4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 #include <linux/trace_events.h>
7 #include <linux/ring_buffer.h>
8 #include <linux/trace_clock.h>
9 #include <linux/trace_seq.h>
10 #include <linux/spinlock.h>
11 #include <linux/irq_work.h>
12 #include <linux/uaccess.h>
13 #include <linux/hardirq.h>
14 #include <linux/kthread.h> /* for self test */
15 #include <linux/kmemcheck.h>
16 #include <linux/module.h>
17 #include <linux/percpu.h>
18 #include <linux/mutex.h>
19 #include <linux/delay.h>
20 #include <linux/slab.h>
21 #include <linux/init.h>
22 #include <linux/hash.h>
23 #include <linux/list.h>
24 #include <linux/cpu.h>
26 #include <asm/local.h>
28 static void update_pages_handler(struct work_struct
*work
);
31 * The ring buffer header is special. We must manually up keep it.
33 int ring_buffer_print_entry_header(struct trace_seq
*s
)
35 trace_seq_puts(s
, "# compressed entry header\n");
36 trace_seq_puts(s
, "\ttype_len : 5 bits\n");
37 trace_seq_puts(s
, "\ttime_delta : 27 bits\n");
38 trace_seq_puts(s
, "\tarray : 32 bits\n");
39 trace_seq_putc(s
, '\n');
40 trace_seq_printf(s
, "\tpadding : type == %d\n",
41 RINGBUF_TYPE_PADDING
);
42 trace_seq_printf(s
, "\ttime_extend : type == %d\n",
43 RINGBUF_TYPE_TIME_EXTEND
);
44 trace_seq_printf(s
, "\tdata max type_len == %d\n",
45 RINGBUF_TYPE_DATA_TYPE_LEN_MAX
);
47 return !trace_seq_has_overflowed(s
);
51 * The ring buffer is made up of a list of pages. A separate list of pages is
52 * allocated for each CPU. A writer may only write to a buffer that is
53 * associated with the CPU it is currently executing on. A reader may read
54 * from any per cpu buffer.
56 * The reader is special. For each per cpu buffer, the reader has its own
57 * reader page. When a reader has read the entire reader page, this reader
58 * page is swapped with another page in the ring buffer.
60 * Now, as long as the writer is off the reader page, the reader can do what
61 * ever it wants with that page. The writer will never write to that page
62 * again (as long as it is out of the ring buffer).
64 * Here's some silly ASCII art.
67 * |reader| RING BUFFER
69 * +------+ +---+ +---+ +---+
78 * |reader| RING BUFFER
79 * |page |------------------v
80 * +------+ +---+ +---+ +---+
89 * |reader| RING BUFFER
90 * |page |------------------v
91 * +------+ +---+ +---+ +---+
96 * +------------------------------+
100 * |buffer| RING BUFFER
101 * |page |------------------v
102 * +------+ +---+ +---+ +---+
104 * | New +---+ +---+ +---+
107 * +------------------------------+
110 * After we make this swap, the reader can hand this page off to the splice
111 * code and be done with it. It can even allocate a new page if it needs to
112 * and swap that into the ring buffer.
114 * We will be using cmpxchg soon to make all this lockless.
118 /* Used for individual buffers (after the counter) */
119 #define RB_BUFFER_OFF (1 << 20)
121 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
123 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
124 #define RB_ALIGNMENT 4U
125 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
126 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
128 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
129 # define RB_FORCE_8BYTE_ALIGNMENT 0
130 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
132 # define RB_FORCE_8BYTE_ALIGNMENT 1
133 # define RB_ARCH_ALIGNMENT 8U
136 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
138 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
139 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
142 RB_LEN_TIME_EXTEND
= 8,
143 RB_LEN_TIME_STAMP
= 16,
146 #define skip_time_extend(event) \
147 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
149 static inline int rb_null_event(struct ring_buffer_event
*event
)
151 return event
->type_len
== RINGBUF_TYPE_PADDING
&& !event
->time_delta
;
154 static void rb_event_set_padding(struct ring_buffer_event
*event
)
156 /* padding has a NULL time_delta */
157 event
->type_len
= RINGBUF_TYPE_PADDING
;
158 event
->time_delta
= 0;
162 rb_event_data_length(struct ring_buffer_event
*event
)
167 length
= event
->type_len
* RB_ALIGNMENT
;
169 length
= event
->array
[0];
170 return length
+ RB_EVNT_HDR_SIZE
;
174 * Return the length of the given event. Will return
175 * the length of the time extend if the event is a
178 static inline unsigned
179 rb_event_length(struct ring_buffer_event
*event
)
181 switch (event
->type_len
) {
182 case RINGBUF_TYPE_PADDING
:
183 if (rb_null_event(event
))
186 return event
->array
[0] + RB_EVNT_HDR_SIZE
;
188 case RINGBUF_TYPE_TIME_EXTEND
:
189 return RB_LEN_TIME_EXTEND
;
191 case RINGBUF_TYPE_TIME_STAMP
:
192 return RB_LEN_TIME_STAMP
;
194 case RINGBUF_TYPE_DATA
:
195 return rb_event_data_length(event
);
204 * Return total length of time extend and data,
205 * or just the event length for all other events.
207 static inline unsigned
208 rb_event_ts_length(struct ring_buffer_event
*event
)
212 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
) {
213 /* time extends include the data event after it */
214 len
= RB_LEN_TIME_EXTEND
;
215 event
= skip_time_extend(event
);
217 return len
+ rb_event_length(event
);
221 * ring_buffer_event_length - return the length of the event
222 * @event: the event to get the length of
224 * Returns the size of the data load of a data event.
225 * If the event is something other than a data event, it
226 * returns the size of the event itself. With the exception
227 * of a TIME EXTEND, where it still returns the size of the
228 * data load of the data event after it.
230 unsigned ring_buffer_event_length(struct ring_buffer_event
*event
)
234 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
235 event
= skip_time_extend(event
);
237 length
= rb_event_length(event
);
238 if (event
->type_len
> RINGBUF_TYPE_DATA_TYPE_LEN_MAX
)
240 length
-= RB_EVNT_HDR_SIZE
;
241 if (length
> RB_MAX_SMALL_DATA
+ sizeof(event
->array
[0]))
242 length
-= sizeof(event
->array
[0]);
245 EXPORT_SYMBOL_GPL(ring_buffer_event_length
);
247 /* inline for ring buffer fast paths */
249 rb_event_data(struct ring_buffer_event
*event
)
251 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
252 event
= skip_time_extend(event
);
253 BUG_ON(event
->type_len
> RINGBUF_TYPE_DATA_TYPE_LEN_MAX
);
254 /* If length is in len field, then array[0] has the data */
256 return (void *)&event
->array
[0];
257 /* Otherwise length is in array[0] and array[1] has the data */
258 return (void *)&event
->array
[1];
262 * ring_buffer_event_data - return the data of the event
263 * @event: the event to get the data from
265 void *ring_buffer_event_data(struct ring_buffer_event
*event
)
267 return rb_event_data(event
);
269 EXPORT_SYMBOL_GPL(ring_buffer_event_data
);
271 #define for_each_buffer_cpu(buffer, cpu) \
272 for_each_cpu(cpu, buffer->cpumask)
275 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
276 #define TS_DELTA_TEST (~TS_MASK)
278 /* Flag when events were overwritten */
279 #define RB_MISSED_EVENTS (1 << 31)
280 /* Missed count stored at end */
281 #define RB_MISSED_STORED (1 << 30)
283 struct buffer_data_page
{
284 u64 time_stamp
; /* page time stamp */
285 local_t commit
; /* write committed index */
286 unsigned char data
[] RB_ALIGN_DATA
; /* data of buffer page */
290 * Note, the buffer_page list must be first. The buffer pages
291 * are allocated in cache lines, which means that each buffer
292 * page will be at the beginning of a cache line, and thus
293 * the least significant bits will be zero. We use this to
294 * add flags in the list struct pointers, to make the ring buffer
298 struct list_head list
; /* list of buffer pages */
299 local_t write
; /* index for next write */
300 unsigned read
; /* index for next read */
301 local_t entries
; /* entries on this page */
302 unsigned long real_end
; /* real end of data */
303 struct buffer_data_page
*page
; /* Actual data page */
307 * The buffer page counters, write and entries, must be reset
308 * atomically when crossing page boundaries. To synchronize this
309 * update, two counters are inserted into the number. One is
310 * the actual counter for the write position or count on the page.
312 * The other is a counter of updaters. Before an update happens
313 * the update partition of the counter is incremented. This will
314 * allow the updater to update the counter atomically.
316 * The counter is 20 bits, and the state data is 12.
318 #define RB_WRITE_MASK 0xfffff
319 #define RB_WRITE_INTCNT (1 << 20)
321 static void rb_init_page(struct buffer_data_page
*bpage
)
323 local_set(&bpage
->commit
, 0);
327 * ring_buffer_page_len - the size of data on the page.
328 * @page: The page to read
330 * Returns the amount of data on the page, including buffer page header.
332 size_t ring_buffer_page_len(void *page
)
334 return local_read(&((struct buffer_data_page
*)page
)->commit
)
339 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
342 static void free_buffer_page(struct buffer_page
*bpage
)
344 free_page((unsigned long)bpage
->page
);
349 * We need to fit the time_stamp delta into 27 bits.
351 static inline int test_time_stamp(u64 delta
)
353 if (delta
& TS_DELTA_TEST
)
358 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
360 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
361 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
363 int ring_buffer_print_page_header(struct trace_seq
*s
)
365 struct buffer_data_page field
;
367 trace_seq_printf(s
, "\tfield: u64 timestamp;\t"
368 "offset:0;\tsize:%u;\tsigned:%u;\n",
369 (unsigned int)sizeof(field
.time_stamp
),
370 (unsigned int)is_signed_type(u64
));
372 trace_seq_printf(s
, "\tfield: local_t commit;\t"
373 "offset:%u;\tsize:%u;\tsigned:%u;\n",
374 (unsigned int)offsetof(typeof(field
), commit
),
375 (unsigned int)sizeof(field
.commit
),
376 (unsigned int)is_signed_type(long));
378 trace_seq_printf(s
, "\tfield: int overwrite;\t"
379 "offset:%u;\tsize:%u;\tsigned:%u;\n",
380 (unsigned int)offsetof(typeof(field
), commit
),
382 (unsigned int)is_signed_type(long));
384 trace_seq_printf(s
, "\tfield: char data;\t"
385 "offset:%u;\tsize:%u;\tsigned:%u;\n",
386 (unsigned int)offsetof(typeof(field
), data
),
387 (unsigned int)BUF_PAGE_SIZE
,
388 (unsigned int)is_signed_type(char));
390 return !trace_seq_has_overflowed(s
);
394 struct irq_work work
;
395 wait_queue_head_t waiters
;
396 wait_queue_head_t full_waiters
;
397 bool waiters_pending
;
398 bool full_waiters_pending
;
403 * Structure to hold event state and handle nested events.
405 struct rb_event_info
{
408 unsigned long length
;
409 struct buffer_page
*tail_page
;
414 * Used for which event context the event is in.
420 * See trace_recursive_lock() comment below for more details.
431 * head_page == tail_page && head == tail then buffer is empty.
433 struct ring_buffer_per_cpu
{
435 atomic_t record_disabled
;
436 struct ring_buffer
*buffer
;
437 raw_spinlock_t reader_lock
; /* serialize readers */
438 arch_spinlock_t lock
;
439 struct lock_class_key lock_key
;
440 unsigned int nr_pages
;
441 unsigned int current_context
;
442 struct list_head
*pages
;
443 struct buffer_page
*head_page
; /* read from head */
444 struct buffer_page
*tail_page
; /* write to tail */
445 struct buffer_page
*commit_page
; /* committed pages */
446 struct buffer_page
*reader_page
;
447 unsigned long lost_events
;
448 unsigned long last_overrun
;
449 local_t entries_bytes
;
452 local_t commit_overrun
;
453 local_t dropped_events
;
457 unsigned long read_bytes
;
460 /* ring buffer pages to update, > 0 to add, < 0 to remove */
461 int nr_pages_to_update
;
462 struct list_head new_pages
; /* new pages to add */
463 struct work_struct update_pages_work
;
464 struct completion update_done
;
466 struct rb_irq_work irq_work
;
472 atomic_t record_disabled
;
473 atomic_t resize_disabled
;
474 cpumask_var_t cpumask
;
476 struct lock_class_key
*reader_lock_key
;
480 struct ring_buffer_per_cpu
**buffers
;
482 #ifdef CONFIG_HOTPLUG_CPU
483 struct notifier_block cpu_notify
;
487 struct rb_irq_work irq_work
;
490 struct ring_buffer_iter
{
491 struct ring_buffer_per_cpu
*cpu_buffer
;
493 struct buffer_page
*head_page
;
494 struct buffer_page
*cache_reader_page
;
495 unsigned long cache_read
;
500 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
502 * Schedules a delayed work to wake up any task that is blocked on the
503 * ring buffer waiters queue.
505 static void rb_wake_up_waiters(struct irq_work
*work
)
507 struct rb_irq_work
*rbwork
= container_of(work
, struct rb_irq_work
, work
);
509 wake_up_all(&rbwork
->waiters
);
510 if (rbwork
->wakeup_full
) {
511 rbwork
->wakeup_full
= false;
512 wake_up_all(&rbwork
->full_waiters
);
517 * ring_buffer_wait - wait for input to the ring buffer
518 * @buffer: buffer to wait on
519 * @cpu: the cpu buffer to wait on
520 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
522 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
523 * as data is added to any of the @buffer's cpu buffers. Otherwise
524 * it will wait for data to be added to a specific cpu buffer.
526 int ring_buffer_wait(struct ring_buffer
*buffer
, int cpu
, bool full
)
528 struct ring_buffer_per_cpu
*uninitialized_var(cpu_buffer
);
530 struct rb_irq_work
*work
;
534 * Depending on what the caller is waiting for, either any
535 * data in any cpu buffer, or a specific buffer, put the
536 * caller on the appropriate wait queue.
538 if (cpu
== RING_BUFFER_ALL_CPUS
) {
539 work
= &buffer
->irq_work
;
540 /* Full only makes sense on per cpu reads */
543 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
545 cpu_buffer
= buffer
->buffers
[cpu
];
546 work
= &cpu_buffer
->irq_work
;
552 prepare_to_wait(&work
->full_waiters
, &wait
, TASK_INTERRUPTIBLE
);
554 prepare_to_wait(&work
->waiters
, &wait
, TASK_INTERRUPTIBLE
);
557 * The events can happen in critical sections where
558 * checking a work queue can cause deadlocks.
559 * After adding a task to the queue, this flag is set
560 * only to notify events to try to wake up the queue
563 * We don't clear it even if the buffer is no longer
564 * empty. The flag only causes the next event to run
565 * irq_work to do the work queue wake up. The worse
566 * that can happen if we race with !trace_empty() is that
567 * an event will cause an irq_work to try to wake up
570 * There's no reason to protect this flag either, as
571 * the work queue and irq_work logic will do the necessary
572 * synchronization for the wake ups. The only thing
573 * that is necessary is that the wake up happens after
574 * a task has been queued. It's OK for spurious wake ups.
577 work
->full_waiters_pending
= true;
579 work
->waiters_pending
= true;
581 if (signal_pending(current
)) {
586 if (cpu
== RING_BUFFER_ALL_CPUS
&& !ring_buffer_empty(buffer
))
589 if (cpu
!= RING_BUFFER_ALL_CPUS
&&
590 !ring_buffer_empty_cpu(buffer
, cpu
)) {
597 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
598 pagebusy
= cpu_buffer
->reader_page
== cpu_buffer
->commit_page
;
599 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
609 finish_wait(&work
->full_waiters
, &wait
);
611 finish_wait(&work
->waiters
, &wait
);
617 * ring_buffer_poll_wait - poll on buffer input
618 * @buffer: buffer to wait on
619 * @cpu: the cpu buffer to wait on
620 * @filp: the file descriptor
621 * @poll_table: The poll descriptor
623 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
624 * as data is added to any of the @buffer's cpu buffers. Otherwise
625 * it will wait for data to be added to a specific cpu buffer.
627 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
630 int ring_buffer_poll_wait(struct ring_buffer
*buffer
, int cpu
,
631 struct file
*filp
, poll_table
*poll_table
)
633 struct ring_buffer_per_cpu
*cpu_buffer
;
634 struct rb_irq_work
*work
;
636 if (cpu
== RING_BUFFER_ALL_CPUS
)
637 work
= &buffer
->irq_work
;
639 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
642 cpu_buffer
= buffer
->buffers
[cpu
];
643 work
= &cpu_buffer
->irq_work
;
646 poll_wait(filp
, &work
->waiters
, poll_table
);
647 work
->waiters_pending
= true;
649 * There's a tight race between setting the waiters_pending and
650 * checking if the ring buffer is empty. Once the waiters_pending bit
651 * is set, the next event will wake the task up, but we can get stuck
652 * if there's only a single event in.
654 * FIXME: Ideally, we need a memory barrier on the writer side as well,
655 * but adding a memory barrier to all events will cause too much of a
656 * performance hit in the fast path. We only need a memory barrier when
657 * the buffer goes from empty to having content. But as this race is
658 * extremely small, and it's not a problem if another event comes in, we
663 if ((cpu
== RING_BUFFER_ALL_CPUS
&& !ring_buffer_empty(buffer
)) ||
664 (cpu
!= RING_BUFFER_ALL_CPUS
&& !ring_buffer_empty_cpu(buffer
, cpu
)))
665 return POLLIN
| POLLRDNORM
;
669 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
670 #define RB_WARN_ON(b, cond) \
672 int _____ret = unlikely(cond); \
674 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
675 struct ring_buffer_per_cpu *__b = \
677 atomic_inc(&__b->buffer->record_disabled); \
679 atomic_inc(&b->record_disabled); \
685 /* Up this if you want to test the TIME_EXTENTS and normalization */
686 #define DEBUG_SHIFT 0
688 static inline u64
rb_time_stamp(struct ring_buffer
*buffer
)
690 /* shift to debug/test normalization and TIME_EXTENTS */
691 return buffer
->clock() << DEBUG_SHIFT
;
694 u64
ring_buffer_time_stamp(struct ring_buffer
*buffer
, int cpu
)
698 preempt_disable_notrace();
699 time
= rb_time_stamp(buffer
);
700 preempt_enable_no_resched_notrace();
704 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp
);
706 void ring_buffer_normalize_time_stamp(struct ring_buffer
*buffer
,
709 /* Just stupid testing the normalize function and deltas */
712 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp
);
715 * Making the ring buffer lockless makes things tricky.
716 * Although writes only happen on the CPU that they are on,
717 * and they only need to worry about interrupts. Reads can
720 * The reader page is always off the ring buffer, but when the
721 * reader finishes with a page, it needs to swap its page with
722 * a new one from the buffer. The reader needs to take from
723 * the head (writes go to the tail). But if a writer is in overwrite
724 * mode and wraps, it must push the head page forward.
726 * Here lies the problem.
728 * The reader must be careful to replace only the head page, and
729 * not another one. As described at the top of the file in the
730 * ASCII art, the reader sets its old page to point to the next
731 * page after head. It then sets the page after head to point to
732 * the old reader page. But if the writer moves the head page
733 * during this operation, the reader could end up with the tail.
735 * We use cmpxchg to help prevent this race. We also do something
736 * special with the page before head. We set the LSB to 1.
738 * When the writer must push the page forward, it will clear the
739 * bit that points to the head page, move the head, and then set
740 * the bit that points to the new head page.
742 * We also don't want an interrupt coming in and moving the head
743 * page on another writer. Thus we use the second LSB to catch
746 * head->list->prev->next bit 1 bit 0
749 * Points to head page 0 1
752 * Note we can not trust the prev pointer of the head page, because:
754 * +----+ +-----+ +-----+
755 * | |------>| T |---X--->| N |
757 * +----+ +-----+ +-----+
760 * +----------| R |----------+ |
764 * Key: ---X--> HEAD flag set in pointer
769 * (see __rb_reserve_next() to see where this happens)
771 * What the above shows is that the reader just swapped out
772 * the reader page with a page in the buffer, but before it
773 * could make the new header point back to the new page added
774 * it was preempted by a writer. The writer moved forward onto
775 * the new page added by the reader and is about to move forward
778 * You can see, it is legitimate for the previous pointer of
779 * the head (or any page) not to point back to itself. But only
783 #define RB_PAGE_NORMAL 0UL
784 #define RB_PAGE_HEAD 1UL
785 #define RB_PAGE_UPDATE 2UL
788 #define RB_FLAG_MASK 3UL
790 /* PAGE_MOVED is not part of the mask */
791 #define RB_PAGE_MOVED 4UL
794 * rb_list_head - remove any bit
796 static struct list_head
*rb_list_head(struct list_head
*list
)
798 unsigned long val
= (unsigned long)list
;
800 return (struct list_head
*)(val
& ~RB_FLAG_MASK
);
804 * rb_is_head_page - test if the given page is the head page
806 * Because the reader may move the head_page pointer, we can
807 * not trust what the head page is (it may be pointing to
808 * the reader page). But if the next page is a header page,
809 * its flags will be non zero.
812 rb_is_head_page(struct ring_buffer_per_cpu
*cpu_buffer
,
813 struct buffer_page
*page
, struct list_head
*list
)
817 val
= (unsigned long)list
->next
;
819 if ((val
& ~RB_FLAG_MASK
) != (unsigned long)&page
->list
)
820 return RB_PAGE_MOVED
;
822 return val
& RB_FLAG_MASK
;
828 * The unique thing about the reader page, is that, if the
829 * writer is ever on it, the previous pointer never points
830 * back to the reader page.
832 static int rb_is_reader_page(struct buffer_page
*page
)
834 struct list_head
*list
= page
->list
.prev
;
836 return rb_list_head(list
->next
) != &page
->list
;
840 * rb_set_list_to_head - set a list_head to be pointing to head.
842 static void rb_set_list_to_head(struct ring_buffer_per_cpu
*cpu_buffer
,
843 struct list_head
*list
)
847 ptr
= (unsigned long *)&list
->next
;
848 *ptr
|= RB_PAGE_HEAD
;
849 *ptr
&= ~RB_PAGE_UPDATE
;
853 * rb_head_page_activate - sets up head page
855 static void rb_head_page_activate(struct ring_buffer_per_cpu
*cpu_buffer
)
857 struct buffer_page
*head
;
859 head
= cpu_buffer
->head_page
;
864 * Set the previous list pointer to have the HEAD flag.
866 rb_set_list_to_head(cpu_buffer
, head
->list
.prev
);
869 static void rb_list_head_clear(struct list_head
*list
)
871 unsigned long *ptr
= (unsigned long *)&list
->next
;
873 *ptr
&= ~RB_FLAG_MASK
;
877 * rb_head_page_dactivate - clears head page ptr (for free list)
880 rb_head_page_deactivate(struct ring_buffer_per_cpu
*cpu_buffer
)
882 struct list_head
*hd
;
884 /* Go through the whole list and clear any pointers found. */
885 rb_list_head_clear(cpu_buffer
->pages
);
887 list_for_each(hd
, cpu_buffer
->pages
)
888 rb_list_head_clear(hd
);
891 static int rb_head_page_set(struct ring_buffer_per_cpu
*cpu_buffer
,
892 struct buffer_page
*head
,
893 struct buffer_page
*prev
,
894 int old_flag
, int new_flag
)
896 struct list_head
*list
;
897 unsigned long val
= (unsigned long)&head
->list
;
902 val
&= ~RB_FLAG_MASK
;
904 ret
= cmpxchg((unsigned long *)&list
->next
,
905 val
| old_flag
, val
| new_flag
);
907 /* check if the reader took the page */
908 if ((ret
& ~RB_FLAG_MASK
) != val
)
909 return RB_PAGE_MOVED
;
911 return ret
& RB_FLAG_MASK
;
914 static int rb_head_page_set_update(struct ring_buffer_per_cpu
*cpu_buffer
,
915 struct buffer_page
*head
,
916 struct buffer_page
*prev
,
919 return rb_head_page_set(cpu_buffer
, head
, prev
,
920 old_flag
, RB_PAGE_UPDATE
);
923 static int rb_head_page_set_head(struct ring_buffer_per_cpu
*cpu_buffer
,
924 struct buffer_page
*head
,
925 struct buffer_page
*prev
,
928 return rb_head_page_set(cpu_buffer
, head
, prev
,
929 old_flag
, RB_PAGE_HEAD
);
932 static int rb_head_page_set_normal(struct ring_buffer_per_cpu
*cpu_buffer
,
933 struct buffer_page
*head
,
934 struct buffer_page
*prev
,
937 return rb_head_page_set(cpu_buffer
, head
, prev
,
938 old_flag
, RB_PAGE_NORMAL
);
941 static inline void rb_inc_page(struct ring_buffer_per_cpu
*cpu_buffer
,
942 struct buffer_page
**bpage
)
944 struct list_head
*p
= rb_list_head((*bpage
)->list
.next
);
946 *bpage
= list_entry(p
, struct buffer_page
, list
);
949 static struct buffer_page
*
950 rb_set_head_page(struct ring_buffer_per_cpu
*cpu_buffer
)
952 struct buffer_page
*head
;
953 struct buffer_page
*page
;
954 struct list_head
*list
;
957 if (RB_WARN_ON(cpu_buffer
, !cpu_buffer
->head_page
))
961 list
= cpu_buffer
->pages
;
962 if (RB_WARN_ON(cpu_buffer
, rb_list_head(list
->prev
->next
) != list
))
965 page
= head
= cpu_buffer
->head_page
;
967 * It is possible that the writer moves the header behind
968 * where we started, and we miss in one loop.
969 * A second loop should grab the header, but we'll do
970 * three loops just because I'm paranoid.
972 for (i
= 0; i
< 3; i
++) {
974 if (rb_is_head_page(cpu_buffer
, page
, page
->list
.prev
)) {
975 cpu_buffer
->head_page
= page
;
978 rb_inc_page(cpu_buffer
, &page
);
979 } while (page
!= head
);
982 RB_WARN_ON(cpu_buffer
, 1);
987 static int rb_head_page_replace(struct buffer_page
*old
,
988 struct buffer_page
*new)
990 unsigned long *ptr
= (unsigned long *)&old
->list
.prev
->next
;
994 val
= *ptr
& ~RB_FLAG_MASK
;
997 ret
= cmpxchg(ptr
, val
, (unsigned long)&new->list
);
1003 * rb_tail_page_update - move the tail page forward
1005 * Returns 1 if moved tail page, 0 if someone else did.
1007 static int rb_tail_page_update(struct ring_buffer_per_cpu
*cpu_buffer
,
1008 struct buffer_page
*tail_page
,
1009 struct buffer_page
*next_page
)
1011 struct buffer_page
*old_tail
;
1012 unsigned long old_entries
;
1013 unsigned long old_write
;
1017 * The tail page now needs to be moved forward.
1019 * We need to reset the tail page, but without messing
1020 * with possible erasing of data brought in by interrupts
1021 * that have moved the tail page and are currently on it.
1023 * We add a counter to the write field to denote this.
1025 old_write
= local_add_return(RB_WRITE_INTCNT
, &next_page
->write
);
1026 old_entries
= local_add_return(RB_WRITE_INTCNT
, &next_page
->entries
);
1029 * Just make sure we have seen our old_write and synchronize
1030 * with any interrupts that come in.
1035 * If the tail page is still the same as what we think
1036 * it is, then it is up to us to update the tail
1039 if (tail_page
== cpu_buffer
->tail_page
) {
1040 /* Zero the write counter */
1041 unsigned long val
= old_write
& ~RB_WRITE_MASK
;
1042 unsigned long eval
= old_entries
& ~RB_WRITE_MASK
;
1045 * This will only succeed if an interrupt did
1046 * not come in and change it. In which case, we
1047 * do not want to modify it.
1049 * We add (void) to let the compiler know that we do not care
1050 * about the return value of these functions. We use the
1051 * cmpxchg to only update if an interrupt did not already
1052 * do it for us. If the cmpxchg fails, we don't care.
1054 (void)local_cmpxchg(&next_page
->write
, old_write
, val
);
1055 (void)local_cmpxchg(&next_page
->entries
, old_entries
, eval
);
1058 * No need to worry about races with clearing out the commit.
1059 * it only can increment when a commit takes place. But that
1060 * only happens in the outer most nested commit.
1062 local_set(&next_page
->page
->commit
, 0);
1064 old_tail
= cmpxchg(&cpu_buffer
->tail_page
,
1065 tail_page
, next_page
);
1067 if (old_tail
== tail_page
)
1074 static int rb_check_bpage(struct ring_buffer_per_cpu
*cpu_buffer
,
1075 struct buffer_page
*bpage
)
1077 unsigned long val
= (unsigned long)bpage
;
1079 if (RB_WARN_ON(cpu_buffer
, val
& RB_FLAG_MASK
))
1086 * rb_check_list - make sure a pointer to a list has the last bits zero
1088 static int rb_check_list(struct ring_buffer_per_cpu
*cpu_buffer
,
1089 struct list_head
*list
)
1091 if (RB_WARN_ON(cpu_buffer
, rb_list_head(list
->prev
) != list
->prev
))
1093 if (RB_WARN_ON(cpu_buffer
, rb_list_head(list
->next
) != list
->next
))
1099 * rb_check_pages - integrity check of buffer pages
1100 * @cpu_buffer: CPU buffer with pages to test
1102 * As a safety measure we check to make sure the data pages have not
1105 static int rb_check_pages(struct ring_buffer_per_cpu
*cpu_buffer
)
1107 struct list_head
*head
= cpu_buffer
->pages
;
1108 struct buffer_page
*bpage
, *tmp
;
1110 /* Reset the head page if it exists */
1111 if (cpu_buffer
->head_page
)
1112 rb_set_head_page(cpu_buffer
);
1114 rb_head_page_deactivate(cpu_buffer
);
1116 if (RB_WARN_ON(cpu_buffer
, head
->next
->prev
!= head
))
1118 if (RB_WARN_ON(cpu_buffer
, head
->prev
->next
!= head
))
1121 if (rb_check_list(cpu_buffer
, head
))
1124 list_for_each_entry_safe(bpage
, tmp
, head
, list
) {
1125 if (RB_WARN_ON(cpu_buffer
,
1126 bpage
->list
.next
->prev
!= &bpage
->list
))
1128 if (RB_WARN_ON(cpu_buffer
,
1129 bpage
->list
.prev
->next
!= &bpage
->list
))
1131 if (rb_check_list(cpu_buffer
, &bpage
->list
))
1135 rb_head_page_activate(cpu_buffer
);
1140 static int __rb_allocate_pages(int nr_pages
, struct list_head
*pages
, int cpu
)
1143 struct buffer_page
*bpage
, *tmp
;
1145 for (i
= 0; i
< nr_pages
; i
++) {
1148 * __GFP_NORETRY flag makes sure that the allocation fails
1149 * gracefully without invoking oom-killer and the system is
1152 bpage
= kzalloc_node(ALIGN(sizeof(*bpage
), cache_line_size()),
1153 GFP_KERNEL
| __GFP_NORETRY
,
1158 list_add(&bpage
->list
, pages
);
1160 page
= alloc_pages_node(cpu_to_node(cpu
),
1161 GFP_KERNEL
| __GFP_NORETRY
, 0);
1164 bpage
->page
= page_address(page
);
1165 rb_init_page(bpage
->page
);
1171 list_for_each_entry_safe(bpage
, tmp
, pages
, list
) {
1172 list_del_init(&bpage
->list
);
1173 free_buffer_page(bpage
);
1179 static int rb_allocate_pages(struct ring_buffer_per_cpu
*cpu_buffer
,
1186 if (__rb_allocate_pages(nr_pages
, &pages
, cpu_buffer
->cpu
))
1190 * The ring buffer page list is a circular list that does not
1191 * start and end with a list head. All page list items point to
1194 cpu_buffer
->pages
= pages
.next
;
1197 cpu_buffer
->nr_pages
= nr_pages
;
1199 rb_check_pages(cpu_buffer
);
1204 static struct ring_buffer_per_cpu
*
1205 rb_allocate_cpu_buffer(struct ring_buffer
*buffer
, int nr_pages
, int cpu
)
1207 struct ring_buffer_per_cpu
*cpu_buffer
;
1208 struct buffer_page
*bpage
;
1212 cpu_buffer
= kzalloc_node(ALIGN(sizeof(*cpu_buffer
), cache_line_size()),
1213 GFP_KERNEL
, cpu_to_node(cpu
));
1217 cpu_buffer
->cpu
= cpu
;
1218 cpu_buffer
->buffer
= buffer
;
1219 raw_spin_lock_init(&cpu_buffer
->reader_lock
);
1220 lockdep_set_class(&cpu_buffer
->reader_lock
, buffer
->reader_lock_key
);
1221 cpu_buffer
->lock
= (arch_spinlock_t
)__ARCH_SPIN_LOCK_UNLOCKED
;
1222 INIT_WORK(&cpu_buffer
->update_pages_work
, update_pages_handler
);
1223 init_completion(&cpu_buffer
->update_done
);
1224 init_irq_work(&cpu_buffer
->irq_work
.work
, rb_wake_up_waiters
);
1225 init_waitqueue_head(&cpu_buffer
->irq_work
.waiters
);
1226 init_waitqueue_head(&cpu_buffer
->irq_work
.full_waiters
);
1228 bpage
= kzalloc_node(ALIGN(sizeof(*bpage
), cache_line_size()),
1229 GFP_KERNEL
, cpu_to_node(cpu
));
1231 goto fail_free_buffer
;
1233 rb_check_bpage(cpu_buffer
, bpage
);
1235 cpu_buffer
->reader_page
= bpage
;
1236 page
= alloc_pages_node(cpu_to_node(cpu
), GFP_KERNEL
, 0);
1238 goto fail_free_reader
;
1239 bpage
->page
= page_address(page
);
1240 rb_init_page(bpage
->page
);
1242 INIT_LIST_HEAD(&cpu_buffer
->reader_page
->list
);
1243 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
1245 ret
= rb_allocate_pages(cpu_buffer
, nr_pages
);
1247 goto fail_free_reader
;
1249 cpu_buffer
->head_page
1250 = list_entry(cpu_buffer
->pages
, struct buffer_page
, list
);
1251 cpu_buffer
->tail_page
= cpu_buffer
->commit_page
= cpu_buffer
->head_page
;
1253 rb_head_page_activate(cpu_buffer
);
1258 free_buffer_page(cpu_buffer
->reader_page
);
1265 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu
*cpu_buffer
)
1267 struct list_head
*head
= cpu_buffer
->pages
;
1268 struct buffer_page
*bpage
, *tmp
;
1270 free_buffer_page(cpu_buffer
->reader_page
);
1272 rb_head_page_deactivate(cpu_buffer
);
1275 list_for_each_entry_safe(bpage
, tmp
, head
, list
) {
1276 list_del_init(&bpage
->list
);
1277 free_buffer_page(bpage
);
1279 bpage
= list_entry(head
, struct buffer_page
, list
);
1280 free_buffer_page(bpage
);
1286 #ifdef CONFIG_HOTPLUG_CPU
1287 static int rb_cpu_notify(struct notifier_block
*self
,
1288 unsigned long action
, void *hcpu
);
1292 * __ring_buffer_alloc - allocate a new ring_buffer
1293 * @size: the size in bytes per cpu that is needed.
1294 * @flags: attributes to set for the ring buffer.
1296 * Currently the only flag that is available is the RB_FL_OVERWRITE
1297 * flag. This flag means that the buffer will overwrite old data
1298 * when the buffer wraps. If this flag is not set, the buffer will
1299 * drop data when the tail hits the head.
1301 struct ring_buffer
*__ring_buffer_alloc(unsigned long size
, unsigned flags
,
1302 struct lock_class_key
*key
)
1304 struct ring_buffer
*buffer
;
1308 /* keep it in its own cache line */
1309 buffer
= kzalloc(ALIGN(sizeof(*buffer
), cache_line_size()),
1314 if (!alloc_cpumask_var(&buffer
->cpumask
, GFP_KERNEL
))
1315 goto fail_free_buffer
;
1317 nr_pages
= DIV_ROUND_UP(size
, BUF_PAGE_SIZE
);
1318 buffer
->flags
= flags
;
1319 buffer
->clock
= trace_clock_local
;
1320 buffer
->reader_lock_key
= key
;
1322 init_irq_work(&buffer
->irq_work
.work
, rb_wake_up_waiters
);
1323 init_waitqueue_head(&buffer
->irq_work
.waiters
);
1325 /* need at least two pages */
1330 * In case of non-hotplug cpu, if the ring-buffer is allocated
1331 * in early initcall, it will not be notified of secondary cpus.
1332 * In that off case, we need to allocate for all possible cpus.
1334 #ifdef CONFIG_HOTPLUG_CPU
1335 cpu_notifier_register_begin();
1336 cpumask_copy(buffer
->cpumask
, cpu_online_mask
);
1338 cpumask_copy(buffer
->cpumask
, cpu_possible_mask
);
1340 buffer
->cpus
= nr_cpu_ids
;
1342 bsize
= sizeof(void *) * nr_cpu_ids
;
1343 buffer
->buffers
= kzalloc(ALIGN(bsize
, cache_line_size()),
1345 if (!buffer
->buffers
)
1346 goto fail_free_cpumask
;
1348 for_each_buffer_cpu(buffer
, cpu
) {
1349 buffer
->buffers
[cpu
] =
1350 rb_allocate_cpu_buffer(buffer
, nr_pages
, cpu
);
1351 if (!buffer
->buffers
[cpu
])
1352 goto fail_free_buffers
;
1355 #ifdef CONFIG_HOTPLUG_CPU
1356 buffer
->cpu_notify
.notifier_call
= rb_cpu_notify
;
1357 buffer
->cpu_notify
.priority
= 0;
1358 __register_cpu_notifier(&buffer
->cpu_notify
);
1359 cpu_notifier_register_done();
1362 mutex_init(&buffer
->mutex
);
1367 for_each_buffer_cpu(buffer
, cpu
) {
1368 if (buffer
->buffers
[cpu
])
1369 rb_free_cpu_buffer(buffer
->buffers
[cpu
]);
1371 kfree(buffer
->buffers
);
1374 free_cpumask_var(buffer
->cpumask
);
1375 #ifdef CONFIG_HOTPLUG_CPU
1376 cpu_notifier_register_done();
1383 EXPORT_SYMBOL_GPL(__ring_buffer_alloc
);
1386 * ring_buffer_free - free a ring buffer.
1387 * @buffer: the buffer to free.
1390 ring_buffer_free(struct ring_buffer
*buffer
)
1394 #ifdef CONFIG_HOTPLUG_CPU
1395 cpu_notifier_register_begin();
1396 __unregister_cpu_notifier(&buffer
->cpu_notify
);
1399 for_each_buffer_cpu(buffer
, cpu
)
1400 rb_free_cpu_buffer(buffer
->buffers
[cpu
]);
1402 #ifdef CONFIG_HOTPLUG_CPU
1403 cpu_notifier_register_done();
1406 kfree(buffer
->buffers
);
1407 free_cpumask_var(buffer
->cpumask
);
1411 EXPORT_SYMBOL_GPL(ring_buffer_free
);
1413 void ring_buffer_set_clock(struct ring_buffer
*buffer
,
1416 buffer
->clock
= clock
;
1419 static void rb_reset_cpu(struct ring_buffer_per_cpu
*cpu_buffer
);
1421 static inline unsigned long rb_page_entries(struct buffer_page
*bpage
)
1423 return local_read(&bpage
->entries
) & RB_WRITE_MASK
;
1426 static inline unsigned long rb_page_write(struct buffer_page
*bpage
)
1428 return local_read(&bpage
->write
) & RB_WRITE_MASK
;
1432 rb_remove_pages(struct ring_buffer_per_cpu
*cpu_buffer
, unsigned int nr_pages
)
1434 struct list_head
*tail_page
, *to_remove
, *next_page
;
1435 struct buffer_page
*to_remove_page
, *tmp_iter_page
;
1436 struct buffer_page
*last_page
, *first_page
;
1437 unsigned int nr_removed
;
1438 unsigned long head_bit
;
1443 raw_spin_lock_irq(&cpu_buffer
->reader_lock
);
1444 atomic_inc(&cpu_buffer
->record_disabled
);
1446 * We don't race with the readers since we have acquired the reader
1447 * lock. We also don't race with writers after disabling recording.
1448 * This makes it easy to figure out the first and the last page to be
1449 * removed from the list. We unlink all the pages in between including
1450 * the first and last pages. This is done in a busy loop so that we
1451 * lose the least number of traces.
1452 * The pages are freed after we restart recording and unlock readers.
1454 tail_page
= &cpu_buffer
->tail_page
->list
;
1457 * tail page might be on reader page, we remove the next page
1458 * from the ring buffer
1460 if (cpu_buffer
->tail_page
== cpu_buffer
->reader_page
)
1461 tail_page
= rb_list_head(tail_page
->next
);
1462 to_remove
= tail_page
;
1464 /* start of pages to remove */
1465 first_page
= list_entry(rb_list_head(to_remove
->next
),
1466 struct buffer_page
, list
);
1468 for (nr_removed
= 0; nr_removed
< nr_pages
; nr_removed
++) {
1469 to_remove
= rb_list_head(to_remove
)->next
;
1470 head_bit
|= (unsigned long)to_remove
& RB_PAGE_HEAD
;
1473 next_page
= rb_list_head(to_remove
)->next
;
1476 * Now we remove all pages between tail_page and next_page.
1477 * Make sure that we have head_bit value preserved for the
1480 tail_page
->next
= (struct list_head
*)((unsigned long)next_page
|
1482 next_page
= rb_list_head(next_page
);
1483 next_page
->prev
= tail_page
;
1485 /* make sure pages points to a valid page in the ring buffer */
1486 cpu_buffer
->pages
= next_page
;
1488 /* update head page */
1490 cpu_buffer
->head_page
= list_entry(next_page
,
1491 struct buffer_page
, list
);
1494 * change read pointer to make sure any read iterators reset
1497 cpu_buffer
->read
= 0;
1499 /* pages are removed, resume tracing and then free the pages */
1500 atomic_dec(&cpu_buffer
->record_disabled
);
1501 raw_spin_unlock_irq(&cpu_buffer
->reader_lock
);
1503 RB_WARN_ON(cpu_buffer
, list_empty(cpu_buffer
->pages
));
1505 /* last buffer page to remove */
1506 last_page
= list_entry(rb_list_head(to_remove
), struct buffer_page
,
1508 tmp_iter_page
= first_page
;
1511 to_remove_page
= tmp_iter_page
;
1512 rb_inc_page(cpu_buffer
, &tmp_iter_page
);
1514 /* update the counters */
1515 page_entries
= rb_page_entries(to_remove_page
);
1518 * If something was added to this page, it was full
1519 * since it is not the tail page. So we deduct the
1520 * bytes consumed in ring buffer from here.
1521 * Increment overrun to account for the lost events.
1523 local_add(page_entries
, &cpu_buffer
->overrun
);
1524 local_sub(BUF_PAGE_SIZE
, &cpu_buffer
->entries_bytes
);
1528 * We have already removed references to this list item, just
1529 * free up the buffer_page and its page
1531 free_buffer_page(to_remove_page
);
1534 } while (to_remove_page
!= last_page
);
1536 RB_WARN_ON(cpu_buffer
, nr_removed
);
1538 return nr_removed
== 0;
1542 rb_insert_pages(struct ring_buffer_per_cpu
*cpu_buffer
)
1544 struct list_head
*pages
= &cpu_buffer
->new_pages
;
1545 int retries
, success
;
1547 raw_spin_lock_irq(&cpu_buffer
->reader_lock
);
1549 * We are holding the reader lock, so the reader page won't be swapped
1550 * in the ring buffer. Now we are racing with the writer trying to
1551 * move head page and the tail page.
1552 * We are going to adapt the reader page update process where:
1553 * 1. We first splice the start and end of list of new pages between
1554 * the head page and its previous page.
1555 * 2. We cmpxchg the prev_page->next to point from head page to the
1556 * start of new pages list.
1557 * 3. Finally, we update the head->prev to the end of new list.
1559 * We will try this process 10 times, to make sure that we don't keep
1565 struct list_head
*head_page
, *prev_page
, *r
;
1566 struct list_head
*last_page
, *first_page
;
1567 struct list_head
*head_page_with_bit
;
1569 head_page
= &rb_set_head_page(cpu_buffer
)->list
;
1572 prev_page
= head_page
->prev
;
1574 first_page
= pages
->next
;
1575 last_page
= pages
->prev
;
1577 head_page_with_bit
= (struct list_head
*)
1578 ((unsigned long)head_page
| RB_PAGE_HEAD
);
1580 last_page
->next
= head_page_with_bit
;
1581 first_page
->prev
= prev_page
;
1583 r
= cmpxchg(&prev_page
->next
, head_page_with_bit
, first_page
);
1585 if (r
== head_page_with_bit
) {
1587 * yay, we replaced the page pointer to our new list,
1588 * now, we just have to update to head page's prev
1589 * pointer to point to end of list
1591 head_page
->prev
= last_page
;
1598 INIT_LIST_HEAD(pages
);
1600 * If we weren't successful in adding in new pages, warn and stop
1603 RB_WARN_ON(cpu_buffer
, !success
);
1604 raw_spin_unlock_irq(&cpu_buffer
->reader_lock
);
1606 /* free pages if they weren't inserted */
1608 struct buffer_page
*bpage
, *tmp
;
1609 list_for_each_entry_safe(bpage
, tmp
, &cpu_buffer
->new_pages
,
1611 list_del_init(&bpage
->list
);
1612 free_buffer_page(bpage
);
1618 static void rb_update_pages(struct ring_buffer_per_cpu
*cpu_buffer
)
1622 if (cpu_buffer
->nr_pages_to_update
> 0)
1623 success
= rb_insert_pages(cpu_buffer
);
1625 success
= rb_remove_pages(cpu_buffer
,
1626 -cpu_buffer
->nr_pages_to_update
);
1629 cpu_buffer
->nr_pages
+= cpu_buffer
->nr_pages_to_update
;
1632 static void update_pages_handler(struct work_struct
*work
)
1634 struct ring_buffer_per_cpu
*cpu_buffer
= container_of(work
,
1635 struct ring_buffer_per_cpu
, update_pages_work
);
1636 rb_update_pages(cpu_buffer
);
1637 complete(&cpu_buffer
->update_done
);
1641 * ring_buffer_resize - resize the ring buffer
1642 * @buffer: the buffer to resize.
1643 * @size: the new size.
1644 * @cpu_id: the cpu buffer to resize
1646 * Minimum size is 2 * BUF_PAGE_SIZE.
1648 * Returns 0 on success and < 0 on failure.
1650 int ring_buffer_resize(struct ring_buffer
*buffer
, unsigned long size
,
1653 struct ring_buffer_per_cpu
*cpu_buffer
;
1658 * Always succeed at resizing a non-existent buffer:
1663 /* Make sure the requested buffer exists */
1664 if (cpu_id
!= RING_BUFFER_ALL_CPUS
&&
1665 !cpumask_test_cpu(cpu_id
, buffer
->cpumask
))
1668 size
= DIV_ROUND_UP(size
, BUF_PAGE_SIZE
);
1669 size
*= BUF_PAGE_SIZE
;
1671 /* we need a minimum of two pages */
1672 if (size
< BUF_PAGE_SIZE
* 2)
1673 size
= BUF_PAGE_SIZE
* 2;
1675 nr_pages
= DIV_ROUND_UP(size
, BUF_PAGE_SIZE
);
1678 * Don't succeed if resizing is disabled, as a reader might be
1679 * manipulating the ring buffer and is expecting a sane state while
1682 if (atomic_read(&buffer
->resize_disabled
))
1685 /* prevent another thread from changing buffer sizes */
1686 mutex_lock(&buffer
->mutex
);
1688 if (cpu_id
== RING_BUFFER_ALL_CPUS
) {
1689 /* calculate the pages to update */
1690 for_each_buffer_cpu(buffer
, cpu
) {
1691 cpu_buffer
= buffer
->buffers
[cpu
];
1693 cpu_buffer
->nr_pages_to_update
= nr_pages
-
1694 cpu_buffer
->nr_pages
;
1696 * nothing more to do for removing pages or no update
1698 if (cpu_buffer
->nr_pages_to_update
<= 0)
1701 * to add pages, make sure all new pages can be
1702 * allocated without receiving ENOMEM
1704 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
1705 if (__rb_allocate_pages(cpu_buffer
->nr_pages_to_update
,
1706 &cpu_buffer
->new_pages
, cpu
)) {
1707 /* not enough memory for new pages */
1715 * Fire off all the required work handlers
1716 * We can't schedule on offline CPUs, but it's not necessary
1717 * since we can change their buffer sizes without any race.
1719 for_each_buffer_cpu(buffer
, cpu
) {
1720 cpu_buffer
= buffer
->buffers
[cpu
];
1721 if (!cpu_buffer
->nr_pages_to_update
)
1724 /* Can't run something on an offline CPU. */
1725 if (!cpu_online(cpu
)) {
1726 rb_update_pages(cpu_buffer
);
1727 cpu_buffer
->nr_pages_to_update
= 0;
1729 schedule_work_on(cpu
,
1730 &cpu_buffer
->update_pages_work
);
1734 /* wait for all the updates to complete */
1735 for_each_buffer_cpu(buffer
, cpu
) {
1736 cpu_buffer
= buffer
->buffers
[cpu
];
1737 if (!cpu_buffer
->nr_pages_to_update
)
1740 if (cpu_online(cpu
))
1741 wait_for_completion(&cpu_buffer
->update_done
);
1742 cpu_buffer
->nr_pages_to_update
= 0;
1747 /* Make sure this CPU has been intitialized */
1748 if (!cpumask_test_cpu(cpu_id
, buffer
->cpumask
))
1751 cpu_buffer
= buffer
->buffers
[cpu_id
];
1753 if (nr_pages
== cpu_buffer
->nr_pages
)
1756 cpu_buffer
->nr_pages_to_update
= nr_pages
-
1757 cpu_buffer
->nr_pages
;
1759 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
1760 if (cpu_buffer
->nr_pages_to_update
> 0 &&
1761 __rb_allocate_pages(cpu_buffer
->nr_pages_to_update
,
1762 &cpu_buffer
->new_pages
, cpu_id
)) {
1769 /* Can't run something on an offline CPU. */
1770 if (!cpu_online(cpu_id
))
1771 rb_update_pages(cpu_buffer
);
1773 schedule_work_on(cpu_id
,
1774 &cpu_buffer
->update_pages_work
);
1775 wait_for_completion(&cpu_buffer
->update_done
);
1778 cpu_buffer
->nr_pages_to_update
= 0;
1784 * The ring buffer resize can happen with the ring buffer
1785 * enabled, so that the update disturbs the tracing as little
1786 * as possible. But if the buffer is disabled, we do not need
1787 * to worry about that, and we can take the time to verify
1788 * that the buffer is not corrupt.
1790 if (atomic_read(&buffer
->record_disabled
)) {
1791 atomic_inc(&buffer
->record_disabled
);
1793 * Even though the buffer was disabled, we must make sure
1794 * that it is truly disabled before calling rb_check_pages.
1795 * There could have been a race between checking
1796 * record_disable and incrementing it.
1798 synchronize_sched();
1799 for_each_buffer_cpu(buffer
, cpu
) {
1800 cpu_buffer
= buffer
->buffers
[cpu
];
1801 rb_check_pages(cpu_buffer
);
1803 atomic_dec(&buffer
->record_disabled
);
1806 mutex_unlock(&buffer
->mutex
);
1810 for_each_buffer_cpu(buffer
, cpu
) {
1811 struct buffer_page
*bpage
, *tmp
;
1813 cpu_buffer
= buffer
->buffers
[cpu
];
1814 cpu_buffer
->nr_pages_to_update
= 0;
1816 if (list_empty(&cpu_buffer
->new_pages
))
1819 list_for_each_entry_safe(bpage
, tmp
, &cpu_buffer
->new_pages
,
1821 list_del_init(&bpage
->list
);
1822 free_buffer_page(bpage
);
1825 mutex_unlock(&buffer
->mutex
);
1828 EXPORT_SYMBOL_GPL(ring_buffer_resize
);
1830 void ring_buffer_change_overwrite(struct ring_buffer
*buffer
, int val
)
1832 mutex_lock(&buffer
->mutex
);
1834 buffer
->flags
|= RB_FL_OVERWRITE
;
1836 buffer
->flags
&= ~RB_FL_OVERWRITE
;
1837 mutex_unlock(&buffer
->mutex
);
1839 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite
);
1841 static inline void *
1842 __rb_data_page_index(struct buffer_data_page
*bpage
, unsigned index
)
1844 return bpage
->data
+ index
;
1847 static inline void *__rb_page_index(struct buffer_page
*bpage
, unsigned index
)
1849 return bpage
->page
->data
+ index
;
1852 static inline struct ring_buffer_event
*
1853 rb_reader_event(struct ring_buffer_per_cpu
*cpu_buffer
)
1855 return __rb_page_index(cpu_buffer
->reader_page
,
1856 cpu_buffer
->reader_page
->read
);
1859 static inline struct ring_buffer_event
*
1860 rb_iter_head_event(struct ring_buffer_iter
*iter
)
1862 return __rb_page_index(iter
->head_page
, iter
->head
);
1865 static inline unsigned rb_page_commit(struct buffer_page
*bpage
)
1867 return local_read(&bpage
->page
->commit
);
1870 /* Size is determined by what has been committed */
1871 static inline unsigned rb_page_size(struct buffer_page
*bpage
)
1873 return rb_page_commit(bpage
);
1876 static inline unsigned
1877 rb_commit_index(struct ring_buffer_per_cpu
*cpu_buffer
)
1879 return rb_page_commit(cpu_buffer
->commit_page
);
1882 static inline unsigned
1883 rb_event_index(struct ring_buffer_event
*event
)
1885 unsigned long addr
= (unsigned long)event
;
1887 return (addr
& ~PAGE_MASK
) - BUF_PAGE_HDR_SIZE
;
1890 static void rb_reset_reader_page(struct ring_buffer_per_cpu
*cpu_buffer
)
1892 cpu_buffer
->read_stamp
= cpu_buffer
->reader_page
->page
->time_stamp
;
1893 cpu_buffer
->reader_page
->read
= 0;
1896 static void rb_inc_iter(struct ring_buffer_iter
*iter
)
1898 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
1901 * The iterator could be on the reader page (it starts there).
1902 * But the head could have moved, since the reader was
1903 * found. Check for this case and assign the iterator
1904 * to the head page instead of next.
1906 if (iter
->head_page
== cpu_buffer
->reader_page
)
1907 iter
->head_page
= rb_set_head_page(cpu_buffer
);
1909 rb_inc_page(cpu_buffer
, &iter
->head_page
);
1911 iter
->read_stamp
= iter
->head_page
->page
->time_stamp
;
1916 * rb_handle_head_page - writer hit the head page
1918 * Returns: +1 to retry page
1923 rb_handle_head_page(struct ring_buffer_per_cpu
*cpu_buffer
,
1924 struct buffer_page
*tail_page
,
1925 struct buffer_page
*next_page
)
1927 struct buffer_page
*new_head
;
1932 entries
= rb_page_entries(next_page
);
1935 * The hard part is here. We need to move the head
1936 * forward, and protect against both readers on
1937 * other CPUs and writers coming in via interrupts.
1939 type
= rb_head_page_set_update(cpu_buffer
, next_page
, tail_page
,
1943 * type can be one of four:
1944 * NORMAL - an interrupt already moved it for us
1945 * HEAD - we are the first to get here.
1946 * UPDATE - we are the interrupt interrupting
1948 * MOVED - a reader on another CPU moved the next
1949 * pointer to its reader page. Give up
1956 * We changed the head to UPDATE, thus
1957 * it is our responsibility to update
1960 local_add(entries
, &cpu_buffer
->overrun
);
1961 local_sub(BUF_PAGE_SIZE
, &cpu_buffer
->entries_bytes
);
1964 * The entries will be zeroed out when we move the
1968 /* still more to do */
1971 case RB_PAGE_UPDATE
:
1973 * This is an interrupt that interrupt the
1974 * previous update. Still more to do.
1977 case RB_PAGE_NORMAL
:
1979 * An interrupt came in before the update
1980 * and processed this for us.
1981 * Nothing left to do.
1986 * The reader is on another CPU and just did
1987 * a swap with our next_page.
1992 RB_WARN_ON(cpu_buffer
, 1); /* WTF??? */
1997 * Now that we are here, the old head pointer is
1998 * set to UPDATE. This will keep the reader from
1999 * swapping the head page with the reader page.
2000 * The reader (on another CPU) will spin till
2003 * We just need to protect against interrupts
2004 * doing the job. We will set the next pointer
2005 * to HEAD. After that, we set the old pointer
2006 * to NORMAL, but only if it was HEAD before.
2007 * otherwise we are an interrupt, and only
2008 * want the outer most commit to reset it.
2010 new_head
= next_page
;
2011 rb_inc_page(cpu_buffer
, &new_head
);
2013 ret
= rb_head_page_set_head(cpu_buffer
, new_head
, next_page
,
2017 * Valid returns are:
2018 * HEAD - an interrupt came in and already set it.
2019 * NORMAL - One of two things:
2020 * 1) We really set it.
2021 * 2) A bunch of interrupts came in and moved
2022 * the page forward again.
2026 case RB_PAGE_NORMAL
:
2030 RB_WARN_ON(cpu_buffer
, 1);
2035 * It is possible that an interrupt came in,
2036 * set the head up, then more interrupts came in
2037 * and moved it again. When we get back here,
2038 * the page would have been set to NORMAL but we
2039 * just set it back to HEAD.
2041 * How do you detect this? Well, if that happened
2042 * the tail page would have moved.
2044 if (ret
== RB_PAGE_NORMAL
) {
2046 * If the tail had moved passed next, then we need
2047 * to reset the pointer.
2049 if (cpu_buffer
->tail_page
!= tail_page
&&
2050 cpu_buffer
->tail_page
!= next_page
)
2051 rb_head_page_set_normal(cpu_buffer
, new_head
,
2057 * If this was the outer most commit (the one that
2058 * changed the original pointer from HEAD to UPDATE),
2059 * then it is up to us to reset it to NORMAL.
2061 if (type
== RB_PAGE_HEAD
) {
2062 ret
= rb_head_page_set_normal(cpu_buffer
, next_page
,
2065 if (RB_WARN_ON(cpu_buffer
,
2066 ret
!= RB_PAGE_UPDATE
))
2074 rb_reset_tail(struct ring_buffer_per_cpu
*cpu_buffer
,
2075 unsigned long tail
, struct rb_event_info
*info
)
2077 struct buffer_page
*tail_page
= info
->tail_page
;
2078 struct ring_buffer_event
*event
;
2079 unsigned long length
= info
->length
;
2082 * Only the event that crossed the page boundary
2083 * must fill the old tail_page with padding.
2085 if (tail
>= BUF_PAGE_SIZE
) {
2087 * If the page was filled, then we still need
2088 * to update the real_end. Reset it to zero
2089 * and the reader will ignore it.
2091 if (tail
== BUF_PAGE_SIZE
)
2092 tail_page
->real_end
= 0;
2094 local_sub(length
, &tail_page
->write
);
2098 event
= __rb_page_index(tail_page
, tail
);
2099 kmemcheck_annotate_bitfield(event
, bitfield
);
2101 /* account for padding bytes */
2102 local_add(BUF_PAGE_SIZE
- tail
, &cpu_buffer
->entries_bytes
);
2105 * Save the original length to the meta data.
2106 * This will be used by the reader to add lost event
2109 tail_page
->real_end
= tail
;
2112 * If this event is bigger than the minimum size, then
2113 * we need to be careful that we don't subtract the
2114 * write counter enough to allow another writer to slip
2116 * We put in a discarded commit instead, to make sure
2117 * that this space is not used again.
2119 * If we are less than the minimum size, we don't need to
2122 if (tail
> (BUF_PAGE_SIZE
- RB_EVNT_MIN_SIZE
)) {
2123 /* No room for any events */
2125 /* Mark the rest of the page with padding */
2126 rb_event_set_padding(event
);
2128 /* Set the write back to the previous setting */
2129 local_sub(length
, &tail_page
->write
);
2133 /* Put in a discarded event */
2134 event
->array
[0] = (BUF_PAGE_SIZE
- tail
) - RB_EVNT_HDR_SIZE
;
2135 event
->type_len
= RINGBUF_TYPE_PADDING
;
2136 /* time delta must be non zero */
2137 event
->time_delta
= 1;
2139 /* Set write to end of buffer */
2140 length
= (tail
+ length
) - BUF_PAGE_SIZE
;
2141 local_sub(length
, &tail_page
->write
);
2145 * This is the slow path, force gcc not to inline it.
2147 static noinline
struct ring_buffer_event
*
2148 rb_move_tail(struct ring_buffer_per_cpu
*cpu_buffer
,
2149 unsigned long tail
, struct rb_event_info
*info
)
2151 struct buffer_page
*tail_page
= info
->tail_page
;
2152 struct buffer_page
*commit_page
= cpu_buffer
->commit_page
;
2153 struct ring_buffer
*buffer
= cpu_buffer
->buffer
;
2154 struct buffer_page
*next_page
;
2158 next_page
= tail_page
;
2160 rb_inc_page(cpu_buffer
, &next_page
);
2163 * If for some reason, we had an interrupt storm that made
2164 * it all the way around the buffer, bail, and warn
2167 if (unlikely(next_page
== commit_page
)) {
2168 local_inc(&cpu_buffer
->commit_overrun
);
2173 * This is where the fun begins!
2175 * We are fighting against races between a reader that
2176 * could be on another CPU trying to swap its reader
2177 * page with the buffer head.
2179 * We are also fighting against interrupts coming in and
2180 * moving the head or tail on us as well.
2182 * If the next page is the head page then we have filled
2183 * the buffer, unless the commit page is still on the
2186 if (rb_is_head_page(cpu_buffer
, next_page
, &tail_page
->list
)) {
2189 * If the commit is not on the reader page, then
2190 * move the header page.
2192 if (!rb_is_reader_page(cpu_buffer
->commit_page
)) {
2194 * If we are not in overwrite mode,
2195 * this is easy, just stop here.
2197 if (!(buffer
->flags
& RB_FL_OVERWRITE
)) {
2198 local_inc(&cpu_buffer
->dropped_events
);
2202 ret
= rb_handle_head_page(cpu_buffer
,
2211 * We need to be careful here too. The
2212 * commit page could still be on the reader
2213 * page. We could have a small buffer, and
2214 * have filled up the buffer with events
2215 * from interrupts and such, and wrapped.
2217 * Note, if the tail page is also the on the
2218 * reader_page, we let it move out.
2220 if (unlikely((cpu_buffer
->commit_page
!=
2221 cpu_buffer
->tail_page
) &&
2222 (cpu_buffer
->commit_page
==
2223 cpu_buffer
->reader_page
))) {
2224 local_inc(&cpu_buffer
->commit_overrun
);
2230 ret
= rb_tail_page_update(cpu_buffer
, tail_page
, next_page
);
2233 * Nested commits always have zero deltas, so
2234 * just reread the time stamp
2236 ts
= rb_time_stamp(buffer
);
2237 next_page
->page
->time_stamp
= ts
;
2242 rb_reset_tail(cpu_buffer
, tail
, info
);
2244 /* fail and let the caller try again */
2245 return ERR_PTR(-EAGAIN
);
2249 rb_reset_tail(cpu_buffer
, tail
, info
);
2254 /* Slow path, do not inline */
2255 static noinline
struct ring_buffer_event
*
2256 rb_add_time_stamp(struct ring_buffer_event
*event
, u64 delta
)
2258 event
->type_len
= RINGBUF_TYPE_TIME_EXTEND
;
2260 /* Not the first event on the page? */
2261 if (rb_event_index(event
)) {
2262 event
->time_delta
= delta
& TS_MASK
;
2263 event
->array
[0] = delta
>> TS_SHIFT
;
2265 /* nope, just zero it */
2266 event
->time_delta
= 0;
2267 event
->array
[0] = 0;
2270 return skip_time_extend(event
);
2273 static inline int rb_event_is_commit(struct ring_buffer_per_cpu
*cpu_buffer
,
2274 struct ring_buffer_event
*event
);
2277 * rb_update_event - update event type and data
2278 * @event: the event to update
2279 * @type: the type of event
2280 * @length: the size of the event field in the ring buffer
2282 * Update the type and data fields of the event. The length
2283 * is the actual size that is written to the ring buffer,
2284 * and with this, we can determine what to place into the
2288 rb_update_event(struct ring_buffer_per_cpu
*cpu_buffer
,
2289 struct ring_buffer_event
*event
,
2290 struct rb_event_info
*info
)
2292 unsigned length
= info
->length
;
2293 u64 delta
= info
->delta
;
2295 /* Only a commit updates the timestamp */
2296 if (unlikely(!rb_event_is_commit(cpu_buffer
, event
)))
2300 * If we need to add a timestamp, then we
2301 * add it to the start of the resevered space.
2303 if (unlikely(info
->add_timestamp
)) {
2304 event
= rb_add_time_stamp(event
, delta
);
2305 length
-= RB_LEN_TIME_EXTEND
;
2309 event
->time_delta
= delta
;
2310 length
-= RB_EVNT_HDR_SIZE
;
2311 if (length
> RB_MAX_SMALL_DATA
|| RB_FORCE_8BYTE_ALIGNMENT
) {
2312 event
->type_len
= 0;
2313 event
->array
[0] = length
;
2315 event
->type_len
= DIV_ROUND_UP(length
, RB_ALIGNMENT
);
2318 static unsigned rb_calculate_event_length(unsigned length
)
2320 struct ring_buffer_event event
; /* Used only for sizeof array */
2322 /* zero length can cause confusions */
2326 if (length
> RB_MAX_SMALL_DATA
|| RB_FORCE_8BYTE_ALIGNMENT
)
2327 length
+= sizeof(event
.array
[0]);
2329 length
+= RB_EVNT_HDR_SIZE
;
2330 length
= ALIGN(length
, RB_ARCH_ALIGNMENT
);
2333 * In case the time delta is larger than the 27 bits for it
2334 * in the header, we need to add a timestamp. If another
2335 * event comes in when trying to discard this one to increase
2336 * the length, then the timestamp will be added in the allocated
2337 * space of this event. If length is bigger than the size needed
2338 * for the TIME_EXTEND, then padding has to be used. The events
2339 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2340 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2341 * As length is a multiple of 4, we only need to worry if it
2342 * is 12 (RB_LEN_TIME_EXTEND + 4).
2344 if (length
== RB_LEN_TIME_EXTEND
+ RB_ALIGNMENT
)
2345 length
+= RB_ALIGNMENT
;
2350 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2351 static inline bool sched_clock_stable(void)
2358 rb_try_to_discard(struct ring_buffer_per_cpu
*cpu_buffer
,
2359 struct ring_buffer_event
*event
)
2361 unsigned long new_index
, old_index
;
2362 struct buffer_page
*bpage
;
2363 unsigned long index
;
2366 new_index
= rb_event_index(event
);
2367 old_index
= new_index
+ rb_event_ts_length(event
);
2368 addr
= (unsigned long)event
;
2371 bpage
= cpu_buffer
->tail_page
;
2373 if (bpage
->page
== (void *)addr
&& rb_page_write(bpage
) == old_index
) {
2374 unsigned long write_mask
=
2375 local_read(&bpage
->write
) & ~RB_WRITE_MASK
;
2376 unsigned long event_length
= rb_event_length(event
);
2378 * This is on the tail page. It is possible that
2379 * a write could come in and move the tail page
2380 * and write to the next page. That is fine
2381 * because we just shorten what is on this page.
2383 old_index
+= write_mask
;
2384 new_index
+= write_mask
;
2385 index
= local_cmpxchg(&bpage
->write
, old_index
, new_index
);
2386 if (index
== old_index
) {
2387 /* update counters */
2388 local_sub(event_length
, &cpu_buffer
->entries_bytes
);
2393 /* could not discard */
2397 static void rb_start_commit(struct ring_buffer_per_cpu
*cpu_buffer
)
2399 local_inc(&cpu_buffer
->committing
);
2400 local_inc(&cpu_buffer
->commits
);
2404 rb_set_commit_to_write(struct ring_buffer_per_cpu
*cpu_buffer
)
2406 unsigned long max_count
;
2409 * We only race with interrupts and NMIs on this CPU.
2410 * If we own the commit event, then we can commit
2411 * all others that interrupted us, since the interruptions
2412 * are in stack format (they finish before they come
2413 * back to us). This allows us to do a simple loop to
2414 * assign the commit to the tail.
2417 max_count
= cpu_buffer
->nr_pages
* 100;
2419 while (cpu_buffer
->commit_page
!= cpu_buffer
->tail_page
) {
2420 if (RB_WARN_ON(cpu_buffer
, !(--max_count
)))
2422 if (RB_WARN_ON(cpu_buffer
,
2423 rb_is_reader_page(cpu_buffer
->tail_page
)))
2425 local_set(&cpu_buffer
->commit_page
->page
->commit
,
2426 rb_page_write(cpu_buffer
->commit_page
));
2427 rb_inc_page(cpu_buffer
, &cpu_buffer
->commit_page
);
2428 cpu_buffer
->write_stamp
=
2429 cpu_buffer
->commit_page
->page
->time_stamp
;
2430 /* add barrier to keep gcc from optimizing too much */
2433 while (rb_commit_index(cpu_buffer
) !=
2434 rb_page_write(cpu_buffer
->commit_page
)) {
2436 local_set(&cpu_buffer
->commit_page
->page
->commit
,
2437 rb_page_write(cpu_buffer
->commit_page
));
2438 RB_WARN_ON(cpu_buffer
,
2439 local_read(&cpu_buffer
->commit_page
->page
->commit
) &
2444 /* again, keep gcc from optimizing */
2448 * If an interrupt came in just after the first while loop
2449 * and pushed the tail page forward, we will be left with
2450 * a dangling commit that will never go forward.
2452 if (unlikely(cpu_buffer
->commit_page
!= cpu_buffer
->tail_page
))
2456 static inline void rb_end_commit(struct ring_buffer_per_cpu
*cpu_buffer
)
2458 unsigned long commits
;
2460 if (RB_WARN_ON(cpu_buffer
,
2461 !local_read(&cpu_buffer
->committing
)))
2465 commits
= local_read(&cpu_buffer
->commits
);
2466 /* synchronize with interrupts */
2468 if (local_read(&cpu_buffer
->committing
) == 1)
2469 rb_set_commit_to_write(cpu_buffer
);
2471 local_dec(&cpu_buffer
->committing
);
2473 /* synchronize with interrupts */
2477 * Need to account for interrupts coming in between the
2478 * updating of the commit page and the clearing of the
2479 * committing counter.
2481 if (unlikely(local_read(&cpu_buffer
->commits
) != commits
) &&
2482 !local_read(&cpu_buffer
->committing
)) {
2483 local_inc(&cpu_buffer
->committing
);
2488 static inline void rb_event_discard(struct ring_buffer_event
*event
)
2490 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
2491 event
= skip_time_extend(event
);
2493 /* array[0] holds the actual length for the discarded event */
2494 event
->array
[0] = rb_event_data_length(event
) - RB_EVNT_HDR_SIZE
;
2495 event
->type_len
= RINGBUF_TYPE_PADDING
;
2496 /* time delta must be non zero */
2497 if (!event
->time_delta
)
2498 event
->time_delta
= 1;
2502 rb_event_is_commit(struct ring_buffer_per_cpu
*cpu_buffer
,
2503 struct ring_buffer_event
*event
)
2505 unsigned long addr
= (unsigned long)event
;
2506 unsigned long index
;
2508 index
= rb_event_index(event
);
2511 return cpu_buffer
->commit_page
->page
== (void *)addr
&&
2512 rb_commit_index(cpu_buffer
) == index
;
2516 rb_update_write_stamp(struct ring_buffer_per_cpu
*cpu_buffer
,
2517 struct ring_buffer_event
*event
)
2522 * The event first in the commit queue updates the
2525 if (rb_event_is_commit(cpu_buffer
, event
)) {
2527 * A commit event that is first on a page
2528 * updates the write timestamp with the page stamp
2530 if (!rb_event_index(event
))
2531 cpu_buffer
->write_stamp
=
2532 cpu_buffer
->commit_page
->page
->time_stamp
;
2533 else if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
) {
2534 delta
= event
->array
[0];
2536 delta
+= event
->time_delta
;
2537 cpu_buffer
->write_stamp
+= delta
;
2539 cpu_buffer
->write_stamp
+= event
->time_delta
;
2543 static void rb_commit(struct ring_buffer_per_cpu
*cpu_buffer
,
2544 struct ring_buffer_event
*event
)
2546 local_inc(&cpu_buffer
->entries
);
2547 rb_update_write_stamp(cpu_buffer
, event
);
2548 rb_end_commit(cpu_buffer
);
2551 static __always_inline
void
2552 rb_wakeups(struct ring_buffer
*buffer
, struct ring_buffer_per_cpu
*cpu_buffer
)
2556 if (buffer
->irq_work
.waiters_pending
) {
2557 buffer
->irq_work
.waiters_pending
= false;
2558 /* irq_work_queue() supplies it's own memory barriers */
2559 irq_work_queue(&buffer
->irq_work
.work
);
2562 if (cpu_buffer
->irq_work
.waiters_pending
) {
2563 cpu_buffer
->irq_work
.waiters_pending
= false;
2564 /* irq_work_queue() supplies it's own memory barriers */
2565 irq_work_queue(&cpu_buffer
->irq_work
.work
);
2568 pagebusy
= cpu_buffer
->reader_page
== cpu_buffer
->commit_page
;
2570 if (!pagebusy
&& cpu_buffer
->irq_work
.full_waiters_pending
) {
2571 cpu_buffer
->irq_work
.wakeup_full
= true;
2572 cpu_buffer
->irq_work
.full_waiters_pending
= false;
2573 /* irq_work_queue() supplies it's own memory barriers */
2574 irq_work_queue(&cpu_buffer
->irq_work
.work
);
2579 * The lock and unlock are done within a preempt disable section.
2580 * The current_context per_cpu variable can only be modified
2581 * by the current task between lock and unlock. But it can
2582 * be modified more than once via an interrupt. To pass this
2583 * information from the lock to the unlock without having to
2584 * access the 'in_interrupt()' functions again (which do show
2585 * a bit of overhead in something as critical as function tracing,
2586 * we use a bitmask trick.
2588 * bit 0 = NMI context
2589 * bit 1 = IRQ context
2590 * bit 2 = SoftIRQ context
2591 * bit 3 = normal context.
2593 * This works because this is the order of contexts that can
2594 * preempt other contexts. A SoftIRQ never preempts an IRQ
2597 * When the context is determined, the corresponding bit is
2598 * checked and set (if it was set, then a recursion of that context
2601 * On unlock, we need to clear this bit. To do so, just subtract
2602 * 1 from the current_context and AND it to itself.
2606 * 101 & 100 = 100 (clearing bit zero)
2609 * 1010 & 1001 = 1000 (clearing bit 1)
2611 * The least significant bit can be cleared this way, and it
2612 * just so happens that it is the same bit corresponding to
2613 * the current context.
2616 static __always_inline
int
2617 trace_recursive_lock(struct ring_buffer_per_cpu
*cpu_buffer
)
2619 unsigned int val
= cpu_buffer
->current_context
;
2622 if (in_interrupt()) {
2628 bit
= RB_CTX_SOFTIRQ
;
2630 bit
= RB_CTX_NORMAL
;
2632 if (unlikely(val
& (1 << bit
)))
2636 cpu_buffer
->current_context
= val
;
2641 static __always_inline
void
2642 trace_recursive_unlock(struct ring_buffer_per_cpu
*cpu_buffer
)
2644 cpu_buffer
->current_context
&= cpu_buffer
->current_context
- 1;
2648 * ring_buffer_unlock_commit - commit a reserved
2649 * @buffer: The buffer to commit to
2650 * @event: The event pointer to commit.
2652 * This commits the data to the ring buffer, and releases any locks held.
2654 * Must be paired with ring_buffer_lock_reserve.
2656 int ring_buffer_unlock_commit(struct ring_buffer
*buffer
,
2657 struct ring_buffer_event
*event
)
2659 struct ring_buffer_per_cpu
*cpu_buffer
;
2660 int cpu
= raw_smp_processor_id();
2662 cpu_buffer
= buffer
->buffers
[cpu
];
2664 rb_commit(cpu_buffer
, event
);
2666 rb_wakeups(buffer
, cpu_buffer
);
2668 trace_recursive_unlock(cpu_buffer
);
2670 preempt_enable_notrace();
2674 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit
);
2676 static noinline
void
2677 rb_handle_timestamp(struct ring_buffer_per_cpu
*cpu_buffer
,
2678 struct rb_event_info
*info
)
2680 WARN_ONCE(info
->delta
> (1ULL << 59),
2681 KERN_WARNING
"Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2682 (unsigned long long)info
->delta
,
2683 (unsigned long long)info
->ts
,
2684 (unsigned long long)cpu_buffer
->write_stamp
,
2685 sched_clock_stable() ? "" :
2686 "If you just came from a suspend/resume,\n"
2687 "please switch to the trace global clock:\n"
2688 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2689 info
->add_timestamp
= 1;
2692 static struct ring_buffer_event
*
2693 __rb_reserve_next(struct ring_buffer_per_cpu
*cpu_buffer
,
2694 struct rb_event_info
*info
)
2696 struct ring_buffer_event
*event
;
2697 struct buffer_page
*tail_page
;
2698 unsigned long tail
, write
;
2701 * If the time delta since the last event is too big to
2702 * hold in the time field of the event, then we append a
2703 * TIME EXTEND event ahead of the data event.
2705 if (unlikely(info
->add_timestamp
))
2706 info
->length
+= RB_LEN_TIME_EXTEND
;
2708 tail_page
= info
->tail_page
= cpu_buffer
->tail_page
;
2709 write
= local_add_return(info
->length
, &tail_page
->write
);
2711 /* set write to only the index of the write */
2712 write
&= RB_WRITE_MASK
;
2713 tail
= write
- info
->length
;
2716 * If this is the first commit on the page, then it has the same
2717 * timestamp as the page itself.
2722 /* See if we shot pass the end of this buffer page */
2723 if (unlikely(write
> BUF_PAGE_SIZE
))
2724 return rb_move_tail(cpu_buffer
, tail
, info
);
2726 /* We reserved something on the buffer */
2728 event
= __rb_page_index(tail_page
, tail
);
2729 kmemcheck_annotate_bitfield(event
, bitfield
);
2730 rb_update_event(cpu_buffer
, event
, info
);
2732 local_inc(&tail_page
->entries
);
2735 * If this is the first commit on the page, then update
2739 tail_page
->page
->time_stamp
= info
->ts
;
2741 /* account for these added bytes */
2742 local_add(info
->length
, &cpu_buffer
->entries_bytes
);
2747 static struct ring_buffer_event
*
2748 rb_reserve_next_event(struct ring_buffer
*buffer
,
2749 struct ring_buffer_per_cpu
*cpu_buffer
,
2750 unsigned long length
)
2752 struct ring_buffer_event
*event
;
2753 struct rb_event_info info
;
2757 rb_start_commit(cpu_buffer
);
2759 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2761 * Due to the ability to swap a cpu buffer from a buffer
2762 * it is possible it was swapped before we committed.
2763 * (committing stops a swap). We check for it here and
2764 * if it happened, we have to fail the write.
2767 if (unlikely(ACCESS_ONCE(cpu_buffer
->buffer
) != buffer
)) {
2768 local_dec(&cpu_buffer
->committing
);
2769 local_dec(&cpu_buffer
->commits
);
2774 info
.length
= rb_calculate_event_length(length
);
2776 info
.add_timestamp
= 0;
2780 * We allow for interrupts to reenter here and do a trace.
2781 * If one does, it will cause this original code to loop
2782 * back here. Even with heavy interrupts happening, this
2783 * should only happen a few times in a row. If this happens
2784 * 1000 times in a row, there must be either an interrupt
2785 * storm or we have something buggy.
2788 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 1000))
2791 info
.ts
= rb_time_stamp(cpu_buffer
->buffer
);
2792 diff
= info
.ts
- cpu_buffer
->write_stamp
;
2794 /* make sure this diff is calculated here */
2797 /* Did the write stamp get updated already? */
2798 if (likely(info
.ts
>= cpu_buffer
->write_stamp
)) {
2800 if (unlikely(test_time_stamp(info
.delta
)))
2801 rb_handle_timestamp(cpu_buffer
, &info
);
2804 event
= __rb_reserve_next(cpu_buffer
, &info
);
2806 if (unlikely(PTR_ERR(event
) == -EAGAIN
))
2815 rb_end_commit(cpu_buffer
);
2820 * ring_buffer_lock_reserve - reserve a part of the buffer
2821 * @buffer: the ring buffer to reserve from
2822 * @length: the length of the data to reserve (excluding event header)
2824 * Returns a reseverd event on the ring buffer to copy directly to.
2825 * The user of this interface will need to get the body to write into
2826 * and can use the ring_buffer_event_data() interface.
2828 * The length is the length of the data needed, not the event length
2829 * which also includes the event header.
2831 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2832 * If NULL is returned, then nothing has been allocated or locked.
2834 struct ring_buffer_event
*
2835 ring_buffer_lock_reserve(struct ring_buffer
*buffer
, unsigned long length
)
2837 struct ring_buffer_per_cpu
*cpu_buffer
;
2838 struct ring_buffer_event
*event
;
2841 /* If we are tracing schedule, we don't want to recurse */
2842 preempt_disable_notrace();
2844 if (unlikely(atomic_read(&buffer
->record_disabled
)))
2847 cpu
= raw_smp_processor_id();
2849 if (unlikely(!cpumask_test_cpu(cpu
, buffer
->cpumask
)))
2852 cpu_buffer
= buffer
->buffers
[cpu
];
2854 if (unlikely(atomic_read(&cpu_buffer
->record_disabled
)))
2857 if (unlikely(length
> BUF_MAX_DATA_SIZE
))
2860 if (unlikely(trace_recursive_lock(cpu_buffer
)))
2863 event
= rb_reserve_next_event(buffer
, cpu_buffer
, length
);
2870 trace_recursive_unlock(cpu_buffer
);
2872 preempt_enable_notrace();
2875 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve
);
2878 * Decrement the entries to the page that an event is on.
2879 * The event does not even need to exist, only the pointer
2880 * to the page it is on. This may only be called before the commit
2884 rb_decrement_entry(struct ring_buffer_per_cpu
*cpu_buffer
,
2885 struct ring_buffer_event
*event
)
2887 unsigned long addr
= (unsigned long)event
;
2888 struct buffer_page
*bpage
= cpu_buffer
->commit_page
;
2889 struct buffer_page
*start
;
2893 /* Do the likely case first */
2894 if (likely(bpage
->page
== (void *)addr
)) {
2895 local_dec(&bpage
->entries
);
2900 * Because the commit page may be on the reader page we
2901 * start with the next page and check the end loop there.
2903 rb_inc_page(cpu_buffer
, &bpage
);
2906 if (bpage
->page
== (void *)addr
) {
2907 local_dec(&bpage
->entries
);
2910 rb_inc_page(cpu_buffer
, &bpage
);
2911 } while (bpage
!= start
);
2913 /* commit not part of this buffer?? */
2914 RB_WARN_ON(cpu_buffer
, 1);
2918 * ring_buffer_commit_discard - discard an event that has not been committed
2919 * @buffer: the ring buffer
2920 * @event: non committed event to discard
2922 * Sometimes an event that is in the ring buffer needs to be ignored.
2923 * This function lets the user discard an event in the ring buffer
2924 * and then that event will not be read later.
2926 * This function only works if it is called before the the item has been
2927 * committed. It will try to free the event from the ring buffer
2928 * if another event has not been added behind it.
2930 * If another event has been added behind it, it will set the event
2931 * up as discarded, and perform the commit.
2933 * If this function is called, do not call ring_buffer_unlock_commit on
2936 void ring_buffer_discard_commit(struct ring_buffer
*buffer
,
2937 struct ring_buffer_event
*event
)
2939 struct ring_buffer_per_cpu
*cpu_buffer
;
2942 /* The event is discarded regardless */
2943 rb_event_discard(event
);
2945 cpu
= smp_processor_id();
2946 cpu_buffer
= buffer
->buffers
[cpu
];
2949 * This must only be called if the event has not been
2950 * committed yet. Thus we can assume that preemption
2951 * is still disabled.
2953 RB_WARN_ON(buffer
, !local_read(&cpu_buffer
->committing
));
2955 rb_decrement_entry(cpu_buffer
, event
);
2956 if (rb_try_to_discard(cpu_buffer
, event
))
2960 * The commit is still visible by the reader, so we
2961 * must still update the timestamp.
2963 rb_update_write_stamp(cpu_buffer
, event
);
2965 rb_end_commit(cpu_buffer
);
2967 trace_recursive_unlock(cpu_buffer
);
2969 preempt_enable_notrace();
2972 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit
);
2975 * ring_buffer_write - write data to the buffer without reserving
2976 * @buffer: The ring buffer to write to.
2977 * @length: The length of the data being written (excluding the event header)
2978 * @data: The data to write to the buffer.
2980 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2981 * one function. If you already have the data to write to the buffer, it
2982 * may be easier to simply call this function.
2984 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2985 * and not the length of the event which would hold the header.
2987 int ring_buffer_write(struct ring_buffer
*buffer
,
2988 unsigned long length
,
2991 struct ring_buffer_per_cpu
*cpu_buffer
;
2992 struct ring_buffer_event
*event
;
2997 preempt_disable_notrace();
2999 if (atomic_read(&buffer
->record_disabled
))
3002 cpu
= raw_smp_processor_id();
3004 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3007 cpu_buffer
= buffer
->buffers
[cpu
];
3009 if (atomic_read(&cpu_buffer
->record_disabled
))
3012 if (length
> BUF_MAX_DATA_SIZE
)
3015 if (unlikely(trace_recursive_lock(cpu_buffer
)))
3018 event
= rb_reserve_next_event(buffer
, cpu_buffer
, length
);
3022 body
= rb_event_data(event
);
3024 memcpy(body
, data
, length
);
3026 rb_commit(cpu_buffer
, event
);
3028 rb_wakeups(buffer
, cpu_buffer
);
3033 trace_recursive_unlock(cpu_buffer
);
3036 preempt_enable_notrace();
3040 EXPORT_SYMBOL_GPL(ring_buffer_write
);
3042 static int rb_per_cpu_empty(struct ring_buffer_per_cpu
*cpu_buffer
)
3044 struct buffer_page
*reader
= cpu_buffer
->reader_page
;
3045 struct buffer_page
*head
= rb_set_head_page(cpu_buffer
);
3046 struct buffer_page
*commit
= cpu_buffer
->commit_page
;
3048 /* In case of error, head will be NULL */
3049 if (unlikely(!head
))
3052 return reader
->read
== rb_page_commit(reader
) &&
3053 (commit
== reader
||
3055 head
->read
== rb_page_commit(commit
)));
3059 * ring_buffer_record_disable - stop all writes into the buffer
3060 * @buffer: The ring buffer to stop writes to.
3062 * This prevents all writes to the buffer. Any attempt to write
3063 * to the buffer after this will fail and return NULL.
3065 * The caller should call synchronize_sched() after this.
3067 void ring_buffer_record_disable(struct ring_buffer
*buffer
)
3069 atomic_inc(&buffer
->record_disabled
);
3071 EXPORT_SYMBOL_GPL(ring_buffer_record_disable
);
3074 * ring_buffer_record_enable - enable writes to the buffer
3075 * @buffer: The ring buffer to enable writes
3077 * Note, multiple disables will need the same number of enables
3078 * to truly enable the writing (much like preempt_disable).
3080 void ring_buffer_record_enable(struct ring_buffer
*buffer
)
3082 atomic_dec(&buffer
->record_disabled
);
3084 EXPORT_SYMBOL_GPL(ring_buffer_record_enable
);
3087 * ring_buffer_record_off - stop all writes into the buffer
3088 * @buffer: The ring buffer to stop writes to.
3090 * This prevents all writes to the buffer. Any attempt to write
3091 * to the buffer after this will fail and return NULL.
3093 * This is different than ring_buffer_record_disable() as
3094 * it works like an on/off switch, where as the disable() version
3095 * must be paired with a enable().
3097 void ring_buffer_record_off(struct ring_buffer
*buffer
)
3100 unsigned int new_rd
;
3103 rd
= atomic_read(&buffer
->record_disabled
);
3104 new_rd
= rd
| RB_BUFFER_OFF
;
3105 } while (atomic_cmpxchg(&buffer
->record_disabled
, rd
, new_rd
) != rd
);
3107 EXPORT_SYMBOL_GPL(ring_buffer_record_off
);
3110 * ring_buffer_record_on - restart writes into the buffer
3111 * @buffer: The ring buffer to start writes to.
3113 * This enables all writes to the buffer that was disabled by
3114 * ring_buffer_record_off().
3116 * This is different than ring_buffer_record_enable() as
3117 * it works like an on/off switch, where as the enable() version
3118 * must be paired with a disable().
3120 void ring_buffer_record_on(struct ring_buffer
*buffer
)
3123 unsigned int new_rd
;
3126 rd
= atomic_read(&buffer
->record_disabled
);
3127 new_rd
= rd
& ~RB_BUFFER_OFF
;
3128 } while (atomic_cmpxchg(&buffer
->record_disabled
, rd
, new_rd
) != rd
);
3130 EXPORT_SYMBOL_GPL(ring_buffer_record_on
);
3133 * ring_buffer_record_is_on - return true if the ring buffer can write
3134 * @buffer: The ring buffer to see if write is enabled
3136 * Returns true if the ring buffer is in a state that it accepts writes.
3138 int ring_buffer_record_is_on(struct ring_buffer
*buffer
)
3140 return !atomic_read(&buffer
->record_disabled
);
3144 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3145 * @buffer: The ring buffer to stop writes to.
3146 * @cpu: The CPU buffer to stop
3148 * This prevents all writes to the buffer. Any attempt to write
3149 * to the buffer after this will fail and return NULL.
3151 * The caller should call synchronize_sched() after this.
3153 void ring_buffer_record_disable_cpu(struct ring_buffer
*buffer
, int cpu
)
3155 struct ring_buffer_per_cpu
*cpu_buffer
;
3157 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3160 cpu_buffer
= buffer
->buffers
[cpu
];
3161 atomic_inc(&cpu_buffer
->record_disabled
);
3163 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu
);
3166 * ring_buffer_record_enable_cpu - enable writes to the buffer
3167 * @buffer: The ring buffer to enable writes
3168 * @cpu: The CPU to enable.
3170 * Note, multiple disables will need the same number of enables
3171 * to truly enable the writing (much like preempt_disable).
3173 void ring_buffer_record_enable_cpu(struct ring_buffer
*buffer
, int cpu
)
3175 struct ring_buffer_per_cpu
*cpu_buffer
;
3177 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3180 cpu_buffer
= buffer
->buffers
[cpu
];
3181 atomic_dec(&cpu_buffer
->record_disabled
);
3183 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu
);
3186 * The total entries in the ring buffer is the running counter
3187 * of entries entered into the ring buffer, minus the sum of
3188 * the entries read from the ring buffer and the number of
3189 * entries that were overwritten.
3191 static inline unsigned long
3192 rb_num_of_entries(struct ring_buffer_per_cpu
*cpu_buffer
)
3194 return local_read(&cpu_buffer
->entries
) -
3195 (local_read(&cpu_buffer
->overrun
) + cpu_buffer
->read
);
3199 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3200 * @buffer: The ring buffer
3201 * @cpu: The per CPU buffer to read from.
3203 u64
ring_buffer_oldest_event_ts(struct ring_buffer
*buffer
, int cpu
)
3205 unsigned long flags
;
3206 struct ring_buffer_per_cpu
*cpu_buffer
;
3207 struct buffer_page
*bpage
;
3210 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3213 cpu_buffer
= buffer
->buffers
[cpu
];
3214 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3216 * if the tail is on reader_page, oldest time stamp is on the reader
3219 if (cpu_buffer
->tail_page
== cpu_buffer
->reader_page
)
3220 bpage
= cpu_buffer
->reader_page
;
3222 bpage
= rb_set_head_page(cpu_buffer
);
3224 ret
= bpage
->page
->time_stamp
;
3225 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3229 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts
);
3232 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3233 * @buffer: The ring buffer
3234 * @cpu: The per CPU buffer to read from.
3236 unsigned long ring_buffer_bytes_cpu(struct ring_buffer
*buffer
, int cpu
)
3238 struct ring_buffer_per_cpu
*cpu_buffer
;
3241 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3244 cpu_buffer
= buffer
->buffers
[cpu
];
3245 ret
= local_read(&cpu_buffer
->entries_bytes
) - cpu_buffer
->read_bytes
;
3249 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu
);
3252 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3253 * @buffer: The ring buffer
3254 * @cpu: The per CPU buffer to get the entries from.
3256 unsigned long ring_buffer_entries_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
];
3265 return rb_num_of_entries(cpu_buffer
);
3267 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu
);
3270 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3271 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3272 * @buffer: The ring buffer
3273 * @cpu: The per CPU buffer to get the number of overruns from
3275 unsigned long ring_buffer_overrun_cpu(struct ring_buffer
*buffer
, int cpu
)
3277 struct ring_buffer_per_cpu
*cpu_buffer
;
3280 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3283 cpu_buffer
= buffer
->buffers
[cpu
];
3284 ret
= local_read(&cpu_buffer
->overrun
);
3288 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu
);
3291 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3292 * commits failing due to the buffer wrapping around while there are uncommitted
3293 * events, such as during an interrupt storm.
3294 * @buffer: The ring buffer
3295 * @cpu: The per CPU buffer to get the number of overruns from
3298 ring_buffer_commit_overrun_cpu(struct ring_buffer
*buffer
, int cpu
)
3300 struct ring_buffer_per_cpu
*cpu_buffer
;
3303 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3306 cpu_buffer
= buffer
->buffers
[cpu
];
3307 ret
= local_read(&cpu_buffer
->commit_overrun
);
3311 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu
);
3314 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3315 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3316 * @buffer: The ring buffer
3317 * @cpu: The per CPU buffer to get the number of overruns from
3320 ring_buffer_dropped_events_cpu(struct ring_buffer
*buffer
, int cpu
)
3322 struct ring_buffer_per_cpu
*cpu_buffer
;
3325 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3328 cpu_buffer
= buffer
->buffers
[cpu
];
3329 ret
= local_read(&cpu_buffer
->dropped_events
);
3333 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu
);
3336 * ring_buffer_read_events_cpu - get the number of events successfully read
3337 * @buffer: The ring buffer
3338 * @cpu: The per CPU buffer to get the number of events read
3341 ring_buffer_read_events_cpu(struct ring_buffer
*buffer
, int cpu
)
3343 struct ring_buffer_per_cpu
*cpu_buffer
;
3345 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3348 cpu_buffer
= buffer
->buffers
[cpu
];
3349 return cpu_buffer
->read
;
3351 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu
);
3354 * ring_buffer_entries - get the number of entries in a buffer
3355 * @buffer: The ring buffer
3357 * Returns the total number of entries in the ring buffer
3360 unsigned long ring_buffer_entries(struct ring_buffer
*buffer
)
3362 struct ring_buffer_per_cpu
*cpu_buffer
;
3363 unsigned long entries
= 0;
3366 /* if you care about this being correct, lock the buffer */
3367 for_each_buffer_cpu(buffer
, cpu
) {
3368 cpu_buffer
= buffer
->buffers
[cpu
];
3369 entries
+= rb_num_of_entries(cpu_buffer
);
3374 EXPORT_SYMBOL_GPL(ring_buffer_entries
);
3377 * ring_buffer_overruns - get the number of overruns in buffer
3378 * @buffer: The ring buffer
3380 * Returns the total number of overruns in the ring buffer
3383 unsigned long ring_buffer_overruns(struct ring_buffer
*buffer
)
3385 struct ring_buffer_per_cpu
*cpu_buffer
;
3386 unsigned long overruns
= 0;
3389 /* if you care about this being correct, lock the buffer */
3390 for_each_buffer_cpu(buffer
, cpu
) {
3391 cpu_buffer
= buffer
->buffers
[cpu
];
3392 overruns
+= local_read(&cpu_buffer
->overrun
);
3397 EXPORT_SYMBOL_GPL(ring_buffer_overruns
);
3399 static void rb_iter_reset(struct ring_buffer_iter
*iter
)
3401 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
3403 /* Iterator usage is expected to have record disabled */
3404 iter
->head_page
= cpu_buffer
->reader_page
;
3405 iter
->head
= cpu_buffer
->reader_page
->read
;
3407 iter
->cache_reader_page
= iter
->head_page
;
3408 iter
->cache_read
= cpu_buffer
->read
;
3411 iter
->read_stamp
= cpu_buffer
->read_stamp
;
3413 iter
->read_stamp
= iter
->head_page
->page
->time_stamp
;
3417 * ring_buffer_iter_reset - reset an iterator
3418 * @iter: The iterator to reset
3420 * Resets the iterator, so that it will start from the beginning
3423 void ring_buffer_iter_reset(struct ring_buffer_iter
*iter
)
3425 struct ring_buffer_per_cpu
*cpu_buffer
;
3426 unsigned long flags
;
3431 cpu_buffer
= iter
->cpu_buffer
;
3433 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3434 rb_iter_reset(iter
);
3435 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3437 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset
);
3440 * ring_buffer_iter_empty - check if an iterator has no more to read
3441 * @iter: The iterator to check
3443 int ring_buffer_iter_empty(struct ring_buffer_iter
*iter
)
3445 struct ring_buffer_per_cpu
*cpu_buffer
;
3447 cpu_buffer
= iter
->cpu_buffer
;
3449 return iter
->head_page
== cpu_buffer
->commit_page
&&
3450 iter
->head
== rb_commit_index(cpu_buffer
);
3452 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty
);
3455 rb_update_read_stamp(struct ring_buffer_per_cpu
*cpu_buffer
,
3456 struct ring_buffer_event
*event
)
3460 switch (event
->type_len
) {
3461 case RINGBUF_TYPE_PADDING
:
3464 case RINGBUF_TYPE_TIME_EXTEND
:
3465 delta
= event
->array
[0];
3467 delta
+= event
->time_delta
;
3468 cpu_buffer
->read_stamp
+= delta
;
3471 case RINGBUF_TYPE_TIME_STAMP
:
3472 /* FIXME: not implemented */
3475 case RINGBUF_TYPE_DATA
:
3476 cpu_buffer
->read_stamp
+= event
->time_delta
;
3486 rb_update_iter_read_stamp(struct ring_buffer_iter
*iter
,
3487 struct ring_buffer_event
*event
)
3491 switch (event
->type_len
) {
3492 case RINGBUF_TYPE_PADDING
:
3495 case RINGBUF_TYPE_TIME_EXTEND
:
3496 delta
= event
->array
[0];
3498 delta
+= event
->time_delta
;
3499 iter
->read_stamp
+= delta
;
3502 case RINGBUF_TYPE_TIME_STAMP
:
3503 /* FIXME: not implemented */
3506 case RINGBUF_TYPE_DATA
:
3507 iter
->read_stamp
+= event
->time_delta
;
3516 static struct buffer_page
*
3517 rb_get_reader_page(struct ring_buffer_per_cpu
*cpu_buffer
)
3519 struct buffer_page
*reader
= NULL
;
3520 unsigned long overwrite
;
3521 unsigned long flags
;
3525 local_irq_save(flags
);
3526 arch_spin_lock(&cpu_buffer
->lock
);
3530 * This should normally only loop twice. But because the
3531 * start of the reader inserts an empty page, it causes
3532 * a case where we will loop three times. There should be no
3533 * reason to loop four times (that I know of).
3535 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 3)) {
3540 reader
= cpu_buffer
->reader_page
;
3542 /* If there's more to read, return this page */
3543 if (cpu_buffer
->reader_page
->read
< rb_page_size(reader
))
3546 /* Never should we have an index greater than the size */
3547 if (RB_WARN_ON(cpu_buffer
,
3548 cpu_buffer
->reader_page
->read
> rb_page_size(reader
)))
3551 /* check if we caught up to the tail */
3553 if (cpu_buffer
->commit_page
== cpu_buffer
->reader_page
)
3556 /* Don't bother swapping if the ring buffer is empty */
3557 if (rb_num_of_entries(cpu_buffer
) == 0)
3561 * Reset the reader page to size zero.
3563 local_set(&cpu_buffer
->reader_page
->write
, 0);
3564 local_set(&cpu_buffer
->reader_page
->entries
, 0);
3565 local_set(&cpu_buffer
->reader_page
->page
->commit
, 0);
3566 cpu_buffer
->reader_page
->real_end
= 0;
3570 * Splice the empty reader page into the list around the head.
3572 reader
= rb_set_head_page(cpu_buffer
);
3575 cpu_buffer
->reader_page
->list
.next
= rb_list_head(reader
->list
.next
);
3576 cpu_buffer
->reader_page
->list
.prev
= reader
->list
.prev
;
3579 * cpu_buffer->pages just needs to point to the buffer, it
3580 * has no specific buffer page to point to. Lets move it out
3581 * of our way so we don't accidentally swap it.
3583 cpu_buffer
->pages
= reader
->list
.prev
;
3585 /* The reader page will be pointing to the new head */
3586 rb_set_list_to_head(cpu_buffer
, &cpu_buffer
->reader_page
->list
);
3589 * We want to make sure we read the overruns after we set up our
3590 * pointers to the next object. The writer side does a
3591 * cmpxchg to cross pages which acts as the mb on the writer
3592 * side. Note, the reader will constantly fail the swap
3593 * while the writer is updating the pointers, so this
3594 * guarantees that the overwrite recorded here is the one we
3595 * want to compare with the last_overrun.
3598 overwrite
= local_read(&(cpu_buffer
->overrun
));
3601 * Here's the tricky part.
3603 * We need to move the pointer past the header page.
3604 * But we can only do that if a writer is not currently
3605 * moving it. The page before the header page has the
3606 * flag bit '1' set if it is pointing to the page we want.
3607 * but if the writer is in the process of moving it
3608 * than it will be '2' or already moved '0'.
3611 ret
= rb_head_page_replace(reader
, cpu_buffer
->reader_page
);
3614 * If we did not convert it, then we must try again.
3620 * Yeah! We succeeded in replacing the page.
3622 * Now make the new head point back to the reader page.
3624 rb_list_head(reader
->list
.next
)->prev
= &cpu_buffer
->reader_page
->list
;
3625 rb_inc_page(cpu_buffer
, &cpu_buffer
->head_page
);
3627 /* Finally update the reader page to the new head */
3628 cpu_buffer
->reader_page
= reader
;
3629 rb_reset_reader_page(cpu_buffer
);
3631 if (overwrite
!= cpu_buffer
->last_overrun
) {
3632 cpu_buffer
->lost_events
= overwrite
- cpu_buffer
->last_overrun
;
3633 cpu_buffer
->last_overrun
= overwrite
;
3639 arch_spin_unlock(&cpu_buffer
->lock
);
3640 local_irq_restore(flags
);
3645 static void rb_advance_reader(struct ring_buffer_per_cpu
*cpu_buffer
)
3647 struct ring_buffer_event
*event
;
3648 struct buffer_page
*reader
;
3651 reader
= rb_get_reader_page(cpu_buffer
);
3653 /* This function should not be called when buffer is empty */
3654 if (RB_WARN_ON(cpu_buffer
, !reader
))
3657 event
= rb_reader_event(cpu_buffer
);
3659 if (event
->type_len
<= RINGBUF_TYPE_DATA_TYPE_LEN_MAX
)
3662 rb_update_read_stamp(cpu_buffer
, event
);
3664 length
= rb_event_length(event
);
3665 cpu_buffer
->reader_page
->read
+= length
;
3668 static void rb_advance_iter(struct ring_buffer_iter
*iter
)
3670 struct ring_buffer_per_cpu
*cpu_buffer
;
3671 struct ring_buffer_event
*event
;
3674 cpu_buffer
= iter
->cpu_buffer
;
3677 * Check if we are at the end of the buffer.
3679 if (iter
->head
>= rb_page_size(iter
->head_page
)) {
3680 /* discarded commits can make the page empty */
3681 if (iter
->head_page
== cpu_buffer
->commit_page
)
3687 event
= rb_iter_head_event(iter
);
3689 length
= rb_event_length(event
);
3692 * This should not be called to advance the header if we are
3693 * at the tail of the buffer.
3695 if (RB_WARN_ON(cpu_buffer
,
3696 (iter
->head_page
== cpu_buffer
->commit_page
) &&
3697 (iter
->head
+ length
> rb_commit_index(cpu_buffer
))))
3700 rb_update_iter_read_stamp(iter
, event
);
3702 iter
->head
+= length
;
3704 /* check for end of page padding */
3705 if ((iter
->head
>= rb_page_size(iter
->head_page
)) &&
3706 (iter
->head_page
!= cpu_buffer
->commit_page
))
3710 static int rb_lost_events(struct ring_buffer_per_cpu
*cpu_buffer
)
3712 return cpu_buffer
->lost_events
;
3715 static struct ring_buffer_event
*
3716 rb_buffer_peek(struct ring_buffer_per_cpu
*cpu_buffer
, u64
*ts
,
3717 unsigned long *lost_events
)
3719 struct ring_buffer_event
*event
;
3720 struct buffer_page
*reader
;
3725 * We repeat when a time extend is encountered.
3726 * Since the time extend is always attached to a data event,
3727 * we should never loop more than once.
3728 * (We never hit the following condition more than twice).
3730 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 2))
3733 reader
= rb_get_reader_page(cpu_buffer
);
3737 event
= rb_reader_event(cpu_buffer
);
3739 switch (event
->type_len
) {
3740 case RINGBUF_TYPE_PADDING
:
3741 if (rb_null_event(event
))
3742 RB_WARN_ON(cpu_buffer
, 1);
3744 * Because the writer could be discarding every
3745 * event it creates (which would probably be bad)
3746 * if we were to go back to "again" then we may never
3747 * catch up, and will trigger the warn on, or lock
3748 * the box. Return the padding, and we will release
3749 * the current locks, and try again.
3753 case RINGBUF_TYPE_TIME_EXTEND
:
3754 /* Internal data, OK to advance */
3755 rb_advance_reader(cpu_buffer
);
3758 case RINGBUF_TYPE_TIME_STAMP
:
3759 /* FIXME: not implemented */
3760 rb_advance_reader(cpu_buffer
);
3763 case RINGBUF_TYPE_DATA
:
3765 *ts
= cpu_buffer
->read_stamp
+ event
->time_delta
;
3766 ring_buffer_normalize_time_stamp(cpu_buffer
->buffer
,
3767 cpu_buffer
->cpu
, ts
);
3770 *lost_events
= rb_lost_events(cpu_buffer
);
3779 EXPORT_SYMBOL_GPL(ring_buffer_peek
);
3781 static struct ring_buffer_event
*
3782 rb_iter_peek(struct ring_buffer_iter
*iter
, u64
*ts
)
3784 struct ring_buffer
*buffer
;
3785 struct ring_buffer_per_cpu
*cpu_buffer
;
3786 struct ring_buffer_event
*event
;
3789 cpu_buffer
= iter
->cpu_buffer
;
3790 buffer
= cpu_buffer
->buffer
;
3793 * Check if someone performed a consuming read to
3794 * the buffer. A consuming read invalidates the iterator
3795 * and we need to reset the iterator in this case.
3797 if (unlikely(iter
->cache_read
!= cpu_buffer
->read
||
3798 iter
->cache_reader_page
!= cpu_buffer
->reader_page
))
3799 rb_iter_reset(iter
);
3802 if (ring_buffer_iter_empty(iter
))
3806 * We repeat when a time extend is encountered or we hit
3807 * the end of the page. Since the time extend is always attached
3808 * to a data event, we should never loop more than three times.
3809 * Once for going to next page, once on time extend, and
3810 * finally once to get the event.
3811 * (We never hit the following condition more than thrice).
3813 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 3))
3816 if (rb_per_cpu_empty(cpu_buffer
))
3819 if (iter
->head
>= rb_page_size(iter
->head_page
)) {
3824 event
= rb_iter_head_event(iter
);
3826 switch (event
->type_len
) {
3827 case RINGBUF_TYPE_PADDING
:
3828 if (rb_null_event(event
)) {
3832 rb_advance_iter(iter
);
3835 case RINGBUF_TYPE_TIME_EXTEND
:
3836 /* Internal data, OK to advance */
3837 rb_advance_iter(iter
);
3840 case RINGBUF_TYPE_TIME_STAMP
:
3841 /* FIXME: not implemented */
3842 rb_advance_iter(iter
);
3845 case RINGBUF_TYPE_DATA
:
3847 *ts
= iter
->read_stamp
+ event
->time_delta
;
3848 ring_buffer_normalize_time_stamp(buffer
,
3849 cpu_buffer
->cpu
, ts
);
3859 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek
);
3861 static inline bool rb_reader_lock(struct ring_buffer_per_cpu
*cpu_buffer
)
3863 if (likely(!in_nmi())) {
3864 raw_spin_lock(&cpu_buffer
->reader_lock
);
3869 * If an NMI die dumps out the content of the ring buffer
3870 * trylock must be used to prevent a deadlock if the NMI
3871 * preempted a task that holds the ring buffer locks. If
3872 * we get the lock then all is fine, if not, then continue
3873 * to do the read, but this can corrupt the ring buffer,
3874 * so it must be permanently disabled from future writes.
3875 * Reading from NMI is a oneshot deal.
3877 if (raw_spin_trylock(&cpu_buffer
->reader_lock
))
3880 /* Continue without locking, but disable the ring buffer */
3881 atomic_inc(&cpu_buffer
->record_disabled
);
3886 rb_reader_unlock(struct ring_buffer_per_cpu
*cpu_buffer
, bool locked
)
3889 raw_spin_unlock(&cpu_buffer
->reader_lock
);
3894 * ring_buffer_peek - peek at the next event to be read
3895 * @buffer: The ring buffer to read
3896 * @cpu: The cpu to peak at
3897 * @ts: The timestamp counter of this event.
3898 * @lost_events: a variable to store if events were lost (may be NULL)
3900 * This will return the event that will be read next, but does
3901 * not consume the data.
3903 struct ring_buffer_event
*
3904 ring_buffer_peek(struct ring_buffer
*buffer
, int cpu
, u64
*ts
,
3905 unsigned long *lost_events
)
3907 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
3908 struct ring_buffer_event
*event
;
3909 unsigned long flags
;
3912 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3916 local_irq_save(flags
);
3917 dolock
= rb_reader_lock(cpu_buffer
);
3918 event
= rb_buffer_peek(cpu_buffer
, ts
, lost_events
);
3919 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3920 rb_advance_reader(cpu_buffer
);
3921 rb_reader_unlock(cpu_buffer
, dolock
);
3922 local_irq_restore(flags
);
3924 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3931 * ring_buffer_iter_peek - peek at the next event to be read
3932 * @iter: The ring buffer iterator
3933 * @ts: The timestamp counter of this event.
3935 * This will return the event that will be read next, but does
3936 * not increment the iterator.
3938 struct ring_buffer_event
*
3939 ring_buffer_iter_peek(struct ring_buffer_iter
*iter
, u64
*ts
)
3941 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
3942 struct ring_buffer_event
*event
;
3943 unsigned long flags
;
3946 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3947 event
= rb_iter_peek(iter
, ts
);
3948 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3950 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3957 * ring_buffer_consume - return an event and consume it
3958 * @buffer: The ring buffer to get the next event from
3959 * @cpu: the cpu to read the buffer from
3960 * @ts: a variable to store the timestamp (may be NULL)
3961 * @lost_events: a variable to store if events were lost (may be NULL)
3963 * Returns the next event in the ring buffer, and that event is consumed.
3964 * Meaning, that sequential reads will keep returning a different event,
3965 * and eventually empty the ring buffer if the producer is slower.
3967 struct ring_buffer_event
*
3968 ring_buffer_consume(struct ring_buffer
*buffer
, int cpu
, u64
*ts
,
3969 unsigned long *lost_events
)
3971 struct ring_buffer_per_cpu
*cpu_buffer
;
3972 struct ring_buffer_event
*event
= NULL
;
3973 unsigned long flags
;
3977 /* might be called in atomic */
3980 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3983 cpu_buffer
= buffer
->buffers
[cpu
];
3984 local_irq_save(flags
);
3985 dolock
= rb_reader_lock(cpu_buffer
);
3987 event
= rb_buffer_peek(cpu_buffer
, ts
, lost_events
);
3989 cpu_buffer
->lost_events
= 0;
3990 rb_advance_reader(cpu_buffer
);
3993 rb_reader_unlock(cpu_buffer
, dolock
);
3994 local_irq_restore(flags
);
3999 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
4004 EXPORT_SYMBOL_GPL(ring_buffer_consume
);
4007 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4008 * @buffer: The ring buffer to read from
4009 * @cpu: The cpu buffer to iterate over
4011 * This performs the initial preparations necessary to iterate
4012 * through the buffer. Memory is allocated, buffer recording
4013 * is disabled, and the iterator pointer is returned to the caller.
4015 * Disabling buffer recordng prevents the reading from being
4016 * corrupted. This is not a consuming read, so a producer is not
4019 * After a sequence of ring_buffer_read_prepare calls, the user is
4020 * expected to make at least one call to ring_buffer_read_prepare_sync.
4021 * Afterwards, ring_buffer_read_start is invoked to get things going
4024 * This overall must be paired with ring_buffer_read_finish.
4026 struct ring_buffer_iter
*
4027 ring_buffer_read_prepare(struct ring_buffer
*buffer
, int cpu
)
4029 struct ring_buffer_per_cpu
*cpu_buffer
;
4030 struct ring_buffer_iter
*iter
;
4032 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4035 iter
= kmalloc(sizeof(*iter
), GFP_KERNEL
);
4039 cpu_buffer
= buffer
->buffers
[cpu
];
4041 iter
->cpu_buffer
= cpu_buffer
;
4043 atomic_inc(&buffer
->resize_disabled
);
4044 atomic_inc(&cpu_buffer
->record_disabled
);
4048 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare
);
4051 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4053 * All previously invoked ring_buffer_read_prepare calls to prepare
4054 * iterators will be synchronized. Afterwards, read_buffer_read_start
4055 * calls on those iterators are allowed.
4058 ring_buffer_read_prepare_sync(void)
4060 synchronize_sched();
4062 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync
);
4065 * ring_buffer_read_start - start a non consuming read of the buffer
4066 * @iter: The iterator returned by ring_buffer_read_prepare
4068 * This finalizes the startup of an iteration through the buffer.
4069 * The iterator comes from a call to ring_buffer_read_prepare and
4070 * an intervening ring_buffer_read_prepare_sync must have been
4073 * Must be paired with ring_buffer_read_finish.
4076 ring_buffer_read_start(struct ring_buffer_iter
*iter
)
4078 struct ring_buffer_per_cpu
*cpu_buffer
;
4079 unsigned long flags
;
4084 cpu_buffer
= iter
->cpu_buffer
;
4086 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4087 arch_spin_lock(&cpu_buffer
->lock
);
4088 rb_iter_reset(iter
);
4089 arch_spin_unlock(&cpu_buffer
->lock
);
4090 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4092 EXPORT_SYMBOL_GPL(ring_buffer_read_start
);
4095 * ring_buffer_read_finish - finish reading the iterator of the buffer
4096 * @iter: The iterator retrieved by ring_buffer_start
4098 * This re-enables the recording to the buffer, and frees the
4102 ring_buffer_read_finish(struct ring_buffer_iter
*iter
)
4104 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
4105 unsigned long flags
;
4108 * Ring buffer is disabled from recording, here's a good place
4109 * to check the integrity of the ring buffer.
4110 * Must prevent readers from trying to read, as the check
4111 * clears the HEAD page and readers require it.
4113 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4114 rb_check_pages(cpu_buffer
);
4115 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4117 atomic_dec(&cpu_buffer
->record_disabled
);
4118 atomic_dec(&cpu_buffer
->buffer
->resize_disabled
);
4121 EXPORT_SYMBOL_GPL(ring_buffer_read_finish
);
4124 * ring_buffer_read - read the next item in the ring buffer by the iterator
4125 * @iter: The ring buffer iterator
4126 * @ts: The time stamp of the event read.
4128 * This reads the next event in the ring buffer and increments the iterator.
4130 struct ring_buffer_event
*
4131 ring_buffer_read(struct ring_buffer_iter
*iter
, u64
*ts
)
4133 struct ring_buffer_event
*event
;
4134 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
4135 unsigned long flags
;
4137 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4139 event
= rb_iter_peek(iter
, ts
);
4143 if (event
->type_len
== RINGBUF_TYPE_PADDING
)
4146 rb_advance_iter(iter
);
4148 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4152 EXPORT_SYMBOL_GPL(ring_buffer_read
);
4155 * ring_buffer_size - return the size of the ring buffer (in bytes)
4156 * @buffer: The ring buffer.
4158 unsigned long ring_buffer_size(struct ring_buffer
*buffer
, int cpu
)
4161 * Earlier, this method returned
4162 * BUF_PAGE_SIZE * buffer->nr_pages
4163 * Since the nr_pages field is now removed, we have converted this to
4164 * return the per cpu buffer value.
4166 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4169 return BUF_PAGE_SIZE
* buffer
->buffers
[cpu
]->nr_pages
;
4171 EXPORT_SYMBOL_GPL(ring_buffer_size
);
4174 rb_reset_cpu(struct ring_buffer_per_cpu
*cpu_buffer
)
4176 rb_head_page_deactivate(cpu_buffer
);
4178 cpu_buffer
->head_page
4179 = list_entry(cpu_buffer
->pages
, struct buffer_page
, list
);
4180 local_set(&cpu_buffer
->head_page
->write
, 0);
4181 local_set(&cpu_buffer
->head_page
->entries
, 0);
4182 local_set(&cpu_buffer
->head_page
->page
->commit
, 0);
4184 cpu_buffer
->head_page
->read
= 0;
4186 cpu_buffer
->tail_page
= cpu_buffer
->head_page
;
4187 cpu_buffer
->commit_page
= cpu_buffer
->head_page
;
4189 INIT_LIST_HEAD(&cpu_buffer
->reader_page
->list
);
4190 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
4191 local_set(&cpu_buffer
->reader_page
->write
, 0);
4192 local_set(&cpu_buffer
->reader_page
->entries
, 0);
4193 local_set(&cpu_buffer
->reader_page
->page
->commit
, 0);
4194 cpu_buffer
->reader_page
->read
= 0;
4196 local_set(&cpu_buffer
->entries_bytes
, 0);
4197 local_set(&cpu_buffer
->overrun
, 0);
4198 local_set(&cpu_buffer
->commit_overrun
, 0);
4199 local_set(&cpu_buffer
->dropped_events
, 0);
4200 local_set(&cpu_buffer
->entries
, 0);
4201 local_set(&cpu_buffer
->committing
, 0);
4202 local_set(&cpu_buffer
->commits
, 0);
4203 cpu_buffer
->read
= 0;
4204 cpu_buffer
->read_bytes
= 0;
4206 cpu_buffer
->write_stamp
= 0;
4207 cpu_buffer
->read_stamp
= 0;
4209 cpu_buffer
->lost_events
= 0;
4210 cpu_buffer
->last_overrun
= 0;
4212 rb_head_page_activate(cpu_buffer
);
4216 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4217 * @buffer: The ring buffer to reset a per cpu buffer of
4218 * @cpu: The CPU buffer to be reset
4220 void ring_buffer_reset_cpu(struct ring_buffer
*buffer
, int cpu
)
4222 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
4223 unsigned long flags
;
4225 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4228 atomic_inc(&buffer
->resize_disabled
);
4229 atomic_inc(&cpu_buffer
->record_disabled
);
4231 /* Make sure all commits have finished */
4232 synchronize_sched();
4234 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4236 if (RB_WARN_ON(cpu_buffer
, local_read(&cpu_buffer
->committing
)))
4239 arch_spin_lock(&cpu_buffer
->lock
);
4241 rb_reset_cpu(cpu_buffer
);
4243 arch_spin_unlock(&cpu_buffer
->lock
);
4246 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4248 atomic_dec(&cpu_buffer
->record_disabled
);
4249 atomic_dec(&buffer
->resize_disabled
);
4251 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu
);
4254 * ring_buffer_reset - reset a ring buffer
4255 * @buffer: The ring buffer to reset all cpu buffers
4257 void ring_buffer_reset(struct ring_buffer
*buffer
)
4261 for_each_buffer_cpu(buffer
, cpu
)
4262 ring_buffer_reset_cpu(buffer
, cpu
);
4264 EXPORT_SYMBOL_GPL(ring_buffer_reset
);
4267 * rind_buffer_empty - is the ring buffer empty?
4268 * @buffer: The ring buffer to test
4270 int ring_buffer_empty(struct ring_buffer
*buffer
)
4272 struct ring_buffer_per_cpu
*cpu_buffer
;
4273 unsigned long flags
;
4278 /* yes this is racy, but if you don't like the race, lock the buffer */
4279 for_each_buffer_cpu(buffer
, cpu
) {
4280 cpu_buffer
= buffer
->buffers
[cpu
];
4281 local_irq_save(flags
);
4282 dolock
= rb_reader_lock(cpu_buffer
);
4283 ret
= rb_per_cpu_empty(cpu_buffer
);
4284 rb_reader_unlock(cpu_buffer
, dolock
);
4285 local_irq_restore(flags
);
4293 EXPORT_SYMBOL_GPL(ring_buffer_empty
);
4296 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4297 * @buffer: The ring buffer
4298 * @cpu: The CPU buffer to test
4300 int ring_buffer_empty_cpu(struct ring_buffer
*buffer
, int cpu
)
4302 struct ring_buffer_per_cpu
*cpu_buffer
;
4303 unsigned long flags
;
4307 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4310 cpu_buffer
= buffer
->buffers
[cpu
];
4311 local_irq_save(flags
);
4312 dolock
= rb_reader_lock(cpu_buffer
);
4313 ret
= rb_per_cpu_empty(cpu_buffer
);
4314 rb_reader_unlock(cpu_buffer
, dolock
);
4315 local_irq_restore(flags
);
4319 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu
);
4321 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4323 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4324 * @buffer_a: One buffer to swap with
4325 * @buffer_b: The other buffer to swap with
4327 * This function is useful for tracers that want to take a "snapshot"
4328 * of a CPU buffer and has another back up buffer lying around.
4329 * it is expected that the tracer handles the cpu buffer not being
4330 * used at the moment.
4332 int ring_buffer_swap_cpu(struct ring_buffer
*buffer_a
,
4333 struct ring_buffer
*buffer_b
, int cpu
)
4335 struct ring_buffer_per_cpu
*cpu_buffer_a
;
4336 struct ring_buffer_per_cpu
*cpu_buffer_b
;
4339 if (!cpumask_test_cpu(cpu
, buffer_a
->cpumask
) ||
4340 !cpumask_test_cpu(cpu
, buffer_b
->cpumask
))
4343 cpu_buffer_a
= buffer_a
->buffers
[cpu
];
4344 cpu_buffer_b
= buffer_b
->buffers
[cpu
];
4346 /* At least make sure the two buffers are somewhat the same */
4347 if (cpu_buffer_a
->nr_pages
!= cpu_buffer_b
->nr_pages
)
4352 if (atomic_read(&buffer_a
->record_disabled
))
4355 if (atomic_read(&buffer_b
->record_disabled
))
4358 if (atomic_read(&cpu_buffer_a
->record_disabled
))
4361 if (atomic_read(&cpu_buffer_b
->record_disabled
))
4365 * We can't do a synchronize_sched here because this
4366 * function can be called in atomic context.
4367 * Normally this will be called from the same CPU as cpu.
4368 * If not it's up to the caller to protect this.
4370 atomic_inc(&cpu_buffer_a
->record_disabled
);
4371 atomic_inc(&cpu_buffer_b
->record_disabled
);
4374 if (local_read(&cpu_buffer_a
->committing
))
4376 if (local_read(&cpu_buffer_b
->committing
))
4379 buffer_a
->buffers
[cpu
] = cpu_buffer_b
;
4380 buffer_b
->buffers
[cpu
] = cpu_buffer_a
;
4382 cpu_buffer_b
->buffer
= buffer_a
;
4383 cpu_buffer_a
->buffer
= buffer_b
;
4388 atomic_dec(&cpu_buffer_a
->record_disabled
);
4389 atomic_dec(&cpu_buffer_b
->record_disabled
);
4393 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu
);
4394 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4397 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4398 * @buffer: the buffer to allocate for.
4399 * @cpu: the cpu buffer to allocate.
4401 * This function is used in conjunction with ring_buffer_read_page.
4402 * When reading a full page from the ring buffer, these functions
4403 * can be used to speed up the process. The calling function should
4404 * allocate a few pages first with this function. Then when it
4405 * needs to get pages from the ring buffer, it passes the result
4406 * of this function into ring_buffer_read_page, which will swap
4407 * the page that was allocated, with the read page of the buffer.
4410 * The page allocated, or NULL on error.
4412 void *ring_buffer_alloc_read_page(struct ring_buffer
*buffer
, int cpu
)
4414 struct buffer_data_page
*bpage
;
4417 page
= alloc_pages_node(cpu_to_node(cpu
),
4418 GFP_KERNEL
| __GFP_NORETRY
, 0);
4422 bpage
= page_address(page
);
4424 rb_init_page(bpage
);
4428 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page
);
4431 * ring_buffer_free_read_page - free an allocated read page
4432 * @buffer: the buffer the page was allocate for
4433 * @data: the page to free
4435 * Free a page allocated from ring_buffer_alloc_read_page.
4437 void ring_buffer_free_read_page(struct ring_buffer
*buffer
, void *data
)
4439 free_page((unsigned long)data
);
4441 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page
);
4444 * ring_buffer_read_page - extract a page from the ring buffer
4445 * @buffer: buffer to extract from
4446 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4447 * @len: amount to extract
4448 * @cpu: the cpu of the buffer to extract
4449 * @full: should the extraction only happen when the page is full.
4451 * This function will pull out a page from the ring buffer and consume it.
4452 * @data_page must be the address of the variable that was returned
4453 * from ring_buffer_alloc_read_page. This is because the page might be used
4454 * to swap with a page in the ring buffer.
4457 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4460 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4462 * process_page(rpage, ret);
4464 * When @full is set, the function will not return true unless
4465 * the writer is off the reader page.
4467 * Note: it is up to the calling functions to handle sleeps and wakeups.
4468 * The ring buffer can be used anywhere in the kernel and can not
4469 * blindly call wake_up. The layer that uses the ring buffer must be
4470 * responsible for that.
4473 * >=0 if data has been transferred, returns the offset of consumed data.
4474 * <0 if no data has been transferred.
4476 int ring_buffer_read_page(struct ring_buffer
*buffer
,
4477 void **data_page
, size_t len
, int cpu
, int full
)
4479 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
4480 struct ring_buffer_event
*event
;
4481 struct buffer_data_page
*bpage
;
4482 struct buffer_page
*reader
;
4483 unsigned long missed_events
;
4484 unsigned long flags
;
4485 unsigned int commit
;
4490 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4494 * If len is not big enough to hold the page header, then
4495 * we can not copy anything.
4497 if (len
<= BUF_PAGE_HDR_SIZE
)
4500 len
-= BUF_PAGE_HDR_SIZE
;
4509 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4511 reader
= rb_get_reader_page(cpu_buffer
);
4515 event
= rb_reader_event(cpu_buffer
);
4517 read
= reader
->read
;
4518 commit
= rb_page_commit(reader
);
4520 /* Check if any events were dropped */
4521 missed_events
= cpu_buffer
->lost_events
;
4524 * If this page has been partially read or
4525 * if len is not big enough to read the rest of the page or
4526 * a writer is still on the page, then
4527 * we must copy the data from the page to the buffer.
4528 * Otherwise, we can simply swap the page with the one passed in.
4530 if (read
|| (len
< (commit
- read
)) ||
4531 cpu_buffer
->reader_page
== cpu_buffer
->commit_page
) {
4532 struct buffer_data_page
*rpage
= cpu_buffer
->reader_page
->page
;
4533 unsigned int rpos
= read
;
4534 unsigned int pos
= 0;
4540 if (len
> (commit
- read
))
4541 len
= (commit
- read
);
4543 /* Always keep the time extend and data together */
4544 size
= rb_event_ts_length(event
);
4549 /* save the current timestamp, since the user will need it */
4550 save_timestamp
= cpu_buffer
->read_stamp
;
4552 /* Need to copy one event at a time */
4554 /* We need the size of one event, because
4555 * rb_advance_reader only advances by one event,
4556 * whereas rb_event_ts_length may include the size of
4557 * one or two events.
4558 * We have already ensured there's enough space if this
4559 * is a time extend. */
4560 size
= rb_event_length(event
);
4561 memcpy(bpage
->data
+ pos
, rpage
->data
+ rpos
, size
);
4565 rb_advance_reader(cpu_buffer
);
4566 rpos
= reader
->read
;
4572 event
= rb_reader_event(cpu_buffer
);
4573 /* Always keep the time extend and data together */
4574 size
= rb_event_ts_length(event
);
4575 } while (len
>= size
);
4578 local_set(&bpage
->commit
, pos
);
4579 bpage
->time_stamp
= save_timestamp
;
4581 /* we copied everything to the beginning */
4584 /* update the entry counter */
4585 cpu_buffer
->read
+= rb_page_entries(reader
);
4586 cpu_buffer
->read_bytes
+= BUF_PAGE_SIZE
;
4588 /* swap the pages */
4589 rb_init_page(bpage
);
4590 bpage
= reader
->page
;
4591 reader
->page
= *data_page
;
4592 local_set(&reader
->write
, 0);
4593 local_set(&reader
->entries
, 0);
4598 * Use the real_end for the data size,
4599 * This gives us a chance to store the lost events
4602 if (reader
->real_end
)
4603 local_set(&bpage
->commit
, reader
->real_end
);
4607 cpu_buffer
->lost_events
= 0;
4609 commit
= local_read(&bpage
->commit
);
4611 * Set a flag in the commit field if we lost events
4613 if (missed_events
) {
4614 /* If there is room at the end of the page to save the
4615 * missed events, then record it there.
4617 if (BUF_PAGE_SIZE
- commit
>= sizeof(missed_events
)) {
4618 memcpy(&bpage
->data
[commit
], &missed_events
,
4619 sizeof(missed_events
));
4620 local_add(RB_MISSED_STORED
, &bpage
->commit
);
4621 commit
+= sizeof(missed_events
);
4623 local_add(RB_MISSED_EVENTS
, &bpage
->commit
);
4627 * This page may be off to user land. Zero it out here.
4629 if (commit
< BUF_PAGE_SIZE
)
4630 memset(&bpage
->data
[commit
], 0, BUF_PAGE_SIZE
- commit
);
4633 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4638 EXPORT_SYMBOL_GPL(ring_buffer_read_page
);
4640 #ifdef CONFIG_HOTPLUG_CPU
4641 static int rb_cpu_notify(struct notifier_block
*self
,
4642 unsigned long action
, void *hcpu
)
4644 struct ring_buffer
*buffer
=
4645 container_of(self
, struct ring_buffer
, cpu_notify
);
4646 long cpu
= (long)hcpu
;
4647 int cpu_i
, nr_pages_same
;
4648 unsigned int nr_pages
;
4651 case CPU_UP_PREPARE
:
4652 case CPU_UP_PREPARE_FROZEN
:
4653 if (cpumask_test_cpu(cpu
, buffer
->cpumask
))
4658 /* check if all cpu sizes are same */
4659 for_each_buffer_cpu(buffer
, cpu_i
) {
4660 /* fill in the size from first enabled cpu */
4662 nr_pages
= buffer
->buffers
[cpu_i
]->nr_pages
;
4663 if (nr_pages
!= buffer
->buffers
[cpu_i
]->nr_pages
) {
4668 /* allocate minimum pages, user can later expand it */
4671 buffer
->buffers
[cpu
] =
4672 rb_allocate_cpu_buffer(buffer
, nr_pages
, cpu
);
4673 if (!buffer
->buffers
[cpu
]) {
4674 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4679 cpumask_set_cpu(cpu
, buffer
->cpumask
);
4681 case CPU_DOWN_PREPARE
:
4682 case CPU_DOWN_PREPARE_FROZEN
:
4685 * If we were to free the buffer, then the user would
4686 * lose any trace that was in the buffer.
4696 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4698 * This is a basic integrity check of the ring buffer.
4699 * Late in the boot cycle this test will run when configured in.
4700 * It will kick off a thread per CPU that will go into a loop
4701 * writing to the per cpu ring buffer various sizes of data.
4702 * Some of the data will be large items, some small.
4704 * Another thread is created that goes into a spin, sending out
4705 * IPIs to the other CPUs to also write into the ring buffer.
4706 * this is to test the nesting ability of the buffer.
4708 * Basic stats are recorded and reported. If something in the
4709 * ring buffer should happen that's not expected, a big warning
4710 * is displayed and all ring buffers are disabled.
4712 static struct task_struct
*rb_threads
[NR_CPUS
] __initdata
;
4714 struct rb_test_data
{
4715 struct ring_buffer
*buffer
;
4716 unsigned long events
;
4717 unsigned long bytes_written
;
4718 unsigned long bytes_alloc
;
4719 unsigned long bytes_dropped
;
4720 unsigned long events_nested
;
4721 unsigned long bytes_written_nested
;
4722 unsigned long bytes_alloc_nested
;
4723 unsigned long bytes_dropped_nested
;
4724 int min_size_nested
;
4725 int max_size_nested
;
4732 static struct rb_test_data rb_data
[NR_CPUS
] __initdata
;
4735 #define RB_TEST_BUFFER_SIZE 1048576
4737 static char rb_string
[] __initdata
=
4738 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4739 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4740 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4742 static bool rb_test_started __initdata
;
4749 static __init
int rb_write_something(struct rb_test_data
*data
, bool nested
)
4751 struct ring_buffer_event
*event
;
4752 struct rb_item
*item
;
4759 /* Have nested writes different that what is written */
4760 cnt
= data
->cnt
+ (nested
? 27 : 0);
4762 /* Multiply cnt by ~e, to make some unique increment */
4763 size
= (data
->cnt
* 68 / 25) % (sizeof(rb_string
) - 1);
4765 len
= size
+ sizeof(struct rb_item
);
4767 started
= rb_test_started
;
4768 /* read rb_test_started before checking buffer enabled */
4771 event
= ring_buffer_lock_reserve(data
->buffer
, len
);
4773 /* Ignore dropped events before test starts. */
4776 data
->bytes_dropped
+= len
;
4778 data
->bytes_dropped_nested
+= len
;
4783 event_len
= ring_buffer_event_length(event
);
4785 if (RB_WARN_ON(data
->buffer
, event_len
< len
))
4788 item
= ring_buffer_event_data(event
);
4790 memcpy(item
->str
, rb_string
, size
);
4793 data
->bytes_alloc_nested
+= event_len
;
4794 data
->bytes_written_nested
+= len
;
4795 data
->events_nested
++;
4796 if (!data
->min_size_nested
|| len
< data
->min_size_nested
)
4797 data
->min_size_nested
= len
;
4798 if (len
> data
->max_size_nested
)
4799 data
->max_size_nested
= len
;
4801 data
->bytes_alloc
+= event_len
;
4802 data
->bytes_written
+= len
;
4804 if (!data
->min_size
|| len
< data
->min_size
)
4805 data
->max_size
= len
;
4806 if (len
> data
->max_size
)
4807 data
->max_size
= len
;
4811 ring_buffer_unlock_commit(data
->buffer
, event
);
4816 static __init
int rb_test(void *arg
)
4818 struct rb_test_data
*data
= arg
;
4820 while (!kthread_should_stop()) {
4821 rb_write_something(data
, false);
4824 set_current_state(TASK_INTERRUPTIBLE
);
4825 /* Now sleep between a min of 100-300us and a max of 1ms */
4826 usleep_range(((data
->cnt
% 3) + 1) * 100, 1000);
4832 static __init
void rb_ipi(void *ignore
)
4834 struct rb_test_data
*data
;
4835 int cpu
= smp_processor_id();
4837 data
= &rb_data
[cpu
];
4838 rb_write_something(data
, true);
4841 static __init
int rb_hammer_test(void *arg
)
4843 while (!kthread_should_stop()) {
4845 /* Send an IPI to all cpus to write data! */
4846 smp_call_function(rb_ipi
, NULL
, 1);
4847 /* No sleep, but for non preempt, let others run */
4854 static __init
int test_ringbuffer(void)
4856 struct task_struct
*rb_hammer
;
4857 struct ring_buffer
*buffer
;
4861 pr_info("Running ring buffer tests...\n");
4863 buffer
= ring_buffer_alloc(RB_TEST_BUFFER_SIZE
, RB_FL_OVERWRITE
);
4864 if (WARN_ON(!buffer
))
4867 /* Disable buffer so that threads can't write to it yet */
4868 ring_buffer_record_off(buffer
);
4870 for_each_online_cpu(cpu
) {
4871 rb_data
[cpu
].buffer
= buffer
;
4872 rb_data
[cpu
].cpu
= cpu
;
4873 rb_data
[cpu
].cnt
= cpu
;
4874 rb_threads
[cpu
] = kthread_create(rb_test
, &rb_data
[cpu
],
4875 "rbtester/%d", cpu
);
4876 if (WARN_ON(!rb_threads
[cpu
])) {
4877 pr_cont("FAILED\n");
4882 kthread_bind(rb_threads
[cpu
], cpu
);
4883 wake_up_process(rb_threads
[cpu
]);
4886 /* Now create the rb hammer! */
4887 rb_hammer
= kthread_run(rb_hammer_test
, NULL
, "rbhammer");
4888 if (WARN_ON(!rb_hammer
)) {
4889 pr_cont("FAILED\n");
4894 ring_buffer_record_on(buffer
);
4896 * Show buffer is enabled before setting rb_test_started.
4897 * Yes there's a small race window where events could be
4898 * dropped and the thread wont catch it. But when a ring
4899 * buffer gets enabled, there will always be some kind of
4900 * delay before other CPUs see it. Thus, we don't care about
4901 * those dropped events. We care about events dropped after
4902 * the threads see that the buffer is active.
4905 rb_test_started
= true;
4907 set_current_state(TASK_INTERRUPTIBLE
);
4908 /* Just run for 10 seconds */;
4909 schedule_timeout(10 * HZ
);
4911 kthread_stop(rb_hammer
);
4914 for_each_online_cpu(cpu
) {
4915 if (!rb_threads
[cpu
])
4917 kthread_stop(rb_threads
[cpu
]);
4920 ring_buffer_free(buffer
);
4925 pr_info("finished\n");
4926 for_each_online_cpu(cpu
) {
4927 struct ring_buffer_event
*event
;
4928 struct rb_test_data
*data
= &rb_data
[cpu
];
4929 struct rb_item
*item
;
4930 unsigned long total_events
;
4931 unsigned long total_dropped
;
4932 unsigned long total_written
;
4933 unsigned long total_alloc
;
4934 unsigned long total_read
= 0;
4935 unsigned long total_size
= 0;
4936 unsigned long total_len
= 0;
4937 unsigned long total_lost
= 0;
4940 int small_event_size
;
4944 total_events
= data
->events
+ data
->events_nested
;
4945 total_written
= data
->bytes_written
+ data
->bytes_written_nested
;
4946 total_alloc
= data
->bytes_alloc
+ data
->bytes_alloc_nested
;
4947 total_dropped
= data
->bytes_dropped
+ data
->bytes_dropped_nested
;
4949 big_event_size
= data
->max_size
+ data
->max_size_nested
;
4950 small_event_size
= data
->min_size
+ data
->min_size_nested
;
4952 pr_info("CPU %d:\n", cpu
);
4953 pr_info(" events: %ld\n", total_events
);
4954 pr_info(" dropped bytes: %ld\n", total_dropped
);
4955 pr_info(" alloced bytes: %ld\n", total_alloc
);
4956 pr_info(" written bytes: %ld\n", total_written
);
4957 pr_info(" biggest event: %d\n", big_event_size
);
4958 pr_info(" smallest event: %d\n", small_event_size
);
4960 if (RB_WARN_ON(buffer
, total_dropped
))
4965 while ((event
= ring_buffer_consume(buffer
, cpu
, NULL
, &lost
))) {
4967 item
= ring_buffer_event_data(event
);
4968 total_len
+= ring_buffer_event_length(event
);
4969 total_size
+= item
->size
+ sizeof(struct rb_item
);
4970 if (memcmp(&item
->str
[0], rb_string
, item
->size
) != 0) {
4971 pr_info("FAILED!\n");
4972 pr_info("buffer had: %.*s\n", item
->size
, item
->str
);
4973 pr_info("expected: %.*s\n", item
->size
, rb_string
);
4974 RB_WARN_ON(buffer
, 1);
4985 pr_info(" read events: %ld\n", total_read
);
4986 pr_info(" lost events: %ld\n", total_lost
);
4987 pr_info(" total events: %ld\n", total_lost
+ total_read
);
4988 pr_info(" recorded len bytes: %ld\n", total_len
);
4989 pr_info(" recorded size bytes: %ld\n", total_size
);
4991 pr_info(" With dropped events, record len and size may not match\n"
4992 " alloced and written from above\n");
4994 if (RB_WARN_ON(buffer
, total_len
!= total_alloc
||
4995 total_size
!= total_written
))
4998 if (RB_WARN_ON(buffer
, total_lost
+ total_read
!= total_events
))
5004 pr_info("Ring buffer PASSED!\n");
5006 ring_buffer_free(buffer
);
5010 late_initcall(test_ringbuffer
);
5011 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */