4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 #include <linux/ring_buffer.h>
7 #include <linux/trace_clock.h>
8 #include <linux/spinlock.h>
9 #include <linux/debugfs.h>
10 #include <linux/uaccess.h>
11 #include <linux/hardirq.h>
12 #include <linux/kmemcheck.h>
13 #include <linux/module.h>
14 #include <linux/percpu.h>
15 #include <linux/mutex.h>
16 #include <linux/slab.h>
17 #include <linux/init.h>
18 #include <linux/hash.h>
19 #include <linux/list.h>
20 #include <linux/cpu.h>
23 #include <asm/local.h>
27 * The ring buffer header is special. We must manually up keep it.
29 int ring_buffer_print_entry_header(struct trace_seq *s)
33 ret = trace_seq_printf(s, "# compressed entry header\n");
34 ret = trace_seq_printf(s, "\ttype_len : 5 bits\n");
35 ret = trace_seq_printf(s, "\ttime_delta : 27 bits\n");
36 ret = trace_seq_printf(s, "\tarray : 32 bits\n");
37 ret = trace_seq_printf(s, "\n");
38 ret = trace_seq_printf(s, "\tpadding : type == %d\n",
39 RINGBUF_TYPE_PADDING);
40 ret = trace_seq_printf(s, "\ttime_extend : type == %d\n",
41 RINGBUF_TYPE_TIME_EXTEND);
42 ret = trace_seq_printf(s, "\tdata max type_len == %d\n",
43 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
49 * The ring buffer is made up of a list of pages. A separate list of pages is
50 * allocated for each CPU. A writer may only write to a buffer that is
51 * associated with the CPU it is currently executing on. A reader may read
52 * from any per cpu buffer.
54 * The reader is special. For each per cpu buffer, the reader has its own
55 * reader page. When a reader has read the entire reader page, this reader
56 * page is swapped with another page in the ring buffer.
58 * Now, as long as the writer is off the reader page, the reader can do what
59 * ever it wants with that page. The writer will never write to that page
60 * again (as long as it is out of the ring buffer).
62 * Here's some silly ASCII art.
65 * |reader| RING BUFFER
67 * +------+ +---+ +---+ +---+
76 * |reader| RING BUFFER
77 * |page |------------------v
78 * +------+ +---+ +---+ +---+
87 * |reader| RING BUFFER
88 * |page |------------------v
89 * +------+ +---+ +---+ +---+
94 * +------------------------------+
98 * |buffer| RING BUFFER
99 * |page |------------------v
100 * +------+ +---+ +---+ +---+
102 * | New +---+ +---+ +---+
105 * +------------------------------+
108 * After we make this swap, the reader can hand this page off to the splice
109 * code and be done with it. It can even allocate a new page if it needs to
110 * and swap that into the ring buffer.
112 * We will be using cmpxchg soon to make all this lockless.
117 * A fast way to enable or disable all ring buffers is to
118 * call tracing_on or tracing_off. Turning off the ring buffers
119 * prevents all ring buffers from being recorded to.
120 * Turning this switch on, makes it OK to write to the
121 * ring buffer, if the ring buffer is enabled itself.
123 * There's three layers that must be on in order to write
124 * to the ring buffer.
126 * 1) This global flag must be set.
127 * 2) The ring buffer must be enabled for recording.
128 * 3) The per cpu buffer must be enabled for recording.
130 * In case of an anomaly, this global flag has a bit set that
131 * will permantly disable all ring buffers.
135 * Global flag to disable all recording to ring buffers
136 * This has two bits: ON, DISABLED
140 * 0 0 : ring buffers are off
141 * 1 0 : ring buffers are on
142 * X 1 : ring buffers are permanently disabled
146 RB_BUFFERS_ON_BIT = 0,
147 RB_BUFFERS_DISABLED_BIT = 1,
151 RB_BUFFERS_ON = 1 << RB_BUFFERS_ON_BIT,
152 RB_BUFFERS_DISABLED = 1 << RB_BUFFERS_DISABLED_BIT,
155 static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON;
157 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
160 * tracing_on - enable all tracing buffers
162 * This function enables all tracing buffers that may have been
163 * disabled with tracing_off.
165 void tracing_on(void)
167 set_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
169 EXPORT_SYMBOL_GPL(tracing_on);
172 * tracing_off - turn off all tracing buffers
174 * This function stops all tracing buffers from recording data.
175 * It does not disable any overhead the tracers themselves may
176 * be causing. This function simply causes all recording to
177 * the ring buffers to fail.
179 void tracing_off(void)
181 clear_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
183 EXPORT_SYMBOL_GPL(tracing_off);
186 * tracing_off_permanent - permanently disable ring buffers
188 * This function, once called, will disable all ring buffers
191 void tracing_off_permanent(void)
193 set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags);
197 * tracing_is_on - show state of ring buffers enabled
199 int tracing_is_on(void)
201 return ring_buffer_flags == RB_BUFFERS_ON;
203 EXPORT_SYMBOL_GPL(tracing_is_on);
205 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
206 #define RB_ALIGNMENT 4U
207 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
208 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
210 #if !defined(CONFIG_64BIT) || defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
211 # define RB_FORCE_8BYTE_ALIGNMENT 0
212 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
214 # define RB_FORCE_8BYTE_ALIGNMENT 1
215 # define RB_ARCH_ALIGNMENT 8U
218 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
219 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
222 RB_LEN_TIME_EXTEND = 8,
223 RB_LEN_TIME_STAMP = 16,
226 #define skip_time_extend(event) \
227 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
229 static inline int rb_null_event(struct ring_buffer_event *event)
231 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
234 static void rb_event_set_padding(struct ring_buffer_event *event)
236 /* padding has a NULL time_delta */
237 event->type_len = RINGBUF_TYPE_PADDING;
238 event->time_delta = 0;
242 rb_event_data_length(struct ring_buffer_event *event)
247 length = event->type_len * RB_ALIGNMENT;
249 length = event->array[0];
250 return length + RB_EVNT_HDR_SIZE;
254 * Return the length of the given event. Will return
255 * the length of the time extend if the event is a
258 static inline unsigned
259 rb_event_length(struct ring_buffer_event *event)
261 switch (event->type_len) {
262 case RINGBUF_TYPE_PADDING:
263 if (rb_null_event(event))
266 return event->array[0] + RB_EVNT_HDR_SIZE;
268 case RINGBUF_TYPE_TIME_EXTEND:
269 return RB_LEN_TIME_EXTEND;
271 case RINGBUF_TYPE_TIME_STAMP:
272 return RB_LEN_TIME_STAMP;
274 case RINGBUF_TYPE_DATA:
275 return rb_event_data_length(event);
284 * Return total length of time extend and data,
285 * or just the event length for all other events.
287 static inline unsigned
288 rb_event_ts_length(struct ring_buffer_event *event)
292 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
293 /* time extends include the data event after it */
294 len = RB_LEN_TIME_EXTEND;
295 event = skip_time_extend(event);
297 return len + rb_event_length(event);
301 * ring_buffer_event_length - return the length of the event
302 * @event: the event to get the length of
304 * Returns the size of the data load of a data event.
305 * If the event is something other than a data event, it
306 * returns the size of the event itself. With the exception
307 * of a TIME EXTEND, where it still returns the size of the
308 * data load of the data event after it.
310 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
314 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
315 event = skip_time_extend(event);
317 length = rb_event_length(event);
318 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
320 length -= RB_EVNT_HDR_SIZE;
321 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
322 length -= sizeof(event->array[0]);
325 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
327 /* inline for ring buffer fast paths */
329 rb_event_data(struct ring_buffer_event *event)
331 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
332 event = skip_time_extend(event);
333 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
334 /* If length is in len field, then array[0] has the data */
336 return (void *)&event->array[0];
337 /* Otherwise length is in array[0] and array[1] has the data */
338 return (void *)&event->array[1];
342 * ring_buffer_event_data - return the data of the event
343 * @event: the event to get the data from
345 void *ring_buffer_event_data(struct ring_buffer_event *event)
347 return rb_event_data(event);
349 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
351 #define for_each_buffer_cpu(buffer, cpu) \
352 for_each_cpu(cpu, buffer->cpumask)
355 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
356 #define TS_DELTA_TEST (~TS_MASK)
358 /* Flag when events were overwritten */
359 #define RB_MISSED_EVENTS (1 << 31)
360 /* Missed count stored at end */
361 #define RB_MISSED_STORED (1 << 30)
363 struct buffer_data_page {
364 u64 time_stamp; /* page time stamp */
365 local_t commit; /* write committed index */
366 unsigned char data[]; /* data of buffer page */
370 * Note, the buffer_page list must be first. The buffer pages
371 * are allocated in cache lines, which means that each buffer
372 * page will be at the beginning of a cache line, and thus
373 * the least significant bits will be zero. We use this to
374 * add flags in the list struct pointers, to make the ring buffer
378 struct list_head list; /* list of buffer pages */
379 local_t write; /* index for next write */
380 unsigned read; /* index for next read */
381 local_t entries; /* entries on this page */
382 unsigned long real_end; /* real end of data */
383 struct buffer_data_page *page; /* Actual data page */
387 * The buffer page counters, write and entries, must be reset
388 * atomically when crossing page boundaries. To synchronize this
389 * update, two counters are inserted into the number. One is
390 * the actual counter for the write position or count on the page.
392 * The other is a counter of updaters. Before an update happens
393 * the update partition of the counter is incremented. This will
394 * allow the updater to update the counter atomically.
396 * The counter is 20 bits, and the state data is 12.
398 #define RB_WRITE_MASK 0xfffff
399 #define RB_WRITE_INTCNT (1 << 20)
401 static void rb_init_page(struct buffer_data_page *bpage)
403 local_set(&bpage->commit, 0);
407 * ring_buffer_page_len - the size of data on the page.
408 * @page: The page to read
410 * Returns the amount of data on the page, including buffer page header.
412 size_t ring_buffer_page_len(void *page)
414 return local_read(&((struct buffer_data_page *)page)->commit)
419 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
422 static void free_buffer_page(struct buffer_page *bpage)
424 free_page((unsigned long)bpage->page);
429 * We need to fit the time_stamp delta into 27 bits.
431 static inline int test_time_stamp(u64 delta)
433 if (delta & TS_DELTA_TEST)
438 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
440 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
441 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
443 int ring_buffer_print_page_header(struct trace_seq *s)
445 struct buffer_data_page field;
448 ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t"
449 "offset:0;\tsize:%u;\tsigned:%u;\n",
450 (unsigned int)sizeof(field.time_stamp),
451 (unsigned int)is_signed_type(u64));
453 ret = trace_seq_printf(s, "\tfield: local_t commit;\t"
454 "offset:%u;\tsize:%u;\tsigned:%u;\n",
455 (unsigned int)offsetof(typeof(field), commit),
456 (unsigned int)sizeof(field.commit),
457 (unsigned int)is_signed_type(long));
459 ret = trace_seq_printf(s, "\tfield: int overwrite;\t"
460 "offset:%u;\tsize:%u;\tsigned:%u;\n",
461 (unsigned int)offsetof(typeof(field), commit),
463 (unsigned int)is_signed_type(long));
465 ret = trace_seq_printf(s, "\tfield: char data;\t"
466 "offset:%u;\tsize:%u;\tsigned:%u;\n",
467 (unsigned int)offsetof(typeof(field), data),
468 (unsigned int)BUF_PAGE_SIZE,
469 (unsigned int)is_signed_type(char));
475 * head_page == tail_page && head == tail then buffer is empty.
477 struct ring_buffer_per_cpu {
479 atomic_t record_disabled;
480 struct ring_buffer *buffer;
481 spinlock_t reader_lock; /* serialize readers */
482 arch_spinlock_t lock;
483 struct lock_class_key lock_key;
484 struct list_head *pages;
485 struct buffer_page *head_page; /* read from head */
486 struct buffer_page *tail_page; /* write to tail */
487 struct buffer_page *commit_page; /* committed pages */
488 struct buffer_page *reader_page;
489 unsigned long lost_events;
490 unsigned long last_overrun;
491 local_t commit_overrun;
505 atomic_t record_disabled;
506 cpumask_var_t cpumask;
508 struct lock_class_key *reader_lock_key;
512 struct ring_buffer_per_cpu **buffers;
514 #ifdef CONFIG_HOTPLUG_CPU
515 struct notifier_block cpu_notify;
520 struct ring_buffer_iter {
521 struct ring_buffer_per_cpu *cpu_buffer;
523 struct buffer_page *head_page;
524 struct buffer_page *cache_reader_page;
525 unsigned long cache_read;
529 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
530 #define RB_WARN_ON(b, cond) \
532 int _____ret = unlikely(cond); \
534 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
535 struct ring_buffer_per_cpu *__b = \
537 atomic_inc(&__b->buffer->record_disabled); \
539 atomic_inc(&b->record_disabled); \
545 /* Up this if you want to test the TIME_EXTENTS and normalization */
546 #define DEBUG_SHIFT 0
548 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
550 /* shift to debug/test normalization and TIME_EXTENTS */
551 return buffer->clock() << DEBUG_SHIFT;
554 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
558 preempt_disable_notrace();
559 time = rb_time_stamp(buffer);
560 preempt_enable_no_resched_notrace();
564 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
566 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
569 /* Just stupid testing the normalize function and deltas */
572 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
575 * Making the ring buffer lockless makes things tricky.
576 * Although writes only happen on the CPU that they are on,
577 * and they only need to worry about interrupts. Reads can
580 * The reader page is always off the ring buffer, but when the
581 * reader finishes with a page, it needs to swap its page with
582 * a new one from the buffer. The reader needs to take from
583 * the head (writes go to the tail). But if a writer is in overwrite
584 * mode and wraps, it must push the head page forward.
586 * Here lies the problem.
588 * The reader must be careful to replace only the head page, and
589 * not another one. As described at the top of the file in the
590 * ASCII art, the reader sets its old page to point to the next
591 * page after head. It then sets the page after head to point to
592 * the old reader page. But if the writer moves the head page
593 * during this operation, the reader could end up with the tail.
595 * We use cmpxchg to help prevent this race. We also do something
596 * special with the page before head. We set the LSB to 1.
598 * When the writer must push the page forward, it will clear the
599 * bit that points to the head page, move the head, and then set
600 * the bit that points to the new head page.
602 * We also don't want an interrupt coming in and moving the head
603 * page on another writer. Thus we use the second LSB to catch
606 * head->list->prev->next bit 1 bit 0
609 * Points to head page 0 1
612 * Note we can not trust the prev pointer of the head page, because:
614 * +----+ +-----+ +-----+
615 * | |------>| T |---X--->| N |
617 * +----+ +-----+ +-----+
620 * +----------| R |----------+ |
624 * Key: ---X--> HEAD flag set in pointer
629 * (see __rb_reserve_next() to see where this happens)
631 * What the above shows is that the reader just swapped out
632 * the reader page with a page in the buffer, but before it
633 * could make the new header point back to the new page added
634 * it was preempted by a writer. The writer moved forward onto
635 * the new page added by the reader and is about to move forward
638 * You can see, it is legitimate for the previous pointer of
639 * the head (or any page) not to point back to itself. But only
643 #define RB_PAGE_NORMAL 0UL
644 #define RB_PAGE_HEAD 1UL
645 #define RB_PAGE_UPDATE 2UL
648 #define RB_FLAG_MASK 3UL
650 /* PAGE_MOVED is not part of the mask */
651 #define RB_PAGE_MOVED 4UL
654 * rb_list_head - remove any bit
656 static struct list_head *rb_list_head(struct list_head *list)
658 unsigned long val = (unsigned long)list;
660 return (struct list_head *)(val & ~RB_FLAG_MASK);
664 * rb_is_head_page - test if the given page is the head page
666 * Because the reader may move the head_page pointer, we can
667 * not trust what the head page is (it may be pointing to
668 * the reader page). But if the next page is a header page,
669 * its flags will be non zero.
672 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
673 struct buffer_page *page, struct list_head *list)
677 val = (unsigned long)list->next;
679 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
680 return RB_PAGE_MOVED;
682 return val & RB_FLAG_MASK;
688 * The unique thing about the reader page, is that, if the
689 * writer is ever on it, the previous pointer never points
690 * back to the reader page.
692 static int rb_is_reader_page(struct buffer_page *page)
694 struct list_head *list = page->list.prev;
696 return rb_list_head(list->next) != &page->list;
700 * rb_set_list_to_head - set a list_head to be pointing to head.
702 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
703 struct list_head *list)
707 ptr = (unsigned long *)&list->next;
708 *ptr |= RB_PAGE_HEAD;
709 *ptr &= ~RB_PAGE_UPDATE;
713 * rb_head_page_activate - sets up head page
715 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
717 struct buffer_page *head;
719 head = cpu_buffer->head_page;
724 * Set the previous list pointer to have the HEAD flag.
726 rb_set_list_to_head(cpu_buffer, head->list.prev);
729 static void rb_list_head_clear(struct list_head *list)
731 unsigned long *ptr = (unsigned long *)&list->next;
733 *ptr &= ~RB_FLAG_MASK;
737 * rb_head_page_dactivate - clears head page ptr (for free list)
740 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
742 struct list_head *hd;
744 /* Go through the whole list and clear any pointers found. */
745 rb_list_head_clear(cpu_buffer->pages);
747 list_for_each(hd, cpu_buffer->pages)
748 rb_list_head_clear(hd);
751 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
752 struct buffer_page *head,
753 struct buffer_page *prev,
754 int old_flag, int new_flag)
756 struct list_head *list;
757 unsigned long val = (unsigned long)&head->list;
762 val &= ~RB_FLAG_MASK;
764 ret = cmpxchg((unsigned long *)&list->next,
765 val | old_flag, val | new_flag);
767 /* check if the reader took the page */
768 if ((ret & ~RB_FLAG_MASK) != val)
769 return RB_PAGE_MOVED;
771 return ret & RB_FLAG_MASK;
774 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
775 struct buffer_page *head,
776 struct buffer_page *prev,
779 return rb_head_page_set(cpu_buffer, head, prev,
780 old_flag, RB_PAGE_UPDATE);
783 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
784 struct buffer_page *head,
785 struct buffer_page *prev,
788 return rb_head_page_set(cpu_buffer, head, prev,
789 old_flag, RB_PAGE_HEAD);
792 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
793 struct buffer_page *head,
794 struct buffer_page *prev,
797 return rb_head_page_set(cpu_buffer, head, prev,
798 old_flag, RB_PAGE_NORMAL);
801 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
802 struct buffer_page **bpage)
804 struct list_head *p = rb_list_head((*bpage)->list.next);
806 *bpage = list_entry(p, struct buffer_page, list);
809 static struct buffer_page *
810 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
812 struct buffer_page *head;
813 struct buffer_page *page;
814 struct list_head *list;
817 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
821 list = cpu_buffer->pages;
822 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
825 page = head = cpu_buffer->head_page;
827 * It is possible that the writer moves the header behind
828 * where we started, and we miss in one loop.
829 * A second loop should grab the header, but we'll do
830 * three loops just because I'm paranoid.
832 for (i = 0; i < 3; i++) {
834 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
835 cpu_buffer->head_page = page;
838 rb_inc_page(cpu_buffer, &page);
839 } while (page != head);
842 RB_WARN_ON(cpu_buffer, 1);
847 static int rb_head_page_replace(struct buffer_page *old,
848 struct buffer_page *new)
850 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
854 val = *ptr & ~RB_FLAG_MASK;
857 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
863 * rb_tail_page_update - move the tail page forward
865 * Returns 1 if moved tail page, 0 if someone else did.
867 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
868 struct buffer_page *tail_page,
869 struct buffer_page *next_page)
871 struct buffer_page *old_tail;
872 unsigned long old_entries;
873 unsigned long old_write;
877 * The tail page now needs to be moved forward.
879 * We need to reset the tail page, but without messing
880 * with possible erasing of data brought in by interrupts
881 * that have moved the tail page and are currently on it.
883 * We add a counter to the write field to denote this.
885 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
886 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
889 * Just make sure we have seen our old_write and synchronize
890 * with any interrupts that come in.
895 * If the tail page is still the same as what we think
896 * it is, then it is up to us to update the tail
899 if (tail_page == cpu_buffer->tail_page) {
900 /* Zero the write counter */
901 unsigned long val = old_write & ~RB_WRITE_MASK;
902 unsigned long eval = old_entries & ~RB_WRITE_MASK;
905 * This will only succeed if an interrupt did
906 * not come in and change it. In which case, we
907 * do not want to modify it.
909 * We add (void) to let the compiler know that we do not care
910 * about the return value of these functions. We use the
911 * cmpxchg to only update if an interrupt did not already
912 * do it for us. If the cmpxchg fails, we don't care.
914 (void)local_cmpxchg(&next_page->write, old_write, val);
915 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
918 * No need to worry about races with clearing out the commit.
919 * it only can increment when a commit takes place. But that
920 * only happens in the outer most nested commit.
922 local_set(&next_page->page->commit, 0);
924 old_tail = cmpxchg(&cpu_buffer->tail_page,
925 tail_page, next_page);
927 if (old_tail == tail_page)
934 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
935 struct buffer_page *bpage)
937 unsigned long val = (unsigned long)bpage;
939 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
946 * rb_check_list - make sure a pointer to a list has the last bits zero
948 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
949 struct list_head *list)
951 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
953 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
959 * check_pages - integrity check of buffer pages
960 * @cpu_buffer: CPU buffer with pages to test
962 * As a safety measure we check to make sure the data pages have not
965 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
967 struct list_head *head = cpu_buffer->pages;
968 struct buffer_page *bpage, *tmp;
970 rb_head_page_deactivate(cpu_buffer);
972 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
974 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
977 if (rb_check_list(cpu_buffer, head))
980 list_for_each_entry_safe(bpage, tmp, head, list) {
981 if (RB_WARN_ON(cpu_buffer,
982 bpage->list.next->prev != &bpage->list))
984 if (RB_WARN_ON(cpu_buffer,
985 bpage->list.prev->next != &bpage->list))
987 if (rb_check_list(cpu_buffer, &bpage->list))
991 rb_head_page_activate(cpu_buffer);
996 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
999 struct buffer_page *bpage, *tmp;
1005 for (i = 0; i < nr_pages; i++) {
1008 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1009 GFP_KERNEL, cpu_to_node(cpu_buffer->cpu));
1013 rb_check_bpage(cpu_buffer, bpage);
1015 list_add(&bpage->list, &pages);
1017 page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu),
1021 bpage->page = page_address(page);
1022 rb_init_page(bpage->page);
1026 * The ring buffer page list is a circular list that does not
1027 * start and end with a list head. All page list items point to
1030 cpu_buffer->pages = pages.next;
1033 rb_check_pages(cpu_buffer);
1038 list_for_each_entry_safe(bpage, tmp, &pages, list) {
1039 list_del_init(&bpage->list);
1040 free_buffer_page(bpage);
1045 static struct ring_buffer_per_cpu *
1046 rb_allocate_cpu_buffer(struct ring_buffer *buffer, int cpu)
1048 struct ring_buffer_per_cpu *cpu_buffer;
1049 struct buffer_page *bpage;
1053 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1054 GFP_KERNEL, cpu_to_node(cpu));
1058 cpu_buffer->cpu = cpu;
1059 cpu_buffer->buffer = buffer;
1060 spin_lock_init(&cpu_buffer->reader_lock);
1061 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1062 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1064 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1065 GFP_KERNEL, cpu_to_node(cpu));
1067 goto fail_free_buffer;
1069 rb_check_bpage(cpu_buffer, bpage);
1071 cpu_buffer->reader_page = bpage;
1072 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1074 goto fail_free_reader;
1075 bpage->page = page_address(page);
1076 rb_init_page(bpage->page);
1078 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1080 ret = rb_allocate_pages(cpu_buffer, buffer->pages);
1082 goto fail_free_reader;
1084 cpu_buffer->head_page
1085 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1086 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1088 rb_head_page_activate(cpu_buffer);
1093 free_buffer_page(cpu_buffer->reader_page);
1100 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1102 struct list_head *head = cpu_buffer->pages;
1103 struct buffer_page *bpage, *tmp;
1105 free_buffer_page(cpu_buffer->reader_page);
1107 rb_head_page_deactivate(cpu_buffer);
1110 list_for_each_entry_safe(bpage, tmp, head, list) {
1111 list_del_init(&bpage->list);
1112 free_buffer_page(bpage);
1114 bpage = list_entry(head, struct buffer_page, list);
1115 free_buffer_page(bpage);
1121 #ifdef CONFIG_HOTPLUG_CPU
1122 static int rb_cpu_notify(struct notifier_block *self,
1123 unsigned long action, void *hcpu);
1127 * ring_buffer_alloc - allocate a new ring_buffer
1128 * @size: the size in bytes per cpu that is needed.
1129 * @flags: attributes to set for the ring buffer.
1131 * Currently the only flag that is available is the RB_FL_OVERWRITE
1132 * flag. This flag means that the buffer will overwrite old data
1133 * when the buffer wraps. If this flag is not set, the buffer will
1134 * drop data when the tail hits the head.
1136 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1137 struct lock_class_key *key)
1139 struct ring_buffer *buffer;
1143 /* keep it in its own cache line */
1144 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1149 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1150 goto fail_free_buffer;
1152 buffer->pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1153 buffer->flags = flags;
1154 buffer->clock = trace_clock_local;
1155 buffer->reader_lock_key = key;
1157 /* need at least two pages */
1158 if (buffer->pages < 2)
1162 * In case of non-hotplug cpu, if the ring-buffer is allocated
1163 * in early initcall, it will not be notified of secondary cpus.
1164 * In that off case, we need to allocate for all possible cpus.
1166 #ifdef CONFIG_HOTPLUG_CPU
1168 cpumask_copy(buffer->cpumask, cpu_online_mask);
1170 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1172 buffer->cpus = nr_cpu_ids;
1174 bsize = sizeof(void *) * nr_cpu_ids;
1175 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1177 if (!buffer->buffers)
1178 goto fail_free_cpumask;
1180 for_each_buffer_cpu(buffer, cpu) {
1181 buffer->buffers[cpu] =
1182 rb_allocate_cpu_buffer(buffer, cpu);
1183 if (!buffer->buffers[cpu])
1184 goto fail_free_buffers;
1187 #ifdef CONFIG_HOTPLUG_CPU
1188 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1189 buffer->cpu_notify.priority = 0;
1190 register_cpu_notifier(&buffer->cpu_notify);
1194 mutex_init(&buffer->mutex);
1199 for_each_buffer_cpu(buffer, cpu) {
1200 if (buffer->buffers[cpu])
1201 rb_free_cpu_buffer(buffer->buffers[cpu]);
1203 kfree(buffer->buffers);
1206 free_cpumask_var(buffer->cpumask);
1213 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1216 * ring_buffer_free - free a ring buffer.
1217 * @buffer: the buffer to free.
1220 ring_buffer_free(struct ring_buffer *buffer)
1226 #ifdef CONFIG_HOTPLUG_CPU
1227 unregister_cpu_notifier(&buffer->cpu_notify);
1230 for_each_buffer_cpu(buffer, cpu)
1231 rb_free_cpu_buffer(buffer->buffers[cpu]);
1235 kfree(buffer->buffers);
1236 free_cpumask_var(buffer->cpumask);
1240 EXPORT_SYMBOL_GPL(ring_buffer_free);
1242 void ring_buffer_set_clock(struct ring_buffer *buffer,
1245 buffer->clock = clock;
1248 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1251 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned nr_pages)
1253 struct buffer_page *bpage;
1254 struct list_head *p;
1257 spin_lock_irq(&cpu_buffer->reader_lock);
1258 rb_head_page_deactivate(cpu_buffer);
1260 for (i = 0; i < nr_pages; i++) {
1261 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1263 p = cpu_buffer->pages->next;
1264 bpage = list_entry(p, struct buffer_page, list);
1265 list_del_init(&bpage->list);
1266 free_buffer_page(bpage);
1268 if (RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages)))
1271 rb_reset_cpu(cpu_buffer);
1272 rb_check_pages(cpu_buffer);
1275 spin_unlock_irq(&cpu_buffer->reader_lock);
1279 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer,
1280 struct list_head *pages, unsigned nr_pages)
1282 struct buffer_page *bpage;
1283 struct list_head *p;
1286 spin_lock_irq(&cpu_buffer->reader_lock);
1287 rb_head_page_deactivate(cpu_buffer);
1289 for (i = 0; i < nr_pages; i++) {
1290 if (RB_WARN_ON(cpu_buffer, list_empty(pages)))
1293 bpage = list_entry(p, struct buffer_page, list);
1294 list_del_init(&bpage->list);
1295 list_add_tail(&bpage->list, cpu_buffer->pages);
1297 rb_reset_cpu(cpu_buffer);
1298 rb_check_pages(cpu_buffer);
1301 spin_unlock_irq(&cpu_buffer->reader_lock);
1305 * ring_buffer_resize - resize the ring buffer
1306 * @buffer: the buffer to resize.
1307 * @size: the new size.
1309 * Minimum size is 2 * BUF_PAGE_SIZE.
1311 * Returns -1 on failure.
1313 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size)
1315 struct ring_buffer_per_cpu *cpu_buffer;
1316 unsigned nr_pages, rm_pages, new_pages;
1317 struct buffer_page *bpage, *tmp;
1318 unsigned long buffer_size;
1323 * Always succeed at resizing a non-existent buffer:
1328 size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1329 size *= BUF_PAGE_SIZE;
1330 buffer_size = buffer->pages * BUF_PAGE_SIZE;
1332 /* we need a minimum of two pages */
1333 if (size < BUF_PAGE_SIZE * 2)
1334 size = BUF_PAGE_SIZE * 2;
1336 if (size == buffer_size)
1339 atomic_inc(&buffer->record_disabled);
1341 /* Make sure all writers are done with this buffer. */
1342 synchronize_sched();
1344 mutex_lock(&buffer->mutex);
1347 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1349 if (size < buffer_size) {
1351 /* easy case, just free pages */
1352 if (RB_WARN_ON(buffer, nr_pages >= buffer->pages))
1355 rm_pages = buffer->pages - nr_pages;
1357 for_each_buffer_cpu(buffer, cpu) {
1358 cpu_buffer = buffer->buffers[cpu];
1359 rb_remove_pages(cpu_buffer, rm_pages);
1365 * This is a bit more difficult. We only want to add pages
1366 * when we can allocate enough for all CPUs. We do this
1367 * by allocating all the pages and storing them on a local
1368 * link list. If we succeed in our allocation, then we
1369 * add these pages to the cpu_buffers. Otherwise we just free
1370 * them all and return -ENOMEM;
1372 if (RB_WARN_ON(buffer, nr_pages <= buffer->pages))
1375 new_pages = nr_pages - buffer->pages;
1377 for_each_buffer_cpu(buffer, cpu) {
1378 for (i = 0; i < new_pages; i++) {
1380 bpage = kzalloc_node(ALIGN(sizeof(*bpage),
1382 GFP_KERNEL, cpu_to_node(cpu));
1385 list_add(&bpage->list, &pages);
1386 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1389 bpage->page = page_address(page);
1390 rb_init_page(bpage->page);
1394 for_each_buffer_cpu(buffer, cpu) {
1395 cpu_buffer = buffer->buffers[cpu];
1396 rb_insert_pages(cpu_buffer, &pages, new_pages);
1399 if (RB_WARN_ON(buffer, !list_empty(&pages)))
1403 buffer->pages = nr_pages;
1405 mutex_unlock(&buffer->mutex);
1407 atomic_dec(&buffer->record_disabled);
1412 list_for_each_entry_safe(bpage, tmp, &pages, list) {
1413 list_del_init(&bpage->list);
1414 free_buffer_page(bpage);
1417 mutex_unlock(&buffer->mutex);
1418 atomic_dec(&buffer->record_disabled);
1422 * Something went totally wrong, and we are too paranoid
1423 * to even clean up the mess.
1427 mutex_unlock(&buffer->mutex);
1428 atomic_dec(&buffer->record_disabled);
1431 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1433 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1435 mutex_lock(&buffer->mutex);
1437 buffer->flags |= RB_FL_OVERWRITE;
1439 buffer->flags &= ~RB_FL_OVERWRITE;
1440 mutex_unlock(&buffer->mutex);
1442 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1444 static inline void *
1445 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1447 return bpage->data + index;
1450 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1452 return bpage->page->data + index;
1455 static inline struct ring_buffer_event *
1456 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1458 return __rb_page_index(cpu_buffer->reader_page,
1459 cpu_buffer->reader_page->read);
1462 static inline struct ring_buffer_event *
1463 rb_iter_head_event(struct ring_buffer_iter *iter)
1465 return __rb_page_index(iter->head_page, iter->head);
1468 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1470 return local_read(&bpage->write) & RB_WRITE_MASK;
1473 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1475 return local_read(&bpage->page->commit);
1478 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1480 return local_read(&bpage->entries) & RB_WRITE_MASK;
1483 /* Size is determined by what has been committed */
1484 static inline unsigned rb_page_size(struct buffer_page *bpage)
1486 return rb_page_commit(bpage);
1489 static inline unsigned
1490 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1492 return rb_page_commit(cpu_buffer->commit_page);
1495 static inline unsigned
1496 rb_event_index(struct ring_buffer_event *event)
1498 unsigned long addr = (unsigned long)event;
1500 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1504 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1505 struct ring_buffer_event *event)
1507 unsigned long addr = (unsigned long)event;
1508 unsigned long index;
1510 index = rb_event_index(event);
1513 return cpu_buffer->commit_page->page == (void *)addr &&
1514 rb_commit_index(cpu_buffer) == index;
1518 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1520 unsigned long max_count;
1523 * We only race with interrupts and NMIs on this CPU.
1524 * If we own the commit event, then we can commit
1525 * all others that interrupted us, since the interruptions
1526 * are in stack format (they finish before they come
1527 * back to us). This allows us to do a simple loop to
1528 * assign the commit to the tail.
1531 max_count = cpu_buffer->buffer->pages * 100;
1533 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1534 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1536 if (RB_WARN_ON(cpu_buffer,
1537 rb_is_reader_page(cpu_buffer->tail_page)))
1539 local_set(&cpu_buffer->commit_page->page->commit,
1540 rb_page_write(cpu_buffer->commit_page));
1541 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1542 cpu_buffer->write_stamp =
1543 cpu_buffer->commit_page->page->time_stamp;
1544 /* add barrier to keep gcc from optimizing too much */
1547 while (rb_commit_index(cpu_buffer) !=
1548 rb_page_write(cpu_buffer->commit_page)) {
1550 local_set(&cpu_buffer->commit_page->page->commit,
1551 rb_page_write(cpu_buffer->commit_page));
1552 RB_WARN_ON(cpu_buffer,
1553 local_read(&cpu_buffer->commit_page->page->commit) &
1558 /* again, keep gcc from optimizing */
1562 * If an interrupt came in just after the first while loop
1563 * and pushed the tail page forward, we will be left with
1564 * a dangling commit that will never go forward.
1566 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1570 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1572 cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1573 cpu_buffer->reader_page->read = 0;
1576 static void rb_inc_iter(struct ring_buffer_iter *iter)
1578 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1581 * The iterator could be on the reader page (it starts there).
1582 * But the head could have moved, since the reader was
1583 * found. Check for this case and assign the iterator
1584 * to the head page instead of next.
1586 if (iter->head_page == cpu_buffer->reader_page)
1587 iter->head_page = rb_set_head_page(cpu_buffer);
1589 rb_inc_page(cpu_buffer, &iter->head_page);
1591 iter->read_stamp = iter->head_page->page->time_stamp;
1595 /* Slow path, do not inline */
1596 static noinline struct ring_buffer_event *
1597 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
1599 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
1601 /* Not the first event on the page? */
1602 if (rb_event_index(event)) {
1603 event->time_delta = delta & TS_MASK;
1604 event->array[0] = delta >> TS_SHIFT;
1606 /* nope, just zero it */
1607 event->time_delta = 0;
1608 event->array[0] = 0;
1611 return skip_time_extend(event);
1615 * ring_buffer_update_event - update event type and data
1616 * @event: the even to update
1617 * @type: the type of event
1618 * @length: the size of the event field in the ring buffer
1620 * Update the type and data fields of the event. The length
1621 * is the actual size that is written to the ring buffer,
1622 * and with this, we can determine what to place into the
1626 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
1627 struct ring_buffer_event *event, unsigned length,
1628 int add_timestamp, u64 delta)
1630 /* Only a commit updates the timestamp */
1631 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
1635 * If we need to add a timestamp, then we
1636 * add it to the start of the resevered space.
1638 if (unlikely(add_timestamp)) {
1639 event = rb_add_time_stamp(event, delta);
1640 length -= RB_LEN_TIME_EXTEND;
1644 event->time_delta = delta;
1645 length -= RB_EVNT_HDR_SIZE;
1646 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
1647 event->type_len = 0;
1648 event->array[0] = length;
1650 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
1654 * rb_handle_head_page - writer hit the head page
1656 * Returns: +1 to retry page
1661 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1662 struct buffer_page *tail_page,
1663 struct buffer_page *next_page)
1665 struct buffer_page *new_head;
1670 entries = rb_page_entries(next_page);
1673 * The hard part is here. We need to move the head
1674 * forward, and protect against both readers on
1675 * other CPUs and writers coming in via interrupts.
1677 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1681 * type can be one of four:
1682 * NORMAL - an interrupt already moved it for us
1683 * HEAD - we are the first to get here.
1684 * UPDATE - we are the interrupt interrupting
1686 * MOVED - a reader on another CPU moved the next
1687 * pointer to its reader page. Give up
1694 * We changed the head to UPDATE, thus
1695 * it is our responsibility to update
1698 local_add(entries, &cpu_buffer->overrun);
1701 * The entries will be zeroed out when we move the
1705 /* still more to do */
1708 case RB_PAGE_UPDATE:
1710 * This is an interrupt that interrupt the
1711 * previous update. Still more to do.
1714 case RB_PAGE_NORMAL:
1716 * An interrupt came in before the update
1717 * and processed this for us.
1718 * Nothing left to do.
1723 * The reader is on another CPU and just did
1724 * a swap with our next_page.
1729 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1734 * Now that we are here, the old head pointer is
1735 * set to UPDATE. This will keep the reader from
1736 * swapping the head page with the reader page.
1737 * The reader (on another CPU) will spin till
1740 * We just need to protect against interrupts
1741 * doing the job. We will set the next pointer
1742 * to HEAD. After that, we set the old pointer
1743 * to NORMAL, but only if it was HEAD before.
1744 * otherwise we are an interrupt, and only
1745 * want the outer most commit to reset it.
1747 new_head = next_page;
1748 rb_inc_page(cpu_buffer, &new_head);
1750 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1754 * Valid returns are:
1755 * HEAD - an interrupt came in and already set it.
1756 * NORMAL - One of two things:
1757 * 1) We really set it.
1758 * 2) A bunch of interrupts came in and moved
1759 * the page forward again.
1763 case RB_PAGE_NORMAL:
1767 RB_WARN_ON(cpu_buffer, 1);
1772 * It is possible that an interrupt came in,
1773 * set the head up, then more interrupts came in
1774 * and moved it again. When we get back here,
1775 * the page would have been set to NORMAL but we
1776 * just set it back to HEAD.
1778 * How do you detect this? Well, if that happened
1779 * the tail page would have moved.
1781 if (ret == RB_PAGE_NORMAL) {
1783 * If the tail had moved passed next, then we need
1784 * to reset the pointer.
1786 if (cpu_buffer->tail_page != tail_page &&
1787 cpu_buffer->tail_page != next_page)
1788 rb_head_page_set_normal(cpu_buffer, new_head,
1794 * If this was the outer most commit (the one that
1795 * changed the original pointer from HEAD to UPDATE),
1796 * then it is up to us to reset it to NORMAL.
1798 if (type == RB_PAGE_HEAD) {
1799 ret = rb_head_page_set_normal(cpu_buffer, next_page,
1802 if (RB_WARN_ON(cpu_buffer,
1803 ret != RB_PAGE_UPDATE))
1810 static unsigned rb_calculate_event_length(unsigned length)
1812 struct ring_buffer_event event; /* Used only for sizeof array */
1814 /* zero length can cause confusions */
1818 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
1819 length += sizeof(event.array[0]);
1821 length += RB_EVNT_HDR_SIZE;
1822 length = ALIGN(length, RB_ARCH_ALIGNMENT);
1828 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
1829 struct buffer_page *tail_page,
1830 unsigned long tail, unsigned long length)
1832 struct ring_buffer_event *event;
1835 * Only the event that crossed the page boundary
1836 * must fill the old tail_page with padding.
1838 if (tail >= BUF_PAGE_SIZE) {
1840 * If the page was filled, then we still need
1841 * to update the real_end. Reset it to zero
1842 * and the reader will ignore it.
1844 if (tail == BUF_PAGE_SIZE)
1845 tail_page->real_end = 0;
1847 local_sub(length, &tail_page->write);
1851 event = __rb_page_index(tail_page, tail);
1852 kmemcheck_annotate_bitfield(event, bitfield);
1855 * Save the original length to the meta data.
1856 * This will be used by the reader to add lost event
1859 tail_page->real_end = tail;
1862 * If this event is bigger than the minimum size, then
1863 * we need to be careful that we don't subtract the
1864 * write counter enough to allow another writer to slip
1866 * We put in a discarded commit instead, to make sure
1867 * that this space is not used again.
1869 * If we are less than the minimum size, we don't need to
1872 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
1873 /* No room for any events */
1875 /* Mark the rest of the page with padding */
1876 rb_event_set_padding(event);
1878 /* Set the write back to the previous setting */
1879 local_sub(length, &tail_page->write);
1883 /* Put in a discarded event */
1884 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
1885 event->type_len = RINGBUF_TYPE_PADDING;
1886 /* time delta must be non zero */
1887 event->time_delta = 1;
1889 /* Set write to end of buffer */
1890 length = (tail + length) - BUF_PAGE_SIZE;
1891 local_sub(length, &tail_page->write);
1895 * This is the slow path, force gcc not to inline it.
1897 static noinline struct ring_buffer_event *
1898 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
1899 unsigned long length, unsigned long tail,
1900 struct buffer_page *tail_page, u64 ts)
1902 struct buffer_page *commit_page = cpu_buffer->commit_page;
1903 struct ring_buffer *buffer = cpu_buffer->buffer;
1904 struct buffer_page *next_page;
1907 next_page = tail_page;
1909 rb_inc_page(cpu_buffer, &next_page);
1912 * If for some reason, we had an interrupt storm that made
1913 * it all the way around the buffer, bail, and warn
1916 if (unlikely(next_page == commit_page)) {
1917 local_inc(&cpu_buffer->commit_overrun);
1922 * This is where the fun begins!
1924 * We are fighting against races between a reader that
1925 * could be on another CPU trying to swap its reader
1926 * page with the buffer head.
1928 * We are also fighting against interrupts coming in and
1929 * moving the head or tail on us as well.
1931 * If the next page is the head page then we have filled
1932 * the buffer, unless the commit page is still on the
1935 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
1938 * If the commit is not on the reader page, then
1939 * move the header page.
1941 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
1943 * If we are not in overwrite mode,
1944 * this is easy, just stop here.
1946 if (!(buffer->flags & RB_FL_OVERWRITE))
1949 ret = rb_handle_head_page(cpu_buffer,
1958 * We need to be careful here too. The
1959 * commit page could still be on the reader
1960 * page. We could have a small buffer, and
1961 * have filled up the buffer with events
1962 * from interrupts and such, and wrapped.
1964 * Note, if the tail page is also the on the
1965 * reader_page, we let it move out.
1967 if (unlikely((cpu_buffer->commit_page !=
1968 cpu_buffer->tail_page) &&
1969 (cpu_buffer->commit_page ==
1970 cpu_buffer->reader_page))) {
1971 local_inc(&cpu_buffer->commit_overrun);
1977 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
1980 * Nested commits always have zero deltas, so
1981 * just reread the time stamp
1983 ts = rb_time_stamp(buffer);
1984 next_page->page->time_stamp = ts;
1989 rb_reset_tail(cpu_buffer, tail_page, tail, length);
1991 /* fail and let the caller try again */
1992 return ERR_PTR(-EAGAIN);
1996 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2001 static struct ring_buffer_event *
2002 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2003 unsigned long length, u64 ts,
2004 u64 delta, int add_timestamp)
2006 struct buffer_page *tail_page;
2007 struct ring_buffer_event *event;
2008 unsigned long tail, write;
2011 * If the time delta since the last event is too big to
2012 * hold in the time field of the event, then we append a
2013 * TIME EXTEND event ahead of the data event.
2015 if (unlikely(add_timestamp))
2016 length += RB_LEN_TIME_EXTEND;
2018 tail_page = cpu_buffer->tail_page;
2019 write = local_add_return(length, &tail_page->write);
2021 /* set write to only the index of the write */
2022 write &= RB_WRITE_MASK;
2023 tail = write - length;
2025 /* See if we shot pass the end of this buffer page */
2026 if (unlikely(write > BUF_PAGE_SIZE))
2027 return rb_move_tail(cpu_buffer, length, tail,
2030 /* We reserved something on the buffer */
2032 event = __rb_page_index(tail_page, tail);
2033 kmemcheck_annotate_bitfield(event, bitfield);
2034 rb_update_event(cpu_buffer, event, length, add_timestamp, delta);
2036 local_inc(&tail_page->entries);
2039 * If this is the first commit on the page, then update
2043 tail_page->page->time_stamp = ts;
2049 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2050 struct ring_buffer_event *event)
2052 unsigned long new_index, old_index;
2053 struct buffer_page *bpage;
2054 unsigned long index;
2057 new_index = rb_event_index(event);
2058 old_index = new_index + rb_event_ts_length(event);
2059 addr = (unsigned long)event;
2062 bpage = cpu_buffer->tail_page;
2064 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2065 unsigned long write_mask =
2066 local_read(&bpage->write) & ~RB_WRITE_MASK;
2068 * This is on the tail page. It is possible that
2069 * a write could come in and move the tail page
2070 * and write to the next page. That is fine
2071 * because we just shorten what is on this page.
2073 old_index += write_mask;
2074 new_index += write_mask;
2075 index = local_cmpxchg(&bpage->write, old_index, new_index);
2076 if (index == old_index)
2080 /* could not discard */
2084 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2086 local_inc(&cpu_buffer->committing);
2087 local_inc(&cpu_buffer->commits);
2090 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2092 unsigned long commits;
2094 if (RB_WARN_ON(cpu_buffer,
2095 !local_read(&cpu_buffer->committing)))
2099 commits = local_read(&cpu_buffer->commits);
2100 /* synchronize with interrupts */
2102 if (local_read(&cpu_buffer->committing) == 1)
2103 rb_set_commit_to_write(cpu_buffer);
2105 local_dec(&cpu_buffer->committing);
2107 /* synchronize with interrupts */
2111 * Need to account for interrupts coming in between the
2112 * updating of the commit page and the clearing of the
2113 * committing counter.
2115 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2116 !local_read(&cpu_buffer->committing)) {
2117 local_inc(&cpu_buffer->committing);
2122 static struct ring_buffer_event *
2123 rb_reserve_next_event(struct ring_buffer *buffer,
2124 struct ring_buffer_per_cpu *cpu_buffer,
2125 unsigned long length)
2127 struct ring_buffer_event *event;
2133 rb_start_commit(cpu_buffer);
2135 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2137 * Due to the ability to swap a cpu buffer from a buffer
2138 * it is possible it was swapped before we committed.
2139 * (committing stops a swap). We check for it here and
2140 * if it happened, we have to fail the write.
2143 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2144 local_dec(&cpu_buffer->committing);
2145 local_dec(&cpu_buffer->commits);
2150 length = rb_calculate_event_length(length);
2156 * We allow for interrupts to reenter here and do a trace.
2157 * If one does, it will cause this original code to loop
2158 * back here. Even with heavy interrupts happening, this
2159 * should only happen a few times in a row. If this happens
2160 * 1000 times in a row, there must be either an interrupt
2161 * storm or we have something buggy.
2164 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2167 ts = rb_time_stamp(cpu_buffer->buffer);
2168 diff = ts - cpu_buffer->write_stamp;
2170 /* make sure this diff is calculated here */
2173 /* Did the write stamp get updated already? */
2174 if (likely(ts >= cpu_buffer->write_stamp)) {
2176 if (unlikely(test_time_stamp(delta))) {
2177 int local_clock_stable = 1;
2178 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2179 local_clock_stable = sched_clock_stable;
2181 WARN_ONCE(delta > (1ULL << 59),
2182 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2183 (unsigned long long)delta,
2184 (unsigned long long)ts,
2185 (unsigned long long)cpu_buffer->write_stamp,
2186 local_clock_stable ? "" :
2187 "If you just came from a suspend/resume,\n"
2188 "please switch to the trace global clock:\n"
2189 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2194 event = __rb_reserve_next(cpu_buffer, length, ts,
2195 delta, add_timestamp);
2196 if (unlikely(PTR_ERR(event) == -EAGAIN))
2205 rb_end_commit(cpu_buffer);
2209 #ifdef CONFIG_TRACING
2211 #define TRACE_RECURSIVE_DEPTH 16
2213 /* Keep this code out of the fast path cache */
2214 static noinline void trace_recursive_fail(void)
2216 /* Disable all tracing before we do anything else */
2217 tracing_off_permanent();
2219 printk_once(KERN_WARNING "Tracing recursion: depth[%ld]:"
2220 "HC[%lu]:SC[%lu]:NMI[%lu]\n",
2221 trace_recursion_buffer(),
2222 hardirq_count() >> HARDIRQ_SHIFT,
2223 softirq_count() >> SOFTIRQ_SHIFT,
2229 static inline int trace_recursive_lock(void)
2231 trace_recursion_inc();
2233 if (likely(trace_recursion_buffer() < TRACE_RECURSIVE_DEPTH))
2236 trace_recursive_fail();
2241 static inline void trace_recursive_unlock(void)
2243 WARN_ON_ONCE(!trace_recursion_buffer());
2245 trace_recursion_dec();
2250 #define trace_recursive_lock() (0)
2251 #define trace_recursive_unlock() do { } while (0)
2256 * ring_buffer_lock_reserve - reserve a part of the buffer
2257 * @buffer: the ring buffer to reserve from
2258 * @length: the length of the data to reserve (excluding event header)
2260 * Returns a reseverd event on the ring buffer to copy directly to.
2261 * The user of this interface will need to get the body to write into
2262 * and can use the ring_buffer_event_data() interface.
2264 * The length is the length of the data needed, not the event length
2265 * which also includes the event header.
2267 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2268 * If NULL is returned, then nothing has been allocated or locked.
2270 struct ring_buffer_event *
2271 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2273 struct ring_buffer_per_cpu *cpu_buffer;
2274 struct ring_buffer_event *event;
2277 if (ring_buffer_flags != RB_BUFFERS_ON)
2280 /* If we are tracing schedule, we don't want to recurse */
2281 preempt_disable_notrace();
2283 if (atomic_read(&buffer->record_disabled))
2286 if (trace_recursive_lock())
2289 cpu = raw_smp_processor_id();
2291 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2294 cpu_buffer = buffer->buffers[cpu];
2296 if (atomic_read(&cpu_buffer->record_disabled))
2299 if (length > BUF_MAX_DATA_SIZE)
2302 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2309 trace_recursive_unlock();
2312 preempt_enable_notrace();
2315 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2318 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2319 struct ring_buffer_event *event)
2324 * The event first in the commit queue updates the
2327 if (rb_event_is_commit(cpu_buffer, event)) {
2329 * A commit event that is first on a page
2330 * updates the write timestamp with the page stamp
2332 if (!rb_event_index(event))
2333 cpu_buffer->write_stamp =
2334 cpu_buffer->commit_page->page->time_stamp;
2335 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2336 delta = event->array[0];
2338 delta += event->time_delta;
2339 cpu_buffer->write_stamp += delta;
2341 cpu_buffer->write_stamp += event->time_delta;
2345 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2346 struct ring_buffer_event *event)
2348 local_inc(&cpu_buffer->entries);
2349 rb_update_write_stamp(cpu_buffer, event);
2350 rb_end_commit(cpu_buffer);
2354 * ring_buffer_unlock_commit - commit a reserved
2355 * @buffer: The buffer to commit to
2356 * @event: The event pointer to commit.
2358 * This commits the data to the ring buffer, and releases any locks held.
2360 * Must be paired with ring_buffer_lock_reserve.
2362 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2363 struct ring_buffer_event *event)
2365 struct ring_buffer_per_cpu *cpu_buffer;
2366 int cpu = raw_smp_processor_id();
2368 cpu_buffer = buffer->buffers[cpu];
2370 rb_commit(cpu_buffer, event);
2372 trace_recursive_unlock();
2374 preempt_enable_notrace();
2378 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2380 static inline void rb_event_discard(struct ring_buffer_event *event)
2382 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2383 event = skip_time_extend(event);
2385 /* array[0] holds the actual length for the discarded event */
2386 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2387 event->type_len = RINGBUF_TYPE_PADDING;
2388 /* time delta must be non zero */
2389 if (!event->time_delta)
2390 event->time_delta = 1;
2394 * Decrement the entries to the page that an event is on.
2395 * The event does not even need to exist, only the pointer
2396 * to the page it is on. This may only be called before the commit
2400 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2401 struct ring_buffer_event *event)
2403 unsigned long addr = (unsigned long)event;
2404 struct buffer_page *bpage = cpu_buffer->commit_page;
2405 struct buffer_page *start;
2409 /* Do the likely case first */
2410 if (likely(bpage->page == (void *)addr)) {
2411 local_dec(&bpage->entries);
2416 * Because the commit page may be on the reader page we
2417 * start with the next page and check the end loop there.
2419 rb_inc_page(cpu_buffer, &bpage);
2422 if (bpage->page == (void *)addr) {
2423 local_dec(&bpage->entries);
2426 rb_inc_page(cpu_buffer, &bpage);
2427 } while (bpage != start);
2429 /* commit not part of this buffer?? */
2430 RB_WARN_ON(cpu_buffer, 1);
2434 * ring_buffer_commit_discard - discard an event that has not been committed
2435 * @buffer: the ring buffer
2436 * @event: non committed event to discard
2438 * Sometimes an event that is in the ring buffer needs to be ignored.
2439 * This function lets the user discard an event in the ring buffer
2440 * and then that event will not be read later.
2442 * This function only works if it is called before the the item has been
2443 * committed. It will try to free the event from the ring buffer
2444 * if another event has not been added behind it.
2446 * If another event has been added behind it, it will set the event
2447 * up as discarded, and perform the commit.
2449 * If this function is called, do not call ring_buffer_unlock_commit on
2452 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2453 struct ring_buffer_event *event)
2455 struct ring_buffer_per_cpu *cpu_buffer;
2458 /* The event is discarded regardless */
2459 rb_event_discard(event);
2461 cpu = smp_processor_id();
2462 cpu_buffer = buffer->buffers[cpu];
2465 * This must only be called if the event has not been
2466 * committed yet. Thus we can assume that preemption
2467 * is still disabled.
2469 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2471 rb_decrement_entry(cpu_buffer, event);
2472 if (rb_try_to_discard(cpu_buffer, event))
2476 * The commit is still visible by the reader, so we
2477 * must still update the timestamp.
2479 rb_update_write_stamp(cpu_buffer, event);
2481 rb_end_commit(cpu_buffer);
2483 trace_recursive_unlock();
2485 preempt_enable_notrace();
2488 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2491 * ring_buffer_write - write data to the buffer without reserving
2492 * @buffer: The ring buffer to write to.
2493 * @length: The length of the data being written (excluding the event header)
2494 * @data: The data to write to the buffer.
2496 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2497 * one function. If you already have the data to write to the buffer, it
2498 * may be easier to simply call this function.
2500 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2501 * and not the length of the event which would hold the header.
2503 int ring_buffer_write(struct ring_buffer *buffer,
2504 unsigned long length,
2507 struct ring_buffer_per_cpu *cpu_buffer;
2508 struct ring_buffer_event *event;
2513 if (ring_buffer_flags != RB_BUFFERS_ON)
2516 preempt_disable_notrace();
2518 if (atomic_read(&buffer->record_disabled))
2521 cpu = raw_smp_processor_id();
2523 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2526 cpu_buffer = buffer->buffers[cpu];
2528 if (atomic_read(&cpu_buffer->record_disabled))
2531 if (length > BUF_MAX_DATA_SIZE)
2534 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2538 body = rb_event_data(event);
2540 memcpy(body, data, length);
2542 rb_commit(cpu_buffer, event);
2546 preempt_enable_notrace();
2550 EXPORT_SYMBOL_GPL(ring_buffer_write);
2552 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
2554 struct buffer_page *reader = cpu_buffer->reader_page;
2555 struct buffer_page *head = rb_set_head_page(cpu_buffer);
2556 struct buffer_page *commit = cpu_buffer->commit_page;
2558 /* In case of error, head will be NULL */
2559 if (unlikely(!head))
2562 return reader->read == rb_page_commit(reader) &&
2563 (commit == reader ||
2565 head->read == rb_page_commit(commit)));
2569 * ring_buffer_record_disable - stop all writes into the buffer
2570 * @buffer: The ring buffer to stop writes to.
2572 * This prevents all writes to the buffer. Any attempt to write
2573 * to the buffer after this will fail and return NULL.
2575 * The caller should call synchronize_sched() after this.
2577 void ring_buffer_record_disable(struct ring_buffer *buffer)
2579 atomic_inc(&buffer->record_disabled);
2581 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
2584 * ring_buffer_record_enable - enable writes to the buffer
2585 * @buffer: The ring buffer to enable writes
2587 * Note, multiple disables will need the same number of enables
2588 * to truly enable the writing (much like preempt_disable).
2590 void ring_buffer_record_enable(struct ring_buffer *buffer)
2592 atomic_dec(&buffer->record_disabled);
2594 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
2597 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
2598 * @buffer: The ring buffer to stop writes to.
2599 * @cpu: The CPU buffer to stop
2601 * This prevents all writes to the buffer. Any attempt to write
2602 * to the buffer after this will fail and return NULL.
2604 * The caller should call synchronize_sched() after this.
2606 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
2608 struct ring_buffer_per_cpu *cpu_buffer;
2610 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2613 cpu_buffer = buffer->buffers[cpu];
2614 atomic_inc(&cpu_buffer->record_disabled);
2616 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
2619 * ring_buffer_record_enable_cpu - enable writes to the buffer
2620 * @buffer: The ring buffer to enable writes
2621 * @cpu: The CPU to enable.
2623 * Note, multiple disables will need the same number of enables
2624 * to truly enable the writing (much like preempt_disable).
2626 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
2628 struct ring_buffer_per_cpu *cpu_buffer;
2630 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2633 cpu_buffer = buffer->buffers[cpu];
2634 atomic_dec(&cpu_buffer->record_disabled);
2636 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
2639 * The total entries in the ring buffer is the running counter
2640 * of entries entered into the ring buffer, minus the sum of
2641 * the entries read from the ring buffer and the number of
2642 * entries that were overwritten.
2644 static inline unsigned long
2645 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
2647 return local_read(&cpu_buffer->entries) -
2648 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
2652 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
2653 * @buffer: The ring buffer
2654 * @cpu: The per CPU buffer to get the entries from.
2656 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
2658 struct ring_buffer_per_cpu *cpu_buffer;
2660 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2663 cpu_buffer = buffer->buffers[cpu];
2665 return rb_num_of_entries(cpu_buffer);
2667 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
2670 * ring_buffer_overrun_cpu - get the number of overruns in a cpu_buffer
2671 * @buffer: The ring buffer
2672 * @cpu: The per CPU buffer to get the number of overruns from
2674 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
2676 struct ring_buffer_per_cpu *cpu_buffer;
2679 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2682 cpu_buffer = buffer->buffers[cpu];
2683 ret = local_read(&cpu_buffer->overrun);
2687 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
2690 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by commits
2691 * @buffer: The ring buffer
2692 * @cpu: The per CPU buffer to get the number of overruns from
2695 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
2697 struct ring_buffer_per_cpu *cpu_buffer;
2700 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2703 cpu_buffer = buffer->buffers[cpu];
2704 ret = local_read(&cpu_buffer->commit_overrun);
2708 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
2711 * ring_buffer_entries - get the number of entries in a buffer
2712 * @buffer: The ring buffer
2714 * Returns the total number of entries in the ring buffer
2717 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
2719 struct ring_buffer_per_cpu *cpu_buffer;
2720 unsigned long entries = 0;
2723 /* if you care about this being correct, lock the buffer */
2724 for_each_buffer_cpu(buffer, cpu) {
2725 cpu_buffer = buffer->buffers[cpu];
2726 entries += rb_num_of_entries(cpu_buffer);
2731 EXPORT_SYMBOL_GPL(ring_buffer_entries);
2734 * ring_buffer_overruns - get the number of overruns in buffer
2735 * @buffer: The ring buffer
2737 * Returns the total number of overruns in the ring buffer
2740 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
2742 struct ring_buffer_per_cpu *cpu_buffer;
2743 unsigned long overruns = 0;
2746 /* if you care about this being correct, lock the buffer */
2747 for_each_buffer_cpu(buffer, cpu) {
2748 cpu_buffer = buffer->buffers[cpu];
2749 overruns += local_read(&cpu_buffer->overrun);
2754 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
2756 static void rb_iter_reset(struct ring_buffer_iter *iter)
2758 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2760 /* Iterator usage is expected to have record disabled */
2761 if (list_empty(&cpu_buffer->reader_page->list)) {
2762 iter->head_page = rb_set_head_page(cpu_buffer);
2763 if (unlikely(!iter->head_page))
2765 iter->head = iter->head_page->read;
2767 iter->head_page = cpu_buffer->reader_page;
2768 iter->head = cpu_buffer->reader_page->read;
2771 iter->read_stamp = cpu_buffer->read_stamp;
2773 iter->read_stamp = iter->head_page->page->time_stamp;
2774 iter->cache_reader_page = cpu_buffer->reader_page;
2775 iter->cache_read = cpu_buffer->read;
2779 * ring_buffer_iter_reset - reset an iterator
2780 * @iter: The iterator to reset
2782 * Resets the iterator, so that it will start from the beginning
2785 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
2787 struct ring_buffer_per_cpu *cpu_buffer;
2788 unsigned long flags;
2793 cpu_buffer = iter->cpu_buffer;
2795 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2796 rb_iter_reset(iter);
2797 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2799 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
2802 * ring_buffer_iter_empty - check if an iterator has no more to read
2803 * @iter: The iterator to check
2805 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
2807 struct ring_buffer_per_cpu *cpu_buffer;
2809 cpu_buffer = iter->cpu_buffer;
2811 return iter->head_page == cpu_buffer->commit_page &&
2812 iter->head == rb_commit_index(cpu_buffer);
2814 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
2817 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2818 struct ring_buffer_event *event)
2822 switch (event->type_len) {
2823 case RINGBUF_TYPE_PADDING:
2826 case RINGBUF_TYPE_TIME_EXTEND:
2827 delta = event->array[0];
2829 delta += event->time_delta;
2830 cpu_buffer->read_stamp += delta;
2833 case RINGBUF_TYPE_TIME_STAMP:
2834 /* FIXME: not implemented */
2837 case RINGBUF_TYPE_DATA:
2838 cpu_buffer->read_stamp += event->time_delta;
2848 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
2849 struct ring_buffer_event *event)
2853 switch (event->type_len) {
2854 case RINGBUF_TYPE_PADDING:
2857 case RINGBUF_TYPE_TIME_EXTEND:
2858 delta = event->array[0];
2860 delta += event->time_delta;
2861 iter->read_stamp += delta;
2864 case RINGBUF_TYPE_TIME_STAMP:
2865 /* FIXME: not implemented */
2868 case RINGBUF_TYPE_DATA:
2869 iter->read_stamp += event->time_delta;
2878 static struct buffer_page *
2879 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
2881 struct buffer_page *reader = NULL;
2882 unsigned long overwrite;
2883 unsigned long flags;
2887 local_irq_save(flags);
2888 arch_spin_lock(&cpu_buffer->lock);
2892 * This should normally only loop twice. But because the
2893 * start of the reader inserts an empty page, it causes
2894 * a case where we will loop three times. There should be no
2895 * reason to loop four times (that I know of).
2897 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
2902 reader = cpu_buffer->reader_page;
2904 /* If there's more to read, return this page */
2905 if (cpu_buffer->reader_page->read < rb_page_size(reader))
2908 /* Never should we have an index greater than the size */
2909 if (RB_WARN_ON(cpu_buffer,
2910 cpu_buffer->reader_page->read > rb_page_size(reader)))
2913 /* check if we caught up to the tail */
2915 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
2919 * Reset the reader page to size zero.
2921 local_set(&cpu_buffer->reader_page->write, 0);
2922 local_set(&cpu_buffer->reader_page->entries, 0);
2923 local_set(&cpu_buffer->reader_page->page->commit, 0);
2924 cpu_buffer->reader_page->real_end = 0;
2928 * Splice the empty reader page into the list around the head.
2930 reader = rb_set_head_page(cpu_buffer);
2931 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
2932 cpu_buffer->reader_page->list.prev = reader->list.prev;
2935 * cpu_buffer->pages just needs to point to the buffer, it
2936 * has no specific buffer page to point to. Lets move it out
2937 * of our way so we don't accidentally swap it.
2939 cpu_buffer->pages = reader->list.prev;
2941 /* The reader page will be pointing to the new head */
2942 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
2945 * We want to make sure we read the overruns after we set up our
2946 * pointers to the next object. The writer side does a
2947 * cmpxchg to cross pages which acts as the mb on the writer
2948 * side. Note, the reader will constantly fail the swap
2949 * while the writer is updating the pointers, so this
2950 * guarantees that the overwrite recorded here is the one we
2951 * want to compare with the last_overrun.
2954 overwrite = local_read(&(cpu_buffer->overrun));
2957 * Here's the tricky part.
2959 * We need to move the pointer past the header page.
2960 * But we can only do that if a writer is not currently
2961 * moving it. The page before the header page has the
2962 * flag bit '1' set if it is pointing to the page we want.
2963 * but if the writer is in the process of moving it
2964 * than it will be '2' or already moved '0'.
2967 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
2970 * If we did not convert it, then we must try again.
2976 * Yeah! We succeeded in replacing the page.
2978 * Now make the new head point back to the reader page.
2980 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
2981 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
2983 /* Finally update the reader page to the new head */
2984 cpu_buffer->reader_page = reader;
2985 rb_reset_reader_page(cpu_buffer);
2987 if (overwrite != cpu_buffer->last_overrun) {
2988 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
2989 cpu_buffer->last_overrun = overwrite;
2995 arch_spin_unlock(&cpu_buffer->lock);
2996 local_irq_restore(flags);
3001 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3003 struct ring_buffer_event *event;
3004 struct buffer_page *reader;
3007 reader = rb_get_reader_page(cpu_buffer);
3009 /* This function should not be called when buffer is empty */
3010 if (RB_WARN_ON(cpu_buffer, !reader))
3013 event = rb_reader_event(cpu_buffer);
3015 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3018 rb_update_read_stamp(cpu_buffer, event);
3020 length = rb_event_length(event);
3021 cpu_buffer->reader_page->read += length;
3024 static void rb_advance_iter(struct ring_buffer_iter *iter)
3026 struct ring_buffer_per_cpu *cpu_buffer;
3027 struct ring_buffer_event *event;
3030 cpu_buffer = iter->cpu_buffer;
3033 * Check if we are at the end of the buffer.
3035 if (iter->head >= rb_page_size(iter->head_page)) {
3036 /* discarded commits can make the page empty */
3037 if (iter->head_page == cpu_buffer->commit_page)
3043 event = rb_iter_head_event(iter);
3045 length = rb_event_length(event);
3048 * This should not be called to advance the header if we are
3049 * at the tail of the buffer.
3051 if (RB_WARN_ON(cpu_buffer,
3052 (iter->head_page == cpu_buffer->commit_page) &&
3053 (iter->head + length > rb_commit_index(cpu_buffer))))
3056 rb_update_iter_read_stamp(iter, event);
3058 iter->head += length;
3060 /* check for end of page padding */
3061 if ((iter->head >= rb_page_size(iter->head_page)) &&
3062 (iter->head_page != cpu_buffer->commit_page))
3063 rb_advance_iter(iter);
3066 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3068 return cpu_buffer->lost_events;
3071 static struct ring_buffer_event *
3072 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3073 unsigned long *lost_events)
3075 struct ring_buffer_event *event;
3076 struct buffer_page *reader;
3081 * We repeat when a time extend is encountered.
3082 * Since the time extend is always attached to a data event,
3083 * we should never loop more than once.
3084 * (We never hit the following condition more than twice).
3086 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3089 reader = rb_get_reader_page(cpu_buffer);
3093 event = rb_reader_event(cpu_buffer);
3095 switch (event->type_len) {
3096 case RINGBUF_TYPE_PADDING:
3097 if (rb_null_event(event))
3098 RB_WARN_ON(cpu_buffer, 1);
3100 * Because the writer could be discarding every
3101 * event it creates (which would probably be bad)
3102 * if we were to go back to "again" then we may never
3103 * catch up, and will trigger the warn on, or lock
3104 * the box. Return the padding, and we will release
3105 * the current locks, and try again.
3109 case RINGBUF_TYPE_TIME_EXTEND:
3110 /* Internal data, OK to advance */
3111 rb_advance_reader(cpu_buffer);
3114 case RINGBUF_TYPE_TIME_STAMP:
3115 /* FIXME: not implemented */
3116 rb_advance_reader(cpu_buffer);
3119 case RINGBUF_TYPE_DATA:
3121 *ts = cpu_buffer->read_stamp + event->time_delta;
3122 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3123 cpu_buffer->cpu, ts);
3126 *lost_events = rb_lost_events(cpu_buffer);
3135 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3137 static struct ring_buffer_event *
3138 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3140 struct ring_buffer *buffer;
3141 struct ring_buffer_per_cpu *cpu_buffer;
3142 struct ring_buffer_event *event;
3145 cpu_buffer = iter->cpu_buffer;
3146 buffer = cpu_buffer->buffer;
3149 * Check if someone performed a consuming read to
3150 * the buffer. A consuming read invalidates the iterator
3151 * and we need to reset the iterator in this case.
3153 if (unlikely(iter->cache_read != cpu_buffer->read ||
3154 iter->cache_reader_page != cpu_buffer->reader_page))
3155 rb_iter_reset(iter);
3158 if (ring_buffer_iter_empty(iter))
3162 * We repeat when a time extend is encountered.
3163 * Since the time extend is always attached to a data event,
3164 * we should never loop more than once.
3165 * (We never hit the following condition more than twice).
3167 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3170 if (rb_per_cpu_empty(cpu_buffer))
3173 if (iter->head >= local_read(&iter->head_page->page->commit)) {
3178 event = rb_iter_head_event(iter);
3180 switch (event->type_len) {
3181 case RINGBUF_TYPE_PADDING:
3182 if (rb_null_event(event)) {
3186 rb_advance_iter(iter);
3189 case RINGBUF_TYPE_TIME_EXTEND:
3190 /* Internal data, OK to advance */
3191 rb_advance_iter(iter);
3194 case RINGBUF_TYPE_TIME_STAMP:
3195 /* FIXME: not implemented */
3196 rb_advance_iter(iter);
3199 case RINGBUF_TYPE_DATA:
3201 *ts = iter->read_stamp + event->time_delta;
3202 ring_buffer_normalize_time_stamp(buffer,
3203 cpu_buffer->cpu, ts);
3213 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3215 static inline int rb_ok_to_lock(void)
3218 * If an NMI die dumps out the content of the ring buffer
3219 * do not grab locks. We also permanently disable the ring
3220 * buffer too. A one time deal is all you get from reading
3221 * the ring buffer from an NMI.
3223 if (likely(!in_nmi()))
3226 tracing_off_permanent();
3231 * ring_buffer_peek - peek at the next event to be read
3232 * @buffer: The ring buffer to read
3233 * @cpu: The cpu to peak at
3234 * @ts: The timestamp counter of this event.
3235 * @lost_events: a variable to store if events were lost (may be NULL)
3237 * This will return the event that will be read next, but does
3238 * not consume the data.
3240 struct ring_buffer_event *
3241 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3242 unsigned long *lost_events)
3244 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3245 struct ring_buffer_event *event;
3246 unsigned long flags;
3249 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3252 dolock = rb_ok_to_lock();
3254 local_irq_save(flags);
3256 spin_lock(&cpu_buffer->reader_lock);
3257 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3258 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3259 rb_advance_reader(cpu_buffer);
3261 spin_unlock(&cpu_buffer->reader_lock);
3262 local_irq_restore(flags);
3264 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3271 * ring_buffer_iter_peek - peek at the next event to be read
3272 * @iter: The ring buffer iterator
3273 * @ts: The timestamp counter of this event.
3275 * This will return the event that will be read next, but does
3276 * not increment the iterator.
3278 struct ring_buffer_event *
3279 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3281 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3282 struct ring_buffer_event *event;
3283 unsigned long flags;
3286 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3287 event = rb_iter_peek(iter, ts);
3288 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3290 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3297 * ring_buffer_consume - return an event and consume it
3298 * @buffer: The ring buffer to get the next event from
3299 * @cpu: the cpu to read the buffer from
3300 * @ts: a variable to store the timestamp (may be NULL)
3301 * @lost_events: a variable to store if events were lost (may be NULL)
3303 * Returns the next event in the ring buffer, and that event is consumed.
3304 * Meaning, that sequential reads will keep returning a different event,
3305 * and eventually empty the ring buffer if the producer is slower.
3307 struct ring_buffer_event *
3308 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3309 unsigned long *lost_events)
3311 struct ring_buffer_per_cpu *cpu_buffer;
3312 struct ring_buffer_event *event = NULL;
3313 unsigned long flags;
3316 dolock = rb_ok_to_lock();
3319 /* might be called in atomic */
3322 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3325 cpu_buffer = buffer->buffers[cpu];
3326 local_irq_save(flags);
3328 spin_lock(&cpu_buffer->reader_lock);
3330 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3332 cpu_buffer->lost_events = 0;
3333 rb_advance_reader(cpu_buffer);
3337 spin_unlock(&cpu_buffer->reader_lock);
3338 local_irq_restore(flags);
3343 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3348 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3351 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3352 * @buffer: The ring buffer to read from
3353 * @cpu: The cpu buffer to iterate over
3355 * This performs the initial preparations necessary to iterate
3356 * through the buffer. Memory is allocated, buffer recording
3357 * is disabled, and the iterator pointer is returned to the caller.
3359 * Disabling buffer recordng prevents the reading from being
3360 * corrupted. This is not a consuming read, so a producer is not
3363 * After a sequence of ring_buffer_read_prepare calls, the user is
3364 * expected to make at least one call to ring_buffer_prepare_sync.
3365 * Afterwards, ring_buffer_read_start is invoked to get things going
3368 * This overall must be paired with ring_buffer_finish.
3370 struct ring_buffer_iter *
3371 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3373 struct ring_buffer_per_cpu *cpu_buffer;
3374 struct ring_buffer_iter *iter;
3376 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3379 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3383 cpu_buffer = buffer->buffers[cpu];
3385 iter->cpu_buffer = cpu_buffer;
3387 atomic_inc(&cpu_buffer->record_disabled);
3391 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
3394 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
3396 * All previously invoked ring_buffer_read_prepare calls to prepare
3397 * iterators will be synchronized. Afterwards, read_buffer_read_start
3398 * calls on those iterators are allowed.
3401 ring_buffer_read_prepare_sync(void)
3403 synchronize_sched();
3405 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
3408 * ring_buffer_read_start - start a non consuming read of the buffer
3409 * @iter: The iterator returned by ring_buffer_read_prepare
3411 * This finalizes the startup of an iteration through the buffer.
3412 * The iterator comes from a call to ring_buffer_read_prepare and
3413 * an intervening ring_buffer_read_prepare_sync must have been
3416 * Must be paired with ring_buffer_finish.
3419 ring_buffer_read_start(struct ring_buffer_iter *iter)
3421 struct ring_buffer_per_cpu *cpu_buffer;
3422 unsigned long flags;
3427 cpu_buffer = iter->cpu_buffer;
3429 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3430 arch_spin_lock(&cpu_buffer->lock);
3431 rb_iter_reset(iter);
3432 arch_spin_unlock(&cpu_buffer->lock);
3433 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3435 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
3438 * ring_buffer_finish - finish reading the iterator of the buffer
3439 * @iter: The iterator retrieved by ring_buffer_start
3441 * This re-enables the recording to the buffer, and frees the
3445 ring_buffer_read_finish(struct ring_buffer_iter *iter)
3447 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3449 atomic_dec(&cpu_buffer->record_disabled);
3452 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
3455 * ring_buffer_read - read the next item in the ring buffer by the iterator
3456 * @iter: The ring buffer iterator
3457 * @ts: The time stamp of the event read.
3459 * This reads the next event in the ring buffer and increments the iterator.
3461 struct ring_buffer_event *
3462 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
3464 struct ring_buffer_event *event;
3465 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3466 unsigned long flags;
3468 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3470 event = rb_iter_peek(iter, ts);
3474 if (event->type_len == RINGBUF_TYPE_PADDING)
3477 rb_advance_iter(iter);
3479 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3483 EXPORT_SYMBOL_GPL(ring_buffer_read);
3486 * ring_buffer_size - return the size of the ring buffer (in bytes)
3487 * @buffer: The ring buffer.
3489 unsigned long ring_buffer_size(struct ring_buffer *buffer)
3491 return BUF_PAGE_SIZE * buffer->pages;
3493 EXPORT_SYMBOL_GPL(ring_buffer_size);
3496 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
3498 rb_head_page_deactivate(cpu_buffer);
3500 cpu_buffer->head_page
3501 = list_entry(cpu_buffer->pages, struct buffer_page, list);
3502 local_set(&cpu_buffer->head_page->write, 0);
3503 local_set(&cpu_buffer->head_page->entries, 0);
3504 local_set(&cpu_buffer->head_page->page->commit, 0);
3506 cpu_buffer->head_page->read = 0;
3508 cpu_buffer->tail_page = cpu_buffer->head_page;
3509 cpu_buffer->commit_page = cpu_buffer->head_page;
3511 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
3512 local_set(&cpu_buffer->reader_page->write, 0);
3513 local_set(&cpu_buffer->reader_page->entries, 0);
3514 local_set(&cpu_buffer->reader_page->page->commit, 0);
3515 cpu_buffer->reader_page->read = 0;
3517 local_set(&cpu_buffer->commit_overrun, 0);
3518 local_set(&cpu_buffer->overrun, 0);
3519 local_set(&cpu_buffer->entries, 0);
3520 local_set(&cpu_buffer->committing, 0);
3521 local_set(&cpu_buffer->commits, 0);
3522 cpu_buffer->read = 0;
3524 cpu_buffer->write_stamp = 0;
3525 cpu_buffer->read_stamp = 0;
3527 cpu_buffer->lost_events = 0;
3528 cpu_buffer->last_overrun = 0;
3530 rb_head_page_activate(cpu_buffer);
3534 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
3535 * @buffer: The ring buffer to reset a per cpu buffer of
3536 * @cpu: The CPU buffer to be reset
3538 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
3540 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3541 unsigned long flags;
3543 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3546 atomic_inc(&cpu_buffer->record_disabled);
3548 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3550 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
3553 arch_spin_lock(&cpu_buffer->lock);
3555 rb_reset_cpu(cpu_buffer);
3557 arch_spin_unlock(&cpu_buffer->lock);
3560 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3562 atomic_dec(&cpu_buffer->record_disabled);
3564 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
3567 * ring_buffer_reset - reset a ring buffer
3568 * @buffer: The ring buffer to reset all cpu buffers
3570 void ring_buffer_reset(struct ring_buffer *buffer)
3574 for_each_buffer_cpu(buffer, cpu)
3575 ring_buffer_reset_cpu(buffer, cpu);
3577 EXPORT_SYMBOL_GPL(ring_buffer_reset);
3580 * rind_buffer_empty - is the ring buffer empty?
3581 * @buffer: The ring buffer to test
3583 int ring_buffer_empty(struct ring_buffer *buffer)
3585 struct ring_buffer_per_cpu *cpu_buffer;
3586 unsigned long flags;
3591 dolock = rb_ok_to_lock();
3593 /* yes this is racy, but if you don't like the race, lock the buffer */
3594 for_each_buffer_cpu(buffer, cpu) {
3595 cpu_buffer = buffer->buffers[cpu];
3596 local_irq_save(flags);
3598 spin_lock(&cpu_buffer->reader_lock);
3599 ret = rb_per_cpu_empty(cpu_buffer);
3601 spin_unlock(&cpu_buffer->reader_lock);
3602 local_irq_restore(flags);
3610 EXPORT_SYMBOL_GPL(ring_buffer_empty);
3613 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
3614 * @buffer: The ring buffer
3615 * @cpu: The CPU buffer to test
3617 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
3619 struct ring_buffer_per_cpu *cpu_buffer;
3620 unsigned long flags;
3624 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3627 dolock = rb_ok_to_lock();
3629 cpu_buffer = buffer->buffers[cpu];
3630 local_irq_save(flags);
3632 spin_lock(&cpu_buffer->reader_lock);
3633 ret = rb_per_cpu_empty(cpu_buffer);
3635 spin_unlock(&cpu_buffer->reader_lock);
3636 local_irq_restore(flags);
3640 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
3642 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3644 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
3645 * @buffer_a: One buffer to swap with
3646 * @buffer_b: The other buffer to swap with
3648 * This function is useful for tracers that want to take a "snapshot"
3649 * of a CPU buffer and has another back up buffer lying around.
3650 * it is expected that the tracer handles the cpu buffer not being
3651 * used at the moment.
3653 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
3654 struct ring_buffer *buffer_b, int cpu)
3656 struct ring_buffer_per_cpu *cpu_buffer_a;
3657 struct ring_buffer_per_cpu *cpu_buffer_b;
3660 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
3661 !cpumask_test_cpu(cpu, buffer_b->cpumask))
3664 /* At least make sure the two buffers are somewhat the same */
3665 if (buffer_a->pages != buffer_b->pages)
3670 if (ring_buffer_flags != RB_BUFFERS_ON)
3673 if (atomic_read(&buffer_a->record_disabled))
3676 if (atomic_read(&buffer_b->record_disabled))
3679 cpu_buffer_a = buffer_a->buffers[cpu];
3680 cpu_buffer_b = buffer_b->buffers[cpu];
3682 if (atomic_read(&cpu_buffer_a->record_disabled))
3685 if (atomic_read(&cpu_buffer_b->record_disabled))
3689 * We can't do a synchronize_sched here because this
3690 * function can be called in atomic context.
3691 * Normally this will be called from the same CPU as cpu.
3692 * If not it's up to the caller to protect this.
3694 atomic_inc(&cpu_buffer_a->record_disabled);
3695 atomic_inc(&cpu_buffer_b->record_disabled);
3698 if (local_read(&cpu_buffer_a->committing))
3700 if (local_read(&cpu_buffer_b->committing))
3703 buffer_a->buffers[cpu] = cpu_buffer_b;
3704 buffer_b->buffers[cpu] = cpu_buffer_a;
3706 cpu_buffer_b->buffer = buffer_a;
3707 cpu_buffer_a->buffer = buffer_b;
3712 atomic_dec(&cpu_buffer_a->record_disabled);
3713 atomic_dec(&cpu_buffer_b->record_disabled);
3717 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
3718 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
3721 * ring_buffer_alloc_read_page - allocate a page to read from buffer
3722 * @buffer: the buffer to allocate for.
3724 * This function is used in conjunction with ring_buffer_read_page.
3725 * When reading a full page from the ring buffer, these functions
3726 * can be used to speed up the process. The calling function should
3727 * allocate a few pages first with this function. Then when it
3728 * needs to get pages from the ring buffer, it passes the result
3729 * of this function into ring_buffer_read_page, which will swap
3730 * the page that was allocated, with the read page of the buffer.
3733 * The page allocated, or NULL on error.
3735 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
3737 struct buffer_data_page *bpage;
3740 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
3744 bpage = page_address(page);
3746 rb_init_page(bpage);
3750 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
3753 * ring_buffer_free_read_page - free an allocated read page
3754 * @buffer: the buffer the page was allocate for
3755 * @data: the page to free
3757 * Free a page allocated from ring_buffer_alloc_read_page.
3759 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
3761 free_page((unsigned long)data);
3763 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
3766 * ring_buffer_read_page - extract a page from the ring buffer
3767 * @buffer: buffer to extract from
3768 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
3769 * @len: amount to extract
3770 * @cpu: the cpu of the buffer to extract
3771 * @full: should the extraction only happen when the page is full.
3773 * This function will pull out a page from the ring buffer and consume it.
3774 * @data_page must be the address of the variable that was returned
3775 * from ring_buffer_alloc_read_page. This is because the page might be used
3776 * to swap with a page in the ring buffer.
3779 * rpage = ring_buffer_alloc_read_page(buffer);
3782 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
3784 * process_page(rpage, ret);
3786 * When @full is set, the function will not return true unless
3787 * the writer is off the reader page.
3789 * Note: it is up to the calling functions to handle sleeps and wakeups.
3790 * The ring buffer can be used anywhere in the kernel and can not
3791 * blindly call wake_up. The layer that uses the ring buffer must be
3792 * responsible for that.
3795 * >=0 if data has been transferred, returns the offset of consumed data.
3796 * <0 if no data has been transferred.
3798 int ring_buffer_read_page(struct ring_buffer *buffer,
3799 void **data_page, size_t len, int cpu, int full)
3801 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3802 struct ring_buffer_event *event;
3803 struct buffer_data_page *bpage;
3804 struct buffer_page *reader;
3805 unsigned long missed_events;
3806 unsigned long flags;
3807 unsigned int commit;
3812 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3816 * If len is not big enough to hold the page header, then
3817 * we can not copy anything.
3819 if (len <= BUF_PAGE_HDR_SIZE)
3822 len -= BUF_PAGE_HDR_SIZE;
3831 spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3833 reader = rb_get_reader_page(cpu_buffer);
3837 event = rb_reader_event(cpu_buffer);
3839 read = reader->read;
3840 commit = rb_page_commit(reader);
3842 /* Check if any events were dropped */
3843 missed_events = cpu_buffer->lost_events;
3846 * If this page has been partially read or
3847 * if len is not big enough to read the rest of the page or
3848 * a writer is still on the page, then
3849 * we must copy the data from the page to the buffer.
3850 * Otherwise, we can simply swap the page with the one passed in.
3852 if (read || (len < (commit - read)) ||
3853 cpu_buffer->reader_page == cpu_buffer->commit_page) {
3854 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
3855 unsigned int rpos = read;
3856 unsigned int pos = 0;
3862 if (len > (commit - read))
3863 len = (commit - read);
3865 /* Always keep the time extend and data together */
3866 size = rb_event_ts_length(event);
3871 /* save the current timestamp, since the user will need it */
3872 save_timestamp = cpu_buffer->read_stamp;
3874 /* Need to copy one event at a time */
3876 /* We need the size of one event, because
3877 * rb_advance_reader only advances by one event,
3878 * whereas rb_event_ts_length may include the size of
3879 * one or two events.
3880 * We have already ensured there's enough space if this
3881 * is a time extend. */
3882 size = rb_event_length(event);
3883 memcpy(bpage->data + pos, rpage->data + rpos, size);
3887 rb_advance_reader(cpu_buffer);
3888 rpos = reader->read;
3894 event = rb_reader_event(cpu_buffer);
3895 /* Always keep the time extend and data together */
3896 size = rb_event_ts_length(event);
3897 } while (len >= size);
3900 local_set(&bpage->commit, pos);
3901 bpage->time_stamp = save_timestamp;
3903 /* we copied everything to the beginning */
3906 /* update the entry counter */
3907 cpu_buffer->read += rb_page_entries(reader);
3909 /* swap the pages */
3910 rb_init_page(bpage);
3911 bpage = reader->page;
3912 reader->page = *data_page;
3913 local_set(&reader->write, 0);
3914 local_set(&reader->entries, 0);
3919 * Use the real_end for the data size,
3920 * This gives us a chance to store the lost events
3923 if (reader->real_end)
3924 local_set(&bpage->commit, reader->real_end);
3928 cpu_buffer->lost_events = 0;
3930 commit = local_read(&bpage->commit);
3932 * Set a flag in the commit field if we lost events
3934 if (missed_events) {
3935 /* If there is room at the end of the page to save the
3936 * missed events, then record it there.
3938 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
3939 memcpy(&bpage->data[commit], &missed_events,
3940 sizeof(missed_events));
3941 local_add(RB_MISSED_STORED, &bpage->commit);
3942 commit += sizeof(missed_events);
3944 local_add(RB_MISSED_EVENTS, &bpage->commit);
3948 * This page may be off to user land. Zero it out here.
3950 if (commit < BUF_PAGE_SIZE)
3951 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
3954 spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3959 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
3961 #ifdef CONFIG_TRACING
3963 rb_simple_read(struct file *filp, char __user *ubuf,
3964 size_t cnt, loff_t *ppos)
3966 unsigned long *p = filp->private_data;
3970 if (test_bit(RB_BUFFERS_DISABLED_BIT, p))
3971 r = sprintf(buf, "permanently disabled\n");
3973 r = sprintf(buf, "%d\n", test_bit(RB_BUFFERS_ON_BIT, p));
3975 return simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
3979 rb_simple_write(struct file *filp, const char __user *ubuf,
3980 size_t cnt, loff_t *ppos)
3982 unsigned long *p = filp->private_data;
3987 if (cnt >= sizeof(buf))
3990 if (copy_from_user(&buf, ubuf, cnt))
3995 ret = strict_strtoul(buf, 10, &val);
4000 set_bit(RB_BUFFERS_ON_BIT, p);
4002 clear_bit(RB_BUFFERS_ON_BIT, p);
4009 static const struct file_operations rb_simple_fops = {
4010 .open = tracing_open_generic,
4011 .read = rb_simple_read,
4012 .write = rb_simple_write,
4013 .llseek = default_llseek,
4017 static __init int rb_init_debugfs(void)
4019 struct dentry *d_tracer;
4021 d_tracer = tracing_init_dentry();
4023 trace_create_file("tracing_on", 0644, d_tracer,
4024 &ring_buffer_flags, &rb_simple_fops);
4029 fs_initcall(rb_init_debugfs);
4032 #ifdef CONFIG_HOTPLUG_CPU
4033 static int rb_cpu_notify(struct notifier_block *self,
4034 unsigned long action, void *hcpu)
4036 struct ring_buffer *buffer =
4037 container_of(self, struct ring_buffer, cpu_notify);
4038 long cpu = (long)hcpu;
4041 case CPU_UP_PREPARE:
4042 case CPU_UP_PREPARE_FROZEN:
4043 if (cpumask_test_cpu(cpu, buffer->cpumask))
4046 buffer->buffers[cpu] =
4047 rb_allocate_cpu_buffer(buffer, cpu);
4048 if (!buffer->buffers[cpu]) {
4049 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4054 cpumask_set_cpu(cpu, buffer->cpumask);
4056 case CPU_DOWN_PREPARE:
4057 case CPU_DOWN_PREPARE_FROZEN:
4060 * If we were to free the buffer, then the user would
4061 * lose any trace that was in the buffer.