1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
7 #include <linux/trace_events.h>
8 #include <linux/ring_buffer.h>
9 #include <linux/trace_clock.h>
10 #include <linux/sched/clock.h>
11 #include <linux/trace_seq.h>
12 #include <linux/spinlock.h>
13 #include <linux/irq_work.h>
14 #include <linux/uaccess.h>
15 #include <linux/hardirq.h>
16 #include <linux/kthread.h> /* for self test */
17 #include <linux/module.h>
18 #include <linux/percpu.h>
19 #include <linux/mutex.h>
20 #include <linux/delay.h>
21 #include <linux/slab.h>
22 #include <linux/init.h>
23 #include <linux/hash.h>
24 #include <linux/list.h>
25 #include <linux/cpu.h>
26 #include <linux/oom.h>
28 #include <asm/local.h>
30 static void update_pages_handler(struct work_struct *work);
33 * The ring buffer header is special. We must manually up keep it.
35 int ring_buffer_print_entry_header(struct trace_seq *s)
37 trace_seq_puts(s, "# compressed entry header\n");
38 trace_seq_puts(s, "\ttype_len : 5 bits\n");
39 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
40 trace_seq_puts(s, "\tarray : 32 bits\n");
41 trace_seq_putc(s, '\n');
42 trace_seq_printf(s, "\tpadding : type == %d\n",
43 RINGBUF_TYPE_PADDING);
44 trace_seq_printf(s, "\ttime_extend : type == %d\n",
45 RINGBUF_TYPE_TIME_EXTEND);
46 trace_seq_printf(s, "\ttime_stamp : type == %d\n",
47 RINGBUF_TYPE_TIME_STAMP);
48 trace_seq_printf(s, "\tdata max type_len == %d\n",
49 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
51 return !trace_seq_has_overflowed(s);
55 * The ring buffer is made up of a list of pages. A separate list of pages is
56 * allocated for each CPU. A writer may only write to a buffer that is
57 * associated with the CPU it is currently executing on. A reader may read
58 * from any per cpu buffer.
60 * The reader is special. For each per cpu buffer, the reader has its own
61 * reader page. When a reader has read the entire reader page, this reader
62 * page is swapped with another page in the ring buffer.
64 * Now, as long as the writer is off the reader page, the reader can do what
65 * ever it wants with that page. The writer will never write to that page
66 * again (as long as it is out of the ring buffer).
68 * Here's some silly ASCII art.
71 * |reader| RING BUFFER
73 * +------+ +---+ +---+ +---+
82 * |reader| RING BUFFER
83 * |page |------------------v
84 * +------+ +---+ +---+ +---+
93 * |reader| RING BUFFER
94 * |page |------------------v
95 * +------+ +---+ +---+ +---+
100 * +------------------------------+
104 * |buffer| RING BUFFER
105 * |page |------------------v
106 * +------+ +---+ +---+ +---+
108 * | New +---+ +---+ +---+
111 * +------------------------------+
114 * After we make this swap, the reader can hand this page off to the splice
115 * code and be done with it. It can even allocate a new page if it needs to
116 * and swap that into the ring buffer.
118 * We will be using cmpxchg soon to make all this lockless.
122 /* Used for individual buffers (after the counter) */
123 #define RB_BUFFER_OFF (1 << 20)
125 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
127 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
128 #define RB_ALIGNMENT 4U
129 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
130 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
132 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
133 # define RB_FORCE_8BYTE_ALIGNMENT 0
134 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
136 # define RB_FORCE_8BYTE_ALIGNMENT 1
137 # define RB_ARCH_ALIGNMENT 8U
140 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
142 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
143 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
146 RB_LEN_TIME_EXTEND = 8,
147 RB_LEN_TIME_STAMP = 8,
150 #define skip_time_extend(event) \
151 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
153 #define extended_time(event) \
154 (event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
156 static inline int rb_null_event(struct ring_buffer_event *event)
158 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
161 static void rb_event_set_padding(struct ring_buffer_event *event)
163 /* padding has a NULL time_delta */
164 event->type_len = RINGBUF_TYPE_PADDING;
165 event->time_delta = 0;
169 rb_event_data_length(struct ring_buffer_event *event)
174 length = event->type_len * RB_ALIGNMENT;
176 length = event->array[0];
177 return length + RB_EVNT_HDR_SIZE;
181 * Return the length of the given event. Will return
182 * the length of the time extend if the event is a
185 static inline unsigned
186 rb_event_length(struct ring_buffer_event *event)
188 switch (event->type_len) {
189 case RINGBUF_TYPE_PADDING:
190 if (rb_null_event(event))
193 return event->array[0] + RB_EVNT_HDR_SIZE;
195 case RINGBUF_TYPE_TIME_EXTEND:
196 return RB_LEN_TIME_EXTEND;
198 case RINGBUF_TYPE_TIME_STAMP:
199 return RB_LEN_TIME_STAMP;
201 case RINGBUF_TYPE_DATA:
202 return rb_event_data_length(event);
211 * Return total length of time extend and data,
212 * or just the event length for all other events.
214 static inline unsigned
215 rb_event_ts_length(struct ring_buffer_event *event)
219 if (extended_time(event)) {
220 /* time extends include the data event after it */
221 len = RB_LEN_TIME_EXTEND;
222 event = skip_time_extend(event);
224 return len + rb_event_length(event);
228 * ring_buffer_event_length - return the length of the event
229 * @event: the event to get the length of
231 * Returns the size of the data load of a data event.
232 * If the event is something other than a data event, it
233 * returns the size of the event itself. With the exception
234 * of a TIME EXTEND, where it still returns the size of the
235 * data load of the data event after it.
237 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
241 if (extended_time(event))
242 event = skip_time_extend(event);
244 length = rb_event_length(event);
245 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
247 length -= RB_EVNT_HDR_SIZE;
248 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
249 length -= sizeof(event->array[0]);
252 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
254 /* inline for ring buffer fast paths */
255 static __always_inline void *
256 rb_event_data(struct ring_buffer_event *event)
258 if (extended_time(event))
259 event = skip_time_extend(event);
260 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
261 /* If length is in len field, then array[0] has the data */
263 return (void *)&event->array[0];
264 /* Otherwise length is in array[0] and array[1] has the data */
265 return (void *)&event->array[1];
269 * ring_buffer_event_data - return the data of the event
270 * @event: the event to get the data from
272 void *ring_buffer_event_data(struct ring_buffer_event *event)
274 return rb_event_data(event);
276 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
278 #define for_each_buffer_cpu(buffer, cpu) \
279 for_each_cpu(cpu, buffer->cpumask)
282 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
283 #define TS_DELTA_TEST (~TS_MASK)
286 * ring_buffer_event_time_stamp - return the event's extended timestamp
287 * @event: the event to get the timestamp of
289 * Returns the extended timestamp associated with a data event.
290 * An extended time_stamp is a 64-bit timestamp represented
291 * internally in a special way that makes the best use of space
292 * contained within a ring buffer event. This function decodes
293 * it and maps it to a straight u64 value.
295 u64 ring_buffer_event_time_stamp(struct ring_buffer_event *event)
299 ts = event->array[0];
301 ts += event->time_delta;
306 /* Flag when events were overwritten */
307 #define RB_MISSED_EVENTS (1 << 31)
308 /* Missed count stored at end */
309 #define RB_MISSED_STORED (1 << 30)
311 #define RB_MISSED_FLAGS (RB_MISSED_EVENTS|RB_MISSED_STORED)
313 struct buffer_data_page {
314 u64 time_stamp; /* page time stamp */
315 local_t commit; /* write committed index */
316 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
320 * Note, the buffer_page list must be first. The buffer pages
321 * are allocated in cache lines, which means that each buffer
322 * page will be at the beginning of a cache line, and thus
323 * the least significant bits will be zero. We use this to
324 * add flags in the list struct pointers, to make the ring buffer
328 struct list_head list; /* list of buffer pages */
329 local_t write; /* index for next write */
330 unsigned read; /* index for next read */
331 local_t entries; /* entries on this page */
332 unsigned long real_end; /* real end of data */
333 struct buffer_data_page *page; /* Actual data page */
337 * The buffer page counters, write and entries, must be reset
338 * atomically when crossing page boundaries. To synchronize this
339 * update, two counters are inserted into the number. One is
340 * the actual counter for the write position or count on the page.
342 * The other is a counter of updaters. Before an update happens
343 * the update partition of the counter is incremented. This will
344 * allow the updater to update the counter atomically.
346 * The counter is 20 bits, and the state data is 12.
348 #define RB_WRITE_MASK 0xfffff
349 #define RB_WRITE_INTCNT (1 << 20)
351 static void rb_init_page(struct buffer_data_page *bpage)
353 local_set(&bpage->commit, 0);
357 * ring_buffer_page_len - the size of data on the page.
358 * @page: The page to read
360 * Returns the amount of data on the page, including buffer page header.
362 size_t ring_buffer_page_len(void *page)
364 struct buffer_data_page *bpage = page;
366 return (local_read(&bpage->commit) & ~RB_MISSED_FLAGS)
371 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
374 static void free_buffer_page(struct buffer_page *bpage)
376 free_page((unsigned long)bpage->page);
381 * We need to fit the time_stamp delta into 27 bits.
383 static inline int test_time_stamp(u64 delta)
385 if (delta & TS_DELTA_TEST)
390 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
392 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
393 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
395 int ring_buffer_print_page_header(struct trace_seq *s)
397 struct buffer_data_page field;
399 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
400 "offset:0;\tsize:%u;\tsigned:%u;\n",
401 (unsigned int)sizeof(field.time_stamp),
402 (unsigned int)is_signed_type(u64));
404 trace_seq_printf(s, "\tfield: local_t commit;\t"
405 "offset:%u;\tsize:%u;\tsigned:%u;\n",
406 (unsigned int)offsetof(typeof(field), commit),
407 (unsigned int)sizeof(field.commit),
408 (unsigned int)is_signed_type(long));
410 trace_seq_printf(s, "\tfield: int overwrite;\t"
411 "offset:%u;\tsize:%u;\tsigned:%u;\n",
412 (unsigned int)offsetof(typeof(field), commit),
414 (unsigned int)is_signed_type(long));
416 trace_seq_printf(s, "\tfield: char data;\t"
417 "offset:%u;\tsize:%u;\tsigned:%u;\n",
418 (unsigned int)offsetof(typeof(field), data),
419 (unsigned int)BUF_PAGE_SIZE,
420 (unsigned int)is_signed_type(char));
422 return !trace_seq_has_overflowed(s);
426 struct irq_work work;
427 wait_queue_head_t waiters;
428 wait_queue_head_t full_waiters;
429 bool waiters_pending;
430 bool full_waiters_pending;
435 * Structure to hold event state and handle nested events.
437 struct rb_event_info {
440 unsigned long length;
441 struct buffer_page *tail_page;
446 * Used for which event context the event is in.
452 * See trace_recursive_lock() comment below for more details.
463 * head_page == tail_page && head == tail then buffer is empty.
465 struct ring_buffer_per_cpu {
467 atomic_t record_disabled;
468 struct ring_buffer *buffer;
469 raw_spinlock_t reader_lock; /* serialize readers */
470 arch_spinlock_t lock;
471 struct lock_class_key lock_key;
472 struct buffer_data_page *free_page;
473 unsigned long nr_pages;
474 unsigned int current_context;
475 struct list_head *pages;
476 struct buffer_page *head_page; /* read from head */
477 struct buffer_page *tail_page; /* write to tail */
478 struct buffer_page *commit_page; /* committed pages */
479 struct buffer_page *reader_page;
480 unsigned long lost_events;
481 unsigned long last_overrun;
483 local_t entries_bytes;
486 local_t commit_overrun;
487 local_t dropped_events;
491 unsigned long read_bytes;
494 /* ring buffer pages to update, > 0 to add, < 0 to remove */
495 long nr_pages_to_update;
496 struct list_head new_pages; /* new pages to add */
497 struct work_struct update_pages_work;
498 struct completion update_done;
500 struct rb_irq_work irq_work;
506 atomic_t record_disabled;
507 atomic_t resize_disabled;
508 cpumask_var_t cpumask;
510 struct lock_class_key *reader_lock_key;
514 struct ring_buffer_per_cpu **buffers;
516 struct hlist_node node;
519 struct rb_irq_work irq_work;
523 struct ring_buffer_iter {
524 struct ring_buffer_per_cpu *cpu_buffer;
526 struct buffer_page *head_page;
527 struct buffer_page *cache_reader_page;
528 unsigned long cache_read;
533 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
535 * Schedules a delayed work to wake up any task that is blocked on the
536 * ring buffer waiters queue.
538 static void rb_wake_up_waiters(struct irq_work *work)
540 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
542 wake_up_all(&rbwork->waiters);
543 if (rbwork->wakeup_full) {
544 rbwork->wakeup_full = false;
545 wake_up_all(&rbwork->full_waiters);
550 * ring_buffer_wait - wait for input to the ring buffer
551 * @buffer: buffer to wait on
552 * @cpu: the cpu buffer to wait on
553 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
555 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
556 * as data is added to any of the @buffer's cpu buffers. Otherwise
557 * it will wait for data to be added to a specific cpu buffer.
559 int ring_buffer_wait(struct ring_buffer *buffer, int cpu, bool full)
561 struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
563 struct rb_irq_work *work;
567 * Depending on what the caller is waiting for, either any
568 * data in any cpu buffer, or a specific buffer, put the
569 * caller on the appropriate wait queue.
571 if (cpu == RING_BUFFER_ALL_CPUS) {
572 work = &buffer->irq_work;
573 /* Full only makes sense on per cpu reads */
576 if (!cpumask_test_cpu(cpu, buffer->cpumask))
578 cpu_buffer = buffer->buffers[cpu];
579 work = &cpu_buffer->irq_work;
585 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
587 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
590 * The events can happen in critical sections where
591 * checking a work queue can cause deadlocks.
592 * After adding a task to the queue, this flag is set
593 * only to notify events to try to wake up the queue
596 * We don't clear it even if the buffer is no longer
597 * empty. The flag only causes the next event to run
598 * irq_work to do the work queue wake up. The worse
599 * that can happen if we race with !trace_empty() is that
600 * an event will cause an irq_work to try to wake up
603 * There's no reason to protect this flag either, as
604 * the work queue and irq_work logic will do the necessary
605 * synchronization for the wake ups. The only thing
606 * that is necessary is that the wake up happens after
607 * a task has been queued. It's OK for spurious wake ups.
610 work->full_waiters_pending = true;
612 work->waiters_pending = true;
614 if (signal_pending(current)) {
619 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
622 if (cpu != RING_BUFFER_ALL_CPUS &&
623 !ring_buffer_empty_cpu(buffer, cpu)) {
630 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
631 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
632 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
642 finish_wait(&work->full_waiters, &wait);
644 finish_wait(&work->waiters, &wait);
650 * ring_buffer_poll_wait - poll on buffer input
651 * @buffer: buffer to wait on
652 * @cpu: the cpu buffer to wait on
653 * @filp: the file descriptor
654 * @poll_table: The poll descriptor
656 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
657 * as data is added to any of the @buffer's cpu buffers. Otherwise
658 * it will wait for data to be added to a specific cpu buffer.
