4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 #include <linux/trace_events.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/uaccess.h>
13 #include <linux/hardirq.h>
14 #include <linux/kthread.h> /* for self test */
15 #include <linux/kmemcheck.h>
16 #include <linux/module.h>
17 #include <linux/percpu.h>
18 #include <linux/mutex.h>
19 #include <linux/delay.h>
20 #include <linux/slab.h>
21 #include <linux/init.h>
22 #include <linux/hash.h>
23 #include <linux/list.h>
24 #include <linux/cpu.h>
26 #include <asm/local.h>
28 static void update_pages_handler(struct work_struct *work);
31 * The ring buffer header is special. We must manually up keep it.
33 int ring_buffer_print_entry_header(struct trace_seq *s)
35 trace_seq_puts(s, "# compressed entry header\n");
36 trace_seq_puts(s, "\ttype_len : 5 bits\n");
37 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
38 trace_seq_puts(s, "\tarray : 32 bits\n");
39 trace_seq_putc(s, '\n');
40 trace_seq_printf(s, "\tpadding : type == %d\n",
41 RINGBUF_TYPE_PADDING);
42 trace_seq_printf(s, "\ttime_extend : type == %d\n",
43 RINGBUF_TYPE_TIME_EXTEND);
44 trace_seq_printf(s, "\tdata max type_len == %d\n",
45 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
47 return !trace_seq_has_overflowed(s);
51 * The ring buffer is made up of a list of pages. A separate list of pages is
52 * allocated for each CPU. A writer may only write to a buffer that is
53 * associated with the CPU it is currently executing on. A reader may read
54 * from any per cpu buffer.
56 * The reader is special. For each per cpu buffer, the reader has its own
57 * reader page. When a reader has read the entire reader page, this reader
58 * page is swapped with another page in the ring buffer.
60 * Now, as long as the writer is off the reader page, the reader can do what
61 * ever it wants with that page. The writer will never write to that page
62 * again (as long as it is out of the ring buffer).
64 * Here's some silly ASCII art.
67 * |reader| RING BUFFER
69 * +------+ +---+ +---+ +---+
78 * |reader| RING BUFFER
79 * |page |------------------v
80 * +------+ +---+ +---+ +---+
89 * |reader| RING BUFFER
90 * |page |------------------v
91 * +------+ +---+ +---+ +---+
96 * +------------------------------+
100 * |buffer| RING BUFFER
101 * |page |------------------v
102 * +------+ +---+ +---+ +---+
104 * | New +---+ +---+ +---+
107 * +------------------------------+
110 * After we make this swap, the reader can hand this page off to the splice
111 * code and be done with it. It can even allocate a new page if it needs to
112 * and swap that into the ring buffer.
114 * We will be using cmpxchg soon to make all this lockless.
119 * A fast way to enable or disable all ring buffers is to
120 * call tracing_on or tracing_off. Turning off the ring buffers
121 * prevents all ring buffers from being recorded to.
122 * Turning this switch on, makes it OK to write to the
123 * ring buffer, if the ring buffer is enabled itself.
125 * There's three layers that must be on in order to write
126 * to the ring buffer.
128 * 1) This global flag must be set.
129 * 2) The ring buffer must be enabled for recording.
130 * 3) The per cpu buffer must be enabled for recording.
132 * In case of an anomaly, this global flag has a bit set that
133 * will permantly disable all ring buffers.
137 * Global flag to disable all recording to ring buffers
138 * This has two bits: ON, DISABLED
142 * 0 0 : ring buffers are off
143 * 1 0 : ring buffers are on
144 * X 1 : ring buffers are permanently disabled
148 RB_BUFFERS_ON_BIT = 0,
149 RB_BUFFERS_DISABLED_BIT = 1,
153 RB_BUFFERS_ON = 1 << RB_BUFFERS_ON_BIT,
154 RB_BUFFERS_DISABLED = 1 << RB_BUFFERS_DISABLED_BIT,
157 static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON;
159 /* Used for individual buffers (after the counter) */
160 #define RB_BUFFER_OFF (1 << 20)
162 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
165 * tracing_off_permanent - permanently disable ring buffers
167 * This function, once called, will disable all ring buffers
170 void tracing_off_permanent(void)
172 set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags);
175 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
176 #define RB_ALIGNMENT 4U
177 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
178 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
180 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
181 # define RB_FORCE_8BYTE_ALIGNMENT 0
182 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
184 # define RB_FORCE_8BYTE_ALIGNMENT 1
185 # define RB_ARCH_ALIGNMENT 8U
188 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
190 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
191 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
194 RB_LEN_TIME_EXTEND = 8,
195 RB_LEN_TIME_STAMP = 16,
198 #define skip_time_extend(event) \
199 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
201 static inline int rb_null_event(struct ring_buffer_event *event)
203 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
206 static void rb_event_set_padding(struct ring_buffer_event *event)
208 /* padding has a NULL time_delta */
209 event->type_len = RINGBUF_TYPE_PADDING;
210 event->time_delta = 0;
214 rb_event_data_length(struct ring_buffer_event *event)
219 length = event->type_len * RB_ALIGNMENT;
221 length = event->array[0];
222 return length + RB_EVNT_HDR_SIZE;
226 * Return the length of the given event. Will return
227 * the length of the time extend if the event is a
230 static inline unsigned
231 rb_event_length(struct ring_buffer_event *event)
233 switch (event->type_len) {
234 case RINGBUF_TYPE_PADDING:
235 if (rb_null_event(event))
238 return event->array[0] + RB_EVNT_HDR_SIZE;
240 case RINGBUF_TYPE_TIME_EXTEND:
241 return RB_LEN_TIME_EXTEND;
243 case RINGBUF_TYPE_TIME_STAMP:
244 return RB_LEN_TIME_STAMP;
246 case RINGBUF_TYPE_DATA:
247 return rb_event_data_length(event);
256 * Return total length of time extend and data,
257 * or just the event length for all other events.
259 static inline unsigned
260 rb_event_ts_length(struct ring_buffer_event *event)
264 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
265 /* time extends include the data event after it */
266 len = RB_LEN_TIME_EXTEND;
267 event = skip_time_extend(event);
269 return len + rb_event_length(event);
273 * ring_buffer_event_length - return the length of the event
274 * @event: the event to get the length of
276 * Returns the size of the data load of a data event.
277 * If the event is something other than a data event, it
278 * returns the size of the event itself. With the exception
279 * of a TIME EXTEND, where it still returns the size of the
280 * data load of the data event after it.
282 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
286 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
287 event = skip_time_extend(event);
289 length = rb_event_length(event);
290 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
292 length -= RB_EVNT_HDR_SIZE;
293 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
294 length -= sizeof(event->array[0]);
297 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
299 /* inline for ring buffer fast paths */
301 rb_event_data(struct ring_buffer_event *event)
303 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
304 event = skip_time_extend(event);
305 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
306 /* If length is in len field, then array[0] has the data */
308 return (void *)&event->array[0];
309 /* Otherwise length is in array[0] and array[1] has the data */
310 return (void *)&event->array[1];
314 * ring_buffer_event_data - return the data of the event
315 * @event: the event to get the data from
317 void *ring_buffer_event_data(struct ring_buffer_event *event)
319 return rb_event_data(event);
321 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
323 #define for_each_buffer_cpu(buffer, cpu) \
324 for_each_cpu(cpu, buffer->cpumask)
327 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
328 #define TS_DELTA_TEST (~TS_MASK)
330 /* Flag when events were overwritten */
331 #define RB_MISSED_EVENTS (1 << 31)
332 /* Missed count stored at end */
333 #define RB_MISSED_STORED (1 << 30)
335 struct buffer_data_page {
336 u64 time_stamp; /* page time stamp */
337 local_t commit; /* write committed index */
338 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
342 * Note, the buffer_page list must be first. The buffer pages
343 * are allocated in cache lines, which means that each buffer
344 * page will be at the beginning of a cache line, and thus
345 * the least significant bits will be zero. We use this to
346 * add flags in the list struct pointers, to make the ring buffer
350 struct list_head list; /* list of buffer pages */
351 local_t write; /* index for next write */
352 unsigned read; /* index for next read */
353 local_t entries; /* entries on this page */
354 unsigned long real_end; /* real end of data */
355 struct buffer_data_page *page; /* Actual data page */
359 * The buffer page counters, write and entries, must be reset
360 * atomically when crossing page boundaries. To synchronize this
361 * update, two counters are inserted into the number. One is
362 * the actual counter for the write position or count on the page.
364 * The other is a counter of updaters. Before an update happens
365 * the update partition of the counter is incremented. This will
366 * allow the updater to update the counter atomically.
368 * The counter is 20 bits, and the state data is 12.
370 #define RB_WRITE_MASK 0xfffff
371 #define RB_WRITE_INTCNT (1 << 20)
373 static void rb_init_page(struct buffer_data_page *bpage)
375 local_set(&bpage->commit, 0);
379 * ring_buffer_page_len - the size of data on the page.
380 * @page: The page to read
382 * Returns the amount of data on the page, including buffer page header.
384 size_t ring_buffer_page_len(void *page)
386 return local_read(&((struct buffer_data_page *)page)->commit)
391 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
394 static void free_buffer_page(struct buffer_page *bpage)
396 free_page((unsigned long)bpage->page);
401 * We need to fit the time_stamp delta into 27 bits.
403 static inline int test_time_stamp(u64 delta)
405 if (delta & TS_DELTA_TEST)
410 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
412 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
413 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
415 int ring_buffer_print_page_header(struct trace_seq *s)
417 struct buffer_data_page field;
419 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
420 "offset:0;\tsize:%u;\tsigned:%u;\n",
421 (unsigned int)sizeof(field.time_stamp),
422 (unsigned int)is_signed_type(u64));
424 trace_seq_printf(s, "\tfield: local_t commit;\t"
425 "offset:%u;\tsize:%u;\tsigned:%u;\n",
426 (unsigned int)offsetof(typeof(field), commit),
427 (unsigned int)sizeof(field.commit),
428 (unsigned int)is_signed_type(long));
430 trace_seq_printf(s, "\tfield: int overwrite;\t"
431 "offset:%u;\tsize:%u;\tsigned:%u;\n",
432 (unsigned int)offsetof(typeof(field), commit),
434 (unsigned int)is_signed_type(long));
436 trace_seq_printf(s, "\tfield: char data;\t"
437 "offset:%u;\tsize:%u;\tsigned:%u;\n",
438 (unsigned int)offsetof(typeof(field), data),
439 (unsigned int)BUF_PAGE_SIZE,
440 (unsigned int)is_signed_type(char));
442 return !trace_seq_has_overflowed(s);
446 struct irq_work work;
447 wait_queue_head_t waiters;
448 wait_queue_head_t full_waiters;
449 bool waiters_pending;
450 bool full_waiters_pending;
455 * head_page == tail_page && head == tail then buffer is empty.
457 struct ring_buffer_per_cpu {
459 atomic_t record_disabled;
460 struct ring_buffer *buffer;
461 raw_spinlock_t reader_lock; /* serialize readers */
462 arch_spinlock_t lock;
463 struct lock_class_key lock_key;
464 unsigned int nr_pages;
465 struct list_head *pages;
466 struct buffer_page *head_page; /* read from head */
467 struct buffer_page *tail_page; /* write to tail */
468 struct buffer_page *commit_page; /* committed pages */
469 struct buffer_page *reader_page;
470 unsigned long lost_events;
471 unsigned long last_overrun;
472 local_t entries_bytes;
475 local_t commit_overrun;
476 local_t dropped_events;
480 unsigned long read_bytes;
483 /* ring buffer pages to update, > 0 to add, < 0 to remove */
484 int nr_pages_to_update;
485 struct list_head new_pages; /* new pages to add */
486 struct work_struct update_pages_work;
487 struct completion update_done;
489 struct rb_irq_work irq_work;
495 atomic_t record_disabled;
496 atomic_t resize_disabled;
497 cpumask_var_t cpumask;
499 struct lock_class_key *reader_lock_key;
503 struct ring_buffer_per_cpu **buffers;
505 #ifdef CONFIG_HOTPLUG_CPU
506 struct notifier_block cpu_notify;
510 struct rb_irq_work irq_work;
513 struct ring_buffer_iter {
514 struct ring_buffer_per_cpu *cpu_buffer;
516 struct buffer_page *head_page;
517 struct buffer_page *cache_reader_page;
518 unsigned long cache_read;
523 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
525 * Schedules a delayed work to wake up any task that is blocked on the
526 * ring buffer waiters queue.
528 static void rb_wake_up_waiters(struct irq_work *work)
530 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
532 wake_up_all(&rbwork->waiters);
533 if (rbwork->wakeup_full) {
534 rbwork->wakeup_full = false;
535 wake_up_all(&rbwork->full_waiters);
540 * ring_buffer_wait - wait for input to the ring buffer
541 * @buffer: buffer to wait on
542 * @cpu: the cpu buffer to wait on
543 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
545 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
546 * as data is added to any of the @buffer's cpu buffers. Otherwise
547 * it will wait for data to be added to a specific cpu buffer.
549 int ring_buffer_wait(struct ring_buffer *buffer, int cpu, bool full)
551 struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
553 struct rb_irq_work *work;
557 * Depending on what the caller is waiting for, either any
558 * data in any cpu buffer, or a specific buffer, put the
559 * caller on the appropriate wait queue.
561 if (cpu == RING_BUFFER_ALL_CPUS) {
562 work = &buffer->irq_work;
563 /* Full only makes sense on per cpu reads */
566 if (!cpumask_test_cpu(cpu, buffer->cpumask))
568 cpu_buffer = buffer->buffers[cpu];
569 work = &cpu_buffer->irq_work;
575 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
577 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
580 * The events can happen in critical sections where
581 * checking a work queue can cause deadlocks.
582 * After adding a task to the queue, this flag is set
583 * only to notify events to try to wake up the queue
586 * We don't clear it even if the buffer is no longer
587 * empty. The flag only causes the next event to run
588 * irq_work to do the work queue wake up. The worse
589 * that can happen if we race with !trace_empty() is that
590 * an event will cause an irq_work to try to wake up
593 * There's no reason to protect this flag either, as
594 * the work queue and irq_work logic will do the necessary
595 * synchronization for the wake ups. The only thing
596 * that is necessary is that the wake up happens after
597 * a task has been queued. It's OK for spurious wake ups.
600 work->full_waiters_pending = true;
602 work->waiters_pending = true;
604 if (signal_pending(current)) {
609 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
612 if (cpu != RING_BUFFER_ALL_CPUS &&
613 !ring_buffer_empty_cpu(buffer, cpu)) {
620 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
621 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
622 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
632 finish_wait(&work->full_waiters, &wait);
634 finish_wait(&work->waiters, &wait);
640 * ring_buffer_poll_wait - poll on buffer input
641 * @buffer: buffer to wait on
642 * @cpu: the cpu buffer to wait on
643 * @filp: the file descriptor
644 * @poll_table: The poll descriptor
646 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
647 * as data is added to any of the @buffer's cpu buffers. Otherwise
648 * it will wait for data to be added to a specific cpu buffer.
650 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
653 int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
654 struct file *filp, poll_table *poll_table)
656 struct ring_buffer_per_cpu *cpu_buffer;
657 struct rb_irq_work *work;
659 if (cpu == RING_BUFFER_ALL_CPUS)
660 work = &buffer->irq_work;
662 if (!cpumask_test_cpu(cpu, buffer->cpumask))
665 cpu_buffer = buffer->buffers[cpu];
666 work = &cpu_buffer->irq_work;
669 poll_wait(filp, &work->waiters, poll_table);
670 work->waiters_pending = true;
672 * There's a tight race between setting the waiters_pending and
673 * checking if the ring buffer is empty. Once the waiters_pending bit
674 * is set, the next event will wake the task up, but we can get stuck
675 * if there's only a single event in.
677 * FIXME: Ideally, we need a memory barrier on the writer side as well,
678 * but adding a memory barrier to all events will cause too much of a
679 * performance hit in the fast path. We only need a memory barrier when
680 * the buffer goes from empty to having content. But as this race is
681 * extremely small, and it's not a problem if another event comes in, we
686 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
687 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
688 return POLLIN | POLLRDNORM;
692 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
693 #define RB_WARN_ON(b, cond) \
695 int _____ret = unlikely(cond); \
697 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
698 struct ring_buffer_per_cpu *__b = \
700 atomic_inc(&__b->buffer->record_disabled); \
702 atomic_inc(&b->record_disabled); \
708 /* Up this if you want to test the TIME_EXTENTS and normalization */
709 #define DEBUG_SHIFT 0
711 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
713 /* shift to debug/test normalization and TIME_EXTENTS */
714 return buffer->clock() << DEBUG_SHIFT;
717 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
721 preempt_disable_notrace();
722 time = rb_time_stamp(buffer);
723 preempt_enable_no_resched_notrace();
727 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
729 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
732 /* Just stupid testing the normalize function and deltas */
735 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
738 * Making the ring buffer lockless makes things tricky.
739 * Although writes only happen on the CPU that they are on,
740 * and they only need to worry about interrupts. Reads can
743 * The reader page is always off the ring buffer, but when the
744 * reader finishes with a page, it needs to swap its page with
745 * a new one from the buffer. The reader needs to take from
746 * the head (writes go to the tail). But if a writer is in overwrite
747 * mode and wraps, it must push the head page forward.
749 * Here lies the problem.
751 * The reader must be careful to replace only the head page, and
752 * not another one. As described at the top of the file in the
753 * ASCII art, the reader sets its old page to point to the next
754 * page after head. It then sets the page after head to point to
755 * the old reader page. But if the writer moves the head page
756 * during this operation, the reader could end up with the tail.
758 * We use cmpxchg to help prevent this race. We also do something
759 * special with the page before head. We set the LSB to 1.
761 * When the writer must push the page forward, it will clear the
762 * bit that points to the head page, move the head, and then set
763 * the bit that points to the new head page.
765 * We also don't want an interrupt coming in and moving the head
766 * page on another writer. Thus we use the second LSB to catch
769 * head->list->prev->next bit 1 bit 0
772 * Points to head page 0 1
775 * Note we can not trust the prev pointer of the head page, because:
777 * +----+ +-----+ +-----+
778 * | |------>| T |---X--->| N |
780 * +----+ +-----+ +-----+
783 * +----------| R |----------+ |
787 * Key: ---X--> HEAD flag set in pointer
792 * (see __rb_reserve_next() to see where this happens)
794 * What the above shows is that the reader just swapped out
795 * the reader page with a page in the buffer, but before it
796 * could make the new header point back to the new page added
797 * it was preempted by a writer. The writer moved forward onto
798 * the new page added by the reader and is about to move forward
801 * You can see, it is legitimate for the previous pointer of
802 * the head (or any page) not to point back to itself. But only
806 #define RB_PAGE_NORMAL 0UL
807 #define RB_PAGE_HEAD 1UL
808 #define RB_PAGE_UPDATE 2UL
811 #define RB_FLAG_MASK 3UL
813 /* PAGE_MOVED is not part of the mask */
814 #define RB_PAGE_MOVED 4UL
817 * rb_list_head - remove any bit
819 static struct list_head *rb_list_head(struct list_head *list)
821 unsigned long val = (unsigned long)list;
823 return (struct list_head *)(val & ~RB_FLAG_MASK);
827 * rb_is_head_page - test if the given page is the head page
829 * Because the reader may move the head_page pointer, we can
830 * not trust what the head page is (it may be pointing to
831 * the reader page). But if the next page is a header page,
832 * its flags will be non zero.
835 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
836 struct buffer_page *page, struct list_head *list)
840 val = (unsigned long)list->next;
842 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
843 return RB_PAGE_MOVED;
845 return val & RB_FLAG_MASK;
851 * The unique thing about the reader page, is that, if the
852 * writer is ever on it, the previous pointer never points
853 * back to the reader page.
855 static int rb_is_reader_page(struct buffer_page *page)
857 struct list_head *list = page->list.prev;
859 return rb_list_head(list->next) != &page->list;
863 * rb_set_list_to_head - set a list_head to be pointing to head.
865 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
866 struct list_head *list)
870 ptr = (unsigned long *)&list->next;
871 *ptr |= RB_PAGE_HEAD;
872 *ptr &= ~RB_PAGE_UPDATE;
876 * rb_head_page_activate - sets up head page
878 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
880 struct buffer_page *head;
882 head = cpu_buffer->head_page;
887 * Set the previous list pointer to have the HEAD flag.
889 rb_set_list_to_head(cpu_buffer, head->list.prev);
892 static void rb_list_head_clear(struct list_head *list)
894 unsigned long *ptr = (unsigned long *)&list->next;
896 *ptr &= ~RB_FLAG_MASK;
900 * rb_head_page_dactivate - clears head page ptr (for free list)
903 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
905 struct list_head *hd;
907 /* Go through the whole list and clear any pointers found. */
908 rb_list_head_clear(cpu_buffer->pages);
910 list_for_each(hd, cpu_buffer->pages)
911 rb_list_head_clear(hd);
914 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
915 struct buffer_page *head,
916 struct buffer_page *prev,
917 int old_flag, int new_flag)
919 struct list_head *list;
920 unsigned long val = (unsigned long)&head->list;
925 val &= ~RB_FLAG_MASK;
927 ret = cmpxchg((unsigned long *)&list->next,
928 val | old_flag, val | new_flag);
930 /* check if the reader took the page */
931 if ((ret & ~RB_FLAG_MASK) != val)
932 return RB_PAGE_MOVED;
934 return ret & RB_FLAG_MASK;
937 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
938 struct buffer_page *head,
939 struct buffer_page *prev,
942 return rb_head_page_set(cpu_buffer, head, prev,
943 old_flag, RB_PAGE_UPDATE);
946 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
947 struct buffer_page *head,
948 struct buffer_page *prev,
951 return rb_head_page_set(cpu_buffer, head, prev,
952 old_flag, RB_PAGE_HEAD);
955 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
956 struct buffer_page *head,
957 struct buffer_page *prev,
960 return rb_head_page_set(cpu_buffer, head, prev,
961 old_flag, RB_PAGE_NORMAL);
964 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
965 struct buffer_page **bpage)
967 struct list_head *p = rb_list_head((*bpage)->list.next);
969 *bpage = list_entry(p, struct buffer_page, list);
972 static struct buffer_page *
973 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
975 struct buffer_page *head;
976 struct buffer_page *page;
977 struct list_head *list;
980 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
984 list = cpu_buffer->pages;
985 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
988 page = head = cpu_buffer->head_page;
990 * It is possible that the writer moves the header behind
991 * where we started, and we miss in one loop.
992 * A second loop should grab the header, but we'll do
993 * three loops just because I'm paranoid.
995 for (i = 0; i < 3; i++) {
997 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
998 cpu_buffer->head_page = page;
1001 rb_inc_page(cpu_buffer, &page);
1002 } while (page != head);
1005 RB_WARN_ON(cpu_buffer, 1);
1010 static int rb_head_page_replace(struct buffer_page *old,
1011 struct buffer_page *new)
1013 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1017 val = *ptr & ~RB_FLAG_MASK;
1018 val |= RB_PAGE_HEAD;
1020 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1026 * rb_tail_page_update - move the tail page forward
1028 * Returns 1 if moved tail page, 0 if someone else did.
1030 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1031 struct buffer_page *tail_page,
1032 struct buffer_page *next_page)
1034 struct buffer_page *old_tail;
1035 unsigned long old_entries;
1036 unsigned long old_write;
1040 * The tail page now needs to be moved forward.
1042 * We need to reset the tail page, but without messing
1043 * with possible erasing of data brought in by interrupts
1044 * that have moved the tail page and are currently on it.
1046 * We add a counter to the write field to denote this.
1048 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1049 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1052 * Just make sure we have seen our old_write and synchronize
1053 * with any interrupts that come in.
1058 * If the tail page is still the same as what we think
1059 * it is, then it is up to us to update the tail
1062 if (tail_page == cpu_buffer->tail_page) {
1063 /* Zero the write counter */
1064 unsigned long val = old_write & ~RB_WRITE_MASK;
1065 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1068 * This will only succeed if an interrupt did
1069 * not come in and change it. In which case, we
1070 * do not want to modify it.
1072 * We add (void) to let the compiler know that we do not care
1073 * about the return value of these functions. We use the
1074 * cmpxchg to only update if an interrupt did not already
1075 * do it for us. If the cmpxchg fails, we don't care.
1077 (void)local_cmpxchg(&next_page->write, old_write, val);
1078 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1081 * No need to worry about races with clearing out the commit.
1082 * it only can increment when a commit takes place. But that
1083 * only happens in the outer most nested commit.
1085 local_set(&next_page->page->commit, 0);
1087 old_tail = cmpxchg(&cpu_buffer->tail_page,
1088 tail_page, next_page);
1090 if (old_tail == tail_page)
1097 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1098 struct buffer_page *bpage)
1100 unsigned long val = (unsigned long)bpage;
1102 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1109 * rb_check_list - make sure a pointer to a list has the last bits zero
1111 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1112 struct list_head *list)
1114 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1116 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1122 * rb_check_pages - integrity check of buffer pages
1123 * @cpu_buffer: CPU buffer with pages to test
1125 * As a safety measure we check to make sure the data pages have not
1128 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1130 struct list_head *head = cpu_buffer->pages;
1131 struct buffer_page *bpage, *tmp;
1133 /* Reset the head page if it exists */
1134 if (cpu_buffer->head_page)
1135 rb_set_head_page(cpu_buffer);
1137 rb_head_page_deactivate(cpu_buffer);
1139 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1141 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1144 if (rb_check_list(cpu_buffer, head))
1147 list_for_each_entry_safe(bpage, tmp, head, list) {
1148 if (RB_WARN_ON(cpu_buffer,
1149 bpage->list.next->prev != &bpage->list))
1151 if (RB_WARN_ON(cpu_buffer,
1152 bpage->list.prev->next != &bpage->list))
1154 if (rb_check_list(cpu_buffer, &bpage->list))
1158 rb_head_page_activate(cpu_buffer);
1163 static int __rb_allocate_pages(int nr_pages, struct list_head *pages, int cpu)
1166 struct buffer_page *bpage, *tmp;
1168 for (i = 0; i < nr_pages; i++) {
1171 * __GFP_NORETRY flag makes sure that the allocation fails
1172 * gracefully without invoking oom-killer and the system is
1175 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1176 GFP_KERNEL | __GFP_NORETRY,
1181 list_add(&bpage->list, pages);
1183 page = alloc_pages_node(cpu_to_node(cpu),
1184 GFP_KERNEL | __GFP_NORETRY, 0);
1187 bpage->page = page_address(page);
1188 rb_init_page(bpage->page);
1194 list_for_each_entry_safe(bpage, tmp, pages, list) {
1195 list_del_init(&bpage->list);
1196 free_buffer_page(bpage);
1202 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1209 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1213 * The ring buffer page list is a circular list that does not
1214 * start and end with a list head. All page list items point to
1217 cpu_buffer->pages = pages.next;
1220 cpu_buffer->nr_pages = nr_pages;
1222 rb_check_pages(cpu_buffer);
1227 static struct ring_buffer_per_cpu *
1228 rb_allocate_cpu_buffer(struct ring_buffer *buffer, int nr_pages, int cpu)
1230 struct ring_buffer_per_cpu *cpu_buffer;
1231 struct buffer_page *bpage;
1235 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1236 GFP_KERNEL, cpu_to_node(cpu));
1240 cpu_buffer->cpu = cpu;
1241 cpu_buffer->buffer = buffer;
1242 raw_spin_lock_init(&cpu_buffer->reader_lock);
1243 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1244 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1245 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1246 init_completion(&cpu_buffer->update_done);
1247 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1248 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1249 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1251 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1252 GFP_KERNEL, cpu_to_node(cpu));
1254 goto fail_free_buffer;
1256 rb_check_bpage(cpu_buffer, bpage);
1258 cpu_buffer->reader_page = bpage;
1259 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1261 goto fail_free_reader;
1262 bpage->page = page_address(page);
1263 rb_init_page(bpage->page);
1265 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1266 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1268 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1270 goto fail_free_reader;
1272 cpu_buffer->head_page
1273 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1274 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1276 rb_head_page_activate(cpu_buffer);
1281 free_buffer_page(cpu_buffer->reader_page);
1288 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1290 struct list_head *head = cpu_buffer->pages;
1291 struct buffer_page *bpage, *tmp;
1293 free_buffer_page(cpu_buffer->reader_page);
1295 rb_head_page_deactivate(cpu_buffer);
1298 list_for_each_entry_safe(bpage, tmp, head, list) {
1299 list_del_init(&bpage->list);
1300 free_buffer_page(bpage);
1302 bpage = list_entry(head, struct buffer_page, list);
1303 free_buffer_page(bpage);
1309 #ifdef CONFIG_HOTPLUG_CPU
1310 static int rb_cpu_notify(struct notifier_block *self,
1311 unsigned long action, void *hcpu);
1315 * __ring_buffer_alloc - allocate a new ring_buffer
1316 * @size: the size in bytes per cpu that is needed.
1317 * @flags: attributes to set for the ring buffer.
1319 * Currently the only flag that is available is the RB_FL_OVERWRITE
1320 * flag. This flag means that the buffer will overwrite old data
1321 * when the buffer wraps. If this flag is not set, the buffer will
1322 * drop data when the tail hits the head.
1324 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1325 struct lock_class_key *key)
1327 struct ring_buffer *buffer;
1331 /* keep it in its own cache line */
1332 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1337 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1338 goto fail_free_buffer;
1340 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1341 buffer->flags = flags;
1342 buffer->clock = trace_clock_local;
1343 buffer->reader_lock_key = key;
1345 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1346 init_waitqueue_head(&buffer->irq_work.waiters);
1348 /* need at least two pages */
1353 * In case of non-hotplug cpu, if the ring-buffer is allocated
1354 * in early initcall, it will not be notified of secondary cpus.
1355 * In that off case, we need to allocate for all possible cpus.
1357 #ifdef CONFIG_HOTPLUG_CPU
1358 cpu_notifier_register_begin();
1359 cpumask_copy(buffer->cpumask, cpu_online_mask);
1361 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1363 buffer->cpus = nr_cpu_ids;
1365 bsize = sizeof(void *) * nr_cpu_ids;
1366 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1368 if (!buffer->buffers)
1369 goto fail_free_cpumask;
1371 for_each_buffer_cpu(buffer, cpu) {
1372 buffer->buffers[cpu] =
1373 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1374 if (!buffer->buffers[cpu])
1375 goto fail_free_buffers;
1378 #ifdef CONFIG_HOTPLUG_CPU
1379 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1380 buffer->cpu_notify.priority = 0;
1381 __register_cpu_notifier(&buffer->cpu_notify);
1382 cpu_notifier_register_done();
1385 mutex_init(&buffer->mutex);
1390 for_each_buffer_cpu(buffer, cpu) {
1391 if (buffer->buffers[cpu])
1392 rb_free_cpu_buffer(buffer->buffers[cpu]);
1394 kfree(buffer->buffers);
1397 free_cpumask_var(buffer->cpumask);
1398 #ifdef CONFIG_HOTPLUG_CPU
1399 cpu_notifier_register_done();
1406 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1409 * ring_buffer_free - free a ring buffer.
1410 * @buffer: the buffer to free.
1413 ring_buffer_free(struct ring_buffer *buffer)
1417 #ifdef CONFIG_HOTPLUG_CPU
1418 cpu_notifier_register_begin();
1419 __unregister_cpu_notifier(&buffer->cpu_notify);
1422 for_each_buffer_cpu(buffer, cpu)
1423 rb_free_cpu_buffer(buffer->buffers[cpu]);
1425 #ifdef CONFIG_HOTPLUG_CPU
1426 cpu_notifier_register_done();
1429 kfree(buffer->buffers);
1430 free_cpumask_var(buffer->cpumask);
1434 EXPORT_SYMBOL_GPL(ring_buffer_free);
1436 void ring_buffer_set_clock(struct ring_buffer *buffer,
1439 buffer->clock = clock;
1442 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1444 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1446 return local_read(&bpage->entries) & RB_WRITE_MASK;
1449 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1451 return local_read(&bpage->write) & RB_WRITE_MASK;
1455 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned int nr_pages)
1457 struct list_head *tail_page, *to_remove, *next_page;
1458 struct buffer_page *to_remove_page, *tmp_iter_page;
1459 struct buffer_page *last_page, *first_page;
1460 unsigned int nr_removed;
1461 unsigned long head_bit;
1466 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1467 atomic_inc(&cpu_buffer->record_disabled);
1469 * We don't race with the readers since we have acquired the reader
1470 * lock. We also don't race with writers after disabling recording.
1471 * This makes it easy to figure out the first and the last page to be
1472 * removed from the list. We unlink all the pages in between including
1473 * the first and last pages. This is done in a busy loop so that we
1474 * lose the least number of traces.
1475 * The pages are freed after we restart recording and unlock readers.
1477 tail_page = &cpu_buffer->tail_page->list;
1480 * tail page might be on reader page, we remove the next page
1481 * from the ring buffer
1483 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1484 tail_page = rb_list_head(tail_page->next);
1485 to_remove = tail_page;
1487 /* start of pages to remove */
1488 first_page = list_entry(rb_list_head(to_remove->next),
1489 struct buffer_page, list);
1491 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1492 to_remove = rb_list_head(to_remove)->next;
1493 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1496 next_page = rb_list_head(to_remove)->next;
1499 * Now we remove all pages between tail_page and next_page.
1500 * Make sure that we have head_bit value preserved for the
1503 tail_page->next = (struct list_head *)((unsigned long)next_page |
1505 next_page = rb_list_head(next_page);
1506 next_page->prev = tail_page;
1508 /* make sure pages points to a valid page in the ring buffer */
1509 cpu_buffer->pages = next_page;
1511 /* update head page */
1513 cpu_buffer->head_page = list_entry(next_page,
1514 struct buffer_page, list);
1517 * change read pointer to make sure any read iterators reset
1520 cpu_buffer->read = 0;
1522 /* pages are removed, resume tracing and then free the pages */
1523 atomic_dec(&cpu_buffer->record_disabled);
1524 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1526 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1528 /* last buffer page to remove */
1529 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1531 tmp_iter_page = first_page;
1534 to_remove_page = tmp_iter_page;
1535 rb_inc_page(cpu_buffer, &tmp_iter_page);
1537 /* update the counters */
1538 page_entries = rb_page_entries(to_remove_page);
1541 * If something was added to this page, it was full
1542 * since it is not the tail page. So we deduct the
1543 * bytes consumed in ring buffer from here.
1544 * Increment overrun to account for the lost events.
1546 local_add(page_entries, &cpu_buffer->overrun);
1547 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1551 * We have already removed references to this list item, just
1552 * free up the buffer_page and its page
1554 free_buffer_page(to_remove_page);
1557 } while (to_remove_page != last_page);
1559 RB_WARN_ON(cpu_buffer, nr_removed);
1561 return nr_removed == 0;
1565 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1567 struct list_head *pages = &cpu_buffer->new_pages;
1568 int retries, success;
1570 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1572 * We are holding the reader lock, so the reader page won't be swapped
1573 * in the ring buffer. Now we are racing with the writer trying to
1574 * move head page and the tail page.
1575 * We are going to adapt the reader page update process where:
1576 * 1. We first splice the start and end of list of new pages between
1577 * the head page and its previous page.
1578 * 2. We cmpxchg the prev_page->next to point from head page to the
1579 * start of new pages list.
1580 * 3. Finally, we update the head->prev to the end of new list.
1582 * We will try this process 10 times, to make sure that we don't keep
1588 struct list_head *head_page, *prev_page, *r;
1589 struct list_head *last_page, *first_page;
1590 struct list_head *head_page_with_bit;
1592 head_page = &rb_set_head_page(cpu_buffer)->list;
1595 prev_page = head_page->prev;
1597 first_page = pages->next;
1598 last_page = pages->prev;
1600 head_page_with_bit = (struct list_head *)
1601 ((unsigned long)head_page | RB_PAGE_HEAD);
1603 last_page->next = head_page_with_bit;
1604 first_page->prev = prev_page;
1606 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1608 if (r == head_page_with_bit) {
1610 * yay, we replaced the page pointer to our new list,
1611 * now, we just have to update to head page's prev
1612 * pointer to point to end of list
1614 head_page->prev = last_page;
1621 INIT_LIST_HEAD(pages);
1623 * If we weren't successful in adding in new pages, warn and stop
1626 RB_WARN_ON(cpu_buffer, !success);
1627 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1629 /* free pages if they weren't inserted */
1631 struct buffer_page *bpage, *tmp;
1632 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1634 list_del_init(&bpage->list);
1635 free_buffer_page(bpage);
1641 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1645 if (cpu_buffer->nr_pages_to_update > 0)
1646 success = rb_insert_pages(cpu_buffer);
1648 success = rb_remove_pages(cpu_buffer,
1649 -cpu_buffer->nr_pages_to_update);
1652 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1655 static void update_pages_handler(struct work_struct *work)
1657 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1658 struct ring_buffer_per_cpu, update_pages_work);
1659 rb_update_pages(cpu_buffer);
1660 complete(&cpu_buffer->update_done);
1664 * ring_buffer_resize - resize the ring buffer
1665 * @buffer: the buffer to resize.
1666 * @size: the new size.
1667 * @cpu_id: the cpu buffer to resize
1669 * Minimum size is 2 * BUF_PAGE_SIZE.
1671 * Returns 0 on success and < 0 on failure.
1673 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1676 struct ring_buffer_per_cpu *cpu_buffer;
1681 * Always succeed at resizing a non-existent buffer:
1686 /* Make sure the requested buffer exists */
1687 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1688 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1691 size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1692 size *= BUF_PAGE_SIZE;
1694 /* we need a minimum of two pages */
1695 if (size < BUF_PAGE_SIZE * 2)
1696 size = BUF_PAGE_SIZE * 2;
1698 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1701 * Don't succeed if resizing is disabled, as a reader might be
1702 * manipulating the ring buffer and is expecting a sane state while
1705 if (atomic_read(&buffer->resize_disabled))
1708 /* prevent another thread from changing buffer sizes */
1709 mutex_lock(&buffer->mutex);
1711 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1712 /* calculate the pages to update */
1713 for_each_buffer_cpu(buffer, cpu) {
1714 cpu_buffer = buffer->buffers[cpu];
1716 cpu_buffer->nr_pages_to_update = nr_pages -
1717 cpu_buffer->nr_pages;
1719 * nothing more to do for removing pages or no update
1721 if (cpu_buffer->nr_pages_to_update <= 0)
1724 * to add pages, make sure all new pages can be
1725 * allocated without receiving ENOMEM
1727 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1728 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1729 &cpu_buffer->new_pages, cpu)) {
1730 /* not enough memory for new pages */
1738 * Fire off all the required work handlers
1739 * We can't schedule on offline CPUs, but it's not necessary
1740 * since we can change their buffer sizes without any race.
1742 for_each_buffer_cpu(buffer, cpu) {
1743 cpu_buffer = buffer->buffers[cpu];
1744 if (!cpu_buffer->nr_pages_to_update)
1747 /* Can't run something on an offline CPU. */
1748 if (!cpu_online(cpu)) {
1749 rb_update_pages(cpu_buffer);
1750 cpu_buffer->nr_pages_to_update = 0;
1752 schedule_work_on(cpu,
1753 &cpu_buffer->update_pages_work);
1757 /* wait for all the updates to complete */
1758 for_each_buffer_cpu(buffer, cpu) {
1759 cpu_buffer = buffer->buffers[cpu];
1760 if (!cpu_buffer->nr_pages_to_update)
1763 if (cpu_online(cpu))
1764 wait_for_completion(&cpu_buffer->update_done);
1765 cpu_buffer->nr_pages_to_update = 0;
1770 /* Make sure this CPU has been intitialized */
1771 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1774 cpu_buffer = buffer->buffers[cpu_id];
1776 if (nr_pages == cpu_buffer->nr_pages)
1779 cpu_buffer->nr_pages_to_update = nr_pages -
1780 cpu_buffer->nr_pages;
1782 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1783 if (cpu_buffer->nr_pages_to_update > 0 &&
1784 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1785 &cpu_buffer->new_pages, cpu_id)) {
1792 /* Can't run something on an offline CPU. */
1793 if (!cpu_online(cpu_id))
1794 rb_update_pages(cpu_buffer);
1796 schedule_work_on(cpu_id,
1797 &cpu_buffer->update_pages_work);
1798 wait_for_completion(&cpu_buffer->update_done);
1801 cpu_buffer->nr_pages_to_update = 0;
1807 * The ring buffer resize can happen with the ring buffer
1808 * enabled, so that the update disturbs the tracing as little
1809 * as possible. But if the buffer is disabled, we do not need
1810 * to worry about that, and we can take the time to verify
1811 * that the buffer is not corrupt.
1813 if (atomic_read(&buffer->record_disabled)) {
1814 atomic_inc(&buffer->record_disabled);
1816 * Even though the buffer was disabled, we must make sure
1817 * that it is truly disabled before calling rb_check_pages.
1818 * There could have been a race between checking
1819 * record_disable and incrementing it.
1821 synchronize_sched();
1822 for_each_buffer_cpu(buffer, cpu) {
1823 cpu_buffer = buffer->buffers[cpu];
1824 rb_check_pages(cpu_buffer);
1826 atomic_dec(&buffer->record_disabled);
1829 mutex_unlock(&buffer->mutex);
1833 for_each_buffer_cpu(buffer, cpu) {
1834 struct buffer_page *bpage, *tmp;
1836 cpu_buffer = buffer->buffers[cpu];
1837 cpu_buffer->nr_pages_to_update = 0;
1839 if (list_empty(&cpu_buffer->new_pages))
1842 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1844 list_del_init(&bpage->list);
1845 free_buffer_page(bpage);
1848 mutex_unlock(&buffer->mutex);
1851 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1853 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1855 mutex_lock(&buffer->mutex);
1857 buffer->flags |= RB_FL_OVERWRITE;
1859 buffer->flags &= ~RB_FL_OVERWRITE;
1860 mutex_unlock(&buffer->mutex);
1862 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1864 static inline void *
1865 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1867 return bpage->data + index;
1870 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1872 return bpage->page->data + index;
1875 static inline struct ring_buffer_event *
1876 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1878 return __rb_page_index(cpu_buffer->reader_page,
1879 cpu_buffer->reader_page->read);
1882 static inline struct ring_buffer_event *
1883 rb_iter_head_event(struct ring_buffer_iter *iter)
1885 return __rb_page_index(iter->head_page, iter->head);
1888 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1890 return local_read(&bpage->page->commit);
1893 /* Size is determined by what has been committed */
1894 static inline unsigned rb_page_size(struct buffer_page *bpage)
1896 return rb_page_commit(bpage);
1899 static inline unsigned
1900 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1902 return rb_page_commit(cpu_buffer->commit_page);
1905 static inline unsigned
1906 rb_event_index(struct ring_buffer_event *event)
1908 unsigned long addr = (unsigned long)event;
1910 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1914 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1915 struct ring_buffer_event *event)
1917 unsigned long addr = (unsigned long)event;
1918 unsigned long index;
1920 index = rb_event_index(event);
1923 return cpu_buffer->commit_page->page == (void *)addr &&
1924 rb_commit_index(cpu_buffer) == index;
1928 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1930 unsigned long max_count;
1933 * We only race with interrupts and NMIs on this CPU.
1934 * If we own the commit event, then we can commit
1935 * all others that interrupted us, since the interruptions
1936 * are in stack format (they finish before they come
1937 * back to us). This allows us to do a simple loop to
1938 * assign the commit to the tail.
1941 max_count = cpu_buffer->nr_pages * 100;
1943 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1944 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1946 if (RB_WARN_ON(cpu_buffer,
1947 rb_is_reader_page(cpu_buffer->tail_page)))
1949 local_set(&cpu_buffer->commit_page->page->commit,
1950 rb_page_write(cpu_buffer->commit_page));
1951 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1952 cpu_buffer->write_stamp =
1953 cpu_buffer->commit_page->page->time_stamp;
1954 /* add barrier to keep gcc from optimizing too much */
1957 while (rb_commit_index(cpu_buffer) !=
1958 rb_page_write(cpu_buffer->commit_page)) {
1960 local_set(&cpu_buffer->commit_page->page->commit,
1961 rb_page_write(cpu_buffer->commit_page));
1962 RB_WARN_ON(cpu_buffer,
1963 local_read(&cpu_buffer->commit_page->page->commit) &
1968 /* again, keep gcc from optimizing */
1972 * If an interrupt came in just after the first while loop
1973 * and pushed the tail page forward, we will be left with
1974 * a dangling commit that will never go forward.
1976 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1980 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1982 cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1983 cpu_buffer->reader_page->read = 0;
1986 static void rb_inc_iter(struct ring_buffer_iter *iter)
1988 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1991 * The iterator could be on the reader page (it starts there).
1992 * But the head could have moved, since the reader was
1993 * found. Check for this case and assign the iterator
1994 * to the head page instead of next.
1996 if (iter->head_page == cpu_buffer->reader_page)
1997 iter->head_page = rb_set_head_page(cpu_buffer);
1999 rb_inc_page(cpu_buffer, &iter->head_page);
2001 iter->read_stamp = iter->head_page->page->time_stamp;
2005 /* Slow path, do not inline */
2006 static noinline struct ring_buffer_event *
2007 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
2009 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2011 /* Not the first event on the page? */
2012 if (rb_event_index(event)) {
2013 event->time_delta = delta & TS_MASK;
2014 event->array[0] = delta >> TS_SHIFT;
2016 /* nope, just zero it */
2017 event->time_delta = 0;
2018 event->array[0] = 0;
2021 return skip_time_extend(event);
2025 * rb_update_event - update event type and data
2026 * @event: the event to update
2027 * @type: the type of event
2028 * @length: the size of the event field in the ring buffer
2030 * Update the type and data fields of the event. The length
2031 * is the actual size that is written to the ring buffer,
2032 * and with this, we can determine what to place into the
2036 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2037 struct ring_buffer_event *event, unsigned length,
2038 int add_timestamp, u64 delta)
2040 /* Only a commit updates the timestamp */
2041 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2045 * If we need to add a timestamp, then we
2046 * add it to the start of the resevered space.
2048 if (unlikely(add_timestamp)) {
2049 event = rb_add_time_stamp(event, delta);
2050 length -= RB_LEN_TIME_EXTEND;
2054 event->time_delta = delta;
2055 length -= RB_EVNT_HDR_SIZE;
2056 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2057 event->type_len = 0;
2058 event->array[0] = length;
2060 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2064 * rb_handle_head_page - writer hit the head page
2066 * Returns: +1 to retry page
2071 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2072 struct buffer_page *tail_page,
2073 struct buffer_page *next_page)
2075 struct buffer_page *new_head;
2080 entries = rb_page_entries(next_page);
2083 * The hard part is here. We need to move the head
2084 * forward, and protect against both readers on
2085 * other CPUs and writers coming in via interrupts.
2087 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2091 * type can be one of four:
2092 * NORMAL - an interrupt already moved it for us
2093 * HEAD - we are the first to get here.
2094 * UPDATE - we are the interrupt interrupting
2096 * MOVED - a reader on another CPU moved the next
2097 * pointer to its reader page. Give up
2104 * We changed the head to UPDATE, thus
2105 * it is our responsibility to update
2108 local_add(entries, &cpu_buffer->overrun);
2109 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2112 * The entries will be zeroed out when we move the
2116 /* still more to do */
2119 case RB_PAGE_UPDATE:
2121 * This is an interrupt that interrupt the
2122 * previous update. Still more to do.
2125 case RB_PAGE_NORMAL:
2127 * An interrupt came in before the update
2128 * and processed this for us.
2129 * Nothing left to do.
2134 * The reader is on another CPU and just did
2135 * a swap with our next_page.
2140 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2145 * Now that we are here, the old head pointer is
2146 * set to UPDATE. This will keep the reader from
2147 * swapping the head page with the reader page.
2148 * The reader (on another CPU) will spin till
2151 * We just need to protect against interrupts
2152 * doing the job. We will set the next pointer
2153 * to HEAD. After that, we set the old pointer
2154 * to NORMAL, but only if it was HEAD before.
2155 * otherwise we are an interrupt, and only
2156 * want the outer most commit to reset it.
2158 new_head = next_page;
2159 rb_inc_page(cpu_buffer, &new_head);
2161 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2165 * Valid returns are:
2166 * HEAD - an interrupt came in and already set it.
2167 * NORMAL - One of two things:
2168 * 1) We really set it.
2169 * 2) A bunch of interrupts came in and moved
2170 * the page forward again.
2174 case RB_PAGE_NORMAL:
2178 RB_WARN_ON(cpu_buffer, 1);
2183 * It is possible that an interrupt came in,
2184 * set the head up, then more interrupts came in
2185 * and moved it again. When we get back here,
2186 * the page would have been set to NORMAL but we
2187 * just set it back to HEAD.
2189 * How do you detect this? Well, if that happened
2190 * the tail page would have moved.
2192 if (ret == RB_PAGE_NORMAL) {
2194 * If the tail had moved passed next, then we need
2195 * to reset the pointer.
2197 if (cpu_buffer->tail_page != tail_page &&
2198 cpu_buffer->tail_page != next_page)
2199 rb_head_page_set_normal(cpu_buffer, new_head,
2205 * If this was the outer most commit (the one that
2206 * changed the original pointer from HEAD to UPDATE),
2207 * then it is up to us to reset it to NORMAL.
2209 if (type == RB_PAGE_HEAD) {
2210 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2213 if (RB_WARN_ON(cpu_buffer,
2214 ret != RB_PAGE_UPDATE))
2221 static unsigned rb_calculate_event_length(unsigned length)
2223 struct ring_buffer_event event; /* Used only for sizeof array */
2225 /* zero length can cause confusions */
2229 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2230 length += sizeof(event.array[0]);
2232 length += RB_EVNT_HDR_SIZE;
2233 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2239 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2240 struct buffer_page *tail_page,
2241 unsigned long tail, unsigned long length)
2243 struct ring_buffer_event *event;
2246 * Only the event that crossed the page boundary
2247 * must fill the old tail_page with padding.
2249 if (tail >= BUF_PAGE_SIZE) {
2251 * If the page was filled, then we still need
2252 * to update the real_end. Reset it to zero
2253 * and the reader will ignore it.
2255 if (tail == BUF_PAGE_SIZE)
2256 tail_page->real_end = 0;
2258 local_sub(length, &tail_page->write);
2262 event = __rb_page_index(tail_page, tail);
2263 kmemcheck_annotate_bitfield(event, bitfield);
2265 /* account for padding bytes */
2266 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2269 * Save the original length to the meta data.
2270 * This will be used by the reader to add lost event
2273 tail_page->real_end = tail;
2276 * If this event is bigger than the minimum size, then
2277 * we need to be careful that we don't subtract the
2278 * write counter enough to allow another writer to slip
2280 * We put in a discarded commit instead, to make sure
2281 * that this space is not used again.
2283 * If we are less than the minimum size, we don't need to
2286 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2287 /* No room for any events */
2289 /* Mark the rest of the page with padding */
2290 rb_event_set_padding(event);
2292 /* Set the write back to the previous setting */
2293 local_sub(length, &tail_page->write);
2297 /* Put in a discarded event */
2298 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2299 event->type_len = RINGBUF_TYPE_PADDING;
2300 /* time delta must be non zero */
2301 event->time_delta = 1;
2303 /* Set write to end of buffer */
2304 length = (tail + length) - BUF_PAGE_SIZE;
2305 local_sub(length, &tail_page->write);
2309 * This is the slow path, force gcc not to inline it.
2311 static noinline struct ring_buffer_event *
2312 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2313 unsigned long length, unsigned long tail,
2314 struct buffer_page *tail_page, u64 ts)
2316 struct buffer_page *commit_page = cpu_buffer->commit_page;
2317 struct ring_buffer *buffer = cpu_buffer->buffer;
2318 struct buffer_page *next_page;
2321 next_page = tail_page;
2323 rb_inc_page(cpu_buffer, &next_page);
2326 * If for some reason, we had an interrupt storm that made
2327 * it all the way around the buffer, bail, and warn
2330 if (unlikely(next_page == commit_page)) {
2331 local_inc(&cpu_buffer->commit_overrun);
2336 * This is where the fun begins!
2338 * We are fighting against races between a reader that
2339 * could be on another CPU trying to swap its reader
2340 * page with the buffer head.
2342 * We are also fighting against interrupts coming in and
2343 * moving the head or tail on us as well.
2345 * If the next page is the head page then we have filled
2346 * the buffer, unless the commit page is still on the
2349 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2352 * If the commit is not on the reader page, then
2353 * move the header page.
2355 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2357 * If we are not in overwrite mode,
2358 * this is easy, just stop here.
2360 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2361 local_inc(&cpu_buffer->dropped_events);
2365 ret = rb_handle_head_page(cpu_buffer,
2374 * We need to be careful here too. The
2375 * commit page could still be on the reader
2376 * page. We could have a small buffer, and
2377 * have filled up the buffer with events
2378 * from interrupts and such, and wrapped.
2380 * Note, if the tail page is also the on the
2381 * reader_page, we let it move out.
2383 if (unlikely((cpu_buffer->commit_page !=
2384 cpu_buffer->tail_page) &&
2385 (cpu_buffer->commit_page ==
2386 cpu_buffer->reader_page))) {
2387 local_inc(&cpu_buffer->commit_overrun);
2393 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
2396 * Nested commits always have zero deltas, so
2397 * just reread the time stamp
2399 ts = rb_time_stamp(buffer);
2400 next_page->page->time_stamp = ts;
2405 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2407 /* fail and let the caller try again */
2408 return ERR_PTR(-EAGAIN);
2412 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2417 static struct ring_buffer_event *
2418 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2419 unsigned long length, u64 ts,
2420 u64 delta, int add_timestamp)
2422 struct buffer_page *tail_page;
2423 struct ring_buffer_event *event;
2424 unsigned long tail, write;
2427 * If the time delta since the last event is too big to
2428 * hold in the time field of the event, then we append a
2429 * TIME EXTEND event ahead of the data event.
2431 if (unlikely(add_timestamp))
2432 length += RB_LEN_TIME_EXTEND;
2434 tail_page = cpu_buffer->tail_page;
2435 write = local_add_return(length, &tail_page->write);
2437 /* set write to only the index of the write */
2438 write &= RB_WRITE_MASK;
2439 tail = write - length;
2442 * If this is the first commit on the page, then it has the same
2443 * timestamp as the page itself.
2448 /* See if we shot pass the end of this buffer page */
2449 if (unlikely(write > BUF_PAGE_SIZE))
2450 return rb_move_tail(cpu_buffer, length, tail,
2453 /* We reserved something on the buffer */
2455 event = __rb_page_index(tail_page, tail);
2456 kmemcheck_annotate_bitfield(event, bitfield);
2457 rb_update_event(cpu_buffer, event, length, add_timestamp, delta);
2459 local_inc(&tail_page->entries);
2462 * If this is the first commit on the page, then update
2466 tail_page->page->time_stamp = ts;
2468 /* account for these added bytes */
2469 local_add(length, &cpu_buffer->entries_bytes);
2475 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2476 struct ring_buffer_event *event)
2478 unsigned long new_index, old_index;
2479 struct buffer_page *bpage;
2480 unsigned long index;
2483 new_index = rb_event_index(event);
2484 old_index = new_index + rb_event_ts_length(event);
2485 addr = (unsigned long)event;
2488 bpage = cpu_buffer->tail_page;
2490 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2491 unsigned long write_mask =
2492 local_read(&bpage->write) & ~RB_WRITE_MASK;
2493 unsigned long event_length = rb_event_length(event);
2495 * This is on the tail page. It is possible that
2496 * a write could come in and move the tail page
2497 * and write to the next page. That is fine
2498 * because we just shorten what is on this page.
2500 old_index += write_mask;
2501 new_index += write_mask;
2502 index = local_cmpxchg(&bpage->write, old_index, new_index);
2503 if (index == old_index) {
2504 /* update counters */
2505 local_sub(event_length, &cpu_buffer->entries_bytes);
2510 /* could not discard */
2514 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2516 local_inc(&cpu_buffer->committing);
2517 local_inc(&cpu_buffer->commits);
2520 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2522 unsigned long commits;
2524 if (RB_WARN_ON(cpu_buffer,
2525 !local_read(&cpu_buffer->committing)))
2529 commits = local_read(&cpu_buffer->commits);
2530 /* synchronize with interrupts */
2532 if (local_read(&cpu_buffer->committing) == 1)
2533 rb_set_commit_to_write(cpu_buffer);
2535 local_dec(&cpu_buffer->committing);
2537 /* synchronize with interrupts */
2541 * Need to account for interrupts coming in between the
2542 * updating of the commit page and the clearing of the
2543 * committing counter.
2545 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2546 !local_read(&cpu_buffer->committing)) {
2547 local_inc(&cpu_buffer->committing);
2552 static struct ring_buffer_event *
2553 rb_reserve_next_event(struct ring_buffer *buffer,
2554 struct ring_buffer_per_cpu *cpu_buffer,
2555 unsigned long length)
2557 struct ring_buffer_event *event;
2563 rb_start_commit(cpu_buffer);
2565 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2567 * Due to the ability to swap a cpu buffer from a buffer
2568 * it is possible it was swapped before we committed.
2569 * (committing stops a swap). We check for it here and
2570 * if it happened, we have to fail the write.
2573 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2574 local_dec(&cpu_buffer->committing);
2575 local_dec(&cpu_buffer->commits);
2580 length = rb_calculate_event_length(length);
2586 * We allow for interrupts to reenter here and do a trace.
2587 * If one does, it will cause this original code to loop
2588 * back here. Even with heavy interrupts happening, this
2589 * should only happen a few times in a row. If this happens
2590 * 1000 times in a row, there must be either an interrupt
2591 * storm or we have something buggy.
2594 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2597 ts = rb_time_stamp(cpu_buffer->buffer);
2598 diff = ts - cpu_buffer->write_stamp;
2600 /* make sure this diff is calculated here */
2603 /* Did the write stamp get updated already? */
2604 if (likely(ts >= cpu_buffer->write_stamp)) {
2606 if (unlikely(test_time_stamp(delta))) {
2607 int local_clock_stable = 1;
2608 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2609 local_clock_stable = sched_clock_stable();
2611 WARN_ONCE(delta > (1ULL << 59),
2612 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2613 (unsigned long long)delta,
2614 (unsigned long long)ts,
2615 (unsigned long long)cpu_buffer->write_stamp,
2616 local_clock_stable ? "" :
2617 "If you just came from a suspend/resume,\n"
2618 "please switch to the trace global clock:\n"
2619 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2624 event = __rb_reserve_next(cpu_buffer, length, ts,
2625 delta, add_timestamp);
2626 if (unlikely(PTR_ERR(event) == -EAGAIN))
2635 rb_end_commit(cpu_buffer);
2639 #ifdef CONFIG_TRACING
2642 * The lock and unlock are done within a preempt disable section.
2643 * The current_context per_cpu variable can only be modified
2644 * by the current task between lock and unlock. But it can
2645 * be modified more than once via an interrupt. To pass this
2646 * information from the lock to the unlock without having to
2647 * access the 'in_interrupt()' functions again (which do show
2648 * a bit of overhead in something as critical as function tracing,
2649 * we use a bitmask trick.
2651 * bit 0 = NMI context
2652 * bit 1 = IRQ context
2653 * bit 2 = SoftIRQ context
2654 * bit 3 = normal context.
2656 * This works because this is the order of contexts that can
2657 * preempt other contexts. A SoftIRQ never preempts an IRQ
2660 * When the context is determined, the corresponding bit is
2661 * checked and set (if it was set, then a recursion of that context
2664 * On unlock, we need to clear this bit. To do so, just subtract
2665 * 1 from the current_context and AND it to itself.
2669 * 101 & 100 = 100 (clearing bit zero)
2672 * 1010 & 1001 = 1000 (clearing bit 1)
2674 * The least significant bit can be cleared this way, and it
2675 * just so happens that it is the same bit corresponding to
2676 * the current context.
2678 static DEFINE_PER_CPU(unsigned int, current_context);
2680 static __always_inline int trace_recursive_lock(void)
2682 unsigned int val = __this_cpu_read(current_context);
2685 if (in_interrupt()) {
2695 if (unlikely(val & (1 << bit)))
2699 __this_cpu_write(current_context, val);
2704 static __always_inline void trace_recursive_unlock(void)
2706 __this_cpu_and(current_context, __this_cpu_read(current_context) - 1);
2711 #define trace_recursive_lock() (0)
2712 #define trace_recursive_unlock() do { } while (0)
2717 * ring_buffer_lock_reserve - reserve a part of the buffer
2718 * @buffer: the ring buffer to reserve from
2719 * @length: the length of the data to reserve (excluding event header)
2721 * Returns a reseverd event on the ring buffer to copy directly to.
2722 * The user of this interface will need to get the body to write into
2723 * and can use the ring_buffer_event_data() interface.
2725 * The length is the length of the data needed, not the event length
2726 * which also includes the event header.
2728 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2729 * If NULL is returned, then nothing has been allocated or locked.
2731 struct ring_buffer_event *
2732 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2734 struct ring_buffer_per_cpu *cpu_buffer;
2735 struct ring_buffer_event *event;
2738 if (ring_buffer_flags != RB_BUFFERS_ON)
2741 /* If we are tracing schedule, we don't want to recurse */
2742 preempt_disable_notrace();
2744 if (unlikely(atomic_read(&buffer->record_disabled)))
2747 if (unlikely(trace_recursive_lock()))
2750 cpu = raw_smp_processor_id();
2752 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2755 cpu_buffer = buffer->buffers[cpu];
2757 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2760 if (unlikely(length > BUF_MAX_DATA_SIZE))
2763 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2770 trace_recursive_unlock();
2773 preempt_enable_notrace();
2776 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2779 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2780 struct ring_buffer_event *event)
2785 * The event first in the commit queue updates the
2788 if (rb_event_is_commit(cpu_buffer, event)) {
2790 * A commit event that is first on a page
2791 * updates the write timestamp with the page stamp
2793 if (!rb_event_index(event))
2794 cpu_buffer->write_stamp =
2795 cpu_buffer->commit_page->page->time_stamp;
2796 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2797 delta = event->array[0];
2799 delta += event->time_delta;
2800 cpu_buffer->write_stamp += delta;
2802 cpu_buffer->write_stamp += event->time_delta;
2806 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2807 struct ring_buffer_event *event)
2809 local_inc(&cpu_buffer->entries);
2810 rb_update_write_stamp(cpu_buffer, event);
2811 rb_end_commit(cpu_buffer);
2814 static __always_inline void
2815 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2819 if (buffer->irq_work.waiters_pending) {
2820 buffer->irq_work.waiters_pending = false;
2821 /* irq_work_queue() supplies it's own memory barriers */
2822 irq_work_queue(&buffer->irq_work.work);
2825 if (cpu_buffer->irq_work.waiters_pending) {
2826 cpu_buffer->irq_work.waiters_pending = false;
2827 /* irq_work_queue() supplies it's own memory barriers */
2828 irq_work_queue(&cpu_buffer->irq_work.work);
2831 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
2833 if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2834 cpu_buffer->irq_work.wakeup_full = true;
2835 cpu_buffer->irq_work.full_waiters_pending = false;
2836 /* irq_work_queue() supplies it's own memory barriers */
2837 irq_work_queue(&cpu_buffer->irq_work.work);
2842 * ring_buffer_unlock_commit - commit a reserved
2843 * @buffer: The buffer to commit to
2844 * @event: The event pointer to commit.
2846 * This commits the data to the ring buffer, and releases any locks held.
2848 * Must be paired with ring_buffer_lock_reserve.
2850 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2851 struct ring_buffer_event *event)
2853 struct ring_buffer_per_cpu *cpu_buffer;
2854 int cpu = raw_smp_processor_id();
2856 cpu_buffer = buffer->buffers[cpu];
2858 rb_commit(cpu_buffer, event);
2860 rb_wakeups(buffer, cpu_buffer);
2862 trace_recursive_unlock();
2864 preempt_enable_notrace();
2868 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2870 static inline void rb_event_discard(struct ring_buffer_event *event)
2872 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2873 event = skip_time_extend(event);
2875 /* array[0] holds the actual length for the discarded event */
2876 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2877 event->type_len = RINGBUF_TYPE_PADDING;
2878 /* time delta must be non zero */
2879 if (!event->time_delta)
2880 event->time_delta = 1;
2884 * Decrement the entries to the page that an event is on.
2885 * The event does not even need to exist, only the pointer
2886 * to the page it is on. This may only be called before the commit
2890 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2891 struct ring_buffer_event *event)
2893 unsigned long addr = (unsigned long)event;
2894 struct buffer_page *bpage = cpu_buffer->commit_page;
2895 struct buffer_page *start;
2899 /* Do the likely case first */
2900 if (likely(bpage->page == (void *)addr)) {
2901 local_dec(&bpage->entries);
2906 * Because the commit page may be on the reader page we
2907 * start with the next page and check the end loop there.
2909 rb_inc_page(cpu_buffer, &bpage);
2912 if (bpage->page == (void *)addr) {
2913 local_dec(&bpage->entries);
2916 rb_inc_page(cpu_buffer, &bpage);
2917 } while (bpage != start);
2919 /* commit not part of this buffer?? */
2920 RB_WARN_ON(cpu_buffer, 1);
2924 * ring_buffer_commit_discard - discard an event that has not been committed
2925 * @buffer: the ring buffer
2926 * @event: non committed event to discard
2928 * Sometimes an event that is in the ring buffer needs to be ignored.
2929 * This function lets the user discard an event in the ring buffer
2930 * and then that event will not be read later.
2932 * This function only works if it is called before the the item has been
2933 * committed. It will try to free the event from the ring buffer
2934 * if another event has not been added behind it.
2936 * If another event has been added behind it, it will set the event
2937 * up as discarded, and perform the commit.
2939 * If this function is called, do not call ring_buffer_unlock_commit on
2942 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2943 struct ring_buffer_event *event)
2945 struct ring_buffer_per_cpu *cpu_buffer;
2948 /* The event is discarded regardless */
2949 rb_event_discard(event);
2951 cpu = smp_processor_id();
2952 cpu_buffer = buffer->buffers[cpu];
2955 * This must only be called if the event has not been
2956 * committed yet. Thus we can assume that preemption
2957 * is still disabled.
2959 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2961 rb_decrement_entry(cpu_buffer, event);
2962 if (rb_try_to_discard(cpu_buffer, event))
2966 * The commit is still visible by the reader, so we
2967 * must still update the timestamp.
2969 rb_update_write_stamp(cpu_buffer, event);
2971 rb_end_commit(cpu_buffer);
2973 trace_recursive_unlock();
2975 preempt_enable_notrace();
2978 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2981 * ring_buffer_write - write data to the buffer without reserving
2982 * @buffer: The ring buffer to write to.
2983 * @length: The length of the data being written (excluding the event header)
2984 * @data: The data to write to the buffer.
2986 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2987 * one function. If you already have the data to write to the buffer, it
2988 * may be easier to simply call this function.
2990 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2991 * and not the length of the event which would hold the header.
2993 int ring_buffer_write(struct ring_buffer *buffer,
2994 unsigned long length,
2997 struct ring_buffer_per_cpu *cpu_buffer;
2998 struct ring_buffer_event *event;
3003 if (ring_buffer_flags != RB_BUFFERS_ON)
3006 preempt_disable_notrace();
3008 if (atomic_read(&buffer->record_disabled))
3011 cpu = raw_smp_processor_id();
3013 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3016 cpu_buffer = buffer->buffers[cpu];
3018 if (atomic_read(&cpu_buffer->record_disabled))
3021 if (length > BUF_MAX_DATA_SIZE)
3024 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3028 body = rb_event_data(event);
3030 memcpy(body, data, length);
3032 rb_commit(cpu_buffer, event);
3034 rb_wakeups(buffer, cpu_buffer);
3038 preempt_enable_notrace();
3042 EXPORT_SYMBOL_GPL(ring_buffer_write);
3044 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3046 struct buffer_page *reader = cpu_buffer->reader_page;
3047 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3048 struct buffer_page *commit = cpu_buffer->commit_page;
3050 /* In case of error, head will be NULL */
3051 if (unlikely(!head))
3054 return reader->read == rb_page_commit(reader) &&
3055 (commit == reader ||
3057 head->read == rb_page_commit(commit)));
3061 * ring_buffer_record_disable - stop all writes into the buffer
3062 * @buffer: The ring buffer to stop writes to.
3064 * This prevents all writes to the buffer. Any attempt to write
3065 * to the buffer after this will fail and return NULL.
3067 * The caller should call synchronize_sched() after this.
3069 void ring_buffer_record_disable(struct ring_buffer *buffer)
3071 atomic_inc(&buffer->record_disabled);
3073 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3076 * ring_buffer_record_enable - enable writes to the buffer
3077 * @buffer: The ring buffer to enable writes
3079 * Note, multiple disables will need the same number of enables
3080 * to truly enable the writing (much like preempt_disable).
3082 void ring_buffer_record_enable(struct ring_buffer *buffer)
3084 atomic_dec(&buffer->record_disabled);
3086 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3089 * ring_buffer_record_off - stop all writes into the buffer
3090 * @buffer: The ring buffer to stop writes to.
3092 * This prevents all writes to the buffer. Any attempt to write
3093 * to the buffer after this will fail and return NULL.
3095 * This is different than ring_buffer_record_disable() as
3096 * it works like an on/off switch, where as the disable() version
3097 * must be paired with a enable().
3099 void ring_buffer_record_off(struct ring_buffer *buffer)
3102 unsigned int new_rd;
3105 rd = atomic_read(&buffer->record_disabled);
3106 new_rd = rd | RB_BUFFER_OFF;
3107 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3109 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3112 * ring_buffer_record_on - restart writes into the buffer
3113 * @buffer: The ring buffer to start writes to.
3115 * This enables all writes to the buffer that was disabled by
3116 * ring_buffer_record_off().
3118 * This is different than ring_buffer_record_enable() as
3119 * it works like an on/off switch, where as the enable() version
3120 * must be paired with a disable().
3122 void ring_buffer_record_on(struct ring_buffer *buffer)
3125 unsigned int new_rd;
3128 rd = atomic_read(&buffer->record_disabled);
3129 new_rd = rd & ~RB_BUFFER_OFF;
3130 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3132 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3135 * ring_buffer_record_is_on - return true if the ring buffer can write
3136 * @buffer: The ring buffer to see if write is enabled
3138 * Returns true if the ring buffer is in a state that it accepts writes.
3140 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3142 return !atomic_read(&buffer->record_disabled);
3146 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3147 * @buffer: The ring buffer to stop writes to.
3148 * @cpu: The CPU buffer to stop
3150 * This prevents all writes to the buffer. Any attempt to write
3151 * to the buffer after this will fail and return NULL.
3153 * The caller should call synchronize_sched() after this.
3155 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3157 struct ring_buffer_per_cpu *cpu_buffer;
3159 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3162 cpu_buffer = buffer->buffers[cpu];
3163 atomic_inc(&cpu_buffer->record_disabled);
3165 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3168 * ring_buffer_record_enable_cpu - enable writes to the buffer
3169 * @buffer: The ring buffer to enable writes
3170 * @cpu: The CPU to enable.
3172 * Note, multiple disables will need the same number of enables
3173 * to truly enable the writing (much like preempt_disable).
3175 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3177 struct ring_buffer_per_cpu *cpu_buffer;
3179 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3182 cpu_buffer = buffer->buffers[cpu];
3183 atomic_dec(&cpu_buffer->record_disabled);
3185 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3188 * The total entries in the ring buffer is the running counter
3189 * of entries entered into the ring buffer, minus the sum of
3190 * the entries read from the ring buffer and the number of
3191 * entries that were overwritten.
3193 static inline unsigned long
3194 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3196 return local_read(&cpu_buffer->entries) -
3197 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3201 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3202 * @buffer: The ring buffer
3203 * @cpu: The per CPU buffer to read from.
3205 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3207 unsigned long flags;
3208 struct ring_buffer_per_cpu *cpu_buffer;
3209 struct buffer_page *bpage;
3212 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3215 cpu_buffer = buffer->buffers[cpu];
3216 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3218 * if the tail is on reader_page, oldest time stamp is on the reader
3221 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3222 bpage = cpu_buffer->reader_page;
3224 bpage = rb_set_head_page(cpu_buffer);
3226 ret = bpage->page->time_stamp;
3227 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3231 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3234 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3235 * @buffer: The ring buffer
3236 * @cpu: The per CPU buffer to read from.
3238 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3240 struct ring_buffer_per_cpu *cpu_buffer;
3243 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3246 cpu_buffer = buffer->buffers[cpu];
3247 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3251 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3254 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3255 * @buffer: The ring buffer
3256 * @cpu: The per CPU buffer to get the entries from.
3258 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3260 struct ring_buffer_per_cpu *cpu_buffer;
3262 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3265 cpu_buffer = buffer->buffers[cpu];
3267 return rb_num_of_entries(cpu_buffer);
3269 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3272 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3273 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3274 * @buffer: The ring buffer
3275 * @cpu: The per CPU buffer to get the number of overruns from
3277 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3279 struct ring_buffer_per_cpu *cpu_buffer;
3282 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3285 cpu_buffer = buffer->buffers[cpu];
3286 ret = local_read(&cpu_buffer->overrun);
3290 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3293 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3294 * commits failing due to the buffer wrapping around while there are uncommitted
3295 * events, such as during an interrupt storm.
3296 * @buffer: The ring buffer
3297 * @cpu: The per CPU buffer to get the number of overruns from
3300 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3302 struct ring_buffer_per_cpu *cpu_buffer;
3305 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3308 cpu_buffer = buffer->buffers[cpu];
3309 ret = local_read(&cpu_buffer->commit_overrun);
3313 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3316 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3317 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3318 * @buffer: The ring buffer
3319 * @cpu: The per CPU buffer to get the number of overruns from
3322 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3324 struct ring_buffer_per_cpu *cpu_buffer;
3327 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3330 cpu_buffer = buffer->buffers[cpu];
3331 ret = local_read(&cpu_buffer->dropped_events);
3335 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3338 * ring_buffer_read_events_cpu - get the number of events successfully read
3339 * @buffer: The ring buffer
3340 * @cpu: The per CPU buffer to get the number of events read
3343 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3345 struct ring_buffer_per_cpu *cpu_buffer;
3347 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3350 cpu_buffer = buffer->buffers[cpu];
3351 return cpu_buffer->read;
3353 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3356 * ring_buffer_entries - get the number of entries in a buffer
3357 * @buffer: The ring buffer
3359 * Returns the total number of entries in the ring buffer
3362 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3364 struct ring_buffer_per_cpu *cpu_buffer;
3365 unsigned long entries = 0;
3368 /* if you care about this being correct, lock the buffer */
3369 for_each_buffer_cpu(buffer, cpu) {
3370 cpu_buffer = buffer->buffers[cpu];
3371 entries += rb_num_of_entries(cpu_buffer);
3376 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3379 * ring_buffer_overruns - get the number of overruns in buffer
3380 * @buffer: The ring buffer
3382 * Returns the total number of overruns in the ring buffer
3385 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3387 struct ring_buffer_per_cpu *cpu_buffer;
3388 unsigned long overruns = 0;
3391 /* if you care about this being correct, lock the buffer */
3392 for_each_buffer_cpu(buffer, cpu) {
3393 cpu_buffer = buffer->buffers[cpu];
3394 overruns += local_read(&cpu_buffer->overrun);
3399 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3401 static void rb_iter_reset(struct ring_buffer_iter *iter)
3403 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3405 /* Iterator usage is expected to have record disabled */
3406 iter->head_page = cpu_buffer->reader_page;
3407 iter->head = cpu_buffer->reader_page->read;
3409 iter->cache_reader_page = iter->head_page;
3410 iter->cache_read = cpu_buffer->read;
3413 iter->read_stamp = cpu_buffer->read_stamp;
3415 iter->read_stamp = iter->head_page->page->time_stamp;
3419 * ring_buffer_iter_reset - reset an iterator
3420 * @iter: The iterator to reset
3422 * Resets the iterator, so that it will start from the beginning
3425 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3427 struct ring_buffer_per_cpu *cpu_buffer;
3428 unsigned long flags;
3433 cpu_buffer = iter->cpu_buffer;
3435 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3436 rb_iter_reset(iter);
3437 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3439 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3442 * ring_buffer_iter_empty - check if an iterator has no more to read
3443 * @iter: The iterator to check
3445 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3447 struct ring_buffer_per_cpu *cpu_buffer;
3449 cpu_buffer = iter->cpu_buffer;
3451 return iter->head_page == cpu_buffer->commit_page &&
3452 iter->head == rb_commit_index(cpu_buffer);
3454 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3457 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3458 struct ring_buffer_event *event)
3462 switch (event->type_len) {
3463 case RINGBUF_TYPE_PADDING:
3466 case RINGBUF_TYPE_TIME_EXTEND:
3467 delta = event->array[0];
3469 delta += event->time_delta;
3470 cpu_buffer->read_stamp += delta;
3473 case RINGBUF_TYPE_TIME_STAMP:
3474 /* FIXME: not implemented */
3477 case RINGBUF_TYPE_DATA:
3478 cpu_buffer->read_stamp += event->time_delta;
3488 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3489 struct ring_buffer_event *event)
3493 switch (event->type_len) {
3494 case RINGBUF_TYPE_PADDING:
3497 case RINGBUF_TYPE_TIME_EXTEND:
3498 delta = event->array[0];
3500 delta += event->time_delta;
3501 iter->read_stamp += delta;
3504 case RINGBUF_TYPE_TIME_STAMP:
3505 /* FIXME: not implemented */
3508 case RINGBUF_TYPE_DATA:
3509 iter->read_stamp += event->time_delta;
3518 static struct buffer_page *
3519 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3521 struct buffer_page *reader = NULL;
3522 unsigned long overwrite;
3523 unsigned long flags;
3527 local_irq_save(flags);
3528 arch_spin_lock(&cpu_buffer->lock);
3532 * This should normally only loop twice. But because the
3533 * start of the reader inserts an empty page, it causes
3534 * a case where we will loop three times. There should be no
3535 * reason to loop four times (that I know of).
3537 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3542 reader = cpu_buffer->reader_page;
3544 /* If there's more to read, return this page */
3545 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3548 /* Never should we have an index greater than the size */
3549 if (RB_WARN_ON(cpu_buffer,
3550 cpu_buffer->reader_page->read > rb_page_size(reader)))
3553 /* check if we caught up to the tail */
3555 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3558 /* Don't bother swapping if the ring buffer is empty */
3559 if (rb_num_of_entries(cpu_buffer) == 0)
3563 * Reset the reader page to size zero.
3565 local_set(&cpu_buffer->reader_page->write, 0);
3566 local_set(&cpu_buffer->reader_page->entries, 0);
3567 local_set(&cpu_buffer->reader_page->page->commit, 0);
3568 cpu_buffer->reader_page->real_end = 0;
3572 * Splice the empty reader page into the list around the head.
3574 reader = rb_set_head_page(cpu_buffer);
3577 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3578 cpu_buffer->reader_page->list.prev = reader->list.prev;
3581 * cpu_buffer->pages just needs to point to the buffer, it
3582 * has no specific buffer page to point to. Lets move it out
3583 * of our way so we don't accidentally swap it.
3585 cpu_buffer->pages = reader->list.prev;
3587 /* The reader page will be pointing to the new head */
3588 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3591 * We want to make sure we read the overruns after we set up our
3592 * pointers to the next object. The writer side does a
3593 * cmpxchg to cross pages which acts as the mb on the writer
3594 * side. Note, the reader will constantly fail the swap
3595 * while the writer is updating the pointers, so this
3596 * guarantees that the overwrite recorded here is the one we
3597 * want to compare with the last_overrun.
3600 overwrite = local_read(&(cpu_buffer->overrun));
3603 * Here's the tricky part.
3605 * We need to move the pointer past the header page.
3606 * But we can only do that if a writer is not currently
3607 * moving it. The page before the header page has the
3608 * flag bit '1' set if it is pointing to the page we want.
3609 * but if the writer is in the process of moving it
3610 * than it will be '2' or already moved '0'.
3613 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3616 * If we did not convert it, then we must try again.
3622 * Yeah! We succeeded in replacing the page.
3624 * Now make the new head point back to the reader page.
3626 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3627 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3629 /* Finally update the reader page to the new head */
3630 cpu_buffer->reader_page = reader;
3631 rb_reset_reader_page(cpu_buffer);
3633 if (overwrite != cpu_buffer->last_overrun) {
3634 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3635 cpu_buffer->last_overrun = overwrite;
3641 arch_spin_unlock(&cpu_buffer->lock);
3642 local_irq_restore(flags);
3647 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3649 struct ring_buffer_event *event;
3650 struct buffer_page *reader;
3653 reader = rb_get_reader_page(cpu_buffer);
3655 /* This function should not be called when buffer is empty */
3656 if (RB_WARN_ON(cpu_buffer, !reader))
3659 event = rb_reader_event(cpu_buffer);
3661 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3664 rb_update_read_stamp(cpu_buffer, event);
3666 length = rb_event_length(event);
3667 cpu_buffer->reader_page->read += length;
3670 static void rb_advance_iter(struct ring_buffer_iter *iter)
3672 struct ring_buffer_per_cpu *cpu_buffer;
3673 struct ring_buffer_event *event;
3676 cpu_buffer = iter->cpu_buffer;
3679 * Check if we are at the end of the buffer.
3681 if (iter->head >= rb_page_size(iter->head_page)) {
3682 /* discarded commits can make the page empty */
3683 if (iter->head_page == cpu_buffer->commit_page)
3689 event = rb_iter_head_event(iter);
3691 length = rb_event_length(event);
3694 * This should not be called to advance the header if we are
3695 * at the tail of the buffer.
3697 if (RB_WARN_ON(cpu_buffer,
3698 (iter->head_page == cpu_buffer->commit_page) &&
3699 (iter->head + length > rb_commit_index(cpu_buffer))))
3702 rb_update_iter_read_stamp(iter, event);
3704 iter->head += length;
3706 /* check for end of page padding */
3707 if ((iter->head >= rb_page_size(iter->head_page)) &&
3708 (iter->head_page != cpu_buffer->commit_page))
3712 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3714 return cpu_buffer->lost_events;
3717 static struct ring_buffer_event *
3718 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3719 unsigned long *lost_events)
3721 struct ring_buffer_event *event;
3722 struct buffer_page *reader;
3727 * We repeat when a time extend is encountered.
3728 * Since the time extend is always attached to a data event,
3729 * we should never loop more than once.
3730 * (We never hit the following condition more than twice).
3732 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3735 reader = rb_get_reader_page(cpu_buffer);
3739 event = rb_reader_event(cpu_buffer);
3741 switch (event->type_len) {
3742 case RINGBUF_TYPE_PADDING:
3743 if (rb_null_event(event))
3744 RB_WARN_ON(cpu_buffer, 1);
3746 * Because the writer could be discarding every
3747 * event it creates (which would probably be bad)
3748 * if we were to go back to "again" then we may never
3749 * catch up, and will trigger the warn on, or lock
3750 * the box. Return the padding, and we will release
3751 * the current locks, and try again.
3755 case RINGBUF_TYPE_TIME_EXTEND:
3756 /* Internal data, OK to advance */
3757 rb_advance_reader(cpu_buffer);
3760 case RINGBUF_TYPE_TIME_STAMP:
3761 /* FIXME: not implemented */
3762 rb_advance_reader(cpu_buffer);
3765 case RINGBUF_TYPE_DATA:
3767 *ts = cpu_buffer->read_stamp + event->time_delta;
3768 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3769 cpu_buffer->cpu, ts);
3772 *lost_events = rb_lost_events(cpu_buffer);
3781 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3783 static struct ring_buffer_event *
3784 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3786 struct ring_buffer *buffer;
3787 struct ring_buffer_per_cpu *cpu_buffer;
3788 struct ring_buffer_event *event;
3791 cpu_buffer = iter->cpu_buffer;
3792 buffer = cpu_buffer->buffer;
3795 * Check if someone performed a consuming read to
3796 * the buffer. A consuming read invalidates the iterator
3797 * and we need to reset the iterator in this case.
3799 if (unlikely(iter->cache_read != cpu_buffer->read ||
3800 iter->cache_reader_page != cpu_buffer->reader_page))
3801 rb_iter_reset(iter);
3804 if (ring_buffer_iter_empty(iter))
3808 * We repeat when a time extend is encountered or we hit
3809 * the end of the page. Since the time extend is always attached
3810 * to a data event, we should never loop more than three times.
3811 * Once for going to next page, once on time extend, and
3812 * finally once to get the event.
3813 * (We never hit the following condition more than thrice).
3815 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3818 if (rb_per_cpu_empty(cpu_buffer))
3821 if (iter->head >= rb_page_size(iter->head_page)) {
3826 event = rb_iter_head_event(iter);
3828 switch (event->type_len) {
3829 case RINGBUF_TYPE_PADDING:
3830 if (rb_null_event(event)) {
3834 rb_advance_iter(iter);
3837 case RINGBUF_TYPE_TIME_EXTEND:
3838 /* Internal data, OK to advance */
3839 rb_advance_iter(iter);
3842 case RINGBUF_TYPE_TIME_STAMP:
3843 /* FIXME: not implemented */
3844 rb_advance_iter(iter);
3847 case RINGBUF_TYPE_DATA:
3849 *ts = iter->read_stamp + event->time_delta;
3850 ring_buffer_normalize_time_stamp(buffer,
3851 cpu_buffer->cpu, ts);
3861 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3863 static inline int rb_ok_to_lock(void)
3866 * If an NMI die dumps out the content of the ring buffer
3867 * do not grab locks. We also permanently disable the ring
3868 * buffer too. A one time deal is all you get from reading
3869 * the ring buffer from an NMI.
3871 if (likely(!in_nmi()))
3874 tracing_off_permanent();
3879 * ring_buffer_peek - peek at the next event to be read
3880 * @buffer: The ring buffer to read
3881 * @cpu: The cpu to peak at
3882 * @ts: The timestamp counter of this event.
3883 * @lost_events: a variable to store if events were lost (may be NULL)
3885 * This will return the event that will be read next, but does
3886 * not consume the data.
3888 struct ring_buffer_event *
3889 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3890 unsigned long *lost_events)
3892 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3893 struct ring_buffer_event *event;
3894 unsigned long flags;
3897 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3900 dolock = rb_ok_to_lock();
3902 local_irq_save(flags);
3904 raw_spin_lock(&cpu_buffer->reader_lock);
3905 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3906 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3907 rb_advance_reader(cpu_buffer);
3909 raw_spin_unlock(&cpu_buffer->reader_lock);
3910 local_irq_restore(flags);
3912 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3919 * ring_buffer_iter_peek - peek at the next event to be read
3920 * @iter: The ring buffer iterator
3921 * @ts: The timestamp counter of this event.
3923 * This will return the event that will be read next, but does
3924 * not increment the iterator.
3926 struct ring_buffer_event *
3927 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3929 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3930 struct ring_buffer_event *event;
3931 unsigned long flags;
3934 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3935 event = rb_iter_peek(iter, ts);
3936 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3938 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3945 * ring_buffer_consume - return an event and consume it
3946 * @buffer: The ring buffer to get the next event from
3947 * @cpu: the cpu to read the buffer from
3948 * @ts: a variable to store the timestamp (may be NULL)
3949 * @lost_events: a variable to store if events were lost (may be NULL)
3951 * Returns the next event in the ring buffer, and that event is consumed.
3952 * Meaning, that sequential reads will keep returning a different event,
3953 * and eventually empty the ring buffer if the producer is slower.
3955 struct ring_buffer_event *
3956 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3957 unsigned long *lost_events)
3959 struct ring_buffer_per_cpu *cpu_buffer;
3960 struct ring_buffer_event *event = NULL;
3961 unsigned long flags;
3964 dolock = rb_ok_to_lock();
3967 /* might be called in atomic */
3970 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3973 cpu_buffer = buffer->buffers[cpu];
3974 local_irq_save(flags);
3976 raw_spin_lock(&cpu_buffer->reader_lock);
3978 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3980 cpu_buffer->lost_events = 0;
3981 rb_advance_reader(cpu_buffer);
3985 raw_spin_unlock(&cpu_buffer->reader_lock);
3986 local_irq_restore(flags);
3991 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3996 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3999 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4000 * @buffer: The ring buffer to read from
4001 * @cpu: The cpu buffer to iterate over
4003 * This performs the initial preparations necessary to iterate
4004 * through the buffer. Memory is allocated, buffer recording
4005 * is disabled, and the iterator pointer is returned to the caller.
4007 * Disabling buffer recordng prevents the reading from being
4008 * corrupted. This is not a consuming read, so a producer is not
4011 * After a sequence of ring_buffer_read_prepare calls, the user is
4012 * expected to make at least one call to ring_buffer_read_prepare_sync.
4013 * Afterwards, ring_buffer_read_start is invoked to get things going
4016 * This overall must be paired with ring_buffer_read_finish.
4018 struct ring_buffer_iter *
4019 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
4021 struct ring_buffer_per_cpu *cpu_buffer;
4022 struct ring_buffer_iter *iter;
4024 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4027 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
4031 cpu_buffer = buffer->buffers[cpu];
4033 iter->cpu_buffer = cpu_buffer;
4035 atomic_inc(&buffer->resize_disabled);
4036 atomic_inc(&cpu_buffer->record_disabled);
4040 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4043 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4045 * All previously invoked ring_buffer_read_prepare calls to prepare
4046 * iterators will be synchronized. Afterwards, read_buffer_read_start
4047 * calls on those iterators are allowed.
4050 ring_buffer_read_prepare_sync(void)
4052 synchronize_sched();
4054 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4057 * ring_buffer_read_start - start a non consuming read of the buffer
4058 * @iter: The iterator returned by ring_buffer_read_prepare
4060 * This finalizes the startup of an iteration through the buffer.
4061 * The iterator comes from a call to ring_buffer_read_prepare and
4062 * an intervening ring_buffer_read_prepare_sync must have been
4065 * Must be paired with ring_buffer_read_finish.
4068 ring_buffer_read_start(struct ring_buffer_iter *iter)
4070 struct ring_buffer_per_cpu *cpu_buffer;
4071 unsigned long flags;
4076 cpu_buffer = iter->cpu_buffer;
4078 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4079 arch_spin_lock(&cpu_buffer->lock);
4080 rb_iter_reset(iter);
4081 arch_spin_unlock(&cpu_buffer->lock);
4082 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4084 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4087 * ring_buffer_read_finish - finish reading the iterator of the buffer
4088 * @iter: The iterator retrieved by ring_buffer_start
4090 * This re-enables the recording to the buffer, and frees the
4094 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4096 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4097 unsigned long flags;
4100 * Ring buffer is disabled from recording, here's a good place
4101 * to check the integrity of the ring buffer.
4102 * Must prevent readers from trying to read, as the check
4103 * clears the HEAD page and readers require it.
4105 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4106 rb_check_pages(cpu_buffer);
4107 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4109 atomic_dec(&cpu_buffer->record_disabled);
4110 atomic_dec(&cpu_buffer->buffer->resize_disabled);
4113 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4116 * ring_buffer_read - read the next item in the ring buffer by the iterator
4117 * @iter: The ring buffer iterator
4118 * @ts: The time stamp of the event read.
4120 * This reads the next event in the ring buffer and increments the iterator.
4122 struct ring_buffer_event *
4123 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4125 struct ring_buffer_event *event;
4126 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4127 unsigned long flags;
4129 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4131 event = rb_iter_peek(iter, ts);
4135 if (event->type_len == RINGBUF_TYPE_PADDING)
4138 rb_advance_iter(iter);
4140 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4144 EXPORT_SYMBOL_GPL(ring_buffer_read);
4147 * ring_buffer_size - return the size of the ring buffer (in bytes)
4148 * @buffer: The ring buffer.
4150 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4153 * Earlier, this method returned
4154 * BUF_PAGE_SIZE * buffer->nr_pages
4155 * Since the nr_pages field is now removed, we have converted this to
4156 * return the per cpu buffer value.
4158 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4161 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4163 EXPORT_SYMBOL_GPL(ring_buffer_size);
4166 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4168 rb_head_page_deactivate(cpu_buffer);
4170 cpu_buffer->head_page
4171 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4172 local_set(&cpu_buffer->head_page->write, 0);
4173 local_set(&cpu_buffer->head_page->entries, 0);
4174 local_set(&cpu_buffer->head_page->page->commit, 0);
4176 cpu_buffer->head_page->read = 0;
4178 cpu_buffer->tail_page = cpu_buffer->head_page;
4179 cpu_buffer->commit_page = cpu_buffer->head_page;
4181 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4182 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4183 local_set(&cpu_buffer->reader_page->write, 0);
4184 local_set(&cpu_buffer->reader_page->entries, 0);
4185 local_set(&cpu_buffer->reader_page->page->commit, 0);
4186 cpu_buffer->reader_page->read = 0;
4188 local_set(&cpu_buffer->entries_bytes, 0);
4189 local_set(&cpu_buffer->overrun, 0);
4190 local_set(&cpu_buffer->commit_overrun, 0);
4191 local_set(&cpu_buffer->dropped_events, 0);
4192 local_set(&cpu_buffer->entries, 0);
4193 local_set(&cpu_buffer->committing, 0);
4194 local_set(&cpu_buffer->commits, 0);
4195 cpu_buffer->read = 0;
4196 cpu_buffer->read_bytes = 0;
4198 cpu_buffer->write_stamp = 0;
4199 cpu_buffer->read_stamp = 0;
4201 cpu_buffer->lost_events = 0;
4202 cpu_buffer->last_overrun = 0;
4204 rb_head_page_activate(cpu_buffer);
4208 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4209 * @buffer: The ring buffer to reset a per cpu buffer of
4210 * @cpu: The CPU buffer to be reset
4212 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4214 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4215 unsigned long flags;
4217 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4220 atomic_inc(&buffer->resize_disabled);
4221 atomic_inc(&cpu_buffer->record_disabled);
4223 /* Make sure all commits have finished */
4224 synchronize_sched();
4226 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4228 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4231 arch_spin_lock(&cpu_buffer->lock);
4233 rb_reset_cpu(cpu_buffer);
4235 arch_spin_unlock(&cpu_buffer->lock);
4238 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4240 atomic_dec(&cpu_buffer->record_disabled);
4241 atomic_dec(&buffer->resize_disabled);
4243 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4246 * ring_buffer_reset - reset a ring buffer
4247 * @buffer: The ring buffer to reset all cpu buffers
4249 void ring_buffer_reset(struct ring_buffer *buffer)
4253 for_each_buffer_cpu(buffer, cpu)
4254 ring_buffer_reset_cpu(buffer, cpu);
4256 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4259 * rind_buffer_empty - is the ring buffer empty?
4260 * @buffer: The ring buffer to test
4262 int ring_buffer_empty(struct ring_buffer *buffer)
4264 struct ring_buffer_per_cpu *cpu_buffer;
4265 unsigned long flags;
4270 dolock = rb_ok_to_lock();
4272 /* yes this is racy, but if you don't like the race, lock the buffer */
4273 for_each_buffer_cpu(buffer, cpu) {
4274 cpu_buffer = buffer->buffers[cpu];
4275 local_irq_save(flags);
4277 raw_spin_lock(&cpu_buffer->reader_lock);
4278 ret = rb_per_cpu_empty(cpu_buffer);
4280 raw_spin_unlock(&cpu_buffer->reader_lock);
4281 local_irq_restore(flags);
4289 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4292 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4293 * @buffer: The ring buffer
4294 * @cpu: The CPU buffer to test
4296 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4298 struct ring_buffer_per_cpu *cpu_buffer;
4299 unsigned long flags;
4303 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4306 dolock = rb_ok_to_lock();
4308 cpu_buffer = buffer->buffers[cpu];
4309 local_irq_save(flags);
4311 raw_spin_lock(&cpu_buffer->reader_lock);
4312 ret = rb_per_cpu_empty(cpu_buffer);
4314 raw_spin_unlock(&cpu_buffer->reader_lock);
4315 local_irq_restore(flags);
4319 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4321 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4323 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4324 * @buffer_a: One buffer to swap with
4325 * @buffer_b: The other buffer to swap with
4327 * This function is useful for tracers that want to take a "snapshot"
4328 * of a CPU buffer and has another back up buffer lying around.
4329 * it is expected that the tracer handles the cpu buffer not being
4330 * used at the moment.
4332 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4333 struct ring_buffer *buffer_b, int cpu)
4335 struct ring_buffer_per_cpu *cpu_buffer_a;
4336 struct ring_buffer_per_cpu *cpu_buffer_b;
4339 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4340 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4343 cpu_buffer_a = buffer_a->buffers[cpu];
4344 cpu_buffer_b = buffer_b->buffers[cpu];
4346 /* At least make sure the two buffers are somewhat the same */
4347 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4352 if (ring_buffer_flags != RB_BUFFERS_ON)
4355 if (atomic_read(&buffer_a->record_disabled))
4358 if (atomic_read(&buffer_b->record_disabled))
4361 if (atomic_read(&cpu_buffer_a->record_disabled))
4364 if (atomic_read(&cpu_buffer_b->record_disabled))
4368 * We can't do a synchronize_sched here because this
4369 * function can be called in atomic context.
4370 * Normally this will be called from the same CPU as cpu.
4371 * If not it's up to the caller to protect this.
4373 atomic_inc(&cpu_buffer_a->record_disabled);
4374 atomic_inc(&cpu_buffer_b->record_disabled);
4377 if (local_read(&cpu_buffer_a->committing))
4379 if (local_read(&cpu_buffer_b->committing))
4382 buffer_a->buffers[cpu] = cpu_buffer_b;
4383 buffer_b->buffers[cpu] = cpu_buffer_a;
4385 cpu_buffer_b->buffer = buffer_a;
4386 cpu_buffer_a->buffer = buffer_b;
4391 atomic_dec(&cpu_buffer_a->record_disabled);
4392 atomic_dec(&cpu_buffer_b->record_disabled);
4396 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4397 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4400 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4401 * @buffer: the buffer to allocate for.
4402 * @cpu: the cpu buffer to allocate.
4404 * This function is used in conjunction with ring_buffer_read_page.
4405 * When reading a full page from the ring buffer, these functions
4406 * can be used to speed up the process. The calling function should
4407 * allocate a few pages first with this function. Then when it
4408 * needs to get pages from the ring buffer, it passes the result
4409 * of this function into ring_buffer_read_page, which will swap
4410 * the page that was allocated, with the read page of the buffer.
4413 * The page allocated, or NULL on error.
4415 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4417 struct buffer_data_page *bpage;
4420 page = alloc_pages_node(cpu_to_node(cpu),
4421 GFP_KERNEL | __GFP_NORETRY, 0);
4425 bpage = page_address(page);
4427 rb_init_page(bpage);
4431 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4434 * ring_buffer_free_read_page - free an allocated read page
4435 * @buffer: the buffer the page was allocate for
4436 * @data: the page to free
4438 * Free a page allocated from ring_buffer_alloc_read_page.
4440 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4442 free_page((unsigned long)data);
4444 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4447 * ring_buffer_read_page - extract a page from the ring buffer
4448 * @buffer: buffer to extract from
4449 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4450 * @len: amount to extract
4451 * @cpu: the cpu of the buffer to extract
4452 * @full: should the extraction only happen when the page is full.
4454 * This function will pull out a page from the ring buffer and consume it.
4455 * @data_page must be the address of the variable that was returned
4456 * from ring_buffer_alloc_read_page. This is because the page might be used
4457 * to swap with a page in the ring buffer.
4460 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4463 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4465 * process_page(rpage, ret);
4467 * When @full is set, the function will not return true unless
4468 * the writer is off the reader page.
4470 * Note: it is up to the calling functions to handle sleeps and wakeups.
4471 * The ring buffer can be used anywhere in the kernel and can not
4472 * blindly call wake_up. The layer that uses the ring buffer must be
4473 * responsible for that.
4476 * >=0 if data has been transferred, returns the offset of consumed data.
4477 * <0 if no data has been transferred.
4479 int ring_buffer_read_page(struct ring_buffer *buffer,
4480 void **data_page, size_t len, int cpu, int full)
4482 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4483 struct ring_buffer_event *event;
4484 struct buffer_data_page *bpage;
4485 struct buffer_page *reader;
4486 unsigned long missed_events;
4487 unsigned long flags;
4488 unsigned int commit;
4493 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4497 * If len is not big enough to hold the page header, then
4498 * we can not copy anything.
4500 if (len <= BUF_PAGE_HDR_SIZE)
4503 len -= BUF_PAGE_HDR_SIZE;
4512 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4514 reader = rb_get_reader_page(cpu_buffer);
4518 event = rb_reader_event(cpu_buffer);
4520 read = reader->read;
4521 commit = rb_page_commit(reader);
4523 /* Check if any events were dropped */
4524 missed_events = cpu_buffer->lost_events;
4527 * If this page has been partially read or
4528 * if len is not big enough to read the rest of the page or
4529 * a writer is still on the page, then
4530 * we must copy the data from the page to the buffer.
4531 * Otherwise, we can simply swap the page with the one passed in.
4533 if (read || (len < (commit - read)) ||
4534 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4535 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4536 unsigned int rpos = read;
4537 unsigned int pos = 0;
4543 if (len > (commit - read))
4544 len = (commit - read);
4546 /* Always keep the time extend and data together */
4547 size = rb_event_ts_length(event);
4552 /* save the current timestamp, since the user will need it */
4553 save_timestamp = cpu_buffer->read_stamp;
4555 /* Need to copy one event at a time */
4557 /* We need the size of one event, because
4558 * rb_advance_reader only advances by one event,
4559 * whereas rb_event_ts_length may include the size of
4560 * one or two events.
4561 * We have already ensured there's enough space if this
4562 * is a time extend. */
4563 size = rb_event_length(event);
4564 memcpy(bpage->data + pos, rpage->data + rpos, size);
4568 rb_advance_reader(cpu_buffer);
4569 rpos = reader->read;
4575 event = rb_reader_event(cpu_buffer);
4576 /* Always keep the time extend and data together */
4577 size = rb_event_ts_length(event);
4578 } while (len >= size);
4581 local_set(&bpage->commit, pos);
4582 bpage->time_stamp = save_timestamp;
4584 /* we copied everything to the beginning */
4587 /* update the entry counter */
4588 cpu_buffer->read += rb_page_entries(reader);
4589 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4591 /* swap the pages */
4592 rb_init_page(bpage);
4593 bpage = reader->page;
4594 reader->page = *data_page;
4595 local_set(&reader->write, 0);
4596 local_set(&reader->entries, 0);
4601 * Use the real_end for the data size,
4602 * This gives us a chance to store the lost events
4605 if (reader->real_end)
4606 local_set(&bpage->commit, reader->real_end);
4610 cpu_buffer->lost_events = 0;
4612 commit = local_read(&bpage->commit);
4614 * Set a flag in the commit field if we lost events
4616 if (missed_events) {
4617 /* If there is room at the end of the page to save the
4618 * missed events, then record it there.
4620 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4621 memcpy(&bpage->data[commit], &missed_events,
4622 sizeof(missed_events));
4623 local_add(RB_MISSED_STORED, &bpage->commit);
4624 commit += sizeof(missed_events);
4626 local_add(RB_MISSED_EVENTS, &bpage->commit);
4630 * This page may be off to user land. Zero it out here.
4632 if (commit < BUF_PAGE_SIZE)
4633 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4636 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4641 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4643 #ifdef CONFIG_HOTPLUG_CPU
4644 static int rb_cpu_notify(struct notifier_block *self,
4645 unsigned long action, void *hcpu)
4647 struct ring_buffer *buffer =
4648 container_of(self, struct ring_buffer, cpu_notify);
4649 long cpu = (long)hcpu;
4650 int cpu_i, nr_pages_same;
4651 unsigned int nr_pages;
4654 case CPU_UP_PREPARE:
4655 case CPU_UP_PREPARE_FROZEN:
4656 if (cpumask_test_cpu(cpu, buffer->cpumask))
4661 /* check if all cpu sizes are same */
4662 for_each_buffer_cpu(buffer, cpu_i) {
4663 /* fill in the size from first enabled cpu */
4665 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4666 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4671 /* allocate minimum pages, user can later expand it */
4674 buffer->buffers[cpu] =
4675 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4676 if (!buffer->buffers[cpu]) {
4677 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4682 cpumask_set_cpu(cpu, buffer->cpumask);
4684 case CPU_DOWN_PREPARE:
4685 case CPU_DOWN_PREPARE_FROZEN:
4688 * If we were to free the buffer, then the user would
4689 * lose any trace that was in the buffer.
4699 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4701 * This is a basic integrity check of the ring buffer.
4702 * Late in the boot cycle this test will run when configured in.
4703 * It will kick off a thread per CPU that will go into a loop
4704 * writing to the per cpu ring buffer various sizes of data.
4705 * Some of the data will be large items, some small.
4707 * Another thread is created that goes into a spin, sending out
4708 * IPIs to the other CPUs to also write into the ring buffer.
4709 * this is to test the nesting ability of the buffer.
4711 * Basic stats are recorded and reported. If something in the
4712 * ring buffer should happen that's not expected, a big warning
4713 * is displayed and all ring buffers are disabled.
4715 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4717 struct rb_test_data {
4718 struct ring_buffer *buffer;
4719 unsigned long events;
4720 unsigned long bytes_written;
4721 unsigned long bytes_alloc;
4722 unsigned long bytes_dropped;
4723 unsigned long events_nested;
4724 unsigned long bytes_written_nested;
4725 unsigned long bytes_alloc_nested;
4726 unsigned long bytes_dropped_nested;
4727 int min_size_nested;
4728 int max_size_nested;
4735 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4738 #define RB_TEST_BUFFER_SIZE 1048576
4740 static char rb_string[] __initdata =
4741 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4742 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4743 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4745 static bool rb_test_started __initdata;
4752 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4754 struct ring_buffer_event *event;
4755 struct rb_item *item;
4762 /* Have nested writes different that what is written */
4763 cnt = data->cnt + (nested ? 27 : 0);
4765 /* Multiply cnt by ~e, to make some unique increment */
4766 size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4768 len = size + sizeof(struct rb_item);
4770 started = rb_test_started;
4771 /* read rb_test_started before checking buffer enabled */
4774 event = ring_buffer_lock_reserve(data->buffer, len);
4776 /* Ignore dropped events before test starts. */
4779 data->bytes_dropped += len;
4781 data->bytes_dropped_nested += len;
4786 event_len = ring_buffer_event_length(event);
4788 if (RB_WARN_ON(data->buffer, event_len < len))
4791 item = ring_buffer_event_data(event);
4793 memcpy(item->str, rb_string, size);
4796 data->bytes_alloc_nested += event_len;
4797 data->bytes_written_nested += len;
4798 data->events_nested++;
4799 if (!data->min_size_nested || len < data->min_size_nested)
4800 data->min_size_nested = len;
4801 if (len > data->max_size_nested)
4802 data->max_size_nested = len;
4804 data->bytes_alloc += event_len;
4805 data->bytes_written += len;
4807 if (!data->min_size || len < data->min_size)
4808 data->max_size = len;
4809 if (len > data->max_size)
4810 data->max_size = len;
4814 ring_buffer_unlock_commit(data->buffer, event);
4819 static __init int rb_test(void *arg)
4821 struct rb_test_data *data = arg;
4823 while (!kthread_should_stop()) {
4824 rb_write_something(data, false);
4827 set_current_state(TASK_INTERRUPTIBLE);
4828 /* Now sleep between a min of 100-300us and a max of 1ms */
4829 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4835 static __init void rb_ipi(void *ignore)
4837 struct rb_test_data *data;
4838 int cpu = smp_processor_id();
4840 data = &rb_data[cpu];
4841 rb_write_something(data, true);
4844 static __init int rb_hammer_test(void *arg)
4846 while (!kthread_should_stop()) {
4848 /* Send an IPI to all cpus to write data! */
4849 smp_call_function(rb_ipi, NULL, 1);
4850 /* No sleep, but for non preempt, let others run */
4857 static __init int test_ringbuffer(void)
4859 struct task_struct *rb_hammer;
4860 struct ring_buffer *buffer;
4864 pr_info("Running ring buffer tests...\n");
4866 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4867 if (WARN_ON(!buffer))
4870 /* Disable buffer so that threads can't write to it yet */
4871 ring_buffer_record_off(buffer);
4873 for_each_online_cpu(cpu) {
4874 rb_data[cpu].buffer = buffer;
4875 rb_data[cpu].cpu = cpu;
4876 rb_data[cpu].cnt = cpu;
4877 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4878 "rbtester/%d", cpu);
4879 if (WARN_ON(!rb_threads[cpu])) {
4880 pr_cont("FAILED\n");
4885 kthread_bind(rb_threads[cpu], cpu);
4886 wake_up_process(rb_threads[cpu]);
4889 /* Now create the rb hammer! */
4890 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4891 if (WARN_ON(!rb_hammer)) {
4892 pr_cont("FAILED\n");
4897 ring_buffer_record_on(buffer);
4899 * Show buffer is enabled before setting rb_test_started.
4900 * Yes there's a small race window where events could be
4901 * dropped and the thread wont catch it. But when a ring
4902 * buffer gets enabled, there will always be some kind of
4903 * delay before other CPUs see it. Thus, we don't care about
4904 * those dropped events. We care about events dropped after
4905 * the threads see that the buffer is active.
4908 rb_test_started = true;
4910 set_current_state(TASK_INTERRUPTIBLE);
4911 /* Just run for 10 seconds */;
4912 schedule_timeout(10 * HZ);
4914 kthread_stop(rb_hammer);
4917 for_each_online_cpu(cpu) {
4918 if (!rb_threads[cpu])
4920 kthread_stop(rb_threads[cpu]);
4923 ring_buffer_free(buffer);
4928 pr_info("finished\n");
4929 for_each_online_cpu(cpu) {
4930 struct ring_buffer_event *event;
4931 struct rb_test_data *data = &rb_data[cpu];
4932 struct rb_item *item;
4933 unsigned long total_events;
4934 unsigned long total_dropped;
4935 unsigned long total_written;
4936 unsigned long total_alloc;
4937 unsigned long total_read = 0;
4938 unsigned long total_size = 0;
4939 unsigned long total_len = 0;
4940 unsigned long total_lost = 0;
4943 int small_event_size;
4947 total_events = data->events + data->events_nested;
4948 total_written = data->bytes_written + data->bytes_written_nested;
4949 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
4950 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
4952 big_event_size = data->max_size + data->max_size_nested;
4953 small_event_size = data->min_size + data->min_size_nested;
4955 pr_info("CPU %d:\n", cpu);
4956 pr_info(" events: %ld\n", total_events);
4957 pr_info(" dropped bytes: %ld\n", total_dropped);
4958 pr_info(" alloced bytes: %ld\n", total_alloc);
4959 pr_info(" written bytes: %ld\n", total_written);
4960 pr_info(" biggest event: %d\n", big_event_size);
4961 pr_info(" smallest event: %d\n", small_event_size);
4963 if (RB_WARN_ON(buffer, total_dropped))
4968 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
4970 item = ring_buffer_event_data(event);
4971 total_len += ring_buffer_event_length(event);
4972 total_size += item->size + sizeof(struct rb_item);
4973 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
4974 pr_info("FAILED!\n");
4975 pr_info("buffer had: %.*s\n", item->size, item->str);
4976 pr_info("expected: %.*s\n", item->size, rb_string);
4977 RB_WARN_ON(buffer, 1);
4988 pr_info(" read events: %ld\n", total_read);
4989 pr_info(" lost events: %ld\n", total_lost);
4990 pr_info(" total events: %ld\n", total_lost + total_read);
4991 pr_info(" recorded len bytes: %ld\n", total_len);
4992 pr_info(" recorded size bytes: %ld\n", total_size);
4994 pr_info(" With dropped events, record len and size may not match\n"
4995 " alloced and written from above\n");
4997 if (RB_WARN_ON(buffer, total_len != total_alloc ||
4998 total_size != total_written))
5001 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5007 pr_info("Ring buffer PASSED!\n");
5009 ring_buffer_free(buffer);
5013 late_initcall(test_ringbuffer);
5014 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */