_buffer->commit_page->page == (void *)addr &&
1443 rb_commit_index(cpu_buffer) == index;
1447 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1449 unsigned long max_count;
1452 * We only race with interrupts and NMIs on this CPU.
1453 * If we own the commit event, then we can commit
1454 * all others that interrupted us, since the interruptions
1455 * are in stack format (they finish before they come
1456 * back to us). This allows us to do a simple loop to
1457 * assign the commit to the tail.
1460 max_count = cpu_buffer->buffer->pages * 100;
1462 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1463 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1465 if (RB_WARN_ON(cpu_buffer,
1466 rb_is_reader_page(cpu_buffer->tail_page)))
1468 local_set(&cpu_buffer->commit_page->page->commit,
1469 rb_page_write(cpu_buffer->commit_page));
1470 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1471 cpu_buffer->write_stamp =
1472 cpu_buffer->commit_page->page->time_stamp;
1473 /* add barrier to keep gcc from optimizing too much */
1476 while (rb_commit_index(cpu_buffer) !=
1477 rb_page_write(cpu_buffer->commit_page)) {
1479 local_set(&cpu_buffer->commit_page->page->commit,
1480 rb_page_write(cpu_buffer->commit_page));
1481 RB_WARN_ON(cpu_buffer,
1482 local_read(&cpu_buffer->commit_page->page->commit) &
1487 /* again, keep gcc from optimizing */
1491 * If an interrupt came in just after the first while loop
1492 * and pushed the tail page forward, we will be left with
1493 * a dangling commit that will never go forward.
1495 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1499 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1501 cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1502 cpu_buffer->reader_page->read = 0;
1505 static void rb_inc_iter(struct ring_buffer_iter *iter)
1507 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1510 * The iterator could be on the reader page (it starts there).
1511 * But the head could have moved, since the reader was
1512 * found. Check for this case and assign the iterator
1513 * to the head page instead of next.
1515 if (iter->head_page == cpu_buffer->reader_page)
1516 iter->head_page = rb_set_head_page(cpu_buffer);
1518 rb_inc_page(cpu_buffer, &iter->head_page);
1520 iter->read_stamp = iter->head_page->page->time_stamp;
1525 * ring_buffer_update_event - update event type and data
1526 * @event: the even to update
1527 * @type: the type of event
1528 * @length: the size of the event field in the ring buffer
1530 * Update the type and data fields of the event. The length
1531 * is the actual size that is written to the ring buffer,
1532 * and with this, we can determine what to place into the
1536 rb_update_event(struct ring_buffer_event *event,
1537 unsigned type, unsigned length)
1539 event->type_len = type;
1543 case RINGBUF_TYPE_PADDING:
1544 case RINGBUF_TYPE_TIME_EXTEND:
1545 case RINGBUF_TYPE_TIME_STAMP:
1549 length -= RB_EVNT_HDR_SIZE;
1550 if (length > RB_MAX_SMALL_DATA)
1551 event->array[0] = length;
1553 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
1561 * rb_handle_head_page - writer hit the head page
1563 * Returns: +1 to retry page
1568 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1569 struct buffer_page *tail_page,
1570 struct buffer_page *next_page)
1572 struct buffer_page *new_head;
1577 entries = rb_page_entries(next_page);
1580 * The hard part is here. We need to move the head
1581 * forward, and protect against both readers on
1582 * other CPUs and writers coming in via interrupts.
1584 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1588 * type can be one of four:
1589 * NORMAL - an interrupt already moved it for us
1590 * HEAD - we are the first to get here.
1591 * UPDATE - we are the interrupt interrupting
1593 * MOVED - a reader on another CPU moved the next
1594 * pointer to its reader page. Give up
1601 * We changed the head to UPDATE, thus
1602 * it is our responsibility to update
1605 local_add(entries, &cpu_buffer->overrun);
1608 * The entries will be zeroed out when we move the
1612 /* still more to do */
1615 case RB_PAGE_UPDATE:
1617 * This is an interrupt that interrupt the
1618 * previous update. Still more to do.
1621 case RB_PAGE_NORMAL:
1623 * An interrupt came in before the update
1624 * and processed this for us.
1625 * Nothing left to do.
1630 * The reader is on another CPU and just did
1631 * a swap with our next_page.
1636 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1641 * Now that we are here, the old head pointer is
1642 * set to UPDATE. This will keep the reader from
1643 * swapping the head page with the reader page.
1644 * The reader (on another CPU) will spin till
1647 * We just need to protect against interrupts
1648 * doing the job. We will set the next pointer
1649 * to HEAD. After that, we set the old pointer
1650 * to NORMAL, but only if it was HEAD before.
1651 * otherwise we are an interrupt, and only
1652 * want the outer most commit to reset it.
1654 new_head = next_page;
1655 rb_inc_page(cpu_buffer, &new_head);
1657 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1661 * Valid returns are:
1662 * HEAD - an interrupt came in and already set it.
1663 * NORMAL - One of two things:
1664 * 1) We really set it.
1665 * 2) A bunch of interrupts came in and moved
1666 * the page forward again.
1670 case RB_PAGE_NORMAL:
1674 RB_WARN_ON(cpu_buffer, 1);
1679 * It is possible that an interrupt came in,
1680 * set the head up, then more interrupts came in
1681 * and moved it again. When we get back here,
1682 * the page would have been set to NORMAL but we
1683 * just set it back to HEAD.
1685 * How do you detect this? Well, if that happened
1686 * the tail page would have moved.
1688 if (ret == RB_PAGE_NORMAL) {
1690 * If the tail had moved passed next, then we need
1691 * to reset the pointer.
1693 if (cpu_buffer->tail_page != tail_page &&
1694 cpu_buffer->tail_page != next_page)
1695 rb_head_page_set_normal(cpu_buffer, new_head,
1701 * If this was the outer most commit (the one that
1702 * changed the original pointer from HEAD to UPDATE),
1703 * then it is up to us to reset it to NORMAL.
1705 if (type == RB_PAGE_HEAD) {
1706 ret = rb_head_page_set_normal(cpu_buffer, next_page,
1709 if (RB_WARN_ON(cpu_buffer,
1710 ret != RB_PAGE_UPDATE))
1717 static unsigned rb_calculate_event_length(unsigned length)
1719 struct ring_buffer_event event; /* Used only for sizeof array */
1721 /* zero length can cause confusions */
1725 if (length > RB_MAX_SMALL_DATA)
1726 length += sizeof(event.array[0]);
1728 length += RB_EVNT_HDR_SIZE;
1729 length = ALIGN(length, RB_ALIGNMENT);
1735 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
1736 struct buffer_page *tail_page,
1737 unsigned long tail, unsigned long length)
1739 struct ring_buffer_event *event;
1742 * Only the event that crossed the page boundary
1743 * must fill the old tail_page with padding.
1745 if (tail >= BUF_PAGE_SIZE) {
1746 local_sub(length, &tail_page->write);
1750 event = __rb_page_index(tail_page, tail);
1751 kmemcheck_annotate_bitfield(event, bitfield);
1754 * If this event is bigger than the minimum size, then
1755 * we need to be careful that we don't subtract the
1756 * write counter enough to allow another writer to slip
1758 * We put in a discarded commit instead, to make sure
1759 * that this space is not used again.