660 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
663 __poll_t ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
664 struct file *filp, poll_table *poll_table)
666 struct ring_buffer_per_cpu *cpu_buffer;
667 struct rb_irq_work *work;
669 if (cpu == RING_BUFFER_ALL_CPUS)
670 work = &buffer->irq_work;
672 if (!cpumask_test_cpu(cpu, buffer->cpumask))
675 cpu_buffer = buffer->buffers[cpu];
676 work = &cpu_buffer->irq_work;
679 poll_wait(filp, &work->waiters, poll_table);
680 work->waiters_pending = true;
682 * There's a tight race between setting the waiters_pending and
683 * checking if the ring buffer is empty. Once the waiters_pending bit
684 * is set, the next event will wake the task up, but we can get stuck
685 * if there's only a single event in.
687 * FIXME: Ideally, we need a memory barrier on the writer side as well,
688 * but adding a memory barrier to all events will cause too much of a
689 * performance hit in the fast path. We only need a memory barrier when
690 * the buffer goes from empty to having content. But as this race is
691 * extremely small, and it's not a problem if another event comes in, we
696 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
697 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
698 return EPOLLIN | EPOLLRDNORM;
702 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
703 #define RB_WARN_ON(b, cond) \
705 int _____ret = unlikely(cond); \
707 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
708 struct ring_buffer_per_cpu *__b = \
710 atomic_inc(&__b->buffer->record_disabled); \
712 atomic_inc(&b->record_disabled); \
718 /* Up this if you want to test the TIME_EXTENTS and normalization */
719 #define DEBUG_SHIFT 0
721 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
723 /* shift to debug/test normalization and TIME_EXTENTS */
724 return buffer->clock() << DEBUG_SHIFT;
727 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
731 preempt_disable_notrace();
732 time = rb_time_stamp(buffer);
733 preempt_enable_no_resched_notrace();
737 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
739 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
742 /* Just stupid testing the normalize function and deltas */
745 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
748 * Making the ring buffer lockless makes things tricky.
749 * Although writes only happen on the CPU that they are on,
750 * and they only need to worry about interrupts. Reads can
753 * The reader page is always off the ring buffer, but when the
754 * reader finishes with a page, it needs to swap its page with
755 * a new one from the buffer. The reader needs to take from
756 * the head (writes go to the tail). But if a writer is in overwrite
757 * mode and wraps, it must push the head page forward.
759 * Here lies the problem.
761 * The reader must be careful to replace only the head page, and
762 * not another one. As described at the top of the file in the
763 * ASCII art, the reader sets its old page to point to the next
764 * page after head. It then sets the page after head to point to
765 * the old reader page. But if the writer moves the head page
766 * during this operation, the reader could end up with the tail.
768 * We use cmpxchg to help prevent this race. We also do something
769 * special with the page before head. We set the LSB to 1.
771 * When the writer must push the page forward, it will clear the
772 * bit that points to the head page, move the head, and then set
773 * the bit that points to the new head page.
775 * We also don't want an interrupt coming in and moving the head
776 * page on another writer. Thus we use the second LSB to catch
779 * head->list->prev->next bit 1 bit 0
782 * Points to head page 0 1
785 * Note we can not trust the prev pointer of the head page, because:
787 * +----+ +-----+ +-----+
788 * | |------>| T |---X--->| N |
790 * +----+ +-----+ +-----+
793 * +----------| R |----------+ |
797 * Key: ---X--> HEAD flag set in pointer
802 * (see __rb_reserve_next() to see where this happens)
804 * What the above shows is that the reader just swapped out
805 * the reader page with a page in the buffer, but before it
806 * could make the new header point back to the new page added
807 * it was preempted by a writer. The writer moved forward onto
808 * the new page added by the reader and is about to move forward
811 * You can see, it is legitimate for the previous pointer of
812 * the head (or any page) not to point back to itself. But only
816 #define RB_PAGE_NORMAL 0UL
817 #define RB_PAGE_HEAD 1UL
818 #define RB_PAGE_UPDATE 2UL
821 #define RB_FLAG_MASK 3UL
823 /* PAGE_MOVED is not part of the mask */
824 #define RB_PAGE_MOVED 4UL
827 * rb_list_head - remove any bit
829 static struct list_head *rb_list_head(struct list_head *list)
831 unsigned long val = (unsigned long)list;
833 return (struct list_head *)(val & ~RB_FLAG_MASK);
837 * rb_is_head_page - test if the given page is the head page
839 * Because the reader may move the head_page pointer, we can
840 * not trust what the head page is (it may be pointing to
841 * the reader page). But if the next page is a header page,
842 * its flags will be non zero.
845 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
846 struct buffer_page *page, struct list_head *list)
850 val = (unsigned long)list->next;
852 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
853 return RB_PAGE_MOVED;
855 return val & RB_FLAG_MASK;
861 * The unique thing about the reader page, is that, if the
862 * writer is ever on it, the previous pointer never points
863 * back to the reader page.
865 static bool rb_is_reader_page(struct buffer_page *page)
867 struct list_head *list = page->list.prev;
869 return rb_list_head(list->next) != &page->list;
873 * rb_set_list_to_head - set a list_head to be pointing to head.
875 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
876 struct list_head *list)
880 ptr = (unsigned long *)&list->next;
881 *ptr |= RB_PAGE_HEAD;
882 *ptr &= ~RB_PAGE_UPDATE;
886 * rb_head_page_activate - sets up head page
888 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
890 struct buffer_page *head;
892 head = cpu_buffer->head_page;
897 * Set the previous list pointer to have the HEAD flag.
899 rb_set_list_to_head(cpu_buffer, head->list.prev);
902 static void rb_list_head_clear(struct list_head *list)
904 unsigned long *ptr = (unsigned long *)&list->next;
906 *ptr &= ~RB_FLAG_MASK;
910 * rb_head_page_deactivate - clears head page ptr (for free list)
913 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
915 struct list_head *hd;
917 /* Go through the whole list and clear any pointers found. */
918 rb_list_head_clear(cpu_buffer->pages);
920 list_for_each(hd, cpu_buffer->pages)
921 rb_list_head_clear(hd);
924 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
925 struct buffer_page *head,
926 struct buffer_page *prev,
927 int old_flag, int new_flag)
929 struct list_head *list;
930 unsigned long val = (unsigned long)&head->list;
935 val &= ~RB_FLAG_MASK;
937 ret = cmpxchg((unsigned long *)&list->next,
938 val | old_flag, val | new_flag);
940 /* check if the reader took the page */
941 if ((ret & ~RB_FLAG_MASK) != val)
942 return RB_PAGE_MOVED;
944 return ret & RB_FLAG_MASK;
947 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
948 struct buffer_page *head,
949 struct buffer_page *prev,
952 return rb_head_page_set(cpu_buffer, head, prev,
953 old_flag, RB_PAGE_UPDATE);
956 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
957 struct buffer_page *head,
958 struct buffer_page *prev,
961 return rb_head_page_set(cpu_buffer, head, prev,
962 old_flag, RB_PAGE_HEAD);
965 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
966 struct buffer_page *head,
967 struct buffer_page *prev,
970 return rb_head_page_set(cpu_buffer, head, prev,
971 old_flag, RB_PAGE_NORMAL);
974 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
975 struct buffer_page **bpage)
977 struct list_head *p = rb_list_head((*bpage)->list.next);
979 *bpage = list_entry(p, struct buffer_page, list);
982 static struct buffer_page *
983 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
985 struct buffer_page *head;
986 struct buffer_page *page;
987 struct list_head *list;
990 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
994 list = cpu_buffer->pages;
995 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
998 page = head = cpu_buffer->head_page;
1000 * It is possible that the writer moves the header behind
1001 * where we started, and we miss in one loop.
1002 * A second loop should grab the header, but we'll do
1003 * three loops just because I'm paranoid.
1005 for (i = 0; i < 3; i++) {
1007 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
1008 cpu_buffer->head_page = page;
1011 rb_inc_page(cpu_buffer, &page);
1012 } while (page != head);
1015 RB_WARN_ON(cpu_buffer, 1);
1020 static int rb_head_page_replace(struct buffer_page *old,
1021 struct buffer_page *new)
1023 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1027 val = *ptr & ~RB_FLAG_MASK;
1028 val |= RB_PAGE_HEAD;
1030 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1036 * rb_tail_page_update - move the tail page forward
1038 static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1039 struct buffer_page *tail_page,
1040 struct buffer_page *next_page)
1042 unsigned long old_entries;
1043 unsigned long old_write;
1046 * The tail page now needs to be moved forward.
1048 * We need to reset the tail page, but without messing
1049 * with possible erasing of data brought in by interrupts
1050 * that have moved the tail page and are currently on it.
1052 * We add a counter to the write field to denote this.
1054 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1055 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1058 * Just make sure we have seen our old_write and synchronize
1059 * with any interrupts that come in.
1064 * If the tail page is still the same as what we think
1065 * it is, then it is up to us to update the tail
1068 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1069 /* Zero the write counter */
1070 unsigned long val = old_write & ~RB_WRITE_MASK;
1071 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1074 * This will only succeed if an interrupt did
1075 * not come in and change it. In which case, we
1076 * do not want to modify it.
1078 * We add (void) to let the compiler know that we do not care
1079 * about the return value of these functions. We use the
1080 * cmpxchg to only update if an interrupt did not already
1081 * do it for us. If the cmpxchg fails, we don't care.
1083 (void)local_cmpxchg(&next_page->write, old_write, val);
1084 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1087 * No need to worry about races with clearing out the commit.
1088 * it only can increment when a commit takes place. But that
1089 * only happens in the outer most nested commit.
1091 local_set(&next_page->page->commit, 0);
1093 /* Again, either we update tail_page or an interrupt does */
1094 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1098 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1099 struct buffer_page *bpage)
1101 unsigned long val = (unsigned long)bpage;
1103 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1110 * rb_check_list - make sure a pointer to a list has the last bits zero
1112 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1113 struct list_head *list)
1115 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1117 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1123 * rb_check_pages - integrity check of buffer pages
1124 * @cpu_buffer: CPU buffer with pages to test
1126 * As a safety measure we check to make sure the data pages have not
1129 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1131 struct list_head *head = cpu_buffer->pages;
1132 struct buffer_page *bpage, *tmp;
1134 /* Reset the head page if it exists */
1135 if (cpu_buffer->head_page)
1136 rb_set_head_page(cpu_buffer);
1138 rb_head_page_deactivate(cpu_buffer);
1140 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1142 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1145 if (rb_check_list(cpu_buffer, head))
1148 list_for_each_entry_safe(bpage, tmp, head, list) {
1149 if (RB_WARN_ON(cpu_buffer,
1150 bpage->list.next->prev != &bpage->list))
1152 if (RB_WARN_ON(cpu_buffer,
1153 bpage->list.prev->next != &bpage->list))
1155 if (rb_check_list(cpu_buffer, &bpage->list))
1159 rb_head_page_activate(cpu_buffer);
1164 static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1166 struct buffer_page *bpage, *tmp;
1167 bool user_thread = current->mm != NULL;
1172 * Check if the available memory is there first.
1173 * Note, si_mem_available() only gives us a rough estimate of available
1174 * memory. It may not be accurate. But we don't care, we just want
1175 * to prevent doing any allocation when it is obvious that it is
1176 * not going to succeed.
1178 i = si_mem_available();
1183 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1184 * gracefully without invoking oom-killer and the system is not
1187 mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1190 * If a user thread allocates too much, and si_mem_available()
1191 * reports there's enough memory, even though there is not.
1192 * Make sure the OOM killer kills this thread. This can happen
1193 * even with RETRY_MAYFAIL because another task may be doing
1194 * an allocation after this task has taken all memory.
1195 * This is the task the OOM killer needs to take out during this
1196 * loop, even if it was triggered by an allocation somewhere else.
1199 set_current_oom_origin();
1200 for (i = 0; i < nr_pages; i++) {
1203 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1204 mflags, cpu_to_node(cpu));
1208 list_add(&bpage->list, pages);
1210 page = alloc_pages_node(cpu_to_node(cpu), mflags, 0);
1213 bpage->page = page_address(page);
1214 rb_init_page(bpage->page);
1216 if (user_thread && fatal_signal_pending(current))
1220 clear_current_oom_origin();
1225 list_for_each_entry_safe(bpage, tmp, pages, list) {
1226 list_del_init(&bpage->list);
1227 free_buffer_page(bpage);
1230 clear_current_oom_origin();
1235 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1236 unsigned long nr_pages)
1242 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1246 * The ring buffer page list is a circular list that does not
1247 * start and end with a list head. All page list items point to
1250 cpu_buffer->pages = pages.next;
1253 cpu_buffer->nr_pages = nr_pages;
1255 rb_check_pages(cpu_buffer);
1260 static struct ring_buffer_per_cpu *
1261 rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1263 struct ring_buffer_per_cpu *cpu_buffer;
1264 struct buffer_page *bpage;
1268 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1269 GFP_KERNEL, cpu_to_node(cpu));
1273 cpu_buffer->cpu = cpu;
1274 cpu_buffer->buffer = buffer;
1275 raw_spin_lock_init(&cpu_buffer->reader_lock);
1276 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1277 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1278 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1279 init_completion(&cpu_buffer->update_done);
1280 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1281 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1282 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1284 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1285 GFP_KERNEL, cpu_to_node(cpu));
1287 goto fail_free_buffer;
1289 rb_check_bpage(cpu_buffer, bpage);
1291 cpu_buffer->reader_page = bpage;
1292 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1294 goto fail_free_reader;
1295 bpage->page = page_address(page);
1296 rb_init_page(bpage->page);
1298 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1299 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1301 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1303 goto fail_free_reader;
1305 cpu_buffer->head_page
1306 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1307 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1309 rb_head_page_activate(cpu_buffer);
1314 free_buffer_page(cpu_buffer->reader_page);
1321 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1323 struct list_head *head = cpu_buffer->pages;
1324 struct buffer_page *bpage, *tmp;
1326 free_buffer_page(cpu_buffer->reader_page);
1328 rb_head_page_deactivate(cpu_buffer);
1331 list_for_each_entry_safe(bpage, tmp, head, list) {
1332 list_del_init(&bpage->list);
1333 free_buffer_page(bpage);
1335 bpage = list_entry(head, struct buffer_page, list);
1336 free_buffer_page(bpage);
1343 * __ring_buffer_alloc - allocate a new ring_buffer
1344 * @size: the size in bytes per cpu that is needed.
1345 * @flags: attributes to set for the ring buffer.
1347 * Currently the only flag that is available is the RB_FL_OVERWRITE
1348 * flag. This flag means that the buffer will overwrite old data
1349 * when the buffer wraps. If this flag is not set, the buffer will
1350 * drop data when the tail hits the head.
1352 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1353 struct lock_class_key *key)
1355 struct ring_buffer *buffer;
1361 /* keep it in its own cache line */
1362 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1367 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1368 goto fail_free_buffer;
1370 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1371 buffer->flags = flags;
1372 buffer->clock = trace_clock_local;
1373 buffer->reader_lock_key = key;
1375 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1376 init_waitqueue_head(&buffer->irq_work.waiters);
1378 /* need at least two pages */
1382 buffer->cpus = nr_cpu_ids;
1384 bsize = sizeof(void *) * nr_cpu_ids;
1385 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1387 if (!buffer->buffers)
1388 goto fail_free_cpumask;
1390 cpu = raw_smp_processor_id();
1391 cpumask_set_cpu(cpu, buffer->cpumask);
1392 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1393 if (!buffer->buffers[cpu])
1394 goto fail_free_buffers;
1396 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1398 goto fail_free_buffers;
1400 mutex_init(&buffer->mutex);
1405 for_each_buffer_cpu(buffer, cpu) {
1406 if (buffer->buffers[cpu])
1407 rb_free_cpu_buffer(buffer->buffers[cpu]);
1409 kfree(buffer->buffers);
1412 free_cpumask_var(buffer->cpumask);
1418 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1421 * ring_buffer_free - free a ring buffer.
1422 * @buffer: the buffer to free.
1425 ring_buffer_free(struct ring_buffer *buffer)
1429 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1431 for_each_buffer_cpu(buffer, cpu)
1432 rb_free_cpu_buffer(buffer->buffers[cpu]);
1434 kfree(buffer->buffers);
1435 free_cpumask_var(buffer->cpumask);
1439 EXPORT_SYMBOL_GPL(ring_buffer_free);
1441 void ring_buffer_set_clock(struct ring_buffer *buffer,
1444 buffer->clock = clock;
1447 void ring_buffer_set_time_stamp_abs(struct ring_buffer *buffer, bool abs)
1449 buffer->time_stamp_abs = abs;
1452 bool ring_buffer_time_stamp_abs(struct ring_buffer *buffer)
1454 return buffer->time_stamp_abs;
1457 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1459 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1461 return local_read(&bpage->entries) & RB_WRITE_MASK;
1464 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1466 return local_read(&bpage->write) & RB_WRITE_MASK;
1470 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1472 struct list_head *tail_page, *to_remove, *next_page;
1473 struct buffer_page *to_remove_page, *tmp_iter_page;
1474 struct buffer_page *last_page, *first_page;
1475 unsigned long nr_removed;
1476 unsigned long head_bit;
1481 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1482 atomic_inc(&cpu_buffer->record_disabled);
1484 * We don't race with the readers since we have acquired the reader
1485 * lock. We also don't race with writers after disabling recording.
1486 * This makes it easy to figure out the first and the last page to be
1487 * removed from the list. We unlink all the pages in between including
1488 * the first and last pages. This is done in a busy loop so that we
1489 * lose the least number of traces.
1490 * The pages are freed after we restart recording and unlock readers.
1492 tail_page = &cpu_buffer->tail_page->list;
1495 * tail page might be on reader page, we remove the next page
1496 * from the ring buffer
1498 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1499 tail_page = rb_list_head(tail_page->next);
1500 to_remove = tail_page;
1502 /* start of pages to remove */
1503 first_page = list_entry(rb_list_head(to_remove->next),
1504 struct buffer_page, list);
1506 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1507 to_remove = rb_list_head(to_remove)->next;
1508 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1511 next_page = rb_list_head(to_remove)->next;
1514 * Now we remove all pages between tail_page and next_page.
1515 * Make sure that we have head_bit value preserved for the
1518 tail_page->next = (struct list_head *)((unsigned long)next_page |
1520 next_page = rb_list_head(next_page);
1521 next_page->prev = tail_page;
1523 /* make sure pages points to a valid page in the ring buffer */
1524 cpu_buffer->pages = next_page;
1526 /* update head page */
1528 cpu_buffer->head_page = list_entry(next_page,
1529 struct buffer_page, list);
1532 * change read pointer to make sure any read iterators reset
1535 cpu_buffer->read = 0;
1537 /* pages are removed, resume tracing and then free the pages */
1538 atomic_dec(&cpu_buffer->record_disabled);
1539 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1541 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1543 /* last buffer page to remove */
1544 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1546 tmp_iter_page = first_page;
1549 to_remove_page = tmp_iter_page;
1550 rb_inc_page(cpu_buffer, &tmp_iter_page);
1552 /* update the counters */
1553 page_entries = rb_page_entries(to_remove_page);
1556 * If something was added to this page, it was full
1557 * since it is not the tail page. So we deduct the
1558 * bytes consumed in ring buffer from here.
1559 * Increment overrun to account for the lost events.
1561 local_add(page_entries, &cpu_buffer->overrun);
1562 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1566 * We have already removed references to this list item, just
1567 * free up the buffer_page and its page
1569 free_buffer_page(to_remove_page);
1572 } while (to_remove_page != last_page);
1574 RB_WARN_ON(cpu_buffer, nr_removed);
1576 return nr_removed == 0;
1580 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1582 struct list_head *pages = &cpu_buffer->new_pages;
1583 int retries, success;
1585 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1587 * We are holding the reader lock, so the reader page won't be swapped
1588 * in the ring buffer. Now we are racing with the writer trying to
1589 * move head page and the tail page.
1590 * We are going to adapt the reader page update process where:
1591 * 1. We first splice the start and end of list of new pages between
1592 * the head page and its previous page.
1593 * 2. We cmpxchg the prev_page->next to point from head page to the
1594 * start of new pages list.
1595 * 3. Finally, we update the head->prev to the end of new list.
1597 * We will try this process 10 times, to make sure that we don't keep
1603 struct list_head *head_page, *prev_page, *r;
1604 struct list_head *last_page, *first_page;
1605 struct list_head *head_page_with_bit;
1607 head_page = &rb_set_head_page(cpu_buffer)->list;
1610 prev_page = head_page->prev;
1612 first_page = pages->next;
1613 last_page = pages->prev;
1615 head_page_with_bit = (struct list_head *)
1616 ((unsigned long)head_page | RB_PAGE_HEAD);
1618 last_page->next = head_page_with_bit;
1619 first_page->prev = prev_page;
1621 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1623 if (r == head_page_with_bit) {
1625 * yay, we replaced the page pointer to our new list,
1626 * now, we just have to update to head page's prev
1627 * pointer to point to end of list
1629 head_page->prev = last_page;
1636 INIT_LIST_HEAD(pages);
1638 * If we weren't successful in adding in new pages, warn and stop
1641 RB_WARN_ON(cpu_buffer, !success);
1642 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1644 /* free pages if they weren't inserted */
1646 struct buffer_page *bpage, *tmp;
1647 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1649 list_del_init(&bpage->list);
1650 free_buffer_page(bpage);
1656 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1660 if (cpu_buffer->nr_pages_to_update > 0)
1661 success = rb_insert_pages(cpu_buffer);
1663 success = rb_remove_pages(cpu_buffer,
1664 -cpu_buffer->nr_pages_to_update);
1667 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1670 static void update_pages_handler(struct work_struct *work)
1672 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1673 struct ring_buffer_per_cpu, update_pages_work);
1674 rb_update_pages(cpu_buffer);
1675 complete(&cpu_buffer->update_done);
1679 * ring_buffer_resize - resize the ring buffer
1680 * @buffer: the buffer to resize.
1681 * @size: the new size.
1682 * @cpu_id: the cpu buffer to resize
1684 * Minimum size is 2 * BUF_PAGE_SIZE.
1686 * Returns 0 on success and < 0 on failure.
1688 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1691 struct ring_buffer_per_cpu *cpu_buffer;
1692 unsigned long nr_pages;
1696 * Always succeed at resizing a non-existent buffer:
1701 /* Make sure the requested buffer exists */
1702 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1703 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1706 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1708 /* we need a minimum of two pages */
1712 size = nr_pages * BUF_PAGE_SIZE;
1715 * Don't succeed if resizing is disabled, as a reader might be
1716 * manipulating the ring buffer and is expecting a sane state while
1719 if (atomic_read(&buffer->resize_disabled))
1722 /* prevent another thread from changing buffer sizes */
1723 mutex_lock(&buffer->mutex);
1725 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1726 /* calculate the pages to update */
1727 for_each_buffer_cpu(buffer, cpu) {
1728 cpu_buffer = buffer->buffers[cpu];
1730 cpu_buffer->nr_pages_to_update = nr_pages -
1731 cpu_buffer->nr_pages;
1733 * nothing more to do for removing pages or no update
1735 if (cpu_buffer->nr_pages_to_update <= 0)
1738 * to add pages, make sure all new pages can be
1739 * allocated without receiving ENOMEM
1741 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1742 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1743 &cpu_buffer->new_pages, cpu)) {
1744 /* not enough memory for new pages */
1752 * Fire off all the required work handlers
1753 * We can't schedule on offline CPUs, but it's not necessary
1754 * since we can change their buffer sizes without any race.
1756 for_each_buffer_cpu(buffer, cpu) {
1757 cpu_buffer = buffer->buffers[cpu];
1758 if (!cpu_buffer->nr_pages_to_update)
1761 /* Can't run something on an offline CPU. */
1762 if (!cpu_online(cpu)) {
1763 rb_update_pages(cpu_buffer);
1764 cpu_buffer->nr_pages_to_update = 0;
1766 schedule_work_on(cpu,
1767 &cpu_buffer->update_pages_work);
1771 /* wait for all the updates to complete */
1772 for_each_buffer_cpu(buffer, cpu) {
1773 cpu_buffer = buffer->buffers[cpu];
1774 if (!cpu_buffer->nr_pages_to_update)
1777 if (cpu_online(cpu))
1778 wait_for_completion(&cpu_buffer->update_done);
1779 cpu_buffer->nr_pages_to_update = 0;
1784 /* Make sure this CPU has been initialized */
1785 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1788 cpu_buffer = buffer->buffers[cpu_id];
1790 if (nr_pages == cpu_buffer->nr_pages)
1793 cpu_buffer->nr_pages_to_update = nr_pages -
1794 cpu_buffer->nr_pages;
1796 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1797 if (cpu_buffer->nr_pages_to_update > 0 &&
1798 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1799 &cpu_buffer->new_pages, cpu_id)) {
1806 /* Can't run something on an offline CPU. */
1807 if (!cpu_online(cpu_id))
1808 rb_update_pages(cpu_buffer);
1810 schedule_work_on(cpu_id,
1811 &cpu_buffer->update_pages_work);
1812 wait_for_completion(&cpu_buffer->update_done);
1815 cpu_buffer->nr_pages_to_update = 0;
1821 * The ring buffer resize can happen with the ring buffer
1822 * enabled, so that the update disturbs the tracing as little
1823 * as possible. But if the buffer is disabled, we do not need
1824 * to worry about that, and we can take the time to verify
1825 * that the buffer is not corrupt.
1827 if (atomic_read(&buffer->record_disabled)) {
1828 atomic_inc(&buffer->record_disabled);
1830 * Even though the buffer was disabled, we must make sure
1831 * that it is truly disabled before calling rb_check_pages.
1832 * There could have been a race between checking
1833 * record_disable and incrementing it.
1835 synchronize_sched();
1836 for_each_buffer_cpu(buffer, cpu) {
1837 cpu_buffer = buffer->buffers[cpu];
1838 rb_check_pages(cpu_buffer);
1840 atomic_dec(&buffer->record_disabled);
1843 mutex_unlock(&buffer->mutex);
1847 for_each_buffer_cpu(buffer, cpu) {
1848 struct buffer_page *bpage, *tmp;
1850 cpu_buffer = buffer->buffers[cpu];
1851 cpu_buffer->nr_pages_to_update = 0;
1853 if (list_empty(&cpu_buffer->new_pages))
1856 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1858 list_del_init(&bpage->list);
1859 free_buffer_page(bpage);
1862 mutex_unlock(&buffer->mutex);
1865 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1867 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1869 mutex_lock(&buffer->mutex);
1871 buffer->flags |= RB_FL_OVERWRITE;
1873 buffer->flags &= ~RB_FL_OVERWRITE;
1874 mutex_unlock(&buffer->mutex);
1876 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1878 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1880 return bpage->page->data + index;
1883 static __always_inline struct ring_buffer_event *
1884 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1886 return __rb_page_index(cpu_buffer->reader_page,
1887 cpu_buffer->reader_page->read);
1890 static __always_inline struct ring_buffer_event *
1891 rb_iter_head_event(struct ring_buffer_iter *iter)
1893 return __rb_page_index(iter->head_page, iter->head);
1896 static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
1898 return local_read(&bpage->page->commit);
1901 /* Size is determined by what has been committed */
1902 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
1904 return rb_page_commit(bpage);
1907 static __always_inline unsigned
1908 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1910 return rb_page_commit(cpu_buffer->commit_page);
1913 static __always_inline unsigned
1914 rb_event_index(struct ring_buffer_event *event)
1916 unsigned long addr = (unsigned long)event;
1918 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1921 static void rb_inc_iter(struct ring_buffer_iter *iter)
1923 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1926 * The iterator could be on the reader page (it starts there).
1927 * But the head could have moved, since the reader was
1928 * found. Check for this case and assign the iterator
1929 * to the head page instead of next.
1931 if (iter->head_page == cpu_buffer->reader_page)
1932 iter->head_page = rb_set_head_page(cpu_buffer);
1934 rb_inc_page(cpu_buffer, &iter->head_page);
1936 iter->read_stamp = iter->head_page->page->time_stamp;
1941 * rb_handle_head_page - writer hit the head page
1943 * Returns: +1 to retry page
1948 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1949 struct buffer_page *tail_page,
1950 struct buffer_page *next_page)
1952 struct buffer_page *new_head;
1957 entries = rb_page_entries(next_page);
1960 * The hard part is here. We need to move the head
1961 * forward, and protect against both readers on
1962 * other CPUs and writers coming in via interrupts.
1964 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1968 * type can be one of four:
1969 * NORMAL - an interrupt already moved it for us
1970 * HEAD - we are the first to get here.
1971 * UPDATE - we are the interrupt interrupting
1973 * MOVED - a reader on another CPU moved the next
1974 * pointer to its reader page. Give up
1981 * We changed the head to UPDATE, thus
1982 * it is our responsibility to update
1985 local_add(entries, &cpu_buffer->overrun);
1986 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1989 * The entries will be zeroed out when we move the
1993 /* still more to do */
1996 case RB_PAGE_UPDATE:
1998 * This is an interrupt that interrupt the
1999 * previous update. Still more to do.
2002 case RB_PAGE_NORMAL:
2004 * An interrupt came in before the update
2005 * and processed this for us.
2006 * Nothing left to do.
2011 * The reader is on another CPU and just did
2012 * a swap with our next_page.
2017 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2022 * Now that we are here, the old head pointer is
2023 * set to UPDATE. This will keep the reader from
2024 * swapping the head page with the reader page.
2025 * The reader (on another CPU) will spin till
2028 * We just need to protect against interrupts
2029 * doing the job. We will set the next pointer
2030 * to HEAD. After that, we set the old pointer
2031 * to NORMAL, but only if it was HEAD before.
2032 * otherwise we are an interrupt, and only
2033 * want the outer most commit to reset it.
2035 new_head = next_page;
2036 rb_inc_page(cpu_buffer, &new_head);
2038 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2042 * Valid returns are:
2043 * HEAD - an interrupt came in and already set it.
2044 * NORMAL - One of two things:
2045 * 1) We really set it.
2046 * 2) A bunch of interrupts came in and moved
2047 * the page forward again.
2051 case RB_PAGE_NORMAL:
2055 RB_WARN_ON(cpu_buffer, 1);
2060 * It is possible that an interrupt came in,
2061 * set the head up, then more interrupts came in
2062 * and moved it again. When we get back here,
2063 * the page would have been set to NORMAL but we
2064 * just set it back to HEAD.
2066 * How do you detect this? Well, if that happened
2067 * the tail page would have moved.
2069 if (ret == RB_PAGE_NORMAL) {
2070 struct buffer_page *buffer_tail_page;
2072 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2074 * If the tail had moved passed next, then we need
2075 * to reset the pointer.
2077 if (buffer_tail_page != tail_page &&
2078 buffer_tail_page != next_page)
2079 rb_head_page_set_normal(cpu_buffer, new_head,
2085 * If this was the outer most commit (the one that
2086 * changed the original pointer from HEAD to UPDATE),
2087 * then it is up to us to reset it to NORMAL.
2089 if (type == RB_PAGE_HEAD) {
2090 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2093 if (RB_WARN_ON(cpu_buffer,
2094 ret != RB_PAGE_UPDATE))
2102 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2103 unsigned long tail, struct rb_event_info *info)
2105 struct buffer_page *tail_page = info->tail_page;
2106 struct ring_buffer_event *event;
2107 unsigned long length = info->length;
2110 * Only the event that crossed the page boundary
2111 * must fill the old tail_page with padding.
2113 if (tail >= BUF_PAGE_SIZE) {
2115 * If the page was filled, then we still need
2116 * to update the real_end. Reset it to zero
2117 * and the reader will ignore it.
2119 if (tail == BUF_PAGE_SIZE)
2120 tail_page->real_end = 0;
2122 local_sub(length, &tail_page->write);
2126 event = __rb_page_index(tail_page, tail);
2128 /* account for padding bytes */
2129 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2132 * Save the original length to the meta data.
2133 * This will be used by the reader to add lost event
2136 tail_page->real_end = tail;
2139 * If this event is bigger than the minimum size, then
2140 * we need to be careful that we don't subtract the
2141 * write counter enough to allow another writer to slip
2143 * We put in a discarded commit instead, to make sure
2144 * that this space is not used again.
2146 * If we are less than the minimum size, we don't need to
2149 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2150 /* No room for any events */
2152 /* Mark the rest of the page with padding */
2153 rb_event_set_padding(event);
2155 /* Set the write back to the previous setting */
2156 local_sub(length, &tail_page->write);
2160 /* Put in a discarded event */
2161 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2162 event->type_len = RINGBUF_TYPE_PADDING;
2163 /* time delta must be non zero */
2164 event->time_delta = 1;
2166 /* Set write to end of buffer */
2167 length = (tail + length) - BUF_PAGE_SIZE;
2168 local_sub(length, &tail_page->write);
2171 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2174 * This is the slow path, force gcc not to inline it.
2176 static noinline struct ring_buffer_event *
2177 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2178 unsigned long tail, struct rb_event_info *info)
2180 struct buffer_page *tail_page = info->tail_page;
2181 struct buffer_page *commit_page = cpu_buffer->commit_page;
2182 struct ring_buffer *buffer = cpu_buffer->buffer;
2183 struct buffer_page *next_page;
2186 next_page = tail_page;
2188 rb_inc_page(cpu_buffer, &next_page);
2191 * If for some reason, we had an interrupt storm that made
2192 * it all the way around the buffer, bail, and warn
2195 if (unlikely(next_page == commit_page)) {
2196 local_inc(&cpu_buffer->commit_overrun);
2201 * This is where the fun begins!
2203 * We are fighting against races between a reader that
2204 * could be on another CPU trying to swap its reader
2205 * page with the buffer head.
2207 * We are also fighting against interrupts coming in and
2208 * moving the head or tail on us as well.
2210 * If the next page is the head page then we have filled
2211 * the buffer, unless the commit page is still on the
2214 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2217 * If the commit is not on the reader page, then
2218 * move the header page.
2220 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2222 * If we are not in overwrite mode,
2223 * this is easy, just stop here.
2225 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2226 local_inc(&cpu_buffer->dropped_events);
2230 ret = rb_handle_head_page(cpu_buffer,
2239 * We need to be careful here too. The
2240 * commit page could still be on the reader
2241 * page. We could have a small buffer, and
2242 * have filled up the buffer with events
2243 * from interrupts and such, and wrapped.
2245 * Note, if the tail page is also the on the
2246 * reader_page, we let it move out.
2248 if (unlikely((cpu_buffer->commit_page !=
2249 cpu_buffer->tail_page) &&
2250 (cpu_buffer->commit_page ==
2251 cpu_buffer->reader_page))) {
2252 local_inc(&cpu_buffer->commit_overrun);
2258 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2262 rb_reset_tail(cpu_buffer, tail, info);
2264 /* Commit what we have for now. */
2265 rb_end_commit(cpu_buffer);
2266 /* rb_end_commit() decs committing */
2267 local_inc(&cpu_buffer->committing);
2269 /* fail and let the caller try again */
2270 return ERR_PTR(-EAGAIN);
2274 rb_reset_tail(cpu_buffer, tail, info);
2279 /* Slow path, do not inline */
2280 static noinline struct ring_buffer_event *
2281 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2284 event->type_len = RINGBUF_TYPE_TIME_STAMP;
2286 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2288 /* Not the first event on the page, or not delta? */
2289 if (abs || rb_event_index(event)) {
2290 event->time_delta = delta & TS_MASK;
2291 event->array[0] = delta >> TS_SHIFT;
2293 /* nope, just zero it */
2294 event->time_delta = 0;
2295 event->array[0] = 0;
2298 return skip_time_extend(event);
2301 static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2302 struct ring_buffer_event *event);
2305 * rb_update_event - update event type and data
2306 * @event: the event to update
2307 * @type: the type of event
2308 * @length: the size of the event field in the ring buffer
2310 * Update the type and data fields of the event. The length
2311 * is the actual size that is written to the ring buffer,
2312 * and with this, we can determine what to place into the
2316 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2317 struct ring_buffer_event *event,
2318 struct rb_event_info *info)
2320 unsigned length = info->length;
2321 u64 delta = info->delta;
2323 /* Only a commit updates the timestamp */
2324 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2328 * If we need to add a timestamp, then we
2329 * add it to the start of the reserved space.
2331 if (unlikely(info->add_timestamp)) {
2332 bool abs = ring_buffer_time_stamp_abs(cpu_buffer->buffer);
2334 event = rb_add_time_stamp(event, info->delta, abs);
2335 length -= RB_LEN_TIME_EXTEND;
2339 event->time_delta = delta;
2340 length -= RB_EVNT_HDR_SIZE;
2341 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2342 event->type_len = 0;
2343 event->array[0] = length;
2345 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2348 static unsigned rb_calculate_event_length(unsigned length)
2350 struct ring_buffer_event event; /* Used only for sizeof array */
2352 /* zero length can cause confusions */
2356 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2357 length += sizeof(event.array[0]);
2359 length += RB_EVNT_HDR_SIZE;
2360 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2363 * In case the time delta is larger than the 27 bits for it
2364 * in the header, we need to add a timestamp. If another
2365 * event comes in when trying to discard this one to increase
2366 * the length, then the timestamp will be added in the allocated
2367 * space of this event. If length is bigger than the size needed
2368 * for the TIME_EXTEND, then padding has to be used. The events
2369 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2370 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2371 * As length is a multiple of 4, we only need to worry if it
2372 * is 12 (RB_LEN_TIME_EXTEND + 4).
2374 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2375 length += RB_ALIGNMENT;
2380 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2381 static inline bool sched_clock_stable(void)
2388 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2389 struct ring_buffer_event *event)
2391 unsigned long new_index, old_index;
2392 struct buffer_page *bpage;
2393 unsigned long index;
2396 new_index = rb_event_index(event);
2397 old_index = new_index + rb_event_ts_length(event);
2398 addr = (unsigned long)event;
2401 bpage = READ_ONCE(cpu_buffer->tail_page);
2403 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2404 unsigned long write_mask =
2405 local_read(&bpage->write) & ~RB_WRITE_MASK;
2406 unsigned long event_length = rb_event_length(event);
2408 * This is on the tail page. It is possible that
2409 * a write could come in and move the tail page
2410 * and write to the next page. That is fine
2411 * because we just shorten what is on this page.
2413 old_index += write_mask;
2414 new_index += write_mask;
2415 index = local_cmpxchg(&bpage->write, old_index, new_index);
2416 if (index == old_index) {
2417 /* update counters */
2418 local_sub(event_length, &cpu_buffer->entries_bytes);
2423 /* could not discard */
2427 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2429 local_inc(&cpu_buffer->committing);
2430 local_inc(&cpu_buffer->commits);
2433 static __always_inline void
2434 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2436 unsigned long max_count;
2439 * We only race with interrupts and NMIs on this CPU.
2440 * If we own the commit event, then we can commit
2441 * all others that interrupted us, since the interruptions
2442 * are in stack format (they finish before they come
2443 * back to us). This allows us to do a simple loop to
2444 * assign the commit to the tail.
2447 max_count = cpu_buffer->nr_pages * 100;
2449 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2450 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2452 if (RB_WARN_ON(cpu_buffer,
2453 rb_is_reader_page(cpu_buffer->tail_page)))
2455 local_set(&cpu_buffer->commit_page->page->commit,
2456 rb_page_write(cpu_buffer->commit_page));
2457 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2458 /* Only update the write stamp if the page has an event */
2459 if (rb_page_write(cpu_buffer->commit_page))
2460 cpu_buffer->write_stamp =
2461 cpu_buffer->commit_page->page->time_stamp;
2462 /* add barrier to keep gcc from optimizing too much */
2465 while (rb_commit_index(cpu_buffer) !=
2466 rb_page_write(cpu_buffer->commit_page)) {
2468 local_set(&cpu_buffer->commit_page->page->commit,
2469 rb_page_write(cpu_buffer->commit_page));
2470 RB_WARN_ON(cpu_buffer,
2471 local_read(&cpu_buffer->commit_page->page->commit) &
2476 /* again, keep gcc from optimizing */
2480 * If an interrupt came in just after the first while loop
2481 * and pushed the tail page forward, we will be left with
2482 * a dangling commit that will never go forward.
2484 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2488 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2490 unsigned long commits;
2492 if (RB_WARN_ON(cpu_buffer,
2493 !local_read(&cpu_buffer->committing)))
2497 commits = local_read(&cpu_buffer->commits);
2498 /* synchronize with interrupts */
2500 if (local_read(&cpu_buffer->committing) == 1)
2501 rb_set_commit_to_write(cpu_buffer);
2503 local_dec(&cpu_buffer->committing);
2505 /* synchronize with interrupts */
2509 * Need to account for interrupts coming in between the
2510 * updating of the commit page and the clearing of the
2511 * committing counter.
2513 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2514 !local_read(&cpu_buffer->committing)) {
2515 local_inc(&cpu_buffer->committing);
2520 static inline void rb_event_discard(struct ring_buffer_event *event)
2522 if (extended_time(event))
2523 event = skip_time_extend(event);
2525 /* array[0] holds the actual length for the discarded event */
2526 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2527 event->type_len = RINGBUF_TYPE_PADDING;
2528 /* time delta must be non zero */
2529 if (!event->time_delta)
2530 event->time_delta = 1;
2533 static __always_inline bool
2534 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2535 struct ring_buffer_event *event)
2537 unsigned long addr = (unsigned long)event;
2538 unsigned long index;
2540 index = rb_event_index(event);
2543 return cpu_buffer->commit_page->page == (void *)addr &&
2544 rb_commit_index(cpu_buffer) == index;
2547 static __always_inline void
2548 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2549 struct ring_buffer_event *event)
2554 * The event first in the commit queue updates the
2557 if (rb_event_is_commit(cpu_buffer, event)) {
2559 * A commit event that is first on a page
2560 * updates the write timestamp with the page stamp
2562 if (!rb_event_index(event))
2563 cpu_buffer->write_stamp =
2564 cpu_buffer->commit_page->page->time_stamp;
2565 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2566 delta = ring_buffer_event_time_stamp(event);
2567 cpu_buffer->write_stamp += delta;
2568 } else if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
2569 delta = ring_buffer_event_time_stamp(event);
2570 cpu_buffer->write_stamp = delta;
2572 cpu_buffer->write_stamp += event->time_delta;
2576 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2577 struct ring_buffer_event *event)
2579 local_inc(&cpu_buffer->entries);
2580 rb_update_write_stamp(cpu_buffer, event);
2581 rb_end_commit(cpu_buffer);
2584 static __always_inline void
2585 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2589 if (buffer->irq_work.waiters_pending) {
2590 buffer->irq_work.waiters_pending = false;
2591 /* irq_work_queue() supplies it's own memory barriers */
2592 irq_work_queue(&buffer->irq_work.work);
2595 if (cpu_buffer->irq_work.waiters_pending) {
2596 cpu_buffer->irq_work.waiters_pending = false;
2597 /* irq_work_queue() supplies it's own memory barriers */
2598 irq_work_queue(&cpu_buffer->irq_work.work);
2601 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
2603 if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2604 cpu_buffer->irq_work.wakeup_full = true;
2605 cpu_buffer->irq_work.full_waiters_pending = false;
2606 /* irq_work_queue() supplies it's own memory barriers */
2607 irq_work_queue(&cpu_buffer->irq_work.work);
2612 * The lock and unlock are done within a preempt disable section.
2613 * The current_context per_cpu variable can only be modified
2614 * by the current task between lock and unlock. But it can
2615 * be modified more than once via an interrupt. To pass this
2616 * information from the lock to the unlock without having to
2617 * access the 'in_interrupt()' functions again (which do show
2618 * a bit of overhead in something as critical as function tracing,
2619 * we use a bitmask trick.
2621 * bit 0 = NMI context
2622 * bit 1 = IRQ context
2623 * bit 2 = SoftIRQ context
2624 * bit 3 = normal context.
2626 * This works because this is the order of contexts that can
2627 * preempt other contexts. A SoftIRQ never preempts an IRQ
2630 * When the context is determined, the corresponding bit is
2631 * checked and set (if it was set, then a recursion of that context
2634 * On unlock, we need to clear this bit. To do so, just subtract
2635 * 1 from the current_context and AND it to itself.
2639 * 101 & 100 = 100 (clearing bit zero)
2642 * 1010 & 1001 = 1000 (clearing bit 1)
2644 * The least significant bit can be cleared this way, and it
2645 * just so happens that it is the same bit corresponding to
2646 * the current context.
2649 static __always_inline int
2650 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2652 unsigned int val = cpu_buffer->current_context;
2653 unsigned long pc = preempt_count();
2656 if (!(pc & (NMI_MASK | HARDIRQ_MASK | SOFTIRQ_OFFSET)))
2657 bit = RB_CTX_NORMAL;
2659 bit = pc & NMI_MASK ? RB_CTX_NMI :
2660 pc & HARDIRQ_MASK ? RB_CTX_IRQ : RB_CTX_SOFTIRQ;
2662 if (unlikely(val & (1 << (bit + cpu_buffer->nest))))
2665 val |= (1 << (bit + cpu_buffer->nest));
2666 cpu_buffer->current_context = val;
2671 static __always_inline void
2672 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2674 cpu_buffer->current_context &=
2675 cpu_buffer->current_context - (1 << cpu_buffer->nest);
2678 /* The recursive locking above uses 4 bits */
2679 #define NESTED_BITS 4
2682 * ring_buffer_nest_start - Allow to trace while nested
2683 * @buffer: The ring buffer to modify
2685 * The ring buffer has a safety mechanism to prevent recursion.
2686 * But there may be a case where a trace needs to be done while
2687 * tracing something else. In this case, calling this function
2688 * will allow this function to nest within a currently active
2689 * ring_buffer_lock_reserve().
2691 * Call this function before calling another ring_buffer_lock_reserve() and
2692 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
2694 void ring_buffer_nest_start(struct ring_buffer *buffer)
2696 struct ring_buffer_per_cpu *cpu_buffer;
2699 /* Enabled by ring_buffer_nest_end() */
2700 preempt_disable_notrace();
2701 cpu = raw_smp_processor_id();
2702 cpu_buffer = buffer->buffers[cpu];
2703 /* This is the shift value for the above recursive locking */
2704 cpu_buffer->nest += NESTED_BITS;
2708 * ring_buffer_nest_end - Allow to trace while nested
2709 * @buffer: The ring buffer to modify
2711 * Must be called after ring_buffer_nest_start() and after the
2712 * ring_buffer_unlock_commit().
2714 void ring_buffer_nest_end(struct ring_buffer *buffer)
2716 struct ring_buffer_per_cpu *cpu_buffer;
2719 /* disabled by ring_buffer_nest_start() */
2720 cpu = raw_smp_processor_id();
2721 cpu_buffer = buffer->buffers[cpu];
2722 /* This is the shift value for the above recursive locking */
2723 cpu_buffer->nest -= NESTED_BITS;
2724 preempt_enable_notrace();
2728 * ring_buffer_unlock_commit - commit a reserved
2729 * @buffer: The buffer to commit to
2730 * @event: The event pointer to commit.
2732 * This commits the data to the ring buffer, and releases any locks held.
2734 * Must be paired with ring_buffer_lock_reserve.
2736 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2737 struct ring_buffer_event *event)
2739 struct ring_buffer_per_cpu *cpu_buffer;
2740 int cpu = raw_smp_processor_id();
2742 cpu_buffer = buffer->buffers[cpu];
2744 rb_commit(cpu_buffer, event);
2746 rb_wakeups(buffer, cpu_buffer);
2748 trace_recursive_unlock(cpu_buffer);
2750 preempt_enable_notrace();
2754 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2756 static noinline void
2757 rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2758 struct rb_event_info *info)
2760 WARN_ONCE(info->delta > (1ULL << 59),
2761 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2762 (unsigned long long)info->delta,
2763 (unsigned long long)info->ts,
2764 (unsigned long long)cpu_buffer->write_stamp,
2765 sched_clock_stable() ? "" :
2766 "If you just came from a suspend/resume,\n"
2767 "please switch to the trace global clock:\n"
2768 " echo global > /sys/kernel/debug/tracing/trace_clock\n"
2769 "or add trace_clock=global to the kernel command line\n");
2770 info->add_timestamp = 1;
2773 static struct ring_buffer_event *
2774 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2775 struct rb_event_info *info)
2777 struct ring_buffer_event *event;
2778 struct buffer_page *tail_page;
2779 unsigned long tail, write;
2782 * If the time delta since the last event is too big to
2783 * hold in the time field of the event, then we append a
2784 * TIME EXTEND event ahead of the data event.
2786 if (unlikely(info->add_timestamp))
2787 info->length += RB_LEN_TIME_EXTEND;
2789 /* Don't let the compiler play games with cpu_buffer->tail_page */
2790 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2791 write = local_add_return(info->length, &tail_page->write);
2793 /* set write to only the index of the write */
2794 write &= RB_WRITE_MASK;
2795 tail = write - info->length;
2798 * If this is the first commit on the page, then it has the same
2799 * timestamp as the page itself.
2801 if (!tail && !ring_buffer_time_stamp_abs(cpu_buffer->buffer))
2804 /* See if we shot pass the end of this buffer page */
2805 if (unlikely(write > BUF_PAGE_SIZE))
2806 return rb_move_tail(cpu_buffer, tail, info);
2808 /* We reserved something on the buffer */
2810 event = __rb_page_index(tail_page, tail);
2811 rb_update_event(cpu_buffer, event, info);
2813 local_inc(&tail_page->entries);
2816 * If this is the first commit on the page, then update
2820 tail_page->page->time_stamp = info->ts;
2822 /* account for these added bytes */
2823 local_add(info->length, &cpu_buffer->entries_bytes);
2828 static __always_inline struct ring_buffer_event *
2829 rb_reserve_next_event(struct ring_buffer *buffer,
2830 struct ring_buffer_per_cpu *cpu_buffer,
2831 unsigned long length)
2833 struct ring_buffer_event *event;
2834 struct rb_event_info info;
2838 rb_start_commit(cpu_buffer);
2840 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2842 * Due to the ability to swap a cpu buffer from a buffer
2843 * it is possible it was swapped before we committed.
2844 * (committing stops a swap). We check for it here and
2845 * if it happened, we have to fail the write.
2848 if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
2849 local_dec(&cpu_buffer->committing);
2850 local_dec(&cpu_buffer->commits);
2855 info.length = rb_calculate_event_length(length);
2857 info.add_timestamp = 0;
2861 * We allow for interrupts to reenter here and do a trace.
2862 * If one does, it will cause this original code to loop
2863 * back here. Even with heavy interrupts happening, this
2864 * should only happen a few times in a row. If this happens
2865 * 1000 times in a row, there must be either an interrupt
2866 * storm or we have something buggy.
2869 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2872 info.ts = rb_time_stamp(cpu_buffer->buffer);
2873 diff = info.ts - cpu_buffer->write_stamp;
2875 /* make sure this diff is calculated here */
2878 if (ring_buffer_time_stamp_abs(buffer)) {
2879 info.delta = info.ts;
2880 rb_handle_timestamp(cpu_buffer, &info);
2881 } else /* Did the write stamp get updated already? */
2882 if (likely(info.ts >= cpu_buffer->write_stamp)) {
2884 if (unlikely(test_time_stamp(info.delta)))
2885 rb_handle_timestamp(cpu_buffer, &info);
2888 event = __rb_reserve_next(cpu_buffer, &info);
2890 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2891 if (info.add_timestamp)
2892 info.length -= RB_LEN_TIME_EXTEND;
2902 rb_end_commit(cpu_buffer);
2907 * ring_buffer_lock_reserve - reserve a part of the buffer
2908 * @buffer: the ring buffer to reserve from
2909 * @length: the length of the data to reserve (excluding event header)
2911 * Returns a reserved event on the ring buffer to copy directly to.
2912 * The user of this interface will need to get the body to write into
2913 * and can use the ring_buffer_event_data() interface.
2915 * The length is the length of the data needed, not the event length
2916 * which also includes the event header.
2918 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2919 * If NULL is returned, then nothing has been allocated or locked.
2921 struct ring_buffer_event *
2922 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2924 struct ring_buffer_per_cpu *cpu_buffer;
2925 struct ring_buffer_event *event;
2928 /* If we are tracing schedule, we don't want to recurse */
2929 preempt_disable_notrace();
2931 if (unlikely(atomic_read(&buffer->record_disabled)))
2934 cpu = raw_smp_processor_id();
2936 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2939 cpu_buffer = buffer->buffers[cpu];
2941 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2944 if (unlikely(length > BUF_MAX_DATA_SIZE))
2947 if (unlikely(trace_recursive_lock(cpu_buffer)))
2950 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2957 trace_recursive_unlock(cpu_buffer);
2959 preempt_enable_notrace();
2962 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2965 * Decrement the entries to the page that an event is on.
2966 * The event does not even need to exist, only the pointer
2967 * to the page it is on. This may only be called before the commit
2971 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2972 struct ring_buffer_event *event)
2974 unsigned long addr = (unsigned long)event;
2975 struct buffer_page *bpage = cpu_buffer->commit_page;
2976 struct buffer_page *start;
2980 /* Do the likely case first */
2981 if (likely(bpage->page == (void *)addr)) {
2982 local_dec(&bpage->entries);
2987 * Because the commit page may be on the reader page we
2988 * start with the next page and check the end loop there.
2990 rb_inc_page(cpu_buffer, &bpage);
2993 if (bpage->page == (void *)addr) {
2994 local_dec(&bpage->entries);
2997 rb_inc_page(cpu_buffer, &bpage);
2998 } while (bpage != start);
3000 /* commit not part of this buffer?? */
3001 RB_WARN_ON(cpu_buffer, 1);
3005 * ring_buffer_commit_discard - discard an event that has not been committed
3006 * @buffer: the ring buffer
3007 * @event: non committed event to discard
3009 * Sometimes an event that is in the ring buffer needs to be ignored.
3010 * This function lets the user discard an event in the ring buffer
3011 * and then that event will not be read later.
3013 * This function only works if it is called before the item has been
3014 * committed. It will try to free the event from the ring buffer
3015 * if another event has not been added behind it.
3017 * If another event has been added behind it, it will set the event
3018 * up as discarded, and perform the commit.
3020 * If this function is called, do not call ring_buffer_unlock_commit on
3023 void ring_buffer_discard_commit(struct ring_buffer *buffer,
3024 struct ring_buffer_event *event)
3026 struct ring_buffer_per_cpu *cpu_buffer;
3029 /* The event is discarded regardless */
3030 rb_event_discard(event);
3032 cpu = smp_processor_id();
3033 cpu_buffer = buffer->buffers[cpu];
3036 * This must only be called if the event has not been
3037 * committed yet. Thus we can assume that preemption
3038 * is still disabled.
3040 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3042 rb_decrement_entry(cpu_buffer, event);
3043 if (rb_try_to_discard(cpu_buffer, event))
3047 * The commit is still visible by the reader, so we
3048 * must still update the timestamp.
3050 rb_update_write_stamp(cpu_buffer, event);
3052 rb_end_commit(cpu_buffer);
3054 trace_recursive_unlock(cpu_buffer);
3056 preempt_enable_notrace();
3059 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3062 * ring_buffer_write - write data to the buffer without reserving
3063 * @buffer: The ring buffer to write to.
3064 * @length: The length of the data being written (excluding the event header)
3065 * @data: The data to write to the buffer.
3067 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3068 * one function. If you already have the data to write to the buffer, it
3069 * may be easier to simply call this function.
3071 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3072 * and not the length of the event which would hold the header.
3074 int ring_buffer_write(struct ring_buffer *buffer,
3075 unsigned long length,
3078 struct ring_buffer_per_cpu *cpu_buffer;
3079 struct ring_buffer_event *event;
3084 preempt_disable_notrace();
3086 if (atomic_read(&buffer->record_disabled))
3089 cpu = raw_smp_processor_id();
3091 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3094 cpu_buffer = buffer->buffers[cpu];
3096 if (atomic_read(&cpu_buffer->record_disabled))
3099 if (length > BUF_MAX_DATA_SIZE)
3102 if (unlikely(trace_recursive_lock(cpu_buffer)))
3105 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3109 body = rb_event_data(event);
3111 memcpy(body, data, length);
3113 rb_commit(cpu_buffer, event);
3115 rb_wakeups(buffer, cpu_buffer);
3120 trace_recursive_unlock(cpu_buffer);
3123 preempt_enable_notrace();
3127 EXPORT_SYMBOL_GPL(ring_buffer_write);
3129 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3131 struct buffer_page *reader = cpu_buffer->reader_page;
3132 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3133 struct buffer_page *commit = cpu_buffer->commit_page;
3135 /* In case of error, head will be NULL */
3136 if (unlikely(!head))
3139 return reader->read == rb_page_commit(reader) &&
3140 (commit == reader ||
3142 head->read == rb_page_commit(commit)));
3146 * ring_buffer_record_disable - stop all writes into the buffer
3147 * @buffer: The ring buffer to stop writes to.
3149 * This prevents all writes to the buffer. Any attempt to write
3150 * to the buffer after this will fail and return NULL.
3152 * The caller should call synchronize_sched() after this.
3154 void ring_buffer_record_disable(struct ring_buffer *buffer)
3156 atomic_inc(&buffer->record_disabled);
3158 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3161 * ring_buffer_record_enable - enable writes to the buffer
3162 * @buffer: The ring buffer to enable writes
3164 * Note, multiple disables will need the same number of enables
3165 * to truly enable the writing (much like preempt_disable).
3167 void ring_buffer_record_enable(struct ring_buffer *buffer)
3169 atomic_dec(&buffer->record_disabled);
3171 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3174 * ring_buffer_record_off - stop all writes into the buffer
3175 * @buffer: The ring buffer to stop writes to.
3177 * This prevents all writes to the buffer. Any attempt to write
3178 * to the buffer after this will fail and return NULL.
3180 * This is different than ring_buffer_record_disable() as
3181 * it works like an on/off switch, where as the disable() version
3182 * must be paired with a enable().
3184 void ring_buffer_record_off(struct ring_buffer *buffer)
3187 unsigned int new_rd;
3190 rd = atomic_read(&buffer->record_disabled);
3191 new_rd = rd | RB_BUFFER_OFF;
3192 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3194 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3197 * ring_buffer_record_on - restart writes into the buffer
3198 * @buffer: The ring buffer to start writes to.
3200 * This enables all writes to the buffer that was disabled by
3201 * ring_buffer_record_off().
3203 * This is different than ring_buffer_record_enable() as
3204 * it works like an on/off switch, where as the enable() version
3205 * must be paired with a disable().
3207 void ring_buffer_record_on(struct ring_buffer *buffer)
3210 unsigned int new_rd;
3213 rd = atomic_read(&buffer->record_disabled);
3214 new_rd = rd & ~RB_BUFFER_OFF;
3215 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3217 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3220 * ring_buffer_record_is_on - return true if the ring buffer can write
3221 * @buffer: The ring buffer to see if write is enabled
3223 * Returns true if the ring buffer is in a state that it accepts writes.
3225 bool ring_buffer_record_is_on(struct ring_buffer *buffer)
3227 return !atomic_read(&buffer->record_disabled);
3231 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
3232 * @buffer: The ring buffer to see if write is set enabled
3234 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
3235 * Note that this does NOT mean it is in a writable state.
3237 * It may return true when the ring buffer has been disabled by
3238 * ring_buffer_record_disable(), as that is a temporary disabling of
3241 bool ring_buffer_record_is_set_on(struct ring_buffer *buffer)
3243 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
3247 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3248 * @buffer: The ring buffer to stop writes to.
3249 * @cpu: The CPU buffer to stop
3251 * This prevents all writes to the buffer. Any attempt to write
3252 * to the buffer after this will fail and return NULL.
3254 * The caller should call synchronize_sched() after this.
3256 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3258 struct ring_buffer_per_cpu *cpu_buffer;
3260 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3263 cpu_buffer = buffer->buffers[cpu];
3264 atomic_inc(&cpu_buffer->record_disabled);
3266 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3269 * ring_buffer_record_enable_cpu - enable writes to the buffer
3270 * @buffer: The ring buffer to enable writes
3271 * @cpu: The CPU to enable.
3273 * Note, multiple disables will need the same number of enables
3274 * to truly enable the writing (much like preempt_disable).
3276 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3278 struct ring_buffer_per_cpu *cpu_buffer;
3280 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3283 cpu_buffer = buffer->buffers[cpu];
3284 atomic_dec(&cpu_buffer->record_disabled);
3286 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3289 * The total entries in the ring buffer is the running counter
3290 * of entries entered into the ring buffer, minus the sum of
3291 * the entries read from the ring buffer and the number of
3292 * entries that were overwritten.
3294 static inline unsigned long
3295 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3297 return local_read(&cpu_buffer->entries) -
3298 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3302 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3303 * @buffer: The ring buffer
3304 * @cpu: The per CPU buffer to read from.
3306 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3308 unsigned long flags;
3309 struct ring_buffer_per_cpu *cpu_buffer;
3310 struct buffer_page *bpage;
3313 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3316 cpu_buffer = buffer->buffers[cpu];
3317 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3319 * if the tail is on reader_page, oldest time stamp is on the reader
3322 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3323 bpage = cpu_buffer->reader_page;
3325 bpage = rb_set_head_page(cpu_buffer);
3327 ret = bpage->page->time_stamp;
3328 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3332 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3335 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3336 * @buffer: The ring buffer
3337 * @cpu: The per CPU buffer to read from.
3339 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3341 struct ring_buffer_per_cpu *cpu_buffer;
3344 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3347 cpu_buffer = buffer->buffers[cpu];
3348 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3352 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3355 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3356 * @buffer: The ring buffer
3357 * @cpu: The per CPU buffer to get the entries from.
3359 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3361 struct ring_buffer_per_cpu *cpu_buffer;
3363 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3366 cpu_buffer = buffer->buffers[cpu];
3368 return rb_num_of_entries(cpu_buffer);
3370 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3373 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3374 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3375 * @buffer: The ring buffer
3376 * @cpu: The per CPU buffer to get the number of overruns from
3378 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3380 struct ring_buffer_per_cpu *cpu_buffer;
3383 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3386 cpu_buffer = buffer->buffers[cpu];
3387 ret = local_read(&cpu_buffer->overrun);
3391 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3394 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3395 * commits failing due to the buffer wrapping around while there are uncommitted
3396 * events, such as during an interrupt storm.
3397 * @buffer: The ring buffer
3398 * @cpu: The per CPU buffer to get the number of overruns from
3401 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3403 struct ring_buffer_per_cpu *cpu_buffer;
3406 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3409 cpu_buffer = buffer->buffers[cpu];
3410 ret = local_read(&cpu_buffer->commit_overrun);
3414 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3417 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3418 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3419 * @buffer: The ring buffer
3420 * @cpu: The per CPU buffer to get the number of overruns from
3423 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3425 struct ring_buffer_per_cpu *cpu_buffer;
3428 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3431 cpu_buffer = buffer->buffers[cpu];
3432 ret = local_read(&cpu_buffer->dropped_events);
3436 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3439 * ring_buffer_read_events_cpu - get the number of events successfully read
3440 * @buffer: The ring buffer
3441 * @cpu: The per CPU buffer to get the number of events read
3444 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3446 struct ring_buffer_per_cpu *cpu_buffer;
3448 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3451 cpu_buffer = buffer->buffers[cpu];
3452 return cpu_buffer->read;
3454 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3457 * ring_buffer_entries - get the number of entries in a buffer
3458 * @buffer: The ring buffer
3460 * Returns the total number of entries in the ring buffer
3463 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3465 struct ring_buffer_per_cpu *cpu_buffer;
3466 unsigned long entries = 0;
3469 /* if you care about this being correct, lock the buffer */
3470 for_each_buffer_cpu(buffer, cpu) {
3471 cpu_buffer = buffer->buffers[cpu];
3472 entries += rb_num_of_entries(cpu_buffer);
3477 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3480 * ring_buffer_overruns - get the number of overruns in buffer
3481 * @buffer: The ring buffer
3483 * Returns the total number of overruns in the ring buffer
3486 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3488 struct ring_buffer_per_cpu *cpu_buffer;
3489 unsigned long overruns = 0;
3492 /* if you care about this being correct, lock the buffer */
3493 for_each_buffer_cpu(buffer, cpu) {
3494 cpu_buffer = buffer->buffers[cpu];
3495 overruns += local_read(&cpu_buffer->overrun);
3500 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3502 static void rb_iter_reset(struct ring_buffer_iter *iter)
3504 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3506 /* Iterator usage is expected to have record disabled */
3507 iter->head_page = cpu_buffer->reader_page;
3508 iter->head = cpu_buffer->reader_page->read;
3510 iter->cache_reader_page = iter->head_page;
3511 iter->cache_read = cpu_buffer->read;
3514 iter->read_stamp = cpu_buffer->read_stamp;
3516 iter->read_stamp = iter->head_page->page->time_stamp;
3520 * ring_buffer_iter_reset - reset an iterator
3521 * @iter: The iterator to reset
3523 * Resets the iterator, so that it will start from the beginning
3526 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3528 struct ring_buffer_per_cpu *cpu_buffer;
3529 unsigned long flags;
3534 cpu_buffer = iter->cpu_buffer;
3536 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3537 rb_iter_reset(iter);
3538 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3540 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3543 * ring_buffer_iter_empty - check if an iterator has no more to read
3544 * @iter: The iterator to check
3546 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3548 struct ring_buffer_per_cpu *cpu_buffer;
3549 struct buffer_page *reader;
3550 struct buffer_page *head_page;
3551 struct buffer_page *commit_page;
3554 cpu_buffer = iter->cpu_buffer;
3556 /* Remember, trace recording is off when iterator is in use */
3557 reader = cpu_buffer->reader_page;
3558 head_page = cpu_buffer->head_page;
3559 commit_page = cpu_buffer->commit_page;
3560 commit = rb_page_commit(commit_page);
3562 return ((iter->head_page == commit_page && iter->head == commit) ||
3563 (iter->head_page == reader && commit_page == head_page &&
3564 head_page->read == commit &&
3565 iter->head == rb_page_commit(cpu_buffer->reader_page)));
3567 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3570 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3571 struct ring_buffer_event *event)
3575 switch (event->type_len) {
3576 case RINGBUF_TYPE_PADDING:
3579 case RINGBUF_TYPE_TIME_EXTEND:
3580 delta = ring_buffer_event_time_stamp(event);
3581 cpu_buffer->read_stamp += delta;
3584 case RINGBUF_TYPE_TIME_STAMP:
3585 delta = ring_buffer_event_time_stamp(event);
3586 cpu_buffer->read_stamp = delta;
3589 case RINGBUF_TYPE_DATA:
3590 cpu_buffer->read_stamp += event->time_delta;
3600 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3601 struct ring_buffer_event *event)
3605 switch (event->type_len) {
3606 case RINGBUF_TYPE_PADDING:
3609 case RINGBUF_TYPE_TIME_EXTEND:
3610 delta = ring_buffer_event_time_stamp(event);
3611 iter->read_stamp += delta;
3614 case RINGBUF_TYPE_TIME_STAMP:
3615 delta = ring_buffer_event_time_stamp(event);
3616 iter->read_stamp = delta;
3619 case RINGBUF_TYPE_DATA:
3620 iter->read_stamp += event->time_delta;
3629 static struct buffer_page *
3630 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3632 struct buffer_page *reader = NULL;
3633 unsigned long overwrite;
3634 unsigned long flags;
3638 local_irq_save(flags);
3639 arch_spin_lock(&cpu_buffer->lock);
3643 * This should normally only loop twice. But because the
3644 * start of the reader inserts an empty page, it causes
3645 * a case where we will loop three times. There should be no
3646 * reason to loop four times (that I know of).
3648 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3653 reader = cpu_buffer->reader_page;
3655 /* If there's more to read, return this page */
3656 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3659 /* Never should we have an index greater than the size */
3660 if (RB_WARN_ON(cpu_buffer,
3661 cpu_buffer->reader_page->read > rb_page_size(reader)))
3664 /* check if we caught up to the tail */
3666 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3669 /* Don't bother swapping if the ring buffer is empty */
3670 if (rb_num_of_entries(cpu_buffer) == 0)
3674 * Reset the reader page to size zero.
3676 local_set(&cpu_buffer->reader_page->write, 0);
3677 local_set(&cpu_buffer->reader_page->entries, 0);
3678 local_set(&cpu_buffer->reader_page->page->commit, 0);
3679 cpu_buffer->reader_page->real_end = 0;
3683 * Splice the empty reader page into the list around the head.
3685 reader = rb_set_head_page(cpu_buffer);
3688 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3689 cpu_buffer->reader_page->list.prev = reader->list.prev;
3692 * cpu_buffer->pages just needs to point to the buffer, it
3693 * has no specific buffer page to point to. Lets move it out
3694 * of our way so we don't accidentally swap it.
3696 cpu_buffer->pages = reader->list.prev;
3698 /* The reader page will be pointing to the new head */
3699 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3702 * We want to make sure we read the overruns after we set up our
3703 * pointers to the next object. The writer side does a
3704 * cmpxchg to cross pages which acts as the mb on the writer
3705 * side. Note, the reader will constantly fail the swap
3706 * while the writer is updating the pointers, so this
3707 * guarantees that the overwrite recorded here is the one we
3708 * want to compare with the last_overrun.
3711 overwrite = local_read(&(cpu_buffer->overrun));
3714 * Here's the tricky part.
3716 * We need to move the pointer past the header page.
3717 * But we can only do that if a writer is not currently
3718 * moving it. The page before the header page has the
3719 * flag bit '1' set if it is pointing to the page we want.
3720 * but if the writer is in the process of moving it
3721 * than it will be '2' or already moved '0'.
3724 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3727 * If we did not convert it, then we must try again.
3733 * Yeah! We succeeded in replacing the page.
3735 * Now make the new head point back to the reader page.
3737 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3738 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3740 /* Finally update the reader page to the new head */
3741 cpu_buffer->reader_page = reader;
3742 cpu_buffer->reader_page->read = 0;
3744 if (overwrite != cpu_buffer->last_overrun) {
3745 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3746 cpu_buffer->last_overrun = overwrite;
3752 /* Update the read_stamp on the first event */
3753 if (reader && reader->read == 0)
3754 cpu_buffer->read_stamp = reader->page->time_stamp;
3756 arch_spin_unlock(&cpu_buffer->lock);
3757 local_irq_restore(flags);
3762 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3764 struct ring_buffer_event *event;
3765 struct buffer_page *reader;
3768 reader = rb_get_reader_page(cpu_buffer);
3770 /* This function should not be called when buffer is empty */
3771 if (RB_WARN_ON(cpu_buffer, !reader))
3774 event = rb_reader_event(cpu_buffer);
3776 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3779 rb_update_read_stamp(cpu_buffer, event);
3781 length = rb_event_length(event);
3782 cpu_buffer->reader_page->read += length;
3785 static void rb_advance_iter(struct ring_buffer_iter *iter)
3787 struct ring_buffer_per_cpu *cpu_buffer;
3788 struct ring_buffer_event *event;
3791 cpu_buffer = iter->cpu_buffer;
3794 * Check if we are at the end of the buffer.
3796 if (iter->head >= rb_page_size(iter->head_page)) {
3797 /* discarded commits can make the page empty */
3798 if (iter->head_page == cpu_buffer->commit_page)
3804 event = rb_iter_head_event(iter);
3806 length = rb_event_length(event);
3809 * This should not be called to advance the header if we are
3810 * at the tail of the buffer.
3812 if (RB_WARN_ON(cpu_buffer,
3813 (iter->head_page == cpu_buffer->commit_page) &&
3814 (iter->head + length > rb_commit_index(cpu_buffer))))
3817 rb_update_iter_read_stamp(iter, event);
3819 iter->head += length;
3821 /* check for end of page padding */
3822 if ((iter->head >= rb_page_size(iter->head_page)) &&
3823 (iter->head_page != cpu_buffer->commit_page))
3827 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3829 return cpu_buffer->lost_events;
3832 static struct ring_buffer_event *
3833 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3834 unsigned long *lost_events)
3836 struct ring_buffer_event *event;
3837 struct buffer_page *reader;
3844 * We repeat when a time extend is encountered.
3845 * Since the time extend is always attached to a data event,
3846 * we should never loop more than once.
3847 * (We never hit the following condition more than twice).
3849 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3852 reader = rb_get_reader_page(cpu_buffer);
3856 event = rb_reader_event(cpu_buffer);
3858 switch (event->type_len) {
3859 case RINGBUF_TYPE_PADDING:
3860 if (rb_null_event(event))
3861 RB_WARN_ON(cpu_buffer, 1);
3863 * Because the writer could be discarding every
3864 * event it creates (which would probably be bad)
3865 * if we were to go back to "again" then we may never
3866 * catch up, and will trigger the warn on, or lock
3867 * the box. Return the padding, and we will release
3868 * the current locks, and try again.
3872 case RINGBUF_TYPE_TIME_EXTEND:
3873 /* Internal data, OK to advance */
3874 rb_advance_reader(cpu_buffer);
3877 case RINGBUF_TYPE_TIME_STAMP:
3879 *ts = ring_buffer_event_time_stamp(event);
3880 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3881 cpu_buffer->cpu, ts);
3883 /* Internal data, OK to advance */
3884 rb_advance_reader(cpu_buffer);
3887 case RINGBUF_TYPE_DATA:
3889 *ts = cpu_buffer->read_stamp + event->time_delta;
3890 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3891 cpu_buffer->cpu, ts);
3894 *lost_events = rb_lost_events(cpu_buffer);
3903 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3905 static struct ring_buffer_event *
3906 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3908 struct ring_buffer *buffer;
3909 struct ring_buffer_per_cpu *cpu_buffer;
3910 struct ring_buffer_event *event;
3916 cpu_buffer = iter->cpu_buffer;
3917 buffer = cpu_buffer->buffer;
3920 * Check if someone performed a consuming read to
3921 * the buffer. A consuming read invalidates the iterator
3922 * and we need to reset the iterator in this case.
3924 if (unlikely(iter->cache_read != cpu_buffer->read ||
3925 iter->cache_reader_page != cpu_buffer->reader_page))
3926 rb_iter_reset(iter);
3929 if (ring_buffer_iter_empty(iter))
3933 * We repeat when a time extend is encountered or we hit
3934 * the end of the page. Since the time extend is always attached
3935 * to a data event, we should never loop more than three times.
3936 * Once for going to next page, once on time extend, and
3937 * finally once to get the event.
3938 * (We never hit the following condition more than thrice).
3940 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3943 if (rb_per_cpu_empty(cpu_buffer))
3946 if (iter->head >= rb_page_size(iter->head_page)) {
3951 event = rb_iter_head_event(iter);
3953 switch (event->type_len) {
3954 case RINGBUF_TYPE_PADDING:
3955 if (rb_null_event(event)) {
3959 rb_advance_iter(iter);
3962 case RINGBUF_TYPE_TIME_EXTEND:
3963 /* Internal data, OK to advance */
3964 rb_advance_iter(iter);
3967 case RINGBUF_TYPE_TIME_STAMP:
3969 *ts = ring_buffer_event_time_stamp(event);
3970 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3971 cpu_buffer->cpu, ts);
3973 /* Internal data, OK to advance */
3974 rb_advance_iter(iter);
3977 case RINGBUF_TYPE_DATA:
3979 *ts = iter->read_stamp + event->time_delta;
3980 ring_buffer_normalize_time_stamp(buffer,
3981 cpu_buffer->cpu, ts);
3991 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3993 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
3995 if (likely(!in_nmi())) {
3996 raw_spin_lock(&cpu_buffer->reader_lock);
4001 * If an NMI die dumps out the content of the ring buffer
4002 * trylock must be used to prevent a deadlock if the NMI
4003 * preempted a task that holds the ring buffer locks. If
4004 * we get the lock then all is fine, if not, then continue
4005 * to do the read, but this can corrupt the ring buffer,
4006 * so it must be permanently disabled from future writes.
4007 * Reading from NMI is a oneshot deal.
4009 if (raw_spin_trylock(&cpu_buffer->reader_lock))
4012 /* Continue without locking, but disable the ring buffer */
4013 atomic_inc(&cpu_buffer->record_disabled);
4018 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4021 raw_spin_unlock(&cpu_buffer->reader_lock);
4026 * ring_buffer_peek - peek at the next event to be read
4027 * @buffer: The ring buffer to read
4028 * @cpu: The cpu to peak at
4029 * @ts: The timestamp counter of this event.
4030 * @lost_events: a variable to store if events were lost (may be NULL)
4032 * This will return the event that will be read next, but does
4033 * not consume the data.
4035 struct ring_buffer_event *
4036 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
4037 unsigned long *lost_events)
4039 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4040 struct ring_buffer_event *event;
4041 unsigned long flags;
4044 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4048 local_irq_save(flags);
4049 dolock = rb_reader_lock(cpu_buffer);
4050 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4051 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4052 rb_advance_reader(cpu_buffer);
4053 rb_reader_unlock(cpu_buffer, dolock);
4054 local_irq_restore(flags);
4056 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4063 * ring_buffer_iter_peek - peek at the next event to be read
4064 * @iter: The ring buffer iterator
4065 * @ts: The timestamp counter of this event.
4067 * This will return the event that will be read next, but does
4068 * not increment the iterator.
4070 struct ring_buffer_event *
4071 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4073 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4074 struct ring_buffer_event *event;
4075 unsigned long flags;
4078 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4079 event = rb_iter_peek(iter, ts);
4080 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4082 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4089 * ring_buffer_consume - return an event and consume it
4090 * @buffer: The ring buffer to get the next event from
4091 * @cpu: the cpu to read the buffer from
4092 * @ts: a variable to store the timestamp (may be NULL)
4093 * @lost_events: a variable to store if events were lost (may be NULL)
4095 * Returns the next event in the ring buffer, and that event is consumed.
4096 * Meaning, that sequential reads will keep returning a different event,
4097 * and eventually empty the ring buffer if the producer is slower.
4099 struct ring_buffer_event *
4100 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
4101 unsigned long *lost_events)
4103 struct ring_buffer_per_cpu *cpu_buffer;
4104 struct ring_buffer_event *event = NULL;
4105 unsigned long flags;
4109 /* might be called in atomic */
4112 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4115 cpu_buffer = buffer->buffers[cpu];
4116 local_irq_save(flags);
4117 dolock = rb_reader_lock(cpu_buffer);
4119 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4121 cpu_buffer->lost_events = 0;
4122 rb_advance_reader(cpu_buffer);
4125 rb_reader_unlock(cpu_buffer, dolock);
4126 local_irq_restore(flags);
4131 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4136 EXPORT_SYMBOL_GPL(ring_buffer_consume);
4139 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4140 * @buffer: The ring buffer to read from
4141 * @cpu: The cpu buffer to iterate over
4143 * This performs the initial preparations necessary to iterate
4144 * through the buffer. Memory is allocated, buffer recording
4145 * is disabled, and the iterator pointer is returned to the caller.
4147 * Disabling buffer recording prevents the reading from being
4148 * corrupted. This is not a consuming read, so a producer is not
4151 * After a sequence of ring_buffer_read_prepare calls, the user is
4152 * expected to make at least one call to ring_buffer_read_prepare_sync.
4153 * Afterwards, ring_buffer_read_start is invoked to get things going
4156 * This overall must be paired with ring_buffer_read_finish.
4158 struct ring_buffer_iter *
4159 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
4161 struct ring_buffer_per_cpu *cpu_buffer;
4162 struct ring_buffer_iter *iter;
4164 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4167 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
4171 cpu_buffer = buffer->buffers[cpu];
4173 iter->cpu_buffer = cpu_buffer;
4175 atomic_inc(&buffer->resize_disabled);
4176 atomic_inc(&cpu_buffer->record_disabled);
4180 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4183 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4185 * All previously invoked ring_buffer_read_prepare calls to prepare
4186 * iterators will be synchronized. Afterwards, read_buffer_read_start
4187 * calls on those iterators are allowed.
4190 ring_buffer_read_prepare_sync(void)
4192 synchronize_sched();
4194 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4197 * ring_buffer_read_start - start a non consuming read of the buffer
4198 * @iter: The iterator returned by ring_buffer_read_prepare
4200 * This finalizes the startup of an iteration through the buffer.
4201 * The iterator comes from a call to ring_buffer_read_prepare and
4202 * an intervening ring_buffer_read_prepare_sync must have been
4205 * Must be paired with ring_buffer_read_finish.
4208 ring_buffer_read_start(struct ring_buffer_iter *iter)
4210 struct ring_buffer_per_cpu *cpu_buffer;
4211 unsigned long flags;
4216 cpu_buffer = iter->cpu_buffer;
4218 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4219 arch_spin_lock(&cpu_buffer->lock);
4220 rb_iter_reset(iter);
4221 arch_spin_unlock(&cpu_buffer->lock);
4222 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4224 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4227 * ring_buffer_read_finish - finish reading the iterator of the buffer
4228 * @iter: The iterator retrieved by ring_buffer_start
4230 * This re-enables the recording to the buffer, and frees the
4234 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4236 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4237 unsigned long flags;
4240 * Ring buffer is disabled from recording, here's a good place
4241 * to check the integrity of the ring buffer.
4242 * Must prevent readers from trying to read, as the check
4243 * clears the HEAD page and readers require it.
4245 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4246 rb_check_pages(cpu_buffer);
4247 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4249 atomic_dec(&cpu_buffer->record_disabled);
4250 atomic_dec(&cpu_buffer->buffer->resize_disabled);
4253 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4256 * ring_buffer_read - read the next item in the ring buffer by the iterator
4257 * @iter: The ring buffer iterator
4258 * @ts: The time stamp of the event read.
4260 * This reads the next event in the ring buffer and increments the iterator.
4262 struct ring_buffer_event *
4263 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4265 struct ring_buffer_event *event;
4266 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4267 unsigned long flags;
4269 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4271 event = rb_iter_peek(iter, ts);
4275 if (event->type_len == RINGBUF_TYPE_PADDING)
4278 rb_advance_iter(iter);
4280 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4284 EXPORT_SYMBOL_GPL(ring_buffer_read);
4287 * ring_buffer_size - return the size of the ring buffer (in bytes)
4288 * @buffer: The ring buffer.
4290 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4293 * Earlier, this method returned
4294 * BUF_PAGE_SIZE * buffer->nr_pages
4295 * Since the nr_pages field is now removed, we have converted this to
4296 * return the per cpu buffer value.
4298 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4301 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4303 EXPORT_SYMBOL_GPL(ring_buffer_size);
4306 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4308 rb_head_page_deactivate(cpu_buffer);
4310 cpu_buffer->head_page
4311 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4312 local_set(&cpu_buffer->head_page->write, 0);
4313 local_set(&cpu_buffer->head_page->entries, 0);
4314 local_set(&cpu_buffer->head_page->page->commit, 0);
4316 cpu_buffer->head_page->read = 0;
4318 cpu_buffer->tail_page = cpu_buffer->head_page;
4319 cpu_buffer->commit_page = cpu_buffer->head_page;
4321 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4322 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4323 local_set(&cpu_buffer->reader_page->write, 0);
4324 local_set(&cpu_buffer->reader_page->entries, 0);
4325 local_set(&cpu_buffer->reader_page->page->commit, 0);
4326 cpu_buffer->reader_page->read = 0;
4328 local_set(&cpu_buffer->entries_bytes, 0);
4329 local_set(&cpu_buffer->overrun, 0);
4330 local_set(&cpu_buffer->commit_overrun, 0);
4331 local_set(&cpu_buffer->dropped_events, 0);
4332 local_set(&cpu_buffer->entries, 0);
4333 local_set(&cpu_buffer->committing, 0);
4334 local_set(&cpu_buffer->commits, 0);
4335 cpu_buffer->read = 0;
4336 cpu_buffer->read_bytes = 0;
4338 cpu_buffer->write_stamp = 0;
4339 cpu_buffer->read_stamp = 0;
4341 cpu_buffer->lost_events = 0;
4342 cpu_buffer->last_overrun = 0;
4344 rb_head_page_activate(cpu_buffer);
4348 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4349 * @buffer: The ring buffer to reset a per cpu buffer of
4350 * @cpu: The CPU buffer to be reset
4352 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4354 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4355 unsigned long flags;
4357 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4360 atomic_inc(&buffer->resize_disabled);
4361 atomic_inc(&cpu_buffer->record_disabled);
4363 /* Make sure all commits have finished */
4364 synchronize_sched();
4366 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4368 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4371 arch_spin_lock(&cpu_buffer->lock);
4373 rb_reset_cpu(cpu_buffer);
4375 arch_spin_unlock(&cpu_buffer->lock);
4378 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4380 atomic_dec(&cpu_buffer->record_disabled);
4381 atomic_dec(&buffer->resize_disabled);
4383 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4386 * ring_buffer_reset - reset a ring buffer
4387 * @buffer: The ring buffer to reset all cpu buffers
4389 void ring_buffer_reset(struct ring_buffer *buffer)
4393 for_each_buffer_cpu(buffer, cpu)
4394 ring_buffer_reset_cpu(buffer, cpu);
4396 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4399 * rind_buffer_empty - is the ring buffer empty?
4400 * @buffer: The ring buffer to test
4402 bool ring_buffer_empty(struct ring_buffer *buffer)
4404 struct ring_buffer_per_cpu *cpu_buffer;
4405 unsigned long flags;
4410 /* yes this is racy, but if you don't like the race, lock the buffer */
4411 for_each_buffer_cpu(buffer, cpu) {
4412 cpu_buffer = buffer->buffers[cpu];
4413 local_irq_save(flags);
4414 dolock = rb_reader_lock(cpu_buffer);
4415 ret = rb_per_cpu_empty(cpu_buffer);
4416 rb_reader_unlock(cpu_buffer, dolock);
4417 local_irq_restore(flags);
4425 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4428 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4429 * @buffer: The ring buffer
4430 * @cpu: The CPU buffer to test
4432 bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4434 struct ring_buffer_per_cpu *cpu_buffer;
4435 unsigned long flags;
4439 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4442 cpu_buffer = buffer->buffers[cpu];
4443 local_irq_save(flags);
4444 dolock = rb_reader_lock(cpu_buffer);
4445 ret = rb_per_cpu_empty(cpu_buffer);
4446 rb_reader_unlock(cpu_buffer, dolock);
4447 local_irq_restore(flags);
4451 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4453 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4455 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4456 * @buffer_a: One buffer to swap with
4457 * @buffer_b: The other buffer to swap with
4459 * This function is useful for tracers that want to take a "snapshot"
4460 * of a CPU buffer and has another back up buffer lying around.
4461 * it is expected that the tracer handles the cpu buffer not being
4462 * used at the moment.
4464 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4465 struct ring_buffer *buffer_b, int cpu)
4467 struct ring_buffer_per_cpu *cpu_buffer_a;
4468 struct ring_buffer_per_cpu *cpu_buffer_b;
4471 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4472 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4475 cpu_buffer_a = buffer_a->buffers[cpu];
4476 cpu_buffer_b = buffer_b->buffers[cpu];
4478 /* At least make sure the two buffers are somewhat the same */
4479 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4484 if (atomic_read(&buffer_a->record_disabled))
4487 if (atomic_read(&buffer_b->record_disabled))
4490 if (atomic_read(&cpu_buffer_a->record_disabled))
4493 if (atomic_read(&cpu_buffer_b->record_disabled))
4497 * We can't do a synchronize_sched here because this
4498 * function can be called in atomic context.
4499 * Normally this will be called from the same CPU as cpu.
4500 * If not it's up to the caller to protect this.
4502 atomic_inc(&cpu_buffer_a->record_disabled);
4503 atomic_inc(&cpu_buffer_b->record_disabled);
4506 if (local_read(&cpu_buffer_a->committing))
4508 if (local_read(&cpu_buffer_b->committing))
4511 buffer_a->buffers[cpu] = cpu_buffer_b;
4512 buffer_b->buffers[cpu] = cpu_buffer_a;
4514 cpu_buffer_b->buffer = buffer_a;
4515 cpu_buffer_a->buffer = buffer_b;
4520 atomic_dec(&cpu_buffer_a->record_disabled);
4521 atomic_dec(&cpu_buffer_b->record_disabled);
4525 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4526 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4529 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4530 * @buffer: the buffer to allocate for.
4531 * @cpu: the cpu buffer to allocate.
4533 * This function is used in conjunction with ring_buffer_read_page.
4534 * When reading a full page from the ring buffer, these functions
4535 * can be used to speed up the process. The calling function should
4536 * allocate a few pages first with this function. Then when it
4537 * needs to get pages from the ring buffer, it passes the result
4538 * of this function into ring_buffer_read_page, which will swap
4539 * the page that was allocated, with the read page of the buffer.
4542 * The page allocated, or ERR_PTR
4544 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4546 struct ring_buffer_per_cpu *cpu_buffer;
4547 struct buffer_data_page *bpage = NULL;
4548 unsigned long flags;
4551 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4552 return ERR_PTR(-ENODEV);
4554 cpu_buffer = buffer->buffers[cpu];
4555 local_irq_save(flags);
4556 arch_spin_lock(&cpu_buffer->lock);
4558 if (cpu_buffer->free_page) {
4559 bpage = cpu_buffer->free_page;
4560 cpu_buffer->free_page = NULL;
4563 arch_spin_unlock(&cpu_buffer->lock);
4564 local_irq_restore(flags);
4569 page = alloc_pages_node(cpu_to_node(cpu),
4570 GFP_KERNEL | __GFP_NORETRY, 0);
4572 return ERR_PTR(-ENOMEM);
4574 bpage = page_address(page);
4577 rb_init_page(bpage);
4581 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4584 * ring_buffer_free_read_page - free an allocated read page
4585 * @buffer: the buffer the page was allocate for
4586 * @cpu: the cpu buffer the page came from
4587 * @data: the page to free
4589 * Free a page allocated from ring_buffer_alloc_read_page.
4591 void ring_buffer_free_read_page(struct ring_buffer *buffer, int cpu, void *data)
4593 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4594 struct buffer_data_page *bpage = data;
4595 struct page *page = virt_to_page(bpage);
4596 unsigned long flags;
4598 /* If the page is still in use someplace else, we can't reuse it */
4599 if (page_ref_count(page) > 1)
4602 local_irq_save(flags);
4603 arch_spin_lock(&cpu_buffer->lock);
4605 if (!cpu_buffer->free_page) {
4606 cpu_buffer->free_page = bpage;
4610 arch_spin_unlock(&cpu_buffer->lock);
4611 local_irq_restore(flags);
4614 free_page((unsigned long)bpage);
4616 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4619 * ring_buffer_read_page - extract a page from the ring buffer
4620 * @buffer: buffer to extract from
4621 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4622 * @len: amount to extract
4623 * @cpu: the cpu of the buffer to extract
4624 * @full: should the extraction only happen when the page is full.
4626 * This function will pull out a page from the ring buffer and consume it.
4627 * @data_page must be the address of the variable that was returned
4628 * from ring_buffer_alloc_read_page. This is because the page might be used
4629 * to swap with a page in the ring buffer.
4632 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4633 * if (IS_ERR(rpage))
4634 * return PTR_ERR(rpage);
4635 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4637 * process_page(rpage, ret);
4639 * When @full is set, the function will not return true unless
4640 * the writer is off the reader page.
4642 * Note: it is up to the calling functions to handle sleeps and wakeups.
4643 * The ring buffer can be used anywhere in the kernel and can not
4644 * blindly call wake_up. The layer that uses the ring buffer must be
4645 * responsible for that.
4648 * >=0 if data has been transferred, returns the offset of consumed data.
4649 * <0 if no data has been transferred.
4651 int ring_buffer_read_page(struct ring_buffer *buffer,
4652 void **data_page, size_t len, int cpu, int full)
4654 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4655 struct ring_buffer_event *event;
4656 struct buffer_data_page *bpage;
4657 struct buffer_page *reader;
4658 unsigned long missed_events;
4659 unsigned long flags;
4660 unsigned int commit;
4665 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4669 * If len is not big enough to hold the page header, then
4670 * we can not copy anything.
4672 if (len <= BUF_PAGE_HDR_SIZE)
4675 len -= BUF_PAGE_HDR_SIZE;
4684 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4686 reader = rb_get_reader_page(cpu_buffer);
4690 event = rb_reader_event(cpu_buffer);
4692 read = reader->read;
4693 commit = rb_page_commit(reader);
4695 /* Check if any events were dropped */
4696 missed_events = cpu_buffer->lost_events;
4699 * If this page has been partially read or
4700 * if len is not big enough to read the rest of the page or
4701 * a writer is still on the page, then
4702 * we must copy the data from the page to the buffer.
4703 * Otherwise, we can simply swap the page with the one passed in.
4705 if (read || (len < (commit - read)) ||
4706 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4707 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4708 unsigned int rpos = read;
4709 unsigned int pos = 0;
4715 if (len > (commit - read))
4716 len = (commit - read);
4718 /* Always keep the time extend and data together */
4719 size = rb_event_ts_length(event);
4724 /* save the current timestamp, since the user will need it */
4725 save_timestamp = cpu_buffer->read_stamp;
4727 /* Need to copy one event at a time */
4729 /* We need the size of one event, because
4730 * rb_advance_reader only advances by one event,
4731 * whereas rb_event_ts_length may include the size of
4732 * one or two events.
4733 * We have already ensured there's enough space if this
4734 * is a time extend. */
4735 size = rb_event_length(event);
4736 memcpy(bpage->data + pos, rpage->data + rpos, size);
4740 rb_advance_reader(cpu_buffer);
4741 rpos = reader->read;
4747 event = rb_reader_event(cpu_buffer);
4748 /* Always keep the time extend and data together */
4749 size = rb_event_ts_length(event);
4750 } while (len >= size);
4753 local_set(&bpage->commit, pos);
4754 bpage->time_stamp = save_timestamp;
4756 /* we copied everything to the beginning */
4759 /* update the entry counter */
4760 cpu_buffer->read += rb_page_entries(reader);
4761 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4763 /* swap the pages */
4764 rb_init_page(bpage);
4765 bpage = reader->page;
4766 reader->page = *data_page;
4767 local_set(&reader->write, 0);
4768 local_set(&reader->entries, 0);
4773 * Use the real_end for the data size,
4774 * This gives us a chance to store the lost events
4777 if (reader->real_end)
4778 local_set(&bpage->commit, reader->real_end);
4782 cpu_buffer->lost_events = 0;
4784 commit = local_read(&bpage->commit);
4786 * Set a flag in the commit field if we lost events
4788 if (missed_events) {
4789 /* If there is room at the end of the page to save the
4790 * missed events, then record it there.
4792 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4793 memcpy(&bpage->data[commit], &missed_events,
4794 sizeof(missed_events));
4795 local_add(RB_MISSED_STORED, &bpage->commit);
4796 commit += sizeof(missed_events);
4798 local_add(RB_MISSED_EVENTS, &bpage->commit);
4802 * This page may be off to user land. Zero it out here.
4804 if (commit < BUF_PAGE_SIZE)
4805 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4808 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4813 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4816 * We only allocate new buffers, never free them if the CPU goes down.
4817 * If we were to free the buffer, then the user would lose any trace that was in
4820 int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
4822 struct ring_buffer *buffer;
4825 unsigned long nr_pages;
4827 buffer = container_of(node, struct ring_buffer, node);
4828 if (cpumask_test_cpu(cpu, buffer->cpumask))
4833 /* check if all cpu sizes are same */
4834 for_each_buffer_cpu(buffer, cpu_i) {
4835 /* fill in the size from first enabled cpu */
4837 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4838 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4843 /* allocate minimum pages, user can later expand it */
4846 buffer->buffers[cpu] =
4847 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4848 if (!buffer->buffers[cpu]) {
4849 WARN(1, "failed to allocate ring buffer on CPU %u\n",
4854 cpumask_set_cpu(cpu, buffer->cpumask);
4858 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4860 * This is a basic integrity check of the ring buffer.
4861 * Late in the boot cycle this test will run when configured in.
4862 * It will kick off a thread per CPU that will go into a loop
4863 * writing to the per cpu ring buffer various sizes of data.
4864 * Some of the data will be large items, some small.
4866 * Another thread is created that goes into a spin, sending out
4867 * IPIs to the other CPUs to also write into the ring buffer.
4868 * this is to test the nesting ability of the buffer.
4870 * Basic stats are recorded and reported. If something in the
4871 * ring buffer should happen that's not expected, a big warning
4872 * is displayed and all ring buffers are disabled.
4874 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4876 struct rb_test_data {
4877 struct ring_buffer *buffer;
4878 unsigned long events;
4879 unsigned long bytes_written;
4880 unsigned long bytes_alloc;
4881 unsigned long bytes_dropped;
4882 unsigned long events_nested;
4883 unsigned long bytes_written_nested;
4884 unsigned long bytes_alloc_nested;
4885 unsigned long bytes_dropped_nested;
4886 int min_size_nested;
4887 int max_size_nested;
4894 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4897 #define RB_TEST_BUFFER_SIZE 1048576
4899 static char rb_string[] __initdata =
4900 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4901 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4902 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4904 static bool rb_test_started __initdata;
4911 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4913 struct ring_buffer_event *event;
4914 struct rb_item *item;
4921 /* Have nested writes different that what is written */
4922 cnt = data->cnt + (nested ? 27 : 0);
4924 /* Multiply cnt by ~e, to make some unique increment */
4925 size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4927 len = size + sizeof(struct rb_item);
4929 started = rb_test_started;
4930 /* read rb_test_started before checking buffer enabled */
4933 event = ring_buffer_lock_reserve(data->buffer, len);
4935 /* Ignore dropped events before test starts. */
4938 data->bytes_dropped += len;
4940 data->bytes_dropped_nested += len;
4945 event_len = ring_buffer_event_length(event);
4947 if (RB_WARN_ON(data->buffer, event_len < len))
4950 item = ring_buffer_event_data(event);
4952 memcpy(item->str, rb_string, size);
4955 data->bytes_alloc_nested += event_len;
4956 data->bytes_written_nested += len;
4957 data->events_nested++;
4958 if (!data->min_size_nested || len < data->min_size_nested)
4959 data->min_size_nested = len;
4960 if (len > data->max_size_nested)
4961 data->max_size_nested = len;
4963 data->bytes_alloc += event_len;
4964 data->bytes_written += len;
4966 if (!data->min_size || len < data->min_size)
4967 data->max_size = len;
4968 if (len > data->max_size)
4969 data->max_size = len;
4973 ring_buffer_unlock_commit(data->buffer, event);
4978 static __init int rb_test(void *arg)
4980 struct rb_test_data *data = arg;
4982 while (!kthread_should_stop()) {
4983 rb_write_something(data, false);
4986 set_current_state(TASK_INTERRUPTIBLE);
4987 /* Now sleep between a min of 100-300us and a max of 1ms */
4988 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4994 static __init void rb_ipi(void *ignore)
4996 struct rb_test_data *data;
4997 int cpu = smp_processor_id();
4999 data = &rb_data[cpu];
5000 rb_write_something(data, true);
5003 static __init int rb_hammer_test(void *arg)
5005 while (!kthread_should_stop()) {
5007 /* Send an IPI to all cpus to write data! */
5008 smp_call_function(rb_ipi, NULL, 1);
5009 /* No sleep, but for non preempt, let others run */
5016 static __init int test_ringbuffer(void)
5018 struct task_struct *rb_hammer;
5019 struct ring_buffer *buffer;
5023 pr_info("Running ring buffer tests...\n");
5025 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
5026 if (WARN_ON(!buffer))
5029 /* Disable buffer so that threads can't write to it yet */
5030 ring_buffer_record_off(buffer);
5032 for_each_online_cpu(cpu) {
5033 rb_data[cpu].buffer = buffer;
5034 rb_data[cpu].cpu = cpu;
5035 rb_data[cpu].cnt = cpu;
5036 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
5037 "rbtester/%d", cpu);
5038 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
5039 pr_cont("FAILED\n");
5040 ret = PTR_ERR(rb_threads[cpu]);
5044 kthread_bind(rb_threads[cpu], cpu);
5045 wake_up_process(rb_threads[cpu]);
5048 /* Now create the rb hammer! */
5049 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
5050 if (WARN_ON(IS_ERR(rb_hammer))) {
5051 pr_cont("FAILED\n");
5052 ret = PTR_ERR(rb_hammer);
5056 ring_buffer_record_on(buffer);
5058 * Show buffer is enabled before setting rb_test_started.
5059 * Yes there's a small race window where events could be
5060 * dropped and the thread wont catch it. But when a ring
5061 * buffer gets enabled, there will always be some kind of
5062 * delay before other CPUs see it. Thus, we don't care about
5063 * those dropped events. We care about events dropped after
5064 * the threads see that the buffer is active.
5067 rb_test_started = true;
5069 set_current_state(TASK_INTERRUPTIBLE);
5070 /* Just run for 10 seconds */;
5071 schedule_timeout(10 * HZ);
5073 kthread_stop(rb_hammer);
5076 for_each_online_cpu(cpu) {
5077 if (!rb_threads[cpu])
5079 kthread_stop(rb_threads[cpu]);
5082 ring_buffer_free(buffer);
5087 pr_info("finished\n");
5088 for_each_online_cpu(cpu) {
5089 struct ring_buffer_event *event;
5090 struct rb_test_data *data = &rb_data[cpu];
5091 struct rb_item *item;
5092 unsigned long total_events;
5093 unsigned long total_dropped;
5094 unsigned long total_written;
5095 unsigned long total_alloc;
5096 unsigned long total_read = 0;
5097 unsigned long total_size = 0;
5098 unsigned long total_len = 0;
5099 unsigned long total_lost = 0;
5102 int small_event_size;
5106 total_events = data->events + data->events_nested;
5107 total_written = data->bytes_written + data->bytes_written_nested;
5108 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
5109 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
5111 big_event_size = data->max_size + data->max_size_nested;
5112 small_event_size = data->min_size + data->min_size_nested;
5114 pr_info("CPU %d:\n", cpu);
5115 pr_info(" events: %ld\n", total_events);
5116 pr_info(" dropped bytes: %ld\n", total_dropped);
5117 pr_info(" alloced bytes: %ld\n", total_alloc);
5118 pr_info(" written bytes: %ld\n", total_written);
5119 pr_info(" biggest event: %d\n", big_event_size);
5120 pr_info(" smallest event: %d\n", small_event_size);
5122 if (RB_WARN_ON(buffer, total_dropped))
5127 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
5129 item = ring_buffer_event_data(event);
5130 total_len += ring_buffer_event_length(event);
5131 total_size += item->size + sizeof(struct rb_item);
5132 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
5133 pr_info("FAILED!\n");
5134 pr_info("buffer had: %.*s\n", item->size, item->str);
5135 pr_info("expected: %.*s\n", item->size, rb_string);
5136 RB_WARN_ON(buffer, 1);
5147 pr_info(" read events: %ld\n", total_read);
5148 pr_info(" lost events: %ld\n", total_lost);
5149 pr_info(" total events: %ld\n", total_lost + total_read);
5150 pr_info(" recorded len bytes: %ld\n", total_len);
5151 pr_info(" recorded size bytes: %ld\n", total_size);
5153 pr_info(" With dropped events, record len and size may not match\n"
5154 " alloced and written from above\n");
5156 if (RB_WARN_ON(buffer, total_len != total_alloc ||
5157 total_size != total_written))
5160 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5166 pr_info("Ring buffer PASSED!\n");
5168 ring_buffer_free(buffer);
5172 late_initcall(test_ringbuffer);
5173 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */