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>
26 static void update_pages_handler(struct work_struct *work);
29 * The ring buffer header is special. We must manually up keep it.
31 int ring_buffer_print_entry_header(struct trace_seq *s)
35 ret = trace_seq_printf(s, "# compressed entry header\n");
36 ret = trace_seq_printf(s, "\ttype_len : 5 bits\n");
37 ret = trace_seq_printf(s, "\ttime_delta : 27 bits\n");
38 ret = trace_seq_printf(s, "\tarray : 32 bits\n");
39 ret = trace_seq_printf(s, "\n");
40 ret = trace_seq_printf(s, "\tpadding : type == %d\n",
41 RINGBUF_TYPE_PADDING);
42 ret = trace_seq_printf(s, "\ttime_extend : type == %d\n",
43 RINGBUF_TYPE_TIME_EXTEND);
44 ret = trace_seq_printf(s, "\tdata max type_len == %d\n",
45 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
51 * The ring buffer is made up of a list of pages. A separate list of pages is
52 * allocated for each CPU. A writer may only write to a buffer that is
53 * associated with the CPU it is currently executing on. A reader may read
54 * from any per cpu buffer.
56 * The reader is special. For each per cpu buffer, the reader has its own
57 * reader page. When a reader has read the entire reader page, this reader
58 * page is swapped with another page in the ring buffer.
60 * Now, as long as the writer is off the reader page, the reader can do what
61 * ever it wants with that page. The writer will never write to that page
62 * again (as long as it is out of the ring buffer).
64 * Here's some silly ASCII art.
67 * |reader| RING BUFFER
69 * +------+ +---+ +---+ +---+
78 * |reader| RING BUFFER
79 * |page |------------------v
80 * +------+ +---+ +---+ +---+
89 * |reader| RING BUFFER
90 * |page |------------------v
91 * +------+ +---+ +---+ +---+
96 * +------------------------------+
100 * |buffer| RING BUFFER
101 * |page |------------------v
102 * +------+ +---+ +---+ +---+
104 * | New +---+ +---+ +---+
107 * +------------------------------+
110 * After we make this swap, the reader can hand this page off to the splice
111 * code and be done with it. It can even allocate a new page if it needs to
112 * and swap that into the ring buffer.
114 * We will be using cmpxchg soon to make all this lockless.
119 * A fast way to enable or disable all ring buffers is to
120 * call tracing_on or tracing_off. Turning off the ring buffers
121 * prevents all ring buffers from being recorded to.
122 * Turning this switch on, makes it OK to write to the
123 * ring buffer, if the ring buffer is enabled itself.
125 * There's three layers that must be on in order to write
126 * to the ring buffer.
128 * 1) This global flag must be set.
129 * 2) The ring buffer must be enabled for recording.
130 * 3) The per cpu buffer must be enabled for recording.
132 * In case of an anomaly, this global flag has a bit set that
133 * will permantly disable all ring buffers.
137 * Global flag to disable all recording to ring buffers
138 * This has two bits: ON, DISABLED
142 * 0 0 : ring buffers are off
143 * 1 0 : ring buffers are on
144 * X 1 : ring buffers are permanently disabled
148 RB_BUFFERS_ON_BIT = 0,
149 RB_BUFFERS_DISABLED_BIT = 1,
153 RB_BUFFERS_ON = 1 << RB_BUFFERS_ON_BIT,
154 RB_BUFFERS_DISABLED = 1 << RB_BUFFERS_DISABLED_BIT,
157 static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON;
159 /* Used for individual buffers (after the counter) */
160 #define RB_BUFFER_OFF (1 << 20)
162 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
165 * tracing_off_permanent - permanently disable ring buffers
167 * This function, once called, will disable all ring buffers
170 void tracing_off_permanent(void)
172 set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags);
175 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
176 #define RB_ALIGNMENT 4U
177 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
178 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
180 #if !defined(CONFIG_64BIT) || defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
181 # define RB_FORCE_8BYTE_ALIGNMENT 0
182 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
184 # define RB_FORCE_8BYTE_ALIGNMENT 1
185 # define RB_ARCH_ALIGNMENT 8U
188 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
189 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
192 RB_LEN_TIME_EXTEND = 8,
193 RB_LEN_TIME_STAMP = 16,
196 #define skip_time_extend(event) \
197 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
199 static inline int rb_null_event(struct ring_buffer_event *event)
201 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
204 static void rb_event_set_padding(struct ring_buffer_event *event)
206 /* padding has a NULL time_delta */
207 event->type_len = RINGBUF_TYPE_PADDING;
208 event->time_delta = 0;
212 rb_event_data_length(struct ring_buffer_event *event)
217 length = event->type_len * RB_ALIGNMENT;
219 length = event->array[0];
220 return length + RB_EVNT_HDR_SIZE;
224 * Return the length of the given event. Will return
225 * the length of the time extend if the event is a
228 static inline unsigned
229 rb_event_length(struct ring_buffer_event *event)
231 switch (event->type_len) {
232 case RINGBUF_TYPE_PADDING:
233 if (rb_null_event(event))
236 return event->array[0] + RB_EVNT_HDR_SIZE;
238 case RINGBUF_TYPE_TIME_EXTEND:
239 return RB_LEN_TIME_EXTEND;
241 case RINGBUF_TYPE_TIME_STAMP:
242 return RB_LEN_TIME_STAMP;
244 case RINGBUF_TYPE_DATA:
245 return rb_event_data_length(event);
254 * Return total length of time extend and data,
255 * or just the event length for all other events.
257 static inline unsigned
258 rb_event_ts_length(struct ring_buffer_event *event)
262 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
263 /* time extends include the data event after it */
264 len = RB_LEN_TIME_EXTEND;
265 event = skip_time_extend(event);
267 return len + rb_event_length(event);
271 * ring_buffer_event_length - return the length of the event
272 * @event: the event to get the length of
274 * Returns the size of the data load of a data event.
275 * If the event is something other than a data event, it
276 * returns the size of the event itself. With the exception
277 * of a TIME EXTEND, where it still returns the size of the
278 * data load of the data event after it.
280 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
284 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
285 event = skip_time_extend(event);
287 length = rb_event_length(event);
288 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
290 length -= RB_EVNT_HDR_SIZE;
291 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
292 length -= sizeof(event->array[0]);
295 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
297 /* inline for ring buffer fast paths */
299 rb_event_data(struct ring_buffer_event *event)
301 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
302 event = skip_time_extend(event);
303 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
304 /* If length is in len field, then array[0] has the data */
306 return (void *)&event->array[0];
307 /* Otherwise length is in array[0] and array[1] has the data */
308 return (void *)&event->array[1];
312 * ring_buffer_event_data - return the data of the event
313 * @event: the event to get the data from
315 void *ring_buffer_event_data(struct ring_buffer_event *event)
317 return rb_event_data(event);
319 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
321 #define for_each_buffer_cpu(buffer, cpu) \
322 for_each_cpu(cpu, buffer->cpumask)
325 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
326 #define TS_DELTA_TEST (~TS_MASK)
328 /* Flag when events were overwritten */
329 #define RB_MISSED_EVENTS (1 << 31)
330 /* Missed count stored at end */
331 #define RB_MISSED_STORED (1 << 30)
333 struct buffer_data_page {
334 u64 time_stamp; /* page time stamp */
335 local_t commit; /* write committed index */
336 unsigned char data[]; /* data of buffer page */
340 * Note, the buffer_page list must be first. The buffer pages
341 * are allocated in cache lines, which means that each buffer
342 * page will be at the beginning of a cache line, and thus
343 * the least significant bits will be zero. We use this to
344 * add flags in the list struct pointers, to make the ring buffer
348 struct list_head list; /* list of buffer pages */
349 local_t write; /* index for next write */
350 unsigned read; /* index for next read */
351 local_t entries; /* entries on this page */
352 unsigned long real_end; /* real end of data */
353 struct buffer_data_page *page; /* Actual data page */
357 * The buffer page counters, write and entries, must be reset
358 * atomically when crossing page boundaries. To synchronize this
359 * update, two counters are inserted into the number. One is
360 * the actual counter for the write position or count on the page.
362 * The other is a counter of updaters. Before an update happens
363 * the update partition of the counter is incremented. This will
364 * allow the updater to update the counter atomically.
366 * The counter is 20 bits, and the state data is 12.
368 #define RB_WRITE_MASK 0xfffff
369 #define RB_WRITE_INTCNT (1 << 20)
371 static void rb_init_page(struct buffer_data_page *bpage)
373 local_set(&bpage->commit, 0);
377 * ring_buffer_page_len - the size of data on the page.
378 * @page: The page to read
380 * Returns the amount of data on the page, including buffer page header.
382 size_t ring_buffer_page_len(void *page)
384 return local_read(&((struct buffer_data_page *)page)->commit)
389 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
392 static void free_buffer_page(struct buffer_page *bpage)
394 free_page((unsigned long)bpage->page);
399 * We need to fit the time_stamp delta into 27 bits.
401 static inline int test_time_stamp(u64 delta)
403 if (delta & TS_DELTA_TEST)
408 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
410 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
411 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
413 int ring_buffer_print_page_header(struct trace_seq *s)
415 struct buffer_data_page field;
418 ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t"
419 "offset:0;\tsize:%u;\tsigned:%u;\n",
420 (unsigned int)sizeof(field.time_stamp),
421 (unsigned int)is_signed_type(u64));
423 ret = trace_seq_printf(s, "\tfield: local_t commit;\t"
424 "offset:%u;\tsize:%u;\tsigned:%u;\n",
425 (unsigned int)offsetof(typeof(field), commit),
426 (unsigned int)sizeof(field.commit),
427 (unsigned int)is_signed_type(long));
429 ret = trace_seq_printf(s, "\tfield: int overwrite;\t"
430 "offset:%u;\tsize:%u;\tsigned:%u;\n",
431 (unsigned int)offsetof(typeof(field), commit),
433 (unsigned int)is_signed_type(long));
435 ret = trace_seq_printf(s, "\tfield: char data;\t"
436 "offset:%u;\tsize:%u;\tsigned:%u;\n",
437 (unsigned int)offsetof(typeof(field), data),
438 (unsigned int)BUF_PAGE_SIZE,
439 (unsigned int)is_signed_type(char));
445 * head_page == tail_page && head == tail then buffer is empty.
447 struct ring_buffer_per_cpu {
449 atomic_t record_disabled;
450 struct ring_buffer *buffer;
451 raw_spinlock_t reader_lock; /* serialize readers */
452 arch_spinlock_t lock;
453 struct lock_class_key lock_key;
454 unsigned int nr_pages;
455 struct list_head *pages;
456 struct buffer_page *head_page; /* read from head */
457 struct buffer_page *tail_page; /* write to tail */
458 struct buffer_page *commit_page; /* committed pages */
459 struct buffer_page *reader_page;
460 unsigned long lost_events;
461 unsigned long last_overrun;
462 local_t entries_bytes;
465 local_t commit_overrun;
466 local_t dropped_events;
470 unsigned long read_bytes;
473 /* ring buffer pages to update, > 0 to add, < 0 to remove */
474 int nr_pages_to_update;
475 struct list_head new_pages; /* new pages to add */
476 struct work_struct update_pages_work;
477 struct completion update_done;
483 atomic_t record_disabled;
484 atomic_t resize_disabled;
485 cpumask_var_t cpumask;
487 struct lock_class_key *reader_lock_key;
491 struct ring_buffer_per_cpu **buffers;
493 #ifdef CONFIG_HOTPLUG_CPU
494 struct notifier_block cpu_notify;
499 struct ring_buffer_iter {
500 struct ring_buffer_per_cpu *cpu_buffer;
502 struct buffer_page *head_page;
503 struct buffer_page *cache_reader_page;
504 unsigned long cache_read;
508 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
509 #define RB_WARN_ON(b, cond) \
511 int _____ret = unlikely(cond); \
513 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
514 struct ring_buffer_per_cpu *__b = \
516 atomic_inc(&__b->buffer->record_disabled); \
518 atomic_inc(&b->record_disabled); \
524 /* Up this if you want to test the TIME_EXTENTS and normalization */
525 #define DEBUG_SHIFT 0
527 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
529 /* shift to debug/test normalization and TIME_EXTENTS */
530 return buffer->clock() << DEBUG_SHIFT;
533 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
537 preempt_disable_notrace();
538 time = rb_time_stamp(buffer);
539 preempt_enable_no_resched_notrace();
543 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
545 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
548 /* Just stupid testing the normalize function and deltas */
551 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
554 * Making the ring buffer lockless makes things tricky.
555 * Although writes only happen on the CPU that they are on,
556 * and they only need to worry about interrupts. Reads can
559 * The reader page is always off the ring buffer, but when the
560 * reader finishes with a page, it needs to swap its page with
561 * a new one from the buffer. The reader needs to take from
562 * the head (writes go to the tail). But if a writer is in overwrite
563 * mode and wraps, it must push the head page forward.
565 * Here lies the problem.
567 * The reader must be careful to replace only the head page, and
568 * not another one. As described at the top of the file in the
569 * ASCII art, the reader sets its old page to point to the next
570 * page after head. It then sets the page after head to point to
571 * the old reader page. But if the writer moves the head page
572 * during this operation, the reader could end up with the tail.
574 * We use cmpxchg to help prevent this race. We also do something
575 * special with the page before head. We set the LSB to 1.
577 * When the writer must push the page forward, it will clear the
578 * bit that points to the head page, move the head, and then set
579 * the bit that points to the new head page.
581 * We also don't want an interrupt coming in and moving the head
582 * page on another writer. Thus we use the second LSB to catch
585 * head->list->prev->next bit 1 bit 0
588 * Points to head page 0 1
591 * Note we can not trust the prev pointer of the head page, because:
593 * +----+ +-----+ +-----+
594 * | |------>| T |---X--->| N |
596 * +----+ +-----+ +-----+
599 * +----------| R |----------+ |
603 * Key: ---X--> HEAD flag set in pointer
608 * (see __rb_reserve_next() to see where this happens)
610 * What the above shows is that the reader just swapped out
611 * the reader page with a page in the buffer, but before it
612 * could make the new header point back to the new page added
613 * it was preempted by a writer. The writer moved forward onto
614 * the new page added by the reader and is about to move forward
617 * You can see, it is legitimate for the previous pointer of
618 * the head (or any page) not to point back to itself. But only
622 #define RB_PAGE_NORMAL 0UL
623 #define RB_PAGE_HEAD 1UL
624 #define RB_PAGE_UPDATE 2UL
627 #define RB_FLAG_MASK 3UL
629 /* PAGE_MOVED is not part of the mask */
630 #define RB_PAGE_MOVED 4UL
633 * rb_list_head - remove any bit
635 static struct list_head *rb_list_head(struct list_head *list)
637 unsigned long val = (unsigned long)list;
639 return (struct list_head *)(val & ~RB_FLAG_MASK);
643 * rb_is_head_page - test if the given page is the head page
645 * Because the reader may move the head_page pointer, we can
646 * not trust what the head page is (it may be pointing to
647 * the reader page). But if the next page is a header page,
648 * its flags will be non zero.
651 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
652 struct buffer_page *page, struct list_head *list)
656 val = (unsigned long)list->next;
658 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
659 return RB_PAGE_MOVED;
661 return val & RB_FLAG_MASK;
667 * The unique thing about the reader page, is that, if the
668 * writer is ever on it, the previous pointer never points
669 * back to the reader page.
671 static int rb_is_reader_page(struct buffer_page *page)
673 struct list_head *list = page->list.prev;
675 return rb_list_head(list->next) != &page->list;
679 * rb_set_list_to_head - set a list_head to be pointing to head.
681 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
682 struct list_head *list)
686 ptr = (unsigned long *)&list->next;
687 *ptr |= RB_PAGE_HEAD;
688 *ptr &= ~RB_PAGE_UPDATE;
692 * rb_head_page_activate - sets up head page
694 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
696 struct buffer_page *head;
698 head = cpu_buffer->head_page;
703 * Set the previous list pointer to have the HEAD flag.
705 rb_set_list_to_head(cpu_buffer, head->list.prev);
708 static void rb_list_head_clear(struct list_head *list)
710 unsigned long *ptr = (unsigned long *)&list->next;
712 *ptr &= ~RB_FLAG_MASK;
716 * rb_head_page_dactivate - clears head page ptr (for free list)
719 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
721 struct list_head *hd;
723 /* Go through the whole list and clear any pointers found. */
724 rb_list_head_clear(cpu_buffer->pages);
726 list_for_each(hd, cpu_buffer->pages)
727 rb_list_head_clear(hd);
730 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
731 struct buffer_page *head,
732 struct buffer_page *prev,
733 int old_flag, int new_flag)
735 struct list_head *list;
736 unsigned long val = (unsigned long)&head->list;
741 val &= ~RB_FLAG_MASK;
743 ret = cmpxchg((unsigned long *)&list->next,
744 val | old_flag, val | new_flag);
746 /* check if the reader took the page */
747 if ((ret & ~RB_FLAG_MASK) != val)
748 return RB_PAGE_MOVED;
750 return ret & RB_FLAG_MASK;
753 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
754 struct buffer_page *head,
755 struct buffer_page *prev,
758 return rb_head_page_set(cpu_buffer, head, prev,
759 old_flag, RB_PAGE_UPDATE);
762 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
763 struct buffer_page *head,
764 struct buffer_page *prev,
767 return rb_head_page_set(cpu_buffer, head, prev,
768 old_flag, RB_PAGE_HEAD);
771 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
772 struct buffer_page *head,
773 struct buffer_page *prev,
776 return rb_head_page_set(cpu_buffer, head, prev,
777 old_flag, RB_PAGE_NORMAL);
780 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
781 struct buffer_page **bpage)
783 struct list_head *p = rb_list_head((*bpage)->list.next);
785 *bpage = list_entry(p, struct buffer_page, list);
788 static struct buffer_page *
789 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
791 struct buffer_page *head;
792 struct buffer_page *page;
793 struct list_head *list;
796 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
800 list = cpu_buffer->pages;
801 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
804 page = head = cpu_buffer->head_page;
806 * It is possible that the writer moves the header behind
807 * where we started, and we miss in one loop.
808 * A second loop should grab the header, but we'll do
809 * three loops just because I'm paranoid.
811 for (i = 0; i < 3; i++) {
813 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
814 cpu_buffer->head_page = page;
817 rb_inc_page(cpu_buffer, &page);
818 } while (page != head);
821 RB_WARN_ON(cpu_buffer, 1);
826 static int rb_head_page_replace(struct buffer_page *old,
827 struct buffer_page *new)
829 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
833 val = *ptr & ~RB_FLAG_MASK;
836 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
842 * rb_tail_page_update - move the tail page forward
844 * Returns 1 if moved tail page, 0 if someone else did.
846 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
847 struct buffer_page *tail_page,
848 struct buffer_page *next_page)
850 struct buffer_page *old_tail;
851 unsigned long old_entries;
852 unsigned long old_write;
856 * The tail page now needs to be moved forward.
858 * We need to reset the tail page, but without messing
859 * with possible erasing of data brought in by interrupts
860 * that have moved the tail page and are currently on it.
862 * We add a counter to the write field to denote this.
864 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
865 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
868 * Just make sure we have seen our old_write and synchronize
869 * with any interrupts that come in.
874 * If the tail page is still the same as what we think
875 * it is, then it is up to us to update the tail
878 if (tail_page == cpu_buffer->tail_page) {
879 /* Zero the write counter */
880 unsigned long val = old_write & ~RB_WRITE_MASK;
881 unsigned long eval = old_entries & ~RB_WRITE_MASK;
884 * This will only succeed if an interrupt did
885 * not come in and change it. In which case, we
886 * do not want to modify it.
888 * We add (void) to let the compiler know that we do not care
889 * about the return value of these functions. We use the
890 * cmpxchg to only update if an interrupt did not already
891 * do it for us. If the cmpxchg fails, we don't care.
893 (void)local_cmpxchg(&next_page->write, old_write, val);
894 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
897 * No need to worry about races with clearing out the commit.
898 * it only can increment when a commit takes place. But that
899 * only happens in the outer most nested commit.
901 local_set(&next_page->page->commit, 0);
903 old_tail = cmpxchg(&cpu_buffer->tail_page,
904 tail_page, next_page);
906 if (old_tail == tail_page)
913 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
914 struct buffer_page *bpage)
916 unsigned long val = (unsigned long)bpage;
918 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
925 * rb_check_list - make sure a pointer to a list has the last bits zero
927 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
928 struct list_head *list)
930 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
932 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
938 * check_pages - integrity check of buffer pages
939 * @cpu_buffer: CPU buffer with pages to test
941 * As a safety measure we check to make sure the data pages have not
944 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
946 struct list_head *head = cpu_buffer->pages;
947 struct buffer_page *bpage, *tmp;
949 /* Reset the head page if it exists */
950 if (cpu_buffer->head_page)
951 rb_set_head_page(cpu_buffer);
953 rb_head_page_deactivate(cpu_buffer);
955 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
957 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
960 if (rb_check_list(cpu_buffer, head))
963 list_for_each_entry_safe(bpage, tmp, head, list) {
964 if (RB_WARN_ON(cpu_buffer,
965 bpage->list.next->prev != &bpage->list))
967 if (RB_WARN_ON(cpu_buffer,
968 bpage->list.prev->next != &bpage->list))
970 if (rb_check_list(cpu_buffer, &bpage->list))
974 rb_head_page_activate(cpu_buffer);
979 static int __rb_allocate_pages(int nr_pages, struct list_head *pages, int cpu)
982 struct buffer_page *bpage, *tmp;
984 for (i = 0; i < nr_pages; i++) {
987 * __GFP_NORETRY flag makes sure that the allocation fails
988 * gracefully without invoking oom-killer and the system is
991 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
992 GFP_KERNEL | __GFP_NORETRY,
997 list_add(&bpage->list, pages);
999 page = alloc_pages_node(cpu_to_node(cpu),
1000 GFP_KERNEL | __GFP_NORETRY, 0);
1003 bpage->page = page_address(page);
1004 rb_init_page(bpage->page);
1010 list_for_each_entry_safe(bpage, tmp, pages, list) {
1011 list_del_init(&bpage->list);
1012 free_buffer_page(bpage);
1018 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1025 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1029 * The ring buffer page list is a circular list that does not
1030 * start and end with a list head. All page list items point to
1033 cpu_buffer->pages = pages.next;
1036 cpu_buffer->nr_pages = nr_pages;
1038 rb_check_pages(cpu_buffer);
1043 static struct ring_buffer_per_cpu *
1044 rb_allocate_cpu_buffer(struct ring_buffer *buffer, int nr_pages, int cpu)
1046 struct ring_buffer_per_cpu *cpu_buffer;
1047 struct buffer_page *bpage;
1051 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1052 GFP_KERNEL, cpu_to_node(cpu));
1056 cpu_buffer->cpu = cpu;
1057 cpu_buffer->buffer = buffer;
1058 raw_spin_lock_init(&cpu_buffer->reader_lock);
1059 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1060 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1061 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1062 init_completion(&cpu_buffer->update_done);
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);
1079 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1081 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1083 goto fail_free_reader;
1085 cpu_buffer->head_page
1086 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1087 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1089 rb_head_page_activate(cpu_buffer);
1094 free_buffer_page(cpu_buffer->reader_page);
1101 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1103 struct list_head *head = cpu_buffer->pages;
1104 struct buffer_page *bpage, *tmp;
1106 free_buffer_page(cpu_buffer->reader_page);
1108 rb_head_page_deactivate(cpu_buffer);
1111 list_for_each_entry_safe(bpage, tmp, head, list) {
1112 list_del_init(&bpage->list);
1113 free_buffer_page(bpage);
1115 bpage = list_entry(head, struct buffer_page, list);
1116 free_buffer_page(bpage);
1122 #ifdef CONFIG_HOTPLUG_CPU
1123 static int rb_cpu_notify(struct notifier_block *self,
1124 unsigned long action, void *hcpu);
1128 * ring_buffer_alloc - allocate a new ring_buffer
1129 * @size: the size in bytes per cpu that is needed.
1130 * @flags: attributes to set for the ring buffer.
1132 * Currently the only flag that is available is the RB_FL_OVERWRITE
1133 * flag. This flag means that the buffer will overwrite old data
1134 * when the buffer wraps. If this flag is not set, the buffer will
1135 * drop data when the tail hits the head.
1137 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1138 struct lock_class_key *key)
1140 struct ring_buffer *buffer;
1144 /* keep it in its own cache line */
1145 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1150 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1151 goto fail_free_buffer;
1153 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1154 buffer->flags = flags;
1155 buffer->clock = trace_clock_local;
1156 buffer->reader_lock_key = key;
1158 /* need at least two pages */
1163 * In case of non-hotplug cpu, if the ring-buffer is allocated
1164 * in early initcall, it will not be notified of secondary cpus.
