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.
118 /* Used for individual buffers (after the counter) */
119 #define RB_BUFFER_OFF (1 << 20)
121 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
123 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
124 #define RB_ALIGNMENT 4U
125 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
126 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
128 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
129 # define RB_FORCE_8BYTE_ALIGNMENT 0
130 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
132 # define RB_FORCE_8BYTE_ALIGNMENT 1
133 # define RB_ARCH_ALIGNMENT 8U
136 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
138 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
139 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
142 RB_LEN_TIME_EXTEND = 8,
143 RB_LEN_TIME_STAMP = 16,
146 #define skip_time_extend(event) \
147 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
149 static inline int rb_null_event(struct ring_buffer_event *event)
151 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
154 static void rb_event_set_padding(struct ring_buffer_event *event)
156 /* padding has a NULL time_delta */
157 event->type_len = RINGBUF_TYPE_PADDING;
158 event->time_delta = 0;
162 rb_event_data_length(struct ring_buffer_event *event)
167 length = event->type_len * RB_ALIGNMENT;
169 length = event->array[0];
170 return length + RB_EVNT_HDR_SIZE;
174 * Return the length of the given event. Will return
175 * the length of the time extend if the event is a
178 static inline unsigned
179 rb_event_length(struct ring_buffer_event *event)
181 switch (event->type_len) {
182 case RINGBUF_TYPE_PADDING:
183 if (rb_null_event(event))
186 return event->array[0] + RB_EVNT_HDR_SIZE;
188 case RINGBUF_TYPE_TIME_EXTEND:
189 return RB_LEN_TIME_EXTEND;
191 case RINGBUF_TYPE_TIME_STAMP:
192 return RB_LEN_TIME_STAMP;
194 case RINGBUF_TYPE_DATA:
195 return rb_event_data_length(event);
204 * Return total length of time extend and data,
205 * or just the event length for all other events.
207 static inline unsigned
208 rb_event_ts_length(struct ring_buffer_event *event)
212 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
213 /* time extends include the data event after it */
214 len = RB_LEN_TIME_EXTEND;
215 event = skip_time_extend(event);
217 return len + rb_event_length(event);
221 * ring_buffer_event_length - return the length of the event
222 * @event: the event to get the length of
224 * Returns the size of the data load of a data event.
225 * If the event is something other than a data event, it
226 * returns the size of the event itself. With the exception
227 * of a TIME EXTEND, where it still returns the size of the
228 * data load of the data event after it.
230 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
234 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
235 event = skip_time_extend(event);
237 length = rb_event_length(event);
238 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
240 length -= RB_EVNT_HDR_SIZE;
241 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
242 length -= sizeof(event->array[0]);
245 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
247 /* inline for ring buffer fast paths */
249 rb_event_data(struct ring_buffer_event *event)
251 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
252 event = skip_time_extend(event);
253 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
254 /* If length is in len field, then array[0] has the data */
256 return (void *)&event->array[0];
257 /* Otherwise length is in array[0] and array[1] has the data */
258 return (void *)&event->array[1];
262 * ring_buffer_event_data - return the data of the event
263 * @event: the event to get the data from
265 void *ring_buffer_event_data(struct ring_buffer_event *event)
267 return rb_event_data(event);
269 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
271 #define for_each_buffer_cpu(buffer, cpu) \
272 for_each_cpu(cpu, buffer->cpumask)
275 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
276 #define TS_DELTA_TEST (~TS_MASK)
278 /* Flag when events were overwritten */
279 #define RB_MISSED_EVENTS (1 << 31)
280 /* Missed count stored at end */
281 #define RB_MISSED_STORED (1 << 30)
283 struct buffer_data_page {
284 u64 time_stamp; /* page time stamp */
285 local_t commit; /* write committed index */
286 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
290 * Note, the buffer_page list must be first. The buffer pages
291 * are allocated in cache lines, which means that each buffer
292 * page will be at the beginning of a cache line, and thus
293 * the least significant bits will be zero. We use this to
294 * add flags in the list struct pointers, to make the ring buffer
298 struct list_head list; /* list of buffer pages */
299 local_t write; /* index for next write */
300 unsigned read; /* index for next read */
301 local_t entries; /* entries on this page */
302 unsigned long real_end; /* real end of data */
303 struct buffer_data_page *page; /* Actual data page */
307 * The buffer page counters, write and entries, must be reset
308 * atomically when crossing page boundaries. To synchronize this
309 * update, two counters are inserted into the number. One is
310 * the actual counter for the write position or count on the page.
312 * The other is a counter of updaters. Before an update happens
313 * the update partition of the counter is incremented. This will
314 * allow the updater to update the counter atomically.
316 * The counter is 20 bits, and the state data is 12.
318 #define RB_WRITE_MASK 0xfffff
319 #define RB_WRITE_INTCNT (1 << 20)
321 static void rb_init_page(struct buffer_data_page *bpage)
323 local_set(&bpage->commit, 0);
327 * ring_buffer_page_len - the size of data on the page.
328 * @page: The page to read
330 * Returns the amount of data on the page, including buffer page header.
332 size_t ring_buffer_page_len(void *page)
334 return local_read(&((struct buffer_data_page *)page)->commit)
339 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
342 static void free_buffer_page(struct buffer_page *bpage)
344 free_page((unsigned long)bpage->page);
349 * We need to fit the time_stamp delta into 27 bits.
351 static inline int test_time_stamp(u64 delta)
353 if (delta & TS_DELTA_TEST)
358 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
360 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
361 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
363 int ring_buffer_print_page_header(struct trace_seq *s)
365 struct buffer_data_page field;
367 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
368 "offset:0;\tsize:%u;\tsigned:%u;\n",
369 (unsigned int)sizeof(field.time_stamp),
370 (unsigned int)is_signed_type(u64));
372 trace_seq_printf(s, "\tfield: local_t commit;\t"
373 "offset:%u;\tsize:%u;\tsigned:%u;\n",
374 (unsigned int)offsetof(typeof(field), commit),
375 (unsigned int)sizeof(field.commit),
376 (unsigned int)is_signed_type(long));
378 trace_seq_printf(s, "\tfield: int overwrite;\t"
379 "offset:%u;\tsize:%u;\tsigned:%u;\n",
380 (unsigned int)offsetof(typeof(field), commit),
382 (unsigned int)is_signed_type(long));
384 trace_seq_printf(s, "\tfield: char data;\t"
385 "offset:%u;\tsize:%u;\tsigned:%u;\n",
386 (unsigned int)offsetof(typeof(field), data),
387 (unsigned int)BUF_PAGE_SIZE,
388 (unsigned int)is_signed_type(char));
390 return !trace_seq_has_overflowed(s);
394 struct irq_work work;
395 wait_queue_head_t waiters;
396 wait_queue_head_t full_waiters;
397 bool waiters_pending;
398 bool full_waiters_pending;
403 * Structure to hold event state and handle nested events.
405 struct rb_event_info {
408 unsigned long length;
409 struct buffer_page *tail_page;
414 * Used for which event context the event is in.
420 * See trace_recursive_lock() comment below for more details.
431 * head_page == tail_page && head == tail then buffer is empty.
433 struct ring_buffer_per_cpu {
435 atomic_t record_disabled;
436 struct ring_buffer *buffer;
437 raw_spinlock_t reader_lock; /* serialize readers */
438 arch_spinlock_t lock;
439 struct lock_class_key lock_key;
440 unsigned long nr_pages;
441 unsigned int current_context;
442 struct list_head *pages;
443 struct buffer_page *head_page; /* read from head */
444 struct buffer_page *tail_page; /* write to tail */
445 struct buffer_page *commit_page; /* committed pages */
446 struct buffer_page *reader_page;
447 unsigned long lost_events;
448 unsigned long last_overrun;
449 local_t entries_bytes;
452 local_t commit_overrun;
453 local_t dropped_events;
457 unsigned long read_bytes;
460 /* ring buffer pages to update, > 0 to add, < 0 to remove */
461 long nr_pages_to_update;
462 struct list_head new_pages; /* new pages to add */
463 struct work_struct update_pages_work;
464 struct completion update_done;
466 struct rb_irq_work irq_work;
472 atomic_t record_disabled;
473 atomic_t resize_disabled;
474 cpumask_var_t cpumask;
476 struct lock_class_key *reader_lock_key;
480 struct ring_buffer_per_cpu **buffers;
482 #ifdef CONFIG_HOTPLUG_CPU
483 struct notifier_block cpu_notify;
487 struct rb_irq_work irq_work;
490 struct ring_buffer_iter {
491 struct ring_buffer_per_cpu *cpu_buffer;
493 struct buffer_page *head_page;
494 struct buffer_page *cache_reader_page;
495 unsigned long cache_read;
500 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
502 * Schedules a delayed work to wake up any task that is blocked on the
503 * ring buffer waiters queue.
505 static void rb_wake_up_waiters(struct irq_work *work)
507 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
509 wake_up_all(&rbwork->waiters);
510 if (rbwork->wakeup_full) {
511 rbwork->wakeup_full = false;
512 wake_up_all(&rbwork->full_waiters);
517 * ring_buffer_wait - wait for input to the ring buffer
518 * @buffer: buffer to wait on
519 * @cpu: the cpu buffer to wait on
520 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
522 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
523 * as data is added to any of the @buffer's cpu buffers. Otherwise
524 * it will wait for data to be added to a specific cpu buffer.
526 int ring_buffer_wait(struct ring_buffer *buffer, int cpu, bool full)
528 struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
530 struct rb_irq_work *work;
534 * Depending on what the caller is waiting for, either any
535 * data in any cpu buffer, or a specific buffer, put the
536 * caller on the appropriate wait queue.
538 if (cpu == RING_BUFFER_ALL_CPUS) {
539 work = &buffer->irq_work;
540 /* Full only makes sense on per cpu reads */
543 if (!cpumask_test_cpu(cpu, buffer->cpumask))
545 cpu_buffer = buffer->buffers[cpu];
546 work = &cpu_buffer->irq_work;
552 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
554 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
557 * The events can happen in critical sections where
558 * checking a work queue can cause deadlocks.
559 * After adding a task to the queue, this flag is set
560 * only to notify events to try to wake up the queue
563 * We don't clear it even if the buffer is no longer
564 * empty. The flag only causes the next event to run
565 * irq_work to do the work queue wake up. The worse
566 * that can happen if we race with !trace_empty() is that
567 * an event will cause an irq_work to try to wake up
570 * There's no reason to protect this flag either, as
571 * the work queue and irq_work logic will do the necessary
572 * synchronization for the wake ups. The only thing
573 * that is necessary is that the wake up happens after
574 * a task has been queued. It's OK for spurious wake ups.
577 work->full_waiters_pending = true;
579 work->waiters_pending = true;
581 if (signal_pending(current)) {
586 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
589 if (cpu != RING_BUFFER_ALL_CPUS &&
590 !ring_buffer_empty_cpu(buffer, cpu)) {
597 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
598 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
599 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
609 finish_wait(&work->full_waiters, &wait);
611 finish_wait(&work->waiters, &wait);
617 * ring_buffer_poll_wait - poll on buffer input
618 * @buffer: buffer to wait on
619 * @cpu: the cpu buffer to wait on
620 * @filp: the file descriptor
621 * @poll_table: The poll descriptor
623 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
624 * as data is added to any of the @buffer's cpu buffers. Otherwise
625 * it will wait for data to be added to a specific cpu buffer.
627 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
630 int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
631 struct file *filp, poll_table *poll_table)
633 struct ring_buffer_per_cpu *cpu_buffer;
634 struct rb_irq_work *work;
636 if (cpu == RING_BUFFER_ALL_CPUS)
637 work = &buffer->irq_work;
639 if (!cpumask_test_cpu(cpu, buffer->cpumask))
642 cpu_buffer = buffer->buffers[cpu];
643 work = &cpu_buffer->irq_work;
646 poll_wait(filp, &work->waiters, poll_table);
647 work->waiters_pending = true;
649 * There's a tight race between setting the waiters_pending and
650 * checking if the ring buffer is empty. Once the waiters_pending bit
651 * is set, the next event will wake the task up, but we can get stuck
652 * if there's only a single event in.
654 * FIXME: Ideally, we need a memory barrier on the writer side as well,
655 * but adding a memory barrier to all events will cause too much of a
656 * performance hit in the fast path. We only need a memory barrier when
657 * the buffer goes from empty to having content. But as this race is
658 * extremely small, and it's not a problem if another event comes in, we
663 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
664 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
665 return POLLIN | POLLRDNORM;
669 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
670 #define RB_WARN_ON(b, cond) \
672 int _____ret = unlikely(cond); \
674 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
675 struct ring_buffer_per_cpu *__b = \
677 atomic_inc(&__b->buffer->record_disabled); \
679 atomic_inc(&b->record_disabled); \
685 /* Up this if you want to test the TIME_EXTENTS and normalization */
686 #define DEBUG_SHIFT 0
688 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
690 /* shift to debug/test normalization and TIME_EXTENTS */
691 return buffer->clock() << DEBUG_SHIFT;
694 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
698 preempt_disable_notrace();
699 time = rb_time_stamp(buffer);
700 preempt_enable_no_resched_notrace();
704 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
706 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
709 /* Just stupid testing the normalize function and deltas */
712 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
715 * Making the ring buffer lockless makes things tricky.
716 * Although writes only happen on the CPU that they are on,
717 * and they only need to worry about interrupts. Reads can
720 * The reader page is always off the ring buffer, but when the
721 * reader finishes with a page, it needs to swap its page with
722 * a new one from the buffer. The reader needs to take from
723 * the head (writes go to the tail). But if a writer is in overwrite
724 * mode and wraps, it must push the head page forward.
726 * Here lies the problem.
728 * The reader must be careful to replace only the head page, and
729 * not another one. As described at the top of the file in the
730 * ASCII art, the reader sets its old page to point to the next
731 * page after head. It then sets the page after head to point to
732 * the old reader page. But if the writer moves the head page
733 * during this operation, the reader could end up with the tail.
735 * We use cmpxchg to help prevent this race. We also do something
736 * special with the page before head. We set the LSB to 1.
738 * When the writer must push the page forward, it will clear the
739 * bit that points to the head page, move the head, and then set
740 * the bit that points to the new head page.
742 * We also don't want an interrupt coming in and moving the head
743 * page on another writer. Thus we use the second LSB to catch
746 * head->list->prev->next bit 1 bit 0
749 * Points to head page 0 1
752 * Note we can not trust the prev pointer of the head page, because:
754 * +----+ +-----+ +-----+
755 * | |------>| T |---X--->| N |
757 * +----+ +-----+ +-----+
760 * +----------| R |----------+ |
764 * Key: ---X--> HEAD flag set in pointer
769 * (see __rb_reserve_next() to see where this happens)
771 * What the above shows is that the reader just swapped out
772 * the reader page with a page in the buffer, but before it
773 * could make the new header point back to the new page added
774 * it was preempted by a writer. The writer moved forward onto
775 * the new page added by the reader and is about to move forward
778 * You can see, it is legitimate for the previous pointer of
779 * the head (or any page) not to point back to itself. But only
783 #define RB_PAGE_NORMAL 0UL
784 #define RB_PAGE_HEAD 1UL
785 #define RB_PAGE_UPDATE 2UL
788 #define RB_FLAG_MASK 3UL
790 /* PAGE_MOVED is not part of the mask */
791 #define RB_PAGE_MOVED 4UL
794 * rb_list_head - remove any bit
796 static struct list_head *rb_list_head(struct list_head *list)
798 unsigned long val = (unsigned long)list;
800 return (struct list_head *)(val & ~RB_FLAG_MASK);
804 * rb_is_head_page - test if the given page is the head page
806 * Because the reader may move the head_page pointer, we can
807 * not trust what the head page is (it may be pointing to
808 * the reader page). But if the next page is a header page,
809 * its flags will be non zero.
812 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
813 struct buffer_page *page, struct list_head *list)
817 val = (unsigned long)list->next;
819 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
820 return RB_PAGE_MOVED;
822 return val & RB_FLAG_MASK;
828 * The unique thing about the reader page, is that, if the
829 * writer is ever on it, the previous pointer never points
830 * back to the reader page.
832 static bool rb_is_reader_page(struct buffer_page *page)
834 struct list_head *list = page->list.prev;
836 return rb_list_head(list->next) != &page->list;
840 * rb_set_list_to_head - set a list_head to be pointing to head.
842 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
843 struct list_head *list)
847 ptr = (unsigned long *)&list->next;
848 *ptr |= RB_PAGE_HEAD;
849 *ptr &= ~RB_PAGE_UPDATE;
853 * rb_head_page_activate - sets up head page
855 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
857 struct buffer_page *head;
859 head = cpu_buffer->head_page;
864 * Set the previous list pointer to have the HEAD flag.
866 rb_set_list_to_head(cpu_buffer, head->list.prev);
869 static void rb_list_head_clear(struct list_head *list)
871 unsigned long *ptr = (unsigned long *)&list->next;
873 *ptr &= ~RB_FLAG_MASK;
877 * rb_head_page_dactivate - clears head page ptr (for free list)
880 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
882 struct list_head *hd;
884 /* Go through the whole list and clear any pointers found. */
885 rb_list_head_clear(cpu_buffer->pages);
887 list_for_each(hd, cpu_buffer->pages)
888 rb_list_head_clear(hd);
891 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
892 struct buffer_page *head,
893 struct buffer_page *prev,
894 int old_flag, int new_flag)
896 struct list_head *list;
897 unsigned long val = (unsigned long)&head->list;
902 val &= ~RB_FLAG_MASK;
904 ret = cmpxchg((unsigned long *)&list->next,
905 val | old_flag, val | new_flag);
907 /* check if the reader took the page */
908 if ((ret & ~RB_FLAG_MASK) != val)
909 return RB_PAGE_MOVED;
911 return ret & RB_FLAG_MASK;
914 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
915 struct buffer_page *head,
916 struct buffer_page *prev,
919 return rb_head_page_set(cpu_buffer, head, prev,
920 old_flag, RB_PAGE_UPDATE);
923 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
924 struct buffer_page *head,
925 struct buffer_page *prev,
928 return rb_head_page_set(cpu_buffer, head, prev,
929 old_flag, RB_PAGE_HEAD);
932 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
933 struct buffer_page *head,
934 struct buffer_page *prev,
937 return rb_head_page_set(cpu_buffer, head, prev,
938 old_flag, RB_PAGE_NORMAL);
941 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
942 struct buffer_page **bpage)
944 struct list_head *p = rb_list_head((*bpage)->list.next);
946 *bpage = list_entry(p, struct buffer_page, list);
949 static struct buffer_page *
950 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
952 struct buffer_page *head;
953 struct buffer_page *page;
954 struct list_head *list;
957 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
961 list = cpu_buffer->pages;
962 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
965 page = head = cpu_buffer->head_page;
967 * It is possible that the writer moves the header behind
968 * where we started, and we miss in one loop.
969 * A second loop should grab the header, but we'll do
970 * three loops just because I'm paranoid.
972 for (i = 0; i < 3; i++) {
974 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
975 cpu_buffer->head_page = page;
978 rb_inc_page(cpu_buffer, &page);
979 } while (page != head);
982 RB_WARN_ON(cpu_buffer, 1);
987 static int rb_head_page_replace(struct buffer_page *old,
988 struct buffer_page *new)
990 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
994 val = *ptr & ~RB_FLAG_MASK;
997 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1003 * rb_tail_page_update - move the tail page forward
1005 * Returns 1 if moved tail page, 0 if someone else did.
1007 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1008 struct buffer_page *tail_page,
1009 struct buffer_page *next_page)
1011 struct buffer_page *old_tail;
1012 unsigned long old_entries;
1013 unsigned long old_write;
1017 * The tail page now needs to be moved forward.
1019 * We need to reset the tail page, but without messing
1020 * with possible erasing of data brought in by interrupts
1021 * that have moved the tail page and are currently on it.
1023 * We add a counter to the write field to denote this.
1025 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1026 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1029 * Just make sure we have seen our old_write and synchronize
1030 * with any interrupts that come in.
1035 * If the tail page is still the same as what we think
1036 * it is, then it is up to us to update the tail
1039 if (tail_page == cpu_buffer->tail_page) {
1040 /* Zero the write counter */
1041 unsigned long val = old_write & ~RB_WRITE_MASK;
1042 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1045 * This will only succeed if an interrupt did
1046 * not come in and change it. In which case, we
1047 * do not want to modify it.
1049 * We add (void) to let the compiler know that we do not care
1050 * about the return value of these functions. We use the
1051 * cmpxchg to only update if an interrupt did not already
1052 * do it for us. If the cmpxchg fails, we don't care.
1054 (void)local_cmpxchg(&next_page->write, old_write, val);
1055 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1058 * No need to worry about races with clearing out the commit.
1059 * it only can increment when a commit takes place. But that
1060 * only happens in the outer most nested commit.
1062 local_set(&next_page->page->commit, 0);
1064 old_tail = cmpxchg(&cpu_buffer->tail_page,
1065 tail_page, next_page);
1067 if (old_tail == tail_page)
1074 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1075 struct buffer_page *bpage)
1077 unsigned long val = (unsigned long)bpage;
1079 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1086 * rb_check_list - make sure a pointer to a list has the last bits zero
1088 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1089 struct list_head *list)
1091 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1093 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1099 * rb_check_pages - integrity check of buffer pages
1100 * @cpu_buffer: CPU buffer with pages to test
1102 * As a safety measure we check to make sure the data pages have not
1105 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1107 struct list_head *head = cpu_buffer->pages;
1108 struct buffer_page *bpage, *tmp;
1110 /* Reset the head page if it exists */
1111 if (cpu_buffer->head_page)
1112 rb_set_head_page(cpu_buffer);
1114 rb_head_page_deactivate(cpu_buffer);
1116 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1118 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1121 if (rb_check_list(cpu_buffer, head))
1124 list_for_each_entry_safe(bpage, tmp, head, list) {
1125 if (RB_WARN_ON(cpu_buffer,
1126 bpage->list.next->prev != &bpage->list))
1128 if (RB_WARN_ON(cpu_buffer,
1129 bpage->list.prev->next != &bpage->list))
1131 if (rb_check_list(cpu_buffer, &bpage->list))
1135 rb_head_page_activate(cpu_buffer);
1140 static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1142 struct buffer_page *bpage, *tmp;
1145 for (i = 0; i < nr_pages; i++) {
1148 * __GFP_NORETRY flag makes sure that the allocation fails
1149 * gracefully without invoking oom-killer and the system is
1152 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1153 GFP_KERNEL | __GFP_NORETRY,
1158 list_add(&bpage->list, pages);
1160 page = alloc_pages_node(cpu_to_node(cpu),
1161 GFP_KERNEL | __GFP_NORETRY, 0);
1164 bpage->page = page_address(page);
1165 rb_init_page(bpage->page);
1171 list_for_each_entry_safe(bpage, tmp, pages, list) {
1172 list_del_init(&bpage->list);
1173 free_buffer_page(bpage);
1179 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1180 unsigned long nr_pages)
1186 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1190 * The ring buffer page list is a circular list that does not
1191 * start and end with a list head. All page list items point to
1194 cpu_buffer->pages = pages.next;
1197 cpu_buffer->nr_pages = nr_pages;
1199 rb_check_pages(cpu_buffer);
1204 static struct ring_buffer_per_cpu *
1205 rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1207 struct ring_buffer_per_cpu *cpu_buffer;
1208 struct buffer_page *bpage;
1212 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1213 GFP_KERNEL, cpu_to_node(cpu));
1217 cpu_buffer->cpu = cpu;
1218 cpu_buffer->buffer = buffer;
1219 raw_spin_lock_init(&cpu_buffer->reader_lock);
1220 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1221 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1222 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1223 init_completion(&cpu_buffer->update_done);
1224 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1225 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1226 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1228 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1229 GFP_KERNEL, cpu_to_node(cpu));
1231 goto fail_free_buffer;
1233 rb_check_bpage(cpu_buffer, bpage);
1235 cpu_buffer->reader_page = bpage;
1236 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1238 goto fail_free_reader;
1239 bpage->page = page_address(page);
1240 rb_init_page(bpage->page);
1242 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1243 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1245 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1247 goto fail_free_reader;
1249 cpu_buffer->head_page
1250 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1251 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1253 rb_head_page_activate(cpu_buffer);
1258 free_buffer_page(cpu_buffer->reader_page);
1265 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1267 struct list_head *head = cpu_buffer->pages;
1268 struct buffer_page *bpage, *tmp;
1270 free_buffer_page(cpu_buffer->reader_page);
1272 rb_head_page_deactivate(cpu_buffer);
1275 list_for_each_entry_safe(bpage, tmp, head, list) {
1276 list_del_init(&bpage->list);
1277 free_buffer_page(bpage);
1279 bpage = list_entry(head, struct buffer_page, list);
1280 free_buffer_page(bpage);
1286 #ifdef CONFIG_HOTPLUG_CPU
1287 static int rb_cpu_notify(struct notifier_block *self,
1288 unsigned long action, void *hcpu);
1292 * __ring_buffer_alloc - allocate a new ring_buffer
1293 * @size: the size in bytes per cpu that is needed.
1294 * @flags: attributes to set for the ring buffer.
1296 * Currently the only flag that is available is the RB_FL_OVERWRITE
1297 * flag. This flag means that the buffer will overwrite old data
1298 * when the buffer wraps. If this flag is not set, the buffer will
1299 * drop data when the tail hits the head.
1301 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1302 struct lock_class_key *key)
1304 struct ring_buffer *buffer;
1309 /* keep it in its own cache line */
1310 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1315 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1316 goto fail_free_buffer;
1318 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1319 buffer->flags = flags;
1320 buffer->clock = trace_clock_local;
1321 buffer->reader_lock_key = key;
1323 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1324 init_waitqueue_head(&buffer->irq_work.waiters);
1326 /* need at least two pages */
1331 * In case of non-hotplug cpu, if the ring-buffer is allocated
1332 * in early initcall, it will not be notified of secondary cpus.
1333 * In that off case, we need to allocate for all possible cpus.
1335 #ifdef CONFIG_HOTPLUG_CPU
1336 cpu_notifier_register_begin();
1337 cpumask_copy(buffer->cpumask, cpu_online_mask);
1339 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1341 buffer->cpus = nr_cpu_ids;
1343 bsize = sizeof(void *) * nr_cpu_ids;
1344 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1346 if (!buffer->buffers)
1347 goto fail_free_cpumask;
1349 for_each_buffer_cpu(buffer, cpu) {
1350 buffer->buffers[cpu] =
1351 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1352 if (!buffer->buffers[cpu])
1353 goto fail_free_buffers;
1356 #ifdef CONFIG_HOTPLUG_CPU
1357 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1358 buffer->cpu_notify.priority = 0;
1359 __register_cpu_notifier(&buffer->cpu_notify);
1360 cpu_notifier_register_done();
1363 mutex_init(&buffer->mutex);
1368 for_each_buffer_cpu(buffer, cpu) {
1369 if (buffer->buffers[cpu])
1370 rb_free_cpu_buffer(buffer->buffers[cpu]);
1372 kfree(buffer->buffers);
1375 free_cpumask_var(buffer->cpumask);
1376 #ifdef CONFIG_HOTPLUG_CPU
1377 cpu_notifier_register_done();
1384 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1387 * ring_buffer_free - free a ring buffer.
1388 * @buffer: the buffer to free.
1391 ring_buffer_free(struct ring_buffer *buffer)
1395 #ifdef CONFIG_HOTPLUG_CPU
1396 cpu_notifier_register_begin();
1397 __unregister_cpu_notifier(&buffer->cpu_notify);
1400 for_each_buffer_cpu(buffer, cpu)
1401 rb_free_cpu_buffer(buffer->buffers[cpu]);
1403 #ifdef CONFIG_HOTPLUG_CPU
1404 cpu_notifier_register_done();
1407 kfree(buffer->buffers);
1408 free_cpumask_var(buffer->cpumask);
1412 EXPORT_SYMBOL_GPL(ring_buffer_free);
1414 void ring_buffer_set_clock(struct ring_buffer *buffer,
1417 buffer->clock = clock;
1420 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1422 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1424 return local_read(&bpage->entries) & RB_WRITE_MASK;
1427 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1429 return local_read(&bpage->write) & RB_WRITE_MASK;
1433 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1435 struct list_head *tail_page, *to_remove, *next_page;
1436 struct buffer_page *to_remove_page, *tmp_iter_page;
1437 struct buffer_page *last_page, *first_page;
1438 unsigned long nr_removed;
1439 unsigned long head_bit;
1444 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1445 atomic_inc(&cpu_buffer->record_disabled);
1447 * We don't race with the readers since we have acquired the reader
1448 * lock. We also don't race with writers after disabling recording.
1449 * This makes it easy to figure out the first and the last page to be
1450 * removed from the list. We unlink all the pages in between including
1451 * the first and last pages. This is done in a busy loop so that we
1452 * lose the least number of traces.
1453 * The pages are freed after we restart recording and unlock readers.
1455 tail_page = &cpu_buffer->tail_page->list;
1458 * tail page might be on reader page, we remove the next page
1459 * from the ring buffer
1461 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1462 tail_page = rb_list_head(tail_page->next);
1463 to_remove = tail_page;
1465 /* start of pages to remove */
1466 first_page = list_entry(rb_list_head(to_remove->next),
1467 struct buffer_page, list);
1469 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1470 to_remove = rb_list_head(to_remove)->next;
1471 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1474 next_page = rb_list_head(to_remove)->next;
1477 * Now we remove all pages between tail_page and next_page.
1478 * Make sure that we have head_bit value preserved for the
1481 tail_page->next = (struct list_head *)((unsigned long)next_page |
1483 next_page = rb_list_head(next_page);
1484 next_page->prev = tail_page;
1486 /* make sure pages points to a valid page in the ring buffer */
1487 cpu_buffer->pages = next_page;
1489 /* update head page */
1491 cpu_buffer->head_page = list_entry(next_page,
1492 struct buffer_page, list);
1495 * change read pointer to make sure any read iterators reset
1498 cpu_buffer->read = 0;
1500 /* pages are removed, resume tracing and then free the pages */
1501 atomic_dec(&cpu_buffer->record_disabled);
1502 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1504 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1506 /* last buffer page to remove */
1507 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1509 tmp_iter_page = first_page;
1512 to_remove_page = tmp_iter_page;
1513 rb_inc_page(cpu_buffer, &tmp_iter_page);
1515 /* update the counters */
1516 page_entries = rb_page_entries(to_remove_page);
1519 * If something was added to this page, it was full
1520 * since it is not the tail page. So we deduct the
1521 * bytes consumed in ring buffer from here.
1522 * Increment overrun to account for the lost events.
1524 local_add(page_entries, &cpu_buffer->overrun);
1525 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1529 * We have already removed references to this list item, just
1530 * free up the buffer_page and its page
1532 free_buffer_page(to_remove_page);
1535 } while (to_remove_page != last_page);
1537 RB_WARN_ON(cpu_buffer, nr_removed);
1539 return nr_removed == 0;
1543 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1545 struct list_head *pages = &cpu_buffer->new_pages;
1546 int retries, success;
1548 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1550 * We are holding the reader lock, so the reader page won't be swapped
1551 * in the ring buffer. Now we are racing with the writer trying to
1552 * move head page and the tail page.
1553 * We are going to adapt the reader page update process where:
1554 * 1. We first splice the start and end of list of new pages between
1555 * the head page and its previous page.
1556 * 2. We cmpxchg the prev_page->next to point from head page to the
1557 * start of new pages list.
1558 * 3. Finally, we update the head->prev to the end of new list.
1560 * We will try this process 10 times, to make sure that we don't keep
1566 struct list_head *head_page, *prev_page, *r;
1567 struct list_head *last_page, *first_page;
1568 struct list_head *head_page_with_bit;
1570 head_page = &rb_set_head_page(cpu_buffer)->list;
1573 prev_page = head_page->prev;
1575 first_page = pages->next;
1576 last_page = pages->prev;
1578 head_page_with_bit = (struct list_head *)
1579 ((unsigned long)head_page | RB_PAGE_HEAD);
1581 last_page->next = head_page_with_bit;
1582 first_page->prev = prev_page;
1584 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1586 if (r == head_page_with_bit) {
1588 * yay, we replaced the page pointer to our new list,
1589 * now, we just have to update to head page's prev
1590 * pointer to point to end of list
1592 head_page->prev = last_page;
1599 INIT_LIST_HEAD(pages);
1601 * If we weren't successful in adding in new pages, warn and stop
1604 RB_WARN_ON(cpu_buffer, !success);
1605 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1607 /* free pages if they weren't inserted */
1609 struct buffer_page *bpage, *tmp;
1610 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1612 list_del_init(&bpage->list);
1613 free_buffer_page(bpage);
1619 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1623 if (cpu_buffer->nr_pages_to_update > 0)
1624 success = rb_insert_pages(cpu_buffer);
1626 success = rb_remove_pages(cpu_buffer,
1627 -cpu_buffer->nr_pages_to_update);
1630 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1633 static void update_pages_handler(struct work_struct *work)
1635 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1636 struct ring_buffer_per_cpu, update_pages_work);
1637 rb_update_pages(cpu_buffer);
1638 complete(&cpu_buffer->update_done);
1642 * ring_buffer_resize - resize the ring buffer
1643 * @buffer: the buffer to resize.
1644 * @size: the new size.
1645 * @cpu_id: the cpu buffer to resize
1647 * Minimum size is 2 * BUF_PAGE_SIZE.
1649 * Returns 0 on success and < 0 on failure.
1651 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1654 struct ring_buffer_per_cpu *cpu_buffer;
1655 unsigned long nr_pages;
1659 * Always succeed at resizing a non-existent buffer:
1664 /* Make sure the requested buffer exists */
1665 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1666 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1669 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1671 /* we need a minimum of two pages */
1675 size = nr_pages * BUF_PAGE_SIZE;
1678 * Don't succeed if resizing is disabled, as a reader might be
1679 * manipulating the ring buffer and is expecting a sane state while
1682 if (atomic_read(&buffer->resize_disabled))
1685 /* prevent another thread from changing buffer sizes */
1686 mutex_lock(&buffer->mutex);
1688 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1689 /* calculate the pages to update */
1690 for_each_buffer_cpu(buffer, cpu) {
1691 cpu_buffer = buffer->buffers[cpu];
1693 cpu_buffer->nr_pages_to_update = nr_pages -
1694 cpu_buffer->nr_pages;
1696 * nothing more to do for removing pages or no update
1698 if (cpu_buffer->nr_pages_to_update <= 0)
1701 * to add pages, make sure all new pages can be
1702 * allocated without receiving ENOMEM
1704 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1705 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1706 &cpu_buffer->new_pages, cpu)) {
1707 /* not enough memory for new pages */
1715 * Fire off all the required work handlers
1716 * We can't schedule on offline CPUs, but it's not necessary
1717 * since we can change their buffer sizes without any race.
1719 for_each_buffer_cpu(buffer, cpu) {
1720 cpu_buffer = buffer->buffers[cpu];
1721 if (!cpu_buffer->nr_pages_to_update)
1724 /* Can't run something on an offline CPU. */
1725 if (!cpu_online(cpu)) {
1726 rb_update_pages(cpu_buffer);
1727 cpu_buffer->nr_pages_to_update = 0;
1729 schedule_work_on(cpu,
1730 &cpu_buffer->update_pages_work);
1734 /* wait for all the updates to complete */
1735 for_each_buffer_cpu(buffer, cpu) {
1736 cpu_buffer = buffer->buffers[cpu];
1737 if (!cpu_buffer->nr_pages_to_update)
1740 if (cpu_online(cpu))
1741 wait_for_completion(&cpu_buffer->update_done);
1742 cpu_buffer->nr_pages_to_update = 0;
1747 /* Make sure this CPU has been intitialized */
1748 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1751 cpu_buffer = buffer->buffers[cpu_id];
1753 if (nr_pages == cpu_buffer->nr_pages)
1756 cpu_buffer->nr_pages_to_update = nr_pages -
1757 cpu_buffer->nr_pages;
1759 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1760 if (cpu_buffer->nr_pages_to_update > 0 &&
1761 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1762 &cpu_buffer->new_pages, cpu_id)) {
1769 /* Can't run something on an offline CPU. */
1770 if (!cpu_online(cpu_id))
1771 rb_update_pages(cpu_buffer);
1773 schedule_work_on(cpu_id,
1774 &cpu_buffer->update_pages_work);
1775 wait_for_completion(&cpu_buffer->update_done);
1778 cpu_buffer->nr_pages_to_update = 0;
1784 * The ring buffer resize can happen with the ring buffer
1785 * enabled, so that the update disturbs the tracing as little
1786 * as possible. But if the buffer is disabled, we do not need
1787 * to worry about that, and we can take the time to verify
1788 * that the buffer is not corrupt.
1790 if (atomic_read(&buffer->record_disabled)) {
1791 atomic_inc(&buffer->record_disabled);
1793 * Even though the buffer was disabled, we must make sure
1794 * that it is truly disabled before calling rb_check_pages.
1795 * There could have been a race between checking
1796 * record_disable and incrementing it.
1798 synchronize_sched();
1799 for_each_buffer_cpu(buffer, cpu) {
1800 cpu_buffer = buffer->buffers[cpu];
1801 rb_check_pages(cpu_buffer);
1803 atomic_dec(&buffer->record_disabled);
1806 mutex_unlock(&buffer->mutex);
1810 for_each_buffer_cpu(buffer, cpu) {
1811 struct buffer_page *bpage, *tmp;
1813 cpu_buffer = buffer->buffers[cpu];
1814 cpu_buffer->nr_pages_to_update = 0;
1816 if (list_empty(&cpu_buffer->new_pages))
1819 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1821 list_del_init(&bpage->list);
1822 free_buffer_page(bpage);
1825 mutex_unlock(&buffer->mutex);
1828 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1830 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1832 mutex_lock(&buffer->mutex);
1834 buffer->flags |= RB_FL_OVERWRITE;
1836 buffer->flags &= ~RB_FL_OVERWRITE;
1837 mutex_unlock(&buffer->mutex);
1839 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1841 static inline void *
1842 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1844 return bpage->data + index;
1847 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1849 return bpage->page->data + index;
1852 static inline struct ring_buffer_event *
1853 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1855 return __rb_page_index(cpu_buffer->reader_page,
1856 cpu_buffer->reader_page->read);
1859 static inline struct ring_buffer_event *
1860 rb_iter_head_event(struct ring_buffer_iter *iter)
1862 return __rb_page_index(iter->head_page, iter->head);
1865 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1867 return local_read(&bpage->page->commit);
1870 /* Size is determined by what has been committed */
1871 static inline unsigned rb_page_size(struct buffer_page *bpage)
1873 return rb_page_commit(bpage);
1876 static inline unsigned
1877 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1879 return rb_page_commit(cpu_buffer->commit_page);
1882 static inline unsigned
1883 rb_event_index(struct ring_buffer_event *event)
1885 unsigned long addr = (unsigned long)event;
1887 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1890 static void rb_inc_iter(struct ring_buffer_iter *iter)
1892 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1895 * The iterator could be on the reader page (it starts there).
1896 * But the head could have moved, since the reader was
1897 * found. Check for this case and assign the iterator
1898 * to the head page instead of next.
1900 if (iter->head_page == cpu_buffer->reader_page)
1901 iter->head_page = rb_set_head_page(cpu_buffer);
1903 rb_inc_page(cpu_buffer, &iter->head_page);
1905 iter->read_stamp = iter->head_page->page->time_stamp;
1910 * rb_handle_head_page - writer hit the head page
1912 * Returns: +1 to retry page
1917 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1918 struct buffer_page *tail_page,
1919 struct buffer_page *next_page)
1921 struct buffer_page *new_head;
1926 entries = rb_page_entries(next_page);
1929 * The hard part is here. We need to move the head
1930 * forward, and protect against both readers on
1931 * other CPUs and writers coming in via interrupts.
1933 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1937 * type can be one of four:
1938 * NORMAL - an interrupt already moved it for us
1939 * HEAD - we are the first to get here.
1940 * UPDATE - we are the interrupt interrupting
1942 * MOVED - a reader on another CPU moved the next
1943 * pointer to its reader page. Give up
1950 * We changed the head to UPDATE, thus
1951 * it is our responsibility to update
1954 local_add(entries, &cpu_buffer->overrun);
1955 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1958 * The entries will be zeroed out when we move the
1962 /* still more to do */
1965 case RB_PAGE_UPDATE:
1967 * This is an interrupt that interrupt the
1968 * previous update. Still more to do.
1971 case RB_PAGE_NORMAL:
1973 * An interrupt came in before the update
1974 * and processed this for us.
1975 * Nothing left to do.
1980 * The reader is on another CPU and just did
1981 * a swap with our next_page.
1986 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1991 * Now that we are here, the old head pointer is
1992 * set to UPDATE. This will keep the reader from
1993 * swapping the head page with the reader page.
1994 * The reader (on another CPU) will spin till
1997 * We just need to protect against interrupts
1998 * doing the job. We will set the next pointer
1999 * to HEAD. After that, we set the old pointer
2000 * to NORMAL, but only if it was HEAD before.
2001 * otherwise we are an interrupt, and only
2002 * want the outer most commit to reset it.
2004 new_head = next_page;
2005 rb_inc_page(cpu_buffer, &new_head);
2007 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2011 * Valid returns are:
2012 * HEAD - an interrupt came in and already set it.
2013 * NORMAL - One of two things:
2014 * 1) We really set it.
2015 * 2) A bunch of interrupts came in and moved
2016 * the page forward again.
2020 case RB_PAGE_NORMAL:
2024 RB_WARN_ON(cpu_buffer, 1);
2029 * It is possible that an interrupt came in,
2030 * set the head up, then more interrupts came in
2031 * and moved it again. When we get back here,
2032 * the page would have been set to NORMAL but we
2033 * just set it back to HEAD.
2035 * How do you detect this? Well, if that happened
2036 * the tail page would have moved.
2038 if (ret == RB_PAGE_NORMAL) {
2040 * If the tail had moved passed next, then we need
2041 * to reset the pointer.
2043 if (cpu_buffer->tail_page != tail_page &&
2044 cpu_buffer->tail_page != next_page)
2045 rb_head_page_set_normal(cpu_buffer, new_head,
2051 * If this was the outer most commit (the one that
2052 * changed the original pointer from HEAD to UPDATE),
2053 * then it is up to us to reset it to NORMAL.
2055 if (type == RB_PAGE_HEAD) {
2056 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2059 if (RB_WARN_ON(cpu_buffer,
2060 ret != RB_PAGE_UPDATE))
2068 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2069 unsigned long tail, struct rb_event_info *info)
2071 struct buffer_page *tail_page = info->tail_page;
2072 struct ring_buffer_event *event;
2073 unsigned long length = info->length;
2076 * Only the event that crossed the page boundary
2077 * must fill the old tail_page with padding.
2079 if (tail >= BUF_PAGE_SIZE) {
2081 * If the page was filled, then we still need
2082 * to update the real_end. Reset it to zero
2083 * and the reader will ignore it.
2085 if (tail == BUF_PAGE_SIZE)
2086 tail_page->real_end = 0;
2088 local_sub(length, &tail_page->write);
2092 event = __rb_page_index(tail_page, tail);
2093 kmemcheck_annotate_bitfield(event, bitfield);
2095 /* account for padding bytes */
2096 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2099 * Save the original length to the meta data.
2100 * This will be used by the reader to add lost event
2103 tail_page->real_end = tail;
2106 * If this event is bigger than the minimum size, then
2107 * we need to be careful that we don't subtract the
2108 * write counter enough to allow another writer to slip
2110 * We put in a discarded commit instead, to make sure
2111 * that this space is not used again.
2113 * If we are less than the minimum size, we don't need to
2116 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2117 /* No room for any events */
2119 /* Mark the rest of the page with padding */
2120 rb_event_set_padding(event);
2122 /* Set the write back to the previous setting */
2123 local_sub(length, &tail_page->write);
2127 /* Put in a discarded event */
2128 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2129 event->type_len = RINGBUF_TYPE_PADDING;
2130 /* time delta must be non zero */
2131 event->time_delta = 1;
2133 /* Set write to end of buffer */
2134 length = (tail + length) - BUF_PAGE_SIZE;
2135 local_sub(length, &tail_page->write);
2139 * This is the slow path, force gcc not to inline it.
2141 static noinline struct ring_buffer_event *
2142 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2143 unsigned long tail, struct rb_event_info *info)
2145 struct buffer_page *tail_page = info->tail_page;
2146 struct buffer_page *commit_page = cpu_buffer->commit_page;
2147 struct ring_buffer *buffer = cpu_buffer->buffer;
2148 struct buffer_page *next_page;
2152 next_page = tail_page;
2154 rb_inc_page(cpu_buffer, &next_page);
2157 * If for some reason, we had an interrupt storm that made
2158 * it all the way around the buffer, bail, and warn
2161 if (unlikely(next_page == commit_page)) {
2162 local_inc(&cpu_buffer->commit_overrun);
2167 * This is where the fun begins!
2169 * We are fighting against races between a reader that
2170 * could be on another CPU trying to swap its reader
2171 * page with the buffer head.
2173 * We are also fighting against interrupts coming in and
2174 * moving the head or tail on us as well.
2176 * If the next page is the head page then we have filled
2177 * the buffer, unless the commit page is still on the
2180 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2183 * If the commit is not on the reader page, then
2184 * move the header page.
2186 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2188 * If we are not in overwrite mode,
2189 * this is easy, just stop here.
2191 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2192 local_inc(&cpu_buffer->dropped_events);
2196 ret = rb_handle_head_page(cpu_buffer,
2205 * We need to be careful here too. The
2206 * commit page could still be on the reader
2207 * page. We could have a small buffer, and
2208 * have filled up the buffer with events
2209 * from interrupts and such, and wrapped.
2211 * Note, if the tail page is also the on the
2212 * reader_page, we let it move out.
2214 if (unlikely((cpu_buffer->commit_page !=
2215 cpu_buffer->tail_page) &&
2216 (cpu_buffer->commit_page ==
2217 cpu_buffer->reader_page))) {
2218 local_inc(&cpu_buffer->commit_overrun);
2224 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
2227 * Nested commits always have zero deltas, so
2228 * just reread the time stamp
2230 ts = rb_time_stamp(buffer);
2231 next_page->page->time_stamp = ts;
2236 rb_reset_tail(cpu_buffer, tail, info);
2238 /* fail and let the caller try again */
2239 return ERR_PTR(-EAGAIN);
2243 rb_reset_tail(cpu_buffer, tail, info);
2248 /* Slow path, do not inline */
2249 static noinline struct ring_buffer_event *
2250 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
2252 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2254 /* Not the first event on the page? */
2255 if (rb_event_index(event)) {
2256 event->time_delta = delta & TS_MASK;
2257 event->array[0] = delta >> TS_SHIFT;
2259 /* nope, just zero it */
2260 event->time_delta = 0;
2261 event->array[0] = 0;
2264 return skip_time_extend(event);
2267 static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2268 struct ring_buffer_event *event);
2271 * rb_update_event - update event type and data
2272 * @event: the event to update
2273 * @type: the type of event
2274 * @length: the size of the event field in the ring buffer
2276 * Update the type and data fields of the event. The length
2277 * is the actual size that is written to the ring buffer,
2278 * and with this, we can determine what to place into the
2282 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2283 struct ring_buffer_event *event,
2284 struct rb_event_info *info)
2286 unsigned length = info->length;
2287 u64 delta = info->delta;
2289 /* Only a commit updates the timestamp */
2290 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2294 * If we need to add a timestamp, then we
2295 * add it to the start of the resevered space.
2297 if (unlikely(info->add_timestamp)) {
2298 event = rb_add_time_stamp(event, delta);
2299 length -= RB_LEN_TIME_EXTEND;
2303 event->time_delta = delta;
2304 length -= RB_EVNT_HDR_SIZE;
2305 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2306 event->type_len = 0;
2307 event->array[0] = length;
2309 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2312 static unsigned rb_calculate_event_length(unsigned length)
2314 struct ring_buffer_event event; /* Used only for sizeof array */
2316 /* zero length can cause confusions */
2320 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2321 length += sizeof(event.array[0]);
2323 length += RB_EVNT_HDR_SIZE;
2324 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2327 * In case the time delta is larger than the 27 bits for it
2328 * in the header, we need to add a timestamp. If another
2329 * event comes in when trying to discard this one to increase
2330 * the length, then the timestamp will be added in the allocated
2331 * space of this event. If length is bigger than the size needed
2332 * for the TIME_EXTEND, then padding has to be used. The events
2333 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2334 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2335 * As length is a multiple of 4, we only need to worry if it
2336 * is 12 (RB_LEN_TIME_EXTEND + 4).
2338 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2339 length += RB_ALIGNMENT;
2344 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2345 static inline bool sched_clock_stable(void)
2352 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2353 struct ring_buffer_event *event)
2355 unsigned long new_index, old_index;
2356 struct buffer_page *bpage;
2357 unsigned long index;
2360 new_index = rb_event_index(event);
2361 old_index = new_index + rb_event_ts_length(event);
2362 addr = (unsigned long)event;
2365 bpage = cpu_buffer->tail_page;
2367 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2368 unsigned long write_mask =
2369 local_read(&bpage->write) & ~RB_WRITE_MASK;
2370 unsigned long event_length = rb_event_length(event);
2372 * This is on the tail page. It is possible that
2373 * a write could come in and move the tail page
2374 * and write to the next page. That is fine
2375 * because we just shorten what is on this page.
2377 old_index += write_mask;
2378 new_index += write_mask;
2379 index = local_cmpxchg(&bpage->write, old_index, new_index);
2380 if (index == old_index) {
2381 /* update counters */
2382 local_sub(event_length, &cpu_buffer->entries_bytes);
2387 /* could not discard */
2391 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2393 local_inc(&cpu_buffer->committing);
2394 local_inc(&cpu_buffer->commits);
2398 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2400 unsigned long max_count;
2403 * We only race with interrupts and NMIs on this CPU.
2404 * If we own the commit event, then we can commit
2405 * all others that interrupted us, since the interruptions
2406 * are in stack format (they finish before they come
2407 * back to us). This allows us to do a simple loop to
2408 * assign the commit to the tail.
2411 max_count = cpu_buffer->nr_pages * 100;
2413 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
2414 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2416 if (RB_WARN_ON(cpu_buffer,
2417 rb_is_reader_page(cpu_buffer->tail_page)))
2419 local_set(&cpu_buffer->commit_page->page->commit,
2420 rb_page_write(cpu_buffer->commit_page));
2421 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2422 cpu_buffer->write_stamp =
2423 cpu_buffer->commit_page->page->time_stamp;
2424 /* add barrier to keep gcc from optimizing too much */
2427 while (rb_commit_index(cpu_buffer) !=
2428 rb_page_write(cpu_buffer->commit_page)) {
2430 local_set(&cpu_buffer->commit_page->page->commit,
2431 rb_page_write(cpu_buffer->commit_page));
2432 RB_WARN_ON(cpu_buffer,
2433 local_read(&cpu_buffer->commit_page->page->commit) &
2438 /* again, keep gcc from optimizing */
2442 * If an interrupt came in just after the first while loop
2443 * and pushed the tail page forward, we will be left with
2444 * a dangling commit that will never go forward.
2446 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
2450 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2452 unsigned long commits;
2454 if (RB_WARN_ON(cpu_buffer,
2455 !local_read(&cpu_buffer->committing)))
2459 commits = local_read(&cpu_buffer->commits);
2460 /* synchronize with interrupts */
2462 if (local_read(&cpu_buffer->committing) == 1)
2463 rb_set_commit_to_write(cpu_buffer);
2465 local_dec(&cpu_buffer->committing);
2467 /* synchronize with interrupts */
2471 * Need to account for interrupts coming in between the
2472 * updating of the commit page and the clearing of the
2473 * committing counter.
2475 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2476 !local_read(&cpu_buffer->committing)) {
2477 local_inc(&cpu_buffer->committing);
2482 static inline void rb_event_discard(struct ring_buffer_event *event)
2484 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2485 event = skip_time_extend(event);
2487 /* array[0] holds the actual length for the discarded event */
2488 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2489 event->type_len = RINGBUF_TYPE_PADDING;
2490 /* time delta must be non zero */
2491 if (!event->time_delta)
2492 event->time_delta = 1;
2496 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2497 struct ring_buffer_event *event)
2499 unsigned long addr = (unsigned long)event;
2500 unsigned long index;
2502 index = rb_event_index(event);
2505 return cpu_buffer->commit_page->page == (void *)addr &&
2506 rb_commit_index(cpu_buffer) == index;
2510 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2511 struct ring_buffer_event *event)
2516 * The event first in the commit queue updates the
2519 if (rb_event_is_commit(cpu_buffer, event)) {
2521 * A commit event that is first on a page
2522 * updates the write timestamp with the page stamp
2524 if (!rb_event_index(event))
2525 cpu_buffer->write_stamp =
2526 cpu_buffer->commit_page->page->time_stamp;
2527 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2528 delta = event->array[0];
2530 delta += event->time_delta;
2531 cpu_buffer->write_stamp += delta;
2533 cpu_buffer->write_stamp += event->time_delta;
2537 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2538 struct ring_buffer_event *event)
2540 local_inc(&cpu_buffer->entries);
2541 rb_update_write_stamp(cpu_buffer, event);
2542 rb_end_commit(cpu_buffer);
2545 static __always_inline void
2546 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2550 if (buffer->irq_work.waiters_pending) {
2551 buffer->irq_work.waiters_pending = false;
2552 /* irq_work_queue() supplies it's own memory barriers */
2553 irq_work_queue(&buffer->irq_work.work);
2556 if (cpu_buffer->irq_work.waiters_pending) {
2557 cpu_buffer->irq_work.waiters_pending = false;
2558 /* irq_work_queue() supplies it's own memory barriers */
2559 irq_work_queue(&cpu_buffer->irq_work.work);
2562 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
2564 if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2565 cpu_buffer->irq_work.wakeup_full = true;
2566 cpu_buffer->irq_work.full_waiters_pending = false;
2567 /* irq_work_queue() supplies it's own memory barriers */
2568 irq_work_queue(&cpu_buffer->irq_work.work);
2573 * The lock and unlock are done within a preempt disable section.
2574 * The current_context per_cpu variable can only be modified
2575 * by the current task between lock and unlock. But it can
2576 * be modified more than once via an interrupt. To pass this
2577 * information from the lock to the unlock without having to
2578 * access the 'in_interrupt()' functions again (which do show
2579 * a bit of overhead in something as critical as function tracing,
2580 * we use a bitmask trick.
2582 * bit 0 = NMI context
2583 * bit 1 = IRQ context
2584 * bit 2 = SoftIRQ context
2585 * bit 3 = normal context.
2587 * This works because this is the order of contexts that can
2588 * preempt other contexts. A SoftIRQ never preempts an IRQ
2591 * When the context is determined, the corresponding bit is
2592 * checked and set (if it was set, then a recursion of that context
2595 * On unlock, we need to clear this bit. To do so, just subtract
2596 * 1 from the current_context and AND it to itself.
2600 * 101 & 100 = 100 (clearing bit zero)
2603 * 1010 & 1001 = 1000 (clearing bit 1)
2605 * The least significant bit can be cleared this way, and it
2606 * just so happens that it is the same bit corresponding to
2607 * the current context.
2610 static __always_inline int
2611 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2613 unsigned int val = cpu_buffer->current_context;
2616 if (in_interrupt()) {
2622 bit = RB_CTX_SOFTIRQ;
2624 bit = RB_CTX_NORMAL;
2626 if (unlikely(val & (1 << bit)))
2630 cpu_buffer->current_context = val;
2635 static __always_inline void
2636 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2638 cpu_buffer->current_context &= cpu_buffer->current_context - 1;
2642 * ring_buffer_unlock_commit - commit a reserved
2643 * @buffer: The buffer to commit to
2644 * @event: The event pointer to commit.
2646 * This commits the data to the ring buffer, and releases any locks held.
2648 * Must be paired with ring_buffer_lock_reserve.
2650 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2651 struct ring_buffer_event *event)
2653 struct ring_buffer_per_cpu *cpu_buffer;
2654 int cpu = raw_smp_processor_id();
2656 cpu_buffer = buffer->buffers[cpu];
2658 rb_commit(cpu_buffer, event);
2660 rb_wakeups(buffer, cpu_buffer);
2662 trace_recursive_unlock(cpu_buffer);
2664 preempt_enable_notrace();
2668 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2670 static noinline void
2671 rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2672 struct rb_event_info *info)
2674 WARN_ONCE(info->delta > (1ULL << 59),
2675 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2676 (unsigned long long)info->delta,
2677 (unsigned long long)info->ts,
2678 (unsigned long long)cpu_buffer->write_stamp,
2679 sched_clock_stable() ? "" :
2680 "If you just came from a suspend/resume,\n"
2681 "please switch to the trace global clock:\n"
2682 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2683 info->add_timestamp = 1;
2686 static struct ring_buffer_event *
2687 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2688 struct rb_event_info *info)
2690 struct ring_buffer_event *event;
2691 struct buffer_page *tail_page;
2692 unsigned long tail, write;
2695 * If the time delta since the last event is too big to
2696 * hold in the time field of the event, then we append a
2697 * TIME EXTEND event ahead of the data event.
2699 if (unlikely(info->add_timestamp))
2700 info->length += RB_LEN_TIME_EXTEND;
2702 tail_page = info->tail_page = cpu_buffer->tail_page;
2703 write = local_add_return(info->length, &tail_page->write);
2705 /* set write to only the index of the write */
2706 write &= RB_WRITE_MASK;
2707 tail = write - info->length;
2710 * If this is the first commit on the page, then it has the same
2711 * timestamp as the page itself.
2716 /* See if we shot pass the end of this buffer page */
2717 if (unlikely(write > BUF_PAGE_SIZE))
2718 return rb_move_tail(cpu_buffer, tail, info);
2720 /* We reserved something on the buffer */
2722 event = __rb_page_index(tail_page, tail);
2723 kmemcheck_annotate_bitfield(event, bitfield);
2724 rb_update_event(cpu_buffer, event, info);
2726 local_inc(&tail_page->entries);
2729 * If this is the first commit on the page, then update
2733 tail_page->page->time_stamp = info->ts;
2735 /* account for these added bytes */
2736 local_add(info->length, &cpu_buffer->entries_bytes);
2741 static struct ring_buffer_event *
2742 rb_reserve_next_event(struct ring_buffer *buffer,
2743 struct ring_buffer_per_cpu *cpu_buffer,
2744 unsigned long length)
2746 struct ring_buffer_event *event;
2747 struct rb_event_info info;
2751 rb_start_commit(cpu_buffer);
2753 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2755 * Due to the ability to swap a cpu buffer from a buffer
2756 * it is possible it was swapped before we committed.
2757 * (committing stops a swap). We check for it here and
2758 * if it happened, we have to fail the write.
2761 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2762 local_dec(&cpu_buffer->committing);
2763 local_dec(&cpu_buffer->commits);
2768 info.length = rb_calculate_event_length(length);
2770 info.add_timestamp = 0;
2774 * We allow for interrupts to reenter here and do a trace.
2775 * If one does, it will cause this original code to loop
2776 * back here. Even with heavy interrupts happening, this
2777 * should only happen a few times in a row. If this happens
2778 * 1000 times in a row, there must be either an interrupt
2779 * storm or we have something buggy.
2782 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2785 info.ts = rb_time_stamp(cpu_buffer->buffer);
2786 diff = info.ts - cpu_buffer->write_stamp;
2788 /* make sure this diff is calculated here */
2791 /* Did the write stamp get updated already? */
2792 if (likely(info.ts >= cpu_buffer->write_stamp)) {
2794 if (unlikely(test_time_stamp(info.delta)))
2795 rb_handle_timestamp(cpu_buffer, &info);
2798 event = __rb_reserve_next(cpu_buffer, &info);
2800 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2801 if (info.add_timestamp)
2802 info.length -= RB_LEN_TIME_EXTEND;
2812 rb_end_commit(cpu_buffer);
2817 * ring_buffer_lock_reserve - reserve a part of the buffer
2818 * @buffer: the ring buffer to reserve from
2819 * @length: the length of the data to reserve (excluding event header)
2821 * Returns a reseverd event on the ring buffer to copy directly to.
2822 * The user of this interface will need to get the body to write into
2823 * and can use the ring_buffer_event_data() interface.
2825 * The length is the length of the data needed, not the event length
2826 * which also includes the event header.
2828 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2829 * If NULL is returned, then nothing has been allocated or locked.
2831 struct ring_buffer_event *
2832 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2834 struct ring_buffer_per_cpu *cpu_buffer;
2835 struct ring_buffer_event *event;
2838 /* If we are tracing schedule, we don't want to recurse */
2839 preempt_disable_notrace();
2841 if (unlikely(atomic_read(&buffer->record_disabled)))
2844 cpu = raw_smp_processor_id();
2846 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2849 cpu_buffer = buffer->buffers[cpu];
2851 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2854 if (unlikely(length > BUF_MAX_DATA_SIZE))
2857 if (unlikely(trace_recursive_lock(cpu_buffer)))
2860 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2867 trace_recursive_unlock(cpu_buffer);
2869 preempt_enable_notrace();
2872 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2875 * Decrement the entries to the page that an event is on.
2876 * The event does not even need to exist, only the pointer
2877 * to the page it is on. This may only be called before the commit
2881 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2882 struct ring_buffer_event *event)
2884 unsigned long addr = (unsigned long)event;
2885 struct buffer_page *bpage = cpu_buffer->commit_page;
2886 struct buffer_page *start;
2890 /* Do the likely case first */
2891 if (likely(bpage->page == (void *)addr)) {
2892 local_dec(&bpage->entries);
2897 * Because the commit page may be on the reader page we
2898 * start with the next page and check the end loop there.
2900 rb_inc_page(cpu_buffer, &bpage);
2903 if (bpage->page == (void *)addr) {
2904 local_dec(&bpage->entries);
2907 rb_inc_page(cpu_buffer, &bpage);
2908 } while (bpage != start);
2910 /* commit not part of this buffer?? */
2911 RB_WARN_ON(cpu_buffer, 1);
2915 * ring_buffer_commit_discard - discard an event that has not been committed
2916 * @buffer: the ring buffer
2917 * @event: non committed event to discard
2919 * Sometimes an event that is in the ring buffer needs to be ignored.
2920 * This function lets the user discard an event in the ring buffer
2921 * and then that event will not be read later.
2923 * This function only works if it is called before the the item has been
2924 * committed. It will try to free the event from the ring buffer
2925 * if another event has not been added behind it.
2927 * If another event has been added behind it, it will set the event
2928 * up as discarded, and perform the commit.
2930 * If this function is called, do not call ring_buffer_unlock_commit on
2933 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2934 struct ring_buffer_event *event)
2936 struct ring_buffer_per_cpu *cpu_buffer;
2939 /* The event is discarded regardless */
2940 rb_event_discard(event);
2942 cpu = smp_processor_id();
2943 cpu_buffer = buffer->buffers[cpu];
2946 * This must only be called if the event has not been
2947 * committed yet. Thus we can assume that preemption
2948 * is still disabled.
2950 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2952 rb_decrement_entry(cpu_buffer, event);
2953 if (rb_try_to_discard(cpu_buffer, event))
2957 * The commit is still visible by the reader, so we
2958 * must still update the timestamp.
2960 rb_update_write_stamp(cpu_buffer, event);
2962 rb_end_commit(cpu_buffer);
2964 trace_recursive_unlock(cpu_buffer);
2966 preempt_enable_notrace();
2969 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2972 * ring_buffer_write - write data to the buffer without reserving
2973 * @buffer: The ring buffer to write to.
2974 * @length: The length of the data being written (excluding the event header)
2975 * @data: The data to write to the buffer.
2977 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2978 * one function. If you already have the data to write to the buffer, it
2979 * may be easier to simply call this function.
2981 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2982 * and not the length of the event which would hold the header.
2984 int ring_buffer_write(struct ring_buffer *buffer,
2985 unsigned long length,
2988 struct ring_buffer_per_cpu *cpu_buffer;
2989 struct ring_buffer_event *event;
2994 preempt_disable_notrace();
2996 if (atomic_read(&buffer->record_disabled))
2999 cpu = raw_smp_processor_id();
3001 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3004 cpu_buffer = buffer->buffers[cpu];
3006 if (atomic_read(&cpu_buffer->record_disabled))
3009 if (length > BUF_MAX_DATA_SIZE)
3012 if (unlikely(trace_recursive_lock(cpu_buffer)))
3015 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3019 body = rb_event_data(event);
3021 memcpy(body, data, length);
3023 rb_commit(cpu_buffer, event);
3025 rb_wakeups(buffer, cpu_buffer);
3030 trace_recursive_unlock(cpu_buffer);
3033 preempt_enable_notrace();
3037 EXPORT_SYMBOL_GPL(ring_buffer_write);
3039 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3041 struct buffer_page *reader = cpu_buffer->reader_page;
3042 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3043 struct buffer_page *commit = cpu_buffer->commit_page;
3045 /* In case of error, head will be NULL */
3046 if (unlikely(!head))
3049 return reader->read == rb_page_commit(reader) &&
3050 (commit == reader ||
3052 head->read == rb_page_commit(commit)));
3056 * ring_buffer_record_disable - stop all writes into the buffer
3057 * @buffer: The ring buffer to stop writes to.
3059 * This prevents all writes to the buffer. Any attempt to write
3060 * to the buffer after this will fail and return NULL.
3062 * The caller should call synchronize_sched() after this.
3064 void ring_buffer_record_disable(struct ring_buffer *buffer)
3066 atomic_inc(&buffer->record_disabled);
3068 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3071 * ring_buffer_record_enable - enable writes to the buffer
3072 * @buffer: The ring buffer to enable writes
3074 * Note, multiple disables will need the same number of enables
3075 * to truly enable the writing (much like preempt_disable).
3077 void ring_buffer_record_enable(struct ring_buffer *buffer)
3079 atomic_dec(&buffer->record_disabled);
3081 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3084 * ring_buffer_record_off - stop all writes into the buffer
3085 * @buffer: The ring buffer to stop writes to.
3087 * This prevents all writes to the buffer. Any attempt to write
3088 * to the buffer after this will fail and return NULL.
3090 * This is different than ring_buffer_record_disable() as
3091 * it works like an on/off switch, where as the disable() version
3092 * must be paired with a enable().
3094 void ring_buffer_record_off(struct ring_buffer *buffer)
3097 unsigned int new_rd;
3100 rd = atomic_read(&buffer->record_disabled);
3101 new_rd = rd | RB_BUFFER_OFF;
3102 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3104 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3107 * ring_buffer_record_on - restart writes into the buffer
3108 * @buffer: The ring buffer to start writes to.
3110 * This enables all writes to the buffer that was disabled by
3111 * ring_buffer_record_off().
3113 * This is different than ring_buffer_record_enable() as
3114 * it works like an on/off switch, where as the enable() version
3115 * must be paired with a disable().
3117 void ring_buffer_record_on(struct ring_buffer *buffer)
3120 unsigned int new_rd;
3123 rd = atomic_read(&buffer->record_disabled);
3124 new_rd = rd & ~RB_BUFFER_OFF;
3125 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3127 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3130 * ring_buffer_record_is_on - return true if the ring buffer can write
3131 * @buffer: The ring buffer to see if write is enabled
3133 * Returns true if the ring buffer is in a state that it accepts writes.
3135 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3137 return !atomic_read(&buffer->record_disabled);
3141 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3142 * @buffer: The ring buffer to stop writes to.
3143 * @cpu: The CPU buffer to stop
3145 * This prevents all writes to the buffer. Any attempt to write
3146 * to the buffer after this will fail and return NULL.
3148 * The caller should call synchronize_sched() after this.
3150 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3152 struct ring_buffer_per_cpu *cpu_buffer;
3154 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3157 cpu_buffer = buffer->buffers[cpu];
3158 atomic_inc(&cpu_buffer->record_disabled);
3160 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3163 * ring_buffer_record_enable_cpu - enable writes to the buffer
3164 * @buffer: The ring buffer to enable writes
3165 * @cpu: The CPU to enable.
3167 * Note, multiple disables will need the same number of enables
3168 * to truly enable the writing (much like preempt_disable).
3170 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3172 struct ring_buffer_per_cpu *cpu_buffer;
3174 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3177 cpu_buffer = buffer->buffers[cpu];
3178 atomic_dec(&cpu_buffer->record_disabled);
3180 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3183 * The total entries in the ring buffer is the running counter
3184 * of entries entered into the ring buffer, minus the sum of
3185 * the entries read from the ring buffer and the number of
3186 * entries that were overwritten.
3188 static inline unsigned long
3189 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3191 return local_read(&cpu_buffer->entries) -
3192 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3196 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3197 * @buffer: The ring buffer
3198 * @cpu: The per CPU buffer to read from.
3200 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3202 unsigned long flags;
3203 struct ring_buffer_per_cpu *cpu_buffer;
3204 struct buffer_page *bpage;
3207 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3210 cpu_buffer = buffer->buffers[cpu];
3211 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3213 * if the tail is on reader_page, oldest time stamp is on the reader
3216 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3217 bpage = cpu_buffer->reader_page;
3219 bpage = rb_set_head_page(cpu_buffer);
3221 ret = bpage->page->time_stamp;
3222 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3226 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3229 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3230 * @buffer: The ring buffer
3231 * @cpu: The per CPU buffer to read from.
3233 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3235 struct ring_buffer_per_cpu *cpu_buffer;
3238 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3241 cpu_buffer = buffer->buffers[cpu];
3242 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3246 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3249 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3250 * @buffer: The ring buffer
3251 * @cpu: The per CPU buffer to get the entries from.
3253 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3255 struct ring_buffer_per_cpu *cpu_buffer;
3257 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3260 cpu_buffer = buffer->buffers[cpu];
3262 return rb_num_of_entries(cpu_buffer);
3264 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3267 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3268 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3269 * @buffer: The ring buffer
3270 * @cpu: The per CPU buffer to get the number of overruns from
3272 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3274 struct ring_buffer_per_cpu *cpu_buffer;
3277 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3280 cpu_buffer = buffer->buffers[cpu];
3281 ret = local_read(&cpu_buffer->overrun);
3285 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3288 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3289 * commits failing due to the buffer wrapping around while there are uncommitted
3290 * events, such as during an interrupt storm.
3291 * @buffer: The ring buffer
3292 * @cpu: The per CPU buffer to get the number of overruns from
3295 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3297 struct ring_buffer_per_cpu *cpu_buffer;
3300 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3303 cpu_buffer = buffer->buffers[cpu];
3304 ret = local_read(&cpu_buffer->commit_overrun);
3308 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3311 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3312 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3313 * @buffer: The ring buffer
3314 * @cpu: The per CPU buffer to get the number of overruns from
3317 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3319 struct ring_buffer_per_cpu *cpu_buffer;
3322 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3325 cpu_buffer = buffer->buffers[cpu];
3326 ret = local_read(&cpu_buffer->dropped_events);
3330 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3333 * ring_buffer_read_events_cpu - get the number of events successfully read
3334 * @buffer: The ring buffer
3335 * @cpu: The per CPU buffer to get the number of events read
3338 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3340 struct ring_buffer_per_cpu *cpu_buffer;
3342 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3345 cpu_buffer = buffer->buffers[cpu];
3346 return cpu_buffer->read;
3348 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3351 * ring_buffer_entries - get the number of entries in a buffer
3352 * @buffer: The ring buffer
3354 * Returns the total number of entries in the ring buffer
3357 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3359 struct ring_buffer_per_cpu *cpu_buffer;
3360 unsigned long entries = 0;
3363 /* if you care about this being correct, lock the buffer */
3364 for_each_buffer_cpu(buffer, cpu) {
3365 cpu_buffer = buffer->buffers[cpu];
3366 entries += rb_num_of_entries(cpu_buffer);
3371 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3374 * ring_buffer_overruns - get the number of overruns in buffer
3375 * @buffer: The ring buffer
3377 * Returns the total number of overruns in the ring buffer
3380 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3382 struct ring_buffer_per_cpu *cpu_buffer;
3383 unsigned long overruns = 0;
3386 /* if you care about this being correct, lock the buffer */
3387 for_each_buffer_cpu(buffer, cpu) {
3388 cpu_buffer = buffer->buffers[cpu];
3389 overruns += local_read(&cpu_buffer->overrun);
3394 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3396 static void rb_iter_reset(struct ring_buffer_iter *iter)
3398 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3400 /* Iterator usage is expected to have record disabled */
3401 iter->head_page = cpu_buffer->reader_page;
3402 iter->head = cpu_buffer->reader_page->read;
3404 iter->cache_reader_page = iter->head_page;
3405 iter->cache_read = cpu_buffer->read;
3408 iter->read_stamp = cpu_buffer->read_stamp;
3410 iter->read_stamp = iter->head_page->page->time_stamp;
3414 * ring_buffer_iter_reset - reset an iterator
3415 * @iter: The iterator to reset
3417 * Resets the iterator, so that it will start from the beginning
3420 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3422 struct ring_buffer_per_cpu *cpu_buffer;
3423 unsigned long flags;
3428 cpu_buffer = iter->cpu_buffer;
3430 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3431 rb_iter_reset(iter);
3432 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3434 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3437 * ring_buffer_iter_empty - check if an iterator has no more to read
3438 * @iter: The iterator to check
3440 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3442 struct ring_buffer_per_cpu *cpu_buffer;
3444 cpu_buffer = iter->cpu_buffer;
3446 return iter->head_page == cpu_buffer->commit_page &&
3447 iter->head == rb_commit_index(cpu_buffer);
3449 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3452 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3453 struct ring_buffer_event *event)
3457 switch (event->type_len) {
3458 case RINGBUF_TYPE_PADDING:
3461 case RINGBUF_TYPE_TIME_EXTEND:
3462 delta = event->array[0];
3464 delta += event->time_delta;
3465 cpu_buffer->read_stamp += delta;
3468 case RINGBUF_TYPE_TIME_STAMP:
3469 /* FIXME: not implemented */
3472 case RINGBUF_TYPE_DATA:
3473 cpu_buffer->read_stamp += event->time_delta;
3483 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3484 struct ring_buffer_event *event)
3488 switch (event->type_len) {
3489 case RINGBUF_TYPE_PADDING:
3492 case RINGBUF_TYPE_TIME_EXTEND:
3493 delta = event->array[0];
3495 delta += event->time_delta;
3496 iter->read_stamp += delta;
3499 case RINGBUF_TYPE_TIME_STAMP:
3500 /* FIXME: not implemented */
3503 case RINGBUF_TYPE_DATA:
3504 iter->read_stamp += event->time_delta;
3513 static struct buffer_page *
3514 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3516 struct buffer_page *reader = NULL;
3517 unsigned long overwrite;
3518 unsigned long flags;
3522 local_irq_save(flags);
3523 arch_spin_lock(&cpu_buffer->lock);
3527 * This should normally only loop twice. But because the
3528 * start of the reader inserts an empty page, it causes
3529 * a case where we will loop three times. There should be no
3530 * reason to loop four times (that I know of).
3532 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3537 reader = cpu_buffer->reader_page;
3539 /* If there's more to read, return this page */
3540 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3543 /* Never should we have an index greater than the size */
3544 if (RB_WARN_ON(cpu_buffer,
3545 cpu_buffer->reader_page->read > rb_page_size(reader)))
3548 /* check if we caught up to the tail */
3550 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3553 /* Don't bother swapping if the ring buffer is empty */
3554 if (rb_num_of_entries(cpu_buffer) == 0)
3558 * Reset the reader page to size zero.
3560 local_set(&cpu_buffer->reader_page->write, 0);
3561 local_set(&cpu_buffer->reader_page->entries, 0);
3562 local_set(&cpu_buffer->reader_page->page->commit, 0);
3563 cpu_buffer->reader_page->real_end = 0;
3567 * Splice the empty reader page into the list around the head.
3569 reader = rb_set_head_page(cpu_buffer);
3572 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3573 cpu_buffer->reader_page->list.prev = reader->list.prev;
3576 * cpu_buffer->pages just needs to point to the buffer, it
3577 * has no specific buffer page to point to. Lets move it out
3578 * of our way so we don't accidentally swap it.
3580 cpu_buffer->pages = reader->list.prev;
3582 /* The reader page will be pointing to the new head */
3583 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3586 * We want to make sure we read the overruns after we set up our
3587 * pointers to the next object. The writer side does a
3588 * cmpxchg to cross pages which acts as the mb on the writer
3589 * side. Note, the reader will constantly fail the swap
3590 * while the writer is updating the pointers, so this
3591 * guarantees that the overwrite recorded here is the one we
3592 * want to compare with the last_overrun.
3595 overwrite = local_read(&(cpu_buffer->overrun));
3598 * Here's the tricky part.
3600 * We need to move the pointer past the header page.
3601 * But we can only do that if a writer is not currently
3602 * moving it. The page before the header page has the
3603 * flag bit '1' set if it is pointing to the page we want.
3604 * but if the writer is in the process of moving it
3605 * than it will be '2' or already moved '0'.
3608 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3611 * If we did not convert it, then we must try again.
3617 * Yeah! We succeeded in replacing the page.
3619 * Now make the new head point back to the reader page.
3621 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3622 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3624 /* Finally update the reader page to the new head */
3625 cpu_buffer->reader_page = reader;
3626 cpu_buffer->reader_page->read = 0;
3628 if (overwrite != cpu_buffer->last_overrun) {
3629 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3630 cpu_buffer->last_overrun = overwrite;
3636 /* Update the read_stamp on the first event */
3637 if (reader && reader->read == 0)
3638 cpu_buffer->read_stamp = reader->page->time_stamp;
3640 arch_spin_unlock(&cpu_buffer->lock);
3641 local_irq_restore(flags);
3646 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3648 struct ring_buffer_event *event;
3649 struct buffer_page *reader;
3652 reader = rb_get_reader_page(cpu_buffer);
3654 /* This function should not be called when buffer is empty */
3655 if (RB_WARN_ON(cpu_buffer, !reader))
3658 event = rb_reader_event(cpu_buffer);
3660 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3663 rb_update_read_stamp(cpu_buffer, event);
3665 length = rb_event_length(event);
3666 cpu_buffer->reader_page->read += length;
3669 static void rb_advance_iter(struct ring_buffer_iter *iter)
3671 struct ring_buffer_per_cpu *cpu_buffer;
3672 struct ring_buffer_event *event;
3675 cpu_buffer = iter->cpu_buffer;
3678 * Check if we are at the end of the buffer.
3680 if (iter->head >= rb_page_size(iter->head_page)) {
3681 /* discarded commits can make the page empty */
3682 if (iter->head_page == cpu_buffer->commit_page)
3688 event = rb_iter_head_event(iter);
3690 length = rb_event_length(event);
3693 * This should not be called to advance the header if we are
3694 * at the tail of the buffer.
3696 if (RB_WARN_ON(cpu_buffer,
3697 (iter->head_page == cpu_buffer->commit_page) &&
3698 (iter->head + length > rb_commit_index(cpu_buffer))))
3701 rb_update_iter_read_stamp(iter, event);
3703 iter->head += length;
3705 /* check for end of page padding */
3706 if ((iter->head >= rb_page_size(iter->head_page)) &&
3707 (iter->head_page != cpu_buffer->commit_page))
3711 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3713 return cpu_buffer->lost_events;
3716 static struct ring_buffer_event *
3717 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3718 unsigned long *lost_events)
3720 struct ring_buffer_event *event;
3721 struct buffer_page *reader;
3726 * We repeat when a time extend is encountered.
3727 * Since the time extend is always attached to a data event,
3728 * we should never loop more than once.
3729 * (We never hit the following condition more than twice).
3731 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3734 reader = rb_get_reader_page(cpu_buffer);
3738 event = rb_reader_event(cpu_buffer);
3740 switch (event->type_len) {
3741 case RINGBUF_TYPE_PADDING:
3742 if (rb_null_event(event))
3743 RB_WARN_ON(cpu_buffer, 1);
3745 * Because the writer could be discarding every
3746 * event it creates (which would probably be bad)
3747 * if we were to go back to "again" then we may never
3748 * catch up, and will trigger the warn on, or lock
3749 * the box. Return the padding, and we will release
3750 * the current locks, and try again.
3754 case RINGBUF_TYPE_TIME_EXTEND:
3755 /* Internal data, OK to advance */
3756 rb_advance_reader(cpu_buffer);
3759 case RINGBUF_TYPE_TIME_STAMP:
3760 /* FIXME: not implemented */
3761 rb_advance_reader(cpu_buffer);
3764 case RINGBUF_TYPE_DATA:
3766 *ts = cpu_buffer->read_stamp + event->time_delta;
3767 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3768 cpu_buffer->cpu, ts);
3771 *lost_events = rb_lost_events(cpu_buffer);
3780 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3782 static struct ring_buffer_event *
3783 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3785 struct ring_buffer *buffer;
3786 struct ring_buffer_per_cpu *cpu_buffer;
3787 struct ring_buffer_event *event;
3790 cpu_buffer = iter->cpu_buffer;
3791 buffer = cpu_buffer->buffer;
3794 * Check if someone performed a consuming read to
3795 * the buffer. A consuming read invalidates the iterator
3796 * and we need to reset the iterator in this case.
3798 if (unlikely(iter->cache_read != cpu_buffer->read ||
3799 iter->cache_reader_page != cpu_buffer->reader_page))
3800 rb_iter_reset(iter);
3803 if (ring_buffer_iter_empty(iter))
3807 * We repeat when a time extend is encountered or we hit
3808 * the end of the page. Since the time extend is always attached
3809 * to a data event, we should never loop more than three times.
3810 * Once for going to next page, once on time extend, and
3811 * finally once to get the event.
3812 * (We never hit the following condition more than thrice).
3814 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3817 if (rb_per_cpu_empty(cpu_buffer))
3820 if (iter->head >= rb_page_size(iter->head_page)) {
3825 event = rb_iter_head_event(iter);
3827 switch (event->type_len) {
3828 case RINGBUF_TYPE_PADDING:
3829 if (rb_null_event(event)) {
3833 rb_advance_iter(iter);
3836 case RINGBUF_TYPE_TIME_EXTEND:
3837 /* Internal data, OK to advance */
3838 rb_advance_iter(iter);
3841 case RINGBUF_TYPE_TIME_STAMP:
3842 /* FIXME: not implemented */
3843 rb_advance_iter(iter);
3846 case RINGBUF_TYPE_DATA:
3848 *ts = iter->read_stamp + event->time_delta;
3849 ring_buffer_normalize_time_stamp(buffer,
3850 cpu_buffer->cpu, ts);
3860 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3862 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
3864 if (likely(!in_nmi())) {
3865 raw_spin_lock(&cpu_buffer->reader_lock);
3870 * If an NMI die dumps out the content of the ring buffer
3871 * trylock must be used to prevent a deadlock if the NMI
3872 * preempted a task that holds the ring buffer locks. If
3873 * we get the lock then all is fine, if not, then continue
3874 * to do the read, but this can corrupt the ring buffer,
3875 * so it must be permanently disabled from future writes.
3876 * Reading from NMI is a oneshot deal.
3878 if (raw_spin_trylock(&cpu_buffer->reader_lock))
3881 /* Continue without locking, but disable the ring buffer */
3882 atomic_inc(&cpu_buffer->record_disabled);
3887 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
3890 raw_spin_unlock(&cpu_buffer->reader_lock);
3895 * ring_buffer_peek - peek at the next event to be read
3896 * @buffer: The ring buffer to read
3897 * @cpu: The cpu to peak at
3898 * @ts: The timestamp counter of this event.
3899 * @lost_events: a variable to store if events were lost (may be NULL)
3901 * This will return the event that will be read next, but does
3902 * not consume the data.
3904 struct ring_buffer_event *
3905 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3906 unsigned long *lost_events)
3908 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3909 struct ring_buffer_event *event;
3910 unsigned long flags;
3913 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3917 local_irq_save(flags);
3918 dolock = rb_reader_lock(cpu_buffer);
3919 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3920 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3921 rb_advance_reader(cpu_buffer);
3922 rb_reader_unlock(cpu_buffer, dolock);
3923 local_irq_restore(flags);
3925 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3932 * ring_buffer_iter_peek - peek at the next event to be read
3933 * @iter: The ring buffer iterator
3934 * @ts: The timestamp counter of this event.
3936 * This will return the event that will be read next, but does
3937 * not increment the iterator.
3939 struct ring_buffer_event *
3940 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3942 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3943 struct ring_buffer_event *event;
3944 unsigned long flags;
3947 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3948 event = rb_iter_peek(iter, ts);
3949 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3951 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3958 * ring_buffer_consume - return an event and consume it
3959 * @buffer: The ring buffer to get the next event from
3960 * @cpu: the cpu to read the buffer from
3961 * @ts: a variable to store the timestamp (may be NULL)
3962 * @lost_events: a variable to store if events were lost (may be NULL)
3964 * Returns the next event in the ring buffer, and that event is consumed.
3965 * Meaning, that sequential reads will keep returning a different event,
3966 * and eventually empty the ring buffer if the producer is slower.
3968 struct ring_buffer_event *
3969 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3970 unsigned long *lost_events)
3972 struct ring_buffer_per_cpu *cpu_buffer;
3973 struct ring_buffer_event *event = NULL;
3974 unsigned long flags;
3978 /* might be called in atomic */
3981 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3984 cpu_buffer = buffer->buffers[cpu];
3985 local_irq_save(flags);
3986 dolock = rb_reader_lock(cpu_buffer);
3988 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3990 cpu_buffer->lost_events = 0;
3991 rb_advance_reader(cpu_buffer);
3994 rb_reader_unlock(cpu_buffer, dolock);
3995 local_irq_restore(flags);
4000 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4005 EXPORT_SYMBOL_GPL(ring_buffer_consume);
4008 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4009 * @buffer: The ring buffer to read from
4010 * @cpu: The cpu buffer to iterate over
4012 * This performs the initial preparations necessary to iterate
4013 * through the buffer. Memory is allocated, buffer recording
4014 * is disabled, and the iterator pointer is returned to the caller.
4016 * Disabling buffer recordng prevents the reading from being
4017 * corrupted. This is not a consuming read, so a producer is not
4020 * After a sequence of ring_buffer_read_prepare calls, the user is
4021 * expected to make at least one call to ring_buffer_read_prepare_sync.
4022 * Afterwards, ring_buffer_read_start is invoked to get things going
4025 * This overall must be paired with ring_buffer_read_finish.
4027 struct ring_buffer_iter *
4028 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
4030 struct ring_buffer_per_cpu *cpu_buffer;
4031 struct ring_buffer_iter *iter;
4033 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4036 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
4040 cpu_buffer = buffer->buffers[cpu];
4042 iter->cpu_buffer = cpu_buffer;
4044 atomic_inc(&buffer->resize_disabled);
4045 atomic_inc(&cpu_buffer->record_disabled);
4049 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4052 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4054 * All previously invoked ring_buffer_read_prepare calls to prepare
4055 * iterators will be synchronized. Afterwards, read_buffer_read_start
4056 * calls on those iterators are allowed.
4059 ring_buffer_read_prepare_sync(void)
4061 synchronize_sched();
4063 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4066 * ring_buffer_read_start - start a non consuming read of the buffer
4067 * @iter: The iterator returned by ring_buffer_read_prepare
4069 * This finalizes the startup of an iteration through the buffer.
4070 * The iterator comes from a call to ring_buffer_read_prepare and
4071 * an intervening ring_buffer_read_prepare_sync must have been
4074 * Must be paired with ring_buffer_read_finish.
4077 ring_buffer_read_start(struct ring_buffer_iter *iter)
4079 struct ring_buffer_per_cpu *cpu_buffer;
4080 unsigned long flags;
4085 cpu_buffer = iter->cpu_buffer;
4087 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4088 arch_spin_lock(&cpu_buffer->lock);
4089 rb_iter_reset(iter);
4090 arch_spin_unlock(&cpu_buffer->lock);
4091 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4093 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4096 * ring_buffer_read_finish - finish reading the iterator of the buffer
4097 * @iter: The iterator retrieved by ring_buffer_start
4099 * This re-enables the recording to the buffer, and frees the
4103 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4105 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4106 unsigned long flags;
4109 * Ring buffer is disabled from recording, here's a good place
4110 * to check the integrity of the ring buffer.
4111 * Must prevent readers from trying to read, as the check
4112 * clears the HEAD page and readers require it.
4114 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4115 rb_check_pages(cpu_buffer);
4116 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4118 atomic_dec(&cpu_buffer->record_disabled);
4119 atomic_dec(&cpu_buffer->buffer->resize_disabled);
4122 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4125 * ring_buffer_read - read the next item in the ring buffer by the iterator
4126 * @iter: The ring buffer iterator
4127 * @ts: The time stamp of the event read.
4129 * This reads the next event in the ring buffer and increments the iterator.
4131 struct ring_buffer_event *
4132 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4134 struct ring_buffer_event *event;
4135 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4136 unsigned long flags;
4138 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4140 event = rb_iter_peek(iter, ts);
4144 if (event->type_len == RINGBUF_TYPE_PADDING)
4147 rb_advance_iter(iter);
4149 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4153 EXPORT_SYMBOL_GPL(ring_buffer_read);
4156 * ring_buffer_size - return the size of the ring buffer (in bytes)
4157 * @buffer: The ring buffer.
4159 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4162 * Earlier, this method returned
4163 * BUF_PAGE_SIZE * buffer->nr_pages
4164 * Since the nr_pages field is now removed, we have converted this to
4165 * return the per cpu buffer value.
4167 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4170 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4172 EXPORT_SYMBOL_GPL(ring_buffer_size);
4175 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4177 rb_head_page_deactivate(cpu_buffer);
4179 cpu_buffer->head_page
4180 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4181 local_set(&cpu_buffer->head_page->write, 0);
4182 local_set(&cpu_buffer->head_page->entries, 0);
4183 local_set(&cpu_buffer->head_page->page->commit, 0);
4185 cpu_buffer->head_page->read = 0;
4187 cpu_buffer->tail_page = cpu_buffer->head_page;
4188 cpu_buffer->commit_page = cpu_buffer->head_page;
4190 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4191 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4192 local_set(&cpu_buffer->reader_page->write, 0);
4193 local_set(&cpu_buffer->reader_page->entries, 0);
4194 local_set(&cpu_buffer->reader_page->page->commit, 0);
4195 cpu_buffer->reader_page->read = 0;
4197 local_set(&cpu_buffer->entries_bytes, 0);
4198 local_set(&cpu_buffer->overrun, 0);
4199 local_set(&cpu_buffer->commit_overrun, 0);
4200 local_set(&cpu_buffer->dropped_events, 0);
4201 local_set(&cpu_buffer->entries, 0);
4202 local_set(&cpu_buffer->committing, 0);
4203 local_set(&cpu_buffer->commits, 0);
4204 cpu_buffer->read = 0;
4205 cpu_buffer->read_bytes = 0;
4207 cpu_buffer->write_stamp = 0;
4208 cpu_buffer->read_stamp = 0;
4210 cpu_buffer->lost_events = 0;
4211 cpu_buffer->last_overrun = 0;
4213 rb_head_page_activate(cpu_buffer);
4217 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4218 * @buffer: The ring buffer to reset a per cpu buffer of
4219 * @cpu: The CPU buffer to be reset
4221 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4223 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4224 unsigned long flags;
4226 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4229 atomic_inc(&buffer->resize_disabled);
4230 atomic_inc(&cpu_buffer->record_disabled);
4232 /* Make sure all commits have finished */
4233 synchronize_sched();
4235 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4237 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4240 arch_spin_lock(&cpu_buffer->lock);
4242 rb_reset_cpu(cpu_buffer);
4244 arch_spin_unlock(&cpu_buffer->lock);
4247 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4249 atomic_dec(&cpu_buffer->record_disabled);
4250 atomic_dec(&buffer->resize_disabled);
4252 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4255 * ring_buffer_reset - reset a ring buffer
4256 * @buffer: The ring buffer to reset all cpu buffers
4258 void ring_buffer_reset(struct ring_buffer *buffer)
4262 for_each_buffer_cpu(buffer, cpu)
4263 ring_buffer_reset_cpu(buffer, cpu);
4265 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4268 * rind_buffer_empty - is the ring buffer empty?
4269 * @buffer: The ring buffer to test
4271 bool ring_buffer_empty(struct ring_buffer *buffer)
4273 struct ring_buffer_per_cpu *cpu_buffer;
4274 unsigned long flags;
4279 /* yes this is racy, but if you don't like the race, lock the buffer */
4280 for_each_buffer_cpu(buffer, cpu) {
4281 cpu_buffer = buffer->buffers[cpu];
4282 local_irq_save(flags);
4283 dolock = rb_reader_lock(cpu_buffer);
4284 ret = rb_per_cpu_empty(cpu_buffer);
4285 rb_reader_unlock(cpu_buffer, dolock);
4286 local_irq_restore(flags);
4294 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4297 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4298 * @buffer: The ring buffer
4299 * @cpu: The CPU buffer to test
4301 bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4303 struct ring_buffer_per_cpu *cpu_buffer;
4304 unsigned long flags;
4308 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4311 cpu_buffer = buffer->buffers[cpu];
4312 local_irq_save(flags);
4313 dolock = rb_reader_lock(cpu_buffer);
4314 ret = rb_per_cpu_empty(cpu_buffer);
4315 rb_reader_unlock(cpu_buffer, dolock);
4316 local_irq_restore(flags);
4320 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4322 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4324 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4325 * @buffer_a: One buffer to swap with
4326 * @buffer_b: The other buffer to swap with
4328 * This function is useful for tracers that want to take a "snapshot"
4329 * of a CPU buffer and has another back up buffer lying around.
4330 * it is expected that the tracer handles the cpu buffer not being
4331 * used at the moment.
4333 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4334 struct ring_buffer *buffer_b, int cpu)
4336 struct ring_buffer_per_cpu *cpu_buffer_a;
4337 struct ring_buffer_per_cpu *cpu_buffer_b;
4340 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4341 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4344 cpu_buffer_a = buffer_a->buffers[cpu];
4345 cpu_buffer_b = buffer_b->buffers[cpu];
4347 /* At least make sure the two buffers are somewhat the same */
4348 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4353 if (atomic_read(&buffer_a->record_disabled))
4356 if (atomic_read(&buffer_b->record_disabled))
4359 if (atomic_read(&cpu_buffer_a->record_disabled))
4362 if (atomic_read(&cpu_buffer_b->record_disabled))
4366 * We can't do a synchronize_sched here because this
4367 * function can be called in atomic context.
4368 * Normally this will be called from the same CPU as cpu.
4369 * If not it's up to the caller to protect this.
4371 atomic_inc(&cpu_buffer_a->record_disabled);
4372 atomic_inc(&cpu_buffer_b->record_disabled);
4375 if (local_read(&cpu_buffer_a->committing))
4377 if (local_read(&cpu_buffer_b->committing))
4380 buffer_a->buffers[cpu] = cpu_buffer_b;
4381 buffer_b->buffers[cpu] = cpu_buffer_a;
4383 cpu_buffer_b->buffer = buffer_a;
4384 cpu_buffer_a->buffer = buffer_b;
4389 atomic_dec(&cpu_buffer_a->record_disabled);
4390 atomic_dec(&cpu_buffer_b->record_disabled);
4394 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4395 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4398 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4399 * @buffer: the buffer to allocate for.
4400 * @cpu: the cpu buffer to allocate.
4402 * This function is used in conjunction with ring_buffer_read_page.
4403 * When reading a full page from the ring buffer, these functions
4404 * can be used to speed up the process. The calling function should
4405 * allocate a few pages first with this function. Then when it
4406 * needs to get pages from the ring buffer, it passes the result
4407 * of this function into ring_buffer_read_page, which will swap
4408 * the page that was allocated, with the read page of the buffer.
4411 * The page allocated, or NULL on error.
4413 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4415 struct buffer_data_page *bpage;
4418 page = alloc_pages_node(cpu_to_node(cpu),
4419 GFP_KERNEL | __GFP_NORETRY, 0);
4423 bpage = page_address(page);
4425 rb_init_page(bpage);
4429 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4432 * ring_buffer_free_read_page - free an allocated read page
4433 * @buffer: the buffer the page was allocate for
4434 * @data: the page to free
4436 * Free a page allocated from ring_buffer_alloc_read_page.
4438 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4440 free_page((unsigned long)data);
4442 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4445 * ring_buffer_read_page - extract a page from the ring buffer
4446 * @buffer: buffer to extract from
4447 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4448 * @len: amount to extract
4449 * @cpu: the cpu of the buffer to extract
4450 * @full: should the extraction only happen when the page is full.
4452 * This function will pull out a page from the ring buffer and consume it.
4453 * @data_page must be the address of the variable that was returned
4454 * from ring_buffer_alloc_read_page. This is because the page might be used
4455 * to swap with a page in the ring buffer.
4458 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4461 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4463 * process_page(rpage, ret);
4465 * When @full is set, the function will not return true unless
4466 * the writer is off the reader page.
4468 * Note: it is up to the calling functions to handle sleeps and wakeups.
4469 * The ring buffer can be used anywhere in the kernel and can not
4470 * blindly call wake_up. The layer that uses the ring buffer must be
4471 * responsible for that.
4474 * >=0 if data has been transferred, returns the offset of consumed data.
4475 * <0 if no data has been transferred.
4477 int ring_buffer_read_page(struct ring_buffer *buffer,
4478 void **data_page, size_t len, int cpu, int full)
4480 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4481 struct ring_buffer_event *event;
4482 struct buffer_data_page *bpage;
4483 struct buffer_page *reader;
4484 unsigned long missed_events;
4485 unsigned long flags;
4486 unsigned int commit;
4491 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4495 * If len is not big enough to hold the page header, then
4496 * we can not copy anything.
4498 if (len <= BUF_PAGE_HDR_SIZE)
4501 len -= BUF_PAGE_HDR_SIZE;
4510 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4512 reader = rb_get_reader_page(cpu_buffer);
4516 event = rb_reader_event(cpu_buffer);
4518 read = reader->read;
4519 commit = rb_page_commit(reader);
4521 /* Check if any events were dropped */
4522 missed_events = cpu_buffer->lost_events;
4525 * If this page has been partially read or
4526 * if len is not big enough to read the rest of the page or
4527 * a writer is still on the page, then
4528 * we must copy the data from the page to the buffer.
4529 * Otherwise, we can simply swap the page with the one passed in.
4531 if (read || (len < (commit - read)) ||
4532 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4533 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4534 unsigned int rpos = read;
4535 unsigned int pos = 0;
4541 if (len > (commit - read))
4542 len = (commit - read);
4544 /* Always keep the time extend and data together */
4545 size = rb_event_ts_length(event);
4550 /* save the current timestamp, since the user will need it */
4551 save_timestamp = cpu_buffer->read_stamp;
4553 /* Need to copy one event at a time */
4555 /* We need the size of one event, because
4556 * rb_advance_reader only advances by one event,
4557 * whereas rb_event_ts_length may include the size of
4558 * one or two events.
4559 * We have already ensured there's enough space if this
4560 * is a time extend. */
4561 size = rb_event_length(event);
4562 memcpy(bpage->data + pos, rpage->data + rpos, size);
4566 rb_advance_reader(cpu_buffer);
4567 rpos = reader->read;
4573 event = rb_reader_event(cpu_buffer);
4574 /* Always keep the time extend and data together */
4575 size = rb_event_ts_length(event);
4576 } while (len >= size);
4579 local_set(&bpage->commit, pos);
4580 bpage->time_stamp = save_timestamp;
4582 /* we copied everything to the beginning */
4585 /* update the entry counter */
4586 cpu_buffer->read += rb_page_entries(reader);
4587 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4589 /* swap the pages */
4590 rb_init_page(bpage);
4591 bpage = reader->page;
4592 reader->page = *data_page;
4593 local_set(&reader->write, 0);
4594 local_set(&reader->entries, 0);
4599 * Use the real_end for the data size,
4600 * This gives us a chance to store the lost events
4603 if (reader->real_end)
4604 local_set(&bpage->commit, reader->real_end);
4608 cpu_buffer->lost_events = 0;
4610 commit = local_read(&bpage->commit);
4612 * Set a flag in the commit field if we lost events
4614 if (missed_events) {
4615 /* If there is room at the end of the page to save the
4616 * missed events, then record it there.
4618 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4619 memcpy(&bpage->data[commit], &missed_events,
4620 sizeof(missed_events));
4621 local_add(RB_MISSED_STORED, &bpage->commit);
4622 commit += sizeof(missed_events);
4624 local_add(RB_MISSED_EVENTS, &bpage->commit);
4628 * This page may be off to user land. Zero it out here.
4630 if (commit < BUF_PAGE_SIZE)
4631 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4634 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4639 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4641 #ifdef CONFIG_HOTPLUG_CPU
4642 static int rb_cpu_notify(struct notifier_block *self,
4643 unsigned long action, void *hcpu)
4645 struct ring_buffer *buffer =
4646 container_of(self, struct ring_buffer, cpu_notify);
4647 long cpu = (long)hcpu;
4650 unsigned long nr_pages;
4653 case CPU_UP_PREPARE:
4654 case CPU_UP_PREPARE_FROZEN:
4655 if (cpumask_test_cpu(cpu, buffer->cpumask))
4660 /* check if all cpu sizes are same */
4661 for_each_buffer_cpu(buffer, cpu_i) {
4662 /* fill in the size from first enabled cpu */
4664 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4665 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4670 /* allocate minimum pages, user can later expand it */
4673 buffer->buffers[cpu] =
4674 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4675 if (!buffer->buffers[cpu]) {
4676 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4681 cpumask_set_cpu(cpu, buffer->cpumask);
4683 case CPU_DOWN_PREPARE:
4684 case CPU_DOWN_PREPARE_FROZEN:
4687 * If we were to free the buffer, then the user would
4688 * lose any trace that was in the buffer.
4698 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4700 * This is a basic integrity check of the ring buffer.
4701 * Late in the boot cycle this test will run when configured in.
4702 * It will kick off a thread per CPU that will go into a loop
4703 * writing to the per cpu ring buffer various sizes of data.
4704 * Some of the data will be large items, some small.
4706 * Another thread is created that goes into a spin, sending out
4707 * IPIs to the other CPUs to also write into the ring buffer.
4708 * this is to test the nesting ability of the buffer.
4710 * Basic stats are recorded and reported. If something in the
4711 * ring buffer should happen that's not expected, a big warning
4712 * is displayed and all ring buffers are disabled.
4714 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4716 struct rb_test_data {
4717 struct ring_buffer *buffer;
4718 unsigned long events;
4719 unsigned long bytes_written;
4720 unsigned long bytes_alloc;
4721 unsigned long bytes_dropped;
4722 unsigned long events_nested;
4723 unsigned long bytes_written_nested;
4724 unsigned long bytes_alloc_nested;
4725 unsigned long bytes_dropped_nested;
4726 int min_size_nested;
4727 int max_size_nested;
4734 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4737 #define RB_TEST_BUFFER_SIZE 1048576
4739 static char rb_string[] __initdata =
4740 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4741 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4742 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4744 static bool rb_test_started __initdata;
4751 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4753 struct ring_buffer_event *event;
4754 struct rb_item *item;
4761 /* Have nested writes different that what is written */
4762 cnt = data->cnt + (nested ? 27 : 0);
4764 /* Multiply cnt by ~e, to make some unique increment */
4765 size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4767 len = size + sizeof(struct rb_item);
4769 started = rb_test_started;
4770 /* read rb_test_started before checking buffer enabled */
4773 event = ring_buffer_lock_reserve(data->buffer, len);
4775 /* Ignore dropped events before test starts. */
4778 data->bytes_dropped += len;
4780 data->bytes_dropped_nested += len;
4785 event_len = ring_buffer_event_length(event);
4787 if (RB_WARN_ON(data->buffer, event_len < len))
4790 item = ring_buffer_event_data(event);
4792 memcpy(item->str, rb_string, size);
4795 data->bytes_alloc_nested += event_len;
4796 data->bytes_written_nested += len;
4797 data->events_nested++;
4798 if (!data->min_size_nested || len < data->min_size_nested)
4799 data->min_size_nested = len;
4800 if (len > data->max_size_nested)
4801 data->max_size_nested = len;
4803 data->bytes_alloc += event_len;
4804 data->bytes_written += len;
4806 if (!data->min_size || len < data->min_size)
4807 data->max_size = len;
4808 if (len > data->max_size)
4809 data->max_size = len;
4813 ring_buffer_unlock_commit(data->buffer, event);
4818 static __init int rb_test(void *arg)
4820 struct rb_test_data *data = arg;
4822 while (!kthread_should_stop()) {
4823 rb_write_something(data, false);
4826 set_current_state(TASK_INTERRUPTIBLE);
4827 /* Now sleep between a min of 100-300us and a max of 1ms */
4828 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4834 static __init void rb_ipi(void *ignore)
4836 struct rb_test_data *data;
4837 int cpu = smp_processor_id();
4839 data = &rb_data[cpu];
4840 rb_write_something(data, true);
4843 static __init int rb_hammer_test(void *arg)
4845 while (!kthread_should_stop()) {
4847 /* Send an IPI to all cpus to write data! */
4848 smp_call_function(rb_ipi, NULL, 1);
4849 /* No sleep, but for non preempt, let others run */
4856 static __init int test_ringbuffer(void)
4858 struct task_struct *rb_hammer;
4859 struct ring_buffer *buffer;
4863 pr_info("Running ring buffer tests...\n");
4865 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4866 if (WARN_ON(!buffer))
4869 /* Disable buffer so that threads can't write to it yet */
4870 ring_buffer_record_off(buffer);
4872 for_each_online_cpu(cpu) {
4873 rb_data[cpu].buffer = buffer;
4874 rb_data[cpu].cpu = cpu;
4875 rb_data[cpu].cnt = cpu;
4876 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4877 "rbtester/%d", cpu);
4878 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
4879 pr_cont("FAILED\n");
4880 ret = PTR_ERR(rb_threads[cpu]);
4884 kthread_bind(rb_threads[cpu], cpu);
4885 wake_up_process(rb_threads[cpu]);
4888 /* Now create the rb hammer! */
4889 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4890 if (WARN_ON(IS_ERR(rb_hammer))) {
4891 pr_cont("FAILED\n");
4892 ret = PTR_ERR(rb_hammer);
4896 ring_buffer_record_on(buffer);
4898 * Show buffer is enabled before setting rb_test_started.
4899 * Yes there's a small race window where events could be
4900 * dropped and the thread wont catch it. But when a ring
4901 * buffer gets enabled, there will always be some kind of
4902 * delay before other CPUs see it. Thus, we don't care about
4903 * those dropped events. We care about events dropped after
4904 * the threads see that the buffer is active.
4907 rb_test_started = true;
4909 set_current_state(TASK_INTERRUPTIBLE);
4910 /* Just run for 10 seconds */;
4911 schedule_timeout(10 * HZ);
4913 kthread_stop(rb_hammer);
4916 for_each_online_cpu(cpu) {
4917 if (!rb_threads[cpu])
4919 kthread_stop(rb_threads[cpu]);
4922 ring_buffer_free(buffer);
4927 pr_info("finished\n");
4928 for_each_online_cpu(cpu) {
4929 struct ring_buffer_event *event;
4930 struct rb_test_data *data = &rb_data[cpu];
4931 struct rb_item *item;
4932 unsigned long total_events;
4933 unsigned long total_dropped;
4934 unsigned long total_written;
4935 unsigned long total_alloc;
4936 unsigned long total_read = 0;
4937 unsigned long total_size = 0;
4938 unsigned long total_len = 0;
4939 unsigned long total_lost = 0;
4942 int small_event_size;
4946 total_events = data->events + data->events_nested;
4947 total_written = data->bytes_written + data->bytes_written_nested;
4948 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
4949 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
4951 big_event_size = data->max_size + data->max_size_nested;
4952 small_event_size = data->min_size + data->min_size_nested;
4954 pr_info("CPU %d:\n", cpu);
4955 pr_info(" events: %ld\n", total_events);
4956 pr_info(" dropped bytes: %ld\n", total_dropped);
4957 pr_info(" alloced bytes: %ld\n", total_alloc);
4958 pr_info(" written bytes: %ld\n", total_written);
4959 pr_info(" biggest event: %d\n", big_event_size);
4960 pr_info(" smallest event: %d\n", small_event_size);
4962 if (RB_WARN_ON(buffer, total_dropped))
4967 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
4969 item = ring_buffer_event_data(event);
4970 total_len += ring_buffer_event_length(event);
4971 total_size += item->size + sizeof(struct rb_item);
4972 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
4973 pr_info("FAILED!\n");
4974 pr_info("buffer had: %.*s\n", item->size, item->str);
4975 pr_info("expected: %.*s\n", item->size, rb_string);
4976 RB_WARN_ON(buffer, 1);
4987 pr_info(" read events: %ld\n", total_read);
4988 pr_info(" lost events: %ld\n", total_lost);
4989 pr_info(" total events: %ld\n", total_lost + total_read);
4990 pr_info(" recorded len bytes: %ld\n", total_len);
4991 pr_info(" recorded size bytes: %ld\n", total_size);
4993 pr_info(" With dropped events, record len and size may not match\n"
4994 " alloced and written from above\n");
4996 if (RB_WARN_ON(buffer, total_len != total_alloc ||
4997 total_size != total_written))
5000 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5006 pr_info("Ring buffer PASSED!\n");
5008 ring_buffer_free(buffer);
5012 late_initcall(test_ringbuffer);
5013 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */