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 kfree(buffer->buffers);
1174 free_cpumask_var(buffer->cpumask);
1178 EXPORT_SYMBOL_GPL(ring_buffer_free);
1180 void ring_buffer_set_clock(struct ring_buffer *buffer,
1183 buffer->clock = clock;
1186 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1189 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned nr_pages)
1191 struct buffer_page *bpage;
1192 struct list_head *p;
1195 atomic_inc(&cpu_buffer->record_disabled);
1196 synchronize_sched();
1198 rb_head_page_deactivate(cpu_buffer);
1200 for (i = 0; i < nr_pages; i++) {
1201 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1203 p = cpu_buffer->pages->next;
1204 bpage = list_entry(p, struct buffer_page, list);
1205 list_del_init(&bpage->list);
1206 free_buffer_page(bpage);
1208 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1211 rb_reset_cpu(cpu_buffer);
1213 rb_check_pages(cpu_buffer);
1215 atomic_dec(&cpu_buffer->record_disabled);
1220 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer,
1221 struct list_head *pages, unsigned nr_pages)
1223 struct buffer_page *bpage;
1224 struct list_head *p;
1227 atomic_inc(&cpu_buffer->record_disabled);
1228 synchronize_sched();
1230 spin_lock_irq(&cpu_buffer->reader_lock);
1231 rb_head_page_deactivate(cpu_buffer);
1233 for (i = 0; i < nr_pages; i++) {
1234 if (RB_WARN_ON(cpu_buffer, list_empty(pages)))
1237 bpage = list_entry(p, struct buffer_page, list);
1238 list_del_init(&bpage->list);
1239 list_add_tail(&bpage->list, cpu_buffer->pages);
1241 rb_reset_cpu(cpu_buffer);
1242 spin_unlock_irq(&cpu_buffer->reader_lock);
1244 rb_check_pages(cpu_buffer);
1246 atomic_dec(&cpu_buffer->record_disabled);
1250 * ring_buffer_resize - resize the ring buffer
1251 * @buffer: the buffer to resize.
1252 * @size: the new size.
1254 * The tracer is responsible for making sure that the buffer is
1255 * not being used while changing the size.
1256 * Note: We may be able to change the above requirement by using
1257 * RCU synchronizations.
1259 * Minimum size is 2 * BUF_PAGE_SIZE.
1261 * Returns -1 on failure.
1263 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size)
1265 struct ring_buffer_per_cpu *cpu_buffer;
1266 unsigned nr_pages, rm_pages, new_pages;
1267 struct buffer_page *bpage, *tmp;
1268 unsigned long buffer_size;
1274 * Always succeed at resizing a non-existent buffer:
1279 size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1280 size *= BUF_PAGE_SIZE;
1281 buffer_size = buffer->pages * BUF_PAGE_SIZE;
1283 /* we need a minimum of two pages */
1284 if (size < BUF_PAGE_SIZE * 2)
1285 size = BUF_PAGE_SIZE * 2;
1287 if (size == buffer_size)
1290 mutex_lock(&buffer->mutex);
1293 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1295 if (size < buffer_size) {
1297 /* easy case, just free pages */
1298 if (RB_WARN_ON(buffer, nr_pages >= buffer->pages))
1301 rm_pages = buffer->pages - nr_pages;
1303 for_each_buffer_cpu(buffer, cpu) {
1304 cpu_buffer = buffer->buffers[cpu];
1305 rb_remove_pages(cpu_buffer, rm_pages);
1311 * This is a bit more difficult. We only want to add pages
1312 * when we can allocate enough for all CPUs. We do this
1313 * by allocating all the pages and storing them on a local
1314 * link list. If we succeed in our allocation, then we
1315 * add these pages to the cpu_buffers. Otherwise we just free
1316 * them all and return -ENOMEM;
1318 if (RB_WARN_ON(buffer, nr_pages <= buffer->pages))
1321 new_pages = nr_pages - buffer->pages;
1323 for_each_buffer_cpu(buffer, cpu) {
1324 for (i = 0; i < new_pages; i++) {
1325 bpage = kzalloc_node(ALIGN(sizeof(*bpage),
1327 GFP_KERNEL, cpu_to_node(cpu));
1330 list_add(&bpage->list, &pages);
1331 addr = __get_free_page(GFP_KERNEL);
1334 bpage->page = (void *)addr;
1335 rb_init_page(bpage->page);
1339 for_each_buffer_cpu(buffer, cpu) {
1340 cpu_buffer = buffer->buffers[cpu];
1341 rb_insert_pages(cpu_buffer, &pages, new_pages);
1344 if (RB_WARN_ON(buffer, !list_empty(&pages)))
1348 buffer->pages = nr_pages;
1350 mutex_unlock(&buffer->mutex);
1355 list_for_each_entry_safe(bpage, tmp, &pages, list) {
1356 list_del_init(&bpage->list);
1357 free_buffer_page(bpage);
1360 mutex_unlock(&buffer->mutex);
1364 * Something went totally wrong, and we are too paranoid
1365 * to even clean up the mess.
1369 mutex_unlock(&buffer->mutex);
1372 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1374 static inline void *
1375 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1377 return bpage->data + index;
1380 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1382 return bpage->page->data + index;
1385 static inline struct ring_buffer_event *
1386 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1388 return __rb_page_index(cpu_buffer->reader_page,
1389 cpu_buffer->reader_page->read);
1392 static inline struct ring_buffer_event *
1393 rb_iter_head_event(struct ring_buffer_iter *iter)
1395 return __rb_page_index(iter->head_page, iter->head);
1398 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1400 return local_read(&bpage->write) & RB_WRITE_MASK;
1403 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1405 return local_read(&bpage->page->commit);
1408 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1410 return local_read(&bpage->entries) & RB_WRITE_MASK;
1413 /* Size is determined by what has been commited */
1414 static inline unsigned rb_page_size(struct buffer_page *bpage)
1416 return rb_page_commit(bpage);
1419 static inline unsigned
1420 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1422 return rb_page_commit(cpu_buffer->commit_page);
1425 static inline unsigned
1426 rb_event_index(struct ring_buffer_event *event)
1428 unsigned long addr = (unsigned long)event;
1430 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1434 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1435 struct ring_buffer_event *event)
1437 unsigned long addr = (unsigned long)event;
1438 unsigned long index;
1440 index = rb_event_index(event);
1443 return cpu_buffer->commit_page->page == (void *)addr &&
1444 rb_commit_index(cpu_buffer) == index;
1448 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1450 unsigned long max_count;
1453 * We only race with interrupts and NMIs on this CPU.
1454 * If we own the commit event, then we can commit
1455 * all others that interrupted us, since the interruptions
1456 * are in stack format (they finish before they come
1457 * back to us). This allows us to do a simple loop to
1458 * assign the commit to the tail.
1461 max_count = cpu_buffer->buffer->pages * 100;
1463 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1464 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1466 if (RB_WARN_ON(cpu_buffer,
1467 rb_is_reader_page(cpu_buffer->tail_page)))
1469 local_set(&cpu_buffer->commit_page->page->commit,
1470 rb_page_write(cpu_buffer->commit_page));
1471 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1472 cpu_buffer->write_stamp =
1473 cpu_buffer->commit_page->page->time_stamp;
1474 /* add barrier to keep gcc from optimizing too much */
1477 while (rb_commit_index(cpu_buffer) !=
1478 rb_page_write(cpu_buffer->commit_page)) {
1480 local_set(&cpu_buffer->commit_page->page->commit,
1481 rb_page_write(cpu_buffer->commit_page));
1482 RB_WARN_ON(cpu_buffer,
1483 local_read(&cpu_buffer->commit_page->page->commit) &
1488 /* again, keep gcc from optimizing */
1492 * If an interrupt came in just after the first while loop
1493 * and pushed the tail page forward, we will be left with
1494 * a dangling commit that will never go forward.
1496 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1500 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1502 cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1503 cpu_buffer->reader_page->read = 0;
1506 static void rb_inc_iter(struct ring_buffer_iter *iter)
1508 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1511 * The iterator could be on the reader page (it starts there).
1512 * But the head could have moved, since the reader was
1513 * found. Check for this case and assign the iterator
1514 * to the head page instead of next.
1516 if (iter->head_page == cpu_buffer->reader_page)
1517 iter->head_page = rb_set_head_page(cpu_buffer);
1519 rb_inc_page(cpu_buffer, &iter->head_page);
1521 iter->read_stamp = iter->head_page->page->time_stamp;
1526 * ring_buffer_update_event - update event type and data
1527 * @event: the even to update
1528 * @type: the type of event
1529 * @length: the size of the event field in the ring buffer
1531 * Update the type and data fields of the event. The length
1532 * is the actual size that is written to the ring buffer,
1533 * and with this, we can determine what to place into the
1537 rb_update_event(struct ring_buffer_event *event,
1538 unsigned type, unsigned length)
1540 event->type_len = type;
1544 case RINGBUF_TYPE_PADDING:
1545 case RINGBUF_TYPE_TIME_EXTEND:
1546 case RINGBUF_TYPE_TIME_STAMP:
1550 length -= RB_EVNT_HDR_SIZE;
1551 if (length > RB_MAX_SMALL_DATA)
1552 event->array[0] = length;
1554 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
1562 * rb_handle_head_page - writer hit the head page
1564 * Returns: +1 to retry page
1569 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1570 struct buffer_page *tail_page,
1571 struct buffer_page *next_page)
1573 struct buffer_page *new_head;
1578 entries = rb_page_entries(next_page);
1581 * The hard part is here. We need to move the head
1582 * forward, and protect against both readers on
1583 * other CPUs and writers coming in via interrupts.
1585 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1589 * type can be one of four:
1590 * NORMAL - an interrupt already moved it for us
1591 * HEAD - we are the first to get here.
1592 * UPDATE - we are the interrupt interrupting
1594 * MOVED - a reader on another CPU moved the next
1595 * pointer to its reader page. Give up
1602 * We changed the head to UPDATE, thus
1603 * it is our responsibility to update
1606 local_add(entries, &cpu_buffer->overrun);
1609 * The entries will be zeroed out when we move the
1613 /* still more to do */
1616 case RB_PAGE_UPDATE:
1618 * This is an interrupt that interrupt the
1619 * previous update. Still more to do.
1622 case RB_PAGE_NORMAL:
1624 * An interrupt came in before the update
1625 * and processed this for us.
1626 * Nothing left to do.
1631 * The reader is on another CPU and just did
1632 * a swap with our next_page.
1637 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1642 * Now that we are here, the old head pointer is
1643 * set to UPDATE. This will keep the reader from
1644 * swapping the head page with the reader page.
1645 * The reader (on another CPU) will spin till
1648 * We just need to protect against interrupts
1649 * doing the job. We will set the next pointer
1650 * to HEAD. After that, we set the old pointer
1651 * to NORMAL, but only if it was HEAD before.
1652 * otherwise we are an interrupt, and only
1653 * want the outer most commit to reset it.
1655 new_head = next_page;
1656 rb_inc_page(cpu_buffer, &new_head);
1658 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1662 * Valid returns are:
1663 * HEAD - an interrupt came in and already set it.
1664 * NORMAL - One of two things:
1665 * 1) We really set it.
1666 * 2) A bunch of interrupts came in and moved
1667 * the page forward again.
1671 case RB_PAGE_NORMAL:
1675 RB_WARN_ON(cpu_buffer, 1);
1680 * It is possible that an interrupt came in,
1681 * set the head up, then more interrupts came in
1682 * and moved it again. When we get back here,
1683 * the page would have been set to NORMAL but we
1684 * just set it back to HEAD.
1686 * How do you detect this? Well, if that happened
1687 * the tail page would have moved.
1689 if (ret == RB_PAGE_NORMAL) {
1691 * If the tail had moved passed next, then we need
1692 * to reset the pointer.
1694 if (cpu_buffer->tail_page != tail_page &&
1695 cpu_buffer->tail_page != next_page)
1696 rb_head_page_set_normal(cpu_buffer, new_head,
1702 * If this was the outer most commit (the one that
1703 * changed the original pointer from HEAD to UPDATE),
1704 * then it is up to us to reset it to NORMAL.
1706 if (type == RB_PAGE_HEAD) {
1707 ret = rb_head_page_set_normal(cpu_buffer, next_page,
1710 if (RB_WARN_ON(cpu_buffer,
1711 ret != RB_PAGE_UPDATE))
1718 static unsigned rb_calculate_event_length(unsigned length)
1720 struct ring_buffer_event event; /* Used only for sizeof array */
1722 /* zero length can cause confusions */
1726 if (length > RB_MAX_SMALL_DATA)
1727 length += sizeof(event.array[0]);
1729 length += RB_EVNT_HDR_SIZE;
1730 length = ALIGN(length, RB_ALIGNMENT);
1736 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
1737 struct buffer_page *tail_page,
1738 unsigned long tail, unsigned long length)
1740 struct ring_buffer_event *event;
1743 * Only the event that crossed the page boundary
1744 * must fill the old tail_page with padding.
1746 if (tail >= BUF_PAGE_SIZE) {
1747 local_sub(length, &tail_page->write);
1751 event = __rb_page_index(tail_page, tail);
1752 kmemcheck_annotate_bitfield(event, bitfield);
1755 * If this event is bigger than the minimum size, then
1756 * we need to be careful that we don't subtract the
1757 * write counter enough to allow another writer to slip
1759 * We put in a discarded commit instead, to make sure
1760 * that this space is not used again.
1762 * If we are less than the minimum size, we don't need to
1765 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
1766 /* No room for any events */
1768 /* Mark the rest of the page with padding */
1769 rb_event_set_padding(event);
1771 /* Set the write back to the previous setting */
1772 local_sub(length, &tail_page->write);
1776 /* Put in a discarded event */
1777 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
1778 event->type_len = RINGBUF_TYPE_PADDING;
1779 /* time delta must be non zero */
1780 event->time_delta = 1;
1781 /* Account for this as an entry */
1782 local_inc(&tail_page->entries);
1783 local_inc(&cpu_buffer->entries);
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, cpu_buffer->cpu);
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_per_cpu *cpu_buffer,
2080 unsigned long length)
2082 struct ring_buffer_event *event;
2087 rb_start_commit(cpu_buffer);
2089 length = rb_calculate_event_length(length);
2092 * We allow for interrupts to reenter here and do a trace.
2093 * If one does, it will cause this original code to loop
2094 * back here. Even with heavy interrupts happening, this
2095 * should only happen a few times in a row. If this happens
2096 * 1000 times in a row, there must be either an interrupt
2097 * storm or we have something buggy.
2100 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2103 ts = rb_time_stamp(cpu_buffer->buffer, cpu_buffer->cpu);
2106 * Only the first commit can update the timestamp.
2107 * Yes there is a race here. If an interrupt comes in
2108 * just after the conditional and it traces too, then it
2109 * will also check the deltas. More than one timestamp may
2110 * also be made. But only the entry that did the actual
2111 * commit will be something other than zero.
2113 if (likely(cpu_buffer->tail_page == cpu_buffer->commit_page &&
2114 rb_page_write(cpu_buffer->tail_page) ==
2115 rb_commit_index(cpu_buffer))) {
2118 diff = ts - cpu_buffer->write_stamp;
2120 /* make sure this diff is calculated here */
2123 /* Did the write stamp get updated already? */
2124 if (unlikely(ts < cpu_buffer->write_stamp))
2128 if (unlikely(test_time_stamp(delta))) {
2130 commit = rb_add_time_stamp(cpu_buffer, &ts, &delta);
2131 if (commit == -EBUSY)
2134 if (commit == -EAGAIN)
2137 RB_WARN_ON(cpu_buffer, commit < 0);
2142 event = __rb_reserve_next(cpu_buffer, 0, length, &ts);
2143 if (unlikely(PTR_ERR(event) == -EAGAIN))
2149 if (!rb_event_is_commit(cpu_buffer, event))
2152 event->time_delta = delta;
2157 rb_end_commit(cpu_buffer);
2161 #ifdef CONFIG_TRACING
2163 #define TRACE_RECURSIVE_DEPTH 16
2165 static int trace_recursive_lock(void)
2167 current->trace_recursion++;
2169 if (likely(current->trace_recursion < TRACE_RECURSIVE_DEPTH))
2172 /* Disable all tracing before we do anything else */
2173 tracing_off_permanent();
2175 printk_once(KERN_WARNING "Tracing recursion: depth[%ld]:"
2176 "HC[%lu]:SC[%lu]:NMI[%lu]\n",
2177 current->trace_recursion,
2178 hardirq_count() >> HARDIRQ_SHIFT,
2179 softirq_count() >> SOFTIRQ_SHIFT,
2186 static void trace_recursive_unlock(void)
2188 WARN_ON_ONCE(!current->trace_recursion);
2190 current->trace_recursion--;
2195 #define trace_recursive_lock() (0)
2196 #define trace_recursive_unlock() do { } while (0)
2200 static DEFINE_PER_CPU(int, rb_need_resched);
2203 * ring_buffer_lock_reserve - reserve a part of the buffer
2204 * @buffer: the ring buffer to reserve from
2205 * @length: the length of the data to reserve (excluding event header)
2207 * Returns a reseverd event on the ring buffer to copy directly to.
2208 * The user of this interface will need to get the body to write into
2209 * and can use the ring_buffer_event_data() interface.
2211 * The length is the length of the data needed, not the event length
2212 * which also includes the event header.
2214 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2215 * If NULL is returned, then nothing has been allocated or locked.
2217 struct ring_buffer_event *
2218 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2220 struct ring_buffer_per_cpu *cpu_buffer;
2221 struct ring_buffer_event *event;
2224 if (ring_buffer_flags != RB_BUFFERS_ON)
2227 if (atomic_read(&buffer->record_disabled))
2230 /* If we are tracing schedule, we don't want to recurse */
2231 resched = ftrace_preempt_disable();
2233 if (trace_recursive_lock())
2236 cpu = raw_smp_processor_id();
2238 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2241 cpu_buffer = buffer->buffers[cpu];
2243 if (atomic_read(&cpu_buffer->record_disabled))
2246 if (length > BUF_MAX_DATA_SIZE)
2249 event = rb_reserve_next_event(cpu_buffer, length);
2254 * Need to store resched state on this cpu.
2255 * Only the first needs to.
2258 if (preempt_count() == 1)
2259 per_cpu(rb_need_resched, cpu) = resched;
2264 trace_recursive_unlock();
2267 ftrace_preempt_enable(resched);
2270 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2272 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2273 struct ring_buffer_event *event)
2275 local_inc(&cpu_buffer->entries);
2278 * The event first in the commit queue updates the
2281 if (rb_event_is_commit(cpu_buffer, event))
2282 cpu_buffer->write_stamp += event->time_delta;
2284 rb_end_commit(cpu_buffer);
2288 * ring_buffer_unlock_commit - commit a reserved
2289 * @buffer: The buffer to commit to
2290 * @event: The event pointer to commit.
2292 * This commits the data to the ring buffer, and releases any locks held.
2294 * Must be paired with ring_buffer_lock_reserve.
2296 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2297 struct ring_buffer_event *event)
2299 struct ring_buffer_per_cpu *cpu_buffer;
2300 int cpu = raw_smp_processor_id();
2302 cpu_buffer = buffer->buffers[cpu];
2304 rb_commit(cpu_buffer, event);
2306 trace_recursive_unlock();
2309 * Only the last preempt count needs to restore preemption.
2311 if (preempt_count() == 1)
2312 ftrace_preempt_enable(per_cpu(rb_need_resched, cpu));
2314 preempt_enable_no_resched_notrace();
2318 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2320 static inline void rb_event_discard(struct ring_buffer_event *event)
2322 /* array[0] holds the actual length for the discarded event */
2323 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2324 event->type_len = RINGBUF_TYPE_PADDING;
2325 /* time delta must be non zero */
2326 if (!event->time_delta)
2327 event->time_delta = 1;
2331 * ring_buffer_event_discard - discard any event in the ring buffer
2332 * @event: the event to discard
2334 * Sometimes a event that is in the ring buffer needs to be ignored.
2335 * This function lets the user discard an event in the ring buffer
2336 * and then that event will not be read later.
2338 * Note, it is up to the user to be careful with this, and protect
2339 * against races. If the user discards an event that has been consumed
2340 * it is possible that it could corrupt the ring buffer.
2342 void ring_buffer_event_discard(struct ring_buffer_event *event)
2344 rb_event_discard(event);
2346 EXPORT_SYMBOL_GPL(ring_buffer_event_discard);
2349 * ring_buffer_commit_discard - discard an event that has not been committed
2350 * @buffer: the ring buffer
2351 * @event: non committed event to discard
2353 * This is similar to ring_buffer_event_discard but must only be
2354 * performed on an event that has not been committed yet. The difference
2355 * is that this will also try to free the event from the ring buffer
2356 * if another event has not been added behind it.
2358 * If another event has been added behind it, it will set the event
2359 * up as discarded, and perform the commit.
2361 * If this function is called, do not call ring_buffer_unlock_commit on
2364 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2365 struct ring_buffer_event *event)
2367 struct ring_buffer_per_cpu *cpu_buffer;
2370 /* The event is discarded regardless */
2371 rb_event_discard(event);
2373 cpu = smp_processor_id();
2374 cpu_buffer = buffer->buffers[cpu];
2377 * This must only be called if the event has not been
2378 * committed yet. Thus we can assume that preemption
2379 * is still disabled.
2381 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2383 if (rb_try_to_discard(cpu_buffer, event))
2387 * The commit is still visible by the reader, so we
2388 * must increment entries.
2390 local_inc(&cpu_buffer->entries);
2392 rb_end_commit(cpu_buffer);
2394 trace_recursive_unlock();
2397 * Only the last preempt count needs to restore preemption.
2399 if (preempt_count() == 1)
2400 ftrace_preempt_enable(per_cpu(rb_need_resched, cpu));
2402 preempt_enable_no_resched_notrace();
2405 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2408 * ring_buffer_write - write data to the buffer without reserving
2409 * @buffer: The ring buffer to write to.
2410 * @length: The length of the data being written (excluding the event header)
2411 * @data: The data to write to the buffer.
2413 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2414 * one function. If you already have the data to write to the buffer, it
2415 * may be easier to simply call this function.
2417 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2418 * and not the length of the event which would hold the header.
2420 int ring_buffer_write(struct ring_buffer *buffer,
2421 unsigned long length,
2424 struct ring_buffer_per_cpu *cpu_buffer;
2425 struct ring_buffer_event *event;
2430 if (ring_buffer_flags != RB_BUFFERS_ON)
2433 if (atomic_read(&buffer->record_disabled))
2436 resched = ftrace_preempt_disable();
2438 cpu = raw_smp_processor_id();
2440 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2443 cpu_buffer = buffer->buffers[cpu];
2445 if (atomic_read(&cpu_buffer->record_disabled))
2448 if (length > BUF_MAX_DATA_SIZE)
2451 event = rb_reserve_next_event(cpu_buffer, length);
2455 body = rb_event_data(event);
2457 memcpy(body, data, length);
2459 rb_commit(cpu_buffer, event);
2463 ftrace_preempt_enable(resched);
2467 EXPORT_SYMBOL_GPL(ring_buffer_write);
2469 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
2471 struct buffer_page *reader = cpu_buffer->reader_page;
2472 struct buffer_page *head = rb_set_head_page(cpu_buffer);
2473 struct buffer_page *commit = cpu_buffer->commit_page;
2475 /* In case of error, head will be NULL */
2476 if (unlikely(!head))
2479 return reader->read == rb_page_commit(reader) &&
2480 (commit == reader ||
2482 head->read == rb_page_commit(commit)));
2486 * ring_buffer_record_disable - stop all writes into the buffer
2487 * @buffer: The ring buffer to stop writes to.
2489 * This prevents all writes to the buffer. Any attempt to write
2490 * to the buffer after this will fail and return NULL.
2492 * The caller should call synchronize_sched() after this.
2494 void ring_buffer_record_disable(struct ring_buffer *buffer)
2496 atomic_inc(&buffer->record_disabled);
2498 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
2501 * ring_buffer_record_enable - enable writes to the buffer
2502 * @buffer: The ring buffer to enable writes
2504 * Note, multiple disables will need the same number of enables
2505 * to truely enable the writing (much like preempt_disable).
2507 void ring_buffer_record_enable(struct ring_buffer *buffer)
2509 atomic_dec(&buffer->record_disabled);
2511 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
2514 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
2515 * @buffer: The ring buffer to stop writes to.
2516 * @cpu: The CPU buffer to stop
2518 * This prevents all writes to the buffer. Any attempt to write
2519 * to the buffer after this will fail and return NULL.
2521 * The caller should call synchronize_sched() after this.
2523 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
2525 struct ring_buffer_per_cpu *cpu_buffer;
2527 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2530 cpu_buffer = buffer->buffers[cpu];
2531 atomic_inc(&cpu_buffer->record_disabled);
2533 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
2536 * ring_buffer_record_enable_cpu - enable writes to the buffer
2537 * @buffer: The ring buffer to enable writes
2538 * @cpu: The CPU to enable.
2540 * Note, multiple disables will need the same number of enables
2541 * to truely enable the writing (much like preempt_disable).
2543 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
2545 struct ring_buffer_per_cpu *cpu_buffer;
2547 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2550 cpu_buffer = buffer->buffers[cpu];
2551 atomic_dec(&cpu_buffer->record_disabled);
2553 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
2556 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
2557 * @buffer: The ring buffer
2558 * @cpu: The per CPU buffer to get the entries from.
2560 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
2562 struct ring_buffer_per_cpu *cpu_buffer;
2565 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2568 cpu_buffer = buffer->buffers[cpu];
2569 ret = (local_read(&cpu_buffer->entries) - local_read(&cpu_buffer->overrun))
2574 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
2577 * ring_buffer_overrun_cpu - get the number of overruns in a cpu_buffer
2578 * @buffer: The ring buffer
2579 * @cpu: The per CPU buffer to get the number of overruns from
2581 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
2583 struct ring_buffer_per_cpu *cpu_buffer;
2586 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2589 cpu_buffer = buffer->buffers[cpu];
2590 ret = local_read(&cpu_buffer->overrun);
2594 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
2597 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by commits
2598 * @buffer: The ring buffer
2599 * @cpu: The per CPU buffer to get the number of overruns from
2602 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
2604 struct ring_buffer_per_cpu *cpu_buffer;
2607 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2610 cpu_buffer = buffer->buffers[cpu];
2611 ret = local_read(&cpu_buffer->commit_overrun);
2615 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
2618 * ring_buffer_entries - get the number of entries in a buffer
2619 * @buffer: The ring buffer
2621 * Returns the total number of entries in the ring buffer
2624 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
2626 struct ring_buffer_per_cpu *cpu_buffer;
2627 unsigned long entries = 0;
2630 /* if you care about this being correct, lock the buffer */
2631 for_each_buffer_cpu(buffer, cpu) {
2632 cpu_buffer = buffer->buffers[cpu];
2633 entries += (local_read(&cpu_buffer->entries) -
2634 local_read(&cpu_buffer->overrun)) - cpu_buffer->read;
2639 EXPORT_SYMBOL_GPL(ring_buffer_entries);
2642 * ring_buffer_overrun_cpu - get the number of overruns in buffer
2643 * @buffer: The ring buffer
2645 * Returns the total number of overruns in the ring buffer
2648 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
2650 struct ring_buffer_per_cpu *cpu_buffer;
2651 unsigned long overruns = 0;
2654 /* if you care about this being correct, lock the buffer */
2655 for_each_buffer_cpu(buffer, cpu) {
2656 cpu_buffer = buffer->buffers[cpu];
2657 overruns += local_read(&cpu_buffer->overrun);
2662 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
2664 static void rb_iter_reset(struct ring_buffer_iter *iter)
2666 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2668 /* Iterator usage is expected to have record disabled */
2669 if (list_empty(&cpu_buffer->reader_page->list)) {
2670 iter->head_page = rb_set_head_page(cpu_buffer);
2671 if (unlikely(!iter->head_page))
2673 iter->head = iter->head_page->read;
2675 iter->head_page = cpu_buffer->reader_page;
2676 iter->head = cpu_buffer->reader_page->read;
2679 iter->read_stamp = cpu_buffer->read_stamp;
2681 iter->read_stamp = iter->head_page->page->time_stamp;
2685 * ring_buffer_iter_reset - reset an iterator
2686 * @iter: The iterator to reset
2688 * Resets the iterator, so that it will start from the beginning
2691 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
2693 struct ring_buffer_per_cpu *cpu_buffer;
2694 unsigned long flags;
2699 cpu_buffer = iter->cpu_buffer;
2701 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2702 rb_iter_reset(iter);
2703 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2705 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
2708 * ring_buffer_iter_empty - check if an iterator has no more to read
2709 * @iter: The iterator to check
2711 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
2713 struct ring_buffer_per_cpu *cpu_buffer;
2715 cpu_buffer = iter->cpu_buffer;
2717 return iter->head_page == cpu_buffer->commit_page &&
2718 iter->head == rb_commit_index(cpu_buffer);
2720 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
2723 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2724 struct ring_buffer_event *event)
2728 switch (event->type_len) {
2729 case RINGBUF_TYPE_PADDING:
2732 case RINGBUF_TYPE_TIME_EXTEND:
2733 delta = event->array[0];
2735 delta += event->time_delta;
2736 cpu_buffer->read_stamp += delta;
2739 case RINGBUF_TYPE_TIME_STAMP:
2740 /* FIXME: not implemented */
2743 case RINGBUF_TYPE_DATA:
2744 cpu_buffer->read_stamp += event->time_delta;
2754 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
2755 struct ring_buffer_event *event)
2759 switch (event->type_len) {
2760 case RINGBUF_TYPE_PADDING:
2763 case RINGBUF_TYPE_TIME_EXTEND:
2764 delta = event->array[0];
2766 delta += event->time_delta;
2767 iter->read_stamp += delta;
2770 case RINGBUF_TYPE_TIME_STAMP:
2771 /* FIXME: not implemented */
2774 case RINGBUF_TYPE_DATA:
2775 iter->read_stamp += event->time_delta;
2784 static struct buffer_page *
2785 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
2787 struct buffer_page *reader = NULL;
2788 unsigned long flags;
2792 local_irq_save(flags);
2793 __raw_spin_lock(&cpu_buffer->lock);
2797 * This should normally only loop twice. But because the
2798 * start of the reader inserts an empty page, it causes
2799 * a case where we will loop three times. There should be no
2800 * reason to loop four times (that I know of).
2802 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
2807 reader = cpu_buffer->reader_page;
2809 /* If there's more to read, return this page */
2810 if (cpu_buffer->reader_page->read < rb_page_size(reader))
2813 /* Never should we have an index greater than the size */
2814 if (RB_WARN_ON(cpu_buffer,
2815 cpu_buffer->reader_page->read > rb_page_size(reader)))
2818 /* check if we caught up to the tail */
2820 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
2824 * Reset the reader page to size zero.
2826 local_set(&cpu_buffer->reader_page->write, 0);
2827 local_set(&cpu_buffer->reader_page->entries, 0);
2828 local_set(&cpu_buffer->reader_page->page->commit, 0);
2832 * Splice the empty reader page into the list around the head.
2834 reader = rb_set_head_page(cpu_buffer);
2835 cpu_buffer->reader_page->list.next = reader->list.next;
2836 cpu_buffer->reader_page->list.prev = reader->list.prev;
2839 * cpu_buffer->pages just needs to point to the buffer, it
2840 * has no specific buffer page to point to. Lets move it out
2841 * of our way so we don't accidently swap it.
2843 cpu_buffer->pages = reader->list.prev;
2845 /* The reader page will be pointing to the new head */
2846 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
2849 * Here's the tricky part.
2851 * We need to move the pointer past the header page.
2852 * But we can only do that if a writer is not currently
2853 * moving it. The page before the header page has the
2854 * flag bit '1' set if it is pointing to the page we want.
2855 * but if the writer is in the process of moving it
2856 * than it will be '2' or already moved '0'.
2859 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
2862 * If we did not convert it, then we must try again.
2868 * Yeah! We succeeded in replacing the page.
2870 * Now make the new head point back to the reader page.
2872 reader->list.next->prev = &cpu_buffer->reader_page->list;
2873 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
2875 /* Finally update the reader page to the new head */
2876 cpu_buffer->reader_page = reader;
2877 rb_reset_reader_page(cpu_buffer);
2882 __raw_spin_unlock(&cpu_buffer->lock);
2883 local_irq_restore(flags);
2888 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
2890 struct ring_buffer_event *event;
2891 struct buffer_page *reader;
2894 reader = rb_get_reader_page(cpu_buffer);
2896 /* This function should not be called when buffer is empty */
2897 if (RB_WARN_ON(cpu_buffer, !reader))
2900 event = rb_reader_event(cpu_buffer);
2902 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX
2903 || rb_discarded_event(event))
2906 rb_update_read_stamp(cpu_buffer, event);
2908 length = rb_event_length(event);
2909 cpu_buffer->reader_page->read += length;
2912 static void rb_advance_iter(struct ring_buffer_iter *iter)
2914 struct ring_buffer *buffer;
2915 struct ring_buffer_per_cpu *cpu_buffer;
2916 struct ring_buffer_event *event;
2919 cpu_buffer = iter->cpu_buffer;
2920 buffer = cpu_buffer->buffer;
2923 * Check if we are at the end of the buffer.
2925 if (iter->head >= rb_page_size(iter->head_page)) {
2926 /* discarded commits can make the page empty */
2927 if (iter->head_page == cpu_buffer->commit_page)
2933 event = rb_iter_head_event(iter);
2935 length = rb_event_length(event);
2938 * This should not be called to advance the header if we are
2939 * at the tail of the buffer.
2941 if (RB_WARN_ON(cpu_buffer,
2942 (iter->head_page == cpu_buffer->commit_page) &&
2943 (iter->head + length > rb_commit_index(cpu_buffer))))
2946 rb_update_iter_read_stamp(iter, event);
2948 iter->head += length;
2950 /* check for end of page padding */
2951 if ((iter->head >= rb_page_size(iter->head_page)) &&
2952 (iter->head_page != cpu_buffer->commit_page))
2953 rb_advance_iter(iter);
2956 static struct ring_buffer_event *
2957 rb_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts)
2959 struct ring_buffer_per_cpu *cpu_buffer;
2960 struct ring_buffer_event *event;
2961 struct buffer_page *reader;
2964 cpu_buffer = buffer->buffers[cpu];
2968 * We repeat when a timestamp is encountered. It is possible
2969 * to get multiple timestamps from an interrupt entering just
2970 * as one timestamp is about to be written, or from discarded
2971 * commits. The most that we can have is the number on a single page.
2973 if (RB_WARN_ON(cpu_buffer, ++nr_loops > RB_TIMESTAMPS_PER_PAGE))
2976 reader = rb_get_reader_page(cpu_buffer);
2980 event = rb_reader_event(cpu_buffer);
2982 switch (event->type_len) {
2983 case RINGBUF_TYPE_PADDING:
2984 if (rb_null_event(event))
2985 RB_WARN_ON(cpu_buffer, 1);
2987 * Because the writer could be discarding every
2988 * event it creates (which would probably be bad)
2989 * if we were to go back to "again" then we may never
2990 * catch up, and will trigger the warn on, or lock
2991 * the box. Return the padding, and we will release
2992 * the current locks, and try again.
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()))
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);
3129 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3130 rb_advance_reader(cpu_buffer);
3132 spin_unlock(&cpu_buffer->reader_lock);
3133 local_irq_restore(flags);
3135 if (event && event->type_len == RINGBUF_TYPE_PADDING) {
3144 * ring_buffer_iter_peek - peek at the next event to be read
3145 * @iter: The ring buffer iterator
3146 * @ts: The timestamp counter of this event.
3148 * This will return the event that will be read next, but does
3149 * not increment the iterator.
3151 struct ring_buffer_event *
3152 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3154 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3155 struct ring_buffer_event *event;
3156 unsigned long flags;
3159 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3160 event = rb_iter_peek(iter, ts);
3161 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3163 if (event && event->type_len == RINGBUF_TYPE_PADDING) {
3172 * ring_buffer_consume - return an event and consume it
3173 * @buffer: The ring buffer to get the next event from
3175 * Returns the next event in the ring buffer, and that event is consumed.
3176 * Meaning, that sequential reads will keep returning a different event,
3177 * and eventually empty the ring buffer if the producer is slower.
3179 struct ring_buffer_event *
3180 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts)
3182 struct ring_buffer_per_cpu *cpu_buffer;
3183 struct ring_buffer_event *event = NULL;
3184 unsigned long flags;
3187 dolock = rb_ok_to_lock();
3190 /* might be called in atomic */
3193 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3196 cpu_buffer = buffer->buffers[cpu];
3197 local_irq_save(flags);
3199 spin_lock(&cpu_buffer->reader_lock);
3201 event = rb_buffer_peek(buffer, cpu, ts);
3203 rb_advance_reader(cpu_buffer);
3206 spin_unlock(&cpu_buffer->reader_lock);
3207 local_irq_restore(flags);
3212 if (event && event->type_len == RINGBUF_TYPE_PADDING) {
3219 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3222 * ring_buffer_read_start - start a non consuming read of the buffer
3223 * @buffer: The ring buffer to read from
3224 * @cpu: The cpu buffer to iterate over
3226 * This starts up an iteration through the buffer. It also disables
3227 * the recording to the buffer until the reading is finished.
3228 * This prevents the reading from being corrupted. This is not
3229 * a consuming read, so a producer is not expected.
3231 * Must be paired with ring_buffer_finish.
3233 struct ring_buffer_iter *
3234 ring_buffer_read_start(struct ring_buffer *buffer, int cpu)
3236 struct ring_buffer_per_cpu *cpu_buffer;
3237 struct ring_buffer_iter *iter;
3238 unsigned long flags;
3240 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3243 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3247 cpu_buffer = buffer->buffers[cpu];
3249 iter->cpu_buffer = cpu_buffer;
3251 atomic_inc(&cpu_buffer->record_disabled);
3252 synchronize_sched();
3254 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3255 __raw_spin_lock(&cpu_buffer->lock);
3256 rb_iter_reset(iter);
3257 __raw_spin_unlock(&cpu_buffer->lock);
3258 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3262 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
3265 * ring_buffer_finish - finish reading the iterator of the buffer
3266 * @iter: The iterator retrieved by ring_buffer_start
3268 * This re-enables the recording to the buffer, and frees the
3272 ring_buffer_read_finish(struct ring_buffer_iter *iter)
3274 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3276 atomic_dec(&cpu_buffer->record_disabled);
3279 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
3282 * ring_buffer_read - read the next item in the ring buffer by the iterator
3283 * @iter: The ring buffer iterator
3284 * @ts: The time stamp of the event read.
3286 * This reads the next event in the ring buffer and increments the iterator.
3288 struct ring_buffer_event *
3289 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
3291 struct ring_buffer_event *event;
3292 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3293 unsigned long flags;
3296 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3297 event = rb_iter_peek(iter, ts);
3301 rb_advance_iter(iter);
3303 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3305 if (event && event->type_len == RINGBUF_TYPE_PADDING) {
3312 EXPORT_SYMBOL_GPL(ring_buffer_read);
3315 * ring_buffer_size - return the size of the ring buffer (in bytes)
3316 * @buffer: The ring buffer.
3318 unsigned long ring_buffer_size(struct ring_buffer *buffer)
3320 return BUF_PAGE_SIZE * buffer->pages;
3322 EXPORT_SYMBOL_GPL(ring_buffer_size);
3325 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
3327 rb_head_page_deactivate(cpu_buffer);
3329 cpu_buffer->head_page
3330 = list_entry(cpu_buffer->pages, struct buffer_page, list);
3331 local_set(&cpu_buffer->head_page->write, 0);
3332 local_set(&cpu_buffer->head_page->entries, 0);
3333 local_set(&cpu_buffer->head_page->page->commit, 0);
3335 cpu_buffer->head_page->read = 0;
3337 cpu_buffer->tail_page = cpu_buffer->head_page;
3338 cpu_buffer->commit_page = cpu_buffer->head_page;
3340 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
3341 local_set(&cpu_buffer->reader_page->write, 0);
3342 local_set(&cpu_buffer->reader_page->entries, 0);
3343 local_set(&cpu_buffer->reader_page->page->commit, 0);
3344 cpu_buffer->reader_page->read = 0;
3346 local_set(&cpu_buffer->commit_overrun, 0);
3347 local_set(&cpu_buffer->overrun, 0);
3348 local_set(&cpu_buffer->entries, 0);
3349 local_set(&cpu_buffer->committing, 0);
3350 local_set(&cpu_buffer->commits, 0);
3351 cpu_buffer->read = 0;
3353 cpu_buffer->write_stamp = 0;
3354 cpu_buffer->read_stamp = 0;
3356 rb_head_page_activate(cpu_buffer);
3360 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
3361 * @buffer: The ring buffer to reset a per cpu buffer of
3362 * @cpu: The CPU buffer to be reset
3364 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
3366 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3367 unsigned long flags;
3369 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3372 atomic_inc(&cpu_buffer->record_disabled);
3374 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3376 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
3379 __raw_spin_lock(&cpu_buffer->lock);
3381 rb_reset_cpu(cpu_buffer);
3383 __raw_spin_unlock(&cpu_buffer->lock);
3386 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3388 atomic_dec(&cpu_buffer->record_disabled);
3390 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
3393 * ring_buffer_reset - reset a ring buffer
3394 * @buffer: The ring buffer to reset all cpu buffers
3396 void ring_buffer_reset(struct ring_buffer *buffer)
3400 for_each_buffer_cpu(buffer, cpu)
3401 ring_buffer_reset_cpu(buffer, cpu);
3403 EXPORT_SYMBOL_GPL(ring_buffer_reset);
3406 * rind_buffer_empty - is the ring buffer empty?
3407 * @buffer: The ring buffer to test
3409 int ring_buffer_empty(struct ring_buffer *buffer)
3411 struct ring_buffer_per_cpu *cpu_buffer;
3412 unsigned long flags;
3417 dolock = rb_ok_to_lock();
3419 /* yes this is racy, but if you don't like the race, lock the buffer */
3420 for_each_buffer_cpu(buffer, cpu) {
3421 cpu_buffer = buffer->buffers[cpu];
3422 local_irq_save(flags);
3424 spin_lock(&cpu_buffer->reader_lock);
3425 ret = rb_per_cpu_empty(cpu_buffer);
3427 spin_unlock(&cpu_buffer->reader_lock);
3428 local_irq_restore(flags);
3436 EXPORT_SYMBOL_GPL(ring_buffer_empty);
3439 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
3440 * @buffer: The ring buffer
3441 * @cpu: The CPU buffer to test
3443 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
3445 struct ring_buffer_per_cpu *cpu_buffer;
3446 unsigned long flags;
3450 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3453 dolock = rb_ok_to_lock();
3455 cpu_buffer = buffer->buffers[cpu];
3456 local_irq_save(flags);
3458 spin_lock(&cpu_buffer->reader_lock);
3459 ret = rb_per_cpu_empty(cpu_buffer);
3461 spin_unlock(&cpu_buffer->reader_lock);
3462 local_irq_restore(flags);
3466 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
3469 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
3470 * @buffer_a: One buffer to swap with
3471 * @buffer_b: The other buffer to swap with
3473 * This function is useful for tracers that want to take a "snapshot"
3474 * of a CPU buffer and has another back up buffer lying around.
3475 * it is expected that the tracer handles the cpu buffer not being
3476 * used at the moment.
3478 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
3479 struct ring_buffer *buffer_b, int cpu)
3481 struct ring_buffer_per_cpu *cpu_buffer_a;
3482 struct ring_buffer_per_cpu *cpu_buffer_b;
3485 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
3486 !cpumask_test_cpu(cpu, buffer_b->cpumask))
3489 /* At least make sure the two buffers are somewhat the same */
3490 if (buffer_a->pages != buffer_b->pages)
3495 if (ring_buffer_flags != RB_BUFFERS_ON)
3498 if (atomic_read(&buffer_a->record_disabled))
3501 if (atomic_read(&buffer_b->record_disabled))
3504 cpu_buffer_a = buffer_a->buffers[cpu];
3505 cpu_buffer_b = buffer_b->buffers[cpu];
3507 if (atomic_read(&cpu_buffer_a->record_disabled))
3510 if (atomic_read(&cpu_buffer_b->record_disabled))
3514 * We can't do a synchronize_sched here because this
3515 * function can be called in atomic context.
3516 * Normally this will be called from the same CPU as cpu.
3517 * If not it's up to the caller to protect this.
3519 atomic_inc(&cpu_buffer_a->record_disabled);
3520 atomic_inc(&cpu_buffer_b->record_disabled);
3522 buffer_a->buffers[cpu] = cpu_buffer_b;
3523 buffer_b->buffers[cpu] = cpu_buffer_a;
3525 cpu_buffer_b->buffer = buffer_a;
3526 cpu_buffer_a->buffer = buffer_b;
3528 atomic_dec(&cpu_buffer_a->record_disabled);
3529 atomic_dec(&cpu_buffer_b->record_disabled);
3535 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
3538 * ring_buffer_alloc_read_page - allocate a page to read from buffer
3539 * @buffer: the buffer to allocate for.
3541 * This function is used in conjunction with ring_buffer_read_page.
3542 * When reading a full page from the ring buffer, these functions
3543 * can be used to speed up the process. The calling function should
3544 * allocate a few pages first with this function. Then when it
3545 * needs to get pages from the ring buffer, it passes the result
3546 * of this function into ring_buffer_read_page, which will swap
3547 * the page that was allocated, with the read page of the buffer.
3550 * The page allocated, or NULL on error.
3552 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer)
3554 struct buffer_data_page *bpage;
3557 addr = __get_free_page(GFP_KERNEL);
3561 bpage = (void *)addr;
3563 rb_init_page(bpage);
3567 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
3570 * ring_buffer_free_read_page - free an allocated read page
3571 * @buffer: the buffer the page was allocate for
3572 * @data: the page to free
3574 * Free a page allocated from ring_buffer_alloc_read_page.
3576 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
3578 free_page((unsigned long)data);
3580 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
3583 * ring_buffer_read_page - extract a page from the ring buffer
3584 * @buffer: buffer to extract from
3585 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
3586 * @len: amount to extract
3587 * @cpu: the cpu of the buffer to extract
3588 * @full: should the extraction only happen when the page is full.
3590 * This function will pull out a page from the ring buffer and consume it.
3591 * @data_page must be the address of the variable that was returned
3592 * from ring_buffer_alloc_read_page. This is because the page might be used
3593 * to swap with a page in the ring buffer.
3596 * rpage = ring_buffer_alloc_read_page(buffer);
3599 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
3601 * process_page(rpage, ret);
3603 * When @full is set, the function will not return true unless
3604 * the writer is off the reader page.
3606 * Note: it is up to the calling functions to handle sleeps and wakeups.
3607 * The ring buffer can be used anywhere in the kernel and can not
3608 * blindly call wake_up. The layer that uses the ring buffer must be
3609 * responsible for that.
3612 * >=0 if data has been transferred, returns the offset of consumed data.
3613 * <0 if no data has been transferred.
3615 int ring_buffer_read_page(struct ring_buffer *buffer,
3616 void **data_page, size_t len, int cpu, int full)
3618 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3619 struct ring_buffer_event *event;
3620 struct buffer_data_page *bpage;
3621 struct buffer_page *reader;
3622 unsigned long flags;
3623 unsigned int commit;
3628 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3632 * If len is not big enough to hold the page header, then
3633 * we can not copy anything.
3635 if (len <= BUF_PAGE_HDR_SIZE)
3638 len -= BUF_PAGE_HDR_SIZE;
3647 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3649 reader = rb_get_reader_page(cpu_buffer);
3653 event = rb_reader_event(cpu_buffer);
3655 read = reader->read;
3656 commit = rb_page_commit(reader);
3659 * If this page has been partially read or
3660 * if len is not big enough to read the rest of the page or
3661 * a writer is still on the page, then
3662 * we must copy the data from the page to the buffer.
3663 * Otherwise, we can simply swap the page with the one passed in.
3665 if (read || (len < (commit - read)) ||
3666 cpu_buffer->reader_page == cpu_buffer->commit_page) {
3667 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
3668 unsigned int rpos = read;
3669 unsigned int pos = 0;
3675 if (len > (commit - read))
3676 len = (commit - read);
3678 size = rb_event_length(event);
3683 /* save the current timestamp, since the user will need it */
3684 save_timestamp = cpu_buffer->read_stamp;
3686 /* Need to copy one event at a time */
3688 memcpy(bpage->data + pos, rpage->data + rpos, size);
3692 rb_advance_reader(cpu_buffer);
3693 rpos = reader->read;
3696 event = rb_reader_event(cpu_buffer);
3697 size = rb_event_length(event);
3698 } while (len > size);
3701 local_set(&bpage->commit, pos);
3702 bpage->time_stamp = save_timestamp;
3704 /* we copied everything to the beginning */
3707 /* update the entry counter */
3708 cpu_buffer->read += rb_page_entries(reader);
3710 /* swap the pages */
3711 rb_init_page(bpage);
3712 bpage = reader->page;
3713 reader->page = *data_page;
3714 local_set(&reader->write, 0);
3715 local_set(&reader->entries, 0);
3722 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3727 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
3729 #ifdef CONFIG_TRACING
3731 rb_simple_read(struct file *filp, char __user *ubuf,
3732 size_t cnt, loff_t *ppos)
3734 unsigned long *p = filp->private_data;
3738 if (test_bit(RB_BUFFERS_DISABLED_BIT, p))
3739 r = sprintf(buf, "permanently disabled\n");
3741 r = sprintf(buf, "%d\n", test_bit(RB_BUFFERS_ON_BIT, p));
3743 return simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
3747 rb_simple_write(struct file *filp, const char __user *ubuf,
3748 size_t cnt, loff_t *ppos)
3750 unsigned long *p = filp->private_data;
3755 if (cnt >= sizeof(buf))
3758 if (copy_from_user(&buf, ubuf, cnt))
3763 ret = strict_strtoul(buf, 10, &val);
3768 set_bit(RB_BUFFERS_ON_BIT, p);
3770 clear_bit(RB_BUFFERS_ON_BIT, p);
3777 static const struct file_operations rb_simple_fops = {
3778 .open = tracing_open_generic,
3779 .read = rb_simple_read,
3780 .write = rb_simple_write,
3784 static __init int rb_init_debugfs(void)
3786 struct dentry *d_tracer;
3788 d_tracer = tracing_init_dentry();
3790 trace_create_file("tracing_on", 0644, d_tracer,
3791 &ring_buffer_flags, &rb_simple_fops);
3796 fs_initcall(rb_init_debugfs);
3799 #ifdef CONFIG_HOTPLUG_CPU
3800 static int rb_cpu_notify(struct notifier_block *self,
3801 unsigned long action, void *hcpu)
3803 struct ring_buffer *buffer =
3804 container_of(self, struct ring_buffer, cpu_notify);
3805 long cpu = (long)hcpu;
3808 case CPU_UP_PREPARE:
3809 case CPU_UP_PREPARE_FROZEN:
3810 if (cpumask_test_cpu(cpu, buffer->cpumask))
3813 buffer->buffers[cpu] =
3814 rb_allocate_cpu_buffer(buffer, cpu);
3815 if (!buffer->buffers[cpu]) {
3816 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
3821 cpumask_set_cpu(cpu, buffer->cpumask);
3823 case CPU_DOWN_PREPARE:
3824 case CPU_DOWN_PREPARE_FROZEN:
3827 * If we were to free the buffer, then the user would
3828 * lose any trace that was in the buffer.