1761 * If we are less than the minimum size, we don't need to
1764 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
1765 /* No room for any events */
1767 /* Mark the rest of the page with padding */
1768 rb_event_set_padding(event);
1770 /* Set the write back to the previous setting */
1771 local_sub(length, &tail_page->write);
1775 /* Put in a discarded event */
1776 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
1777 event->type_len = RINGBUF_TYPE_PADDING;
1778 /* time delta must be non zero */
1779 event->time_delta = 1;
1780 /* Account for this as an entry */
1781 local_inc(&tail_page->entries);
1782 local_inc(&cpu_buffer->entries);
1784 /* Set write to end of buffer */
1785 length = (tail + length) - BUF_PAGE_SIZE;
1786 local_sub(length, &tail_page->write);
1789 static struct ring_buffer_event *
1790 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
1791 unsigned long length, unsigned long tail,
1792 struct buffer_page *commit_page,
1793 struct buffer_page *tail_page, u64 *ts)
1795 struct ring_buffer *buffer = cpu_buffer->buffer;
1796 struct buffer_page *next_page;
1799 next_page = tail_page;
1801 rb_inc_page(cpu_buffer, &next_page);
1804 * If for some reason, we had an interrupt storm that made
1805 * it all the way around the buffer, bail, and warn
1808 if (unlikely(next_page == commit_page)) {
1809 local_inc(&cpu_buffer->commit_overrun);
1814 * This is where the fun begins!
1816 * We are fighting against races between a reader that
1817 * could be on another CPU trying to swap its reader
1818 * page with the buffer head.
1820 * We are also fighting against interrupts coming in and
1821 * moving the head or tail on us as well.
1823 * If the next page is the head page then we have filled
1824 * the buffer, unless the commit page is still on the
1827 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
1830 * If the commit is not on the reader page, then
1831 * move the header page.
1833 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
1835 * If we are not in overwrite mode,
1836 * this is easy, just stop here.
1838 if (!(buffer->flags & RB_FL_OVERWRITE))
1841 ret = rb_handle_head_page(cpu_buffer,
1850 * We need to be careful here too. The
1851 * commit page could still be on the reader
1852 * page. We could have a small buffer, and
1853 * have filled up the buffer with events
1854 * from interrupts and such, and wrapped.
1856 * Note, if the tail page is also the on the
1857 * reader_page, we let it move out.
1859 if (unlikely((cpu_buffer->commit_page !=
1860 cpu_buffer->tail_page) &&
1861 (cpu_buffer->commit_page ==
1862 cpu_buffer->reader_page))) {
1863 local_inc(&cpu_buffer->commit_overrun);
1869 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
1872 * Nested commits always have zero deltas, so
1873 * just reread the time stamp
1875 *ts = rb_time_stamp(buffer, cpu_buffer->cpu);
1876 next_page->page->time_stamp = *ts;
1881 rb_reset_tail(cpu_buffer, tail_page, tail, length);
1883 /* fail and let the caller try again */
1884 return ERR_PTR(-EAGAIN);
1888 rb_reset_tail(cpu_buffer, tail_page, tail, length);
1893 static struct ring_buffer_event *
1894 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
1895 unsigned type, unsigned long length, u64 *ts)
1897 struct buffer_page *tail_page, *commit_page;
1898 struct ring_buffer_event *event;
1899 unsigned long tail, write;
1901 commit_page = cpu_buffer->commit_page;
1902 /* we just need to protect against interrupts */
1904 tail_page = cpu_buffer->tail_page;
1905 write = local_add_return(length, &tail_page->write);
1907 /* set write to only the index of the write */
1908 write &= RB_WRITE_MASK;
1909 tail = write - length;
1911 /* See if we shot pass the end of this buffer page */
1912 if (write > BUF_PAGE_SIZE)
1913 return rb_move_tail(cpu_buffer, length, tail,
1914 commit_page, tail_page, ts);
1916 /* We reserved something on the buffer */
1918 event = __rb_page_index(tail_page, tail);
1919 kmemcheck_annotate_bitfield(event, bitfield);
1920 rb_update_event(event, type, length);
1922 /* The passed in type is zero for DATA */
1924 local_inc(&tail_page->entries);
1927 * If this is the first commit on the page, then update
1931 tail_page->page->time_stamp = *ts;
1937 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
1938 struct ring_buffer_event *event)
1940 unsigned long new_index, old_index;
1941 struct buffer_page *bpage;
1942 unsigned long index;
1945 new_index = rb_event_index(event);
1946 old_index = new_index + rb_event_length(event);
1947 addr = (unsigned long)event;
1950 bpage = cpu_buffer->tail_page;
1952 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
1953 unsigned long write_mask =
1954 local_read(&bpage->write) & ~RB_WRITE_MASK;
1956 * This is on the tail page. It is possible that
1957 * a write could come in and move the tail page
1958 * and write to the next page. That is fine
1959 * because we just shorten what is on this page.
1961 old_index += write_mask;
1962 new_index += write_mask;
1963 index = local_cmpxchg(&bpage->write, old_index, new_index);
1964 if (index == old_index)
1968 /* could not discard */
1973 rb_add_time_stamp(struct ring_buffer_per_cpu *cpu_buffer,
1974 u64 *ts, u64 *delta)
1976 struct ring_buffer_event *event;
1980 if (unlikely(*delta > (1ULL << 59) && !once++)) {
1981 printk(KERN_WARNING "Delta way too big! %llu"
1982 " ts=%llu write stamp = %llu\n",
1983 (unsigned long long)*delta,
1984 (unsigned long long)*ts,
1985 (unsigned long long)cpu_buffer->write_stamp);
1990 * The delta is too big, we to add a
1993 event = __rb_reserve_next(cpu_buffer,
1994 RINGBUF_TYPE_TIME_EXTEND,
2000 if (PTR_ERR(event) == -EAGAIN)
2003 /* Only a commited time event can update the write stamp */
2004 if (rb_event_is_commit(cpu_buffer, event)) {
2006 * If this is the first on the page, then it was
2007 * updated with the page itself. Try to discard it
2008 * and if we can't just make it zero.
2010 if (rb_event_index(event)) {
2011 event->time_delta = *delta & TS_MASK;
2012 event->array[0] = *delta >> TS_SHIFT;
2014 /* try to discard, since we do not need this */
2015 if (!rb_try_to_discard(cpu_buffer, event)) {
2016 /* nope, just zero it */
2017 event->time_delta = 0;
2018 event->array[0] = 0;
2021 cpu_buffer->write_stamp = *ts;
2022 /* let the caller know this was the commit */
2025 /* Try to discard the event */
2026 if (!rb_try_to_discard(cpu_buffer, event)) {
2027 /* Darn, this is just wasted space */
2028 event->time_delta = 0;
2029 event->array[0] = 0;
2039 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2041 local_inc(&cpu_buffer->committing);
2042 local_inc(&cpu_buffer->commits);
2045 static void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2047 unsigned long commits;
2049 if (RB_WARN_ON(cpu_buffer,
2050 !local_read(&cpu_buffer->committing)))
2054 commits = local_read(&cpu_buffer->commits);
2055 /* synchronize with interrupts */
2057 if (local_read(&cpu_buffer->committing) == 1)
2058 rb_set_commit_to_write(cpu_buffer);
2060 local_dec(&cpu_buffer->committing);
2062 /* synchronize with interrupts */
2066 * Need to account for interrupts coming in between the
2067 * updating of the commit page and the clearing of the
2068 * committing counter.
2070 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2071 !local_read(&cpu_buffer->committing)) {
2072 local_inc(&cpu_buffer->committing);
2077 static struct ring_buffer_event *
2078 rb_reserve_next_event(struct ring_buffer_per_cpu *cpu_buffer,
2079 unsigned long length)
2081 struct ring_buffer_event *event;
2086 rb_start_commit(cpu_buffer);
2088 length = rb_calculate_event_length(length);
2091 * We allow for interrupts to reenter here and do a trace.
2092 * If one does, it will cause this original code to loop
2093 * back here. Even with heavy interrupts happening, this
2094 * should only happen a few times in a row. If this happens
2095 * 1000 times in a row, there must be either an interrupt
2096 * storm or we have something buggy.
2099 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2102 ts = rb_time_stamp(cpu_buffer->buffer, cpu_buffer->cpu);
2105 * Only the first commit can update the timestamp.
2106 * Yes there is a race here. If an interrupt comes in
2107 * just after the conditional and it traces too, then it
2108 * will also check the deltas. More than one timestamp may
2109 * also be made. But only the entry that did the actual
2110 * commit will be something other than zero.
2112 if (likely(cpu_buffer->tail_page == cpu_buffer->commit_page &&
2113 rb_page_write(cpu_buffer->tail_page) ==
2114 rb_commit_index(cpu_buffer))) {
2117 diff = ts - cpu_buffer->write_stamp;
2119 /* make sure this diff is calculated here */
2122 /* Did the write stamp get updated already? */
2123 if (unlikely(ts < cpu_buffer->write_stamp))
2127 if (unlikely(test_time_stamp(delta))) {
2129 commit = rb_add_time_stamp(cpu_buffer, &ts, &delta);
2130 if (commit == -EBUSY)
2133 if (commit == -EAGAIN)
2136 RB_WARN_ON(cpu_buffer, commit < 0);
2141 event = __rb_reserve_next(cpu_buffer, 0, length, &ts);
2142 if (unlikely(PTR_ERR(event) == -EAGAIN))
2148 if (!rb_event_is_commit(cpu_buffer, event))
2151 event->time_delta = delta;
2156 rb_end_commit(cpu_buffer);
2160 #ifdef CONFIG_TRACING
2162 #define TRACE_RECURSIVE_DEPTH 16
2164 static int trace_recursive_lock(void)
2166 current->trace_recursion++;
2168 if (likely(current->trace_recursion < TRACE_RECURSIVE_DEPTH))
2171 /* Disable all tracing before we do anything else */
2172 tracing_off_permanent();
2174 printk_once(KERN_WARNING "Tracing recursion: depth[%ld]:"
2175 "HC[%lu]:SC[%lu]:NMI[%lu]\n",
2176 current->trace_recursion,
2177 hardirq_count() >> HARDIRQ_SHIFT,
2178 softirq_count() >> SOFTIRQ_SHIFT,
2185 static void trace_recursive_unlock(void)
2187 WARN_ON_ONCE(!current->trace_recursion);
2189 current->trace_recursion--;
2194 #define trace_recursive_lock() (0)
2195 #define trace_recursive_unlock() do { } while (0)
2199 static DEFINE_PER_CPU(int, rb_need_resched);
2202 * ring_buffer_lock_reserve - reserve a part of the buffer
2203 * @buffer: the ring buffer to reserve from
2204 * @length: the length of the data to reserve (excluding event header)
2206 * Returns a reseverd event on the ring buffer to copy directly to.
2207 * The user of this interface will need to get the body to write into
2208 * and can use the ring_buffer_event_data() interface.
2210 * The length is the length of the data needed, not the event length
2211 * which also includes the event header.
2213 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2214 * If NULL is returned, then nothing has been allocated or locked.
2216 struct ring_buffer_event *
2217 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2219 struct ring_buffer_per_cpu *cpu_buffer;
2220 struct ring_buffer_event *event;
2223 if (ring_buffer_flags != RB_BUFFERS_ON)
2226 if (atomic_read(&buffer->record_disabled))
2229 /* If we are tracing schedule, we don't want to recurse */
2230 resched = ftrace_preempt_disable();
2232 if (trace_recursive_lock())
2235 cpu = raw_smp_processor_id();
2237 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2240 cpu_buffer = buffer->buffers[cpu];
2242 if (atomic_read(&cpu_buffer->record_disabled))
2245 if (length > BUF_MAX_DATA_SIZE)
2248 event = rb_reserve_next_event(cpu_buffer, length);
2253 * Need to store resched state on this cpu.
2254 * Only the first needs to.
2257 if (preempt_count() == 1)
2258 per_cpu(rb_need_resched, cpu) = resched;
2263 trace_recursive_unlock();
2266 ftrace_preempt_enable(resched);
2269 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2271 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2272 struct ring_buffer_event *event)
2274 local_inc(&cpu_buffer->entries);
2277 * The event first in the commit queue updates the
2280 if (rb_event_is_commit(cpu_buffer, event))
2281 cpu_buffer->write_stamp += event->time_delta;
2283 rb_end_commit(cpu_buffer);
2287 * ring_buffer_unlock_commit - commit a reserved
2288 * @buffer: The buffer to commit to
2289 * @event: The event pointer to commit.
2291 * This commits the data to the ring buffer, and releases any locks held.
2293 * Must be paired with ring_buffer_lock_reserve.
2295 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2296 struct ring_buffer_event *event)
2298 struct ring_buffer_per_cpu *cpu_buffer;
2299 int cpu = raw_smp_processor_id();
2301 cpu_buffer = buffer->buffers[cpu];
2303 rb_commit(cpu_buffer, event);
2305 trace_recursive_unlock();
2308 * Only the last preempt count needs to restore preemption.
2310 if (preempt_count() == 1)
2311 ftrace_preempt_enable(per_cpu(rb_need_resched, cpu));
2313 preempt_enable_no_resched_notrace();
2317 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2319 static inline void rb_event_discard(struct ring_buffer_event *event)
2321 /* array[0] holds the actual length for the discarded event */
2322 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2323 event->type_len = RINGBUF_TYPE_PADDING;
2324 /* time delta must be non zero */
2325 if (!event->time_delta)
2326 event->time_delta = 1;
2330 * ring_buffer_event_discard - discard any event in the ring buffer
2331 * @event: the event to discard
2333 * Sometimes a event that is in the ring buffer needs to be ignored.
2334 * This function lets the user discard an event in the ring buffer
2335 * and then that event will not be read later.
2337 * Note, it is up to the user to be careful with this, and protect
2338 * against races. If the user discards an event that has been consumed
2339 * it is possible that it could corrupt the ring buffer.
2341 void ring_buffer_event_discard(struct ring_buffer_event *event)
2343 rb_event_discard(event);
2345 EXPORT_SYMBOL_GPL(ring_buffer_event_discard);
2348 * ring_buffer_commit_discard - discard an event that has not been committed
2349 * @buffer: the ring buffer
2350 * @event: non committed event to discard
2352 * This is similar to ring_buffer_event_discard but must only be
2353 * performed on an event that has not been committed yet. The difference
2354 * is that this will also try to free the event from the ring buffer
2355 * if another event has not been added behind it.
2357 * If another event has been added behind it, it will set the event
2358 * up as discarded, and perform the commit.
2360 * If this function is called, do not call ring_buffer_unlock_commit on
2363 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2364 struct ring_buffer_event *event)
2366 struct ring_buffer_per_cpu *cpu_buffer;
2369 /* The event is discarded regardless */
2370 rb_event_discard(event);
2372 cpu = smp_processor_id();
2373 cpu_buffer = buffer->buffers[cpu];
2376 * This must only be called if the event has not been
2377 * committed yet. Thus we can assume that preemption
2378 * is still disabled.
2380 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2382 if (!rb_try_to_discard(cpu_buffer, event))
2386 * The commit is still visible by the reader, so we
2387 * must increment entries.
2389 local_inc(&cpu_buffer->entries);
2391 rb_end_commit(cpu_buffer);
2393 trace_recursive_unlock();
2396 * Only the last preempt count needs to restore preemption.
2398 if (preempt_count() == 1)
2399 ftrace_preempt_enable(per_cpu(rb_need_resched, cpu));
2401 preempt_enable_no_resched_notrace();
2404 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2407 * ring_buffer_write - write data to the buffer without reserving
2408 * @buffer: The ring buffer to write to.
2409 * @length: The length of the data being written (excluding the event header)
2410 * @data: The data to write to the buffer.
2412 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2413 * one function. If you already have the data to write to the buffer, it
2414 * may be easier to simply call this function.
2416 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2417 * and not the length of the event which would hold the header.
2419 int ring_buffer_write(struct ring_buffer *buffer,
2420 unsigned long length,
2423 struct ring_buffer_per_cpu *cpu_buffer;
2424 struct ring_buffer_event *event;
2429 if (ring_buffer_flags != RB_BUFFERS_ON)
2432 if (atomic_read(&buffer->record_disabled))
2435 resched = ftrace_preempt_disable();
2437 cpu = raw_smp_processor_id();
2439 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2442 cpu_buffer = buffer->buffers[cpu];
2444 if (atomic_read(&cpu_buffer->record_disabled))
2447 if (length > BUF_MAX_DATA_SIZE)
2450 event = rb_reserve_next_event(cpu_buffer, length);
2454 body = rb_event_data(event);
2456 memcpy(body, data, length);
2458 rb_commit(cpu_buffer, event);
2462 ftrace_preempt_enable(resched);
2466 EXPORT_SYMBOL_GPL(ring_buffer_write);
2468 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
2470 struct buffer_page *reader = cpu_buffer->reader_page;
2471 struct buffer_page *head = rb_set_head_page(cpu_buffer);
2472 struct buffer_page *commit = cpu_buffer->commit_page;
2474 /* In case of error, head will be NULL */
2475 if (unlikely(!head))
2478 return reader->read == rb_page_commit(reader) &&
2479 (commit == reader ||
2481 head->read == rb_page_commit(commit)));
2485 * ring_buffer_record_disable - stop all writes into the buffer
2486 * @buffer: The ring buffer to stop writes to.
2488 * This prevents all writes to the buffer. Any attempt to write
2489 * to the buffer after this will fail and return NULL.
2491 * The caller should call synchronize_sched() after this.
2493 void ring_buffer_record_disable(struct ring_buffer *buffer)
2495 atomic_inc(&buffer->record_disabled);
2497 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
2500 * ring_buffer_record_enable - enable writes to the buffer
2501 * @buffer: The ring buffer to enable writes
2503 * Note, multiple disables will need the same number of enables
2504 * to truely enable the writing (much like preempt_disable).
2506 void ring_buffer_record_enable(struct ring_buffer *buffer)
2508 atomic_dec(&buffer->record_disabled);
2510 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
2513 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
2514 * @buffer: The ring buffer to stop writes to.
2515 * @cpu: The CPU buffer to stop
2517 * This prevents all writes to the buffer. Any attempt to write
2518 * to the buffer after this will fail and return NULL.
2520 * The caller should call synchronize_sched() after this.
2522 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
2524 struct ring_buffer_per_cpu *cpu_buffer;
2526 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2529 cpu_buffer = buffer->buffers[cpu];
2530 atomic_inc(&cpu_buffer->record_disabled);
2532 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
2535 * ring_buffer_record_enable_cpu - enable writes to the buffer
2536 * @buffer: The ring buffer to enable writes
2537 * @cpu: The CPU to enable.
2539 * Note, multiple disables will need the same number of enables
2540 * to truely enable the writing (much like preempt_disable).
2542 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
2544 struct ring_buffer_per_cpu *cpu_buffer;
2546 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2549 cpu_buffer = buffer->buffers[cpu];
2550 atomic_dec(&cpu_buffer->record_disabled);
2552 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
2555 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
2556 * @buffer: The ring buffer
2557 * @cpu: The per CPU buffer to get the entries from.
2559 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
2561 struct ring_buffer_per_cpu *cpu_buffer;
2564 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2567 cpu_buffer = buffer->buffers[cpu];
2568 ret = (local_read(&cpu_buffer->entries) - local_read(&cpu_buffer->overrun))
2573 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
2576 * ring_buffer_overrun_cpu - get the number of overruns in a cpu_buffer
2577 * @buffer: The ring buffer
2578 * @cpu: The per CPU buffer to get the number of overruns from
2580 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
2582 struct ring_buffer_per_cpu *cpu_buffer;
2585 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2588 cpu_buffer = buffer->buffers[cpu];
2589 ret = local_read(&cpu_buffer->overrun);
2593 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
2596 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by commits
2597 * @buffer: The ring buffer
2598 * @cpu: The per CPU buffer to get the number of overruns from
2601 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
2603 struct ring_buffer_per_cpu *cpu_buffer;
2606 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2609 cpu_buffer = buffer->buffers[cpu];
2610 ret = local_read(&cpu_buffer->commit_overrun);
2614 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
2617 * ring_buffer_entries - get the number of entries in a buffer
2618 * @buffer: The ring buffer
2620 * Returns the total number of entries in the ring buffer
2623 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
2625 struct ring_buffer_per_cpu *cpu_buffer;
2626 unsigned long entries = 0;
2629 /* if you care about this being correct, lock the buffer */
2630 for_each_buffer_cpu(buffer, cpu) {
2631 cpu_buffer = buffer->buffers[cpu];
2632 entries += (local_read(&cpu_buffer->entries) -
2633 local_read(&cpu_buffer->overrun)) - cpu_buffer->read;
2638 EXPORT_SYMBOL_GPL(ring_buffer_entries);
2641 * ring_buffer_overrun_cpu - get the number of overruns in buffer
2642 * @buffer: The ring buffer
2644 * Returns the total number of overruns in the ring buffer
2647 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
2649 struct ring_buffer_per_cpu *cpu_buffer;
2650 unsigned long overruns = 0;
2653 /* if you care about this being correct, lock the buffer */
2654 for_each_buffer_cpu(buffer, cpu) {
2655 cpu_buffer = buffer->buffers[cpu];
2656 overruns += local_read(&cpu_buffer->overrun);
2661 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
2663 static void rb_iter_reset(struct ring_buffer_iter *iter)
2665 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2667 /* Iterator usage is expected to have record disabled */
2668 if (list_empty(&cpu_buffer->reader_page->list)) {
2669 iter->head_page = rb_set_head_page(cpu_buffer);
2670 if (unlikely(!iter->head_page))
2672 iter->head = iter->head_page->read;
2674 iter->head_page = cpu_buffer->reader_page;
2675 iter->head = cpu_buffer->reader_page->read;
2678 iter->read_stamp = cpu_buffer->read_stamp;
2680 iter->read_stamp = iter->head_page->page->time_stamp;
2684 * ring_buffer_iter_reset - reset an iterator
2685 * @iter: The iterator to reset
2687 * Resets the iterator, so that it will start from the beginning
2690 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
2692 struct ring_buffer_per_cpu *cpu_buffer;
2693 unsigned long flags;
2698 cpu_buffer = iter->cpu_buffer;
2700 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2701 rb_iter_reset(iter);
2702 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2704 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
2707 * ring_buffer_iter_empty - check if an iterator has no more to read
2708 * @iter: The iterator to check
2710 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
2712 struct ring_buffer_per_cpu *cpu_buffer;
2714 cpu_buffer = iter->cpu_buffer;
2716 return iter->head_page == cpu_buffer->commit_page &&
2717 iter->head == rb_commit_index(cpu_buffer);
2719 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
2722 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2723 struct ring_buffer_event *event)
2727 switch (event->type_len) {
2728 case RINGBUF_TYPE_PADDING:
2731 case RINGBUF_TYPE_TIME_EXTEND:
2732 delta = event->array[0];
2734 delta += event->time_delta;
2735 cpu_buffer->read_stamp += delta;
2738 case RINGBUF_TYPE_TIME_STAMP:
2739 /* FIXME: not implemented */
2742 case RINGBUF_TYPE_DATA:
2743 cpu_buffer->read_stamp += event->time_delta;
2753 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
2754 struct ring_buffer_event *event)
2758 switch (event->type_len) {
2759 case RINGBUF_TYPE_PADDING:
2762 case RINGBUF_TYPE_TIME_EXTEND:
2763 delta = event->array[0];
2765 delta += event->time_delta;
2766 iter->read_stamp += delta;
2769 case RINGBUF_TYPE_TIME_STAMP:
2770 /* FIXME: not implemented */
2773 case RINGBUF_TYPE_DATA:
2774 iter->read_stamp += event->time_delta;
2783 static struct buffer_page *
2784 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
2786 struct buffer_page *reader = NULL;
2787 unsigned long flags;
2791 local_irq_save(flags);
2792 __raw_spin_lock(&cpu_buffer->lock);
2796 * This should normally only loop twice. But because the
2797 * start of the reader inserts an empty page, it causes
2798 * a case where we will loop three times. There should be no
2799 * reason to loop four times (that I know of).
2801 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
2806 reader = cpu_buffer->reader_page;
2808 /* If there's more to read, return this page */
2809 if (cpu_buffer->reader_page->read < rb_page_size(reader))
2812 /* Never should we have an index greater than the size */
2813 if (RB_WARN_ON(cpu_buffer,
2814 cpu_buffer->reader_page->read > rb_page_size(reader)))
2817 /* check if we caught up to the tail */
2819 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
2823 * Reset the reader page to size zero.
2825 local_set(&cpu_buffer->reader_page->write, 0);
2826 local_set(&cpu_buffer->reader_page->entries, 0);
2827 local_set(&cpu_buffer->reader_page->page->commit, 0);
2831 * Splice the empty reader page into the list around the head.
2833 reader = rb_set_head_page(cpu_buffer);
2834 cpu_buffer->reader_page->list.next = reader->list.next;
2835 cpu_buffer->reader_page->list.prev = reader->list.prev;
2838 * cpu_buffer->pages just needs to point to the buffer, it
2839 * has no specific buffer page to point to. Lets move it out
2840 * of our way so we don't accidently swap it.
2842 cpu_buffer->pages = reader->list.prev;
2844 /* The reader page will be pointing to the new head */
2845 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
2848 * Here's the tricky part.
2850 * We need to move the pointer past the header page.
2851 * But we can only do that if a writer is not currently
2852 * moving it. The page before the header page has the
2853 * flag bit '1' set if it is pointing to the page we want.
2854 * but if the writer is in the process of moving it
2855 * than it will be '2' or already moved '0'.
2858 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
2861 * If we did not convert it, then we must try again.
2867 * Yeah! We succeeded in replacing the page.
2869 * Now make the new head point back to the reader page.
2871 reader->list.next->prev = &cpu_buffer->reader_page->list;
2872 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
2874 /* Finally update the reader page to the new head */
2875 cpu_buffer->reader_page = reader;
2876 rb_reset_reader_page(cpu_buffer);
2881 __raw_spin_unlock(&cpu_buffer->lock);
2882 local_irq_restore(flags);
2887 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
2889 struct ring_buffer_event *event;
2890 struct buffer_page *reader;
2893 reader = rb_get_reader_page(cpu_buffer);
2895 /* This function should not be called when buffer is empty */
2896 if (RB_WARN_ON(cpu_buffer, !reader))
2899 event = rb_reader_event(cpu_buffer);
2901 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX
2902 || rb_discarded_event(event))
2905 rb_update_read_stamp(cpu_buffer, event);
2907 length = rb_event_length(event);
2908 cpu_buffer->reader_page->read += length;
2911 static void rb_advance_iter(struct ring_buffer_iter *iter)
2913 struct ring_buffer *buffer;
2914 struct ring_buffer_per_cpu *cpu_buffer;
2915 struct ring_buffer_event *event;
2918 cpu_buffer = iter->cpu_buffer;
2919 buffer = cpu_buffer->buffer;
2922 * Check if we are at the end of the buffer.
2924 if (iter->head >= rb_page_size(iter->head_page)) {
2925 /* discarded commits can make the page empty */
2926 if (iter->head_page == cpu_buffer->commit_page)
2932 event = rb_iter_head_event(iter);
2934 length = rb_event_length(event);
2937 * This should not be called to advance the header if we are
2938 * at the tail of the buffer.
2940 if (RB_WARN_ON(cpu_buffer,
2941 (iter->head_page == cpu_buffer->commit_page) &&
2942 (iter->head + length > rb_commit_index(cpu_buffer))))
2945 rb_update_iter_read_stamp(iter, event);
2947 iter->head += length;
2949 /* check for end of page padding */
2950 if ((iter->head >= rb_page_size(iter->head_page)) &&
2951 (iter->head_page != cpu_buffer->commit_page))
2952 rb_advance_iter(iter);
2955 static struct ring_buffer_event *
2956 rb_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts)
2958 struct ring_buffer_per_cpu *cpu_buffer;
2959 struct ring_buffer_event *event;
2960 struct buffer_page *reader;
2963 cpu_buffer = buffer->buffers[cpu];
2967 * We repeat when a timestamp is encountered. It is possible
2968 * to get multiple timestamps from an interrupt entering just
2969 * as one timestamp is about to be written, or from discarded
2970 * commits. The most that we can have is the number on a single page.
2972 if (RB_WARN_ON(cpu_buffer, ++nr_loops > RB_TIMESTAMPS_PER_PAGE))
2975 reader = rb_get_reader_page(cpu_buffer);
2979 event = rb_reader_event(cpu_buffer);
2981 switch (event->type_len) {
2982 case RINGBUF_TYPE_PADDING:
2983 if (rb_null_event(event))
2984 RB_WARN_ON(cpu_buffer, 1);
2986 * Because the writer could be discarding every
2987 * event it creates (which would probably be bad)
2988 * if we were to go back to "again" then we may never
2989 * catch up, and will trigger the warn on, or lock
2990 * the box. Return the padding, and we will release
2991 * the current locks, and try again.
2993 rb_advance_reader(cpu_buffer);
2996 case RINGBUF_TYPE_TIME_EXTEND:
2997 /* Internal data, OK to advance */
2998 rb_advance_reader(cpu_buffer);
3001 case RINGBUF_TYPE_TIME_STAMP:
3002 /* FIXME: not implemented */
3003 rb_advance_reader(cpu_buffer);
3006 case RINGBUF_TYPE_DATA:
3008 *ts = cpu_buffer->read_stamp + event->time_delta;
3009 ring_buffer_normalize_time_stamp(buffer,
3010 cpu_buffer->cpu, ts);
3020 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3022 static struct ring_buffer_event *
3023 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3025 struct ring_buffer *buffer;
3026 struct ring_buffer_per_cpu *cpu_buffer;
3027 struct ring_buffer_event *event;
3030 if (ring_buffer_iter_empty(iter))
3033 cpu_buffer = iter->cpu_buffer;
3034 buffer = cpu_buffer->buffer;
3038 * We repeat when a timestamp is encountered.
3039 * We can get multiple timestamps by nested interrupts or also
3040 * if filtering is on (discarding commits). Since discarding
3041 * commits can be frequent we can get a lot of timestamps.
3042 * But we limit them by not adding timestamps if they begin
3043 * at the start of a page.
3045 if (RB_WARN_ON(cpu_buffer, ++nr_loops > RB_TIMESTAMPS_PER_PAGE))
3048 if (rb_per_cpu_empty(cpu_buffer))
3051 event = rb_iter_head_event(iter);
3053 switch (event->type_len) {
3054 case RINGBUF_TYPE_PADDING:
3055 if (rb_null_event(event)) {
3059 rb_advance_iter(iter);
3062 case RINGBUF_TYPE_TIME_EXTEND:
3063 /* Internal data, OK to advance */
3064 rb_advance_iter(iter);
3067 case RINGBUF_TYPE_TIME_STAMP:
3068 /* FIXME: not implemented */
3069 rb_advance_iter(iter);
3072 case RINGBUF_TYPE_DATA:
3074 *ts = iter->read_stamp + event->time_delta;
3075 ring_buffer_normalize_time_stamp(buffer,
3076 cpu_buffer->cpu, ts);
3086 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3088 static inline int rb_ok_to_lock(void)
3091 * If an NMI die dumps out the content of the ring buffer
3092 * do not grab locks. We also permanently disable the ring
3093 * buffer too. A one time deal is all you get from reading
3094 * the ring buffer from an NMI.
3096 if (likely(!in_nmi() && !oops_in_progress))
3099 tracing_off_permanent();
3104 * ring_buffer_peek - peek at the next event to be read
3105 * @buffer: The ring buffer to read
3106 * @cpu: The cpu to peak at
3107 * @ts: The timestamp counter of this event.
3109 * This will return the event that will be read next, but does
3110 * not consume the data.
3112 struct ring_buffer_event *
3113 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts)
3115 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3116 struct ring_buffer_event *event;
3117 unsigned long flags;
3120 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3123 dolock = rb_ok_to_lock();
3125 local_irq_save(flags);
3127 spin_lock(&cpu_buffer->reader_lock);
3128 event = rb_buffer_peek(buffer, cpu, ts);
3130 spin_unlock(&cpu_buffer->reader_lock);
3131 local_irq_restore(flags);
3133 if (event && event->type_len == RINGBUF_TYPE_PADDING) {
3142 * ring_buffer_iter_peek - peek at the next event to be read
3143 * @iter: The ring buffer iterator
3144 * @ts: The timestamp counter of this event.
3146 * This will return the event that will be read next, but does
3147 * not increment the iterator.
3149 struct ring_buffer_event *
3150 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3152 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3153 struct ring_buffer_event *event;
3154 unsigned long flags;
3157 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3158 event = rb_iter_peek(iter, ts);
3159 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3161 if (event && event->type_len == RINGBUF_TYPE_PADDING) {
3170 * ring_buffer_consume - return an event and consume it
3171 * @buffer: The ring buffer to get the next event from
3173 * Returns the next event in the ring buffer, and that event is consumed.
3174 * Meaning, that sequential reads will keep returning a different event,
3175 * and eventually empty the ring buffer if the producer is slower.
3177 struct ring_buffer_event *
3178 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts)
3180 struct ring_buffer_per_cpu *cpu_buffer;
3181 struct ring_buffer_event *event = NULL;
3182 unsigned long flags;
3185 dolock = rb_ok_to_lock();
3188 /* might be called in atomic */
3191 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3194 cpu_buffer = buffer->buffers[cpu];
3195 local_irq_save(flags);
3197 spin_lock(&cpu_buffer->reader_lock);
3199 event = rb_buffer_peek(buffer, cpu, ts);
3203 rb_advance_reader(cpu_buffer);
3207 spin_unlock(&cpu_buffer->reader_lock);
3208 local_irq_restore(flags);
3213 if (event && event->type_len == RINGBUF_TYPE_PADDING) {
3220 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3223 * ring_buffer_read_start - start a non consuming read of the buffer
3224 * @buffer: The ring buffer to read from
3225 * @cpu: The cpu buffer to iterate over
3227 * This starts up an iteration through the buffer. It also disables
3228 * the recording to the buffer until the reading is finished.
3229 * This prevents the reading from being corrupted. This is not
3230 * a consuming read, so a producer is not expected.
3232 * Must be paired with ring_buffer_finish.
3234 struct ring_buffer_iter *
3235 ring_buffer_read_start(struct ring_buffer *buffer, int cpu)
3237 struct ring_buffer_per_cpu *cpu_buffer;
3238 struct ring_buffer_iter *iter;
3239 unsigned long flags;
3241 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3244 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3248 cpu_buffer = buffer->buffers[cpu];
3250 iter->cpu_buffer = cpu_buffer;
3252 atomic_inc(&cpu_buffer->record_disabled);
3253 synchronize_sched();
3255 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3256 __raw_spin_lock(&cpu_buffer->lock);
3257 rb_iter_reset(iter);
3258 __raw_spin_unlock(&cpu_buffer->lock);
3259 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3263 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
3266 * ring_buffer_finish - finish reading the iterator of the buffer
3267 * @iter: The iterator retrieved by ring_buffer_start
3269 * This re-enables the recording to the buffer, and frees the
3273 ring_buffer_read_finish(struct ring_buffer_iter *iter)
3275 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3277 atomic_dec(&cpu_buffer->record_disabled);
3280 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
3283 * ring_buffer_read - read the next item in the ring buffer by the iterator
3284 * @iter: The ring buffer iterator
3285 * @ts: The time stamp of the event read.
3287 * This reads the next event in the ring buffer and increments the iterator.
3289 struct ring_buffer_event *
3290 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
3292 struct ring_buffer_event *event;
3293 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3294 unsigned long flags;
3297 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3298 event = rb_iter_peek(iter, ts);
3302 rb_advance_iter(iter);
3304 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3306 if (event && event->type_len == RINGBUF_TYPE_PADDING) {
3313 EXPORT_SYMBOL_GPL(ring_buffer_read);
3316 * ring_buffer_size - return the size of the ring buffer (in bytes)
3317 * @buffer: The ring buffer.
3319 unsigned long ring_buffer_size(struct ring_buffer *buffer)
3321 return BUF_PAGE_SIZE * buffer->pages;
3323 EXPORT_SYMBOL_GPL(ring_buffer_size);
3326 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
3328 rb_head_page_deactivate(cpu_buffer);
3330 cpu_buffer->head_page
3331 = list_entry(cpu_buffer->pages, struct buffer_page, list);
3332 local_set(&cpu_buffer->head_page->write, 0);
3333 local_set(&cpu_buffer->head_page->entries, 0);
3334 local_set(&cpu_buffer->head_page->page->commit, 0);
3336 cpu_buffer->head_page->read = 0;
3338 cpu_buffer->tail_page = cpu_buffer->head_page;
3339 cpu_buffer->commit_page = cpu_buffer->head_page;
3341 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
3342 local_set(&cpu_buffer->reader_page->write, 0);
3343 local_set(&cpu_buffer->reader_page->entries, 0);
3344 local_set(&cpu_buffer->reader_page->page->commit, 0);
3345 cpu_buffer->reader_page->read = 0;
3347 local_set(&cpu_buffer->commit_overrun, 0);
3348 local_set(&cpu_buffer->overrun, 0);
3349 local_set(&cpu_buffer->entries, 0);
3350 local_set(&cpu_buffer->committing, 0);
3351 local_set(&cpu_buffer->commits, 0);
3352 cpu_buffer->read = 0;
3354 cpu_buffer->write_stamp = 0;
3355 cpu_buffer->read_stamp = 0;
3357 rb_head_page_activate(cpu_buffer);
3361 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
3362 * @buffer: The ring buffer to reset a per cpu buffer of
3363 * @cpu: The CPU buffer to be reset
3365 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
3367 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3368 unsigned long flags;
3370 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3373 atomic_inc(&cpu_buffer->record_disabled);
3375 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3377 __raw_spin_lock(&cpu_buffer->lock);
3379 rb_reset_cpu(cpu_buffer);
3381 __raw_spin_unlock(&cpu_buffer->lock);
3383 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3385 atomic_dec(&cpu_buffer->record_disabled);
3387 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
3390 * ring_buffer_reset - reset a ring buffer
3391 * @buffer: The ring buffer to reset all cpu buffers
3393 void ring_buffer_reset(struct ring_buffer *buffer)
3397 for_each_buffer_cpu(buffer, cpu)
3398 ring_buffer_reset_cpu(buffer, cpu);
3400 EXPORT_SYMBOL_GPL(ring_buffer_reset);
3403 * rind_buffer_empty - is the ring buffer empty?
3404 * @buffer: The ring buffer to test
3406 int ring_buffer_empty(struct ring_buffer *buffer)
3408 struct ring_buffer_per_cpu *cpu_buffer;
3409 unsigned long flags;
3414 dolock = rb_ok_to_lock();
3416 /* yes this is racy, but if you don't like the race, lock the buffer */
3417 for_each_buffer_cpu(buffer, cpu) {
3418 cpu_buffer = buffer->buffers[cpu];
3419 local_irq_save(flags);
3421 spin_lock(&cpu_buffer->reader_lock);
3422 ret = rb_per_cpu_empty(cpu_buffer);
3424 spin_unlock(&cpu_buffer->reader_lock);
3425 local_irq_restore(flags);
3433 EXPORT_SYMBOL_GPL(ring_buffer_empty);
3436 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
3437 * @buffer: The ring buffer
3438 * @cpu: The CPU buffer to test
3440 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
3442 struct ring_buffer_per_cpu *cpu_buffer;
3443 unsigned long flags;
3447 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3450 dolock = rb_ok_to_lock();
3452 cpu_buffer = buffer->buffers[cpu];
3453 local_irq_save(flags);
3455 spin_lock(&cpu_buffer->reader_lock);
3456 ret = rb_per_cpu_empty(cpu_buffer);
3458 spin_unlock(&cpu_buffer->reader_lock);
3459 local_irq_restore(flags);
3463 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
3466 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
3467 * @buffer_a: One buffer to swap with
3468 * @buffer_b: The other buffer to swap with
3470 * This function is useful for tracers that want to take a "snapshot"
3471 * of a CPU buffer and has another back up buffer lying around.
3472 * it is expected that the tracer handles the cpu buffer not being
3473 * used at the moment.
3475 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
3476 struct ring_buffer *buffer_b, int cpu)
3478 struct ring_buffer_per_cpu *cpu_buffer_a;
3479 struct ring_buffer_per_cpu *cpu_buffer_b;
3482 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
3483 !cpumask_test_cpu(cpu, buffer_b->cpumask))
3486 /* At least make sure the two buffers are somewhat the same */
3487 if (buffer_a->pages != buffer_b->pages)
3492 if (ring_buffer_flags != RB_BUFFERS_ON)
3495 if (atomic_read(&buffer_a->record_disabled))
3498 if (atomic_read(&buffer_b->record_disabled))
3501 cpu_buffer_a = buffer_a->buffers[cpu];
3502 cpu_buffer_b = buffer_b->buffers[cpu];
3504 if (atomic_read(&cpu_buffer_a->record_disabled))
3507 if (atomic_read(&cpu_buffer_b->record_disabled))
3511 * We can't do a synchronize_sched here because this
3512 * function can be called in atomic context.
3513 * Normally this will be called from the same CPU as cpu.
3514 * If not it's up to the caller to protect this.
3516 atomic_inc(&cpu_buffer_a->record_disabled);
3517 atomic_inc(&cpu_buffer_b->record_disabled);
3519 buffer_a->buffers[cpu] = cpu_buffer_b;
3520 buffer_b->buffers[cpu] = cpu_buffer_a;
3522 cpu_buffer_b->buffer = buffer_a;
3523 cpu_buffer_a->buffer = buffer_b;
3525 atomic_dec(&cpu_buffer_a->record_disabled);
3526 atomic_dec(&cpu_buffer_b->record_disabled);
3532 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
3535 * ring_buffer_alloc_read_page - allocate a page to read from buffer
3536 * @buffer: the buffer to allocate for.
3538 * This function is used in conjunction with ring_buffer_read_page.
3539 * When reading a full page from the ring buffer, these functions
3540 * can be used to speed up the process. The calling function should
3541 * allocate a few pages first with this function. Then when it
3542 * needs to get pages from the ring buffer, it passes the result
3543 * of this function into ring_buffer_read_page, which will swap
3544 * the page that was allocated, with the read page of the buffer.
3547 * The page allocated, or NULL on error.
3549 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer)
3551 struct buffer_data_page *bpage;
3554 addr = __get_free_page(GFP_KERNEL);
3558 bpage = (void *)addr;
3560 rb_init_page(bpage);
3564 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
3567 * ring_buffer_free_read_page - free an allocated read page
3568 * @buffer: the buffer the page was allocate for
3569 * @data: the page to free
3571 * Free a page allocated from ring_buffer_alloc_read_page.
3573 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
3575 free_page((unsigned long)data);
3577 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
3580 * ring_buffer_read_page - extract a page from the ring buffer
3581 * @buffer: buffer to extract from
3582 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
3583 * @len: amount to extract
3584 * @cpu: the cpu of the buffer to extract
3585 * @full: should the extraction only happen when the page is full.
3587 * This function will pull out a page from the ring buffer and consume it.
3588 * @data_page must be the address of the variable that was returned
3589 * from ring_buffer_alloc_read_page. This is because the page might be used
3590 * to swap with a page in the ring buffer.
3593 * rpage = ring_buffer_alloc_read_page(buffer);
3596 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
3598 * process_page(rpage, ret);
3600 * When @full is set, the function will not return true unless
3601 * the writer is off the reader page.
3603 * Note: it is up to the calling functions to handle sleeps and wakeups.
3604 * The ring buffer can be used anywhere in the kernel and can not
3605 * blindly call wake_up. The layer that uses the ring buffer must be
3606 * responsible for that.
3609 * >=0 if data has been transferred, returns the offset of consumed data.
3610 * <0 if no data has been transferred.
3612 int ring_buffer_read_page(struct ring_buffer *buffer,
3613 void **data_page, size_t len, int cpu, int full)
3615 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3616 struct ring_buffer_event *event;
3617 struct buffer_data_page *bpage;
3618 struct buffer_page *reader;
3619 unsigned long flags;
3620 unsigned int commit;
3625 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3629 * If len is not big enough to hold the page header, then
3630 * we can not copy anything.
3632 if (len <= BUF_PAGE_HDR_SIZE)
3635 len -= BUF_PAGE_HDR_SIZE;
3644 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3646 reader = rb_get_reader_page(cpu_buffer);
3650 event = rb_reader_event(cpu_buffer);
3652 read = reader->read;
3653 commit = rb_page_commit(reader);
3656 * If this page has been partially read or
3657 * if len is not big enough to read the rest of the page or
3658 * a writer is still on the page, then
3659 * we must copy the data from the page to the buffer.
3660 * Otherwise, we can simply swap the page with the one passed in.
3662 if (read || (len < (commit - read)) ||
3663 cpu_buffer->reader_page == cpu_buffer->commit_page) {
3664 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
3665 unsigned int rpos = read;
3666 unsigned int pos = 0;
3672 if (len > (commit - read))
3673 len = (commit - read);
3675 size = rb_event_length(event);
3680 /* save the current timestamp, since the user will need it */
3681 save_timestamp = cpu_buffer->read_stamp;
3683 /* Need to copy one event at a time */
3685 memcpy(bpage->data + pos, rpage->data + rpos, size);
3689 rb_advance_reader(cpu_buffer);
3690 rpos = reader->read;
3693 event = rb_reader_event(cpu_buffer);
3694 size = rb_event_length(event);
3695 } while (len > size);
3698 local_set(&bpage->commit, pos);
3699 bpage->time_stamp = save_timestamp;
3701 /* we copied everything to the beginning */
3704 /* update the entry counter */
3705 cpu_buffer->read += rb_page_entries(reader);
3707 /* swap the pages */
3708 rb_init_page(bpage);
3709 bpage = reader->page;
3710 reader->page = *data_page;
3711 local_set(&reader->write, 0);
3712 local_set(&reader->entries, 0);
3719 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3724 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
3726 #ifdef CONFIG_TRACING
3728 rb_simple_read(struct file *filp, char __user *ubuf,
3729 size_t cnt, loff_t *ppos)
3731 unsigned long *p = filp->private_data;
3735 if (test_bit(RB_BUFFERS_DISABLED_BIT, p))
3736 r = sprintf(buf, "permanently disabled\n");
3738 r = sprintf(buf, "%d\n", test_bit(RB_BUFFERS_ON_BIT, p));
3740 return simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
3744 rb_simple_write(struct file *filp, const char __user *ubuf,
3745 size_t cnt, loff_t *ppos)
3747 unsigned long *p = filp->private_data;
3752 if (cnt >= sizeof(buf))
3755 if (copy_from_user(&buf, ubuf, cnt))
3760 ret = strict_strtoul(buf, 10, &val);
3765 set_bit(RB_BUFFERS_ON_BIT, p);
3767 clear_bit(RB_BUFFERS_ON_BIT, p);
3774 static const struct file_operations rb_simple_fops = {
3775 .open = tracing_open_generic,
3776 .read = rb_simple_read,
3777 .write = rb_simple_write,
3781 static __init int rb_init_debugfs(void)
3783 struct dentry *d_tracer;
3785 d_tracer = tracing_init_dentry();
3787 trace_create_file("tracing_on", 0644, d_tracer,
3788 &ring_buffer_flags, &rb_simple_fops);
3793 fs_initcall(rb_init_debugfs);
3796 #ifdef CONFIG_HOTPLUG_CPU
3797 static int rb_cpu_notify(struct notifier_block *self,
3798 unsigned long action, void *hcpu)
3800 struct ring_buffer *buffer =
3801 container_of(self, struct ring_buffer, cpu_notify);
3802 long cpu = (long)hcpu;
3805 case CPU_UP_PREPARE:
3806 case CPU_UP_PREPARE_FROZEN:
3807 if (cpumask_test_cpu(cpu, buffer->cpumask))
3810 buffer->buffers[cpu] =
3811 rb_allocate_cpu_buffer(buffer, cpu);
3812 if (!buffer->buffers[cpu]) {
3813 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
3818 cpumask_set_cpu(cpu, buffer->cpumask);
3820 case CPU_DOWN_PREPARE:
3821 case CPU_DOWN_PREPARE_FROZEN:
3824 * If we were to free the buffer, then the user would
3825 * lose any trace that was in the buffer.