1165 * In that off case, we need to allocate for all possible cpus.
1167 #ifdef CONFIG_HOTPLUG_CPU
1169 cpumask_copy(buffer->cpumask, cpu_online_mask);
1171 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1173 buffer->cpus = nr_cpu_ids;
1175 bsize = sizeof(void *) * nr_cpu_ids;
1176 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1178 if (!buffer->buffers)
1179 goto fail_free_cpumask;
1181 for_each_buffer_cpu(buffer, cpu) {
1182 buffer->buffers[cpu] =
1183 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1184 if (!buffer->buffers[cpu])
1185 goto fail_free_buffers;
1188 #ifdef CONFIG_HOTPLUG_CPU
1189 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1190 buffer->cpu_notify.priority = 0;
1191 register_cpu_notifier(&buffer->cpu_notify);
1195 mutex_init(&buffer->mutex);
1200 for_each_buffer_cpu(buffer, cpu) {
1201 if (buffer->buffers[cpu])
1202 rb_free_cpu_buffer(buffer->buffers[cpu]);
1204 kfree(buffer->buffers);
1207 free_cpumask_var(buffer->cpumask);
1214 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1217 * ring_buffer_free - free a ring buffer.
1218 * @buffer: the buffer to free.
1221 ring_buffer_free(struct ring_buffer *buffer)
1227 #ifdef CONFIG_HOTPLUG_CPU
1228 unregister_cpu_notifier(&buffer->cpu_notify);
1231 for_each_buffer_cpu(buffer, cpu)
1232 rb_free_cpu_buffer(buffer->buffers[cpu]);
1236 kfree(buffer->buffers);
1237 free_cpumask_var(buffer->cpumask);
1241 EXPORT_SYMBOL_GPL(ring_buffer_free);
1243 void ring_buffer_set_clock(struct ring_buffer *buffer,
1246 buffer->clock = clock;
1249 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1251 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1253 return local_read(&bpage->entries) & RB_WRITE_MASK;
1256 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1258 return local_read(&bpage->write) & RB_WRITE_MASK;
1262 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned int nr_pages)
1264 struct list_head *tail_page, *to_remove, *next_page;
1265 struct buffer_page *to_remove_page, *tmp_iter_page;
1266 struct buffer_page *last_page, *first_page;
1267 unsigned int nr_removed;
1268 unsigned long head_bit;
1273 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1274 atomic_inc(&cpu_buffer->record_disabled);
1276 * We don't race with the readers since we have acquired the reader
1277 * lock. We also don't race with writers after disabling recording.
1278 * This makes it easy to figure out the first and the last page to be
1279 * removed from the list. We unlink all the pages in between including
1280 * the first and last pages. This is done in a busy loop so that we
1281 * lose the least number of traces.
1282 * The pages are freed after we restart recording and unlock readers.
1284 tail_page = &cpu_buffer->tail_page->list;
1287 * tail page might be on reader page, we remove the next page
1288 * from the ring buffer
1290 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1291 tail_page = rb_list_head(tail_page->next);
1292 to_remove = tail_page;
1294 /* start of pages to remove */
1295 first_page = list_entry(rb_list_head(to_remove->next),
1296 struct buffer_page, list);
1298 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1299 to_remove = rb_list_head(to_remove)->next;
1300 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1303 next_page = rb_list_head(to_remove)->next;
1306 * Now we remove all pages between tail_page and next_page.
1307 * Make sure that we have head_bit value preserved for the
1310 tail_page->next = (struct list_head *)((unsigned long)next_page |
1312 next_page = rb_list_head(next_page);
1313 next_page->prev = tail_page;
1315 /* make sure pages points to a valid page in the ring buffer */
1316 cpu_buffer->pages = next_page;
1318 /* update head page */
1320 cpu_buffer->head_page = list_entry(next_page,
1321 struct buffer_page, list);
1324 * change read pointer to make sure any read iterators reset
1327 cpu_buffer->read = 0;
1329 /* pages are removed, resume tracing and then free the pages */
1330 atomic_dec(&cpu_buffer->record_disabled);
1331 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1333 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1335 /* last buffer page to remove */
1336 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1338 tmp_iter_page = first_page;
1341 to_remove_page = tmp_iter_page;
1342 rb_inc_page(cpu_buffer, &tmp_iter_page);
1344 /* update the counters */
1345 page_entries = rb_page_entries(to_remove_page);
1348 * If something was added to this page, it was full
1349 * since it is not the tail page. So we deduct the
1350 * bytes consumed in ring buffer from here.
1351 * Increment overrun to account for the lost events.
1353 local_add(page_entries, &cpu_buffer->overrun);
1354 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1358 * We have already removed references to this list item, just
1359 * free up the buffer_page and its page
1361 free_buffer_page(to_remove_page);
1364 } while (to_remove_page != last_page);
1366 RB_WARN_ON(cpu_buffer, nr_removed);
1368 return nr_removed == 0;
1372 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1374 struct list_head *pages = &cpu_buffer->new_pages;
1375 int retries, success;
1377 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1379 * We are holding the reader lock, so the reader page won't be swapped
1380 * in the ring buffer. Now we are racing with the writer trying to
1381 * move head page and the tail page.
1382 * We are going to adapt the reader page update process where:
1383 * 1. We first splice the start and end of list of new pages between
1384 * the head page and its previous page.
1385 * 2. We cmpxchg the prev_page->next to point from head page to the
1386 * start of new pages list.
1387 * 3. Finally, we update the head->prev to the end of new list.
1389 * We will try this process 10 times, to make sure that we don't keep
1395 struct list_head *head_page, *prev_page, *r;
1396 struct list_head *last_page, *first_page;
1397 struct list_head *head_page_with_bit;
1399 head_page = &rb_set_head_page(cpu_buffer)->list;
1402 prev_page = head_page->prev;
1404 first_page = pages->next;
1405 last_page = pages->prev;
1407 head_page_with_bit = (struct list_head *)
1408 ((unsigned long)head_page | RB_PAGE_HEAD);
1410 last_page->next = head_page_with_bit;
1411 first_page->prev = prev_page;
1413 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1415 if (r == head_page_with_bit) {
1417 * yay, we replaced the page pointer to our new list,
1418 * now, we just have to update to head page's prev
1419 * pointer to point to end of list
1421 head_page->prev = last_page;
1428 INIT_LIST_HEAD(pages);
1430 * If we weren't successful in adding in new pages, warn and stop
1433 RB_WARN_ON(cpu_buffer, !success);
1434 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1436 /* free pages if they weren't inserted */
1438 struct buffer_page *bpage, *tmp;
1439 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1441 list_del_init(&bpage->list);
1442 free_buffer_page(bpage);
1448 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1452 if (cpu_buffer->nr_pages_to_update > 0)
1453 success = rb_insert_pages(cpu_buffer);
1455 success = rb_remove_pages(cpu_buffer,
1456 -cpu_buffer->nr_pages_to_update);
1459 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1462 static void update_pages_handler(struct work_struct *work)
1464 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1465 struct ring_buffer_per_cpu, update_pages_work);
1466 rb_update_pages(cpu_buffer);
1467 complete(&cpu_buffer->update_done);
1471 * ring_buffer_resize - resize the ring buffer
1472 * @buffer: the buffer to resize.
1473 * @size: the new size.
1475 * Minimum size is 2 * BUF_PAGE_SIZE.
1477 * Returns 0 on success and < 0 on failure.
1479 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1482 struct ring_buffer_per_cpu *cpu_buffer;
1487 * Always succeed at resizing a non-existent buffer:
1492 /* Make sure the requested buffer exists */
1493 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1494 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1497 size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1498 size *= BUF_PAGE_SIZE;
1500 /* we need a minimum of two pages */
1501 if (size < BUF_PAGE_SIZE * 2)
1502 size = BUF_PAGE_SIZE * 2;
1504 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1507 * Don't succeed if resizing is disabled, as a reader might be
1508 * manipulating the ring buffer and is expecting a sane state while
1511 if (atomic_read(&buffer->resize_disabled))
1514 /* prevent another thread from changing buffer sizes */
1515 mutex_lock(&buffer->mutex);
1517 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1518 /* calculate the pages to update */
1519 for_each_buffer_cpu(buffer, cpu) {
1520 cpu_buffer = buffer->buffers[cpu];
1522 cpu_buffer->nr_pages_to_update = nr_pages -
1523 cpu_buffer->nr_pages;
1525 * nothing more to do for removing pages or no update
1527 if (cpu_buffer->nr_pages_to_update <= 0)
1530 * to add pages, make sure all new pages can be
1531 * allocated without receiving ENOMEM
1533 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1534 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1535 &cpu_buffer->new_pages, cpu)) {
1536 /* not enough memory for new pages */
1544 * Fire off all the required work handlers
1545 * We can't schedule on offline CPUs, but it's not necessary
1546 * since we can change their buffer sizes without any race.
1548 for_each_buffer_cpu(buffer, cpu) {
1549 cpu_buffer = buffer->buffers[cpu];
1550 if (!cpu_buffer->nr_pages_to_update)
1553 if (cpu_online(cpu))
1554 schedule_work_on(cpu,
1555 &cpu_buffer->update_pages_work);
1557 rb_update_pages(cpu_buffer);
1560 /* wait for all the updates to complete */
1561 for_each_buffer_cpu(buffer, cpu) {
1562 cpu_buffer = buffer->buffers[cpu];
1563 if (!cpu_buffer->nr_pages_to_update)
1566 if (cpu_online(cpu))
1567 wait_for_completion(&cpu_buffer->update_done);
1568 cpu_buffer->nr_pages_to_update = 0;
1573 /* Make sure this CPU has been intitialized */
1574 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1577 cpu_buffer = buffer->buffers[cpu_id];
1579 if (nr_pages == cpu_buffer->nr_pages)
1582 cpu_buffer->nr_pages_to_update = nr_pages -
1583 cpu_buffer->nr_pages;
1585 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1586 if (cpu_buffer->nr_pages_to_update > 0 &&
1587 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1588 &cpu_buffer->new_pages, cpu_id)) {
1595 if (cpu_online(cpu_id)) {
1596 schedule_work_on(cpu_id,
1597 &cpu_buffer->update_pages_work);
1598 wait_for_completion(&cpu_buffer->update_done);
1600 rb_update_pages(cpu_buffer);
1602 cpu_buffer->nr_pages_to_update = 0;
1608 * The ring buffer resize can happen with the ring buffer
1609 * enabled, so that the update disturbs the tracing as little
1610 * as possible. But if the buffer is disabled, we do not need
1611 * to worry about that, and we can take the time to verify
1612 * that the buffer is not corrupt.
1614 if (atomic_read(&buffer->record_disabled)) {
1615 atomic_inc(&buffer->record_disabled);
1617 * Even though the buffer was disabled, we must make sure
1618 * that it is truly disabled before calling rb_check_pages.
1619 * There could have been a race between checking
1620 * record_disable and incrementing it.
1622 synchronize_sched();
1623 for_each_buffer_cpu(buffer, cpu) {
1624 cpu_buffer = buffer->buffers[cpu];
1625 rb_check_pages(cpu_buffer);
1627 atomic_dec(&buffer->record_disabled);
1630 mutex_unlock(&buffer->mutex);
1634 for_each_buffer_cpu(buffer, cpu) {
1635 struct buffer_page *bpage, *tmp;
1637 cpu_buffer = buffer->buffers[cpu];
1638 cpu_buffer->nr_pages_to_update = 0;
1640 if (list_empty(&cpu_buffer->new_pages))
1643 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1645 list_del_init(&bpage->list);
1646 free_buffer_page(bpage);
1649 mutex_unlock(&buffer->mutex);
1652 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1654 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1656 mutex_lock(&buffer->mutex);
1658 buffer->flags |= RB_FL_OVERWRITE;
1660 buffer->flags &= ~RB_FL_OVERWRITE;
1661 mutex_unlock(&buffer->mutex);
1663 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1665 static inline void *
1666 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1668 return bpage->data + index;
1671 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1673 return bpage->page->data + index;
1676 static inline struct ring_buffer_event *
1677 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1679 return __rb_page_index(cpu_buffer->reader_page,
1680 cpu_buffer->reader_page->read);
1683 static inline struct ring_buffer_event *
1684 rb_iter_head_event(struct ring_buffer_iter *iter)
1686 return __rb_page_index(iter->head_page, iter->head);
1689 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1691 return local_read(&bpage->page->commit);
1694 /* Size is determined by what has been committed */
1695 static inline unsigned rb_page_size(struct buffer_page *bpage)
1697 return rb_page_commit(bpage);
1700 static inline unsigned
1701 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1703 return rb_page_commit(cpu_buffer->commit_page);
1706 static inline unsigned
1707 rb_event_index(struct ring_buffer_event *event)
1709 unsigned long addr = (unsigned long)event;
1711 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1715 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1716 struct ring_buffer_event *event)
1718 unsigned long addr = (unsigned long)event;
1719 unsigned long index;
1721 index = rb_event_index(event);
1724 return cpu_buffer->commit_page->page == (void *)addr &&
1725 rb_commit_index(cpu_buffer) == index;
1729 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1731 unsigned long max_count;
1734 * We only race with interrupts and NMIs on this CPU.
1735 * If we own the commit event, then we can commit
1736 * all others that interrupted us, since the interruptions
1737 * are in stack format (they finish before they come
1738 * back to us). This allows us to do a simple loop to
1739 * assign the commit to the tail.
1742 max_count = cpu_buffer->nr_pages * 100;
1744 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1745 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1747 if (RB_WARN_ON(cpu_buffer,
1748 rb_is_reader_page(cpu_buffer->tail_page)))
1750 local_set(&cpu_buffer->commit_page->page->commit,
1751 rb_page_write(cpu_buffer->commit_page));
1752 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1753 cpu_buffer->write_stamp =
1754 cpu_buffer->commit_page->page->time_stamp;
1755 /* add barrier to keep gcc from optimizing too much */
1758 while (rb_commit_index(cpu_buffer) !=
1759 rb_page_write(cpu_buffer->commit_page)) {
1761 local_set(&cpu_buffer->commit_page->page->commit,
1762 rb_page_write(cpu_buffer->commit_page));
1763 RB_WARN_ON(cpu_buffer,
1764 local_read(&cpu_buffer->commit_page->page->commit) &
1769 /* again, keep gcc from optimizing */
1773 * If an interrupt came in just after the first while loop
1774 * and pushed the tail page forward, we will be left with
1775 * a dangling commit that will never go forward.
1777 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1781 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1783 cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1784 cpu_buffer->reader_page->read = 0;
1787 static void rb_inc_iter(struct ring_buffer_iter *iter)
1789 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1792 * The iterator could be on the reader page (it starts there).
1793 * But the head could have moved, since the reader was
1794 * found. Check for this case and assign the iterator
1795 * to the head page instead of next.
1797 if (iter->head_page == cpu_buffer->reader_page)
1798 iter->head_page = rb_set_head_page(cpu_buffer);
1800 rb_inc_page(cpu_buffer, &iter->head_page);
1802 iter->read_stamp = iter->head_page->page->time_stamp;
1806 /* Slow path, do not inline */
1807 static noinline struct ring_buffer_event *
1808 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
1810 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
1812 /* Not the first event on the page? */
1813 if (rb_event_index(event)) {
1814 event->time_delta = delta & TS_MASK;
1815 event->array[0] = delta >> TS_SHIFT;
1817 /* nope, just zero it */
1818 event->time_delta = 0;
1819 event->array[0] = 0;
1822 return skip_time_extend(event);
1826 * rb_update_event - update event type and data
1827 * @event: the even to update
1828 * @type: the type of event
1829 * @length: the size of the event field in the ring buffer
1831 * Update the type and data fields of the event. The length
1832 * is the actual size that is written to the ring buffer,
1833 * and with this, we can determine what to place into the
1837 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
1838 struct ring_buffer_event *event, unsigned length,
1839 int add_timestamp, u64 delta)
1841 /* Only a commit updates the timestamp */
1842 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
1846 * If we need to add a timestamp, then we
1847 * add it to the start of the resevered space.
1849 if (unlikely(add_timestamp)) {
1850 event = rb_add_time_stamp(event, delta);
1851 length -= RB_LEN_TIME_EXTEND;
1855 event->time_delta = delta;
1856 length -= RB_EVNT_HDR_SIZE;
1857 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
1858 event->type_len = 0;
1859 event->array[0] = length;
1861 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
1865 * rb_handle_head_page - writer hit the head page
1867 * Returns: +1 to retry page
1872 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1873 struct buffer_page *tail_page,
1874 struct buffer_page *next_page)
1876 struct buffer_page *new_head;
1881 entries = rb_page_entries(next_page);
1884 * The hard part is here. We need to move the head
1885 * forward, and protect against both readers on
1886 * other CPUs and writers coming in via interrupts.
1888 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1892 * type can be one of four:
1893 * NORMAL - an interrupt already moved it for us
1894 * HEAD - we are the first to get here.
1895 * UPDATE - we are the interrupt interrupting
1897 * MOVED - a reader on another CPU moved the next
1898 * pointer to its reader page. Give up
1905 * We changed the head to UPDATE, thus
1906 * it is our responsibility to update
1909 local_add(entries, &cpu_buffer->overrun);
1910 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1913 * The entries will be zeroed out when we move the
1917 /* still more to do */
1920 case RB_PAGE_UPDATE:
1922 * This is an interrupt that interrupt the
1923 * previous update. Still more to do.
1926 case RB_PAGE_NORMAL:
1928 * An interrupt came in before the update
1929 * and processed this for us.
1930 * Nothing left to do.
1935 * The reader is on another CPU and just did
1936 * a swap with our next_page.
1941 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1946 * Now that we are here, the old head pointer is
1947 * set to UPDATE. This will keep the reader from
1948 * swapping the head page with the reader page.
1949 * The reader (on another CPU) will spin till
1952 * We just need to protect against interrupts
1953 * doing the job. We will set the next pointer
1954 * to HEAD. After that, we set the old pointer
1955 * to NORMAL, but only if it was HEAD before.
1956 * otherwise we are an interrupt, and only
1957 * want the outer most commit to reset it.
1959 new_head = next_page;
1960 rb_inc_page(cpu_buffer, &new_head);
1962 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1966 * Valid returns are:
1967 * HEAD - an interrupt came in and already set it.
1968 * NORMAL - One of two things:
1969 * 1) We really set it.
1970 * 2) A bunch of interrupts came in and moved
1971 * the page forward again.
1975 case RB_PAGE_NORMAL:
1979 RB_WARN_ON(cpu_buffer, 1);
1984 * It is possible that an interrupt came in,
1985 * set the head up, then more interrupts came in
1986 * and moved it again. When we get back here,
1987 * the page would have been set to NORMAL but we
1988 * just set it back to HEAD.
1990 * How do you detect this? Well, if that happened
1991 * the tail page would have moved.
1993 if (ret == RB_PAGE_NORMAL) {
1995 * If the tail had moved passed next, then we need
1996 * to reset the pointer.
1998 if (cpu_buffer->tail_page != tail_page &&
1999 cpu_buffer->tail_page != next_page)
2000 rb_head_page_set_normal(cpu_buffer, new_head,
2006 * If this was the outer most commit (the one that
2007 * changed the original pointer from HEAD to UPDATE),
2008 * then it is up to us to reset it to NORMAL.
2010 if (type == RB_PAGE_HEAD) {
2011 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2014 if (RB_WARN_ON(cpu_buffer,
2015 ret != RB_PAGE_UPDATE))
2022 static unsigned rb_calculate_event_length(unsigned length)
2024 struct ring_buffer_event event; /* Used only for sizeof array */
2026 /* zero length can cause confusions */
2030 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2031 length += sizeof(event.array[0]);
2033 length += RB_EVNT_HDR_SIZE;
2034 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2040 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2041 struct buffer_page *tail_page,
2042 unsigned long tail, unsigned long length)
2044 struct ring_buffer_event *event;
2047 * Only the event that crossed the page boundary
2048 * must fill the old tail_page with padding.
2050 if (tail >= BUF_PAGE_SIZE) {
2052 * If the page was filled, then we still need
2053 * to update the real_end. Reset it to zero
2054 * and the reader will ignore it.
2056 if (tail == BUF_PAGE_SIZE)
2057 tail_page->real_end = 0;
2059 local_sub(length, &tail_page->write);
2063 event = __rb_page_index(tail_page, tail);
2064 kmemcheck_annotate_bitfield(event, bitfield);
2066 /* account for padding bytes */
2067 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2070 * Save the original length to the meta data.
2071 * This will be used by the reader to add lost event
2074 tail_page->real_end = tail;
2077 * If this event is bigger than the minimum size, then
2078 * we need to be careful that we don't subtract the
2079 * write counter enough to allow another writer to slip
2081 * We put in a discarded commit instead, to make sure
2082 * that this space is not used again.
2084 * If we are less than the minimum size, we don't need to
2087 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2088 /* No room for any events */
2090 /* Mark the rest of the page with padding */
2091 rb_event_set_padding(event);
2093 /* Set the write back to the previous setting */
2094 local_sub(length, &tail_page->write);
2098 /* Put in a discarded event */
2099 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2100 event->type_len = RINGBUF_TYPE_PADDING;
2101 /* time delta must be non zero */
2102 event->time_delta = 1;
2104 /* Set write to end of buffer */
2105 length = (tail + length) - BUF_PAGE_SIZE;
2106 local_sub(length, &tail_page->write);
2110 * This is the slow path, force gcc not to inline it.
2112 static noinline struct ring_buffer_event *
2113 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2114 unsigned long length, unsigned long tail,
2115 struct buffer_page *tail_page, u64 ts)
2117 struct buffer_page *commit_page = cpu_buffer->commit_page;
2118 struct ring_buffer *buffer = cpu_buffer->buffer;
2119 struct buffer_page *next_page;
2122 next_page = tail_page;
2124 rb_inc_page(cpu_buffer, &next_page);
2127 * If for some reason, we had an interrupt storm that made
2128 * it all the way around the buffer, bail, and warn
2131 if (unlikely(next_page == commit_page)) {
2132 local_inc(&cpu_buffer->commit_overrun);
2137 * This is where the fun begins!
2139 * We are fighting against races between a reader that
2140 * could be on another CPU trying to swap its reader
2141 * page with the buffer head.
2143 * We are also fighting against interrupts coming in and
2144 * moving the head or tail on us as well.
2146 * If the next page is the head page then we have filled
2147 * the buffer, unless the commit page is still on the
2150 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2153 * If the commit is not on the reader page, then
2154 * move the header page.
2156 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2158 * If we are not in overwrite mode,
2159 * this is easy, just stop here.
2161 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2162 local_inc(&cpu_buffer->dropped_events);
2166 ret = rb_handle_head_page(cpu_buffer,
2175 * We need to be careful here too. The
2176 * commit page could still be on the reader
2177 * page. We could have a small buffer, and
2178 * have filled up the buffer with events
2179 * from interrupts and such, and wrapped.
2181 * Note, if the tail page is also the on the
2182 * reader_page, we let it move out.
2184 if (unlikely((cpu_buffer->commit_page !=
2185 cpu_buffer->tail_page) &&
2186 (cpu_buffer->commit_page ==
2187 cpu_buffer->reader_page))) {
2188 local_inc(&cpu_buffer->commit_overrun);
2194 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
2197 * Nested commits always have zero deltas, so
2198 * just reread the time stamp
2200 ts = rb_time_stamp(buffer);
2201 next_page->page->time_stamp = ts;
2206 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2208 /* fail and let the caller try again */
2209 return ERR_PTR(-EAGAIN);
2213 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2218 static struct ring_buffer_event *
2219 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2220 unsigned long length, u64 ts,
2221 u64 delta, int add_timestamp)
2223 struct buffer_page *tail_page;
2224 struct ring_buffer_event *event;
2225 unsigned long tail, write;
2228 * If the time delta since the last event is too big to
2229 * hold in the time field of the event, then we append a
2230 * TIME EXTEND event ahead of the data event.
2232 if (unlikely(add_timestamp))
2233 length += RB_LEN_TIME_EXTEND;
2235 tail_page = cpu_buffer->tail_page;
2236 write = local_add_return(length, &tail_page->write);
2238 /* set write to only the index of the write */
2239 write &= RB_WRITE_MASK;
2240 tail = write - length;
2242 /* See if we shot pass the end of this buffer page */
2243 if (unlikely(write > BUF_PAGE_SIZE))
2244 return rb_move_tail(cpu_buffer, length, tail,
2247 /* We reserved something on the buffer */
2249 event = __rb_page_index(tail_page, tail);
2250 kmemcheck_annotate_bitfield(event, bitfield);
2251 rb_update_event(cpu_buffer, event, length, add_timestamp, delta);
2253 local_inc(&tail_page->entries);
2256 * If this is the first commit on the page, then update
2260 tail_page->page->time_stamp = ts;
2262 /* account for these added bytes */
2263 local_add(length, &cpu_buffer->entries_bytes);
2269 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2270 struct ring_buffer_event *event)
2272 unsigned long new_index, old_index;
2273 struct buffer_page *bpage;
2274 unsigned long index;
2277 new_index = rb_event_index(event);
2278 old_index = new_index + rb_event_ts_length(event);
2279 addr = (unsigned long)event;
2282 bpage = cpu_buffer->tail_page;
2284 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2285 unsigned long write_mask =
2286 local_read(&bpage->write) & ~RB_WRITE_MASK;
2287 unsigned long event_length = rb_event_length(event);
2289 * This is on the tail page. It is possible that
2290 * a write could come in and move the tail page
2291 * and write to the next page. That is fine
2292 * because we just shorten what is on this page.
2294 old_index += write_mask;
2295 new_index += write_mask;
2296 index = local_cmpxchg(&bpage->write, old_index, new_index);
2297 if (index == old_index) {
2298 /* update counters */
2299 local_sub(event_length, &cpu_buffer->entries_bytes);
2304 /* could not discard */
2308 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2310 local_inc(&cpu_buffer->committing);
2311 local_inc(&cpu_buffer->commits);
2314 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2316 unsigned long commits;
2318 if (RB_WARN_ON(cpu_buffer,
2319 !local_read(&cpu_buffer->committing)))
2323 commits = local_read(&cpu_buffer->commits);
2324 /* synchronize with interrupts */
2326 if (local_read(&cpu_buffer->committing) == 1)
2327 rb_set_commit_to_write(cpu_buffer);
2329 local_dec(&cpu_buffer->committing);
2331 /* synchronize with interrupts */
2335 * Need to account for interrupts coming in between the
2336 * updating of the commit page and the clearing of the
2337 * committing counter.
2339 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2340 !local_read(&cpu_buffer->committing)) {
2341 local_inc(&cpu_buffer->committing);
2346 static struct ring_buffer_event *
2347 rb_reserve_next_event(struct ring_buffer *buffer,
2348 struct ring_buffer_per_cpu *cpu_buffer,
2349 unsigned long length)
2351 struct ring_buffer_event *event;
2357 rb_start_commit(cpu_buffer);
2359 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2361 * Due to the ability to swap a cpu buffer from a buffer
2362 * it is possible it was swapped before we committed.
2363 * (committing stops a swap). We check for it here and
2364 * if it happened, we have to fail the write.
2367 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2368 local_dec(&cpu_buffer->committing);
2369 local_dec(&cpu_buffer->commits);
2374 length = rb_calculate_event_length(length);
2380 * We allow for interrupts to reenter here and do a trace.
2381 * If one does, it will cause this original code to loop
2382 * back here. Even with heavy interrupts happening, this
2383 * should only happen a few times in a row. If this happens
2384 * 1000 times in a row, there must be either an interrupt
2385 * storm or we have something buggy.
2388 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2391 ts = rb_time_stamp(cpu_buffer->buffer);
2392 diff = ts - cpu_buffer->write_stamp;
2394 /* make sure this diff is calculated here */
2397 /* Did the write stamp get updated already? */
2398 if (likely(ts >= cpu_buffer->write_stamp)) {
2400 if (unlikely(test_time_stamp(delta))) {
2401 int local_clock_stable = 1;
2402 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2403 local_clock_stable = sched_clock_stable;
2405 WARN_ONCE(delta > (1ULL << 59),
2406 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2407 (unsigned long long)delta,
2408 (unsigned long long)ts,
2409 (unsigned long long)cpu_buffer->write_stamp,
2410 local_clock_stable ? "" :
2411 "If you just came from a suspend/resume,\n"
2412 "please switch to the trace global clock:\n"
2413 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2418 event = __rb_reserve_next(cpu_buffer, length, ts,
2419 delta, add_timestamp);
2420 if (unlikely(PTR_ERR(event) == -EAGAIN))
2429 rb_end_commit(cpu_buffer);
2433 #ifdef CONFIG_TRACING
2436 * The lock and unlock are done within a preempt disable section.
2437 * The current_context per_cpu variable can only be modified
2438 * by the current task between lock and unlock. But it can
2439 * be modified more than once via an interrupt. To pass this
2440 * information from the lock to the unlock without having to
2441 * access the 'in_interrupt()' functions again (which do show
2442 * a bit of overhead in something as critical as function tracing,
2443 * we use a bitmask trick.
2445 * bit 0 = NMI context
2446 * bit 1 = IRQ context
2447 * bit 2 = SoftIRQ context
2448 * bit 3 = normal context.
2450 * This works because this is the order of contexts that can
2451 * preempt other contexts. A SoftIRQ never preempts an IRQ
2454 * When the context is determined, the corresponding bit is
2455 * checked and set (if it was set, then a recursion of that context
2458 * On unlock, we need to clear this bit. To do so, just subtract
2459 * 1 from the current_context and AND it to itself.
2463 * 101 & 100 = 100 (clearing bit zero)
2466 * 1010 & 1001 = 1000 (clearing bit 1)
2468 * The least significant bit can be cleared this way, and it
2469 * just so happens that it is the same bit corresponding to
2470 * the current context.
2472 static DEFINE_PER_CPU(unsigned int, current_context);
2474 static __always_inline int trace_recursive_lock(void)
2476 unsigned int val = this_cpu_read(current_context);
2479 if (in_interrupt()) {
2489 if (unlikely(val & (1 << bit)))
2493 this_cpu_write(current_context, val);
2498 static __always_inline void trace_recursive_unlock(void)
2500 unsigned int val = this_cpu_read(current_context);
2503 val &= this_cpu_read(current_context);
2504 this_cpu_write(current_context, val);
2509 #define trace_recursive_lock() (0)
2510 #define trace_recursive_unlock() do { } while (0)
2515 * ring_buffer_lock_reserve - reserve a part of the buffer
2516 * @buffer: the ring buffer to reserve from
2517 * @length: the length of the data to reserve (excluding event header)
2519 * Returns a reseverd event on the ring buffer to copy directly to.
2520 * The user of this interface will need to get the body to write into
2521 * and can use the ring_buffer_event_data() interface.
2523 * The length is the length of the data needed, not the event length
2524 * which also includes the event header.
2526 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2527 * If NULL is returned, then nothing has been allocated or locked.
2529 struct ring_buffer_event *
2530 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2532 struct ring_buffer_per_cpu *cpu_buffer;
2533 struct ring_buffer_event *event;
2536 if (ring_buffer_flags != RB_BUFFERS_ON)
2539 /* If we are tracing schedule, we don't want to recurse */
2540 preempt_disable_notrace();
2542 if (atomic_read(&buffer->record_disabled))
2545 if (trace_recursive_lock())
2548 cpu = raw_smp_processor_id();
2550 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2553 cpu_buffer = buffer->buffers[cpu];
2555 if (atomic_read(&cpu_buffer->record_disabled))
2558 if (length > BUF_MAX_DATA_SIZE)
2561 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2568 trace_recursive_unlock();
2571 preempt_enable_notrace();
2574 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2577 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2578 struct ring_buffer_event *event)
2583 * The event first in the commit queue updates the
2586 if (rb_event_is_commit(cpu_buffer, event)) {
2588 * A commit event that is first on a page
2589 * updates the write timestamp with the page stamp
2591 if (!rb_event_index(event))
2592 cpu_buffer->write_stamp =
2593 cpu_buffer->commit_page->page->time_stamp;
2594 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2595 delta = event->array[0];
2597 delta += event->time_delta;
2598 cpu_buffer->write_stamp += delta;
2600 cpu_buffer->write_stamp += event->time_delta;
2604 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2605 struct ring_buffer_event *event)
2607 local_inc(&cpu_buffer->entries);
2608 rb_update_write_stamp(cpu_buffer, event);
2609 rb_end_commit(cpu_buffer);
2613 * ring_buffer_unlock_commit - commit a reserved
2614 * @buffer: The buffer to commit to
2615 * @event: The event pointer to commit.
2617 * This commits the data to the ring buffer, and releases any locks held.
2619 * Must be paired with ring_buffer_lock_reserve.
2621 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2622 struct ring_buffer_event *event)
2624 struct ring_buffer_per_cpu *cpu_buffer;
2625 int cpu = raw_smp_processor_id();
2627 cpu_buffer = buffer->buffers[cpu];
2629 rb_commit(cpu_buffer, event);
2631 trace_recursive_unlock();
2633 preempt_enable_notrace();
2637 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2639 static inline void rb_event_discard(struct ring_buffer_event *event)
2641 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2642 event = skip_time_extend(event);
2644 /* array[0] holds the actual length for the discarded event */
2645 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2646 event->type_len = RINGBUF_TYPE_PADDING;
2647 /* time delta must be non zero */
2648 if (!event->time_delta)
2649 event->time_delta = 1;
2653 * Decrement the entries to the page that an event is on.
2654 * The event does not even need to exist, only the pointer
2655 * to the page it is on. This may only be called before the commit
2659 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2660 struct ring_buffer_event *event)
2662 unsigned long addr = (unsigned long)event;
2663 struct buffer_page *bpage = cpu_buffer->commit_page;
2664 struct buffer_page *start;
2668 /* Do the likely case first */
2669 if (likely(bpage->page == (void *)addr)) {
2670 local_dec(&bpage->entries);
2675 * Because the commit page may be on the reader page we
2676 * start with the next page and check the end loop there.
2678 rb_inc_page(cpu_buffer, &bpage);
2681 if (bpage->page == (void *)addr) {
2682 local_dec(&bpage->entries);
2685 rb_inc_page(cpu_buffer, &bpage);
2686 } while (bpage != start);
2688 /* commit not part of this buffer?? */
2689 RB_WARN_ON(cpu_buffer, 1);
2693 * ring_buffer_commit_discard - discard an event that has not been committed
2694 * @buffer: the ring buffer
2695 * @event: non committed event to discard
2697 * Sometimes an event that is in the ring buffer needs to be ignored.
2698 * This function lets the user discard an event in the ring buffer
2699 * and then that event will not be read later.
2701 * This function only works if it is called before the the item has been
2702 * committed. It will try to free the event from the ring buffer
2703 * if another event has not been added behind it.
2705 * If another event has been added behind it, it will set the event
2706 * up as discarded, and perform the commit.
2708 * If this function is called, do not call ring_buffer_unlock_commit on
2711 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2712 struct ring_buffer_event *event)
2714 struct ring_buffer_per_cpu *cpu_buffer;
2717 /* The event is discarded regardless */
2718 rb_event_discard(event);
2720 cpu = smp_processor_id();
2721 cpu_buffer = buffer->buffers[cpu];
2724 * This must only be called if the event has not been
2725 * committed yet. Thus we can assume that preemption
2726 * is still disabled.
2728 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2730 rb_decrement_entry(cpu_buffer, event);
2731 if (rb_try_to_discard(cpu_buffer, event))
2735 * The commit is still visible by the reader, so we
2736 * must still update the timestamp.
2738 rb_update_write_stamp(cpu_buffer, event);
2740 rb_end_commit(cpu_buffer);
2742 trace_recursive_unlock();
2744 preempt_enable_notrace();
2747 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2750 * ring_buffer_write - write data to the buffer without reserving
2751 * @buffer: The ring buffer to write to.
2752 * @length: The length of the data being written (excluding the event header)
2753 * @data: The data to write to the buffer.
2755 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2756 * one function. If you already have the data to write to the buffer, it
2757 * may be easier to simply call this function.
2759 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2760 * and not the length of the event which would hold the header.
2762 int ring_buffer_write(struct ring_buffer *buffer,
2763 unsigned long length,
2766 struct ring_buffer_per_cpu *cpu_buffer;
2767 struct ring_buffer_event *event;
2772 if (ring_buffer_flags != RB_BUFFERS_ON)
2775 preempt_disable_notrace();
2777 if (atomic_read(&buffer->record_disabled))
2780 cpu = raw_smp_processor_id();
2782 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2785 cpu_buffer = buffer->buffers[cpu];
2787 if (atomic_read(&cpu_buffer->record_disabled))
2790 if (length > BUF_MAX_DATA_SIZE)
2793 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2797 body = rb_event_data(event);
2799 memcpy(body, data, length);
2801 rb_commit(cpu_buffer, event);
2805 preempt_enable_notrace();
2809 EXPORT_SYMBOL_GPL(ring_buffer_write);
2811 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
2813 struct buffer_page *reader = cpu_buffer->reader_page;
2814 struct buffer_page *head = rb_set_head_page(cpu_buffer);
2815 struct buffer_page *commit = cpu_buffer->commit_page;
2817 /* In case of error, head will be NULL */
2818 if (unlikely(!head))
2821 return reader->read == rb_page_commit(reader) &&
2822 (commit == reader ||
2824 head->read == rb_page_commit(commit)));
2828 * ring_buffer_record_disable - stop all writes into the buffer
2829 * @buffer: The ring buffer to stop writes to.
2831 * This prevents all writes to the buffer. Any attempt to write
2832 * to the buffer after this will fail and return NULL.
2834 * The caller should call synchronize_sched() after this.
2836 void ring_buffer_record_disable(struct ring_buffer *buffer)
2838 atomic_inc(&buffer->record_disabled);
2840 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
2843 * ring_buffer_record_enable - enable writes to the buffer
2844 * @buffer: The ring buffer to enable writes
2846 * Note, multiple disables will need the same number of enables
2847 * to truly enable the writing (much like preempt_disable).
2849 void ring_buffer_record_enable(struct ring_buffer *buffer)
2851 atomic_dec(&buffer->record_disabled);
2853 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
2856 * ring_buffer_record_off - stop all writes into the buffer
2857 * @buffer: The ring buffer to stop writes to.
2859 * This prevents all writes to the buffer. Any attempt to write
2860 * to the buffer after this will fail and return NULL.
2862 * This is different than ring_buffer_record_disable() as
2863 * it works like an on/off switch, where as the disable() version
2864 * must be paired with a enable().
2866 void ring_buffer_record_off(struct ring_buffer *buffer)
2869 unsigned int new_rd;
2872 rd = atomic_read(&buffer->record_disabled);
2873 new_rd = rd | RB_BUFFER_OFF;
2874 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
2876 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
2879 * ring_buffer_record_on - restart writes into the buffer
2880 * @buffer: The ring buffer to start writes to.
2882 * This enables all writes to the buffer that was disabled by
2883 * ring_buffer_record_off().
2885 * This is different than ring_buffer_record_enable() as
2886 * it works like an on/off switch, where as the enable() version
2887 * must be paired with a disable().
2889 void ring_buffer_record_on(struct ring_buffer *buffer)
2892 unsigned int new_rd;
2895 rd = atomic_read(&buffer->record_disabled);
2896 new_rd = rd & ~RB_BUFFER_OFF;
2897 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
2899 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
2902 * ring_buffer_record_is_on - return true if the ring buffer can write
2903 * @buffer: The ring buffer to see if write is enabled
2905 * Returns true if the ring buffer is in a state that it accepts writes.
2907 int ring_buffer_record_is_on(struct ring_buffer *buffer)
2909 return !atomic_read(&buffer->record_disabled);
2913 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
2914 * @buffer: The ring buffer to stop writes to.
2915 * @cpu: The CPU buffer to stop
2917 * This prevents all writes to the buffer. Any attempt to write
2918 * to the buffer after this will fail and return NULL.
2920 * The caller should call synchronize_sched() after this.
2922 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
2924 struct ring_buffer_per_cpu *cpu_buffer;
2926 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2929 cpu_buffer = buffer->buffers[cpu];
2930 atomic_inc(&cpu_buffer->record_disabled);
2932 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
2935 * ring_buffer_record_enable_cpu - enable writes to the buffer
2936 * @buffer: The ring buffer to enable writes
2937 * @cpu: The CPU to enable.
2939 * Note, multiple disables will need the same number of enables
2940 * to truly enable the writing (much like preempt_disable).
2942 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
2944 struct ring_buffer_per_cpu *cpu_buffer;
2946 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2949 cpu_buffer = buffer->buffers[cpu];
2950 atomic_dec(&cpu_buffer->record_disabled);
2952 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
2955 * The total entries in the ring buffer is the running counter
2956 * of entries entered into the ring buffer, minus the sum of
2957 * the entries read from the ring buffer and the number of
2958 * entries that were overwritten.
2960 static inline unsigned long
2961 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
2963 return local_read(&cpu_buffer->entries) -
2964 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
2968 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
2969 * @buffer: The ring buffer
2970 * @cpu: The per CPU buffer to read from.
2972 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
2974 unsigned long flags;
2975 struct ring_buffer_per_cpu *cpu_buffer;
2976 struct buffer_page *bpage;
2979 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2982 cpu_buffer = buffer->buffers[cpu];
2983 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2985 * if the tail is on reader_page, oldest time stamp is on the reader
2988 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
2989 bpage = cpu_buffer->reader_page;
2991 bpage = rb_set_head_page(cpu_buffer);
2993 ret = bpage->page->time_stamp;
2994 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2998 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3001 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3002 * @buffer: The ring buffer
3003 * @cpu: The per CPU buffer to read from.
3005 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3007 struct ring_buffer_per_cpu *cpu_buffer;
3010 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3013 cpu_buffer = buffer->buffers[cpu];
3014 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3018 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3021 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3022 * @buffer: The ring buffer
3023 * @cpu: The per CPU buffer to get the entries from.
3025 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3027 struct ring_buffer_per_cpu *cpu_buffer;
3029 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3032 cpu_buffer = buffer->buffers[cpu];
3034 return rb_num_of_entries(cpu_buffer);
3036 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3039 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3040 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3041 * @buffer: The ring buffer
3042 * @cpu: The per CPU buffer to get the number of overruns from
3044 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3046 struct ring_buffer_per_cpu *cpu_buffer;
3049 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3052 cpu_buffer = buffer->buffers[cpu];
3053 ret = local_read(&cpu_buffer->overrun);
3057 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3060 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3061 * commits failing due to the buffer wrapping around while there are uncommitted
3062 * events, such as during an interrupt storm.
3063 * @buffer: The ring buffer
3064 * @cpu: The per CPU buffer to get the number of overruns from
3067 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3069 struct ring_buffer_per_cpu *cpu_buffer;
3072 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3075 cpu_buffer = buffer->buffers[cpu];
3076 ret = local_read(&cpu_buffer->commit_overrun);
3080 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3083 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3084 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3085 * @buffer: The ring buffer
3086 * @cpu: The per CPU buffer to get the number of overruns from
3089 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3091 struct ring_buffer_per_cpu *cpu_buffer;
3094 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3097 cpu_buffer = buffer->buffers[cpu];
3098 ret = local_read(&cpu_buffer->dropped_events);
3102 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3105 * ring_buffer_entries - get the number of entries in a buffer
3106 * @buffer: The ring buffer
3108 * Returns the total number of entries in the ring buffer
3111 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3113 struct ring_buffer_per_cpu *cpu_buffer;
3114 unsigned long entries = 0;
3117 /* if you care about this being correct, lock the buffer */
3118 for_each_buffer_cpu(buffer, cpu) {
3119 cpu_buffer = buffer->buffers[cpu];
3120 entries += rb_num_of_entries(cpu_buffer);
3125 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3128 * ring_buffer_overruns - get the number of overruns in buffer
3129 * @buffer: The ring buffer
3131 * Returns the total number of overruns in the ring buffer
3134 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3136 struct ring_buffer_per_cpu *cpu_buffer;
3137 unsigned long overruns = 0;
3140 /* if you care about this being correct, lock the buffer */
3141 for_each_buffer_cpu(buffer, cpu) {
3142 cpu_buffer = buffer->buffers[cpu];
3143 overruns += local_read(&cpu_buffer->overrun);
3148 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3150 static void rb_iter_reset(struct ring_buffer_iter *iter)
3152 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3154 /* Iterator usage is expected to have record disabled */
3155 if (list_empty(&cpu_buffer->reader_page->list)) {
3156 iter->head_page = rb_set_head_page(cpu_buffer);
3157 if (unlikely(!iter->head_page))
3159 iter->head = iter->head_page->read;
3161 iter->head_page = cpu_buffer->reader_page;
3162 iter->head = cpu_buffer->reader_page->read;
3165 iter->read_stamp = cpu_buffer->read_stamp;
3167 iter->read_stamp = iter->head_page->page->time_stamp;
3168 iter->cache_reader_page = cpu_buffer->reader_page;
3169 iter->cache_read = cpu_buffer->read;
3173 * ring_buffer_iter_reset - reset an iterator
3174 * @iter: The iterator to reset
3176 * Resets the iterator, so that it will start from the beginning
3179 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3181 struct ring_buffer_per_cpu *cpu_buffer;
3182 unsigned long flags;
3187 cpu_buffer = iter->cpu_buffer;
3189 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3190 rb_iter_reset(iter);
3191 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3193 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3196 * ring_buffer_iter_empty - check if an iterator has no more to read
3197 * @iter: The iterator to check
3199 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3201 struct ring_buffer_per_cpu *cpu_buffer;
3203 cpu_buffer = iter->cpu_buffer;
3205 return iter->head_page == cpu_buffer->commit_page &&
3206 iter->head == rb_commit_index(cpu_buffer);
3208 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3211 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3212 struct ring_buffer_event *event)
3216 switch (event->type_len) {
3217 case RINGBUF_TYPE_PADDING:
3220 case RINGBUF_TYPE_TIME_EXTEND:
3221 delta = event->array[0];
3223 delta += event->time_delta;
3224 cpu_buffer->read_stamp += delta;
3227 case RINGBUF_TYPE_TIME_STAMP:
3228 /* FIXME: not implemented */
3231 case RINGBUF_TYPE_DATA:
3232 cpu_buffer->read_stamp += event->time_delta;
3242 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3243 struct ring_buffer_event *event)
3247 switch (event->type_len) {
3248 case RINGBUF_TYPE_PADDING:
3251 case RINGBUF_TYPE_TIME_EXTEND:
3252 delta = event->array[0];
3254 delta += event->time_delta;
3255 iter->read_stamp += delta;
3258 case RINGBUF_TYPE_TIME_STAMP:
3259 /* FIXME: not implemented */
3262 case RINGBUF_TYPE_DATA:
3263 iter->read_stamp += event->time_delta;
3272 static struct buffer_page *
3273 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3275 struct buffer_page *reader = NULL;
3276 unsigned long overwrite;
3277 unsigned long flags;
3281 local_irq_save(flags);
3282 arch_spin_lock(&cpu_buffer->lock);
3286 * This should normally only loop twice. But because the
3287 * start of the reader inserts an empty page, it causes
3288 * a case where we will loop three times. There should be no
3289 * reason to loop four times (that I know of).
3291 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3296 reader = cpu_buffer->reader_page;
3298 /* If there's more to read, return this page */
3299 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3302 /* Never should we have an index greater than the size */
3303 if (RB_WARN_ON(cpu_buffer,
3304 cpu_buffer->reader_page->read > rb_page_size(reader)))
3307 /* check if we caught up to the tail */
3309 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3312 /* Don't bother swapping if the ring buffer is empty */
3313 if (rb_num_of_entries(cpu_buffer) == 0)
3317 * Reset the reader page to size zero.
3319 local_set(&cpu_buffer->reader_page->write, 0);
3320 local_set(&cpu_buffer->reader_page->entries, 0);
3321 local_set(&cpu_buffer->reader_page->page->commit, 0);
3322 cpu_buffer->reader_page->real_end = 0;
3326 * Splice the empty reader page into the list around the head.
3328 reader = rb_set_head_page(cpu_buffer);
3331 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3332 cpu_buffer->reader_page->list.prev = reader->list.prev;
3335 * cpu_buffer->pages just needs to point to the buffer, it
3336 * has no specific buffer page to point to. Lets move it out
3337 * of our way so we don't accidentally swap it.
3339 cpu_buffer->pages = reader->list.prev;
3341 /* The reader page will be pointing to the new head */
3342 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3345 * We want to make sure we read the overruns after we set up our
3346 * pointers to the next object. The writer side does a
3347 * cmpxchg to cross pages which acts as the mb on the writer
3348 * side. Note, the reader will constantly fail the swap
3349 * while the writer is updating the pointers, so this
3350 * guarantees that the overwrite recorded here is the one we
3351 * want to compare with the last_overrun.
3354 overwrite = local_read(&(cpu_buffer->overrun));
3357 * Here's the tricky part.
3359 * We need to move the pointer past the header page.
3360 * But we can only do that if a writer is not currently
3361 * moving it. The page before the header page has the
3362 * flag bit '1' set if it is pointing to the page we want.
3363 * but if the writer is in the process of moving it
3364 * than it will be '2' or already moved '0'.
3367 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3370 * If we did not convert it, then we must try again.
3376 * Yeah! We succeeded in replacing the page.
3378 * Now make the new head point back to the reader page.
3380 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3381 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3383 /* Finally update the reader page to the new head */
3384 cpu_buffer->reader_page = reader;
3385 rb_reset_reader_page(cpu_buffer);
3387 if (overwrite != cpu_buffer->last_overrun) {
3388 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3389 cpu_buffer->last_overrun = overwrite;
3395 arch_spin_unlock(&cpu_buffer->lock);
3396 local_irq_restore(flags);
3401 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3403 struct ring_buffer_event *event;
3404 struct buffer_page *reader;
3407 reader = rb_get_reader_page(cpu_buffer);
3409 /* This function should not be called when buffer is empty */
3410 if (RB_WARN_ON(cpu_buffer, !reader))
3413 event = rb_reader_event(cpu_buffer);
3415 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3418 rb_update_read_stamp(cpu_buffer, event);
3420 length = rb_event_length(event);
3421 cpu_buffer->reader_page->read += length;
3424 static void rb_advance_iter(struct ring_buffer_iter *iter)
3426 struct ring_buffer_per_cpu *cpu_buffer;
3427 struct ring_buffer_event *event;
3430 cpu_buffer = iter->cpu_buffer;
3433 * Check if we are at the end of the buffer.
3435 if (iter->head >= rb_page_size(iter->head_page)) {
3436 /* discarded commits can make the page empty */
3437 if (iter->head_page == cpu_buffer->commit_page)
3443 event = rb_iter_head_event(iter);
3445 length = rb_event_length(event);
3448 * This should not be called to advance the header if we are
3449 * at the tail of the buffer.
3451 if (RB_WARN_ON(cpu_buffer,
3452 (iter->head_page == cpu_buffer->commit_page) &&
3453 (iter->head + length > rb_commit_index(cpu_buffer))))
3456 rb_update_iter_read_stamp(iter, event);
3458 iter->head += length;
3460 /* check for end of page padding */
3461 if ((iter->head >= rb_page_size(iter->head_page)) &&
3462 (iter->head_page != cpu_buffer->commit_page))
3466 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3468 return cpu_buffer->lost_events;
3471 static struct ring_buffer_event *
3472 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3473 unsigned long *lost_events)
3475 struct ring_buffer_event *event;
3476 struct buffer_page *reader;
3481 * We repeat when a time extend is encountered.
3482 * Since the time extend is always attached to a data event,
3483 * we should never loop more than once.
3484 * (We never hit the following condition more than twice).
3486 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3489 reader = rb_get_reader_page(cpu_buffer);
3493 event = rb_reader_event(cpu_buffer);
3495 switch (event->type_len) {
3496 case RINGBUF_TYPE_PADDING:
3497 if (rb_null_event(event))
3498 RB_WARN_ON(cpu_buffer, 1);
3500 * Because the writer could be discarding every
3501 * event it creates (which would probably be bad)
3502 * if we were to go back to "again" then we may never
3503 * catch up, and will trigger the warn on, or lock
3504 * the box. Return the padding, and we will release
3505 * the current locks, and try again.
3509 case RINGBUF_TYPE_TIME_EXTEND:
3510 /* Internal data, OK to advance */
3511 rb_advance_reader(cpu_buffer);
3514 case RINGBUF_TYPE_TIME_STAMP:
3515 /* FIXME: not implemented */
3516 rb_advance_reader(cpu_buffer);
3519 case RINGBUF_TYPE_DATA:
3521 *ts = cpu_buffer->read_stamp + event->time_delta;
3522 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3523 cpu_buffer->cpu, ts);
3526 *lost_events = rb_lost_events(cpu_buffer);
3535 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3537 static struct ring_buffer_event *
3538 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3540 struct ring_buffer *buffer;
3541 struct ring_buffer_per_cpu *cpu_buffer;
3542 struct ring_buffer_event *event;
3545 cpu_buffer = iter->cpu_buffer;
3546 buffer = cpu_buffer->buffer;
3549 * Check if someone performed a consuming read to
3550 * the buffer. A consuming read invalidates the iterator
3551 * and we need to reset the iterator in this case.
3553 if (unlikely(iter->cache_read != cpu_buffer->read ||
3554 iter->cache_reader_page != cpu_buffer->reader_page))
3555 rb_iter_reset(iter);
3558 if (ring_buffer_iter_empty(iter))
3562 * We repeat when a time extend is encountered.
3563 * Since the time extend is always attached to a data event,
3564 * we should never loop more than once.
3565 * (We never hit the following condition more than twice).
3567 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3570 if (rb_per_cpu_empty(cpu_buffer))
3573 if (iter->head >= local_read(&iter->head_page->page->commit)) {
3578 event = rb_iter_head_event(iter);
3580 switch (event->type_len) {
3581 case RINGBUF_TYPE_PADDING:
3582 if (rb_null_event(event)) {
3586 rb_advance_iter(iter);
3589 case RINGBUF_TYPE_TIME_EXTEND:
3590 /* Internal data, OK to advance */
3591 rb_advance_iter(iter);
3594 case RINGBUF_TYPE_TIME_STAMP:
3595 /* FIXME: not implemented */
3596 rb_advance_iter(iter);
3599 case RINGBUF_TYPE_DATA:
3601 *ts = iter->read_stamp + event->time_delta;
3602 ring_buffer_normalize_time_stamp(buffer,
3603 cpu_buffer->cpu, ts);
3613 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3615 static inline int rb_ok_to_lock(void)
3618 * If an NMI die dumps out the content of the ring buffer
3619 * do not grab locks. We also permanently disable the ring
3620 * buffer too. A one time deal is all you get from reading
3621 * the ring buffer from an NMI.
3623 if (likely(!in_nmi()))
3626 tracing_off_permanent();
3631 * ring_buffer_peek - peek at the next event to be read
3632 * @buffer: The ring buffer to read
3633 * @cpu: The cpu to peak at
3634 * @ts: The timestamp counter of this event.
3635 * @lost_events: a variable to store if events were lost (may be NULL)
3637 * This will return the event that will be read next, but does
3638 * not consume the data.
3640 struct ring_buffer_event *
3641 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3642 unsigned long *lost_events)
3644 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3645 struct ring_buffer_event *event;
3646 unsigned long flags;
3649 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3652 dolock = rb_ok_to_lock();
3654 local_irq_save(flags);
3656 raw_spin_lock(&cpu_buffer->reader_lock);
3657 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3658 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3659 rb_advance_reader(cpu_buffer);
3661 raw_spin_unlock(&cpu_buffer->reader_lock);
3662 local_irq_restore(flags);
3664 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3671 * ring_buffer_iter_peek - peek at the next event to be read
3672 * @iter: The ring buffer iterator
3673 * @ts: The timestamp counter of this event.
3675 * This will return the event that will be read next, but does
3676 * not increment the iterator.
3678 struct ring_buffer_event *
3679 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3681 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3682 struct ring_buffer_event *event;
3683 unsigned long flags;
3686 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3687 event = rb_iter_peek(iter, ts);
3688 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3690 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3697 * ring_buffer_consume - return an event and consume it
3698 * @buffer: The ring buffer to get the next event from
3699 * @cpu: the cpu to read the buffer from
3700 * @ts: a variable to store the timestamp (may be NULL)
3701 * @lost_events: a variable to store if events were lost (may be NULL)
3703 * Returns the next event in the ring buffer, and that event is consumed.
3704 * Meaning, that sequential reads will keep returning a different event,
3705 * and eventually empty the ring buffer if the producer is slower.
3707 struct ring_buffer_event *
3708 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3709 unsigned long *lost_events)
3711 struct ring_buffer_per_cpu *cpu_buffer;
3712 struct ring_buffer_event *event = NULL;
3713 unsigned long flags;
3716 dolock = rb_ok_to_lock();
3719 /* might be called in atomic */
3722 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3725 cpu_buffer = buffer->buffers[cpu];
3726 local_irq_save(flags);
3728 raw_spin_lock(&cpu_buffer->reader_lock);
3730 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3732 cpu_buffer->lost_events = 0;
3733 rb_advance_reader(cpu_buffer);
3737 raw_spin_unlock(&cpu_buffer->reader_lock);
3738 local_irq_restore(flags);
3743 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3748 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3751 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3752 * @buffer: The ring buffer to read from
3753 * @cpu: The cpu buffer to iterate over
3755 * This performs the initial preparations necessary to iterate
3756 * through the buffer. Memory is allocated, buffer recording
3757 * is disabled, and the iterator pointer is returned to the caller.
3759 * Disabling buffer recordng prevents the reading from being
3760 * corrupted. This is not a consuming read, so a producer is not
3763 * After a sequence of ring_buffer_read_prepare calls, the user is
3764 * expected to make at least one call to ring_buffer_prepare_sync.
3765 * Afterwards, ring_buffer_read_start is invoked to get things going
3768 * This overall must be paired with ring_buffer_finish.
3770 struct ring_buffer_iter *
3771 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3773 struct ring_buffer_per_cpu *cpu_buffer;
3774 struct ring_buffer_iter *iter;
3776 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3779 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3783 cpu_buffer = buffer->buffers[cpu];
3785 iter->cpu_buffer = cpu_buffer;
3787 atomic_inc(&buffer->resize_disabled);
3788 atomic_inc(&cpu_buffer->record_disabled);
3792 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
3795 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
3797 * All previously invoked ring_buffer_read_prepare calls to prepare
3798 * iterators will be synchronized. Afterwards, read_buffer_read_start
3799 * calls on those iterators are allowed.
3802 ring_buffer_read_prepare_sync(void)
3804 synchronize_sched();
3806 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
3809 * ring_buffer_read_start - start a non consuming read of the buffer
3810 * @iter: The iterator returned by ring_buffer_read_prepare
3812 * This finalizes the startup of an iteration through the buffer.
3813 * The iterator comes from a call to ring_buffer_read_prepare and
3814 * an intervening ring_buffer_read_prepare_sync must have been
3817 * Must be paired with ring_buffer_finish.
3820 ring_buffer_read_start(struct ring_buffer_iter *iter)
3822 struct ring_buffer_per_cpu *cpu_buffer;
3823 unsigned long flags;
3828 cpu_buffer = iter->cpu_buffer;
3830 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3831 arch_spin_lock(&cpu_buffer->lock);
3832 rb_iter_reset(iter);
3833 arch_spin_unlock(&cpu_buffer->lock);
3834 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3836 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
3839 * ring_buffer_finish - finish reading the iterator of the buffer
3840 * @iter: The iterator retrieved by ring_buffer_start
3842 * This re-enables the recording to the buffer, and frees the
3846 ring_buffer_read_finish(struct ring_buffer_iter *iter)
3848 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3849 unsigned long flags;
3852 * Ring buffer is disabled from recording, here's a good place
3853 * to check the integrity of the ring buffer.
3854 * Must prevent readers from trying to read, as the check
3855 * clears the HEAD page and readers require it.
3857 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3858 rb_check_pages(cpu_buffer);
3859 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3861 atomic_dec(&cpu_buffer->record_disabled);
3862 atomic_dec(&cpu_buffer->buffer->resize_disabled);
3865 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
3868 * ring_buffer_read - read the next item in the ring buffer by the iterator
3869 * @iter: The ring buffer iterator
3870 * @ts: The time stamp of the event read.
3872 * This reads the next event in the ring buffer and increments the iterator.
3874 struct ring_buffer_event *
3875 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
3877 struct ring_buffer_event *event;
3878 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3879 unsigned long flags;
3881 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3883 event = rb_iter_peek(iter, ts);
3887 if (event->type_len == RINGBUF_TYPE_PADDING)
3890 rb_advance_iter(iter);
3892 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3896 EXPORT_SYMBOL_GPL(ring_buffer_read);
3899 * ring_buffer_size - return the size of the ring buffer (in bytes)
3900 * @buffer: The ring buffer.
3902 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
3905 * Earlier, this method returned
3906 * BUF_PAGE_SIZE * buffer->nr_pages
3907 * Since the nr_pages field is now removed, we have converted this to
3908 * return the per cpu buffer value.
3910 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3913 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
3915 EXPORT_SYMBOL_GPL(ring_buffer_size);
3918 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
3920 rb_head_page_deactivate(cpu_buffer);
3922 cpu_buffer->head_page
3923 = list_entry(cpu_buffer->pages, struct buffer_page, list);
3924 local_set(&cpu_buffer->head_page->write, 0);
3925 local_set(&cpu_buffer->head_page->entries, 0);
3926 local_set(&cpu_buffer->head_page->page->commit, 0);
3928 cpu_buffer->head_page->read = 0;
3930 cpu_buffer->tail_page = cpu_buffer->head_page;
3931 cpu_buffer->commit_page = cpu_buffer->head_page;
3933 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
3934 INIT_LIST_HEAD(&cpu_buffer->new_pages);
3935 local_set(&cpu_buffer->reader_page->write, 0);
3936 local_set(&cpu_buffer->reader_page->entries, 0);
3937 local_set(&cpu_buffer->reader_page->page->commit, 0);
3938 cpu_buffer->reader_page->read = 0;
3940 local_set(&cpu_buffer->entries_bytes, 0);
3941 local_set(&cpu_buffer->overrun, 0);
3942 local_set(&cpu_buffer->commit_overrun, 0);
3943 local_set(&cpu_buffer->dropped_events, 0);
3944 local_set(&cpu_buffer->entries, 0);
3945 local_set(&cpu_buffer->committing, 0);
3946 local_set(&cpu_buffer->commits, 0);
3947 cpu_buffer->read = 0;
3948 cpu_buffer->read_bytes = 0;
3950 cpu_buffer->write_stamp = 0;
3951 cpu_buffer->read_stamp = 0;
3953 cpu_buffer->lost_events = 0;
3954 cpu_buffer->last_overrun = 0;
3956 rb_head_page_activate(cpu_buffer);
3960 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
3961 * @buffer: The ring buffer to reset a per cpu buffer of
3962 * @cpu: The CPU buffer to be reset
3964 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
3966 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3967 unsigned long flags;
3969 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3972 atomic_inc(&buffer->resize_disabled);
3973 atomic_inc(&cpu_buffer->record_disabled);
3975 /* Make sure all commits have finished */
3976 synchronize_sched();
3978 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3980 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
3983 arch_spin_lock(&cpu_buffer->lock);
3985 rb_reset_cpu(cpu_buffer);
3987 arch_spin_unlock(&cpu_buffer->lock);
3990 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3992 atomic_dec(&cpu_buffer->record_disabled);
3993 atomic_dec(&buffer->resize_disabled);
3995 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
3998 * ring_buffer_reset - reset a ring buffer
3999 * @buffer: The ring buffer to reset all cpu buffers
4001 void ring_buffer_reset(struct ring_buffer *buffer)
4005 for_each_buffer_cpu(buffer, cpu)
4006 ring_buffer_reset_cpu(buffer, cpu);
4008 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4011 * rind_buffer_empty - is the ring buffer empty?
4012 * @buffer: The ring buffer to test
4014 int ring_buffer_empty(struct ring_buffer *buffer)
4016 struct ring_buffer_per_cpu *cpu_buffer;
4017 unsigned long flags;
4022 dolock = rb_ok_to_lock();
4024 /* yes this is racy, but if you don't like the race, lock the buffer */
4025 for_each_buffer_cpu(buffer, cpu) {
4026 cpu_buffer = buffer->buffers[cpu];
4027 local_irq_save(flags);
4029 raw_spin_lock(&cpu_buffer->reader_lock);
4030 ret = rb_per_cpu_empty(cpu_buffer);
4032 raw_spin_unlock(&cpu_buffer->reader_lock);
4033 local_irq_restore(flags);
4041 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4044 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4045 * @buffer: The ring buffer
4046 * @cpu: The CPU buffer to test
4048 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4050 struct ring_buffer_per_cpu *cpu_buffer;
4051 unsigned long flags;
4055 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4058 dolock = rb_ok_to_lock();
4060 cpu_buffer = buffer->buffers[cpu];
4061 local_irq_save(flags);
4063 raw_spin_lock(&cpu_buffer->reader_lock);
4064 ret = rb_per_cpu_empty(cpu_buffer);
4066 raw_spin_unlock(&cpu_buffer->reader_lock);
4067 local_irq_restore(flags);
4071 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4073 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4075 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4076 * @buffer_a: One buffer to swap with
4077 * @buffer_b: The other buffer to swap with
4079 * This function is useful for tracers that want to take a "snapshot"
4080 * of a CPU buffer and has another back up buffer lying around.
4081 * it is expected that the tracer handles the cpu buffer not being
4082 * used at the moment.
4084 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4085 struct ring_buffer *buffer_b, int cpu)
4087 struct ring_buffer_per_cpu *cpu_buffer_a;
4088 struct ring_buffer_per_cpu *cpu_buffer_b;
4091 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4092 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4095 cpu_buffer_a = buffer_a->buffers[cpu];
4096 cpu_buffer_b = buffer_b->buffers[cpu];
4098 /* At least make sure the two buffers are somewhat the same */
4099 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4104 if (ring_buffer_flags != RB_BUFFERS_ON)
4107 if (atomic_read(&buffer_a->record_disabled))
4110 if (atomic_read(&buffer_b->record_disabled))
4113 if (atomic_read(&cpu_buffer_a->record_disabled))
4116 if (atomic_read(&cpu_buffer_b->record_disabled))
4120 * We can't do a synchronize_sched here because this
4121 * function can be called in atomic context.
4122 * Normally this will be called from the same CPU as cpu.
4123 * If not it's up to the caller to protect this.
4125 atomic_inc(&cpu_buffer_a->record_disabled);
4126 atomic_inc(&cpu_buffer_b->record_disabled);
4129 if (local_read(&cpu_buffer_a->committing))
4131 if (local_read(&cpu_buffer_b->committing))
4134 buffer_a->buffers[cpu] = cpu_buffer_b;
4135 buffer_b->buffers[cpu] = cpu_buffer_a;
4137 cpu_buffer_b->buffer = buffer_a;
4138 cpu_buffer_a->buffer = buffer_b;
4143 atomic_dec(&cpu_buffer_a->record_disabled);
4144 atomic_dec(&cpu_buffer_b->record_disabled);
4148 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4149 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4152 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4153 * @buffer: the buffer to allocate for.
4155 * This function is used in conjunction with ring_buffer_read_page.
4156 * When reading a full page from the ring buffer, these functions
4157 * can be used to speed up the process. The calling function should
4158 * allocate a few pages first with this function. Then when it
4159 * needs to get pages from the ring buffer, it passes the result
4160 * of this function into ring_buffer_read_page, which will swap
4161 * the page that was allocated, with the read page of the buffer.
4164 * The page allocated, or NULL on error.
4166 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4168 struct buffer_data_page *bpage;
4171 page = alloc_pages_node(cpu_to_node(cpu),
4172 GFP_KERNEL | __GFP_NORETRY, 0);
4176 bpage = page_address(page);
4178 rb_init_page(bpage);
4182 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4185 * ring_buffer_free_read_page - free an allocated read page
4186 * @buffer: the buffer the page was allocate for
4187 * @data: the page to free
4189 * Free a page allocated from ring_buffer_alloc_read_page.
4191 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4193 free_page((unsigned long)data);
4195 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4198 * ring_buffer_read_page - extract a page from the ring buffer
4199 * @buffer: buffer to extract from
4200 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4201 * @len: amount to extract
4202 * @cpu: the cpu of the buffer to extract
4203 * @full: should the extraction only happen when the page is full.
4205 * This function will pull out a page from the ring buffer and consume it.
4206 * @data_page must be the address of the variable that was returned
4207 * from ring_buffer_alloc_read_page. This is because the page might be used
4208 * to swap with a page in the ring buffer.
4211 * rpage = ring_buffer_alloc_read_page(buffer);
4214 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4216 * process_page(rpage, ret);
4218 * When @full is set, the function will not return true unless
4219 * the writer is off the reader page.
4221 * Note: it is up to the calling functions to handle sleeps and wakeups.
4222 * The ring buffer can be used anywhere in the kernel and can not
4223 * blindly call wake_up. The layer that uses the ring buffer must be
4224 * responsible for that.
4227 * >=0 if data has been transferred, returns the offset of consumed data.
4228 * <0 if no data has been transferred.
4230 int ring_buffer_read_page(struct ring_buffer *buffer,
4231 void **data_page, size_t len, int cpu, int full)
4233 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4234 struct ring_buffer_event *event;
4235 struct buffer_data_page *bpage;
4236 struct buffer_page *reader;
4237 unsigned long missed_events;
4238 unsigned long flags;
4239 unsigned int commit;
4244 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4248 * If len is not big enough to hold the page header, then
4249 * we can not copy anything.
4251 if (len <= BUF_PAGE_HDR_SIZE)
4254 len -= BUF_PAGE_HDR_SIZE;
4263 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4265 reader = rb_get_reader_page(cpu_buffer);
4269 event = rb_reader_event(cpu_buffer);
4271 read = reader->read;
4272 commit = rb_page_commit(reader);
4274 /* Check if any events were dropped */
4275 missed_events = cpu_buffer->lost_events;
4278 * If this page has been partially read or
4279 * if len is not big enough to read the rest of the page or
4280 * a writer is still on the page, then
4281 * we must copy the data from the page to the buffer.
4282 * Otherwise, we can simply swap the page with the one passed in.
4284 if (read || (len < (commit - read)) ||
4285 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4286 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4287 unsigned int rpos = read;
4288 unsigned int pos = 0;
4294 if (len > (commit - read))
4295 len = (commit - read);
4297 /* Always keep the time extend and data together */
4298 size = rb_event_ts_length(event);
4303 /* save the current timestamp, since the user will need it */
4304 save_timestamp = cpu_buffer->read_stamp;
4306 /* Need to copy one event at a time */
4308 /* We need the size of one event, because
4309 * rb_advance_reader only advances by one event,
4310 * whereas rb_event_ts_length may include the size of
4311 * one or two events.
4312 * We have already ensured there's enough space if this
4313 * is a time extend. */
4314 size = rb_event_length(event);
4315 memcpy(bpage->data + pos, rpage->data + rpos, size);
4319 rb_advance_reader(cpu_buffer);
4320 rpos = reader->read;
4326 event = rb_reader_event(cpu_buffer);
4327 /* Always keep the time extend and data together */
4328 size = rb_event_ts_length(event);
4329 } while (len >= size);
4332 local_set(&bpage->commit, pos);
4333 bpage->time_stamp = save_timestamp;
4335 /* we copied everything to the beginning */
4338 /* update the entry counter */
4339 cpu_buffer->read += rb_page_entries(reader);
4340 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4342 /* swap the pages */
4343 rb_init_page(bpage);
4344 bpage = reader->page;
4345 reader->page = *data_page;
4346 local_set(&reader->write, 0);
4347 local_set(&reader->entries, 0);
4352 * Use the real_end for the data size,
4353 * This gives us a chance to store the lost events
4356 if (reader->real_end)
4357 local_set(&bpage->commit, reader->real_end);
4361 cpu_buffer->lost_events = 0;
4363 commit = local_read(&bpage->commit);
4365 * Set a flag in the commit field if we lost events
4367 if (missed_events) {
4368 /* If there is room at the end of the page to save the
4369 * missed events, then record it there.
4371 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4372 memcpy(&bpage->data[commit], &missed_events,
4373 sizeof(missed_events));
4374 local_add(RB_MISSED_STORED, &bpage->commit);
4375 commit += sizeof(missed_events);
4377 local_add(RB_MISSED_EVENTS, &bpage->commit);
4381 * This page may be off to user land. Zero it out here.
4383 if (commit < BUF_PAGE_SIZE)
4384 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4387 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4392 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4394 #ifdef CONFIG_HOTPLUG_CPU
4395 static int rb_cpu_notify(struct notifier_block *self,
4396 unsigned long action, void *hcpu)
4398 struct ring_buffer *buffer =
4399 container_of(self, struct ring_buffer, cpu_notify);
4400 long cpu = (long)hcpu;
4401 int cpu_i, nr_pages_same;
4402 unsigned int nr_pages;
4405 case CPU_UP_PREPARE:
4406 case CPU_UP_PREPARE_FROZEN:
4407 if (cpumask_test_cpu(cpu, buffer->cpumask))
4412 /* check if all cpu sizes are same */
4413 for_each_buffer_cpu(buffer, cpu_i) {
4414 /* fill in the size from first enabled cpu */
4416 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4417 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4422 /* allocate minimum pages, user can later expand it */
4425 buffer->buffers[cpu] =
4426 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4427 if (!buffer->buffers[cpu]) {
4428 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4433 cpumask_set_cpu(cpu, buffer->cpumask);
4435 case CPU_DOWN_PREPARE:
4436 case CPU_DOWN_PREPARE_FROZEN:
4439 * If we were to free the buffer, then the user would
4440 * lose any trace that was in the buffer.