4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 #include <linux/ftrace_event.h>
7 #include <linux/ring_buffer.h>
8 #include <linux/trace_clock.h>
9 #include <linux/trace_seq.h>
10 #include <linux/spinlock.h>
11 #include <linux/irq_work.h>
12 #include <linux/debugfs.h>
13 #include <linux/uaccess.h>
14 #include <linux/hardirq.h>
15 #include <linux/kthread.h> /* for self test */
16 #include <linux/kmemcheck.h>
17 #include <linux/module.h>
18 #include <linux/percpu.h>
19 #include <linux/mutex.h>
20 #include <linux/delay.h>
21 #include <linux/slab.h>
22 #include <linux/init.h>
23 #include <linux/hash.h>
24 #include <linux/list.h>
25 #include <linux/cpu.h>
28 #include <asm/local.h>
30 static void update_pages_handler(struct work_struct *work);
33 * The ring buffer header is special. We must manually up keep it.
35 int ring_buffer_print_entry_header(struct trace_seq *s)
39 ret = trace_seq_puts(s, "# compressed entry header\n");
40 ret = trace_seq_puts(s, "\ttype_len : 5 bits\n");
41 ret = trace_seq_puts(s, "\ttime_delta : 27 bits\n");
42 ret = trace_seq_puts(s, "\tarray : 32 bits\n");
43 ret = trace_seq_putc(s, '\n');
44 ret = trace_seq_printf(s, "\tpadding : type == %d\n",
45 RINGBUF_TYPE_PADDING);
46 ret = trace_seq_printf(s, "\ttime_extend : type == %d\n",
47 RINGBUF_TYPE_TIME_EXTEND);
48 ret = trace_seq_printf(s, "\tdata max type_len == %d\n",
49 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
55 * The ring buffer is made up of a list of pages. A separate list of pages is
56 * allocated for each CPU. A writer may only write to a buffer that is
57 * associated with the CPU it is currently executing on. A reader may read
58 * from any per cpu buffer.
60 * The reader is special. For each per cpu buffer, the reader has its own
61 * reader page. When a reader has read the entire reader page, this reader
62 * page is swapped with another page in the ring buffer.
64 * Now, as long as the writer is off the reader page, the reader can do what
65 * ever it wants with that page. The writer will never write to that page
66 * again (as long as it is out of the ring buffer).
68 * Here's some silly ASCII art.
71 * |reader| RING BUFFER
73 * +------+ +---+ +---+ +---+
82 * |reader| RING BUFFER
83 * |page |------------------v
84 * +------+ +---+ +---+ +---+
93 * |reader| RING BUFFER
94 * |page |------------------v
95 * +------+ +---+ +---+ +---+
100 * +------------------------------+
104 * |buffer| RING BUFFER
105 * |page |------------------v
106 * +------+ +---+ +---+ +---+
108 * | New +---+ +---+ +---+
111 * +------------------------------+
114 * After we make this swap, the reader can hand this page off to the splice
115 * code and be done with it. It can even allocate a new page if it needs to
116 * and swap that into the ring buffer.
118 * We will be using cmpxchg soon to make all this lockless.
123 * A fast way to enable or disable all ring buffers is to
124 * call tracing_on or tracing_off. Turning off the ring buffers
125 * prevents all ring buffers from being recorded to.
126 * Turning this switch on, makes it OK to write to the
127 * ring buffer, if the ring buffer is enabled itself.
129 * There's three layers that must be on in order to write
130 * to the ring buffer.
132 * 1) This global flag must be set.
133 * 2) The ring buffer must be enabled for recording.
134 * 3) The per cpu buffer must be enabled for recording.
136 * In case of an anomaly, this global flag has a bit set that
137 * will permantly disable all ring buffers.
141 * Global flag to disable all recording to ring buffers
142 * This has two bits: ON, DISABLED
146 * 0 0 : ring buffers are off
147 * 1 0 : ring buffers are on
148 * X 1 : ring buffers are permanently disabled
152 RB_BUFFERS_ON_BIT = 0,
153 RB_BUFFERS_DISABLED_BIT = 1,
157 RB_BUFFERS_ON = 1 << RB_BUFFERS_ON_BIT,
158 RB_BUFFERS_DISABLED = 1 << RB_BUFFERS_DISABLED_BIT,
161 static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON;
163 /* Used for individual buffers (after the counter) */
164 #define RB_BUFFER_OFF (1 << 20)
166 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
169 * tracing_off_permanent - permanently disable ring buffers
171 * This function, once called, will disable all ring buffers
174 void tracing_off_permanent(void)
176 set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags);
179 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
180 #define RB_ALIGNMENT 4U
181 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
182 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
184 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
185 # define RB_FORCE_8BYTE_ALIGNMENT 0
186 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
188 # define RB_FORCE_8BYTE_ALIGNMENT 1
189 # define RB_ARCH_ALIGNMENT 8U
192 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
194 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
195 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
198 RB_LEN_TIME_EXTEND = 8,
199 RB_LEN_TIME_STAMP = 16,
202 #define skip_time_extend(event) \
203 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
205 static inline int rb_null_event(struct ring_buffer_event *event)
207 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
210 static void rb_event_set_padding(struct ring_buffer_event *event)
212 /* padding has a NULL time_delta */
213 event->type_len = RINGBUF_TYPE_PADDING;
214 event->time_delta = 0;
218 rb_event_data_length(struct ring_buffer_event *event)
223 length = event->type_len * RB_ALIGNMENT;
225 length = event->array[0];
226 return length + RB_EVNT_HDR_SIZE;
230 * Return the length of the given event. Will return
231 * the length of the time extend if the event is a
234 static inline unsigned
235 rb_event_length(struct ring_buffer_event *event)
237 switch (event->type_len) {
238 case RINGBUF_TYPE_PADDING:
239 if (rb_null_event(event))
242 return event->array[0] + RB_EVNT_HDR_SIZE;
244 case RINGBUF_TYPE_TIME_EXTEND:
245 return RB_LEN_TIME_EXTEND;
247 case RINGBUF_TYPE_TIME_STAMP:
248 return RB_LEN_TIME_STAMP;
250 case RINGBUF_TYPE_DATA:
251 return rb_event_data_length(event);
260 * Return total length of time extend and data,
261 * or just the event length for all other events.
263 static inline unsigned
264 rb_event_ts_length(struct ring_buffer_event *event)
268 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
269 /* time extends include the data event after it */
270 len = RB_LEN_TIME_EXTEND;
271 event = skip_time_extend(event);
273 return len + rb_event_length(event);
277 * ring_buffer_event_length - return the length of the event
278 * @event: the event to get the length of
280 * Returns the size of the data load of a data event.
281 * If the event is something other than a data event, it
282 * returns the size of the event itself. With the exception
283 * of a TIME EXTEND, where it still returns the size of the
284 * data load of the data event after it.
286 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
290 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
291 event = skip_time_extend(event);
293 length = rb_event_length(event);
294 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
296 length -= RB_EVNT_HDR_SIZE;
297 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
298 length -= sizeof(event->array[0]);
301 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
303 /* inline for ring buffer fast paths */
305 rb_event_data(struct ring_buffer_event *event)
307 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
308 event = skip_time_extend(event);
309 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
310 /* If length is in len field, then array[0] has the data */
312 return (void *)&event->array[0];
313 /* Otherwise length is in array[0] and array[1] has the data */
314 return (void *)&event->array[1];
318 * ring_buffer_event_data - return the data of the event
319 * @event: the event to get the data from
321 void *ring_buffer_event_data(struct ring_buffer_event *event)
323 return rb_event_data(event);
325 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
327 #define for_each_buffer_cpu(buffer, cpu) \
328 for_each_cpu(cpu, buffer->cpumask)
331 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
332 #define TS_DELTA_TEST (~TS_MASK)
334 /* Flag when events were overwritten */
335 #define RB_MISSED_EVENTS (1 << 31)
336 /* Missed count stored at end */
337 #define RB_MISSED_STORED (1 << 30)
339 struct buffer_data_page {
340 u64 time_stamp; /* page time stamp */
341 local_t commit; /* write committed index */
342 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
346 * Note, the buffer_page list must be first. The buffer pages
347 * are allocated in cache lines, which means that each buffer
348 * page will be at the beginning of a cache line, and thus
349 * the least significant bits will be zero. We use this to
350 * add flags in the list struct pointers, to make the ring buffer
354 struct list_head list; /* list of buffer pages */
355 local_t write; /* index for next write */
356 unsigned read; /* index for next read */
357 local_t entries; /* entries on this page */
358 unsigned long real_end; /* real end of data */
359 struct buffer_data_page *page; /* Actual data page */
363 * The buffer page counters, write and entries, must be reset
364 * atomically when crossing page boundaries. To synchronize this
365 * update, two counters are inserted into the number. One is
366 * the actual counter for the write position or count on the page.
368 * The other is a counter of updaters. Before an update happens
369 * the update partition of the counter is incremented. This will
370 * allow the updater to update the counter atomically.
372 * The counter is 20 bits, and the state data is 12.
374 #define RB_WRITE_MASK 0xfffff
375 #define RB_WRITE_INTCNT (1 << 20)
377 static void rb_init_page(struct buffer_data_page *bpage)
379 local_set(&bpage->commit, 0);
383 * ring_buffer_page_len - the size of data on the page.
384 * @page: The page to read
386 * Returns the amount of data on the page, including buffer page header.
388 size_t ring_buffer_page_len(void *page)
390 return local_read(&((struct buffer_data_page *)page)->commit)
395 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
398 static void free_buffer_page(struct buffer_page *bpage)
400 free_page((unsigned long)bpage->page);
405 * We need to fit the time_stamp delta into 27 bits.
407 static inline int test_time_stamp(u64 delta)
409 if (delta & TS_DELTA_TEST)
414 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
416 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
417 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
419 int ring_buffer_print_page_header(struct trace_seq *s)
421 struct buffer_data_page field;
424 ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t"
425 "offset:0;\tsize:%u;\tsigned:%u;\n",
426 (unsigned int)sizeof(field.time_stamp),
427 (unsigned int)is_signed_type(u64));
429 ret = trace_seq_printf(s, "\tfield: local_t commit;\t"
430 "offset:%u;\tsize:%u;\tsigned:%u;\n",
431 (unsigned int)offsetof(typeof(field), commit),
432 (unsigned int)sizeof(field.commit),
433 (unsigned int)is_signed_type(long));
435 ret = trace_seq_printf(s, "\tfield: int overwrite;\t"
436 "offset:%u;\tsize:%u;\tsigned:%u;\n",
437 (unsigned int)offsetof(typeof(field), commit),
439 (unsigned int)is_signed_type(long));
441 ret = trace_seq_printf(s, "\tfield: char data;\t"
442 "offset:%u;\tsize:%u;\tsigned:%u;\n",
443 (unsigned int)offsetof(typeof(field), data),
444 (unsigned int)BUF_PAGE_SIZE,
445 (unsigned int)is_signed_type(char));
451 struct irq_work work;
452 wait_queue_head_t waiters;
453 bool waiters_pending;
457 * head_page == tail_page && head == tail then buffer is empty.
459 struct ring_buffer_per_cpu {
461 atomic_t record_disabled;
462 struct ring_buffer *buffer;
463 raw_spinlock_t reader_lock; /* serialize readers */
464 arch_spinlock_t lock;
465 struct lock_class_key lock_key;
466 unsigned int nr_pages;
467 struct list_head *pages;
468 struct buffer_page *head_page; /* read from head */
469 struct buffer_page *tail_page; /* write to tail */
470 struct buffer_page *commit_page; /* committed pages */
471 struct buffer_page *reader_page;
472 unsigned long lost_events;
473 unsigned long last_overrun;
474 local_t entries_bytes;
477 local_t commit_overrun;
478 local_t dropped_events;
482 unsigned long read_bytes;
485 /* ring buffer pages to update, > 0 to add, < 0 to remove */
486 int nr_pages_to_update;
487 struct list_head new_pages; /* new pages to add */
488 struct work_struct update_pages_work;
489 struct completion update_done;
491 struct rb_irq_work irq_work;
497 atomic_t record_disabled;
498 atomic_t resize_disabled;
499 cpumask_var_t cpumask;
501 struct lock_class_key *reader_lock_key;
505 struct ring_buffer_per_cpu **buffers;
507 #ifdef CONFIG_HOTPLUG_CPU
508 struct notifier_block cpu_notify;
512 struct rb_irq_work irq_work;
515 struct ring_buffer_iter {
516 struct ring_buffer_per_cpu *cpu_buffer;
518 struct buffer_page *head_page;
519 struct buffer_page *cache_reader_page;
520 unsigned long cache_read;
525 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
527 * Schedules a delayed work to wake up any task that is blocked on the
528 * ring buffer waiters queue.
530 static void rb_wake_up_waiters(struct irq_work *work)
532 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
534 wake_up_all(&rbwork->waiters);
538 * ring_buffer_wait - wait for input to the ring buffer
539 * @buffer: buffer to wait on
540 * @cpu: the cpu buffer to wait on
542 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
543 * as data is added to any of the @buffer's cpu buffers. Otherwise
544 * it will wait for data to be added to a specific cpu buffer.
546 int ring_buffer_wait(struct ring_buffer *buffer, int cpu)
548 struct ring_buffer_per_cpu *cpu_buffer;
550 struct rb_irq_work *work;
553 * Depending on what the caller is waiting for, either any
554 * data in any cpu buffer, or a specific buffer, put the
555 * caller on the appropriate wait queue.
557 if (cpu == RING_BUFFER_ALL_CPUS)
558 work = &buffer->irq_work;
560 if (!cpumask_test_cpu(cpu, buffer->cpumask))
562 cpu_buffer = buffer->buffers[cpu];
563 work = &cpu_buffer->irq_work;
567 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
570 * The events can happen in critical sections where
571 * checking a work queue can cause deadlocks.
572 * After adding a task to the queue, this flag is set
573 * only to notify events to try to wake up the queue
576 * We don't clear it even if the buffer is no longer
577 * empty. The flag only causes the next event to run
578 * irq_work to do the work queue wake up. The worse
579 * that can happen if we race with !trace_empty() is that
580 * an event will cause an irq_work to try to wake up
583 * There's no reason to protect this flag either, as
584 * the work queue and irq_work logic will do the necessary
585 * synchronization for the wake ups. The only thing
586 * that is necessary is that the wake up happens after
587 * a task has been queued. It's OK for spurious wake ups.
589 work->waiters_pending = true;
591 if ((cpu == RING_BUFFER_ALL_CPUS && ring_buffer_empty(buffer)) ||
592 (cpu != RING_BUFFER_ALL_CPUS && ring_buffer_empty_cpu(buffer, cpu)))
595 finish_wait(&work->waiters, &wait);
600 * ring_buffer_poll_wait - poll on buffer input
601 * @buffer: buffer to wait on
602 * @cpu: the cpu buffer to wait on
603 * @filp: the file descriptor
604 * @poll_table: The poll descriptor
606 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
607 * as data is added to any of the @buffer's cpu buffers. Otherwise
608 * it will wait for data to be added to a specific cpu buffer.
610 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
613 int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
614 struct file *filp, poll_table *poll_table)
616 struct ring_buffer_per_cpu *cpu_buffer;
617 struct rb_irq_work *work;
619 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
620 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
621 return POLLIN | POLLRDNORM;
623 if (cpu == RING_BUFFER_ALL_CPUS)
624 work = &buffer->irq_work;
626 if (!cpumask_test_cpu(cpu, buffer->cpumask))
629 cpu_buffer = buffer->buffers[cpu];
630 work = &cpu_buffer->irq_work;
633 work->waiters_pending = true;
634 poll_wait(filp, &work->waiters, poll_table);
636 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
637 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
638 return POLLIN | POLLRDNORM;
642 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
643 #define RB_WARN_ON(b, cond) \
645 int _____ret = unlikely(cond); \
647 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
648 struct ring_buffer_per_cpu *__b = \
650 atomic_inc(&__b->buffer->record_disabled); \
652 atomic_inc(&b->record_disabled); \
658 /* Up this if you want to test the TIME_EXTENTS and normalization */
659 #define DEBUG_SHIFT 0
661 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
663 /* shift to debug/test normalization and TIME_EXTENTS */
664 return buffer->clock() << DEBUG_SHIFT;
667 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
671 preempt_disable_notrace();
672 time = rb_time_stamp(buffer);
673 preempt_enable_no_resched_notrace();
677 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
679 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
682 /* Just stupid testing the normalize function and deltas */
685 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
688 * Making the ring buffer lockless makes things tricky.
689 * Although writes only happen on the CPU that they are on,
690 * and they only need to worry about interrupts. Reads can
693 * The reader page is always off the ring buffer, but when the
694 * reader finishes with a page, it needs to swap its page with
695 * a new one from the buffer. The reader needs to take from
696 * the head (writes go to the tail). But if a writer is in overwrite
697 * mode and wraps, it must push the head page forward.
699 * Here lies the problem.
701 * The reader must be careful to replace only the head page, and
702 * not another one. As described at the top of the file in the
703 * ASCII art, the reader sets its old page to point to the next
704 * page after head. It then sets the page after head to point to
705 * the old reader page. But if the writer moves the head page
706 * during this operation, the reader could end up with the tail.
708 * We use cmpxchg to help prevent this race. We also do something
709 * special with the page before head. We set the LSB to 1.
711 * When the writer must push the page forward, it will clear the
712 * bit that points to the head page, move the head, and then set
713 * the bit that points to the new head page.
715 * We also don't want an interrupt coming in and moving the head
716 * page on another writer. Thus we use the second LSB to catch
719 * head->list->prev->next bit 1 bit 0
722 * Points to head page 0 1
725 * Note we can not trust the prev pointer of the head page, because:
727 * +----+ +-----+ +-----+
728 * | |------>| T |---X--->| N |
730 * +----+ +-----+ +-----+
733 * +----------| R |----------+ |
737 * Key: ---X--> HEAD flag set in pointer
742 * (see __rb_reserve_next() to see where this happens)
744 * What the above shows is that the reader just swapped out
745 * the reader page with a page in the buffer, but before it
746 * could make the new header point back to the new page added
747 * it was preempted by a writer. The writer moved forward onto
748 * the new page added by the reader and is about to move forward
751 * You can see, it is legitimate for the previous pointer of
752 * the head (or any page) not to point back to itself. But only
756 #define RB_PAGE_NORMAL 0UL
757 #define RB_PAGE_HEAD 1UL
758 #define RB_PAGE_UPDATE 2UL
761 #define RB_FLAG_MASK 3UL
763 /* PAGE_MOVED is not part of the mask */
764 #define RB_PAGE_MOVED 4UL
767 * rb_list_head - remove any bit
769 static struct list_head *rb_list_head(struct list_head *list)
771 unsigned long val = (unsigned long)list;
773 return (struct list_head *)(val & ~RB_FLAG_MASK);
777 * rb_is_head_page - test if the given page is the head page
779 * Because the reader may move the head_page pointer, we can
780 * not trust what the head page is (it may be pointing to
781 * the reader page). But if the next page is a header page,
782 * its flags will be non zero.
785 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
786 struct buffer_page *page, struct list_head *list)
790 val = (unsigned long)list->next;
792 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
793 return RB_PAGE_MOVED;
795 return val & RB_FLAG_MASK;
801 * The unique thing about the reader page, is that, if the
802 * writer is ever on it, the previous pointer never points
803 * back to the reader page.
805 static int rb_is_reader_page(struct buffer_page *page)
807 struct list_head *list = page->list.prev;
809 return rb_list_head(list->next) != &page->list;
813 * rb_set_list_to_head - set a list_head to be pointing to head.
815 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
816 struct list_head *list)
820 ptr = (unsigned long *)&list->next;
821 *ptr |= RB_PAGE_HEAD;
822 *ptr &= ~RB_PAGE_UPDATE;
826 * rb_head_page_activate - sets up head page
828 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
830 struct buffer_page *head;
832 head = cpu_buffer->head_page;
837 * Set the previous list pointer to have the HEAD flag.
839 rb_set_list_to_head(cpu_buffer, head->list.prev);
842 static void rb_list_head_clear(struct list_head *list)
844 unsigned long *ptr = (unsigned long *)&list->next;
846 *ptr &= ~RB_FLAG_MASK;
850 * rb_head_page_dactivate - clears head page ptr (for free list)
853 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
855 struct list_head *hd;
857 /* Go through the whole list and clear any pointers found. */
858 rb_list_head_clear(cpu_buffer->pages);
860 list_for_each(hd, cpu_buffer->pages)
861 rb_list_head_clear(hd);
864 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
865 struct buffer_page *head,
866 struct buffer_page *prev,
867 int old_flag, int new_flag)
869 struct list_head *list;
870 unsigned long val = (unsigned long)&head->list;
875 val &= ~RB_FLAG_MASK;
877 ret = cmpxchg((unsigned long *)&list->next,
878 val | old_flag, val | new_flag);
880 /* check if the reader took the page */
881 if ((ret & ~RB_FLAG_MASK) != val)
882 return RB_PAGE_MOVED;
884 return ret & RB_FLAG_MASK;
887 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
888 struct buffer_page *head,
889 struct buffer_page *prev,
892 return rb_head_page_set(cpu_buffer, head, prev,
893 old_flag, RB_PAGE_UPDATE);
896 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
897 struct buffer_page *head,
898 struct buffer_page *prev,
901 return rb_head_page_set(cpu_buffer, head, prev,
902 old_flag, RB_PAGE_HEAD);
905 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
906 struct buffer_page *head,
907 struct buffer_page *prev,
910 return rb_head_page_set(cpu_buffer, head, prev,
911 old_flag, RB_PAGE_NORMAL);
914 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
915 struct buffer_page **bpage)
917 struct list_head *p = rb_list_head((*bpage)->list.next);
919 *bpage = list_entry(p, struct buffer_page, list);
922 static struct buffer_page *
923 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
925 struct buffer_page *head;
926 struct buffer_page *page;
927 struct list_head *list;
930 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
934 list = cpu_buffer->pages;
935 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
938 page = head = cpu_buffer->head_page;
940 * It is possible that the writer moves the header behind
941 * where we started, and we miss in one loop.
942 * A second loop should grab the header, but we'll do
943 * three loops just because I'm paranoid.
945 for (i = 0; i < 3; i++) {
947 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
948 cpu_buffer->head_page = page;
951 rb_inc_page(cpu_buffer, &page);
952 } while (page != head);
955 RB_WARN_ON(cpu_buffer, 1);
960 static int rb_head_page_replace(struct buffer_page *old,
961 struct buffer_page *new)
963 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
967 val = *ptr & ~RB_FLAG_MASK;
970 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
976 * rb_tail_page_update - move the tail page forward
978 * Returns 1 if moved tail page, 0 if someone else did.
980 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
981 struct buffer_page *tail_page,
982 struct buffer_page *next_page)
984 struct buffer_page *old_tail;
985 unsigned long old_entries;
986 unsigned long old_write;
990 * The tail page now needs to be moved forward.
992 * We need to reset the tail page, but without messing
993 * with possible erasing of data brought in by interrupts
994 * that have moved the tail page and are currently on it.
996 * We add a counter to the write field to denote this.
998 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
999 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1002 * Just make sure we have seen our old_write and synchronize
1003 * with any interrupts that come in.
1008 * If the tail page is still the same as what we think
1009 * it is, then it is up to us to update the tail
1012 if (tail_page == cpu_buffer->tail_page) {
1013 /* Zero the write counter */
1014 unsigned long val = old_write & ~RB_WRITE_MASK;
1015 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1018 * This will only succeed if an interrupt did
1019 * not come in and change it. In which case, we
1020 * do not want to modify it.
1022 * We add (void) to let the compiler know that we do not care
1023 * about the return value of these functions. We use the
1024 * cmpxchg to only update if an interrupt did not already
1025 * do it for us. If the cmpxchg fails, we don't care.
1027 (void)local_cmpxchg(&next_page->write, old_write, val);
1028 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1031 * No need to worry about races with clearing out the commit.
1032 * it only can increment when a commit takes place. But that
1033 * only happens in the outer most nested commit.
1035 local_set(&next_page->page->commit, 0);
1037 old_tail = cmpxchg(&cpu_buffer->tail_page,
1038 tail_page, next_page);
1040 if (old_tail == tail_page)
1047 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1048 struct buffer_page *bpage)
1050 unsigned long val = (unsigned long)bpage;
1052 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1059 * rb_check_list - make sure a pointer to a list has the last bits zero
1061 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1062 struct list_head *list)
1064 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1066 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1072 * rb_check_pages - integrity check of buffer pages
1073 * @cpu_buffer: CPU buffer with pages to test
1075 * As a safety measure we check to make sure the data pages have not
1078 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1080 struct list_head *head = cpu_buffer->pages;
1081 struct buffer_page *bpage, *tmp;
1083 /* Reset the head page if it exists */
1084 if (cpu_buffer->head_page)
1085 rb_set_head_page(cpu_buffer);
1087 rb_head_page_deactivate(cpu_buffer);
1089 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1091 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1094 if (rb_check_list(cpu_buffer, head))
1097 list_for_each_entry_safe(bpage, tmp, head, list) {
1098 if (RB_WARN_ON(cpu_buffer,
1099 bpage->list.next->prev != &bpage->list))
1101 if (RB_WARN_ON(cpu_buffer,
1102 bpage->list.prev->next != &bpage->list))
1104 if (rb_check_list(cpu_buffer, &bpage->list))
1108 rb_head_page_activate(cpu_buffer);
1113 static int __rb_allocate_pages(int nr_pages, struct list_head *pages, int cpu)
1116 struct buffer_page *bpage, *tmp;
1118 for (i = 0; i < nr_pages; i++) {
1121 * __GFP_NORETRY flag makes sure that the allocation fails
1122 * gracefully without invoking oom-killer and the system is
1125 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1126 GFP_KERNEL | __GFP_NORETRY,
1131 list_add(&bpage->list, pages);
1133 page = alloc_pages_node(cpu_to_node(cpu),
1134 GFP_KERNEL | __GFP_NORETRY, 0);
1137 bpage->page = page_address(page);
1138 rb_init_page(bpage->page);
1144 list_for_each_entry_safe(bpage, tmp, pages, list) {
1145 list_del_init(&bpage->list);
1146 free_buffer_page(bpage);
1152 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1159 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1163 * The ring buffer page list is a circular list that does not
1164 * start and end with a list head. All page list items point to
1167 cpu_buffer->pages = pages.next;
1170 cpu_buffer->nr_pages = nr_pages;
1172 rb_check_pages(cpu_buffer);
1177 static struct ring_buffer_per_cpu *
1178 rb_allocate_cpu_buffer(struct ring_buffer *buffer, int nr_pages, int cpu)
1180 struct ring_buffer_per_cpu *cpu_buffer;
1181 struct buffer_page *bpage;
1185 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1186 GFP_KERNEL, cpu_to_node(cpu));
1190 cpu_buffer->cpu = cpu;
1191 cpu_buffer->buffer = buffer;
1192 raw_spin_lock_init(&cpu_buffer->reader_lock);
1193 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1194 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1195 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1196 init_completion(&cpu_buffer->update_done);
1197 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1198 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1200 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1201 GFP_KERNEL, cpu_to_node(cpu));
1203 goto fail_free_buffer;
1205 rb_check_bpage(cpu_buffer, bpage);
1207 cpu_buffer->reader_page = bpage;
1208 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1210 goto fail_free_reader;
1211 bpage->page = page_address(page);
1212 rb_init_page(bpage->page);
1214 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1215 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1217 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1219 goto fail_free_reader;
1221 cpu_buffer->head_page
1222 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1223 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1225 rb_head_page_activate(cpu_buffer);
1230 free_buffer_page(cpu_buffer->reader_page);
1237 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1239 struct list_head *head = cpu_buffer->pages;
1240 struct buffer_page *bpage, *tmp;
1242 free_buffer_page(cpu_buffer->reader_page);
1244 rb_head_page_deactivate(cpu_buffer);
1247 list_for_each_entry_safe(bpage, tmp, head, list) {
1248 list_del_init(&bpage->list);
1249 free_buffer_page(bpage);
1251 bpage = list_entry(head, struct buffer_page, list);
1252 free_buffer_page(bpage);
1258 #ifdef CONFIG_HOTPLUG_CPU
1259 static int rb_cpu_notify(struct notifier_block *self,
1260 unsigned long action, void *hcpu);
1264 * __ring_buffer_alloc - allocate a new ring_buffer
1265 * @size: the size in bytes per cpu that is needed.
1266 * @flags: attributes to set for the ring buffer.
1268 * Currently the only flag that is available is the RB_FL_OVERWRITE
1269 * flag. This flag means that the buffer will overwrite old data
1270 * when the buffer wraps. If this flag is not set, the buffer will
1271 * drop data when the tail hits the head.
1273 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1274 struct lock_class_key *key)
1276 struct ring_buffer *buffer;
1280 /* keep it in its own cache line */
1281 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1286 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1287 goto fail_free_buffer;
1289 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1290 buffer->flags = flags;
1291 buffer->clock = trace_clock_local;
1292 buffer->reader_lock_key = key;
1294 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1295 init_waitqueue_head(&buffer->irq_work.waiters);
1297 /* need at least two pages */
1302 * In case of non-hotplug cpu, if the ring-buffer is allocated
1303 * in early initcall, it will not be notified of secondary cpus.
1304 * In that off case, we need to allocate for all possible cpus.
1306 #ifdef CONFIG_HOTPLUG_CPU
1307 cpu_notifier_register_begin();
1308 cpumask_copy(buffer->cpumask, cpu_online_mask);
1310 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1312 buffer->cpus = nr_cpu_ids;
1314 bsize = sizeof(void *) * nr_cpu_ids;
1315 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1317 if (!buffer->buffers)
1318 goto fail_free_cpumask;
1320 for_each_buffer_cpu(buffer, cpu) {
1321 buffer->buffers[cpu] =
1322 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1323 if (!buffer->buffers[cpu])
1324 goto fail_free_buffers;
1327 #ifdef CONFIG_HOTPLUG_CPU
1328 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1329 buffer->cpu_notify.priority = 0;
1330 __register_cpu_notifier(&buffer->cpu_notify);
1331 cpu_notifier_register_done();
1334 mutex_init(&buffer->mutex);
1339 for_each_buffer_cpu(buffer, cpu) {
1340 if (buffer->buffers[cpu])
1341 rb_free_cpu_buffer(buffer->buffers[cpu]);
1343 kfree(buffer->buffers);
1346 free_cpumask_var(buffer->cpumask);
1347 #ifdef CONFIG_HOTPLUG_CPU
1348 cpu_notifier_register_done();
1355 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1358 * ring_buffer_free - free a ring buffer.
1359 * @buffer: the buffer to free.
1362 ring_buffer_free(struct ring_buffer *buffer)
1366 #ifdef CONFIG_HOTPLUG_CPU
1367 cpu_notifier_register_begin();
1368 __unregister_cpu_notifier(&buffer->cpu_notify);
1371 for_each_buffer_cpu(buffer, cpu)
1372 rb_free_cpu_buffer(buffer->buffers[cpu]);
1374 #ifdef CONFIG_HOTPLUG_CPU
1375 cpu_notifier_register_done();
1378 kfree(buffer->buffers);
1379 free_cpumask_var(buffer->cpumask);
1383 EXPORT_SYMBOL_GPL(ring_buffer_free);
1385 void ring_buffer_set_clock(struct ring_buffer *buffer,
1388 buffer->clock = clock;
1391 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1393 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1395 return local_read(&bpage->entries) & RB_WRITE_MASK;
1398 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1400 return local_read(&bpage->write) & RB_WRITE_MASK;
1404 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned int nr_pages)
1406 struct list_head *tail_page, *to_remove, *next_page;
1407 struct buffer_page *to_remove_page, *tmp_iter_page;
1408 struct buffer_page *last_page, *first_page;
1409 unsigned int nr_removed;
1410 unsigned long head_bit;
1415 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1416 atomic_inc(&cpu_buffer->record_disabled);
1418 * We don't race with the readers since we have acquired the reader
1419 * lock. We also don't race with writers after disabling recording.
1420 * This makes it easy to figure out the first and the last page to be
1421 * removed from the list. We unlink all the pages in between including
1422 * the first and last pages. This is done in a busy loop so that we
1423 * lose the least number of traces.
1424 * The pages are freed after we restart recording and unlock readers.
1426 tail_page = &cpu_buffer->tail_page->list;
1429 * tail page might be on reader page, we remove the next page
1430 * from the ring buffer
1432 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1433 tail_page = rb_list_head(tail_page->next);
1434 to_remove = tail_page;
1436 /* start of pages to remove */
1437 first_page = list_entry(rb_list_head(to_remove->next),
1438 struct buffer_page, list);
1440 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1441 to_remove = rb_list_head(to_remove)->next;
1442 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1445 next_page = rb_list_head(to_remove)->next;
1448 * Now we remove all pages between tail_page and next_page.
1449 * Make sure that we have head_bit value preserved for the
1452 tail_page->next = (struct list_head *)((unsigned long)next_page |
1454 next_page = rb_list_head(next_page);
1455 next_page->prev = tail_page;
1457 /* make sure pages points to a valid page in the ring buffer */
1458 cpu_buffer->pages = next_page;
1460 /* update head page */
1462 cpu_buffer->head_page = list_entry(next_page,
1463 struct buffer_page, list);
1466 * change read pointer to make sure any read iterators reset
1469 cpu_buffer->read = 0;
1471 /* pages are removed, resume tracing and then free the pages */
1472 atomic_dec(&cpu_buffer->record_disabled);
1473 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1475 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1477 /* last buffer page to remove */
1478 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1480 tmp_iter_page = first_page;
1483 to_remove_page = tmp_iter_page;
1484 rb_inc_page(cpu_buffer, &tmp_iter_page);
1486 /* update the counters */
1487 page_entries = rb_page_entries(to_remove_page);
1490 * If something was added to this page, it was full
1491 * since it is not the tail page. So we deduct the
1492 * bytes consumed in ring buffer from here.
1493 * Increment overrun to account for the lost events.
1495 local_add(page_entries, &cpu_buffer->overrun);
1496 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1500 * We have already removed references to this list item, just
1501 * free up the buffer_page and its page
1503 free_buffer_page(to_remove_page);
1506 } while (to_remove_page != last_page);
1508 RB_WARN_ON(cpu_buffer, nr_removed);
1510 return nr_removed == 0;
1514 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1516 struct list_head *pages = &cpu_buffer->new_pages;
1517 int retries, success;
1519 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1521 * We are holding the reader lock, so the reader page won't be swapped
1522 * in the ring buffer. Now we are racing with the writer trying to
1523 * move head page and the tail page.
1524 * We are going to adapt the reader page update process where:
1525 * 1. We first splice the start and end of list of new pages between
1526 * the head page and its previous page.
1527 * 2. We cmpxchg the prev_page->next to point from head page to the
1528 * start of new pages list.
1529 * 3. Finally, we update the head->prev to the end of new list.
1531 * We will try this process 10 times, to make sure that we don't keep
1537 struct list_head *head_page, *prev_page, *r;
1538 struct list_head *last_page, *first_page;
1539 struct list_head *head_page_with_bit;
1541 head_page = &rb_set_head_page(cpu_buffer)->list;
1544 prev_page = head_page->prev;
1546 first_page = pages->next;
1547 last_page = pages->prev;
1549 head_page_with_bit = (struct list_head *)
1550 ((unsigned long)head_page | RB_PAGE_HEAD);
1552 last_page->next = head_page_with_bit;
1553 first_page->prev = prev_page;
1555 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1557 if (r == head_page_with_bit) {
1559 * yay, we replaced the page pointer to our new list,
1560 * now, we just have to update to head page's prev
1561 * pointer to point to end of list
1563 head_page->prev = last_page;
1570 INIT_LIST_HEAD(pages);
1572 * If we weren't successful in adding in new pages, warn and stop
1575 RB_WARN_ON(cpu_buffer, !success);
1576 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1578 /* free pages if they weren't inserted */
1580 struct buffer_page *bpage, *tmp;
1581 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1583 list_del_init(&bpage->list);
1584 free_buffer_page(bpage);
1590 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1594 if (cpu_buffer->nr_pages_to_update > 0)
1595 success = rb_insert_pages(cpu_buffer);
1597 success = rb_remove_pages(cpu_buffer,
1598 -cpu_buffer->nr_pages_to_update);
1601 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1604 static void update_pages_handler(struct work_struct *work)
1606 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1607 struct ring_buffer_per_cpu, update_pages_work);
1608 rb_update_pages(cpu_buffer);
1609 complete(&cpu_buffer->update_done);
1613 * ring_buffer_resize - resize the ring buffer
1614 * @buffer: the buffer to resize.
1615 * @size: the new size.
1616 * @cpu_id: the cpu buffer to resize
1618 * Minimum size is 2 * BUF_PAGE_SIZE.
1620 * Returns 0 on success and < 0 on failure.
1622 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1625 struct ring_buffer_per_cpu *cpu_buffer;
1630 * Always succeed at resizing a non-existent buffer:
1635 /* Make sure the requested buffer exists */
1636 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1637 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1640 size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1641 size *= BUF_PAGE_SIZE;
1643 /* we need a minimum of two pages */
1644 if (size < BUF_PAGE_SIZE * 2)
1645 size = BUF_PAGE_SIZE * 2;
1647 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1650 * Don't succeed if resizing is disabled, as a reader might be
1651 * manipulating the ring buffer and is expecting a sane state while
1654 if (atomic_read(&buffer->resize_disabled))
1657 /* prevent another thread from changing buffer sizes */
1658 mutex_lock(&buffer->mutex);
1660 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1661 /* calculate the pages to update */
1662 for_each_buffer_cpu(buffer, cpu) {
1663 cpu_buffer = buffer->buffers[cpu];
1665 cpu_buffer->nr_pages_to_update = nr_pages -
1666 cpu_buffer->nr_pages;
1668 * nothing more to do for removing pages or no update
1670 if (cpu_buffer->nr_pages_to_update <= 0)
1673 * to add pages, make sure all new pages can be
1674 * allocated without receiving ENOMEM
1676 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1677 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1678 &cpu_buffer->new_pages, cpu)) {
1679 /* not enough memory for new pages */
1687 * Fire off all the required work handlers
1688 * We can't schedule on offline CPUs, but it's not necessary
1689 * since we can change their buffer sizes without any race.
1691 for_each_buffer_cpu(buffer, cpu) {
1692 cpu_buffer = buffer->buffers[cpu];
1693 if (!cpu_buffer->nr_pages_to_update)
1696 /* The update must run on the CPU that is being updated. */
1698 if (cpu == smp_processor_id() || !cpu_online(cpu)) {
1699 rb_update_pages(cpu_buffer);
1700 cpu_buffer->nr_pages_to_update = 0;
1703 * Can not disable preemption for schedule_work_on()
1707 schedule_work_on(cpu,
1708 &cpu_buffer->update_pages_work);
1714 /* wait for all the updates to complete */
1715 for_each_buffer_cpu(buffer, cpu) {
1716 cpu_buffer = buffer->buffers[cpu];
1717 if (!cpu_buffer->nr_pages_to_update)
1720 if (cpu_online(cpu))
1721 wait_for_completion(&cpu_buffer->update_done);
1722 cpu_buffer->nr_pages_to_update = 0;
1727 /* Make sure this CPU has been intitialized */
1728 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1731 cpu_buffer = buffer->buffers[cpu_id];
1733 if (nr_pages == cpu_buffer->nr_pages)
1736 cpu_buffer->nr_pages_to_update = nr_pages -
1737 cpu_buffer->nr_pages;
1739 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1740 if (cpu_buffer->nr_pages_to_update > 0 &&
1741 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1742 &cpu_buffer->new_pages, cpu_id)) {
1750 /* The update must run on the CPU that is being updated. */
1751 if (cpu_id == smp_processor_id() || !cpu_online(cpu_id))
1752 rb_update_pages(cpu_buffer);
1755 * Can not disable preemption for schedule_work_on()
1759 schedule_work_on(cpu_id,
1760 &cpu_buffer->update_pages_work);
1761 wait_for_completion(&cpu_buffer->update_done);
1766 cpu_buffer->nr_pages_to_update = 0;
1772 * The ring buffer resize can happen with the ring buffer
1773 * enabled, so that the update disturbs the tracing as little
1774 * as possible. But if the buffer is disabled, we do not need
1775 * to worry about that, and we can take the time to verify
1776 * that the buffer is not corrupt.
1778 if (atomic_read(&buffer->record_disabled)) {
1779 atomic_inc(&buffer->record_disabled);
1781 * Even though the buffer was disabled, we must make sure
1782 * that it is truly disabled before calling rb_check_pages.
1783 * There could have been a race between checking
1784 * record_disable and incrementing it.
1786 synchronize_sched();
1787 for_each_buffer_cpu(buffer, cpu) {
1788 cpu_buffer = buffer->buffers[cpu];
1789 rb_check_pages(cpu_buffer);
1791 atomic_dec(&buffer->record_disabled);
1794 mutex_unlock(&buffer->mutex);
1798 for_each_buffer_cpu(buffer, cpu) {
1799 struct buffer_page *bpage, *tmp;
1801 cpu_buffer = buffer->buffers[cpu];
1802 cpu_buffer->nr_pages_to_update = 0;
1804 if (list_empty(&cpu_buffer->new_pages))
1807 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1809 list_del_init(&bpage->list);
1810 free_buffer_page(bpage);
1813 mutex_unlock(&buffer->mutex);
1816 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1818 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1820 mutex_lock(&buffer->mutex);
1822 buffer->flags |= RB_FL_OVERWRITE;
1824 buffer->flags &= ~RB_FL_OVERWRITE;
1825 mutex_unlock(&buffer->mutex);
1827 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1829 static inline void *
1830 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1832 return bpage->data + index;
1835 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1837 return bpage->page->data + index;
1840 static inline struct ring_buffer_event *
1841 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1843 return __rb_page_index(cpu_buffer->reader_page,
1844 cpu_buffer->reader_page->read);
1847 static inline struct ring_buffer_event *
1848 rb_iter_head_event(struct ring_buffer_iter *iter)
1850 return __rb_page_index(iter->head_page, iter->head);
1853 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1855 return local_read(&bpage->page->commit);
1858 /* Size is determined by what has been committed */
1859 static inline unsigned rb_page_size(struct buffer_page *bpage)
1861 return rb_page_commit(bpage);
1864 static inline unsigned
1865 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1867 return rb_page_commit(cpu_buffer->commit_page);
1870 static inline unsigned
1871 rb_event_index(struct ring_buffer_event *event)
1873 unsigned long addr = (unsigned long)event;
1875 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1879 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1880 struct ring_buffer_event *event)
1882 unsigned long addr = (unsigned long)event;
1883 unsigned long index;
1885 index = rb_event_index(event);
1888 return cpu_buffer->commit_page->page == (void *)addr &&
1889 rb_commit_index(cpu_buffer) == index;
1893 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1895 unsigned long max_count;
1898 * We only race with interrupts and NMIs on this CPU.
1899 * If we own the commit event, then we can commit
1900 * all others that interrupted us, since the interruptions
1901 * are in stack format (they finish before they come
1902 * back to us). This allows us to do a simple loop to
1903 * assign the commit to the tail.
1906 max_count = cpu_buffer->nr_pages * 100;
1908 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1909 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1911 if (RB_WARN_ON(cpu_buffer,
1912 rb_is_reader_page(cpu_buffer->tail_page)))
1914 local_set(&cpu_buffer->commit_page->page->commit,
1915 rb_page_write(cpu_buffer->commit_page));
1916 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1917 cpu_buffer->write_stamp =
1918 cpu_buffer->commit_page->page->time_stamp;
1919 /* add barrier to keep gcc from optimizing too much */
1922 while (rb_commit_index(cpu_buffer) !=
1923 rb_page_write(cpu_buffer->commit_page)) {
1925 local_set(&cpu_buffer->commit_page->page->commit,
1926 rb_page_write(cpu_buffer->commit_page));
1927 RB_WARN_ON(cpu_buffer,
1928 local_read(&cpu_buffer->commit_page->page->commit) &
1933 /* again, keep gcc from optimizing */
1937 * If an interrupt came in just after the first while loop
1938 * and pushed the tail page forward, we will be left with
1939 * a dangling commit that will never go forward.
1941 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1945 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1947 cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1948 cpu_buffer->reader_page->read = 0;
1951 static void rb_inc_iter(struct ring_buffer_iter *iter)
1953 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1956 * The iterator could be on the reader page (it starts there).
1957 * But the head could have moved, since the reader was
1958 * found. Check for this case and assign the iterator
1959 * to the head page instead of next.
1961 if (iter->head_page == cpu_buffer->reader_page)
1962 iter->head_page = rb_set_head_page(cpu_buffer);
1964 rb_inc_page(cpu_buffer, &iter->head_page);
1966 iter->read_stamp = iter->head_page->page->time_stamp;
1970 /* Slow path, do not inline */
1971 static noinline struct ring_buffer_event *
1972 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
1974 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
1976 /* Not the first event on the page? */
1977 if (rb_event_index(event)) {
1978 event->time_delta = delta & TS_MASK;
1979 event->array[0] = delta >> TS_SHIFT;
1981 /* nope, just zero it */
1982 event->time_delta = 0;
1983 event->array[0] = 0;
1986 return skip_time_extend(event);
1990 * rb_update_event - update event type and data
1991 * @event: the even to update
1992 * @type: the type of event
1993 * @length: the size of the event field in the ring buffer
1995 * Update the type and data fields of the event. The length
1996 * is the actual size that is written to the ring buffer,
1997 * and with this, we can determine what to place into the
2001 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2002 struct ring_buffer_event *event, unsigned length,
2003 int add_timestamp, u64 delta)
2005 /* Only a commit updates the timestamp */
2006 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2010 * If we need to add a timestamp, then we
2011 * add it to the start of the resevered space.
2013 if (unlikely(add_timestamp)) {
2014 event = rb_add_time_stamp(event, delta);
2015 length -= RB_LEN_TIME_EXTEND;
2019 event->time_delta = delta;
2020 length -= RB_EVNT_HDR_SIZE;
2021 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2022 event->type_len = 0;
2023 event->array[0] = length;
2025 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2029 * rb_handle_head_page - writer hit the head page
2031 * Returns: +1 to retry page
2036 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2037 struct buffer_page *tail_page,
2038 struct buffer_page *next_page)
2040 struct buffer_page *new_head;
2045 entries = rb_page_entries(next_page);
2048 * The hard part is here. We need to move the head
2049 * forward, and protect against both readers on
2050 * other CPUs and writers coming in via interrupts.
2052 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2056 * type can be one of four:
2057 * NORMAL - an interrupt already moved it for us
2058 * HEAD - we are the first to get here.
2059 * UPDATE - we are the interrupt interrupting
2061 * MOVED - a reader on another CPU moved the next
2062 * pointer to its reader page. Give up
2069 * We changed the head to UPDATE, thus
2070 * it is our responsibility to update
2073 local_add(entries, &cpu_buffer->overrun);
2074 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2077 * The entries will be zeroed out when we move the
2081 /* still more to do */
2084 case RB_PAGE_UPDATE:
2086 * This is an interrupt that interrupt the
2087 * previous update. Still more to do.
2090 case RB_PAGE_NORMAL:
2092 * An interrupt came in before the update
2093 * and processed this for us.
2094 * Nothing left to do.
2099 * The reader is on another CPU and just did
2100 * a swap with our next_page.
2105 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2110 * Now that we are here, the old head pointer is
2111 * set to UPDATE. This will keep the reader from
2112 * swapping the head page with the reader page.
2113 * The reader (on another CPU) will spin till
2116 * We just need to protect against interrupts
2117 * doing the job. We will set the next pointer
2118 * to HEAD. After that, we set the old pointer
2119 * to NORMAL, but only if it was HEAD before.
2120 * otherwise we are an interrupt, and only
2121 * want the outer most commit to reset it.
2123 new_head = next_page;
2124 rb_inc_page(cpu_buffer, &new_head);
2126 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2130 * Valid returns are:
2131 * HEAD - an interrupt came in and already set it.
2132 * NORMAL - One of two things:
2133 * 1) We really set it.
2134 * 2) A bunch of interrupts came in and moved
2135 * the page forward again.
2139 case RB_PAGE_NORMAL:
2143 RB_WARN_ON(cpu_buffer, 1);
2148 * It is possible that an interrupt came in,
2149 * set the head up, then more interrupts came in
2150 * and moved it again. When we get back here,
2151 * the page would have been set to NORMAL but we
2152 * just set it back to HEAD.
2154 * How do you detect this? Well, if that happened
2155 * the tail page would have moved.
2157 if (ret == RB_PAGE_NORMAL) {
2159 * If the tail had moved passed next, then we need
2160 * to reset the pointer.
2162 if (cpu_buffer->tail_page != tail_page &&
2163 cpu_buffer->tail_page != next_page)
2164 rb_head_page_set_normal(cpu_buffer, new_head,
2170 * If this was the outer most commit (the one that
2171 * changed the original pointer from HEAD to UPDATE),
2172 * then it is up to us to reset it to NORMAL.
2174 if (type == RB_PAGE_HEAD) {
2175 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2178 if (RB_WARN_ON(cpu_buffer,
2179 ret != RB_PAGE_UPDATE))
2186 static unsigned rb_calculate_event_length(unsigned length)
2188 struct ring_buffer_event event; /* Used only for sizeof array */
2190 /* zero length can cause confusions */
2194 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2195 length += sizeof(event.array[0]);
2197 length += RB_EVNT_HDR_SIZE;
2198 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2204 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2205 struct buffer_page *tail_page,
2206 unsigned long tail, unsigned long length)
2208 struct ring_buffer_event *event;
2211 * Only the event that crossed the page boundary
2212 * must fill the old tail_page with padding.
2214 if (tail >= BUF_PAGE_SIZE) {
2216 * If the page was filled, then we still need
2217 * to update the real_end. Reset it to zero
2218 * and the reader will ignore it.
2220 if (tail == BUF_PAGE_SIZE)
2221 tail_page->real_end = 0;
2223 local_sub(length, &tail_page->write);
2227 event = __rb_page_index(tail_page, tail);
2228 kmemcheck_annotate_bitfield(event, bitfield);
2230 /* account for padding bytes */
2231 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2234 * Save the original length to the meta data.
2235 * This will be used by the reader to add lost event
2238 tail_page->real_end = tail;
2241 * If this event is bigger than the minimum size, then
2242 * we need to be careful that we don't subtract the
2243 * write counter enough to allow another writer to slip
2245 * We put in a discarded commit instead, to make sure
2246 * that this space is not used again.
2248 * If we are less than the minimum size, we don't need to
2251 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2252 /* No room for any events */
2254 /* Mark the rest of the page with padding */
2255 rb_event_set_padding(event);
2257 /* Set the write back to the previous setting */
2258 local_sub(length, &tail_page->write);
2262 /* Put in a discarded event */
2263 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2264 event->type_len = RINGBUF_TYPE_PADDING;
2265 /* time delta must be non zero */
2266 event->time_delta = 1;
2268 /* Set write to end of buffer */
2269 length = (tail + length) - BUF_PAGE_SIZE;
2270 local_sub(length, &tail_page->write);
2274 * This is the slow path, force gcc not to inline it.
2276 static noinline struct ring_buffer_event *
2277 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2278 unsigned long length, unsigned long tail,
2279 struct buffer_page *tail_page, u64 ts)
2281 struct buffer_page *commit_page = cpu_buffer->commit_page;
2282 struct ring_buffer *buffer = cpu_buffer->buffer;
2283 struct buffer_page *next_page;
2286 next_page = tail_page;
2288 rb_inc_page(cpu_buffer, &next_page);
2291 * If for some reason, we had an interrupt storm that made
2292 * it all the way around the buffer, bail, and warn
2295 if (unlikely(next_page == commit_page)) {
2296 local_inc(&cpu_buffer->commit_overrun);
2301 * This is where the fun begins!
2303 * We are fighting against races between a reader that
2304 * could be on another CPU trying to swap its reader
2305 * page with the buffer head.
2307 * We are also fighting against interrupts coming in and
2308 * moving the head or tail on us as well.
2310 * If the next page is the head page then we have filled
2311 * the buffer, unless the commit page is still on the
2314 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2317 * If the commit is not on the reader page, then
2318 * move the header page.
2320 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2322 * If we are not in overwrite mode,
2323 * this is easy, just stop here.
2325 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2326 local_inc(&cpu_buffer->dropped_events);
2330 ret = rb_handle_head_page(cpu_buffer,
2339 * We need to be careful here too. The
2340 * commit page could still be on the reader
2341 * page. We could have a small buffer, and
2342 * have filled up the buffer with events
2343 * from interrupts and such, and wrapped.
2345 * Note, if the tail page is also the on the
2346 * reader_page, we let it move out.
2348 if (unlikely((cpu_buffer->commit_page !=
2349 cpu_buffer->tail_page) &&
2350 (cpu_buffer->commit_page ==
2351 cpu_buffer->reader_page))) {
2352 local_inc(&cpu_buffer->commit_overrun);
2358 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
2361 * Nested commits always have zero deltas, so
2362 * just reread the time stamp
2364 ts = rb_time_stamp(buffer);
2365 next_page->page->time_stamp = ts;
2370 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2372 /* fail and let the caller try again */
2373 return ERR_PTR(-EAGAIN);
2377 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2382 static struct ring_buffer_event *
2383 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2384 unsigned long length, u64 ts,
2385 u64 delta, int add_timestamp)
2387 struct buffer_page *tail_page;
2388 struct ring_buffer_event *event;
2389 unsigned long tail, write;
2392 * If the time delta since the last event is too big to
2393 * hold in the time field of the event, then we append a
2394 * TIME EXTEND event ahead of the data event.
2396 if (unlikely(add_timestamp))
2397 length += RB_LEN_TIME_EXTEND;
2399 tail_page = cpu_buffer->tail_page;
2400 write = local_add_return(length, &tail_page->write);
2402 /* set write to only the index of the write */
2403 write &= RB_WRITE_MASK;
2404 tail = write - length;
2407 * If this is the first commit on the page, then it has the same
2408 * timestamp as the page itself.
2413 /* See if we shot pass the end of this buffer page */
2414 if (unlikely(write > BUF_PAGE_SIZE))
2415 return rb_move_tail(cpu_buffer, length, tail,
2418 /* We reserved something on the buffer */
2420 event = __rb_page_index(tail_page, tail);
2421 kmemcheck_annotate_bitfield(event, bitfield);
2422 rb_update_event(cpu_buffer, event, length, add_timestamp, delta);
2424 local_inc(&tail_page->entries);
2427 * If this is the first commit on the page, then update
2431 tail_page->page->time_stamp = ts;
2433 /* account for these added bytes */
2434 local_add(length, &cpu_buffer->entries_bytes);
2440 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2441 struct ring_buffer_event *event)
2443 unsigned long new_index, old_index;
2444 struct buffer_page *bpage;
2445 unsigned long index;
2448 new_index = rb_event_index(event);
2449 old_index = new_index + rb_event_ts_length(event);
2450 addr = (unsigned long)event;
2453 bpage = cpu_buffer->tail_page;
2455 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2456 unsigned long write_mask =
2457 local_read(&bpage->write) & ~RB_WRITE_MASK;
2458 unsigned long event_length = rb_event_length(event);
2460 * This is on the tail page. It is possible that
2461 * a write could come in and move the tail page
2462 * and write to the next page. That is fine
2463 * because we just shorten what is on this page.
2465 old_index += write_mask;
2466 new_index += write_mask;
2467 index = local_cmpxchg(&bpage->write, old_index, new_index);
2468 if (index == old_index) {
2469 /* update counters */
2470 local_sub(event_length, &cpu_buffer->entries_bytes);
2475 /* could not discard */
2479 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2481 local_inc(&cpu_buffer->committing);
2482 local_inc(&cpu_buffer->commits);
2485 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2487 unsigned long commits;
2489 if (RB_WARN_ON(cpu_buffer,
2490 !local_read(&cpu_buffer->committing)))
2494 commits = local_read(&cpu_buffer->commits);
2495 /* synchronize with interrupts */
2497 if (local_read(&cpu_buffer->committing) == 1)
2498 rb_set_commit_to_write(cpu_buffer);
2500 local_dec(&cpu_buffer->committing);
2502 /* synchronize with interrupts */
2506 * Need to account for interrupts coming in between the
2507 * updating of the commit page and the clearing of the
2508 * committing counter.
2510 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2511 !local_read(&cpu_buffer->committing)) {
2512 local_inc(&cpu_buffer->committing);
2517 static struct ring_buffer_event *
2518 rb_reserve_next_event(struct ring_buffer *buffer,
2519 struct ring_buffer_per_cpu *cpu_buffer,
2520 unsigned long length)
2522 struct ring_buffer_event *event;
2528 rb_start_commit(cpu_buffer);
2530 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2532 * Due to the ability to swap a cpu buffer from a buffer
2533 * it is possible it was swapped before we committed.
2534 * (committing stops a swap). We check for it here and
2535 * if it happened, we have to fail the write.
2538 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2539 local_dec(&cpu_buffer->committing);
2540 local_dec(&cpu_buffer->commits);
2545 length = rb_calculate_event_length(length);
2551 * We allow for interrupts to reenter here and do a trace.
2552 * If one does, it will cause this original code to loop
2553 * back here. Even with heavy interrupts happening, this
2554 * should only happen a few times in a row. If this happens
2555 * 1000 times in a row, there must be either an interrupt
2556 * storm or we have something buggy.
2559 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2562 ts = rb_time_stamp(cpu_buffer->buffer);
2563 diff = ts - cpu_buffer->write_stamp;
2565 /* make sure this diff is calculated here */
2568 /* Did the write stamp get updated already? */
2569 if (likely(ts >= cpu_buffer->write_stamp)) {
2571 if (unlikely(test_time_stamp(delta))) {
2572 int local_clock_stable = 1;
2573 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2574 local_clock_stable = sched_clock_stable();
2576 WARN_ONCE(delta > (1ULL << 59),
2577 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2578 (unsigned long long)delta,
2579 (unsigned long long)ts,
2580 (unsigned long long)cpu_buffer->write_stamp,
2581 local_clock_stable ? "" :
2582 "If you just came from a suspend/resume,\n"
2583 "please switch to the trace global clock:\n"
2584 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2589 event = __rb_reserve_next(cpu_buffer, length, ts,
2590 delta, add_timestamp);
2591 if (unlikely(PTR_ERR(event) == -EAGAIN))
2600 rb_end_commit(cpu_buffer);
2604 #ifdef CONFIG_TRACING
2607 * The lock and unlock are done within a preempt disable section.
2608 * The current_context per_cpu variable can only be modified
2609 * by the current task between lock and unlock. But it can
2610 * be modified more than once via an interrupt. To pass this
2611 * information from the lock to the unlock without having to
2612 * access the 'in_interrupt()' functions again (which do show
2613 * a bit of overhead in something as critical as function tracing,
2614 * we use a bitmask trick.
2616 * bit 0 = NMI context
2617 * bit 1 = IRQ context
2618 * bit 2 = SoftIRQ context
2619 * bit 3 = normal context.
2621 * This works because this is the order of contexts that can
2622 * preempt other contexts. A SoftIRQ never preempts an IRQ
2625 * When the context is determined, the corresponding bit is
2626 * checked and set (if it was set, then a recursion of that context
2629 * On unlock, we need to clear this bit. To do so, just subtract
2630 * 1 from the current_context and AND it to itself.
2634 * 101 & 100 = 100 (clearing bit zero)
2637 * 1010 & 1001 = 1000 (clearing bit 1)
2639 * The least significant bit can be cleared this way, and it
2640 * just so happens that it is the same bit corresponding to
2641 * the current context.
2643 static DEFINE_PER_CPU(unsigned int, current_context);
2645 static __always_inline int trace_recursive_lock(void)
2647 unsigned int val = this_cpu_read(current_context);
2650 if (in_interrupt()) {
2660 if (unlikely(val & (1 << bit)))
2664 this_cpu_write(current_context, val);
2669 static __always_inline void trace_recursive_unlock(void)
2671 unsigned int val = this_cpu_read(current_context);
2674 val &= this_cpu_read(current_context);
2675 this_cpu_write(current_context, val);
2680 #define trace_recursive_lock() (0)
2681 #define trace_recursive_unlock() do { } while (0)
2686 * ring_buffer_lock_reserve - reserve a part of the buffer
2687 * @buffer: the ring buffer to reserve from
2688 * @length: the length of the data to reserve (excluding event header)
2690 * Returns a reseverd event on the ring buffer to copy directly to.
2691 * The user of this interface will need to get the body to write into
2692 * and can use the ring_buffer_event_data() interface.
2694 * The length is the length of the data needed, not the event length
2695 * which also includes the event header.
2697 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2698 * If NULL is returned, then nothing has been allocated or locked.
2700 struct ring_buffer_event *
2701 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2703 struct ring_buffer_per_cpu *cpu_buffer;
2704 struct ring_buffer_event *event;
2707 if (ring_buffer_flags != RB_BUFFERS_ON)
2710 /* If we are tracing schedule, we don't want to recurse */
2711 preempt_disable_notrace();
2713 if (atomic_read(&buffer->record_disabled))
2716 if (trace_recursive_lock())
2719 cpu = raw_smp_processor_id();
2721 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2724 cpu_buffer = buffer->buffers[cpu];
2726 if (atomic_read(&cpu_buffer->record_disabled))
2729 if (length > BUF_MAX_DATA_SIZE)
2732 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2739 trace_recursive_unlock();
2742 preempt_enable_notrace();
2745 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2748 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2749 struct ring_buffer_event *event)
2754 * The event first in the commit queue updates the
2757 if (rb_event_is_commit(cpu_buffer, event)) {
2759 * A commit event that is first on a page
2760 * updates the write timestamp with the page stamp
2762 if (!rb_event_index(event))
2763 cpu_buffer->write_stamp =
2764 cpu_buffer->commit_page->page->time_stamp;
2765 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2766 delta = event->array[0];
2768 delta += event->time_delta;
2769 cpu_buffer->write_stamp += delta;
2771 cpu_buffer->write_stamp += event->time_delta;
2775 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2776 struct ring_buffer_event *event)
2778 local_inc(&cpu_buffer->entries);
2779 rb_update_write_stamp(cpu_buffer, event);
2780 rb_end_commit(cpu_buffer);
2783 static __always_inline void
2784 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2786 if (buffer->irq_work.waiters_pending) {
2787 buffer->irq_work.waiters_pending = false;
2788 /* irq_work_queue() supplies it's own memory barriers */
2789 irq_work_queue(&buffer->irq_work.work);
2792 if (cpu_buffer->irq_work.waiters_pending) {
2793 cpu_buffer->irq_work.waiters_pending = false;
2794 /* irq_work_queue() supplies it's own memory barriers */
2795 irq_work_queue(&cpu_buffer->irq_work.work);
2800 * ring_buffer_unlock_commit - commit a reserved
2801 * @buffer: The buffer to commit to
2802 * @event: The event pointer to commit.
2804 * This commits the data to the ring buffer, and releases any locks held.
2806 * Must be paired with ring_buffer_lock_reserve.
2808 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2809 struct ring_buffer_event *event)
2811 struct ring_buffer_per_cpu *cpu_buffer;
2812 int cpu = raw_smp_processor_id();
2814 cpu_buffer = buffer->buffers[cpu];
2816 rb_commit(cpu_buffer, event);
2818 rb_wakeups(buffer, cpu_buffer);
2820 trace_recursive_unlock();
2822 preempt_enable_notrace();
2826 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2828 static inline void rb_event_discard(struct ring_buffer_event *event)
2830 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2831 event = skip_time_extend(event);
2833 /* array[0] holds the actual length for the discarded event */
2834 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2835 event->type_len = RINGBUF_TYPE_PADDING;
2836 /* time delta must be non zero */
2837 if (!event->time_delta)
2838 event->time_delta = 1;
2842 * Decrement the entries to the page that an event is on.
2843 * The event does not even need to exist, only the pointer
2844 * to the page it is on. This may only be called before the commit
2848 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2849 struct ring_buffer_event *event)
2851 unsigned long addr = (unsigned long)event;
2852 struct buffer_page *bpage = cpu_buffer->commit_page;
2853 struct buffer_page *start;
2857 /* Do the likely case first */
2858 if (likely(bpage->page == (void *)addr)) {
2859 local_dec(&bpage->entries);
2864 * Because the commit page may be on the reader page we
2865 * start with the next page and check the end loop there.
2867 rb_inc_page(cpu_buffer, &bpage);
2870 if (bpage->page == (void *)addr) {
2871 local_dec(&bpage->entries);
2874 rb_inc_page(cpu_buffer, &bpage);
2875 } while (bpage != start);
2877 /* commit not part of this buffer?? */
2878 RB_WARN_ON(cpu_buffer, 1);
2882 * ring_buffer_commit_discard - discard an event that has not been committed
2883 * @buffer: the ring buffer
2884 * @event: non committed event to discard
2886 * Sometimes an event that is in the ring buffer needs to be ignored.
2887 * This function lets the user discard an event in the ring buffer
2888 * and then that event will not be read later.
2890 * This function only works if it is called before the the item has been
2891 * committed. It will try to free the event from the ring buffer
2892 * if another event has not been added behind it.
2894 * If another event has been added behind it, it will set the event
2895 * up as discarded, and perform the commit.
2897 * If this function is called, do not call ring_buffer_unlock_commit on
2900 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2901 struct ring_buffer_event *event)
2903 struct ring_buffer_per_cpu *cpu_buffer;
2906 /* The event is discarded regardless */
2907 rb_event_discard(event);
2909 cpu = smp_processor_id();
2910 cpu_buffer = buffer->buffers[cpu];
2913 * This must only be called if the event has not been
2914 * committed yet. Thus we can assume that preemption
2915 * is still disabled.
2917 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2919 rb_decrement_entry(cpu_buffer, event);
2920 if (rb_try_to_discard(cpu_buffer, event))
2924 * The commit is still visible by the reader, so we
2925 * must still update the timestamp.
2927 rb_update_write_stamp(cpu_buffer, event);
2929 rb_end_commit(cpu_buffer);
2931 trace_recursive_unlock();
2933 preempt_enable_notrace();
2936 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2939 * ring_buffer_write - write data to the buffer without reserving
2940 * @buffer: The ring buffer to write to.
2941 * @length: The length of the data being written (excluding the event header)
2942 * @data: The data to write to the buffer.
2944 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2945 * one function. If you already have the data to write to the buffer, it
2946 * may be easier to simply call this function.
2948 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2949 * and not the length of the event which would hold the header.
2951 int ring_buffer_write(struct ring_buffer *buffer,
2952 unsigned long length,
2955 struct ring_buffer_per_cpu *cpu_buffer;
2956 struct ring_buffer_event *event;
2961 if (ring_buffer_flags != RB_BUFFERS_ON)
2964 preempt_disable_notrace();
2966 if (atomic_read(&buffer->record_disabled))
2969 cpu = raw_smp_processor_id();
2971 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2974 cpu_buffer = buffer->buffers[cpu];
2976 if (atomic_read(&cpu_buffer->record_disabled))
2979 if (length > BUF_MAX_DATA_SIZE)
2982 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2986 body = rb_event_data(event);
2988 memcpy(body, data, length);
2990 rb_commit(cpu_buffer, event);
2992 rb_wakeups(buffer, cpu_buffer);
2996 preempt_enable_notrace();
3000 EXPORT_SYMBOL_GPL(ring_buffer_write);
3002 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3004 struct buffer_page *reader = cpu_buffer->reader_page;
3005 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3006 struct buffer_page *commit = cpu_buffer->commit_page;
3008 /* In case of error, head will be NULL */
3009 if (unlikely(!head))
3012 return reader->read == rb_page_commit(reader) &&
3013 (commit == reader ||
3015 head->read == rb_page_commit(commit)));
3019 * ring_buffer_record_disable - stop all writes into the buffer
3020 * @buffer: The ring buffer to stop writes to.
3022 * This prevents all writes to the buffer. Any attempt to write
3023 * to the buffer after this will fail and return NULL.
3025 * The caller should call synchronize_sched() after this.
3027 void ring_buffer_record_disable(struct ring_buffer *buffer)
3029 atomic_inc(&buffer->record_disabled);
3031 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3034 * ring_buffer_record_enable - enable writes to the buffer
3035 * @buffer: The ring buffer to enable writes
3037 * Note, multiple disables will need the same number of enables
3038 * to truly enable the writing (much like preempt_disable).
3040 void ring_buffer_record_enable(struct ring_buffer *buffer)
3042 atomic_dec(&buffer->record_disabled);
3044 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3047 * ring_buffer_record_off - stop all writes into the buffer
3048 * @buffer: The ring buffer to stop writes to.
3050 * This prevents all writes to the buffer. Any attempt to write
3051 * to the buffer after this will fail and return NULL.
3053 * This is different than ring_buffer_record_disable() as
3054 * it works like an on/off switch, where as the disable() version
3055 * must be paired with a enable().
3057 void ring_buffer_record_off(struct ring_buffer *buffer)
3060 unsigned int new_rd;
3063 rd = atomic_read(&buffer->record_disabled);
3064 new_rd = rd | RB_BUFFER_OFF;
3065 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3067 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3070 * ring_buffer_record_on - restart writes into the buffer
3071 * @buffer: The ring buffer to start writes to.
3073 * This enables all writes to the buffer that was disabled by
3074 * ring_buffer_record_off().
3076 * This is different than ring_buffer_record_enable() as
3077 * it works like an on/off switch, where as the enable() version
3078 * must be paired with a disable().
3080 void ring_buffer_record_on(struct ring_buffer *buffer)
3083 unsigned int new_rd;
3086 rd = atomic_read(&buffer->record_disabled);
3087 new_rd = rd & ~RB_BUFFER_OFF;
3088 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3090 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3093 * ring_buffer_record_is_on - return true if the ring buffer can write
3094 * @buffer: The ring buffer to see if write is enabled
3096 * Returns true if the ring buffer is in a state that it accepts writes.
3098 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3100 return !atomic_read(&buffer->record_disabled);
3104 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3105 * @buffer: The ring buffer to stop writes to.
3106 * @cpu: The CPU buffer to stop
3108 * This prevents all writes to the buffer. Any attempt to write
3109 * to the buffer after this will fail and return NULL.
3111 * The caller should call synchronize_sched() after this.
3113 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3115 struct ring_buffer_per_cpu *cpu_buffer;
3117 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3120 cpu_buffer = buffer->buffers[cpu];
3121 atomic_inc(&cpu_buffer->record_disabled);
3123 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3126 * ring_buffer_record_enable_cpu - enable writes to the buffer
3127 * @buffer: The ring buffer to enable writes
3128 * @cpu: The CPU to enable.
3130 * Note, multiple disables will need the same number of enables
3131 * to truly enable the writing (much like preempt_disable).
3133 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3135 struct ring_buffer_per_cpu *cpu_buffer;
3137 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3140 cpu_buffer = buffer->buffers[cpu];
3141 atomic_dec(&cpu_buffer->record_disabled);
3143 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3146 * The total entries in the ring buffer is the running counter
3147 * of entries entered into the ring buffer, minus the sum of
3148 * the entries read from the ring buffer and the number of
3149 * entries that were overwritten.
3151 static inline unsigned long
3152 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3154 return local_read(&cpu_buffer->entries) -
3155 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3159 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3160 * @buffer: The ring buffer
3161 * @cpu: The per CPU buffer to read from.
3163 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3165 unsigned long flags;
3166 struct ring_buffer_per_cpu *cpu_buffer;
3167 struct buffer_page *bpage;
3170 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3173 cpu_buffer = buffer->buffers[cpu];
3174 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3176 * if the tail is on reader_page, oldest time stamp is on the reader
3179 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3180 bpage = cpu_buffer->reader_page;
3182 bpage = rb_set_head_page(cpu_buffer);
3184 ret = bpage->page->time_stamp;
3185 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3189 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3192 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3193 * @buffer: The ring buffer
3194 * @cpu: The per CPU buffer to read from.
3196 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3198 struct ring_buffer_per_cpu *cpu_buffer;
3201 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3204 cpu_buffer = buffer->buffers[cpu];
3205 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3209 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3212 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3213 * @buffer: The ring buffer
3214 * @cpu: The per CPU buffer to get the entries from.
3216 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3218 struct ring_buffer_per_cpu *cpu_buffer;
3220 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3223 cpu_buffer = buffer->buffers[cpu];
3225 return rb_num_of_entries(cpu_buffer);
3227 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3230 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3231 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3232 * @buffer: The ring buffer
3233 * @cpu: The per CPU buffer to get the number of overruns from
3235 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3237 struct ring_buffer_per_cpu *cpu_buffer;
3240 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3243 cpu_buffer = buffer->buffers[cpu];
3244 ret = local_read(&cpu_buffer->overrun);
3248 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3251 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3252 * commits failing due to the buffer wrapping around while there are uncommitted
3253 * events, such as during an interrupt storm.
3254 * @buffer: The ring buffer
3255 * @cpu: The per CPU buffer to get the number of overruns from
3258 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3260 struct ring_buffer_per_cpu *cpu_buffer;
3263 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3266 cpu_buffer = buffer->buffers[cpu];
3267 ret = local_read(&cpu_buffer->commit_overrun);
3271 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3274 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3275 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3276 * @buffer: The ring buffer
3277 * @cpu: The per CPU buffer to get the number of overruns from
3280 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3282 struct ring_buffer_per_cpu *cpu_buffer;
3285 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3288 cpu_buffer = buffer->buffers[cpu];
3289 ret = local_read(&cpu_buffer->dropped_events);
3293 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3296 * ring_buffer_read_events_cpu - get the number of events successfully read
3297 * @buffer: The ring buffer
3298 * @cpu: The per CPU buffer to get the number of events read
3301 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3303 struct ring_buffer_per_cpu *cpu_buffer;
3305 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3308 cpu_buffer = buffer->buffers[cpu];
3309 return cpu_buffer->read;
3311 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3314 * ring_buffer_entries - get the number of entries in a buffer
3315 * @buffer: The ring buffer
3317 * Returns the total number of entries in the ring buffer
3320 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3322 struct ring_buffer_per_cpu *cpu_buffer;
3323 unsigned long entries = 0;
3326 /* if you care about this being correct, lock the buffer */
3327 for_each_buffer_cpu(buffer, cpu) {
3328 cpu_buffer = buffer->buffers[cpu];
3329 entries += rb_num_of_entries(cpu_buffer);
3334 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3337 * ring_buffer_overruns - get the number of overruns in buffer
3338 * @buffer: The ring buffer
3340 * Returns the total number of overruns in the ring buffer
3343 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3345 struct ring_buffer_per_cpu *cpu_buffer;
3346 unsigned long overruns = 0;
3349 /* if you care about this being correct, lock the buffer */
3350 for_each_buffer_cpu(buffer, cpu) {
3351 cpu_buffer = buffer->buffers[cpu];
3352 overruns += local_read(&cpu_buffer->overrun);
3357 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3359 static void rb_iter_reset(struct ring_buffer_iter *iter)
3361 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3363 /* Iterator usage is expected to have record disabled */
3364 if (list_empty(&cpu_buffer->reader_page->list)) {
3365 iter->head_page = rb_set_head_page(cpu_buffer);
3366 if (unlikely(!iter->head_page))
3368 iter->head = iter->head_page->read;
3370 iter->head_page = cpu_buffer->reader_page;
3371 iter->head = cpu_buffer->reader_page->read;
3374 iter->read_stamp = cpu_buffer->read_stamp;
3376 iter->read_stamp = iter->head_page->page->time_stamp;
3377 iter->cache_reader_page = cpu_buffer->reader_page;
3378 iter->cache_read = cpu_buffer->read;
3382 * ring_buffer_iter_reset - reset an iterator
3383 * @iter: The iterator to reset
3385 * Resets the iterator, so that it will start from the beginning
3388 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3390 struct ring_buffer_per_cpu *cpu_buffer;
3391 unsigned long flags;
3396 cpu_buffer = iter->cpu_buffer;
3398 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3399 rb_iter_reset(iter);
3400 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3402 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3405 * ring_buffer_iter_empty - check if an iterator has no more to read
3406 * @iter: The iterator to check
3408 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3410 struct ring_buffer_per_cpu *cpu_buffer;
3412 cpu_buffer = iter->cpu_buffer;
3414 return iter->head_page == cpu_buffer->commit_page &&
3415 iter->head == rb_commit_index(cpu_buffer);
3417 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3420 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3421 struct ring_buffer_event *event)
3425 switch (event->type_len) {
3426 case RINGBUF_TYPE_PADDING:
3429 case RINGBUF_TYPE_TIME_EXTEND:
3430 delta = event->array[0];
3432 delta += event->time_delta;
3433 cpu_buffer->read_stamp += delta;
3436 case RINGBUF_TYPE_TIME_STAMP:
3437 /* FIXME: not implemented */
3440 case RINGBUF_TYPE_DATA:
3441 cpu_buffer->read_stamp += event->time_delta;
3451 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3452 struct ring_buffer_event *event)
3456 switch (event->type_len) {
3457 case RINGBUF_TYPE_PADDING:
3460 case RINGBUF_TYPE_TIME_EXTEND:
3461 delta = event->array[0];
3463 delta += event->time_delta;
3464 iter->read_stamp += delta;
3467 case RINGBUF_TYPE_TIME_STAMP:
3468 /* FIXME: not implemented */
3471 case RINGBUF_TYPE_DATA:
3472 iter->read_stamp += event->time_delta;
3481 static struct buffer_page *
3482 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3484 struct buffer_page *reader = NULL;
3485 unsigned long overwrite;
3486 unsigned long flags;
3490 local_irq_save(flags);
3491 arch_spin_lock(&cpu_buffer->lock);
3495 * This should normally only loop twice. But because the
3496 * start of the reader inserts an empty page, it causes
3497 * a case where we will loop three times. There should be no
3498 * reason to loop four times (that I know of).
3500 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3505 reader = cpu_buffer->reader_page;
3507 /* If there's more to read, return this page */
3508 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3511 /* Never should we have an index greater than the size */
3512 if (RB_WARN_ON(cpu_buffer,
3513 cpu_buffer->reader_page->read > rb_page_size(reader)))
3516 /* check if we caught up to the tail */
3518 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3521 /* Don't bother swapping if the ring buffer is empty */
3522 if (rb_num_of_entries(cpu_buffer) == 0)
3526 * Reset the reader page to size zero.
3528 local_set(&cpu_buffer->reader_page->write, 0);
3529 local_set(&cpu_buffer->reader_page->entries, 0);
3530 local_set(&cpu_buffer->reader_page->page->commit, 0);
3531 cpu_buffer->reader_page->real_end = 0;
3535 * Splice the empty reader page into the list around the head.
3537 reader = rb_set_head_page(cpu_buffer);
3540 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3541 cpu_buffer->reader_page->list.prev = reader->list.prev;
3544 * cpu_buffer->pages just needs to point to the buffer, it
3545 * has no specific buffer page to point to. Lets move it out
3546 * of our way so we don't accidentally swap it.
3548 cpu_buffer->pages = reader->list.prev;
3550 /* The reader page will be pointing to the new head */
3551 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3554 * We want to make sure we read the overruns after we set up our
3555 * pointers to the next object. The writer side does a
3556 * cmpxchg to cross pages which acts as the mb on the writer
3557 * side. Note, the reader will constantly fail the swap
3558 * while the writer is updating the pointers, so this
3559 * guarantees that the overwrite recorded here is the one we
3560 * want to compare with the last_overrun.
3563 overwrite = local_read(&(cpu_buffer->overrun));
3566 * Here's the tricky part.
3568 * We need to move the pointer past the header page.
3569 * But we can only do that if a writer is not currently
3570 * moving it. The page before the header page has the
3571 * flag bit '1' set if it is pointing to the page we want.
3572 * but if the writer is in the process of moving it
3573 * than it will be '2' or already moved '0'.
3576 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3579 * If we did not convert it, then we must try again.
3585 * Yeah! We succeeded in replacing the page.
3587 * Now make the new head point back to the reader page.
3589 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3590 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3592 /* Finally update the reader page to the new head */
3593 cpu_buffer->reader_page = reader;
3594 rb_reset_reader_page(cpu_buffer);
3596 if (overwrite != cpu_buffer->last_overrun) {
3597 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3598 cpu_buffer->last_overrun = overwrite;
3604 arch_spin_unlock(&cpu_buffer->lock);
3605 local_irq_restore(flags);
3610 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3612 struct ring_buffer_event *event;
3613 struct buffer_page *reader;
3616 reader = rb_get_reader_page(cpu_buffer);
3618 /* This function should not be called when buffer is empty */
3619 if (RB_WARN_ON(cpu_buffer, !reader))
3622 event = rb_reader_event(cpu_buffer);
3624 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3627 rb_update_read_stamp(cpu_buffer, event);
3629 length = rb_event_length(event);
3630 cpu_buffer->reader_page->read += length;
3633 static void rb_advance_iter(struct ring_buffer_iter *iter)
3635 struct ring_buffer_per_cpu *cpu_buffer;
3636 struct ring_buffer_event *event;
3639 cpu_buffer = iter->cpu_buffer;
3642 * Check if we are at the end of the buffer.
3644 if (iter->head >= rb_page_size(iter->head_page)) {
3645 /* discarded commits can make the page empty */
3646 if (iter->head_page == cpu_buffer->commit_page)
3652 event = rb_iter_head_event(iter);
3654 length = rb_event_length(event);
3657 * This should not be called to advance the header if we are
3658 * at the tail of the buffer.
3660 if (RB_WARN_ON(cpu_buffer,
3661 (iter->head_page == cpu_buffer->commit_page) &&
3662 (iter->head + length > rb_commit_index(cpu_buffer))))
3665 rb_update_iter_read_stamp(iter, event);
3667 iter->head += length;
3669 /* check for end of page padding */
3670 if ((iter->head >= rb_page_size(iter->head_page)) &&
3671 (iter->head_page != cpu_buffer->commit_page))
3675 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3677 return cpu_buffer->lost_events;
3680 static struct ring_buffer_event *
3681 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3682 unsigned long *lost_events)
3684 struct ring_buffer_event *event;
3685 struct buffer_page *reader;
3690 * We repeat when a time extend is encountered.
3691 * Since the time extend is always attached to a data event,
3692 * we should never loop more than once.
3693 * (We never hit the following condition more than twice).
3695 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3698 reader = rb_get_reader_page(cpu_buffer);
3702 event = rb_reader_event(cpu_buffer);
3704 switch (event->type_len) {
3705 case RINGBUF_TYPE_PADDING:
3706 if (rb_null_event(event))
3707 RB_WARN_ON(cpu_buffer, 1);
3709 * Because the writer could be discarding every
3710 * event it creates (which would probably be bad)
3711 * if we were to go back to "again" then we may never
3712 * catch up, and will trigger the warn on, or lock
3713 * the box. Return the padding, and we will release
3714 * the current locks, and try again.
3718 case RINGBUF_TYPE_TIME_EXTEND:
3719 /* Internal data, OK to advance */
3720 rb_advance_reader(cpu_buffer);
3723 case RINGBUF_TYPE_TIME_STAMP:
3724 /* FIXME: not implemented */
3725 rb_advance_reader(cpu_buffer);
3728 case RINGBUF_TYPE_DATA:
3730 *ts = cpu_buffer->read_stamp + event->time_delta;
3731 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3732 cpu_buffer->cpu, ts);
3735 *lost_events = rb_lost_events(cpu_buffer);
3744 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3746 static struct ring_buffer_event *
3747 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3749 struct ring_buffer *buffer;
3750 struct ring_buffer_per_cpu *cpu_buffer;
3751 struct ring_buffer_event *event;
3754 cpu_buffer = iter->cpu_buffer;
3755 buffer = cpu_buffer->buffer;
3758 * Check if someone performed a consuming read to
3759 * the buffer. A consuming read invalidates the iterator
3760 * and we need to reset the iterator in this case.
3762 if (unlikely(iter->cache_read != cpu_buffer->read ||
3763 iter->cache_reader_page != cpu_buffer->reader_page))
3764 rb_iter_reset(iter);
3767 if (ring_buffer_iter_empty(iter))
3771 * We repeat when a time extend is encountered.
3772 * Since the time extend is always attached to a data event,
3773 * we should never loop more than once.
3774 * (We never hit the following condition more than twice).
3776 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3779 if (rb_per_cpu_empty(cpu_buffer))
3782 if (iter->head >= local_read(&iter->head_page->page->commit)) {
3787 event = rb_iter_head_event(iter);
3789 switch (event->type_len) {
3790 case RINGBUF_TYPE_PADDING:
3791 if (rb_null_event(event)) {
3795 rb_advance_iter(iter);
3798 case RINGBUF_TYPE_TIME_EXTEND:
3799 /* Internal data, OK to advance */
3800 rb_advance_iter(iter);
3803 case RINGBUF_TYPE_TIME_STAMP:
3804 /* FIXME: not implemented */
3805 rb_advance_iter(iter);
3808 case RINGBUF_TYPE_DATA:
3810 *ts = iter->read_stamp + event->time_delta;
3811 ring_buffer_normalize_time_stamp(buffer,
3812 cpu_buffer->cpu, ts);
3822 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3824 static inline int rb_ok_to_lock(void)
3827 * If an NMI die dumps out the content of the ring buffer
3828 * do not grab locks. We also permanently disable the ring
3829 * buffer too. A one time deal is all you get from reading
3830 * the ring buffer from an NMI.
3832 if (likely(!in_nmi()))
3835 tracing_off_permanent();
3840 * ring_buffer_peek - peek at the next event to be read
3841 * @buffer: The ring buffer to read
3842 * @cpu: The cpu to peak at
3843 * @ts: The timestamp counter of this event.
3844 * @lost_events: a variable to store if events were lost (may be NULL)
3846 * This will return the event that will be read next, but does
3847 * not consume the data.
3849 struct ring_buffer_event *
3850 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3851 unsigned long *lost_events)
3853 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3854 struct ring_buffer_event *event;
3855 unsigned long flags;
3858 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3861 dolock = rb_ok_to_lock();
3863 local_irq_save(flags);
3865 raw_spin_lock(&cpu_buffer->reader_lock);
3866 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3867 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3868 rb_advance_reader(cpu_buffer);
3870 raw_spin_unlock(&cpu_buffer->reader_lock);
3871 local_irq_restore(flags);
3873 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3880 * ring_buffer_iter_peek - peek at the next event to be read
3881 * @iter: The ring buffer iterator
3882 * @ts: The timestamp counter of this event.
3884 * This will return the event that will be read next, but does
3885 * not increment the iterator.
3887 struct ring_buffer_event *
3888 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3890 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3891 struct ring_buffer_event *event;
3892 unsigned long flags;
3895 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3896 event = rb_iter_peek(iter, ts);
3897 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3899 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3906 * ring_buffer_consume - return an event and consume it
3907 * @buffer: The ring buffer to get the next event from
3908 * @cpu: the cpu to read the buffer from
3909 * @ts: a variable to store the timestamp (may be NULL)
3910 * @lost_events: a variable to store if events were lost (may be NULL)
3912 * Returns the next event in the ring buffer, and that event is consumed.
3913 * Meaning, that sequential reads will keep returning a different event,
3914 * and eventually empty the ring buffer if the producer is slower.
3916 struct ring_buffer_event *
3917 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3918 unsigned long *lost_events)
3920 struct ring_buffer_per_cpu *cpu_buffer;
3921 struct ring_buffer_event *event = NULL;
3922 unsigned long flags;
3925 dolock = rb_ok_to_lock();
3928 /* might be called in atomic */
3931 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3934 cpu_buffer = buffer->buffers[cpu];
3935 local_irq_save(flags);
3937 raw_spin_lock(&cpu_buffer->reader_lock);
3939 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3941 cpu_buffer->lost_events = 0;
3942 rb_advance_reader(cpu_buffer);
3946 raw_spin_unlock(&cpu_buffer->reader_lock);
3947 local_irq_restore(flags);
3952 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3957 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3960 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3961 * @buffer: The ring buffer to read from
3962 * @cpu: The cpu buffer to iterate over
3964 * This performs the initial preparations necessary to iterate
3965 * through the buffer. Memory is allocated, buffer recording
3966 * is disabled, and the iterator pointer is returned to the caller.
3968 * Disabling buffer recordng prevents the reading from being
3969 * corrupted. This is not a consuming read, so a producer is not
3972 * After a sequence of ring_buffer_read_prepare calls, the user is
3973 * expected to make at least one call to ring_buffer_read_prepare_sync.
3974 * Afterwards, ring_buffer_read_start is invoked to get things going
3977 * This overall must be paired with ring_buffer_read_finish.
3979 struct ring_buffer_iter *
3980 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3982 struct ring_buffer_per_cpu *cpu_buffer;
3983 struct ring_buffer_iter *iter;
3985 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3988 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3992 cpu_buffer = buffer->buffers[cpu];
3994 iter->cpu_buffer = cpu_buffer;
3996 atomic_inc(&buffer->resize_disabled);
3997 atomic_inc(&cpu_buffer->record_disabled);
4001 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4004 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4006 * All previously invoked ring_buffer_read_prepare calls to prepare
4007 * iterators will be synchronized. Afterwards, read_buffer_read_start
4008 * calls on those iterators are allowed.
4011 ring_buffer_read_prepare_sync(void)
4013 synchronize_sched();
4015 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4018 * ring_buffer_read_start - start a non consuming read of the buffer
4019 * @iter: The iterator returned by ring_buffer_read_prepare
4021 * This finalizes the startup of an iteration through the buffer.
4022 * The iterator comes from a call to ring_buffer_read_prepare and
4023 * an intervening ring_buffer_read_prepare_sync must have been
4026 * Must be paired with ring_buffer_read_finish.
4029 ring_buffer_read_start(struct ring_buffer_iter *iter)
4031 struct ring_buffer_per_cpu *cpu_buffer;
4032 unsigned long flags;
4037 cpu_buffer = iter->cpu_buffer;
4039 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4040 arch_spin_lock(&cpu_buffer->lock);
4041 rb_iter_reset(iter);
4042 arch_spin_unlock(&cpu_buffer->lock);
4043 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4045 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4048 * ring_buffer_read_finish - finish reading the iterator of the buffer
4049 * @iter: The iterator retrieved by ring_buffer_start
4051 * This re-enables the recording to the buffer, and frees the
4055 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4057 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4058 unsigned long flags;
4061 * Ring buffer is disabled from recording, here's a good place
4062 * to check the integrity of the ring buffer.
4063 * Must prevent readers from trying to read, as the check
4064 * clears the HEAD page and readers require it.
4066 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4067 rb_check_pages(cpu_buffer);
4068 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4070 atomic_dec(&cpu_buffer->record_disabled);
4071 atomic_dec(&cpu_buffer->buffer->resize_disabled);
4074 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4077 * ring_buffer_read - read the next item in the ring buffer by the iterator
4078 * @iter: The ring buffer iterator
4079 * @ts: The time stamp of the event read.
4081 * This reads the next event in the ring buffer and increments the iterator.
4083 struct ring_buffer_event *
4084 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4086 struct ring_buffer_event *event;
4087 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4088 unsigned long flags;
4090 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4092 event = rb_iter_peek(iter, ts);
4096 if (event->type_len == RINGBUF_TYPE_PADDING)
4099 rb_advance_iter(iter);
4101 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4105 EXPORT_SYMBOL_GPL(ring_buffer_read);
4108 * ring_buffer_size - return the size of the ring buffer (in bytes)
4109 * @buffer: The ring buffer.
4111 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4114 * Earlier, this method returned
4115 * BUF_PAGE_SIZE * buffer->nr_pages
4116 * Since the nr_pages field is now removed, we have converted this to
4117 * return the per cpu buffer value.
4119 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4122 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4124 EXPORT_SYMBOL_GPL(ring_buffer_size);
4127 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4129 rb_head_page_deactivate(cpu_buffer);
4131 cpu_buffer->head_page
4132 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4133 local_set(&cpu_buffer->head_page->write, 0);
4134 local_set(&cpu_buffer->head_page->entries, 0);
4135 local_set(&cpu_buffer->head_page->page->commit, 0);
4137 cpu_buffer->head_page->read = 0;
4139 cpu_buffer->tail_page = cpu_buffer->head_page;
4140 cpu_buffer->commit_page = cpu_buffer->head_page;
4142 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4143 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4144 local_set(&cpu_buffer->reader_page->write, 0);
4145 local_set(&cpu_buffer->reader_page->entries, 0);
4146 local_set(&cpu_buffer->reader_page->page->commit, 0);
4147 cpu_buffer->reader_page->read = 0;
4149 local_set(&cpu_buffer->entries_bytes, 0);
4150 local_set(&cpu_buffer->overrun, 0);
4151 local_set(&cpu_buffer->commit_overrun, 0);
4152 local_set(&cpu_buffer->dropped_events, 0);
4153 local_set(&cpu_buffer->entries, 0);
4154 local_set(&cpu_buffer->committing, 0);
4155 local_set(&cpu_buffer->commits, 0);
4156 cpu_buffer->read = 0;
4157 cpu_buffer->read_bytes = 0;
4159 cpu_buffer->write_stamp = 0;
4160 cpu_buffer->read_stamp = 0;
4162 cpu_buffer->lost_events = 0;
4163 cpu_buffer->last_overrun = 0;
4165 rb_head_page_activate(cpu_buffer);
4169 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4170 * @buffer: The ring buffer to reset a per cpu buffer of
4171 * @cpu: The CPU buffer to be reset
4173 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4175 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4176 unsigned long flags;
4178 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4181 atomic_inc(&buffer->resize_disabled);
4182 atomic_inc(&cpu_buffer->record_disabled);
4184 /* Make sure all commits have finished */
4185 synchronize_sched();
4187 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4189 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4192 arch_spin_lock(&cpu_buffer->lock);
4194 rb_reset_cpu(cpu_buffer);
4196 arch_spin_unlock(&cpu_buffer->lock);
4199 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4201 atomic_dec(&cpu_buffer->record_disabled);
4202 atomic_dec(&buffer->resize_disabled);
4204 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4207 * ring_buffer_reset - reset a ring buffer
4208 * @buffer: The ring buffer to reset all cpu buffers
4210 void ring_buffer_reset(struct ring_buffer *buffer)
4214 for_each_buffer_cpu(buffer, cpu)
4215 ring_buffer_reset_cpu(buffer, cpu);
4217 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4220 * rind_buffer_empty - is the ring buffer empty?
4221 * @buffer: The ring buffer to test
4223 int ring_buffer_empty(struct ring_buffer *buffer)
4225 struct ring_buffer_per_cpu *cpu_buffer;
4226 unsigned long flags;
4231 dolock = rb_ok_to_lock();
4233 /* yes this is racy, but if you don't like the race, lock the buffer */
4234 for_each_buffer_cpu(buffer, cpu) {
4235 cpu_buffer = buffer->buffers[cpu];
4236 local_irq_save(flags);
4238 raw_spin_lock(&cpu_buffer->reader_lock);
4239 ret = rb_per_cpu_empty(cpu_buffer);
4241 raw_spin_unlock(&cpu_buffer->reader_lock);
4242 local_irq_restore(flags);
4250 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4253 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4254 * @buffer: The ring buffer
4255 * @cpu: The CPU buffer to test
4257 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4259 struct ring_buffer_per_cpu *cpu_buffer;
4260 unsigned long flags;
4264 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4267 dolock = rb_ok_to_lock();
4269 cpu_buffer = buffer->buffers[cpu];
4270 local_irq_save(flags);
4272 raw_spin_lock(&cpu_buffer->reader_lock);
4273 ret = rb_per_cpu_empty(cpu_buffer);
4275 raw_spin_unlock(&cpu_buffer->reader_lock);
4276 local_irq_restore(flags);
4280 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4282 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4284 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4285 * @buffer_a: One buffer to swap with
4286 * @buffer_b: The other buffer to swap with
4288 * This function is useful for tracers that want to take a "snapshot"
4289 * of a CPU buffer and has another back up buffer lying around.
4290 * it is expected that the tracer handles the cpu buffer not being
4291 * used at the moment.
4293 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4294 struct ring_buffer *buffer_b, int cpu)
4296 struct ring_buffer_per_cpu *cpu_buffer_a;
4297 struct ring_buffer_per_cpu *cpu_buffer_b;
4300 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4301 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4304 cpu_buffer_a = buffer_a->buffers[cpu];
4305 cpu_buffer_b = buffer_b->buffers[cpu];
4307 /* At least make sure the two buffers are somewhat the same */
4308 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4313 if (ring_buffer_flags != RB_BUFFERS_ON)
4316 if (atomic_read(&buffer_a->record_disabled))
4319 if (atomic_read(&buffer_b->record_disabled))
4322 if (atomic_read(&cpu_buffer_a->record_disabled))
4325 if (atomic_read(&cpu_buffer_b->record_disabled))
4329 * We can't do a synchronize_sched here because this
4330 * function can be called in atomic context.
4331 * Normally this will be called from the same CPU as cpu.
4332 * If not it's up to the caller to protect this.
4334 atomic_inc(&cpu_buffer_a->record_disabled);
4335 atomic_inc(&cpu_buffer_b->record_disabled);
4338 if (local_read(&cpu_buffer_a->committing))
4340 if (local_read(&cpu_buffer_b->committing))
4343 buffer_a->buffers[cpu] = cpu_buffer_b;
4344 buffer_b->buffers[cpu] = cpu_buffer_a;
4346 cpu_buffer_b->buffer = buffer_a;
4347 cpu_buffer_a->buffer = buffer_b;
4352 atomic_dec(&cpu_buffer_a->record_disabled);
4353 atomic_dec(&cpu_buffer_b->record_disabled);
4357 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4358 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4361 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4362 * @buffer: the buffer to allocate for.
4363 * @cpu: the cpu buffer to allocate.
4365 * This function is used in conjunction with ring_buffer_read_page.
4366 * When reading a full page from the ring buffer, these functions
4367 * can be used to speed up the process. The calling function should
4368 * allocate a few pages first with this function. Then when it
4369 * needs to get pages from the ring buffer, it passes the result
4370 * of this function into ring_buffer_read_page, which will swap
4371 * the page that was allocated, with the read page of the buffer.
4374 * The page allocated, or NULL on error.
4376 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4378 struct buffer_data_page *bpage;
4381 page = alloc_pages_node(cpu_to_node(cpu),
4382 GFP_KERNEL | __GFP_NORETRY, 0);
4386 bpage = page_address(page);
4388 rb_init_page(bpage);
4392 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4395 * ring_buffer_free_read_page - free an allocated read page
4396 * @buffer: the buffer the page was allocate for
4397 * @data: the page to free
4399 * Free a page allocated from ring_buffer_alloc_read_page.
4401 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4403 free_page((unsigned long)data);
4405 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4408 * ring_buffer_read_page - extract a page from the ring buffer
4409 * @buffer: buffer to extract from
4410 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4411 * @len: amount to extract
4412 * @cpu: the cpu of the buffer to extract
4413 * @full: should the extraction only happen when the page is full.
4415 * This function will pull out a page from the ring buffer and consume it.
4416 * @data_page must be the address of the variable that was returned
4417 * from ring_buffer_alloc_read_page. This is because the page might be used
4418 * to swap with a page in the ring buffer.
4421 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4424 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4426 * process_page(rpage, ret);
4428 * When @full is set, the function will not return true unless
4429 * the writer is off the reader page.
4431 * Note: it is up to the calling functions to handle sleeps and wakeups.
4432 * The ring buffer can be used anywhere in the kernel and can not
4433 * blindly call wake_up. The layer that uses the ring buffer must be
4434 * responsible for that.
4437 * >=0 if data has been transferred, returns the offset of consumed data.
4438 * <0 if no data has been transferred.
4440 int ring_buffer_read_page(struct ring_buffer *buffer,
4441 void **data_page, size_t len, int cpu, int full)
4443 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4444 struct ring_buffer_event *event;
4445 struct buffer_data_page *bpage;
4446 struct buffer_page *reader;
4447 unsigned long missed_events;
4448 unsigned long flags;
4449 unsigned int commit;
4454 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4458 * If len is not big enough to hold the page header, then
4459 * we can not copy anything.
4461 if (len <= BUF_PAGE_HDR_SIZE)
4464 len -= BUF_PAGE_HDR_SIZE;
4473 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4475 reader = rb_get_reader_page(cpu_buffer);
4479 event = rb_reader_event(cpu_buffer);
4481 read = reader->read;
4482 commit = rb_page_commit(reader);
4484 /* Check if any events were dropped */
4485 missed_events = cpu_buffer->lost_events;
4488 * If this page has been partially read or
4489 * if len is not big enough to read the rest of the page or
4490 * a writer is still on the page, then
4491 * we must copy the data from the page to the buffer.
4492 * Otherwise, we can simply swap the page with the one passed in.
4494 if (read || (len < (commit - read)) ||
4495 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4496 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4497 unsigned int rpos = read;
4498 unsigned int pos = 0;
4504 if (len > (commit - read))
4505 len = (commit - read);
4507 /* Always keep the time extend and data together */
4508 size = rb_event_ts_length(event);
4513 /* save the current timestamp, since the user will need it */
4514 save_timestamp = cpu_buffer->read_stamp;
4516 /* Need to copy one event at a time */
4518 /* We need the size of one event, because
4519 * rb_advance_reader only advances by one event,
4520 * whereas rb_event_ts_length may include the size of
4521 * one or two events.
4522 * We have already ensured there's enough space if this
4523 * is a time extend. */
4524 size = rb_event_length(event);
4525 memcpy(bpage->data + pos, rpage->data + rpos, size);
4529 rb_advance_reader(cpu_buffer);
4530 rpos = reader->read;
4536 event = rb_reader_event(cpu_buffer);
4537 /* Always keep the time extend and data together */
4538 size = rb_event_ts_length(event);
4539 } while (len >= size);
4542 local_set(&bpage->commit, pos);
4543 bpage->time_stamp = save_timestamp;
4545 /* we copied everything to the beginning */
4548 /* update the entry counter */
4549 cpu_buffer->read += rb_page_entries(reader);
4550 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4552 /* swap the pages */
4553 rb_init_page(bpage);
4554 bpage = reader->page;
4555 reader->page = *data_page;
4556 local_set(&reader->write, 0);
4557 local_set(&reader->entries, 0);
4562 * Use the real_end for the data size,
4563 * This gives us a chance to store the lost events
4566 if (reader->real_end)
4567 local_set(&bpage->commit, reader->real_end);
4571 cpu_buffer->lost_events = 0;
4573 commit = local_read(&bpage->commit);
4575 * Set a flag in the commit field if we lost events
4577 if (missed_events) {
4578 /* If there is room at the end of the page to save the
4579 * missed events, then record it there.
4581 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4582 memcpy(&bpage->data[commit], &missed_events,
4583 sizeof(missed_events));
4584 local_add(RB_MISSED_STORED, &bpage->commit);
4585 commit += sizeof(missed_events);
4587 local_add(RB_MISSED_EVENTS, &bpage->commit);
4591 * This page may be off to user land. Zero it out here.
4593 if (commit < BUF_PAGE_SIZE)
4594 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4597 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4602 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4604 #ifdef CONFIG_HOTPLUG_CPU
4605 static int rb_cpu_notify(struct notifier_block *self,
4606 unsigned long action, void *hcpu)
4608 struct ring_buffer *buffer =
4609 container_of(self, struct ring_buffer, cpu_notify);
4610 long cpu = (long)hcpu;
4611 int cpu_i, nr_pages_same;
4612 unsigned int nr_pages;
4615 case CPU_UP_PREPARE:
4616 case CPU_UP_PREPARE_FROZEN:
4617 if (cpumask_test_cpu(cpu, buffer->cpumask))
4622 /* check if all cpu sizes are same */
4623 for_each_buffer_cpu(buffer, cpu_i) {
4624 /* fill in the size from first enabled cpu */
4626 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4627 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4632 /* allocate minimum pages, user can later expand it */
4635 buffer->buffers[cpu] =
4636 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4637 if (!buffer->buffers[cpu]) {
4638 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4643 cpumask_set_cpu(cpu, buffer->cpumask);
4645 case CPU_DOWN_PREPARE:
4646 case CPU_DOWN_PREPARE_FROZEN:
4649 * If we were to free the buffer, then the user would
4650 * lose any trace that was in the buffer.
4660 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4662 * This is a basic integrity check of the ring buffer.
4663 * Late in the boot cycle this test will run when configured in.
4664 * It will kick off a thread per CPU that will go into a loop
4665 * writing to the per cpu ring buffer various sizes of data.
4666 * Some of the data will be large items, some small.
4668 * Another thread is created that goes into a spin, sending out
4669 * IPIs to the other CPUs to also write into the ring buffer.
4670 * this is to test the nesting ability of the buffer.
4672 * Basic stats are recorded and reported. If something in the
4673 * ring buffer should happen that's not expected, a big warning
4674 * is displayed and all ring buffers are disabled.
4676 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4678 struct rb_test_data {
4679 struct ring_buffer *buffer;
4680 unsigned long events;
4681 unsigned long bytes_written;
4682 unsigned long bytes_alloc;
4683 unsigned long bytes_dropped;
4684 unsigned long events_nested;
4685 unsigned long bytes_written_nested;
4686 unsigned long bytes_alloc_nested;
4687 unsigned long bytes_dropped_nested;
4688 int min_size_nested;
4689 int max_size_nested;
4696 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4699 #define RB_TEST_BUFFER_SIZE 1048576
4701 static char rb_string[] __initdata =
4702 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4703 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4704 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4706 static bool rb_test_started __initdata;
4713 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4715 struct ring_buffer_event *event;
4716 struct rb_item *item;
4723 /* Have nested writes different that what is written */
4724 cnt = data->cnt + (nested ? 27 : 0);
4726 /* Multiply cnt by ~e, to make some unique increment */
4727 size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4729 len = size + sizeof(struct rb_item);
4731 started = rb_test_started;
4732 /* read rb_test_started before checking buffer enabled */
4735 event = ring_buffer_lock_reserve(data->buffer, len);
4737 /* Ignore dropped events before test starts. */
4740 data->bytes_dropped += len;
4742 data->bytes_dropped_nested += len;
4747 event_len = ring_buffer_event_length(event);
4749 if (RB_WARN_ON(data->buffer, event_len < len))
4752 item = ring_buffer_event_data(event);
4754 memcpy(item->str, rb_string, size);
4757 data->bytes_alloc_nested += event_len;
4758 data->bytes_written_nested += len;
4759 data->events_nested++;
4760 if (!data->min_size_nested || len < data->min_size_nested)
4761 data->min_size_nested = len;
4762 if (len > data->max_size_nested)
4763 data->max_size_nested = len;
4765 data->bytes_alloc += event_len;
4766 data->bytes_written += len;
4768 if (!data->min_size || len < data->min_size)
4769 data->max_size = len;
4770 if (len > data->max_size)
4771 data->max_size = len;
4775 ring_buffer_unlock_commit(data->buffer, event);
4780 static __init int rb_test(void *arg)
4782 struct rb_test_data *data = arg;
4784 while (!kthread_should_stop()) {
4785 rb_write_something(data, false);
4788 set_current_state(TASK_INTERRUPTIBLE);
4789 /* Now sleep between a min of 100-300us and a max of 1ms */
4790 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4796 static __init void rb_ipi(void *ignore)
4798 struct rb_test_data *data;
4799 int cpu = smp_processor_id();
4801 data = &rb_data[cpu];
4802 rb_write_something(data, true);
4805 static __init int rb_hammer_test(void *arg)
4807 while (!kthread_should_stop()) {
4809 /* Send an IPI to all cpus to write data! */
4810 smp_call_function(rb_ipi, NULL, 1);
4811 /* No sleep, but for non preempt, let others run */
4818 static __init int test_ringbuffer(void)
4820 struct task_struct *rb_hammer;
4821 struct ring_buffer *buffer;
4825 pr_info("Running ring buffer tests...\n");
4827 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4828 if (WARN_ON(!buffer))
4831 /* Disable buffer so that threads can't write to it yet */
4832 ring_buffer_record_off(buffer);
4834 for_each_online_cpu(cpu) {
4835 rb_data[cpu].buffer = buffer;
4836 rb_data[cpu].cpu = cpu;
4837 rb_data[cpu].cnt = cpu;
4838 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4839 "rbtester/%d", cpu);
4840 if (WARN_ON(!rb_threads[cpu])) {
4841 pr_cont("FAILED\n");
4846 kthread_bind(rb_threads[cpu], cpu);
4847 wake_up_process(rb_threads[cpu]);
4850 /* Now create the rb hammer! */
4851 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4852 if (WARN_ON(!rb_hammer)) {
4853 pr_cont("FAILED\n");
4858 ring_buffer_record_on(buffer);
4860 * Show buffer is enabled before setting rb_test_started.
4861 * Yes there's a small race window where events could be
4862 * dropped and the thread wont catch it. But when a ring
4863 * buffer gets enabled, there will always be some kind of
4864 * delay before other CPUs see it. Thus, we don't care about
4865 * those dropped events. We care about events dropped after
4866 * the threads see that the buffer is active.
4869 rb_test_started = true;
4871 set_current_state(TASK_INTERRUPTIBLE);
4872 /* Just run for 10 seconds */;
4873 schedule_timeout(10 * HZ);
4875 kthread_stop(rb_hammer);
4878 for_each_online_cpu(cpu) {
4879 if (!rb_threads[cpu])
4881 kthread_stop(rb_threads[cpu]);
4884 ring_buffer_free(buffer);
4889 pr_info("finished\n");
4890 for_each_online_cpu(cpu) {
4891 struct ring_buffer_event *event;
4892 struct rb_test_data *data = &rb_data[cpu];
4893 struct rb_item *item;
4894 unsigned long total_events;
4895 unsigned long total_dropped;
4896 unsigned long total_written;
4897 unsigned long total_alloc;
4898 unsigned long total_read = 0;
4899 unsigned long total_size = 0;
4900 unsigned long total_len = 0;
4901 unsigned long total_lost = 0;
4904 int small_event_size;
4908 total_events = data->events + data->events_nested;
4909 total_written = data->bytes_written + data->bytes_written_nested;
4910 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
4911 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
4913 big_event_size = data->max_size + data->max_size_nested;
4914 small_event_size = data->min_size + data->min_size_nested;
4916 pr_info("CPU %d:\n", cpu);
4917 pr_info(" events: %ld\n", total_events);
4918 pr_info(" dropped bytes: %ld\n", total_dropped);
4919 pr_info(" alloced bytes: %ld\n", total_alloc);
4920 pr_info(" written bytes: %ld\n", total_written);
4921 pr_info(" biggest event: %d\n", big_event_size);
4922 pr_info(" smallest event: %d\n", small_event_size);
4924 if (RB_WARN_ON(buffer, total_dropped))
4929 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
4931 item = ring_buffer_event_data(event);
4932 total_len += ring_buffer_event_length(event);
4933 total_size += item->size + sizeof(struct rb_item);
4934 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
4935 pr_info("FAILED!\n");
4936 pr_info("buffer had: %.*s\n", item->size, item->str);
4937 pr_info("expected: %.*s\n", item->size, rb_string);
4938 RB_WARN_ON(buffer, 1);
4949 pr_info(" read events: %ld\n", total_read);
4950 pr_info(" lost events: %ld\n", total_lost);
4951 pr_info(" total events: %ld\n", total_lost + total_read);
4952 pr_info(" recorded len bytes: %ld\n", total_len);
4953 pr_info(" recorded size bytes: %ld\n", total_size);
4955 pr_info(" With dropped events, record len and size may not match\n"
4956 " alloced and written from above\n");
4958 if (RB_WARN_ON(buffer, total_len != total_alloc ||
4959 total_size != total_written))
4962 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
4968 pr_info("Ring buffer PASSED!\n");
4970 ring_buffer_free(buffer);
4974 late_initcall(test_ringbuffer);
4975 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */