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/sched/clock.h>
10 #include <linux/trace_seq.h>
11 #include <linux/spinlock.h>
12 #include <linux/irq_work.h>
13 #include <linux/uaccess.h>
14 #include <linux/hardirq.h>
15 #include <linux/kthread.h> /* for self test */
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 */
248 static __always_inline void *
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 #define RB_MISSED_FLAGS (RB_MISSED_EVENTS|RB_MISSED_STORED)
285 struct buffer_data_page {
286 u64 time_stamp; /* page time stamp */
287 local_t commit; /* write committed index */
288 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
292 * Note, the buffer_page list must be first. The buffer pages
293 * are allocated in cache lines, which means that each buffer
294 * page will be at the beginning of a cache line, and thus
295 * the least significant bits will be zero. We use this to
296 * add flags in the list struct pointers, to make the ring buffer
300 struct list_head list; /* list of buffer pages */
301 local_t write; /* index for next write */
302 unsigned read; /* index for next read */
303 local_t entries; /* entries on this page */
304 unsigned long real_end; /* real end of data */
305 struct buffer_data_page *page; /* Actual data page */
309 * The buffer page counters, write and entries, must be reset
310 * atomically when crossing page boundaries. To synchronize this
311 * update, two counters are inserted into the number. One is
312 * the actual counter for the write position or count on the page.
314 * The other is a counter of updaters. Before an update happens
315 * the update partition of the counter is incremented. This will
316 * allow the updater to update the counter atomically.
318 * The counter is 20 bits, and the state data is 12.
320 #define RB_WRITE_MASK 0xfffff
321 #define RB_WRITE_INTCNT (1 << 20)
323 static void rb_init_page(struct buffer_data_page *bpage)
325 local_set(&bpage->commit, 0);
329 * ring_buffer_page_len - the size of data on the page.
330 * @page: The page to read
332 * Returns the amount of data on the page, including buffer page header.
334 size_t ring_buffer_page_len(void *page)
336 struct buffer_data_page *bpage = page;
338 return (local_read(&bpage->commit) & ~RB_MISSED_FLAGS)
343 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
346 static void free_buffer_page(struct buffer_page *bpage)
348 free_page((unsigned long)bpage->page);
353 * We need to fit the time_stamp delta into 27 bits.
355 static inline int test_time_stamp(u64 delta)
357 if (delta & TS_DELTA_TEST)
362 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
364 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
365 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
367 int ring_buffer_print_page_header(struct trace_seq *s)
369 struct buffer_data_page field;
371 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
372 "offset:0;\tsize:%u;\tsigned:%u;\n",
373 (unsigned int)sizeof(field.time_stamp),
374 (unsigned int)is_signed_type(u64));
376 trace_seq_printf(s, "\tfield: local_t commit;\t"
377 "offset:%u;\tsize:%u;\tsigned:%u;\n",
378 (unsigned int)offsetof(typeof(field), commit),
379 (unsigned int)sizeof(field.commit),
380 (unsigned int)is_signed_type(long));
382 trace_seq_printf(s, "\tfield: int overwrite;\t"
383 "offset:%u;\tsize:%u;\tsigned:%u;\n",
384 (unsigned int)offsetof(typeof(field), commit),
386 (unsigned int)is_signed_type(long));
388 trace_seq_printf(s, "\tfield: char data;\t"
389 "offset:%u;\tsize:%u;\tsigned:%u;\n",
390 (unsigned int)offsetof(typeof(field), data),
391 (unsigned int)BUF_PAGE_SIZE,
392 (unsigned int)is_signed_type(char));
394 return !trace_seq_has_overflowed(s);
398 struct irq_work work;
399 wait_queue_head_t waiters;
400 wait_queue_head_t full_waiters;
401 bool waiters_pending;
402 bool full_waiters_pending;
407 * Structure to hold event state and handle nested events.
409 struct rb_event_info {
412 unsigned long length;
413 struct buffer_page *tail_page;
418 * Used for which event context the event is in.
424 * See trace_recursive_lock() comment below for more details.
435 * head_page == tail_page && head == tail then buffer is empty.
437 struct ring_buffer_per_cpu {
439 atomic_t record_disabled;
440 struct ring_buffer *buffer;
441 raw_spinlock_t reader_lock; /* serialize readers */
442 arch_spinlock_t lock;
443 struct lock_class_key lock_key;
444 struct buffer_data_page *free_page;
445 unsigned long nr_pages;
446 unsigned int current_context;
447 struct list_head *pages;
448 struct buffer_page *head_page; /* read from head */
449 struct buffer_page *tail_page; /* write to tail */
450 struct buffer_page *commit_page; /* committed pages */
451 struct buffer_page *reader_page;
452 unsigned long lost_events;
453 unsigned long last_overrun;
454 local_t entries_bytes;
457 local_t commit_overrun;
458 local_t dropped_events;
462 unsigned long read_bytes;
465 /* ring buffer pages to update, > 0 to add, < 0 to remove */
466 long nr_pages_to_update;
467 struct list_head new_pages; /* new pages to add */
468 struct work_struct update_pages_work;
469 struct completion update_done;
471 struct rb_irq_work irq_work;
477 atomic_t record_disabled;
478 atomic_t resize_disabled;
479 cpumask_var_t cpumask;
481 struct lock_class_key *reader_lock_key;
485 struct ring_buffer_per_cpu **buffers;
487 struct hlist_node node;
490 struct rb_irq_work irq_work;
493 struct ring_buffer_iter {
494 struct ring_buffer_per_cpu *cpu_buffer;
496 struct buffer_page *head_page;
497 struct buffer_page *cache_reader_page;
498 unsigned long cache_read;
503 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
505 * Schedules a delayed work to wake up any task that is blocked on the
506 * ring buffer waiters queue.
508 static void rb_wake_up_waiters(struct irq_work *work)
510 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
512 wake_up_all(&rbwork->waiters);
513 if (rbwork->wakeup_full) {
514 rbwork->wakeup_full = false;
515 wake_up_all(&rbwork->full_waiters);
520 * ring_buffer_wait - wait for input to the ring buffer
521 * @buffer: buffer to wait on
522 * @cpu: the cpu buffer to wait on
523 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
525 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
526 * as data is added to any of the @buffer's cpu buffers. Otherwise
527 * it will wait for data to be added to a specific cpu buffer.
529 int ring_buffer_wait(struct ring_buffer *buffer, int cpu, bool full)
531 struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
533 struct rb_irq_work *work;
537 * Depending on what the caller is waiting for, either any
538 * data in any cpu buffer, or a specific buffer, put the
539 * caller on the appropriate wait queue.
541 if (cpu == RING_BUFFER_ALL_CPUS) {
542 work = &buffer->irq_work;
543 /* Full only makes sense on per cpu reads */
546 if (!cpumask_test_cpu(cpu, buffer->cpumask))
548 cpu_buffer = buffer->buffers[cpu];
549 work = &cpu_buffer->irq_work;
555 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
557 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
560 * The events can happen in critical sections where
561 * checking a work queue can cause deadlocks.
562 * After adding a task to the queue, this flag is set
563 * only to notify events to try to wake up the queue
566 * We don't clear it even if the buffer is no longer
567 * empty. The flag only causes the next event to run
568 * irq_work to do the work queue wake up. The worse
569 * that can happen if we race with !trace_empty() is that
570 * an event will cause an irq_work to try to wake up
573 * There's no reason to protect this flag either, as
574 * the work queue and irq_work logic will do the necessary
575 * synchronization for the wake ups. The only thing
576 * that is necessary is that the wake up happens after
577 * a task has been queued. It's OK for spurious wake ups.
580 work->full_waiters_pending = true;
582 work->waiters_pending = true;
584 if (signal_pending(current)) {
589 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
592 if (cpu != RING_BUFFER_ALL_CPUS &&
593 !ring_buffer_empty_cpu(buffer, cpu)) {
600 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
601 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
602 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
612 finish_wait(&work->full_waiters, &wait);
614 finish_wait(&work->waiters, &wait);
620 * ring_buffer_poll_wait - poll on buffer input
621 * @buffer: buffer to wait on
622 * @cpu: the cpu buffer to wait on
623 * @filp: the file descriptor
624 * @poll_table: The poll descriptor
626 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
627 * as data is added to any of the @buffer's cpu buffers. Otherwise
628 * it will wait for data to be added to a specific cpu buffer.
630 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
633 int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
634 struct file *filp, poll_table *poll_table)
636 struct ring_buffer_per_cpu *cpu_buffer;
637 struct rb_irq_work *work;
639 if (cpu == RING_BUFFER_ALL_CPUS)
640 work = &buffer->irq_work;
642 if (!cpumask_test_cpu(cpu, buffer->cpumask))
645 cpu_buffer = buffer->buffers[cpu];
646 work = &cpu_buffer->irq_work;
649 poll_wait(filp, &work->waiters, poll_table);
650 work->waiters_pending = true;
652 * There's a tight race between setting the waiters_pending and
653 * checking if the ring buffer is empty. Once the waiters_pending bit
654 * is set, the next event will wake the task up, but we can get stuck
655 * if there's only a single event in.
657 * FIXME: Ideally, we need a memory barrier on the writer side as well,
658 * but adding a memory barrier to all events will cause too much of a
659 * performance hit in the fast path. We only need a memory barrier when
660 * the buffer goes from empty to having content. But as this race is
661 * extremely small, and it's not a problem if another event comes in, we
666 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
667 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
668 return POLLIN | POLLRDNORM;
672 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
673 #define RB_WARN_ON(b, cond) \
675 int _____ret = unlikely(cond); \
677 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
678 struct ring_buffer_per_cpu *__b = \
680 atomic_inc(&__b->buffer->record_disabled); \
682 atomic_inc(&b->record_disabled); \
688 /* Up this if you want to test the TIME_EXTENTS and normalization */
689 #define DEBUG_SHIFT 0
691 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
693 /* shift to debug/test normalization and TIME_EXTENTS */
694 return buffer->clock() << DEBUG_SHIFT;
697 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
701 preempt_disable_notrace();
702 time = rb_time_stamp(buffer);
703 preempt_enable_no_resched_notrace();
707 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
709 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
712 /* Just stupid testing the normalize function and deltas */
715 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
718 * Making the ring buffer lockless makes things tricky.
719 * Although writes only happen on the CPU that they are on,
720 * and they only need to worry about interrupts. Reads can
723 * The reader page is always off the ring buffer, but when the
724 * reader finishes with a page, it needs to swap its page with
725 * a new one from the buffer. The reader needs to take from
726 * the head (writes go to the tail). But if a writer is in overwrite
727 * mode and wraps, it must push the head page forward.
729 * Here lies the problem.
731 * The reader must be careful to replace only the head page, and
732 * not another one. As described at the top of the file in the
733 * ASCII art, the reader sets its old page to point to the next
734 * page after head. It then sets the page after head to point to
735 * the old reader page. But if the writer moves the head page
736 * during this operation, the reader could end up with the tail.
738 * We use cmpxchg to help prevent this race. We also do something
739 * special with the page before head. We set the LSB to 1.
741 * When the writer must push the page forward, it will clear the
742 * bit that points to the head page, move the head, and then set
743 * the bit that points to the new head page.
745 * We also don't want an interrupt coming in and moving the head
746 * page on another writer. Thus we use the second LSB to catch
749 * head->list->prev->next bit 1 bit 0
752 * Points to head page 0 1
755 * Note we can not trust the prev pointer of the head page, because:
757 * +----+ +-----+ +-----+
758 * | |------>| T |---X--->| N |
760 * +----+ +-----+ +-----+
763 * +----------| R |----------+ |
767 * Key: ---X--> HEAD flag set in pointer
772 * (see __rb_reserve_next() to see where this happens)
774 * What the above shows is that the reader just swapped out
775 * the reader page with a page in the buffer, but before it
776 * could make the new header point back to the new page added
777 * it was preempted by a writer. The writer moved forward onto
778 * the new page added by the reader and is about to move forward
781 * You can see, it is legitimate for the previous pointer of
782 * the head (or any page) not to point back to itself. But only
786 #define RB_PAGE_NORMAL 0UL
787 #define RB_PAGE_HEAD 1UL
788 #define RB_PAGE_UPDATE 2UL
791 #define RB_FLAG_MASK 3UL
793 /* PAGE_MOVED is not part of the mask */
794 #define RB_PAGE_MOVED 4UL
797 * rb_list_head - remove any bit
799 static struct list_head *rb_list_head(struct list_head *list)
801 unsigned long val = (unsigned long)list;
803 return (struct list_head *)(val & ~RB_FLAG_MASK);
807 * rb_is_head_page - test if the given page is the head page
809 * Because the reader may move the head_page pointer, we can
810 * not trust what the head page is (it may be pointing to
811 * the reader page). But if the next page is a header page,
812 * its flags will be non zero.
815 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
816 struct buffer_page *page, struct list_head *list)
820 val = (unsigned long)list->next;
822 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
823 return RB_PAGE_MOVED;
825 return val & RB_FLAG_MASK;
831 * The unique thing about the reader page, is that, if the
832 * writer is ever on it, the previous pointer never points
833 * back to the reader page.
835 static bool rb_is_reader_page(struct buffer_page *page)
837 struct list_head *list = page->list.prev;
839 return rb_list_head(list->next) != &page->list;
843 * rb_set_list_to_head - set a list_head to be pointing to head.
845 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
846 struct list_head *list)
850 ptr = (unsigned long *)&list->next;
851 *ptr |= RB_PAGE_HEAD;
852 *ptr &= ~RB_PAGE_UPDATE;
856 * rb_head_page_activate - sets up head page
858 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
860 struct buffer_page *head;
862 head = cpu_buffer->head_page;
867 * Set the previous list pointer to have the HEAD flag.
869 rb_set_list_to_head(cpu_buffer, head->list.prev);
872 static void rb_list_head_clear(struct list_head *list)
874 unsigned long *ptr = (unsigned long *)&list->next;
876 *ptr &= ~RB_FLAG_MASK;
880 * rb_head_page_dactivate - clears head page ptr (for free list)
883 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
885 struct list_head *hd;
887 /* Go through the whole list and clear any pointers found. */
888 rb_list_head_clear(cpu_buffer->pages);
890 list_for_each(hd, cpu_buffer->pages)
891 rb_list_head_clear(hd);
894 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
895 struct buffer_page *head,
896 struct buffer_page *prev,
897 int old_flag, int new_flag)
899 struct list_head *list;
900 unsigned long val = (unsigned long)&head->list;
905 val &= ~RB_FLAG_MASK;
907 ret = cmpxchg((unsigned long *)&list->next,
908 val | old_flag, val | new_flag);
910 /* check if the reader took the page */
911 if ((ret & ~RB_FLAG_MASK) != val)
912 return RB_PAGE_MOVED;
914 return ret & RB_FLAG_MASK;
917 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
918 struct buffer_page *head,
919 struct buffer_page *prev,
922 return rb_head_page_set(cpu_buffer, head, prev,
923 old_flag, RB_PAGE_UPDATE);
926 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
927 struct buffer_page *head,
928 struct buffer_page *prev,
931 return rb_head_page_set(cpu_buffer, head, prev,
932 old_flag, RB_PAGE_HEAD);
935 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
936 struct buffer_page *head,
937 struct buffer_page *prev,
940 return rb_head_page_set(cpu_buffer, head, prev,
941 old_flag, RB_PAGE_NORMAL);
944 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
945 struct buffer_page **bpage)
947 struct list_head *p = rb_list_head((*bpage)->list.next);
949 *bpage = list_entry(p, struct buffer_page, list);
952 static struct buffer_page *
953 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
955 struct buffer_page *head;
956 struct buffer_page *page;
957 struct list_head *list;
960 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
964 list = cpu_buffer->pages;
965 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
968 page = head = cpu_buffer->head_page;
970 * It is possible that the writer moves the header behind
971 * where we started, and we miss in one loop.
972 * A second loop should grab the header, but we'll do
973 * three loops just because I'm paranoid.
975 for (i = 0; i < 3; i++) {
977 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
978 cpu_buffer->head_page = page;
981 rb_inc_page(cpu_buffer, &page);
982 } while (page != head);
985 RB_WARN_ON(cpu_buffer, 1);
990 static int rb_head_page_replace(struct buffer_page *old,
991 struct buffer_page *new)
993 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
997 val = *ptr & ~RB_FLAG_MASK;
1000 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1006 * rb_tail_page_update - move the tail page forward
1008 static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1009 struct buffer_page *tail_page,
1010 struct buffer_page *next_page)
1012 unsigned long old_entries;
1013 unsigned long old_write;
1016 * The tail page now needs to be moved forward.
1018 * We need to reset the tail page, but without messing
1019 * with possible erasing of data brought in by interrupts
1020 * that have moved the tail page and are currently on it.
1022 * We add a counter to the write field to denote this.
1024 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1025 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1028 * Just make sure we have seen our old_write and synchronize
1029 * with any interrupts that come in.
1034 * If the tail page is still the same as what we think
1035 * it is, then it is up to us to update the tail
1038 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1039 /* Zero the write counter */
1040 unsigned long val = old_write & ~RB_WRITE_MASK;
1041 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1044 * This will only succeed if an interrupt did
1045 * not come in and change it. In which case, we
1046 * do not want to modify it.
1048 * We add (void) to let the compiler know that we do not care
1049 * about the return value of these functions. We use the
1050 * cmpxchg to only update if an interrupt did not already
1051 * do it for us. If the cmpxchg fails, we don't care.
1053 (void)local_cmpxchg(&next_page->write, old_write, val);
1054 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1057 * No need to worry about races with clearing out the commit.
1058 * it only can increment when a commit takes place. But that
1059 * only happens in the outer most nested commit.
1061 local_set(&next_page->page->commit, 0);
1063 /* Again, either we update tail_page or an interrupt does */
1064 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1068 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1069 struct buffer_page *bpage)
1071 unsigned long val = (unsigned long)bpage;
1073 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1080 * rb_check_list - make sure a pointer to a list has the last bits zero
1082 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1083 struct list_head *list)
1085 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1087 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1093 * rb_check_pages - integrity check of buffer pages
1094 * @cpu_buffer: CPU buffer with pages to test
1096 * As a safety measure we check to make sure the data pages have not
1099 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1101 struct list_head *head = cpu_buffer->pages;
1102 struct buffer_page *bpage, *tmp;
1104 /* Reset the head page if it exists */
1105 if (cpu_buffer->head_page)
1106 rb_set_head_page(cpu_buffer);
1108 rb_head_page_deactivate(cpu_buffer);
1110 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1112 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1115 if (rb_check_list(cpu_buffer, head))
1118 list_for_each_entry_safe(bpage, tmp, head, list) {
1119 if (RB_WARN_ON(cpu_buffer,
1120 bpage->list.next->prev != &bpage->list))
1122 if (RB_WARN_ON(cpu_buffer,
1123 bpage->list.prev->next != &bpage->list))
1125 if (rb_check_list(cpu_buffer, &bpage->list))
1129 rb_head_page_activate(cpu_buffer);
1134 static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1136 struct buffer_page *bpage, *tmp;
1139 /* Check if the available memory is there first */
1140 i = si_mem_available();
1144 for (i = 0; i < nr_pages; i++) {
1147 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1148 * gracefully without invoking oom-killer and the system is not
1151 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1152 GFP_KERNEL | __GFP_RETRY_MAYFAIL,
1157 list_add(&bpage->list, pages);
1159 page = alloc_pages_node(cpu_to_node(cpu),
1160 GFP_KERNEL | __GFP_RETRY_MAYFAIL, 0);
1163 bpage->page = page_address(page);
1164 rb_init_page(bpage->page);
1170 list_for_each_entry_safe(bpage, tmp, pages, list) {
1171 list_del_init(&bpage->list);
1172 free_buffer_page(bpage);
1178 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1179 unsigned long nr_pages)
1185 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1189 * The ring buffer page list is a circular list that does not
1190 * start and end with a list head. All page list items point to
1193 cpu_buffer->pages = pages.next;
1196 cpu_buffer->nr_pages = nr_pages;
1198 rb_check_pages(cpu_buffer);
1203 static struct ring_buffer_per_cpu *
1204 rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1206 struct ring_buffer_per_cpu *cpu_buffer;
1207 struct buffer_page *bpage;
1211 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1212 GFP_KERNEL, cpu_to_node(cpu));
1216 cpu_buffer->cpu = cpu;
1217 cpu_buffer->buffer = buffer;
1218 raw_spin_lock_init(&cpu_buffer->reader_lock);
1219 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1220 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1221 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1222 init_completion(&cpu_buffer->update_done);
1223 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1224 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1225 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1227 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1228 GFP_KERNEL, cpu_to_node(cpu));
1230 goto fail_free_buffer;
1232 rb_check_bpage(cpu_buffer, bpage);
1234 cpu_buffer->reader_page = bpage;
1235 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1237 goto fail_free_reader;
1238 bpage->page = page_address(page);
1239 rb_init_page(bpage->page);
1241 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1242 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1244 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1246 goto fail_free_reader;
1248 cpu_buffer->head_page
1249 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1250 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1252 rb_head_page_activate(cpu_buffer);
1257 free_buffer_page(cpu_buffer->reader_page);
1264 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1266 struct list_head *head = cpu_buffer->pages;
1267 struct buffer_page *bpage, *tmp;
1269 free_buffer_page(cpu_buffer->reader_page);
1271 rb_head_page_deactivate(cpu_buffer);
1274 list_for_each_entry_safe(bpage, tmp, head, list) {
1275 list_del_init(&bpage->list);
1276 free_buffer_page(bpage);
1278 bpage = list_entry(head, struct buffer_page, list);
1279 free_buffer_page(bpage);
1286 * __ring_buffer_alloc - allocate a new ring_buffer
1287 * @size: the size in bytes per cpu that is needed.
1288 * @flags: attributes to set for the ring buffer.
1290 * Currently the only flag that is available is the RB_FL_OVERWRITE
1291 * flag. This flag means that the buffer will overwrite old data
1292 * when the buffer wraps. If this flag is not set, the buffer will
1293 * drop data when the tail hits the head.
1295 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1296 struct lock_class_key *key)
1298 struct ring_buffer *buffer;
1304 /* keep it in its own cache line */
1305 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1310 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1311 goto fail_free_buffer;
1313 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1314 buffer->flags = flags;
1315 buffer->clock = trace_clock_local;
1316 buffer->reader_lock_key = key;
1318 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1319 init_waitqueue_head(&buffer->irq_work.waiters);
1321 /* need at least two pages */
1325 buffer->cpus = nr_cpu_ids;
1327 bsize = sizeof(void *) * nr_cpu_ids;
1328 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1330 if (!buffer->buffers)
1331 goto fail_free_cpumask;
1333 cpu = raw_smp_processor_id();
1334 cpumask_set_cpu(cpu, buffer->cpumask);
1335 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1336 if (!buffer->buffers[cpu])
1337 goto fail_free_buffers;
1339 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1341 goto fail_free_buffers;
1343 mutex_init(&buffer->mutex);
1348 for_each_buffer_cpu(buffer, cpu) {
1349 if (buffer->buffers[cpu])
1350 rb_free_cpu_buffer(buffer->buffers[cpu]);
1352 kfree(buffer->buffers);
1355 free_cpumask_var(buffer->cpumask);
1361 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1364 * ring_buffer_free - free a ring buffer.
1365 * @buffer: the buffer to free.
1368 ring_buffer_free(struct ring_buffer *buffer)
1372 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1374 for_each_buffer_cpu(buffer, cpu)
1375 rb_free_cpu_buffer(buffer->buffers[cpu]);
1377 kfree(buffer->buffers);
1378 free_cpumask_var(buffer->cpumask);
1382 EXPORT_SYMBOL_GPL(ring_buffer_free);
1384 void ring_buffer_set_clock(struct ring_buffer *buffer,
1387 buffer->clock = clock;
1390 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1392 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1394 return local_read(&bpage->entries) & RB_WRITE_MASK;
1397 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1399 return local_read(&bpage->write) & RB_WRITE_MASK;
1403 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1405 struct list_head *tail_page, *to_remove, *next_page;
1406 struct buffer_page *to_remove_page, *tmp_iter_page;
1407 struct buffer_page *last_page, *first_page;
1408 unsigned long nr_removed;
1409 unsigned long head_bit;
1414 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1415 atomic_inc(&cpu_buffer->record_disabled);
1417 * We don't race with the readers since we have acquired the reader
1418 * lock. We also don't race with writers after disabling recording.
1419 * This makes it easy to figure out the first and the last page to be
1420 * removed from the list. We unlink all the pages in between including
1421 * the first and last pages. This is done in a busy loop so that we
1422 * lose the least number of traces.
1423 * The pages are freed after we restart recording and unlock readers.
1425 tail_page = &cpu_buffer->tail_page->list;
1428 * tail page might be on reader page, we remove the next page
1429 * from the ring buffer
1431 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1432 tail_page = rb_list_head(tail_page->next);
1433 to_remove = tail_page;
1435 /* start of pages to remove */
1436 first_page = list_entry(rb_list_head(to_remove->next),
1437 struct buffer_page, list);
1439 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1440 to_remove = rb_list_head(to_remove)->next;
1441 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1444 next_page = rb_list_head(to_remove)->next;
1447 * Now we remove all pages between tail_page and next_page.
1448 * Make sure that we have head_bit value preserved for the
1451 tail_page->next = (struct list_head *)((unsigned long)next_page |
1453 next_page = rb_list_head(next_page);
1454 next_page->prev = tail_page;
1456 /* make sure pages points to a valid page in the ring buffer */
1457 cpu_buffer->pages = next_page;
1459 /* update head page */
1461 cpu_buffer->head_page = list_entry(next_page,
1462 struct buffer_page, list);
1465 * change read pointer to make sure any read iterators reset
1468 cpu_buffer->read = 0;
1470 /* pages are removed, resume tracing and then free the pages */
1471 atomic_dec(&cpu_buffer->record_disabled);
1472 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1474 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1476 /* last buffer page to remove */
1477 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1479 tmp_iter_page = first_page;
1482 to_remove_page = tmp_iter_page;
1483 rb_inc_page(cpu_buffer, &tmp_iter_page);
1485 /* update the counters */
1486 page_entries = rb_page_entries(to_remove_page);
1489 * If something was added to this page, it was full
1490 * since it is not the tail page. So we deduct the
1491 * bytes consumed in ring buffer from here.
1492 * Increment overrun to account for the lost events.
1494 local_add(page_entries, &cpu_buffer->overrun);
1495 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1499 * We have already removed references to this list item, just
1500 * free up the buffer_page and its page
1502 free_buffer_page(to_remove_page);
1505 } while (to_remove_page != last_page);
1507 RB_WARN_ON(cpu_buffer, nr_removed);
1509 return nr_removed == 0;
1513 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1515 struct list_head *pages = &cpu_buffer->new_pages;
1516 int retries, success;
1518 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1520 * We are holding the reader lock, so the reader page won't be swapped
1521 * in the ring buffer. Now we are racing with the writer trying to
1522 * move head page and the tail page.
1523 * We are going to adapt the reader page update process where:
1524 * 1. We first splice the start and end of list of new pages between
1525 * the head page and its previous page.
1526 * 2. We cmpxchg the prev_page->next to point from head page to the
1527 * start of new pages list.
1528 * 3. Finally, we update the head->prev to the end of new list.
1530 * We will try this process 10 times, to make sure that we don't keep
1536 struct list_head *head_page, *prev_page, *r;
1537 struct list_head *last_page, *first_page;
1538 struct list_head *head_page_with_bit;
1540 head_page = &rb_set_head_page(cpu_buffer)->list;
1543 prev_page = head_page->prev;
1545 first_page = pages->next;
1546 last_page = pages->prev;
1548 head_page_with_bit = (struct list_head *)
1549 ((unsigned long)head_page | RB_PAGE_HEAD);
1551 last_page->next = head_page_with_bit;
1552 first_page->prev = prev_page;
1554 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1556 if (r == head_page_with_bit) {
1558 * yay, we replaced the page pointer to our new list,
1559 * now, we just have to update to head page's prev
1560 * pointer to point to end of list
1562 head_page->prev = last_page;
1569 INIT_LIST_HEAD(pages);
1571 * If we weren't successful in adding in new pages, warn and stop
1574 RB_WARN_ON(cpu_buffer, !success);
1575 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1577 /* free pages if they weren't inserted */
1579 struct buffer_page *bpage, *tmp;
1580 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1582 list_del_init(&bpage->list);
1583 free_buffer_page(bpage);
1589 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1593 if (cpu_buffer->nr_pages_to_update > 0)
1594 success = rb_insert_pages(cpu_buffer);
1596 success = rb_remove_pages(cpu_buffer,
1597 -cpu_buffer->nr_pages_to_update);
1600 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1603 static void update_pages_handler(struct work_struct *work)
1605 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1606 struct ring_buffer_per_cpu, update_pages_work);
1607 rb_update_pages(cpu_buffer);
1608 complete(&cpu_buffer->update_done);
1612 * ring_buffer_resize - resize the ring buffer
1613 * @buffer: the buffer to resize.
1614 * @size: the new size.
1615 * @cpu_id: the cpu buffer to resize
1617 * Minimum size is 2 * BUF_PAGE_SIZE.
1619 * Returns 0 on success and < 0 on failure.
1621 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1624 struct ring_buffer_per_cpu *cpu_buffer;
1625 unsigned long nr_pages;
1629 * Always succeed at resizing a non-existent buffer:
1634 /* Make sure the requested buffer exists */
1635 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1636 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1639 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1641 /* we need a minimum of two pages */
1645 size = nr_pages * BUF_PAGE_SIZE;
1648 * Don't succeed if resizing is disabled, as a reader might be
1649 * manipulating the ring buffer and is expecting a sane state while
1652 if (atomic_read(&buffer->resize_disabled))
1655 /* prevent another thread from changing buffer sizes */
1656 mutex_lock(&buffer->mutex);
1658 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1659 /* calculate the pages to update */
1660 for_each_buffer_cpu(buffer, cpu) {
1661 cpu_buffer = buffer->buffers[cpu];
1663 cpu_buffer->nr_pages_to_update = nr_pages -
1664 cpu_buffer->nr_pages;
1666 * nothing more to do for removing pages or no update
1668 if (cpu_buffer->nr_pages_to_update <= 0)
1671 * to add pages, make sure all new pages can be
1672 * allocated without receiving ENOMEM
1674 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1675 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1676 &cpu_buffer->new_pages, cpu)) {
1677 /* not enough memory for new pages */
1685 * Fire off all the required work handlers
1686 * We can't schedule on offline CPUs, but it's not necessary
1687 * since we can change their buffer sizes without any race.
1689 for_each_buffer_cpu(buffer, cpu) {
1690 cpu_buffer = buffer->buffers[cpu];
1691 if (!cpu_buffer->nr_pages_to_update)
1694 /* Can't run something on an offline CPU. */
1695 if (!cpu_online(cpu)) {
1696 rb_update_pages(cpu_buffer);
1697 cpu_buffer->nr_pages_to_update = 0;
1699 schedule_work_on(cpu,
1700 &cpu_buffer->update_pages_work);
1704 /* wait for all the updates to complete */
1705 for_each_buffer_cpu(buffer, cpu) {
1706 cpu_buffer = buffer->buffers[cpu];
1707 if (!cpu_buffer->nr_pages_to_update)
1710 if (cpu_online(cpu))
1711 wait_for_completion(&cpu_buffer->update_done);
1712 cpu_buffer->nr_pages_to_update = 0;
1717 /* Make sure this CPU has been intitialized */
1718 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1721 cpu_buffer = buffer->buffers[cpu_id];
1723 if (nr_pages == cpu_buffer->nr_pages)
1726 cpu_buffer->nr_pages_to_update = nr_pages -
1727 cpu_buffer->nr_pages;
1729 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1730 if (cpu_buffer->nr_pages_to_update > 0 &&
1731 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1732 &cpu_buffer->new_pages, cpu_id)) {
1739 /* Can't run something on an offline CPU. */
1740 if (!cpu_online(cpu_id))
1741 rb_update_pages(cpu_buffer);
1743 schedule_work_on(cpu_id,
1744 &cpu_buffer->update_pages_work);
1745 wait_for_completion(&cpu_buffer->update_done);
1748 cpu_buffer->nr_pages_to_update = 0;
1754 * The ring buffer resize can happen with the ring buffer
1755 * enabled, so that the update disturbs the tracing as little
1756 * as possible. But if the buffer is disabled, we do not need
1757 * to worry about that, and we can take the time to verify
1758 * that the buffer is not corrupt.
1760 if (atomic_read(&buffer->record_disabled)) {
1761 atomic_inc(&buffer->record_disabled);
1763 * Even though the buffer was disabled, we must make sure
1764 * that it is truly disabled before calling rb_check_pages.
1765 * There could have been a race between checking
1766 * record_disable and incrementing it.
1768 synchronize_sched();
1769 for_each_buffer_cpu(buffer, cpu) {
1770 cpu_buffer = buffer->buffers[cpu];
1771 rb_check_pages(cpu_buffer);
1773 atomic_dec(&buffer->record_disabled);
1776 mutex_unlock(&buffer->mutex);
1780 for_each_buffer_cpu(buffer, cpu) {
1781 struct buffer_page *bpage, *tmp;
1783 cpu_buffer = buffer->buffers[cpu];
1784 cpu_buffer->nr_pages_to_update = 0;
1786 if (list_empty(&cpu_buffer->new_pages))
1789 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1791 list_del_init(&bpage->list);
1792 free_buffer_page(bpage);
1795 mutex_unlock(&buffer->mutex);
1798 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1800 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1802 mutex_lock(&buffer->mutex);
1804 buffer->flags |= RB_FL_OVERWRITE;
1806 buffer->flags &= ~RB_FL_OVERWRITE;
1807 mutex_unlock(&buffer->mutex);
1809 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1811 static __always_inline void *
1812 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1814 return bpage->data + index;
1817 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1819 return bpage->page->data + index;
1822 static __always_inline struct ring_buffer_event *
1823 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1825 return __rb_page_index(cpu_buffer->reader_page,
1826 cpu_buffer->reader_page->read);
1829 static __always_inline struct ring_buffer_event *
1830 rb_iter_head_event(struct ring_buffer_iter *iter)
1832 return __rb_page_index(iter->head_page, iter->head);
1835 static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
1837 return local_read(&bpage->page->commit);
1840 /* Size is determined by what has been committed */
1841 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
1843 return rb_page_commit(bpage);
1846 static __always_inline unsigned
1847 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1849 return rb_page_commit(cpu_buffer->commit_page);
1852 static __always_inline unsigned
1853 rb_event_index(struct ring_buffer_event *event)
1855 unsigned long addr = (unsigned long)event;
1857 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1860 static void rb_inc_iter(struct ring_buffer_iter *iter)
1862 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1865 * The iterator could be on the reader page (it starts there).
1866 * But the head could have moved, since the reader was
1867 * found. Check for this case and assign the iterator
1868 * to the head page instead of next.
1870 if (iter->head_page == cpu_buffer->reader_page)
1871 iter->head_page = rb_set_head_page(cpu_buffer);
1873 rb_inc_page(cpu_buffer, &iter->head_page);
1875 iter->read_stamp = iter->head_page->page->time_stamp;
1880 * rb_handle_head_page - writer hit the head page
1882 * Returns: +1 to retry page
1887 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1888 struct buffer_page *tail_page,
1889 struct buffer_page *next_page)
1891 struct buffer_page *new_head;
1896 entries = rb_page_entries(next_page);
1899 * The hard part is here. We need to move the head
1900 * forward, and protect against both readers on
1901 * other CPUs and writers coming in via interrupts.
1903 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1907 * type can be one of four:
1908 * NORMAL - an interrupt already moved it for us
1909 * HEAD - we are the first to get here.
1910 * UPDATE - we are the interrupt interrupting
1912 * MOVED - a reader on another CPU moved the next
1913 * pointer to its reader page. Give up
1920 * We changed the head to UPDATE, thus
1921 * it is our responsibility to update
1924 local_add(entries, &cpu_buffer->overrun);
1925 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1928 * The entries will be zeroed out when we move the
1932 /* still more to do */
1935 case RB_PAGE_UPDATE:
1937 * This is an interrupt that interrupt the
1938 * previous update. Still more to do.
1941 case RB_PAGE_NORMAL:
1943 * An interrupt came in before the update
1944 * and processed this for us.
1945 * Nothing left to do.
1950 * The reader is on another CPU and just did
1951 * a swap with our next_page.
1956 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1961 * Now that we are here, the old head pointer is
1962 * set to UPDATE. This will keep the reader from
1963 * swapping the head page with the reader page.
1964 * The reader (on another CPU) will spin till
1967 * We just need to protect against interrupts
1968 * doing the job. We will set the next pointer
1969 * to HEAD. After that, we set the old pointer
1970 * to NORMAL, but only if it was HEAD before.
1971 * otherwise we are an interrupt, and only
1972 * want the outer most commit to reset it.
1974 new_head = next_page;
1975 rb_inc_page(cpu_buffer, &new_head);
1977 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1981 * Valid returns are:
1982 * HEAD - an interrupt came in and already set it.
1983 * NORMAL - One of two things:
1984 * 1) We really set it.
1985 * 2) A bunch of interrupts came in and moved
1986 * the page forward again.
1990 case RB_PAGE_NORMAL:
1994 RB_WARN_ON(cpu_buffer, 1);
1999 * It is possible that an interrupt came in,
2000 * set the head up, then more interrupts came in
2001 * and moved it again. When we get back here,
2002 * the page would have been set to NORMAL but we
2003 * just set it back to HEAD.
2005 * How do you detect this? Well, if that happened
2006 * the tail page would have moved.
2008 if (ret == RB_PAGE_NORMAL) {
2009 struct buffer_page *buffer_tail_page;
2011 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2013 * If the tail had moved passed next, then we need
2014 * to reset the pointer.
2016 if (buffer_tail_page != tail_page &&
2017 buffer_tail_page != next_page)
2018 rb_head_page_set_normal(cpu_buffer, new_head,
2024 * If this was the outer most commit (the one that
2025 * changed the original pointer from HEAD to UPDATE),
2026 * then it is up to us to reset it to NORMAL.
2028 if (type == RB_PAGE_HEAD) {
2029 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2032 if (RB_WARN_ON(cpu_buffer,
2033 ret != RB_PAGE_UPDATE))
2041 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2042 unsigned long tail, struct rb_event_info *info)
2044 struct buffer_page *tail_page = info->tail_page;
2045 struct ring_buffer_event *event;
2046 unsigned long length = info->length;
2049 * Only the event that crossed the page boundary
2050 * must fill the old tail_page with padding.
2052 if (tail >= BUF_PAGE_SIZE) {
2054 * If the page was filled, then we still need
2055 * to update the real_end. Reset it to zero
2056 * and the reader will ignore it.
2058 if (tail == BUF_PAGE_SIZE)
2059 tail_page->real_end = 0;
2061 local_sub(length, &tail_page->write);
2065 event = __rb_page_index(tail_page, tail);
2067 /* account for padding bytes */
2068 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2071 * Save the original length to the meta data.
2072 * This will be used by the reader to add lost event
2075 tail_page->real_end = tail;
2078 * If this event is bigger than the minimum size, then
2079 * we need to be careful that we don't subtract the
2080 * write counter enough to allow another writer to slip
2082 * We put in a discarded commit instead, to make sure
2083 * that this space is not used again.
2085 * If we are less than the minimum size, we don't need to
2088 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2089 /* No room for any events */
2091 /* Mark the rest of the page with padding */
2092 rb_event_set_padding(event);
2094 /* Set the write back to the previous setting */
2095 local_sub(length, &tail_page->write);
2099 /* Put in a discarded event */
2100 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2101 event->type_len = RINGBUF_TYPE_PADDING;
2102 /* time delta must be non zero */
2103 event->time_delta = 1;
2105 /* Set write to end of buffer */
2106 length = (tail + length) - BUF_PAGE_SIZE;
2107 local_sub(length, &tail_page->write);
2110 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2113 * This is the slow path, force gcc not to inline it.
2115 static noinline struct ring_buffer_event *
2116 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2117 unsigned long tail, struct rb_event_info *info)
2119 struct buffer_page *tail_page = info->tail_page;
2120 struct buffer_page *commit_page = cpu_buffer->commit_page;
2121 struct ring_buffer *buffer = cpu_buffer->buffer;
2122 struct buffer_page *next_page;
2125 next_page = tail_page;
2127 rb_inc_page(cpu_buffer, &next_page);
2130 * If for some reason, we had an interrupt storm that made
2131 * it all the way around the buffer, bail, and warn
2134 if (unlikely(next_page == commit_page)) {
2135 local_inc(&cpu_buffer->commit_overrun);
2140 * This is where the fun begins!
2142 * We are fighting against races between a reader that
2143 * could be on another CPU trying to swap its reader
2144 * page with the buffer head.
2146 * We are also fighting against interrupts coming in and
2147 * moving the head or tail on us as well.
2149 * If the next page is the head page then we have filled
2150 * the buffer, unless the commit page is still on the
2153 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2156 * If the commit is not on the reader page, then
2157 * move the header page.
2159 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2161 * If we are not in overwrite mode,
2162 * this is easy, just stop here.
2164 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2165 local_inc(&cpu_buffer->dropped_events);
2169 ret = rb_handle_head_page(cpu_buffer,
2178 * We need to be careful here too. The
2179 * commit page could still be on the reader
2180 * page. We could have a small buffer, and
2181 * have filled up the buffer with events
2182 * from interrupts and such, and wrapped.
2184 * Note, if the tail page is also the on the
2185 * reader_page, we let it move out.
2187 if (unlikely((cpu_buffer->commit_page !=
2188 cpu_buffer->tail_page) &&
2189 (cpu_buffer->commit_page ==
2190 cpu_buffer->reader_page))) {
2191 local_inc(&cpu_buffer->commit_overrun);
2197 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2201 rb_reset_tail(cpu_buffer, tail, info);
2203 /* Commit what we have for now. */
2204 rb_end_commit(cpu_buffer);
2205 /* rb_end_commit() decs committing */
2206 local_inc(&cpu_buffer->committing);
2208 /* fail and let the caller try again */
2209 return ERR_PTR(-EAGAIN);
2213 rb_reset_tail(cpu_buffer, tail, info);
2218 /* Slow path, do not inline */
2219 static noinline struct ring_buffer_event *
2220 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
2222 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2224 /* Not the first event on the page? */
2225 if (rb_event_index(event)) {
2226 event->time_delta = delta & TS_MASK;
2227 event->array[0] = delta >> TS_SHIFT;
2229 /* nope, just zero it */
2230 event->time_delta = 0;
2231 event->array[0] = 0;
2234 return skip_time_extend(event);
2237 static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2238 struct ring_buffer_event *event);
2241 * rb_update_event - update event type and data
2242 * @event: the event to update
2243 * @type: the type of event
2244 * @length: the size of the event field in the ring buffer
2246 * Update the type and data fields of the event. The length
2247 * is the actual size that is written to the ring buffer,
2248 * and with this, we can determine what to place into the
2252 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2253 struct ring_buffer_event *event,
2254 struct rb_event_info *info)
2256 unsigned length = info->length;
2257 u64 delta = info->delta;
2259 /* Only a commit updates the timestamp */
2260 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2264 * If we need to add a timestamp, then we
2265 * add it to the start of the resevered space.
2267 if (unlikely(info->add_timestamp)) {
2268 event = rb_add_time_stamp(event, delta);
2269 length -= RB_LEN_TIME_EXTEND;
2273 event->time_delta = delta;
2274 length -= RB_EVNT_HDR_SIZE;
2275 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2276 event->type_len = 0;
2277 event->array[0] = length;
2279 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2282 static unsigned rb_calculate_event_length(unsigned length)
2284 struct ring_buffer_event event; /* Used only for sizeof array */
2286 /* zero length can cause confusions */
2290 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2291 length += sizeof(event.array[0]);
2293 length += RB_EVNT_HDR_SIZE;
2294 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2297 * In case the time delta is larger than the 27 bits for it
2298 * in the header, we need to add a timestamp. If another
2299 * event comes in when trying to discard this one to increase
2300 * the length, then the timestamp will be added in the allocated
2301 * space of this event. If length is bigger than the size needed
2302 * for the TIME_EXTEND, then padding has to be used. The events
2303 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2304 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2305 * As length is a multiple of 4, we only need to worry if it
2306 * is 12 (RB_LEN_TIME_EXTEND + 4).
2308 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2309 length += RB_ALIGNMENT;
2314 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2315 static inline bool sched_clock_stable(void)
2322 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2323 struct ring_buffer_event *event)
2325 unsigned long new_index, old_index;
2326 struct buffer_page *bpage;
2327 unsigned long index;
2330 new_index = rb_event_index(event);
2331 old_index = new_index + rb_event_ts_length(event);
2332 addr = (unsigned long)event;
2335 bpage = READ_ONCE(cpu_buffer->tail_page);
2337 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2338 unsigned long write_mask =
2339 local_read(&bpage->write) & ~RB_WRITE_MASK;
2340 unsigned long event_length = rb_event_length(event);
2342 * This is on the tail page. It is possible that
2343 * a write could come in and move the tail page
2344 * and write to the next page. That is fine
2345 * because we just shorten what is on this page.
2347 old_index += write_mask;
2348 new_index += write_mask;
2349 index = local_cmpxchg(&bpage->write, old_index, new_index);
2350 if (index == old_index) {
2351 /* update counters */
2352 local_sub(event_length, &cpu_buffer->entries_bytes);
2357 /* could not discard */
2361 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2363 local_inc(&cpu_buffer->committing);
2364 local_inc(&cpu_buffer->commits);
2367 static __always_inline void
2368 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2370 unsigned long max_count;
2373 * We only race with interrupts and NMIs on this CPU.
2374 * If we own the commit event, then we can commit
2375 * all others that interrupted us, since the interruptions
2376 * are in stack format (they finish before they come
2377 * back to us). This allows us to do a simple loop to
2378 * assign the commit to the tail.
2381 max_count = cpu_buffer->nr_pages * 100;
2383 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2384 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2386 if (RB_WARN_ON(cpu_buffer,
2387 rb_is_reader_page(cpu_buffer->tail_page)))
2389 local_set(&cpu_buffer->commit_page->page->commit,
2390 rb_page_write(cpu_buffer->commit_page));
2391 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2392 /* Only update the write stamp if the page has an event */
2393 if (rb_page_write(cpu_buffer->commit_page))
2394 cpu_buffer->write_stamp =
2395 cpu_buffer->commit_page->page->time_stamp;
2396 /* add barrier to keep gcc from optimizing too much */
2399 while (rb_commit_index(cpu_buffer) !=
2400 rb_page_write(cpu_buffer->commit_page)) {
2402 local_set(&cpu_buffer->commit_page->page->commit,
2403 rb_page_write(cpu_buffer->commit_page));
2404 RB_WARN_ON(cpu_buffer,
2405 local_read(&cpu_buffer->commit_page->page->commit) &
2410 /* again, keep gcc from optimizing */
2414 * If an interrupt came in just after the first while loop
2415 * and pushed the tail page forward, we will be left with
2416 * a dangling commit that will never go forward.
2418 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2422 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2424 unsigned long commits;
2426 if (RB_WARN_ON(cpu_buffer,
2427 !local_read(&cpu_buffer->committing)))
2431 commits = local_read(&cpu_buffer->commits);
2432 /* synchronize with interrupts */
2434 if (local_read(&cpu_buffer->committing) == 1)
2435 rb_set_commit_to_write(cpu_buffer);
2437 local_dec(&cpu_buffer->committing);
2439 /* synchronize with interrupts */
2443 * Need to account for interrupts coming in between the
2444 * updating of the commit page and the clearing of the
2445 * committing counter.
2447 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2448 !local_read(&cpu_buffer->committing)) {
2449 local_inc(&cpu_buffer->committing);
2454 static inline void rb_event_discard(struct ring_buffer_event *event)
2456 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2457 event = skip_time_extend(event);
2459 /* array[0] holds the actual length for the discarded event */
2460 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2461 event->type_len = RINGBUF_TYPE_PADDING;
2462 /* time delta must be non zero */
2463 if (!event->time_delta)
2464 event->time_delta = 1;
2467 static __always_inline bool
2468 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2469 struct ring_buffer_event *event)
2471 unsigned long addr = (unsigned long)event;
2472 unsigned long index;
2474 index = rb_event_index(event);
2477 return cpu_buffer->commit_page->page == (void *)addr &&
2478 rb_commit_index(cpu_buffer) == index;
2481 static __always_inline void
2482 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2483 struct ring_buffer_event *event)
2488 * The event first in the commit queue updates the
2491 if (rb_event_is_commit(cpu_buffer, event)) {
2493 * A commit event that is first on a page
2494 * updates the write timestamp with the page stamp
2496 if (!rb_event_index(event))
2497 cpu_buffer->write_stamp =
2498 cpu_buffer->commit_page->page->time_stamp;
2499 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2500 delta = event->array[0];
2502 delta += event->time_delta;
2503 cpu_buffer->write_stamp += delta;
2505 cpu_buffer->write_stamp += event->time_delta;
2509 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2510 struct ring_buffer_event *event)
2512 local_inc(&cpu_buffer->entries);
2513 rb_update_write_stamp(cpu_buffer, event);
2514 rb_end_commit(cpu_buffer);
2517 static __always_inline void
2518 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2522 if (buffer->irq_work.waiters_pending) {
2523 buffer->irq_work.waiters_pending = false;
2524 /* irq_work_queue() supplies it's own memory barriers */
2525 irq_work_queue(&buffer->irq_work.work);
2528 if (cpu_buffer->irq_work.waiters_pending) {
2529 cpu_buffer->irq_work.waiters_pending = false;
2530 /* irq_work_queue() supplies it's own memory barriers */
2531 irq_work_queue(&cpu_buffer->irq_work.work);
2534 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
2536 if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2537 cpu_buffer->irq_work.wakeup_full = true;
2538 cpu_buffer->irq_work.full_waiters_pending = false;
2539 /* irq_work_queue() supplies it's own memory barriers */
2540 irq_work_queue(&cpu_buffer->irq_work.work);
2545 * The lock and unlock are done within a preempt disable section.
2546 * The current_context per_cpu variable can only be modified
2547 * by the current task between lock and unlock. But it can
2548 * be modified more than once via an interrupt. To pass this
2549 * information from the lock to the unlock without having to
2550 * access the 'in_interrupt()' functions again (which do show
2551 * a bit of overhead in something as critical as function tracing,
2552 * we use a bitmask trick.
2554 * bit 0 = NMI context
2555 * bit 1 = IRQ context
2556 * bit 2 = SoftIRQ context
2557 * bit 3 = normal context.
2559 * This works because this is the order of contexts that can
2560 * preempt other contexts. A SoftIRQ never preempts an IRQ
2563 * When the context is determined, the corresponding bit is
2564 * checked and set (if it was set, then a recursion of that context
2567 * On unlock, we need to clear this bit. To do so, just subtract
2568 * 1 from the current_context and AND it to itself.
2572 * 101 & 100 = 100 (clearing bit zero)
2575 * 1010 & 1001 = 1000 (clearing bit 1)
2577 * The least significant bit can be cleared this way, and it
2578 * just so happens that it is the same bit corresponding to
2579 * the current context.
2582 static __always_inline int
2583 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2585 unsigned int val = cpu_buffer->current_context;
2588 if (in_interrupt()) {
2594 bit = RB_CTX_SOFTIRQ;
2596 bit = RB_CTX_NORMAL;
2598 if (unlikely(val & (1 << bit)))
2602 cpu_buffer->current_context = val;
2607 static __always_inline void
2608 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2610 cpu_buffer->current_context &= cpu_buffer->current_context - 1;
2614 * ring_buffer_unlock_commit - commit a reserved
2615 * @buffer: The buffer to commit to
2616 * @event: The event pointer to commit.
2618 * This commits the data to the ring buffer, and releases any locks held.
2620 * Must be paired with ring_buffer_lock_reserve.
2622 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2623 struct ring_buffer_event *event)
2625 struct ring_buffer_per_cpu *cpu_buffer;
2626 int cpu = raw_smp_processor_id();
2628 cpu_buffer = buffer->buffers[cpu];
2630 rb_commit(cpu_buffer, event);
2632 rb_wakeups(buffer, cpu_buffer);
2634 trace_recursive_unlock(cpu_buffer);
2636 preempt_enable_notrace();
2640 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2642 static noinline void
2643 rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2644 struct rb_event_info *info)
2646 WARN_ONCE(info->delta > (1ULL << 59),
2647 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2648 (unsigned long long)info->delta,
2649 (unsigned long long)info->ts,
2650 (unsigned long long)cpu_buffer->write_stamp,
2651 sched_clock_stable() ? "" :
2652 "If you just came from a suspend/resume,\n"
2653 "please switch to the trace global clock:\n"
2654 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2655 info->add_timestamp = 1;
2658 static struct ring_buffer_event *
2659 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2660 struct rb_event_info *info)
2662 struct ring_buffer_event *event;
2663 struct buffer_page *tail_page;
2664 unsigned long tail, write;
2667 * If the time delta since the last event is too big to
2668 * hold in the time field of the event, then we append a
2669 * TIME EXTEND event ahead of the data event.
2671 if (unlikely(info->add_timestamp))
2672 info->length += RB_LEN_TIME_EXTEND;
2674 /* Don't let the compiler play games with cpu_buffer->tail_page */
2675 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2676 write = local_add_return(info->length, &tail_page->write);
2678 /* set write to only the index of the write */
2679 write &= RB_WRITE_MASK;
2680 tail = write - info->length;
2683 * If this is the first commit on the page, then it has the same
2684 * timestamp as the page itself.
2689 /* See if we shot pass the end of this buffer page */
2690 if (unlikely(write > BUF_PAGE_SIZE))
2691 return rb_move_tail(cpu_buffer, tail, info);
2693 /* We reserved something on the buffer */
2695 event = __rb_page_index(tail_page, tail);
2696 rb_update_event(cpu_buffer, event, info);
2698 local_inc(&tail_page->entries);
2701 * If this is the first commit on the page, then update
2705 tail_page->page->time_stamp = info->ts;
2707 /* account for these added bytes */
2708 local_add(info->length, &cpu_buffer->entries_bytes);
2713 static __always_inline struct ring_buffer_event *
2714 rb_reserve_next_event(struct ring_buffer *buffer,
2715 struct ring_buffer_per_cpu *cpu_buffer,
2716 unsigned long length)
2718 struct ring_buffer_event *event;
2719 struct rb_event_info info;
2723 rb_start_commit(cpu_buffer);
2725 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2727 * Due to the ability to swap a cpu buffer from a buffer
2728 * it is possible it was swapped before we committed.
2729 * (committing stops a swap). We check for it here and
2730 * if it happened, we have to fail the write.
2733 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2734 local_dec(&cpu_buffer->committing);
2735 local_dec(&cpu_buffer->commits);
2740 info.length = rb_calculate_event_length(length);
2742 info.add_timestamp = 0;
2746 * We allow for interrupts to reenter here and do a trace.
2747 * If one does, it will cause this original code to loop
2748 * back here. Even with heavy interrupts happening, this
2749 * should only happen a few times in a row. If this happens
2750 * 1000 times in a row, there must be either an interrupt
2751 * storm or we have something buggy.
2754 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2757 info.ts = rb_time_stamp(cpu_buffer->buffer);
2758 diff = info.ts - cpu_buffer->write_stamp;
2760 /* make sure this diff is calculated here */
2763 /* Did the write stamp get updated already? */
2764 if (likely(info.ts >= cpu_buffer->write_stamp)) {
2766 if (unlikely(test_time_stamp(info.delta)))
2767 rb_handle_timestamp(cpu_buffer, &info);
2770 event = __rb_reserve_next(cpu_buffer, &info);
2772 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2773 if (info.add_timestamp)
2774 info.length -= RB_LEN_TIME_EXTEND;
2784 rb_end_commit(cpu_buffer);
2789 * ring_buffer_lock_reserve - reserve a part of the buffer
2790 * @buffer: the ring buffer to reserve from
2791 * @length: the length of the data to reserve (excluding event header)
2793 * Returns a reseverd event on the ring buffer to copy directly to.
2794 * The user of this interface will need to get the body to write into
2795 * and can use the ring_buffer_event_data() interface.
2797 * The length is the length of the data needed, not the event length
2798 * which also includes the event header.
2800 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2801 * If NULL is returned, then nothing has been allocated or locked.
2803 struct ring_buffer_event *
2804 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2806 struct ring_buffer_per_cpu *cpu_buffer;
2807 struct ring_buffer_event *event;
2810 /* If we are tracing schedule, we don't want to recurse */
2811 preempt_disable_notrace();
2813 if (unlikely(atomic_read(&buffer->record_disabled)))
2816 cpu = raw_smp_processor_id();
2818 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2821 cpu_buffer = buffer->buffers[cpu];
2823 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2826 if (unlikely(length > BUF_MAX_DATA_SIZE))
2829 if (unlikely(trace_recursive_lock(cpu_buffer)))
2832 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2839 trace_recursive_unlock(cpu_buffer);
2841 preempt_enable_notrace();
2844 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2847 * Decrement the entries to the page that an event is on.
2848 * The event does not even need to exist, only the pointer
2849 * to the page it is on. This may only be called before the commit
2853 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2854 struct ring_buffer_event *event)
2856 unsigned long addr = (unsigned long)event;
2857 struct buffer_page *bpage = cpu_buffer->commit_page;
2858 struct buffer_page *start;
2862 /* Do the likely case first */
2863 if (likely(bpage->page == (void *)addr)) {
2864 local_dec(&bpage->entries);
2869 * Because the commit page may be on the reader page we
2870 * start with the next page and check the end loop there.
2872 rb_inc_page(cpu_buffer, &bpage);
2875 if (bpage->page == (void *)addr) {
2876 local_dec(&bpage->entries);
2879 rb_inc_page(cpu_buffer, &bpage);
2880 } while (bpage != start);
2882 /* commit not part of this buffer?? */
2883 RB_WARN_ON(cpu_buffer, 1);
2887 * ring_buffer_commit_discard - discard an event that has not been committed
2888 * @buffer: the ring buffer
2889 * @event: non committed event to discard
2891 * Sometimes an event that is in the ring buffer needs to be ignored.
2892 * This function lets the user discard an event in the ring buffer
2893 * and then that event will not be read later.
2895 * This function only works if it is called before the the item has been
2896 * committed. It will try to free the event from the ring buffer
2897 * if another event has not been added behind it.
2899 * If another event has been added behind it, it will set the event
2900 * up as discarded, and perform the commit.
2902 * If this function is called, do not call ring_buffer_unlock_commit on
2905 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2906 struct ring_buffer_event *event)
2908 struct ring_buffer_per_cpu *cpu_buffer;
2911 /* The event is discarded regardless */
2912 rb_event_discard(event);
2914 cpu = smp_processor_id();
2915 cpu_buffer = buffer->buffers[cpu];
2918 * This must only be called if the event has not been
2919 * committed yet. Thus we can assume that preemption
2920 * is still disabled.
2922 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2924 rb_decrement_entry(cpu_buffer, event);
2925 if (rb_try_to_discard(cpu_buffer, event))
2929 * The commit is still visible by the reader, so we
2930 * must still update the timestamp.
2932 rb_update_write_stamp(cpu_buffer, event);
2934 rb_end_commit(cpu_buffer);
2936 trace_recursive_unlock(cpu_buffer);
2938 preempt_enable_notrace();
2941 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2944 * ring_buffer_write - write data to the buffer without reserving
2945 * @buffer: The ring buffer to write to.
2946 * @length: The length of the data being written (excluding the event header)
2947 * @data: The data to write to the buffer.
2949 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2950 * one function. If you already have the data to write to the buffer, it
2951 * may be easier to simply call this function.
2953 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2954 * and not the length of the event which would hold the header.
2956 int ring_buffer_write(struct ring_buffer *buffer,
2957 unsigned long length,
2960 struct ring_buffer_per_cpu *cpu_buffer;
2961 struct ring_buffer_event *event;
2966 preempt_disable_notrace();
2968 if (atomic_read(&buffer->record_disabled))
2971 cpu = raw_smp_processor_id();
2973 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2976 cpu_buffer = buffer->buffers[cpu];
2978 if (atomic_read(&cpu_buffer->record_disabled))
2981 if (length > BUF_MAX_DATA_SIZE)
2984 if (unlikely(trace_recursive_lock(cpu_buffer)))
2987 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2991 body = rb_event_data(event);
2993 memcpy(body, data, length);
2995 rb_commit(cpu_buffer, event);
2997 rb_wakeups(buffer, cpu_buffer);
3002 trace_recursive_unlock(cpu_buffer);
3005 preempt_enable_notrace();
3009 EXPORT_SYMBOL_GPL(ring_buffer_write);
3011 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3013 struct buffer_page *reader = cpu_buffer->reader_page;
3014 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3015 struct buffer_page *commit = cpu_buffer->commit_page;
3017 /* In case of error, head will be NULL */
3018 if (unlikely(!head))
3021 return reader->read == rb_page_commit(reader) &&
3022 (commit == reader ||
3024 head->read == rb_page_commit(commit)));
3028 * ring_buffer_record_disable - stop all writes into the buffer
3029 * @buffer: The ring buffer to stop writes to.
3031 * This prevents all writes to the buffer. Any attempt to write
3032 * to the buffer after this will fail and return NULL.
3034 * The caller should call synchronize_sched() after this.
3036 void ring_buffer_record_disable(struct ring_buffer *buffer)
3038 atomic_inc(&buffer->record_disabled);
3040 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3043 * ring_buffer_record_enable - enable writes to the buffer
3044 * @buffer: The ring buffer to enable writes
3046 * Note, multiple disables will need the same number of enables
3047 * to truly enable the writing (much like preempt_disable).
3049 void ring_buffer_record_enable(struct ring_buffer *buffer)
3051 atomic_dec(&buffer->record_disabled);
3053 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3056 * ring_buffer_record_off - 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 * This is different than ring_buffer_record_disable() as
3063 * it works like an on/off switch, where as the disable() version
3064 * must be paired with a enable().
3066 void ring_buffer_record_off(struct ring_buffer *buffer)
3069 unsigned int new_rd;
3072 rd = atomic_read(&buffer->record_disabled);
3073 new_rd = rd | RB_BUFFER_OFF;
3074 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3076 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3079 * ring_buffer_record_on - restart writes into the buffer
3080 * @buffer: The ring buffer to start writes to.
3082 * This enables all writes to the buffer that was disabled by
3083 * ring_buffer_record_off().
3085 * This is different than ring_buffer_record_enable() as
3086 * it works like an on/off switch, where as the enable() version
3087 * must be paired with a disable().
3089 void ring_buffer_record_on(struct ring_buffer *buffer)
3092 unsigned int new_rd;
3095 rd = atomic_read(&buffer->record_disabled);
3096 new_rd = rd & ~RB_BUFFER_OFF;
3097 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3099 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3102 * ring_buffer_record_is_on - return true if the ring buffer can write
3103 * @buffer: The ring buffer to see if write is enabled
3105 * Returns true if the ring buffer is in a state that it accepts writes.
3107 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3109 return !atomic_read(&buffer->record_disabled);
3113 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3114 * @buffer: The ring buffer to stop writes to.
3115 * @cpu: The CPU buffer to stop
3117 * This prevents all writes to the buffer. Any attempt to write
3118 * to the buffer after this will fail and return NULL.
3120 * The caller should call synchronize_sched() after this.
3122 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3124 struct ring_buffer_per_cpu *cpu_buffer;
3126 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3129 cpu_buffer = buffer->buffers[cpu];
3130 atomic_inc(&cpu_buffer->record_disabled);
3132 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3135 * ring_buffer_record_enable_cpu - enable writes to the buffer
3136 * @buffer: The ring buffer to enable writes
3137 * @cpu: The CPU to enable.
3139 * Note, multiple disables will need the same number of enables
3140 * to truly enable the writing (much like preempt_disable).
3142 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3144 struct ring_buffer_per_cpu *cpu_buffer;
3146 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3149 cpu_buffer = buffer->buffers[cpu];
3150 atomic_dec(&cpu_buffer->record_disabled);
3152 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3155 * The total entries in the ring buffer is the running counter
3156 * of entries entered into the ring buffer, minus the sum of
3157 * the entries read from the ring buffer and the number of
3158 * entries that were overwritten.
3160 static inline unsigned long
3161 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3163 return local_read(&cpu_buffer->entries) -
3164 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3168 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3169 * @buffer: The ring buffer
3170 * @cpu: The per CPU buffer to read from.
3172 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3174 unsigned long flags;
3175 struct ring_buffer_per_cpu *cpu_buffer;
3176 struct buffer_page *bpage;
3179 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3182 cpu_buffer = buffer->buffers[cpu];
3183 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3185 * if the tail is on reader_page, oldest time stamp is on the reader
3188 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3189 bpage = cpu_buffer->reader_page;
3191 bpage = rb_set_head_page(cpu_buffer);
3193 ret = bpage->page->time_stamp;
3194 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3198 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3201 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3202 * @buffer: The ring buffer
3203 * @cpu: The per CPU buffer to read from.
3205 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3207 struct ring_buffer_per_cpu *cpu_buffer;
3210 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3213 cpu_buffer = buffer->buffers[cpu];
3214 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3218 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3221 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3222 * @buffer: The ring buffer
3223 * @cpu: The per CPU buffer to get the entries from.
3225 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3227 struct ring_buffer_per_cpu *cpu_buffer;
3229 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3232 cpu_buffer = buffer->buffers[cpu];
3234 return rb_num_of_entries(cpu_buffer);
3236 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3239 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3240 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3241 * @buffer: The ring buffer
3242 * @cpu: The per CPU buffer to get the number of overruns from
3244 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3246 struct ring_buffer_per_cpu *cpu_buffer;
3249 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3252 cpu_buffer = buffer->buffers[cpu];
3253 ret = local_read(&cpu_buffer->overrun);
3257 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3260 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3261 * commits failing due to the buffer wrapping around while there are uncommitted
3262 * events, such as during an interrupt storm.
3263 * @buffer: The ring buffer
3264 * @cpu: The per CPU buffer to get the number of overruns from
3267 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3269 struct ring_buffer_per_cpu *cpu_buffer;
3272 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3275 cpu_buffer = buffer->buffers[cpu];
3276 ret = local_read(&cpu_buffer->commit_overrun);
3280 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3283 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3284 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3285 * @buffer: The ring buffer
3286 * @cpu: The per CPU buffer to get the number of overruns from
3289 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3291 struct ring_buffer_per_cpu *cpu_buffer;
3294 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3297 cpu_buffer = buffer->buffers[cpu];
3298 ret = local_read(&cpu_buffer->dropped_events);
3302 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3305 * ring_buffer_read_events_cpu - get the number of events successfully read
3306 * @buffer: The ring buffer
3307 * @cpu: The per CPU buffer to get the number of events read
3310 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3312 struct ring_buffer_per_cpu *cpu_buffer;
3314 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3317 cpu_buffer = buffer->buffers[cpu];
3318 return cpu_buffer->read;
3320 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3323 * ring_buffer_entries - get the number of entries in a buffer
3324 * @buffer: The ring buffer
3326 * Returns the total number of entries in the ring buffer
3329 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3331 struct ring_buffer_per_cpu *cpu_buffer;
3332 unsigned long entries = 0;
3335 /* if you care about this being correct, lock the buffer */
3336 for_each_buffer_cpu(buffer, cpu) {
3337 cpu_buffer = buffer->buffers[cpu];
3338 entries += rb_num_of_entries(cpu_buffer);
3343 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3346 * ring_buffer_overruns - get the number of overruns in buffer
3347 * @buffer: The ring buffer
3349 * Returns the total number of overruns in the ring buffer
3352 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3354 struct ring_buffer_per_cpu *cpu_buffer;
3355 unsigned long overruns = 0;
3358 /* if you care about this being correct, lock the buffer */
3359 for_each_buffer_cpu(buffer, cpu) {
3360 cpu_buffer = buffer->buffers[cpu];
3361 overruns += local_read(&cpu_buffer->overrun);
3366 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3368 static void rb_iter_reset(struct ring_buffer_iter *iter)
3370 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3372 /* Iterator usage is expected to have record disabled */
3373 iter->head_page = cpu_buffer->reader_page;
3374 iter->head = cpu_buffer->reader_page->read;
3376 iter->cache_reader_page = iter->head_page;
3377 iter->cache_read = cpu_buffer->read;
3380 iter->read_stamp = cpu_buffer->read_stamp;
3382 iter->read_stamp = iter->head_page->page->time_stamp;
3386 * ring_buffer_iter_reset - reset an iterator
3387 * @iter: The iterator to reset
3389 * Resets the iterator, so that it will start from the beginning
3392 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3394 struct ring_buffer_per_cpu *cpu_buffer;
3395 unsigned long flags;
3400 cpu_buffer = iter->cpu_buffer;
3402 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3403 rb_iter_reset(iter);
3404 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3406 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3409 * ring_buffer_iter_empty - check if an iterator has no more to read
3410 * @iter: The iterator to check
3412 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3414 struct ring_buffer_per_cpu *cpu_buffer;
3415 struct buffer_page *reader;
3416 struct buffer_page *head_page;
3417 struct buffer_page *commit_page;
3420 cpu_buffer = iter->cpu_buffer;
3422 /* Remember, trace recording is off when iterator is in use */
3423 reader = cpu_buffer->reader_page;
3424 head_page = cpu_buffer->head_page;
3425 commit_page = cpu_buffer->commit_page;
3426 commit = rb_page_commit(commit_page);
3428 return ((iter->head_page == commit_page && iter->head == commit) ||
3429 (iter->head_page == reader && commit_page == head_page &&
3430 head_page->read == commit &&
3431 iter->head == rb_page_commit(cpu_buffer->reader_page)));
3433 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3436 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3437 struct ring_buffer_event *event)
3441 switch (event->type_len) {
3442 case RINGBUF_TYPE_PADDING:
3445 case RINGBUF_TYPE_TIME_EXTEND:
3446 delta = event->array[0];
3448 delta += event->time_delta;
3449 cpu_buffer->read_stamp += delta;
3452 case RINGBUF_TYPE_TIME_STAMP:
3453 /* FIXME: not implemented */
3456 case RINGBUF_TYPE_DATA:
3457 cpu_buffer->read_stamp += event->time_delta;
3467 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3468 struct ring_buffer_event *event)
3472 switch (event->type_len) {
3473 case RINGBUF_TYPE_PADDING:
3476 case RINGBUF_TYPE_TIME_EXTEND:
3477 delta = event->array[0];
3479 delta += event->time_delta;
3480 iter->read_stamp += delta;
3483 case RINGBUF_TYPE_TIME_STAMP:
3484 /* FIXME: not implemented */
3487 case RINGBUF_TYPE_DATA:
3488 iter->read_stamp += event->time_delta;
3497 static struct buffer_page *
3498 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3500 struct buffer_page *reader = NULL;
3501 unsigned long overwrite;
3502 unsigned long flags;
3506 local_irq_save(flags);
3507 arch_spin_lock(&cpu_buffer->lock);
3511 * This should normally only loop twice. But because the
3512 * start of the reader inserts an empty page, it causes
3513 * a case where we will loop three times. There should be no
3514 * reason to loop four times (that I know of).
3516 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3521 reader = cpu_buffer->reader_page;
3523 /* If there's more to read, return this page */
3524 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3527 /* Never should we have an index greater than the size */
3528 if (RB_WARN_ON(cpu_buffer,
3529 cpu_buffer->reader_page->read > rb_page_size(reader)))
3532 /* check if we caught up to the tail */
3534 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3537 /* Don't bother swapping if the ring buffer is empty */
3538 if (rb_num_of_entries(cpu_buffer) == 0)
3542 * Reset the reader page to size zero.
3544 local_set(&cpu_buffer->reader_page->write, 0);
3545 local_set(&cpu_buffer->reader_page->entries, 0);
3546 local_set(&cpu_buffer->reader_page->page->commit, 0);
3547 cpu_buffer->reader_page->real_end = 0;
3551 * Splice the empty reader page into the list around the head.
3553 reader = rb_set_head_page(cpu_buffer);
3556 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3557 cpu_buffer->reader_page->list.prev = reader->list.prev;
3560 * cpu_buffer->pages just needs to point to the buffer, it
3561 * has no specific buffer page to point to. Lets move it out
3562 * of our way so we don't accidentally swap it.
3564 cpu_buffer->pages = reader->list.prev;
3566 /* The reader page will be pointing to the new head */
3567 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3570 * We want to make sure we read the overruns after we set up our
3571 * pointers to the next object. The writer side does a
3572 * cmpxchg to cross pages which acts as the mb on the writer
3573 * side. Note, the reader will constantly fail the swap
3574 * while the writer is updating the pointers, so this
3575 * guarantees that the overwrite recorded here is the one we
3576 * want to compare with the last_overrun.
3579 overwrite = local_read(&(cpu_buffer->overrun));
3582 * Here's the tricky part.
3584 * We need to move the pointer past the header page.
3585 * But we can only do that if a writer is not currently
3586 * moving it. The page before the header page has the
3587 * flag bit '1' set if it is pointing to the page we want.
3588 * but if the writer is in the process of moving it
3589 * than it will be '2' or already moved '0'.
3592 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3595 * If we did not convert it, then we must try again.
3601 * Yeah! We succeeded in replacing the page.
3603 * Now make the new head point back to the reader page.
3605 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3606 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3608 /* Finally update the reader page to the new head */
3609 cpu_buffer->reader_page = reader;
3610 cpu_buffer->reader_page->read = 0;
3612 if (overwrite != cpu_buffer->last_overrun) {
3613 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3614 cpu_buffer->last_overrun = overwrite;
3620 /* Update the read_stamp on the first event */
3621 if (reader && reader->read == 0)
3622 cpu_buffer->read_stamp = reader->page->time_stamp;
3624 arch_spin_unlock(&cpu_buffer->lock);
3625 local_irq_restore(flags);
3630 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3632 struct ring_buffer_event *event;
3633 struct buffer_page *reader;
3636 reader = rb_get_reader_page(cpu_buffer);
3638 /* This function should not be called when buffer is empty */
3639 if (RB_WARN_ON(cpu_buffer, !reader))
3642 event = rb_reader_event(cpu_buffer);
3644 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3647 rb_update_read_stamp(cpu_buffer, event);
3649 length = rb_event_length(event);
3650 cpu_buffer->reader_page->read += length;
3653 static void rb_advance_iter(struct ring_buffer_iter *iter)
3655 struct ring_buffer_per_cpu *cpu_buffer;
3656 struct ring_buffer_event *event;
3659 cpu_buffer = iter->cpu_buffer;
3662 * Check if we are at the end of the buffer.
3664 if (iter->head >= rb_page_size(iter->head_page)) {
3665 /* discarded commits can make the page empty */
3666 if (iter->head_page == cpu_buffer->commit_page)
3672 event = rb_iter_head_event(iter);
3674 length = rb_event_length(event);
3677 * This should not be called to advance the header if we are
3678 * at the tail of the buffer.
3680 if (RB_WARN_ON(cpu_buffer,
3681 (iter->head_page == cpu_buffer->commit_page) &&
3682 (iter->head + length > rb_commit_index(cpu_buffer))))
3685 rb_update_iter_read_stamp(iter, event);
3687 iter->head += length;
3689 /* check for end of page padding */
3690 if ((iter->head >= rb_page_size(iter->head_page)) &&
3691 (iter->head_page != cpu_buffer->commit_page))
3695 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3697 return cpu_buffer->lost_events;
3700 static struct ring_buffer_event *
3701 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3702 unsigned long *lost_events)
3704 struct ring_buffer_event *event;
3705 struct buffer_page *reader;
3710 * We repeat when a time extend is encountered.
3711 * Since the time extend is always attached to a data event,
3712 * we should never loop more than once.
3713 * (We never hit the following condition more than twice).
3715 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3718 reader = rb_get_reader_page(cpu_buffer);
3722 event = rb_reader_event(cpu_buffer);
3724 switch (event->type_len) {
3725 case RINGBUF_TYPE_PADDING:
3726 if (rb_null_event(event))
3727 RB_WARN_ON(cpu_buffer, 1);
3729 * Because the writer could be discarding every
3730 * event it creates (which would probably be bad)
3731 * if we were to go back to "again" then we may never
3732 * catch up, and will trigger the warn on, or lock
3733 * the box. Return the padding, and we will release
3734 * the current locks, and try again.
3738 case RINGBUF_TYPE_TIME_EXTEND:
3739 /* Internal data, OK to advance */
3740 rb_advance_reader(cpu_buffer);
3743 case RINGBUF_TYPE_TIME_STAMP:
3744 /* FIXME: not implemented */
3745 rb_advance_reader(cpu_buffer);
3748 case RINGBUF_TYPE_DATA:
3750 *ts = cpu_buffer->read_stamp + event->time_delta;
3751 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3752 cpu_buffer->cpu, ts);
3755 *lost_events = rb_lost_events(cpu_buffer);
3764 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3766 static struct ring_buffer_event *
3767 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3769 struct ring_buffer *buffer;
3770 struct ring_buffer_per_cpu *cpu_buffer;
3771 struct ring_buffer_event *event;
3774 cpu_buffer = iter->cpu_buffer;
3775 buffer = cpu_buffer->buffer;
3778 * Check if someone performed a consuming read to
3779 * the buffer. A consuming read invalidates the iterator
3780 * and we need to reset the iterator in this case.
3782 if (unlikely(iter->cache_read != cpu_buffer->read ||
3783 iter->cache_reader_page != cpu_buffer->reader_page))
3784 rb_iter_reset(iter);
3787 if (ring_buffer_iter_empty(iter))
3791 * We repeat when a time extend is encountered or we hit
3792 * the end of the page. Since the time extend is always attached
3793 * to a data event, we should never loop more than three times.
3794 * Once for going to next page, once on time extend, and
3795 * finally once to get the event.
3796 * (We never hit the following condition more than thrice).
3798 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3801 if (rb_per_cpu_empty(cpu_buffer))
3804 if (iter->head >= rb_page_size(iter->head_page)) {
3809 event = rb_iter_head_event(iter);
3811 switch (event->type_len) {
3812 case RINGBUF_TYPE_PADDING:
3813 if (rb_null_event(event)) {
3817 rb_advance_iter(iter);
3820 case RINGBUF_TYPE_TIME_EXTEND:
3821 /* Internal data, OK to advance */
3822 rb_advance_iter(iter);
3825 case RINGBUF_TYPE_TIME_STAMP:
3826 /* FIXME: not implemented */
3827 rb_advance_iter(iter);
3830 case RINGBUF_TYPE_DATA:
3832 *ts = iter->read_stamp + event->time_delta;
3833 ring_buffer_normalize_time_stamp(buffer,
3834 cpu_buffer->cpu, ts);
3844 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3846 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
3848 if (likely(!in_nmi())) {
3849 raw_spin_lock(&cpu_buffer->reader_lock);
3854 * If an NMI die dumps out the content of the ring buffer
3855 * trylock must be used to prevent a deadlock if the NMI
3856 * preempted a task that holds the ring buffer locks. If
3857 * we get the lock then all is fine, if not, then continue
3858 * to do the read, but this can corrupt the ring buffer,
3859 * so it must be permanently disabled from future writes.
3860 * Reading from NMI is a oneshot deal.
3862 if (raw_spin_trylock(&cpu_buffer->reader_lock))
3865 /* Continue without locking, but disable the ring buffer */
3866 atomic_inc(&cpu_buffer->record_disabled);
3871 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
3874 raw_spin_unlock(&cpu_buffer->reader_lock);
3879 * ring_buffer_peek - peek at the next event to be read
3880 * @buffer: The ring buffer to read
3881 * @cpu: The cpu to peak at
3882 * @ts: The timestamp counter of this event.
3883 * @lost_events: a variable to store if events were lost (may be NULL)
3885 * This will return the event that will be read next, but does
3886 * not consume the data.
3888 struct ring_buffer_event *
3889 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3890 unsigned long *lost_events)
3892 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3893 struct ring_buffer_event *event;
3894 unsigned long flags;
3897 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3901 local_irq_save(flags);
3902 dolock = rb_reader_lock(cpu_buffer);
3903 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3904 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3905 rb_advance_reader(cpu_buffer);
3906 rb_reader_unlock(cpu_buffer, dolock);
3907 local_irq_restore(flags);
3909 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3916 * ring_buffer_iter_peek - peek at the next event to be read
3917 * @iter: The ring buffer iterator
3918 * @ts: The timestamp counter of this event.
3920 * This will return the event that will be read next, but does
3921 * not increment the iterator.
3923 struct ring_buffer_event *
3924 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3926 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3927 struct ring_buffer_event *event;
3928 unsigned long flags;
3931 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3932 event = rb_iter_peek(iter, ts);
3933 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3935 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3942 * ring_buffer_consume - return an event and consume it
3943 * @buffer: The ring buffer to get the next event from
3944 * @cpu: the cpu to read the buffer from
3945 * @ts: a variable to store the timestamp (may be NULL)
3946 * @lost_events: a variable to store if events were lost (may be NULL)
3948 * Returns the next event in the ring buffer, and that event is consumed.
3949 * Meaning, that sequential reads will keep returning a different event,
3950 * and eventually empty the ring buffer if the producer is slower.
3952 struct ring_buffer_event *
3953 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3954 unsigned long *lost_events)
3956 struct ring_buffer_per_cpu *cpu_buffer;
3957 struct ring_buffer_event *event = NULL;
3958 unsigned long flags;
3962 /* might be called in atomic */
3965 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3968 cpu_buffer = buffer->buffers[cpu];
3969 local_irq_save(flags);
3970 dolock = rb_reader_lock(cpu_buffer);
3972 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3974 cpu_buffer->lost_events = 0;
3975 rb_advance_reader(cpu_buffer);
3978 rb_reader_unlock(cpu_buffer, dolock);
3979 local_irq_restore(flags);
3984 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3989 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3992 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3993 * @buffer: The ring buffer to read from
3994 * @cpu: The cpu buffer to iterate over
3996 * This performs the initial preparations necessary to iterate
3997 * through the buffer. Memory is allocated, buffer recording
3998 * is disabled, and the iterator pointer is returned to the caller.
4000 * Disabling buffer recordng prevents the reading from being
4001 * corrupted. This is not a consuming read, so a producer is not
4004 * After a sequence of ring_buffer_read_prepare calls, the user is
4005 * expected to make at least one call to ring_buffer_read_prepare_sync.
4006 * Afterwards, ring_buffer_read_start is invoked to get things going
4009 * This overall must be paired with ring_buffer_read_finish.
4011 struct ring_buffer_iter *
4012 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
4014 struct ring_buffer_per_cpu *cpu_buffer;
4015 struct ring_buffer_iter *iter;
4017 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4020 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
4024 cpu_buffer = buffer->buffers[cpu];
4026 iter->cpu_buffer = cpu_buffer;
4028 atomic_inc(&buffer->resize_disabled);
4029 atomic_inc(&cpu_buffer->record_disabled);
4033 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4036 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4038 * All previously invoked ring_buffer_read_prepare calls to prepare
4039 * iterators will be synchronized. Afterwards, read_buffer_read_start
4040 * calls on those iterators are allowed.
4043 ring_buffer_read_prepare_sync(void)
4045 synchronize_sched();
4047 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4050 * ring_buffer_read_start - start a non consuming read of the buffer
4051 * @iter: The iterator returned by ring_buffer_read_prepare
4053 * This finalizes the startup of an iteration through the buffer.
4054 * The iterator comes from a call to ring_buffer_read_prepare and
4055 * an intervening ring_buffer_read_prepare_sync must have been
4058 * Must be paired with ring_buffer_read_finish.
4061 ring_buffer_read_start(struct ring_buffer_iter *iter)
4063 struct ring_buffer_per_cpu *cpu_buffer;
4064 unsigned long flags;
4069 cpu_buffer = iter->cpu_buffer;
4071 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4072 arch_spin_lock(&cpu_buffer->lock);
4073 rb_iter_reset(iter);
4074 arch_spin_unlock(&cpu_buffer->lock);
4075 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4077 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4080 * ring_buffer_read_finish - finish reading the iterator of the buffer
4081 * @iter: The iterator retrieved by ring_buffer_start
4083 * This re-enables the recording to the buffer, and frees the
4087 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4089 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4090 unsigned long flags;
4093 * Ring buffer is disabled from recording, here's a good place
4094 * to check the integrity of the ring buffer.
4095 * Must prevent readers from trying to read, as the check
4096 * clears the HEAD page and readers require it.
4098 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4099 rb_check_pages(cpu_buffer);
4100 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4102 atomic_dec(&cpu_buffer->record_disabled);
4103 atomic_dec(&cpu_buffer->buffer->resize_disabled);
4106 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4109 * ring_buffer_read - read the next item in the ring buffer by the iterator
4110 * @iter: The ring buffer iterator
4111 * @ts: The time stamp of the event read.
4113 * This reads the next event in the ring buffer and increments the iterator.
4115 struct ring_buffer_event *
4116 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4118 struct ring_buffer_event *event;
4119 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4120 unsigned long flags;
4122 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4124 event = rb_iter_peek(iter, ts);
4128 if (event->type_len == RINGBUF_TYPE_PADDING)
4131 rb_advance_iter(iter);
4133 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4137 EXPORT_SYMBOL_GPL(ring_buffer_read);
4140 * ring_buffer_size - return the size of the ring buffer (in bytes)
4141 * @buffer: The ring buffer.
4143 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4146 * Earlier, this method returned
4147 * BUF_PAGE_SIZE * buffer->nr_pages
4148 * Since the nr_pages field is now removed, we have converted this to
4149 * return the per cpu buffer value.
4151 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4154 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4156 EXPORT_SYMBOL_GPL(ring_buffer_size);
4159 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4161 rb_head_page_deactivate(cpu_buffer);
4163 cpu_buffer->head_page
4164 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4165 local_set(&cpu_buffer->head_page->write, 0);
4166 local_set(&cpu_buffer->head_page->entries, 0);
4167 local_set(&cpu_buffer->head_page->page->commit, 0);
4169 cpu_buffer->head_page->read = 0;
4171 cpu_buffer->tail_page = cpu_buffer->head_page;
4172 cpu_buffer->commit_page = cpu_buffer->head_page;
4174 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4175 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4176 local_set(&cpu_buffer->reader_page->write, 0);
4177 local_set(&cpu_buffer->reader_page->entries, 0);
4178 local_set(&cpu_buffer->reader_page->page->commit, 0);
4179 cpu_buffer->reader_page->read = 0;
4181 local_set(&cpu_buffer->entries_bytes, 0);
4182 local_set(&cpu_buffer->overrun, 0);
4183 local_set(&cpu_buffer->commit_overrun, 0);
4184 local_set(&cpu_buffer->dropped_events, 0);
4185 local_set(&cpu_buffer->entries, 0);
4186 local_set(&cpu_buffer->committing, 0);
4187 local_set(&cpu_buffer->commits, 0);
4188 cpu_buffer->read = 0;
4189 cpu_buffer->read_bytes = 0;
4191 cpu_buffer->write_stamp = 0;
4192 cpu_buffer->read_stamp = 0;
4194 cpu_buffer->lost_events = 0;
4195 cpu_buffer->last_overrun = 0;
4197 rb_head_page_activate(cpu_buffer);
4201 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4202 * @buffer: The ring buffer to reset a per cpu buffer of
4203 * @cpu: The CPU buffer to be reset
4205 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4207 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4208 unsigned long flags;
4210 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4213 atomic_inc(&buffer->resize_disabled);
4214 atomic_inc(&cpu_buffer->record_disabled);
4216 /* Make sure all commits have finished */
4217 synchronize_sched();
4219 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4221 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4224 arch_spin_lock(&cpu_buffer->lock);
4226 rb_reset_cpu(cpu_buffer);
4228 arch_spin_unlock(&cpu_buffer->lock);
4231 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4233 atomic_dec(&cpu_buffer->record_disabled);
4234 atomic_dec(&buffer->resize_disabled);
4236 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4239 * ring_buffer_reset - reset a ring buffer
4240 * @buffer: The ring buffer to reset all cpu buffers
4242 void ring_buffer_reset(struct ring_buffer *buffer)
4246 for_each_buffer_cpu(buffer, cpu)
4247 ring_buffer_reset_cpu(buffer, cpu);
4249 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4252 * rind_buffer_empty - is the ring buffer empty?
4253 * @buffer: The ring buffer to test
4255 bool ring_buffer_empty(struct ring_buffer *buffer)
4257 struct ring_buffer_per_cpu *cpu_buffer;
4258 unsigned long flags;
4263 /* yes this is racy, but if you don't like the race, lock the buffer */
4264 for_each_buffer_cpu(buffer, cpu) {
4265 cpu_buffer = buffer->buffers[cpu];
4266 local_irq_save(flags);
4267 dolock = rb_reader_lock(cpu_buffer);
4268 ret = rb_per_cpu_empty(cpu_buffer);
4269 rb_reader_unlock(cpu_buffer, dolock);
4270 local_irq_restore(flags);
4278 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4281 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4282 * @buffer: The ring buffer
4283 * @cpu: The CPU buffer to test
4285 bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4287 struct ring_buffer_per_cpu *cpu_buffer;
4288 unsigned long flags;
4292 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4295 cpu_buffer = buffer->buffers[cpu];
4296 local_irq_save(flags);
4297 dolock = rb_reader_lock(cpu_buffer);
4298 ret = rb_per_cpu_empty(cpu_buffer);
4299 rb_reader_unlock(cpu_buffer, dolock);
4300 local_irq_restore(flags);
4304 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4306 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4308 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4309 * @buffer_a: One buffer to swap with
4310 * @buffer_b: The other buffer to swap with
4312 * This function is useful for tracers that want to take a "snapshot"
4313 * of a CPU buffer and has another back up buffer lying around.
4314 * it is expected that the tracer handles the cpu buffer not being
4315 * used at the moment.
4317 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4318 struct ring_buffer *buffer_b, int cpu)
4320 struct ring_buffer_per_cpu *cpu_buffer_a;
4321 struct ring_buffer_per_cpu *cpu_buffer_b;
4324 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4325 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4328 cpu_buffer_a = buffer_a->buffers[cpu];
4329 cpu_buffer_b = buffer_b->buffers[cpu];
4331 /* At least make sure the two buffers are somewhat the same */
4332 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4337 if (atomic_read(&buffer_a->record_disabled))
4340 if (atomic_read(&buffer_b->record_disabled))
4343 if (atomic_read(&cpu_buffer_a->record_disabled))
4346 if (atomic_read(&cpu_buffer_b->record_disabled))
4350 * We can't do a synchronize_sched here because this
4351 * function can be called in atomic context.
4352 * Normally this will be called from the same CPU as cpu.
4353 * If not it's up to the caller to protect this.
4355 atomic_inc(&cpu_buffer_a->record_disabled);
4356 atomic_inc(&cpu_buffer_b->record_disabled);
4359 if (local_read(&cpu_buffer_a->committing))
4361 if (local_read(&cpu_buffer_b->committing))
4364 buffer_a->buffers[cpu] = cpu_buffer_b;
4365 buffer_b->buffers[cpu] = cpu_buffer_a;
4367 cpu_buffer_b->buffer = buffer_a;
4368 cpu_buffer_a->buffer = buffer_b;
4373 atomic_dec(&cpu_buffer_a->record_disabled);
4374 atomic_dec(&cpu_buffer_b->record_disabled);
4378 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4379 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4382 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4383 * @buffer: the buffer to allocate for.
4384 * @cpu: the cpu buffer to allocate.
4386 * This function is used in conjunction with ring_buffer_read_page.
4387 * When reading a full page from the ring buffer, these functions
4388 * can be used to speed up the process. The calling function should
4389 * allocate a few pages first with this function. Then when it
4390 * needs to get pages from the ring buffer, it passes the result
4391 * of this function into ring_buffer_read_page, which will swap
4392 * the page that was allocated, with the read page of the buffer.
4395 * The page allocated, or ERR_PTR
4397 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4399 struct ring_buffer_per_cpu *cpu_buffer;
4400 struct buffer_data_page *bpage = NULL;
4401 unsigned long flags;
4404 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4405 return ERR_PTR(-ENODEV);
4407 cpu_buffer = buffer->buffers[cpu];
4408 local_irq_save(flags);
4409 arch_spin_lock(&cpu_buffer->lock);
4411 if (cpu_buffer->free_page) {
4412 bpage = cpu_buffer->free_page;
4413 cpu_buffer->free_page = NULL;
4416 arch_spin_unlock(&cpu_buffer->lock);
4417 local_irq_restore(flags);
4422 page = alloc_pages_node(cpu_to_node(cpu),
4423 GFP_KERNEL | __GFP_NORETRY, 0);
4425 return ERR_PTR(-ENOMEM);
4427 bpage = page_address(page);
4430 rb_init_page(bpage);
4434 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4437 * ring_buffer_free_read_page - free an allocated read page
4438 * @buffer: the buffer the page was allocate for
4439 * @cpu: the cpu buffer the page came from
4440 * @data: the page to free
4442 * Free a page allocated from ring_buffer_alloc_read_page.
4444 void ring_buffer_free_read_page(struct ring_buffer *buffer, int cpu, void *data)
4446 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4447 struct buffer_data_page *bpage = data;
4448 struct page *page = virt_to_page(bpage);
4449 unsigned long flags;
4451 /* If the page is still in use someplace else, we can't reuse it */
4452 if (page_ref_count(page) > 1)
4455 local_irq_save(flags);
4456 arch_spin_lock(&cpu_buffer->lock);
4458 if (!cpu_buffer->free_page) {
4459 cpu_buffer->free_page = bpage;
4463 arch_spin_unlock(&cpu_buffer->lock);
4464 local_irq_restore(flags);
4467 free_page((unsigned long)bpage);
4469 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4472 * ring_buffer_read_page - extract a page from the ring buffer
4473 * @buffer: buffer to extract from
4474 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4475 * @len: amount to extract
4476 * @cpu: the cpu of the buffer to extract
4477 * @full: should the extraction only happen when the page is full.
4479 * This function will pull out a page from the ring buffer and consume it.
4480 * @data_page must be the address of the variable that was returned
4481 * from ring_buffer_alloc_read_page. This is because the page might be used
4482 * to swap with a page in the ring buffer.
4485 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4486 * if (IS_ERR(rpage))
4487 * return PTR_ERR(rpage);
4488 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4490 * process_page(rpage, ret);
4492 * When @full is set, the function will not return true unless
4493 * the writer is off the reader page.
4495 * Note: it is up to the calling functions to handle sleeps and wakeups.
4496 * The ring buffer can be used anywhere in the kernel and can not
4497 * blindly call wake_up. The layer that uses the ring buffer must be
4498 * responsible for that.
4501 * >=0 if data has been transferred, returns the offset of consumed data.
4502 * <0 if no data has been transferred.
4504 int ring_buffer_read_page(struct ring_buffer *buffer,
4505 void **data_page, size_t len, int cpu, int full)
4507 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4508 struct ring_buffer_event *event;
4509 struct buffer_data_page *bpage;
4510 struct buffer_page *reader;
4511 unsigned long missed_events;
4512 unsigned long flags;
4513 unsigned int commit;
4518 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4522 * If len is not big enough to hold the page header, then
4523 * we can not copy anything.
4525 if (len <= BUF_PAGE_HDR_SIZE)
4528 len -= BUF_PAGE_HDR_SIZE;
4537 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4539 reader = rb_get_reader_page(cpu_buffer);
4543 event = rb_reader_event(cpu_buffer);
4545 read = reader->read;
4546 commit = rb_page_commit(reader);
4548 /* Check if any events were dropped */
4549 missed_events = cpu_buffer->lost_events;
4552 * If this page has been partially read or
4553 * if len is not big enough to read the rest of the page or
4554 * a writer is still on the page, then
4555 * we must copy the data from the page to the buffer.
4556 * Otherwise, we can simply swap the page with the one passed in.
4558 if (read || (len < (commit - read)) ||
4559 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4560 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4561 unsigned int rpos = read;
4562 unsigned int pos = 0;
4568 if (len > (commit - read))
4569 len = (commit - read);
4571 /* Always keep the time extend and data together */
4572 size = rb_event_ts_length(event);
4577 /* save the current timestamp, since the user will need it */
4578 save_timestamp = cpu_buffer->read_stamp;
4580 /* Need to copy one event at a time */
4582 /* We need the size of one event, because
4583 * rb_advance_reader only advances by one event,
4584 * whereas rb_event_ts_length may include the size of
4585 * one or two events.
4586 * We have already ensured there's enough space if this
4587 * is a time extend. */
4588 size = rb_event_length(event);
4589 memcpy(bpage->data + pos, rpage->data + rpos, size);
4593 rb_advance_reader(cpu_buffer);
4594 rpos = reader->read;
4600 event = rb_reader_event(cpu_buffer);
4601 /* Always keep the time extend and data together */
4602 size = rb_event_ts_length(event);
4603 } while (len >= size);
4606 local_set(&bpage->commit, pos);
4607 bpage->time_stamp = save_timestamp;
4609 /* we copied everything to the beginning */
4612 /* update the entry counter */
4613 cpu_buffer->read += rb_page_entries(reader);
4614 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4616 /* swap the pages */
4617 rb_init_page(bpage);
4618 bpage = reader->page;
4619 reader->page = *data_page;
4620 local_set(&reader->write, 0);
4621 local_set(&reader->entries, 0);
4626 * Use the real_end for the data size,
4627 * This gives us a chance to store the lost events
4630 if (reader->real_end)
4631 local_set(&bpage->commit, reader->real_end);
4635 cpu_buffer->lost_events = 0;
4637 commit = local_read(&bpage->commit);
4639 * Set a flag in the commit field if we lost events
4641 if (missed_events) {
4642 /* If there is room at the end of the page to save the
4643 * missed events, then record it there.
4645 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4646 memcpy(&bpage->data[commit], &missed_events,
4647 sizeof(missed_events));
4648 local_add(RB_MISSED_STORED, &bpage->commit);
4649 commit += sizeof(missed_events);
4651 local_add(RB_MISSED_EVENTS, &bpage->commit);
4655 * This page may be off to user land. Zero it out here.
4657 if (commit < BUF_PAGE_SIZE)
4658 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4661 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4666 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4669 * We only allocate new buffers, never free them if the CPU goes down.
4670 * If we were to free the buffer, then the user would lose any trace that was in
4673 int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
4675 struct ring_buffer *buffer;
4678 unsigned long nr_pages;
4680 buffer = container_of(node, struct ring_buffer, node);
4681 if (cpumask_test_cpu(cpu, buffer->cpumask))
4686 /* check if all cpu sizes are same */
4687 for_each_buffer_cpu(buffer, cpu_i) {
4688 /* fill in the size from first enabled cpu */
4690 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4691 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4696 /* allocate minimum pages, user can later expand it */
4699 buffer->buffers[cpu] =
4700 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4701 if (!buffer->buffers[cpu]) {
4702 WARN(1, "failed to allocate ring buffer on CPU %u\n",
4707 cpumask_set_cpu(cpu, buffer->cpumask);
4711 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4713 * This is a basic integrity check of the ring buffer.
4714 * Late in the boot cycle this test will run when configured in.
4715 * It will kick off a thread per CPU that will go into a loop
4716 * writing to the per cpu ring buffer various sizes of data.
4717 * Some of the data will be large items, some small.
4719 * Another thread is created that goes into a spin, sending out
4720 * IPIs to the other CPUs to also write into the ring buffer.
4721 * this is to test the nesting ability of the buffer.
4723 * Basic stats are recorded and reported. If something in the
4724 * ring buffer should happen that's not expected, a big warning
4725 * is displayed and all ring buffers are disabled.
4727 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4729 struct rb_test_data {
4730 struct ring_buffer *buffer;
4731 unsigned long events;
4732 unsigned long bytes_written;
4733 unsigned long bytes_alloc;
4734 unsigned long bytes_dropped;
4735 unsigned long events_nested;
4736 unsigned long bytes_written_nested;
4737 unsigned long bytes_alloc_nested;
4738 unsigned long bytes_dropped_nested;
4739 int min_size_nested;
4740 int max_size_nested;
4747 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4750 #define RB_TEST_BUFFER_SIZE 1048576
4752 static char rb_string[] __initdata =
4753 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4754 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4755 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4757 static bool rb_test_started __initdata;
4764 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4766 struct ring_buffer_event *event;
4767 struct rb_item *item;
4774 /* Have nested writes different that what is written */
4775 cnt = data->cnt + (nested ? 27 : 0);
4777 /* Multiply cnt by ~e, to make some unique increment */
4778 size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4780 len = size + sizeof(struct rb_item);
4782 started = rb_test_started;
4783 /* read rb_test_started before checking buffer enabled */
4786 event = ring_buffer_lock_reserve(data->buffer, len);
4788 /* Ignore dropped events before test starts. */
4791 data->bytes_dropped += len;
4793 data->bytes_dropped_nested += len;
4798 event_len = ring_buffer_event_length(event);
4800 if (RB_WARN_ON(data->buffer, event_len < len))
4803 item = ring_buffer_event_data(event);
4805 memcpy(item->str, rb_string, size);
4808 data->bytes_alloc_nested += event_len;
4809 data->bytes_written_nested += len;
4810 data->events_nested++;
4811 if (!data->min_size_nested || len < data->min_size_nested)
4812 data->min_size_nested = len;
4813 if (len > data->max_size_nested)
4814 data->max_size_nested = len;
4816 data->bytes_alloc += event_len;
4817 data->bytes_written += len;
4819 if (!data->min_size || len < data->min_size)
4820 data->max_size = len;
4821 if (len > data->max_size)
4822 data->max_size = len;
4826 ring_buffer_unlock_commit(data->buffer, event);
4831 static __init int rb_test(void *arg)
4833 struct rb_test_data *data = arg;
4835 while (!kthread_should_stop()) {
4836 rb_write_something(data, false);
4839 set_current_state(TASK_INTERRUPTIBLE);
4840 /* Now sleep between a min of 100-300us and a max of 1ms */
4841 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4847 static __init void rb_ipi(void *ignore)
4849 struct rb_test_data *data;
4850 int cpu = smp_processor_id();
4852 data = &rb_data[cpu];
4853 rb_write_something(data, true);
4856 static __init int rb_hammer_test(void *arg)
4858 while (!kthread_should_stop()) {
4860 /* Send an IPI to all cpus to write data! */
4861 smp_call_function(rb_ipi, NULL, 1);
4862 /* No sleep, but for non preempt, let others run */
4869 static __init int test_ringbuffer(void)
4871 struct task_struct *rb_hammer;
4872 struct ring_buffer *buffer;
4876 pr_info("Running ring buffer tests...\n");
4878 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4879 if (WARN_ON(!buffer))
4882 /* Disable buffer so that threads can't write to it yet */
4883 ring_buffer_record_off(buffer);
4885 for_each_online_cpu(cpu) {
4886 rb_data[cpu].buffer = buffer;
4887 rb_data[cpu].cpu = cpu;
4888 rb_data[cpu].cnt = cpu;
4889 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4890 "rbtester/%d", cpu);
4891 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
4892 pr_cont("FAILED\n");
4893 ret = PTR_ERR(rb_threads[cpu]);
4897 kthread_bind(rb_threads[cpu], cpu);
4898 wake_up_process(rb_threads[cpu]);
4901 /* Now create the rb hammer! */
4902 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4903 if (WARN_ON(IS_ERR(rb_hammer))) {
4904 pr_cont("FAILED\n");
4905 ret = PTR_ERR(rb_hammer);
4909 ring_buffer_record_on(buffer);
4911 * Show buffer is enabled before setting rb_test_started.
4912 * Yes there's a small race window where events could be
4913 * dropped and the thread wont catch it. But when a ring
4914 * buffer gets enabled, there will always be some kind of
4915 * delay before other CPUs see it. Thus, we don't care about
4916 * those dropped events. We care about events dropped after
4917 * the threads see that the buffer is active.
4920 rb_test_started = true;
4922 set_current_state(TASK_INTERRUPTIBLE);
4923 /* Just run for 10 seconds */;
4924 schedule_timeout(10 * HZ);
4926 kthread_stop(rb_hammer);
4929 for_each_online_cpu(cpu) {
4930 if (!rb_threads[cpu])
4932 kthread_stop(rb_threads[cpu]);
4935 ring_buffer_free(buffer);
4940 pr_info("finished\n");
4941 for_each_online_cpu(cpu) {
4942 struct ring_buffer_event *event;
4943 struct rb_test_data *data = &rb_data[cpu];
4944 struct rb_item *item;
4945 unsigned long total_events;
4946 unsigned long total_dropped;
4947 unsigned long total_written;
4948 unsigned long total_alloc;
4949 unsigned long total_read = 0;
4950 unsigned long total_size = 0;
4951 unsigned long total_len = 0;
4952 unsigned long total_lost = 0;
4955 int small_event_size;
4959 total_events = data->events + data->events_nested;
4960 total_written = data->bytes_written + data->bytes_written_nested;
4961 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
4962 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
4964 big_event_size = data->max_size + data->max_size_nested;
4965 small_event_size = data->min_size + data->min_size_nested;
4967 pr_info("CPU %d:\n", cpu);
4968 pr_info(" events: %ld\n", total_events);
4969 pr_info(" dropped bytes: %ld\n", total_dropped);
4970 pr_info(" alloced bytes: %ld\n", total_alloc);
4971 pr_info(" written bytes: %ld\n", total_written);
4972 pr_info(" biggest event: %d\n", big_event_size);
4973 pr_info(" smallest event: %d\n", small_event_size);
4975 if (RB_WARN_ON(buffer, total_dropped))
4980 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
4982 item = ring_buffer_event_data(event);
4983 total_len += ring_buffer_event_length(event);
4984 total_size += item->size + sizeof(struct rb_item);
4985 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
4986 pr_info("FAILED!\n");
4987 pr_info("buffer had: %.*s\n", item->size, item->str);
4988 pr_info("expected: %.*s\n", item->size, rb_string);
4989 RB_WARN_ON(buffer, 1);
5000 pr_info(" read events: %ld\n", total_read);
5001 pr_info(" lost events: %ld\n", total_lost);
5002 pr_info(" total events: %ld\n", total_lost + total_read);
5003 pr_info(" recorded len bytes: %ld\n", total_len);
5004 pr_info(" recorded size bytes: %ld\n", total_size);
5006 pr_info(" With dropped events, record len and size may not match\n"
5007 " alloced and written from above\n");
5009 if (RB_WARN_ON(buffer, total_len != total_alloc ||
5010 total_size != total_written))
5013 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5019 pr_info("Ring buffer PASSED!\n");
5021 ring_buffer_free(buffer);
5025 late_initcall(test_ringbuffer);
5026 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */