4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 #include <linux/trace_events.h>
7 #include <linux/ring_buffer.h>
8 #include <linux/trace_clock.h>
9 #include <linux/trace_seq.h>
10 #include <linux/spinlock.h>
11 #include <linux/irq_work.h>
12 #include <linux/uaccess.h>
13 #include <linux/hardirq.h>
14 #include <linux/kthread.h> /* for self test */
15 #include <linux/kmemcheck.h>
16 #include <linux/module.h>
17 #include <linux/percpu.h>
18 #include <linux/mutex.h>
19 #include <linux/delay.h>
20 #include <linux/slab.h>
21 #include <linux/init.h>
22 #include <linux/hash.h>
23 #include <linux/list.h>
24 #include <linux/cpu.h>
26 #include <asm/local.h>
28 static void update_pages_handler(struct work_struct
*work
);
31 * The ring buffer header is special. We must manually up keep it.
33 int ring_buffer_print_entry_header(struct trace_seq
*s
)
35 trace_seq_puts(s
, "# compressed entry header\n");
36 trace_seq_puts(s
, "\ttype_len : 5 bits\n");
37 trace_seq_puts(s
, "\ttime_delta : 27 bits\n");
38 trace_seq_puts(s
, "\tarray : 32 bits\n");
39 trace_seq_putc(s
, '\n');
40 trace_seq_printf(s
, "\tpadding : type == %d\n",
41 RINGBUF_TYPE_PADDING
);
42 trace_seq_printf(s
, "\ttime_extend : type == %d\n",
43 RINGBUF_TYPE_TIME_EXTEND
);
44 trace_seq_printf(s
, "\tdata max type_len == %d\n",
45 RINGBUF_TYPE_DATA_TYPE_LEN_MAX
);
47 return !trace_seq_has_overflowed(s
);
51 * The ring buffer is made up of a list of pages. A separate list of pages is
52 * allocated for each CPU. A writer may only write to a buffer that is
53 * associated with the CPU it is currently executing on. A reader may read
54 * from any per cpu buffer.
56 * The reader is special. For each per cpu buffer, the reader has its own
57 * reader page. When a reader has read the entire reader page, this reader
58 * page is swapped with another page in the ring buffer.
60 * Now, as long as the writer is off the reader page, the reader can do what
61 * ever it wants with that page. The writer will never write to that page
62 * again (as long as it is out of the ring buffer).
64 * Here's some silly ASCII art.
67 * |reader| RING BUFFER
69 * +------+ +---+ +---+ +---+
78 * |reader| RING BUFFER
79 * |page |------------------v
80 * +------+ +---+ +---+ +---+
89 * |reader| RING BUFFER
90 * |page |------------------v
91 * +------+ +---+ +---+ +---+
96 * +------------------------------+
100 * |buffer| RING BUFFER
101 * |page |------------------v
102 * +------+ +---+ +---+ +---+
104 * | New +---+ +---+ +---+
107 * +------------------------------+
110 * After we make this swap, the reader can hand this page off to the splice
111 * code and be done with it. It can even allocate a new page if it needs to
112 * and swap that into the ring buffer.
114 * We will be using cmpxchg soon to make all this lockless.
119 * A fast way to enable or disable all ring buffers is to
120 * call tracing_on or tracing_off. Turning off the ring buffers
121 * prevents all ring buffers from being recorded to.
122 * Turning this switch on, makes it OK to write to the
123 * ring buffer, if the ring buffer is enabled itself.
125 * There's three layers that must be on in order to write
126 * to the ring buffer.
128 * 1) This global flag must be set.
129 * 2) The ring buffer must be enabled for recording.
130 * 3) The per cpu buffer must be enabled for recording.
132 * In case of an anomaly, this global flag has a bit set that
133 * will permantly disable all ring buffers.
137 * Global flag to disable all recording to ring buffers
138 * This has two bits: ON, DISABLED
142 * 0 0 : ring buffers are off
143 * 1 0 : ring buffers are on
144 * X 1 : ring buffers are permanently disabled
148 RB_BUFFERS_ON_BIT
= 0,
149 RB_BUFFERS_DISABLED_BIT
= 1,
153 RB_BUFFERS_ON
= 1 << RB_BUFFERS_ON_BIT
,
154 RB_BUFFERS_DISABLED
= 1 << RB_BUFFERS_DISABLED_BIT
,
157 static unsigned long ring_buffer_flags __read_mostly
= RB_BUFFERS_ON
;
159 /* Used for individual buffers (after the counter) */
160 #define RB_BUFFER_OFF (1 << 20)
162 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
165 * tracing_off_permanent - permanently disable ring buffers
167 * This function, once called, will disable all ring buffers
170 void tracing_off_permanent(void)
172 set_bit(RB_BUFFERS_DISABLED_BIT
, &ring_buffer_flags
);
175 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
176 #define RB_ALIGNMENT 4U
177 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
178 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
180 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
181 # define RB_FORCE_8BYTE_ALIGNMENT 0
182 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
184 # define RB_FORCE_8BYTE_ALIGNMENT 1
185 # define RB_ARCH_ALIGNMENT 8U
188 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
190 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
191 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
194 RB_LEN_TIME_EXTEND
= 8,
195 RB_LEN_TIME_STAMP
= 16,
198 #define skip_time_extend(event) \
199 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
201 static inline int rb_null_event(struct ring_buffer_event
*event
)
203 return event
->type_len
== RINGBUF_TYPE_PADDING
&& !event
->time_delta
;
206 static void rb_event_set_padding(struct ring_buffer_event
*event
)
208 /* padding has a NULL time_delta */
209 event
->type_len
= RINGBUF_TYPE_PADDING
;
210 event
->time_delta
= 0;
214 rb_event_data_length(struct ring_buffer_event
*event
)
219 length
= event
->type_len
* RB_ALIGNMENT
;
221 length
= event
->array
[0];
222 return length
+ RB_EVNT_HDR_SIZE
;
226 * Return the length of the given event. Will return
227 * the length of the time extend if the event is a
230 static inline unsigned
231 rb_event_length(struct ring_buffer_event
*event
)
233 switch (event
->type_len
) {
234 case RINGBUF_TYPE_PADDING
:
235 if (rb_null_event(event
))
238 return event
->array
[0] + RB_EVNT_HDR_SIZE
;
240 case RINGBUF_TYPE_TIME_EXTEND
:
241 return RB_LEN_TIME_EXTEND
;
243 case RINGBUF_TYPE_TIME_STAMP
:
244 return RB_LEN_TIME_STAMP
;
246 case RINGBUF_TYPE_DATA
:
247 return rb_event_data_length(event
);
256 * Return total length of time extend and data,
257 * or just the event length for all other events.
259 static inline unsigned
260 rb_event_ts_length(struct ring_buffer_event
*event
)
264 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
) {
265 /* time extends include the data event after it */
266 len
= RB_LEN_TIME_EXTEND
;
267 event
= skip_time_extend(event
);
269 return len
+ rb_event_length(event
);
273 * ring_buffer_event_length - return the length of the event
274 * @event: the event to get the length of
276 * Returns the size of the data load of a data event.
277 * If the event is something other than a data event, it
278 * returns the size of the event itself. With the exception
279 * of a TIME EXTEND, where it still returns the size of the
280 * data load of the data event after it.
282 unsigned ring_buffer_event_length(struct ring_buffer_event
*event
)
286 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
287 event
= skip_time_extend(event
);
289 length
= rb_event_length(event
);
290 if (event
->type_len
> RINGBUF_TYPE_DATA_TYPE_LEN_MAX
)
292 length
-= RB_EVNT_HDR_SIZE
;
293 if (length
> RB_MAX_SMALL_DATA
+ sizeof(event
->array
[0]))
294 length
-= sizeof(event
->array
[0]);
297 EXPORT_SYMBOL_GPL(ring_buffer_event_length
);
299 /* inline for ring buffer fast paths */
301 rb_event_data(struct ring_buffer_event
*event
)
303 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
304 event
= skip_time_extend(event
);
305 BUG_ON(event
->type_len
> RINGBUF_TYPE_DATA_TYPE_LEN_MAX
);
306 /* If length is in len field, then array[0] has the data */
308 return (void *)&event
->array
[0];
309 /* Otherwise length is in array[0] and array[1] has the data */
310 return (void *)&event
->array
[1];
314 * ring_buffer_event_data - return the data of the event
315 * @event: the event to get the data from
317 void *ring_buffer_event_data(struct ring_buffer_event
*event
)
319 return rb_event_data(event
);
321 EXPORT_SYMBOL_GPL(ring_buffer_event_data
);
323 #define for_each_buffer_cpu(buffer, cpu) \
324 for_each_cpu(cpu, buffer->cpumask)
327 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
328 #define TS_DELTA_TEST (~TS_MASK)
330 /* Flag when events were overwritten */
331 #define RB_MISSED_EVENTS (1 << 31)
332 /* Missed count stored at end */
333 #define RB_MISSED_STORED (1 << 30)
335 struct buffer_data_page
{
336 u64 time_stamp
; /* page time stamp */
337 local_t commit
; /* write committed index */
338 unsigned char data
[] RB_ALIGN_DATA
; /* data of buffer page */
342 * Note, the buffer_page list must be first. The buffer pages
343 * are allocated in cache lines, which means that each buffer
344 * page will be at the beginning of a cache line, and thus
345 * the least significant bits will be zero. We use this to
346 * add flags in the list struct pointers, to make the ring buffer
350 struct list_head list
; /* list of buffer pages */
351 local_t write
; /* index for next write */
352 unsigned read
; /* index for next read */
353 local_t entries
; /* entries on this page */
354 unsigned long real_end
; /* real end of data */
355 struct buffer_data_page
*page
; /* Actual data page */
359 * The buffer page counters, write and entries, must be reset
360 * atomically when crossing page boundaries. To synchronize this
361 * update, two counters are inserted into the number. One is
362 * the actual counter for the write position or count on the page.
364 * The other is a counter of updaters. Before an update happens
365 * the update partition of the counter is incremented. This will
366 * allow the updater to update the counter atomically.
368 * The counter is 20 bits, and the state data is 12.
370 #define RB_WRITE_MASK 0xfffff
371 #define RB_WRITE_INTCNT (1 << 20)
373 static void rb_init_page(struct buffer_data_page
*bpage
)
375 local_set(&bpage
->commit
, 0);
379 * ring_buffer_page_len - the size of data on the page.
380 * @page: The page to read
382 * Returns the amount of data on the page, including buffer page header.
384 size_t ring_buffer_page_len(void *page
)
386 return local_read(&((struct buffer_data_page
*)page
)->commit
)
391 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
394 static void free_buffer_page(struct buffer_page
*bpage
)
396 free_page((unsigned long)bpage
->page
);
401 * We need to fit the time_stamp delta into 27 bits.
403 static inline int test_time_stamp(u64 delta
)
405 if (delta
& TS_DELTA_TEST
)
410 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
412 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
413 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
415 int ring_buffer_print_page_header(struct trace_seq
*s
)
417 struct buffer_data_page field
;
419 trace_seq_printf(s
, "\tfield: u64 timestamp;\t"
420 "offset:0;\tsize:%u;\tsigned:%u;\n",
421 (unsigned int)sizeof(field
.time_stamp
),
422 (unsigned int)is_signed_type(u64
));
424 trace_seq_printf(s
, "\tfield: local_t commit;\t"
425 "offset:%u;\tsize:%u;\tsigned:%u;\n",
426 (unsigned int)offsetof(typeof(field
), commit
),
427 (unsigned int)sizeof(field
.commit
),
428 (unsigned int)is_signed_type(long));
430 trace_seq_printf(s
, "\tfield: int overwrite;\t"
431 "offset:%u;\tsize:%u;\tsigned:%u;\n",
432 (unsigned int)offsetof(typeof(field
), commit
),
434 (unsigned int)is_signed_type(long));
436 trace_seq_printf(s
, "\tfield: char data;\t"
437 "offset:%u;\tsize:%u;\tsigned:%u;\n",
438 (unsigned int)offsetof(typeof(field
), data
),
439 (unsigned int)BUF_PAGE_SIZE
,
440 (unsigned int)is_signed_type(char));
442 return !trace_seq_has_overflowed(s
);
446 struct irq_work work
;
447 wait_queue_head_t waiters
;
448 wait_queue_head_t full_waiters
;
449 bool waiters_pending
;
450 bool full_waiters_pending
;
455 * head_page == tail_page && head == tail then buffer is empty.
457 struct ring_buffer_per_cpu
{
459 atomic_t record_disabled
;
460 struct ring_buffer
*buffer
;
461 raw_spinlock_t reader_lock
; /* serialize readers */
462 arch_spinlock_t lock
;
463 struct lock_class_key lock_key
;
464 unsigned int nr_pages
;
465 unsigned int current_context
;
466 struct list_head
*pages
;
467 struct buffer_page
*head_page
; /* read from head */
468 struct buffer_page
*tail_page
; /* write to tail */
469 struct buffer_page
*commit_page
; /* committed pages */
470 struct buffer_page
*reader_page
;
471 unsigned long lost_events
;
472 unsigned long last_overrun
;
473 local_t entries_bytes
;
476 local_t commit_overrun
;
477 local_t dropped_events
;
481 unsigned long read_bytes
;
484 /* ring buffer pages to update, > 0 to add, < 0 to remove */
485 int nr_pages_to_update
;
486 struct list_head new_pages
; /* new pages to add */
487 struct work_struct update_pages_work
;
488 struct completion update_done
;
490 struct rb_irq_work irq_work
;
496 atomic_t record_disabled
;
497 atomic_t resize_disabled
;
498 cpumask_var_t cpumask
;
500 struct lock_class_key
*reader_lock_key
;
504 struct ring_buffer_per_cpu
**buffers
;
506 #ifdef CONFIG_HOTPLUG_CPU
507 struct notifier_block cpu_notify
;
511 struct rb_irq_work irq_work
;
514 struct ring_buffer_iter
{
515 struct ring_buffer_per_cpu
*cpu_buffer
;
517 struct buffer_page
*head_page
;
518 struct buffer_page
*cache_reader_page
;
519 unsigned long cache_read
;
524 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
526 * Schedules a delayed work to wake up any task that is blocked on the
527 * ring buffer waiters queue.
529 static void rb_wake_up_waiters(struct irq_work
*work
)
531 struct rb_irq_work
*rbwork
= container_of(work
, struct rb_irq_work
, work
);
533 wake_up_all(&rbwork
->waiters
);
534 if (rbwork
->wakeup_full
) {
535 rbwork
->wakeup_full
= false;
536 wake_up_all(&rbwork
->full_waiters
);
541 * ring_buffer_wait - wait for input to the ring buffer
542 * @buffer: buffer to wait on
543 * @cpu: the cpu buffer to wait on
544 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
546 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
547 * as data is added to any of the @buffer's cpu buffers. Otherwise
548 * it will wait for data to be added to a specific cpu buffer.
550 int ring_buffer_wait(struct ring_buffer
*buffer
, int cpu
, bool full
)
552 struct ring_buffer_per_cpu
*uninitialized_var(cpu_buffer
);
554 struct rb_irq_work
*work
;
558 * Depending on what the caller is waiting for, either any
559 * data in any cpu buffer, or a specific buffer, put the
560 * caller on the appropriate wait queue.
562 if (cpu
== RING_BUFFER_ALL_CPUS
) {
563 work
= &buffer
->irq_work
;
564 /* Full only makes sense on per cpu reads */
567 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
569 cpu_buffer
= buffer
->buffers
[cpu
];
570 work
= &cpu_buffer
->irq_work
;
576 prepare_to_wait(&work
->full_waiters
, &wait
, TASK_INTERRUPTIBLE
);
578 prepare_to_wait(&work
->waiters
, &wait
, TASK_INTERRUPTIBLE
);
581 * The events can happen in critical sections where
582 * checking a work queue can cause deadlocks.
583 * After adding a task to the queue, this flag is set
584 * only to notify events to try to wake up the queue
587 * We don't clear it even if the buffer is no longer
588 * empty. The flag only causes the next event to run
589 * irq_work to do the work queue wake up. The worse
590 * that can happen if we race with !trace_empty() is that
591 * an event will cause an irq_work to try to wake up
594 * There's no reason to protect this flag either, as
595 * the work queue and irq_work logic will do the necessary
596 * synchronization for the wake ups. The only thing
597 * that is necessary is that the wake up happens after
598 * a task has been queued. It's OK for spurious wake ups.
601 work
->full_waiters_pending
= true;
603 work
->waiters_pending
= true;
605 if (signal_pending(current
)) {
610 if (cpu
== RING_BUFFER_ALL_CPUS
&& !ring_buffer_empty(buffer
))
613 if (cpu
!= RING_BUFFER_ALL_CPUS
&&
614 !ring_buffer_empty_cpu(buffer
, cpu
)) {
621 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
622 pagebusy
= cpu_buffer
->reader_page
== cpu_buffer
->commit_page
;
623 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
633 finish_wait(&work
->full_waiters
, &wait
);
635 finish_wait(&work
->waiters
, &wait
);
641 * ring_buffer_poll_wait - poll on buffer input
642 * @buffer: buffer to wait on
643 * @cpu: the cpu buffer to wait on
644 * @filp: the file descriptor
645 * @poll_table: The poll descriptor
647 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
648 * as data is added to any of the @buffer's cpu buffers. Otherwise
649 * it will wait for data to be added to a specific cpu buffer.
651 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
654 int ring_buffer_poll_wait(struct ring_buffer
*buffer
, int cpu
,
655 struct file
*filp
, poll_table
*poll_table
)
657 struct ring_buffer_per_cpu
*cpu_buffer
;
658 struct rb_irq_work
*work
;
660 if (cpu
== RING_BUFFER_ALL_CPUS
)
661 work
= &buffer
->irq_work
;
663 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
666 cpu_buffer
= buffer
->buffers
[cpu
];
667 work
= &cpu_buffer
->irq_work
;
670 poll_wait(filp
, &work
->waiters
, poll_table
);
671 work
->waiters_pending
= true;
673 * There's a tight race between setting the waiters_pending and
674 * checking if the ring buffer is empty. Once the waiters_pending bit
675 * is set, the next event will wake the task up, but we can get stuck
676 * if there's only a single event in.
678 * FIXME: Ideally, we need a memory barrier on the writer side as well,
679 * but adding a memory barrier to all events will cause too much of a
680 * performance hit in the fast path. We only need a memory barrier when
681 * the buffer goes from empty to having content. But as this race is
682 * extremely small, and it's not a problem if another event comes in, we
687 if ((cpu
== RING_BUFFER_ALL_CPUS
&& !ring_buffer_empty(buffer
)) ||
688 (cpu
!= RING_BUFFER_ALL_CPUS
&& !ring_buffer_empty_cpu(buffer
, cpu
)))
689 return POLLIN
| POLLRDNORM
;
693 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
694 #define RB_WARN_ON(b, cond) \
696 int _____ret = unlikely(cond); \
698 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
699 struct ring_buffer_per_cpu *__b = \
701 atomic_inc(&__b->buffer->record_disabled); \
703 atomic_inc(&b->record_disabled); \
709 /* Up this if you want to test the TIME_EXTENTS and normalization */
710 #define DEBUG_SHIFT 0
712 static inline u64
rb_time_stamp(struct ring_buffer
*buffer
)
714 /* shift to debug/test normalization and TIME_EXTENTS */
715 return buffer
->clock() << DEBUG_SHIFT
;
718 u64
ring_buffer_time_stamp(struct ring_buffer
*buffer
, int cpu
)
722 preempt_disable_notrace();
723 time
= rb_time_stamp(buffer
);
724 preempt_enable_no_resched_notrace();
728 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp
);
730 void ring_buffer_normalize_time_stamp(struct ring_buffer
*buffer
,
733 /* Just stupid testing the normalize function and deltas */
736 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp
);
739 * Making the ring buffer lockless makes things tricky.
740 * Although writes only happen on the CPU that they are on,
741 * and they only need to worry about interrupts. Reads can
744 * The reader page is always off the ring buffer, but when the
745 * reader finishes with a page, it needs to swap its page with
746 * a new one from the buffer. The reader needs to take from
747 * the head (writes go to the tail). But if a writer is in overwrite
748 * mode and wraps, it must push the head page forward.
750 * Here lies the problem.
752 * The reader must be careful to replace only the head page, and
753 * not another one. As described at the top of the file in the
754 * ASCII art, the reader sets its old page to point to the next
755 * page after head. It then sets the page after head to point to
756 * the old reader page. But if the writer moves the head page
757 * during this operation, the reader could end up with the tail.
759 * We use cmpxchg to help prevent this race. We also do something
760 * special with the page before head. We set the LSB to 1.
762 * When the writer must push the page forward, it will clear the
763 * bit that points to the head page, move the head, and then set
764 * the bit that points to the new head page.
766 * We also don't want an interrupt coming in and moving the head
767 * page on another writer. Thus we use the second LSB to catch
770 * head->list->prev->next bit 1 bit 0
773 * Points to head page 0 1
776 * Note we can not trust the prev pointer of the head page, because:
778 * +----+ +-----+ +-----+
779 * | |------>| T |---X--->| N |
781 * +----+ +-----+ +-----+
784 * +----------| R |----------+ |
788 * Key: ---X--> HEAD flag set in pointer
793 * (see __rb_reserve_next() to see where this happens)
795 * What the above shows is that the reader just swapped out
796 * the reader page with a page in the buffer, but before it
797 * could make the new header point back to the new page added
798 * it was preempted by a writer. The writer moved forward onto
799 * the new page added by the reader and is about to move forward
802 * You can see, it is legitimate for the previous pointer of
803 * the head (or any page) not to point back to itself. But only
807 #define RB_PAGE_NORMAL 0UL
808 #define RB_PAGE_HEAD 1UL
809 #define RB_PAGE_UPDATE 2UL
812 #define RB_FLAG_MASK 3UL
814 /* PAGE_MOVED is not part of the mask */
815 #define RB_PAGE_MOVED 4UL
818 * rb_list_head - remove any bit
820 static struct list_head
*rb_list_head(struct list_head
*list
)
822 unsigned long val
= (unsigned long)list
;
824 return (struct list_head
*)(val
& ~RB_FLAG_MASK
);
828 * rb_is_head_page - test if the given page is the head page
830 * Because the reader may move the head_page pointer, we can
831 * not trust what the head page is (it may be pointing to
832 * the reader page). But if the next page is a header page,
833 * its flags will be non zero.
836 rb_is_head_page(struct ring_buffer_per_cpu
*cpu_buffer
,
837 struct buffer_page
*page
, struct list_head
*list
)
841 val
= (unsigned long)list
->next
;
843 if ((val
& ~RB_FLAG_MASK
) != (unsigned long)&page
->list
)
844 return RB_PAGE_MOVED
;
846 return val
& RB_FLAG_MASK
;
852 * The unique thing about the reader page, is that, if the
853 * writer is ever on it, the previous pointer never points
854 * back to the reader page.
856 static int rb_is_reader_page(struct buffer_page
*page
)
858 struct list_head
*list
= page
->list
.prev
;
860 return rb_list_head(list
->next
) != &page
->list
;
864 * rb_set_list_to_head - set a list_head to be pointing to head.
866 static void rb_set_list_to_head(struct ring_buffer_per_cpu
*cpu_buffer
,
867 struct list_head
*list
)
871 ptr
= (unsigned long *)&list
->next
;
872 *ptr
|= RB_PAGE_HEAD
;
873 *ptr
&= ~RB_PAGE_UPDATE
;
877 * rb_head_page_activate - sets up head page
879 static void rb_head_page_activate(struct ring_buffer_per_cpu
*cpu_buffer
)
881 struct buffer_page
*head
;
883 head
= cpu_buffer
->head_page
;
888 * Set the previous list pointer to have the HEAD flag.
890 rb_set_list_to_head(cpu_buffer
, head
->list
.prev
);
893 static void rb_list_head_clear(struct list_head
*list
)
895 unsigned long *ptr
= (unsigned long *)&list
->next
;
897 *ptr
&= ~RB_FLAG_MASK
;
901 * rb_head_page_dactivate - clears head page ptr (for free list)
904 rb_head_page_deactivate(struct ring_buffer_per_cpu
*cpu_buffer
)
906 struct list_head
*hd
;
908 /* Go through the whole list and clear any pointers found. */
909 rb_list_head_clear(cpu_buffer
->pages
);
911 list_for_each(hd
, cpu_buffer
->pages
)
912 rb_list_head_clear(hd
);
915 static int rb_head_page_set(struct ring_buffer_per_cpu
*cpu_buffer
,
916 struct buffer_page
*head
,
917 struct buffer_page
*prev
,
918 int old_flag
, int new_flag
)
920 struct list_head
*list
;
921 unsigned long val
= (unsigned long)&head
->list
;
926 val
&= ~RB_FLAG_MASK
;
928 ret
= cmpxchg((unsigned long *)&list
->next
,
929 val
| old_flag
, val
| new_flag
);
931 /* check if the reader took the page */
932 if ((ret
& ~RB_FLAG_MASK
) != val
)
933 return RB_PAGE_MOVED
;
935 return ret
& RB_FLAG_MASK
;
938 static int rb_head_page_set_update(struct ring_buffer_per_cpu
*cpu_buffer
,
939 struct buffer_page
*head
,
940 struct buffer_page
*prev
,
943 return rb_head_page_set(cpu_buffer
, head
, prev
,
944 old_flag
, RB_PAGE_UPDATE
);
947 static int rb_head_page_set_head(struct ring_buffer_per_cpu
*cpu_buffer
,
948 struct buffer_page
*head
,
949 struct buffer_page
*prev
,
952 return rb_head_page_set(cpu_buffer
, head
, prev
,
953 old_flag
, RB_PAGE_HEAD
);
956 static int rb_head_page_set_normal(struct ring_buffer_per_cpu
*cpu_buffer
,
957 struct buffer_page
*head
,
958 struct buffer_page
*prev
,
961 return rb_head_page_set(cpu_buffer
, head
, prev
,
962 old_flag
, RB_PAGE_NORMAL
);
965 static inline void rb_inc_page(struct ring_buffer_per_cpu
*cpu_buffer
,
966 struct buffer_page
**bpage
)
968 struct list_head
*p
= rb_list_head((*bpage
)->list
.next
);
970 *bpage
= list_entry(p
, struct buffer_page
, list
);
973 static struct buffer_page
*
974 rb_set_head_page(struct ring_buffer_per_cpu
*cpu_buffer
)
976 struct buffer_page
*head
;
977 struct buffer_page
*page
;
978 struct list_head
*list
;
981 if (RB_WARN_ON(cpu_buffer
, !cpu_buffer
->head_page
))
985 list
= cpu_buffer
->pages
;
986 if (RB_WARN_ON(cpu_buffer
, rb_list_head(list
->prev
->next
) != list
))
989 page
= head
= cpu_buffer
->head_page
;
991 * It is possible that the writer moves the header behind
992 * where we started, and we miss in one loop.
993 * A second loop should grab the header, but we'll do
994 * three loops just because I'm paranoid.
996 for (i
= 0; i
< 3; i
++) {
998 if (rb_is_head_page(cpu_buffer
, page
, page
->list
.prev
)) {
999 cpu_buffer
->head_page
= page
;
1002 rb_inc_page(cpu_buffer
, &page
);
1003 } while (page
!= head
);
1006 RB_WARN_ON(cpu_buffer
, 1);
1011 static int rb_head_page_replace(struct buffer_page
*old
,
1012 struct buffer_page
*new)
1014 unsigned long *ptr
= (unsigned long *)&old
->list
.prev
->next
;
1018 val
= *ptr
& ~RB_FLAG_MASK
;
1019 val
|= RB_PAGE_HEAD
;
1021 ret
= cmpxchg(ptr
, val
, (unsigned long)&new->list
);
1027 * rb_tail_page_update - move the tail page forward
1029 * Returns 1 if moved tail page, 0 if someone else did.
1031 static int rb_tail_page_update(struct ring_buffer_per_cpu
*cpu_buffer
,
1032 struct buffer_page
*tail_page
,
1033 struct buffer_page
*next_page
)
1035 struct buffer_page
*old_tail
;
1036 unsigned long old_entries
;
1037 unsigned long old_write
;
1041 * The tail page now needs to be moved forward.
1043 * We need to reset the tail page, but without messing
1044 * with possible erasing of data brought in by interrupts
1045 * that have moved the tail page and are currently on it.
1047 * We add a counter to the write field to denote this.
1049 old_write
= local_add_return(RB_WRITE_INTCNT
, &next_page
->write
);
1050 old_entries
= local_add_return(RB_WRITE_INTCNT
, &next_page
->entries
);
1053 * Just make sure we have seen our old_write and synchronize
1054 * with any interrupts that come in.
1059 * If the tail page is still the same as what we think
1060 * it is, then it is up to us to update the tail
1063 if (tail_page
== cpu_buffer
->tail_page
) {
1064 /* Zero the write counter */
1065 unsigned long val
= old_write
& ~RB_WRITE_MASK
;
1066 unsigned long eval
= old_entries
& ~RB_WRITE_MASK
;
1069 * This will only succeed if an interrupt did
1070 * not come in and change it. In which case, we
1071 * do not want to modify it.
1073 * We add (void) to let the compiler know that we do not care
1074 * about the return value of these functions. We use the
1075 * cmpxchg to only update if an interrupt did not already
1076 * do it for us. If the cmpxchg fails, we don't care.
1078 (void)local_cmpxchg(&next_page
->write
, old_write
, val
);
1079 (void)local_cmpxchg(&next_page
->entries
, old_entries
, eval
);
1082 * No need to worry about races with clearing out the commit.
1083 * it only can increment when a commit takes place. But that
1084 * only happens in the outer most nested commit.
1086 local_set(&next_page
->page
->commit
, 0);
1088 old_tail
= cmpxchg(&cpu_buffer
->tail_page
,
1089 tail_page
, next_page
);
1091 if (old_tail
== tail_page
)
1098 static int rb_check_bpage(struct ring_buffer_per_cpu
*cpu_buffer
,
1099 struct buffer_page
*bpage
)
1101 unsigned long val
= (unsigned long)bpage
;
1103 if (RB_WARN_ON(cpu_buffer
, val
& RB_FLAG_MASK
))
1110 * rb_check_list - make sure a pointer to a list has the last bits zero
1112 static int rb_check_list(struct ring_buffer_per_cpu
*cpu_buffer
,
1113 struct list_head
*list
)
1115 if (RB_WARN_ON(cpu_buffer
, rb_list_head(list
->prev
) != list
->prev
))
1117 if (RB_WARN_ON(cpu_buffer
, rb_list_head(list
->next
) != list
->next
))
1123 * rb_check_pages - integrity check of buffer pages
1124 * @cpu_buffer: CPU buffer with pages to test
1126 * As a safety measure we check to make sure the data pages have not
1129 static int rb_check_pages(struct ring_buffer_per_cpu
*cpu_buffer
)
1131 struct list_head
*head
= cpu_buffer
->pages
;
1132 struct buffer_page
*bpage
, *tmp
;
1134 /* Reset the head page if it exists */
1135 if (cpu_buffer
->head_page
)
1136 rb_set_head_page(cpu_buffer
);
1138 rb_head_page_deactivate(cpu_buffer
);
1140 if (RB_WARN_ON(cpu_buffer
, head
->next
->prev
!= head
))
1142 if (RB_WARN_ON(cpu_buffer
, head
->prev
->next
!= head
))
1145 if (rb_check_list(cpu_buffer
, head
))
1148 list_for_each_entry_safe(bpage
, tmp
, head
, list
) {
1149 if (RB_WARN_ON(cpu_buffer
,
1150 bpage
->list
.next
->prev
!= &bpage
->list
))
1152 if (RB_WARN_ON(cpu_buffer
,
1153 bpage
->list
.prev
->next
!= &bpage
->list
))
1155 if (rb_check_list(cpu_buffer
, &bpage
->list
))
1159 rb_head_page_activate(cpu_buffer
);
1164 static int __rb_allocate_pages(int nr_pages
, struct list_head
*pages
, int cpu
)
1167 struct buffer_page
*bpage
, *tmp
;
1169 for (i
= 0; i
< nr_pages
; i
++) {
1172 * __GFP_NORETRY flag makes sure that the allocation fails
1173 * gracefully without invoking oom-killer and the system is
1176 bpage
= kzalloc_node(ALIGN(sizeof(*bpage
), cache_line_size()),
1177 GFP_KERNEL
| __GFP_NORETRY
,
1182 list_add(&bpage
->list
, pages
);
1184 page
= alloc_pages_node(cpu_to_node(cpu
),
1185 GFP_KERNEL
| __GFP_NORETRY
, 0);
1188 bpage
->page
= page_address(page
);
1189 rb_init_page(bpage
->page
);
1195 list_for_each_entry_safe(bpage
, tmp
, pages
, list
) {
1196 list_del_init(&bpage
->list
);
1197 free_buffer_page(bpage
);
1203 static int rb_allocate_pages(struct ring_buffer_per_cpu
*cpu_buffer
,
1210 if (__rb_allocate_pages(nr_pages
, &pages
, cpu_buffer
->cpu
))
1214 * The ring buffer page list is a circular list that does not
1215 * start and end with a list head. All page list items point to
1218 cpu_buffer
->pages
= pages
.next
;
1221 cpu_buffer
->nr_pages
= nr_pages
;
1223 rb_check_pages(cpu_buffer
);
1228 static struct ring_buffer_per_cpu
*
1229 rb_allocate_cpu_buffer(struct ring_buffer
*buffer
, int nr_pages
, int cpu
)
1231 struct ring_buffer_per_cpu
*cpu_buffer
;
1232 struct buffer_page
*bpage
;
1236 cpu_buffer
= kzalloc_node(ALIGN(sizeof(*cpu_buffer
), cache_line_size()),
1237 GFP_KERNEL
, cpu_to_node(cpu
));
1241 cpu_buffer
->cpu
= cpu
;
1242 cpu_buffer
->buffer
= buffer
;
1243 raw_spin_lock_init(&cpu_buffer
->reader_lock
);
1244 lockdep_set_class(&cpu_buffer
->reader_lock
, buffer
->reader_lock_key
);
1245 cpu_buffer
->lock
= (arch_spinlock_t
)__ARCH_SPIN_LOCK_UNLOCKED
;
1246 INIT_WORK(&cpu_buffer
->update_pages_work
, update_pages_handler
);
1247 init_completion(&cpu_buffer
->update_done
);
1248 init_irq_work(&cpu_buffer
->irq_work
.work
, rb_wake_up_waiters
);
1249 init_waitqueue_head(&cpu_buffer
->irq_work
.waiters
);
1250 init_waitqueue_head(&cpu_buffer
->irq_work
.full_waiters
);
1252 bpage
= kzalloc_node(ALIGN(sizeof(*bpage
), cache_line_size()),
1253 GFP_KERNEL
, cpu_to_node(cpu
));
1255 goto fail_free_buffer
;
1257 rb_check_bpage(cpu_buffer
, bpage
);
1259 cpu_buffer
->reader_page
= bpage
;
1260 page
= alloc_pages_node(cpu_to_node(cpu
), GFP_KERNEL
, 0);
1262 goto fail_free_reader
;
1263 bpage
->page
= page_address(page
);
1264 rb_init_page(bpage
->page
);
1266 INIT_LIST_HEAD(&cpu_buffer
->reader_page
->list
);
1267 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
1269 ret
= rb_allocate_pages(cpu_buffer
, nr_pages
);
1271 goto fail_free_reader
;
1273 cpu_buffer
->head_page
1274 = list_entry(cpu_buffer
->pages
, struct buffer_page
, list
);
1275 cpu_buffer
->tail_page
= cpu_buffer
->commit_page
= cpu_buffer
->head_page
;
1277 rb_head_page_activate(cpu_buffer
);
1282 free_buffer_page(cpu_buffer
->reader_page
);
1289 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu
*cpu_buffer
)
1291 struct list_head
*head
= cpu_buffer
->pages
;
1292 struct buffer_page
*bpage
, *tmp
;
1294 free_buffer_page(cpu_buffer
->reader_page
);
1296 rb_head_page_deactivate(cpu_buffer
);
1299 list_for_each_entry_safe(bpage
, tmp
, head
, list
) {
1300 list_del_init(&bpage
->list
);
1301 free_buffer_page(bpage
);
1303 bpage
= list_entry(head
, struct buffer_page
, list
);
1304 free_buffer_page(bpage
);
1310 #ifdef CONFIG_HOTPLUG_CPU
1311 static int rb_cpu_notify(struct notifier_block
*self
,
1312 unsigned long action
, void *hcpu
);
1316 * __ring_buffer_alloc - allocate a new ring_buffer
1317 * @size: the size in bytes per cpu that is needed.
1318 * @flags: attributes to set for the ring buffer.
1320 * Currently the only flag that is available is the RB_FL_OVERWRITE
1321 * flag. This flag means that the buffer will overwrite old data
1322 * when the buffer wraps. If this flag is not set, the buffer will
1323 * drop data when the tail hits the head.
1325 struct ring_buffer
*__ring_buffer_alloc(unsigned long size
, unsigned flags
,
1326 struct lock_class_key
*key
)
1328 struct ring_buffer
*buffer
;
1332 /* keep it in its own cache line */
1333 buffer
= kzalloc(ALIGN(sizeof(*buffer
), cache_line_size()),
1338 if (!alloc_cpumask_var(&buffer
->cpumask
, GFP_KERNEL
))
1339 goto fail_free_buffer
;
1341 nr_pages
= DIV_ROUND_UP(size
, BUF_PAGE_SIZE
);
1342 buffer
->flags
= flags
;
1343 buffer
->clock
= trace_clock_local
;
1344 buffer
->reader_lock_key
= key
;
1346 init_irq_work(&buffer
->irq_work
.work
, rb_wake_up_waiters
);
1347 init_waitqueue_head(&buffer
->irq_work
.waiters
);
1349 /* need at least two pages */
1354 * In case of non-hotplug cpu, if the ring-buffer is allocated
1355 * in early initcall, it will not be notified of secondary cpus.
1356 * In that off case, we need to allocate for all possible cpus.
1358 #ifdef CONFIG_HOTPLUG_CPU
1359 cpu_notifier_register_begin();
1360 cpumask_copy(buffer
->cpumask
, cpu_online_mask
);
1362 cpumask_copy(buffer
->cpumask
, cpu_possible_mask
);
1364 buffer
->cpus
= nr_cpu_ids
;
1366 bsize
= sizeof(void *) * nr_cpu_ids
;
1367 buffer
->buffers
= kzalloc(ALIGN(bsize
, cache_line_size()),
1369 if (!buffer
->buffers
)
1370 goto fail_free_cpumask
;
1372 for_each_buffer_cpu(buffer
, cpu
) {
1373 buffer
->buffers
[cpu
] =
1374 rb_allocate_cpu_buffer(buffer
, nr_pages
, cpu
);
1375 if (!buffer
->buffers
[cpu
])
1376 goto fail_free_buffers
;
1379 #ifdef CONFIG_HOTPLUG_CPU
1380 buffer
->cpu_notify
.notifier_call
= rb_cpu_notify
;
1381 buffer
->cpu_notify
.priority
= 0;
1382 __register_cpu_notifier(&buffer
->cpu_notify
);
1383 cpu_notifier_register_done();
1386 mutex_init(&buffer
->mutex
);
1391 for_each_buffer_cpu(buffer
, cpu
) {
1392 if (buffer
->buffers
[cpu
])
1393 rb_free_cpu_buffer(buffer
->buffers
[cpu
]);
1395 kfree(buffer
->buffers
);
1398 free_cpumask_var(buffer
->cpumask
);
1399 #ifdef CONFIG_HOTPLUG_CPU
1400 cpu_notifier_register_done();
1407 EXPORT_SYMBOL_GPL(__ring_buffer_alloc
);
1410 * ring_buffer_free - free a ring buffer.
1411 * @buffer: the buffer to free.
1414 ring_buffer_free(struct ring_buffer
*buffer
)
1418 #ifdef CONFIG_HOTPLUG_CPU
1419 cpu_notifier_register_begin();
1420 __unregister_cpu_notifier(&buffer
->cpu_notify
);
1423 for_each_buffer_cpu(buffer
, cpu
)
1424 rb_free_cpu_buffer(buffer
->buffers
[cpu
]);
1426 #ifdef CONFIG_HOTPLUG_CPU
1427 cpu_notifier_register_done();
1430 kfree(buffer
->buffers
);
1431 free_cpumask_var(buffer
->cpumask
);
1435 EXPORT_SYMBOL_GPL(ring_buffer_free
);
1437 void ring_buffer_set_clock(struct ring_buffer
*buffer
,
1440 buffer
->clock
= clock
;
1443 static void rb_reset_cpu(struct ring_buffer_per_cpu
*cpu_buffer
);
1445 static inline unsigned long rb_page_entries(struct buffer_page
*bpage
)
1447 return local_read(&bpage
->entries
) & RB_WRITE_MASK
;
1450 static inline unsigned long rb_page_write(struct buffer_page
*bpage
)
1452 return local_read(&bpage
->write
) & RB_WRITE_MASK
;
1456 rb_remove_pages(struct ring_buffer_per_cpu
*cpu_buffer
, unsigned int nr_pages
)
1458 struct list_head
*tail_page
, *to_remove
, *next_page
;
1459 struct buffer_page
*to_remove_page
, *tmp_iter_page
;
1460 struct buffer_page
*last_page
, *first_page
;
1461 unsigned int nr_removed
;
1462 unsigned long head_bit
;
1467 raw_spin_lock_irq(&cpu_buffer
->reader_lock
);
1468 atomic_inc(&cpu_buffer
->record_disabled
);
1470 * We don't race with the readers since we have acquired the reader
1471 * lock. We also don't race with writers after disabling recording.
1472 * This makes it easy to figure out the first and the last page to be
1473 * removed from the list. We unlink all the pages in between including
1474 * the first and last pages. This is done in a busy loop so that we
1475 * lose the least number of traces.
1476 * The pages are freed after we restart recording and unlock readers.
1478 tail_page
= &cpu_buffer
->tail_page
->list
;
1481 * tail page might be on reader page, we remove the next page
1482 * from the ring buffer
1484 if (cpu_buffer
->tail_page
== cpu_buffer
->reader_page
)
1485 tail_page
= rb_list_head(tail_page
->next
);
1486 to_remove
= tail_page
;
1488 /* start of pages to remove */
1489 first_page
= list_entry(rb_list_head(to_remove
->next
),
1490 struct buffer_page
, list
);
1492 for (nr_removed
= 0; nr_removed
< nr_pages
; nr_removed
++) {
1493 to_remove
= rb_list_head(to_remove
)->next
;
1494 head_bit
|= (unsigned long)to_remove
& RB_PAGE_HEAD
;
1497 next_page
= rb_list_head(to_remove
)->next
;
1500 * Now we remove all pages between tail_page and next_page.
1501 * Make sure that we have head_bit value preserved for the
1504 tail_page
->next
= (struct list_head
*)((unsigned long)next_page
|
1506 next_page
= rb_list_head(next_page
);
1507 next_page
->prev
= tail_page
;
1509 /* make sure pages points to a valid page in the ring buffer */
1510 cpu_buffer
->pages
= next_page
;
1512 /* update head page */
1514 cpu_buffer
->head_page
= list_entry(next_page
,
1515 struct buffer_page
, list
);
1518 * change read pointer to make sure any read iterators reset
1521 cpu_buffer
->read
= 0;
1523 /* pages are removed, resume tracing and then free the pages */
1524 atomic_dec(&cpu_buffer
->record_disabled
);
1525 raw_spin_unlock_irq(&cpu_buffer
->reader_lock
);
1527 RB_WARN_ON(cpu_buffer
, list_empty(cpu_buffer
->pages
));
1529 /* last buffer page to remove */
1530 last_page
= list_entry(rb_list_head(to_remove
), struct buffer_page
,
1532 tmp_iter_page
= first_page
;
1535 to_remove_page
= tmp_iter_page
;
1536 rb_inc_page(cpu_buffer
, &tmp_iter_page
);
1538 /* update the counters */
1539 page_entries
= rb_page_entries(to_remove_page
);
1542 * If something was added to this page, it was full
1543 * since it is not the tail page. So we deduct the
1544 * bytes consumed in ring buffer from here.
1545 * Increment overrun to account for the lost events.
1547 local_add(page_entries
, &cpu_buffer
->overrun
);
1548 local_sub(BUF_PAGE_SIZE
, &cpu_buffer
->entries_bytes
);
1552 * We have already removed references to this list item, just
1553 * free up the buffer_page and its page
1555 free_buffer_page(to_remove_page
);
1558 } while (to_remove_page
!= last_page
);
1560 RB_WARN_ON(cpu_buffer
, nr_removed
);
1562 return nr_removed
== 0;
1566 rb_insert_pages(struct ring_buffer_per_cpu
*cpu_buffer
)
1568 struct list_head
*pages
= &cpu_buffer
->new_pages
;
1569 int retries
, success
;
1571 raw_spin_lock_irq(&cpu_buffer
->reader_lock
);
1573 * We are holding the reader lock, so the reader page won't be swapped
1574 * in the ring buffer. Now we are racing with the writer trying to
1575 * move head page and the tail page.
1576 * We are going to adapt the reader page update process where:
1577 * 1. We first splice the start and end of list of new pages between
1578 * the head page and its previous page.
1579 * 2. We cmpxchg the prev_page->next to point from head page to the
1580 * start of new pages list.
1581 * 3. Finally, we update the head->prev to the end of new list.
1583 * We will try this process 10 times, to make sure that we don't keep
1589 struct list_head
*head_page
, *prev_page
, *r
;
1590 struct list_head
*last_page
, *first_page
;
1591 struct list_head
*head_page_with_bit
;
1593 head_page
= &rb_set_head_page(cpu_buffer
)->list
;
1596 prev_page
= head_page
->prev
;
1598 first_page
= pages
->next
;
1599 last_page
= pages
->prev
;
1601 head_page_with_bit
= (struct list_head
*)
1602 ((unsigned long)head_page
| RB_PAGE_HEAD
);
1604 last_page
->next
= head_page_with_bit
;
1605 first_page
->prev
= prev_page
;
1607 r
= cmpxchg(&prev_page
->next
, head_page_with_bit
, first_page
);
1609 if (r
== head_page_with_bit
) {
1611 * yay, we replaced the page pointer to our new list,
1612 * now, we just have to update to head page's prev
1613 * pointer to point to end of list
1615 head_page
->prev
= last_page
;
1622 INIT_LIST_HEAD(pages
);
1624 * If we weren't successful in adding in new pages, warn and stop
1627 RB_WARN_ON(cpu_buffer
, !success
);
1628 raw_spin_unlock_irq(&cpu_buffer
->reader_lock
);
1630 /* free pages if they weren't inserted */
1632 struct buffer_page
*bpage
, *tmp
;
1633 list_for_each_entry_safe(bpage
, tmp
, &cpu_buffer
->new_pages
,
1635 list_del_init(&bpage
->list
);
1636 free_buffer_page(bpage
);
1642 static void rb_update_pages(struct ring_buffer_per_cpu
*cpu_buffer
)
1646 if (cpu_buffer
->nr_pages_to_update
> 0)
1647 success
= rb_insert_pages(cpu_buffer
);
1649 success
= rb_remove_pages(cpu_buffer
,
1650 -cpu_buffer
->nr_pages_to_update
);
1653 cpu_buffer
->nr_pages
+= cpu_buffer
->nr_pages_to_update
;
1656 static void update_pages_handler(struct work_struct
*work
)
1658 struct ring_buffer_per_cpu
*cpu_buffer
= container_of(work
,
1659 struct ring_buffer_per_cpu
, update_pages_work
);
1660 rb_update_pages(cpu_buffer
);
1661 complete(&cpu_buffer
->update_done
);
1665 * ring_buffer_resize - resize the ring buffer
1666 * @buffer: the buffer to resize.
1667 * @size: the new size.
1668 * @cpu_id: the cpu buffer to resize
1670 * Minimum size is 2 * BUF_PAGE_SIZE.
1672 * Returns 0 on success and < 0 on failure.
1674 int ring_buffer_resize(struct ring_buffer
*buffer
, unsigned long size
,
1677 struct ring_buffer_per_cpu
*cpu_buffer
;
1682 * Always succeed at resizing a non-existent buffer:
1687 /* Make sure the requested buffer exists */
1688 if (cpu_id
!= RING_BUFFER_ALL_CPUS
&&
1689 !cpumask_test_cpu(cpu_id
, buffer
->cpumask
))
1692 size
= DIV_ROUND_UP(size
, BUF_PAGE_SIZE
);
1693 size
*= BUF_PAGE_SIZE
;
1695 /* we need a minimum of two pages */
1696 if (size
< BUF_PAGE_SIZE
* 2)
1697 size
= BUF_PAGE_SIZE
* 2;
1699 nr_pages
= DIV_ROUND_UP(size
, BUF_PAGE_SIZE
);
1702 * Don't succeed if resizing is disabled, as a reader might be
1703 * manipulating the ring buffer and is expecting a sane state while
1706 if (atomic_read(&buffer
->resize_disabled
))
1709 /* prevent another thread from changing buffer sizes */
1710 mutex_lock(&buffer
->mutex
);
1712 if (cpu_id
== RING_BUFFER_ALL_CPUS
) {
1713 /* calculate the pages to update */
1714 for_each_buffer_cpu(buffer
, cpu
) {
1715 cpu_buffer
= buffer
->buffers
[cpu
];
1717 cpu_buffer
->nr_pages_to_update
= nr_pages
-
1718 cpu_buffer
->nr_pages
;
1720 * nothing more to do for removing pages or no update
1722 if (cpu_buffer
->nr_pages_to_update
<= 0)
1725 * to add pages, make sure all new pages can be
1726 * allocated without receiving ENOMEM
1728 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
1729 if (__rb_allocate_pages(cpu_buffer
->nr_pages_to_update
,
1730 &cpu_buffer
->new_pages
, cpu
)) {
1731 /* not enough memory for new pages */
1739 * Fire off all the required work handlers
1740 * We can't schedule on offline CPUs, but it's not necessary
1741 * since we can change their buffer sizes without any race.
1743 for_each_buffer_cpu(buffer
, cpu
) {
1744 cpu_buffer
= buffer
->buffers
[cpu
];
1745 if (!cpu_buffer
->nr_pages_to_update
)
1748 /* Can't run something on an offline CPU. */
1749 if (!cpu_online(cpu
)) {
1750 rb_update_pages(cpu_buffer
);
1751 cpu_buffer
->nr_pages_to_update
= 0;
1753 schedule_work_on(cpu
,
1754 &cpu_buffer
->update_pages_work
);
1758 /* wait for all the updates to complete */
1759 for_each_buffer_cpu(buffer
, cpu
) {
1760 cpu_buffer
= buffer
->buffers
[cpu
];
1761 if (!cpu_buffer
->nr_pages_to_update
)
1764 if (cpu_online(cpu
))
1765 wait_for_completion(&cpu_buffer
->update_done
);
1766 cpu_buffer
->nr_pages_to_update
= 0;
1771 /* Make sure this CPU has been intitialized */
1772 if (!cpumask_test_cpu(cpu_id
, buffer
->cpumask
))
1775 cpu_buffer
= buffer
->buffers
[cpu_id
];
1777 if (nr_pages
== cpu_buffer
->nr_pages
)
1780 cpu_buffer
->nr_pages_to_update
= nr_pages
-
1781 cpu_buffer
->nr_pages
;
1783 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
1784 if (cpu_buffer
->nr_pages_to_update
> 0 &&
1785 __rb_allocate_pages(cpu_buffer
->nr_pages_to_update
,
1786 &cpu_buffer
->new_pages
, cpu_id
)) {
1793 /* Can't run something on an offline CPU. */
1794 if (!cpu_online(cpu_id
))
1795 rb_update_pages(cpu_buffer
);
1797 schedule_work_on(cpu_id
,
1798 &cpu_buffer
->update_pages_work
);
1799 wait_for_completion(&cpu_buffer
->update_done
);
1802 cpu_buffer
->nr_pages_to_update
= 0;
1808 * The ring buffer resize can happen with the ring buffer
1809 * enabled, so that the update disturbs the tracing as little
1810 * as possible. But if the buffer is disabled, we do not need
1811 * to worry about that, and we can take the time to verify
1812 * that the buffer is not corrupt.
1814 if (atomic_read(&buffer
->record_disabled
)) {
1815 atomic_inc(&buffer
->record_disabled
);
1817 * Even though the buffer was disabled, we must make sure
1818 * that it is truly disabled before calling rb_check_pages.
1819 * There could have been a race between checking
1820 * record_disable and incrementing it.
1822 synchronize_sched();
1823 for_each_buffer_cpu(buffer
, cpu
) {
1824 cpu_buffer
= buffer
->buffers
[cpu
];
1825 rb_check_pages(cpu_buffer
);
1827 atomic_dec(&buffer
->record_disabled
);
1830 mutex_unlock(&buffer
->mutex
);
1834 for_each_buffer_cpu(buffer
, cpu
) {
1835 struct buffer_page
*bpage
, *tmp
;
1837 cpu_buffer
= buffer
->buffers
[cpu
];
1838 cpu_buffer
->nr_pages_to_update
= 0;
1840 if (list_empty(&cpu_buffer
->new_pages
))
1843 list_for_each_entry_safe(bpage
, tmp
, &cpu_buffer
->new_pages
,
1845 list_del_init(&bpage
->list
);
1846 free_buffer_page(bpage
);
1849 mutex_unlock(&buffer
->mutex
);
1852 EXPORT_SYMBOL_GPL(ring_buffer_resize
);
1854 void ring_buffer_change_overwrite(struct ring_buffer
*buffer
, int val
)
1856 mutex_lock(&buffer
->mutex
);
1858 buffer
->flags
|= RB_FL_OVERWRITE
;
1860 buffer
->flags
&= ~RB_FL_OVERWRITE
;
1861 mutex_unlock(&buffer
->mutex
);
1863 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite
);
1865 static inline void *
1866 __rb_data_page_index(struct buffer_data_page
*bpage
, unsigned index
)
1868 return bpage
->data
+ index
;
1871 static inline void *__rb_page_index(struct buffer_page
*bpage
, unsigned index
)
1873 return bpage
->page
->data
+ index
;
1876 static inline struct ring_buffer_event
*
1877 rb_reader_event(struct ring_buffer_per_cpu
*cpu_buffer
)
1879 return __rb_page_index(cpu_buffer
->reader_page
,
1880 cpu_buffer
->reader_page
->read
);
1883 static inline struct ring_buffer_event
*
1884 rb_iter_head_event(struct ring_buffer_iter
*iter
)
1886 return __rb_page_index(iter
->head_page
, iter
->head
);
1889 static inline unsigned rb_page_commit(struct buffer_page
*bpage
)
1891 return local_read(&bpage
->page
->commit
);
1894 /* Size is determined by what has been committed */
1895 static inline unsigned rb_page_size(struct buffer_page
*bpage
)
1897 return rb_page_commit(bpage
);
1900 static inline unsigned
1901 rb_commit_index(struct ring_buffer_per_cpu
*cpu_buffer
)
1903 return rb_page_commit(cpu_buffer
->commit_page
);
1906 static inline unsigned
1907 rb_event_index(struct ring_buffer_event
*event
)
1909 unsigned long addr
= (unsigned long)event
;
1911 return (addr
& ~PAGE_MASK
) - BUF_PAGE_HDR_SIZE
;
1915 rb_event_is_commit(struct ring_buffer_per_cpu
*cpu_buffer
,
1916 struct ring_buffer_event
*event
)
1918 unsigned long addr
= (unsigned long)event
;
1919 unsigned long index
;
1921 index
= rb_event_index(event
);
1924 return cpu_buffer
->commit_page
->page
== (void *)addr
&&
1925 rb_commit_index(cpu_buffer
) == index
;
1929 rb_set_commit_to_write(struct ring_buffer_per_cpu
*cpu_buffer
)
1931 unsigned long max_count
;
1934 * We only race with interrupts and NMIs on this CPU.
1935 * If we own the commit event, then we can commit
1936 * all others that interrupted us, since the interruptions
1937 * are in stack format (they finish before they come
1938 * back to us). This allows us to do a simple loop to
1939 * assign the commit to the tail.
1942 max_count
= cpu_buffer
->nr_pages
* 100;
1944 while (cpu_buffer
->commit_page
!= cpu_buffer
->tail_page
) {
1945 if (RB_WARN_ON(cpu_buffer
, !(--max_count
)))
1947 if (RB_WARN_ON(cpu_buffer
,
1948 rb_is_reader_page(cpu_buffer
->tail_page
)))
1950 local_set(&cpu_buffer
->commit_page
->page
->commit
,
1951 rb_page_write(cpu_buffer
->commit_page
));
1952 rb_inc_page(cpu_buffer
, &cpu_buffer
->commit_page
);
1953 cpu_buffer
->write_stamp
=
1954 cpu_buffer
->commit_page
->page
->time_stamp
;
1955 /* add barrier to keep gcc from optimizing too much */
1958 while (rb_commit_index(cpu_buffer
) !=
1959 rb_page_write(cpu_buffer
->commit_page
)) {
1961 local_set(&cpu_buffer
->commit_page
->page
->commit
,
1962 rb_page_write(cpu_buffer
->commit_page
));
1963 RB_WARN_ON(cpu_buffer
,
1964 local_read(&cpu_buffer
->commit_page
->page
->commit
) &
1969 /* again, keep gcc from optimizing */
1973 * If an interrupt came in just after the first while loop
1974 * and pushed the tail page forward, we will be left with
1975 * a dangling commit that will never go forward.
1977 if (unlikely(cpu_buffer
->commit_page
!= cpu_buffer
->tail_page
))
1981 static void rb_reset_reader_page(struct ring_buffer_per_cpu
*cpu_buffer
)
1983 cpu_buffer
->read_stamp
= cpu_buffer
->reader_page
->page
->time_stamp
;
1984 cpu_buffer
->reader_page
->read
= 0;
1987 static void rb_inc_iter(struct ring_buffer_iter
*iter
)
1989 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
1992 * The iterator could be on the reader page (it starts there).
1993 * But the head could have moved, since the reader was
1994 * found. Check for this case and assign the iterator
1995 * to the head page instead of next.
1997 if (iter
->head_page
== cpu_buffer
->reader_page
)
1998 iter
->head_page
= rb_set_head_page(cpu_buffer
);
2000 rb_inc_page(cpu_buffer
, &iter
->head_page
);
2002 iter
->read_stamp
= iter
->head_page
->page
->time_stamp
;
2006 /* Slow path, do not inline */
2007 static noinline
struct ring_buffer_event
*
2008 rb_add_time_stamp(struct ring_buffer_event
*event
, u64 delta
)
2010 event
->type_len
= RINGBUF_TYPE_TIME_EXTEND
;
2012 /* Not the first event on the page? */
2013 if (rb_event_index(event
)) {
2014 event
->time_delta
= delta
& TS_MASK
;
2015 event
->array
[0] = delta
>> TS_SHIFT
;
2017 /* nope, just zero it */
2018 event
->time_delta
= 0;
2019 event
->array
[0] = 0;
2022 return skip_time_extend(event
);
2026 * rb_update_event - update event type and data
2027 * @event: the event to update
2028 * @type: the type of event
2029 * @length: the size of the event field in the ring buffer
2031 * Update the type and data fields of the event. The length
2032 * is the actual size that is written to the ring buffer,
2033 * and with this, we can determine what to place into the
2037 rb_update_event(struct ring_buffer_per_cpu
*cpu_buffer
,
2038 struct ring_buffer_event
*event
, unsigned length
,
2039 int add_timestamp
, u64 delta
)
2041 /* Only a commit updates the timestamp */
2042 if (unlikely(!rb_event_is_commit(cpu_buffer
, event
)))
2046 * If we need to add a timestamp, then we
2047 * add it to the start of the resevered space.
2049 if (unlikely(add_timestamp
)) {
2050 event
= rb_add_time_stamp(event
, delta
);
2051 length
-= RB_LEN_TIME_EXTEND
;
2055 event
->time_delta
= delta
;
2056 length
-= RB_EVNT_HDR_SIZE
;
2057 if (length
> RB_MAX_SMALL_DATA
|| RB_FORCE_8BYTE_ALIGNMENT
) {
2058 event
->type_len
= 0;
2059 event
->array
[0] = length
;
2061 event
->type_len
= DIV_ROUND_UP(length
, RB_ALIGNMENT
);
2065 * rb_handle_head_page - writer hit the head page
2067 * Returns: +1 to retry page
2072 rb_handle_head_page(struct ring_buffer_per_cpu
*cpu_buffer
,
2073 struct buffer_page
*tail_page
,
2074 struct buffer_page
*next_page
)
2076 struct buffer_page
*new_head
;
2081 entries
= rb_page_entries(next_page
);
2084 * The hard part is here. We need to move the head
2085 * forward, and protect against both readers on
2086 * other CPUs and writers coming in via interrupts.
2088 type
= rb_head_page_set_update(cpu_buffer
, next_page
, tail_page
,
2092 * type can be one of four:
2093 * NORMAL - an interrupt already moved it for us
2094 * HEAD - we are the first to get here.
2095 * UPDATE - we are the interrupt interrupting
2097 * MOVED - a reader on another CPU moved the next
2098 * pointer to its reader page. Give up
2105 * We changed the head to UPDATE, thus
2106 * it is our responsibility to update
2109 local_add(entries
, &cpu_buffer
->overrun
);
2110 local_sub(BUF_PAGE_SIZE
, &cpu_buffer
->entries_bytes
);
2113 * The entries will be zeroed out when we move the
2117 /* still more to do */
2120 case RB_PAGE_UPDATE
:
2122 * This is an interrupt that interrupt the
2123 * previous update. Still more to do.
2126 case RB_PAGE_NORMAL
:
2128 * An interrupt came in before the update
2129 * and processed this for us.
2130 * Nothing left to do.
2135 * The reader is on another CPU and just did
2136 * a swap with our next_page.
2141 RB_WARN_ON(cpu_buffer
, 1); /* WTF??? */
2146 * Now that we are here, the old head pointer is
2147 * set to UPDATE. This will keep the reader from
2148 * swapping the head page with the reader page.
2149 * The reader (on another CPU) will spin till
2152 * We just need to protect against interrupts
2153 * doing the job. We will set the next pointer
2154 * to HEAD. After that, we set the old pointer
2155 * to NORMAL, but only if it was HEAD before.
2156 * otherwise we are an interrupt, and only
2157 * want the outer most commit to reset it.
2159 new_head
= next_page
;
2160 rb_inc_page(cpu_buffer
, &new_head
);
2162 ret
= rb_head_page_set_head(cpu_buffer
, new_head
, next_page
,
2166 * Valid returns are:
2167 * HEAD - an interrupt came in and already set it.
2168 * NORMAL - One of two things:
2169 * 1) We really set it.
2170 * 2) A bunch of interrupts came in and moved
2171 * the page forward again.
2175 case RB_PAGE_NORMAL
:
2179 RB_WARN_ON(cpu_buffer
, 1);
2184 * It is possible that an interrupt came in,
2185 * set the head up, then more interrupts came in
2186 * and moved it again. When we get back here,
2187 * the page would have been set to NORMAL but we
2188 * just set it back to HEAD.
2190 * How do you detect this? Well, if that happened
2191 * the tail page would have moved.
2193 if (ret
== RB_PAGE_NORMAL
) {
2195 * If the tail had moved passed next, then we need
2196 * to reset the pointer.
2198 if (cpu_buffer
->tail_page
!= tail_page
&&
2199 cpu_buffer
->tail_page
!= next_page
)
2200 rb_head_page_set_normal(cpu_buffer
, new_head
,
2206 * If this was the outer most commit (the one that
2207 * changed the original pointer from HEAD to UPDATE),
2208 * then it is up to us to reset it to NORMAL.
2210 if (type
== RB_PAGE_HEAD
) {
2211 ret
= rb_head_page_set_normal(cpu_buffer
, next_page
,
2214 if (RB_WARN_ON(cpu_buffer
,
2215 ret
!= RB_PAGE_UPDATE
))
2222 static unsigned rb_calculate_event_length(unsigned length
)
2224 struct ring_buffer_event event
; /* Used only for sizeof array */
2226 /* zero length can cause confusions */
2230 if (length
> RB_MAX_SMALL_DATA
|| RB_FORCE_8BYTE_ALIGNMENT
)
2231 length
+= sizeof(event
.array
[0]);
2233 length
+= RB_EVNT_HDR_SIZE
;
2234 length
= ALIGN(length
, RB_ARCH_ALIGNMENT
);
2240 rb_reset_tail(struct ring_buffer_per_cpu
*cpu_buffer
,
2241 struct buffer_page
*tail_page
,
2242 unsigned long tail
, unsigned long length
)
2244 struct ring_buffer_event
*event
;
2247 * Only the event that crossed the page boundary
2248 * must fill the old tail_page with padding.
2250 if (tail
>= BUF_PAGE_SIZE
) {
2252 * If the page was filled, then we still need
2253 * to update the real_end. Reset it to zero
2254 * and the reader will ignore it.
2256 if (tail
== BUF_PAGE_SIZE
)
2257 tail_page
->real_end
= 0;
2259 local_sub(length
, &tail_page
->write
);
2263 event
= __rb_page_index(tail_page
, tail
);
2264 kmemcheck_annotate_bitfield(event
, bitfield
);
2266 /* account for padding bytes */
2267 local_add(BUF_PAGE_SIZE
- tail
, &cpu_buffer
->entries_bytes
);
2270 * Save the original length to the meta data.
2271 * This will be used by the reader to add lost event
2274 tail_page
->real_end
= tail
;
2277 * If this event is bigger than the minimum size, then
2278 * we need to be careful that we don't subtract the
2279 * write counter enough to allow another writer to slip
2281 * We put in a discarded commit instead, to make sure
2282 * that this space is not used again.
2284 * If we are less than the minimum size, we don't need to
2287 if (tail
> (BUF_PAGE_SIZE
- RB_EVNT_MIN_SIZE
)) {
2288 /* No room for any events */
2290 /* Mark the rest of the page with padding */
2291 rb_event_set_padding(event
);
2293 /* Set the write back to the previous setting */
2294 local_sub(length
, &tail_page
->write
);
2298 /* Put in a discarded event */
2299 event
->array
[0] = (BUF_PAGE_SIZE
- tail
) - RB_EVNT_HDR_SIZE
;
2300 event
->type_len
= RINGBUF_TYPE_PADDING
;
2301 /* time delta must be non zero */
2302 event
->time_delta
= 1;
2304 /* Set write to end of buffer */
2305 length
= (tail
+ length
) - BUF_PAGE_SIZE
;
2306 local_sub(length
, &tail_page
->write
);
2310 * This is the slow path, force gcc not to inline it.
2312 static noinline
struct ring_buffer_event
*
2313 rb_move_tail(struct ring_buffer_per_cpu
*cpu_buffer
,
2314 unsigned long length
, unsigned long tail
,
2315 struct buffer_page
*tail_page
, u64 ts
)
2317 struct buffer_page
*commit_page
= cpu_buffer
->commit_page
;
2318 struct ring_buffer
*buffer
= cpu_buffer
->buffer
;
2319 struct buffer_page
*next_page
;
2322 next_page
= tail_page
;
2324 rb_inc_page(cpu_buffer
, &next_page
);
2327 * If for some reason, we had an interrupt storm that made
2328 * it all the way around the buffer, bail, and warn
2331 if (unlikely(next_page
== commit_page
)) {
2332 local_inc(&cpu_buffer
->commit_overrun
);
2337 * This is where the fun begins!
2339 * We are fighting against races between a reader that
2340 * could be on another CPU trying to swap its reader
2341 * page with the buffer head.
2343 * We are also fighting against interrupts coming in and
2344 * moving the head or tail on us as well.
2346 * If the next page is the head page then we have filled
2347 * the buffer, unless the commit page is still on the
2350 if (rb_is_head_page(cpu_buffer
, next_page
, &tail_page
->list
)) {
2353 * If the commit is not on the reader page, then
2354 * move the header page.
2356 if (!rb_is_reader_page(cpu_buffer
->commit_page
)) {
2358 * If we are not in overwrite mode,
2359 * this is easy, just stop here.
2361 if (!(buffer
->flags
& RB_FL_OVERWRITE
)) {
2362 local_inc(&cpu_buffer
->dropped_events
);
2366 ret
= rb_handle_head_page(cpu_buffer
,
2375 * We need to be careful here too. The
2376 * commit page could still be on the reader
2377 * page. We could have a small buffer, and
2378 * have filled up the buffer with events
2379 * from interrupts and such, and wrapped.
2381 * Note, if the tail page is also the on the
2382 * reader_page, we let it move out.
2384 if (unlikely((cpu_buffer
->commit_page
!=
2385 cpu_buffer
->tail_page
) &&
2386 (cpu_buffer
->commit_page
==
2387 cpu_buffer
->reader_page
))) {
2388 local_inc(&cpu_buffer
->commit_overrun
);
2394 ret
= rb_tail_page_update(cpu_buffer
, tail_page
, next_page
);
2397 * Nested commits always have zero deltas, so
2398 * just reread the time stamp
2400 ts
= rb_time_stamp(buffer
);
2401 next_page
->page
->time_stamp
= ts
;
2406 rb_reset_tail(cpu_buffer
, tail_page
, tail
, length
);
2408 /* fail and let the caller try again */
2409 return ERR_PTR(-EAGAIN
);
2413 rb_reset_tail(cpu_buffer
, tail_page
, tail
, length
);
2418 static struct ring_buffer_event
*
2419 __rb_reserve_next(struct ring_buffer_per_cpu
*cpu_buffer
,
2420 unsigned long length
, u64 ts
,
2421 u64 delta
, int add_timestamp
)
2423 struct buffer_page
*tail_page
;
2424 struct ring_buffer_event
*event
;
2425 unsigned long tail
, write
;
2428 * If the time delta since the last event is too big to
2429 * hold in the time field of the event, then we append a
2430 * TIME EXTEND event ahead of the data event.
2432 if (unlikely(add_timestamp
))
2433 length
+= RB_LEN_TIME_EXTEND
;
2435 tail_page
= cpu_buffer
->tail_page
;
2436 write
= local_add_return(length
, &tail_page
->write
);
2438 /* set write to only the index of the write */
2439 write
&= RB_WRITE_MASK
;
2440 tail
= write
- length
;
2443 * If this is the first commit on the page, then it has the same
2444 * timestamp as the page itself.
2449 /* See if we shot pass the end of this buffer page */
2450 if (unlikely(write
> BUF_PAGE_SIZE
))
2451 return rb_move_tail(cpu_buffer
, length
, tail
,
2454 /* We reserved something on the buffer */
2456 event
= __rb_page_index(tail_page
, tail
);
2457 kmemcheck_annotate_bitfield(event
, bitfield
);
2458 rb_update_event(cpu_buffer
, event
, length
, add_timestamp
, delta
);
2460 local_inc(&tail_page
->entries
);
2463 * If this is the first commit on the page, then update
2467 tail_page
->page
->time_stamp
= ts
;
2469 /* account for these added bytes */
2470 local_add(length
, &cpu_buffer
->entries_bytes
);
2476 rb_try_to_discard(struct ring_buffer_per_cpu
*cpu_buffer
,
2477 struct ring_buffer_event
*event
)
2479 unsigned long new_index
, old_index
;
2480 struct buffer_page
*bpage
;
2481 unsigned long index
;
2484 new_index
= rb_event_index(event
);
2485 old_index
= new_index
+ rb_event_ts_length(event
);
2486 addr
= (unsigned long)event
;
2489 bpage
= cpu_buffer
->tail_page
;
2491 if (bpage
->page
== (void *)addr
&& rb_page_write(bpage
) == old_index
) {
2492 unsigned long write_mask
=
2493 local_read(&bpage
->write
) & ~RB_WRITE_MASK
;
2494 unsigned long event_length
= rb_event_length(event
);
2496 * This is on the tail page. It is possible that
2497 * a write could come in and move the tail page
2498 * and write to the next page. That is fine
2499 * because we just shorten what is on this page.
2501 old_index
+= write_mask
;
2502 new_index
+= write_mask
;
2503 index
= local_cmpxchg(&bpage
->write
, old_index
, new_index
);
2504 if (index
== old_index
) {
2505 /* update counters */
2506 local_sub(event_length
, &cpu_buffer
->entries_bytes
);
2511 /* could not discard */
2515 static void rb_start_commit(struct ring_buffer_per_cpu
*cpu_buffer
)
2517 local_inc(&cpu_buffer
->committing
);
2518 local_inc(&cpu_buffer
->commits
);
2521 static inline void rb_end_commit(struct ring_buffer_per_cpu
*cpu_buffer
)
2523 unsigned long commits
;
2525 if (RB_WARN_ON(cpu_buffer
,
2526 !local_read(&cpu_buffer
->committing
)))
2530 commits
= local_read(&cpu_buffer
->commits
);
2531 /* synchronize with interrupts */
2533 if (local_read(&cpu_buffer
->committing
) == 1)
2534 rb_set_commit_to_write(cpu_buffer
);
2536 local_dec(&cpu_buffer
->committing
);
2538 /* synchronize with interrupts */
2542 * Need to account for interrupts coming in between the
2543 * updating of the commit page and the clearing of the
2544 * committing counter.
2546 if (unlikely(local_read(&cpu_buffer
->commits
) != commits
) &&
2547 !local_read(&cpu_buffer
->committing
)) {
2548 local_inc(&cpu_buffer
->committing
);
2553 static struct ring_buffer_event
*
2554 rb_reserve_next_event(struct ring_buffer
*buffer
,
2555 struct ring_buffer_per_cpu
*cpu_buffer
,
2556 unsigned long length
)
2558 struct ring_buffer_event
*event
;
2564 rb_start_commit(cpu_buffer
);
2566 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2568 * Due to the ability to swap a cpu buffer from a buffer
2569 * it is possible it was swapped before we committed.
2570 * (committing stops a swap). We check for it here and
2571 * if it happened, we have to fail the write.
2574 if (unlikely(ACCESS_ONCE(cpu_buffer
->buffer
) != buffer
)) {
2575 local_dec(&cpu_buffer
->committing
);
2576 local_dec(&cpu_buffer
->commits
);
2581 length
= rb_calculate_event_length(length
);
2587 * We allow for interrupts to reenter here and do a trace.
2588 * If one does, it will cause this original code to loop
2589 * back here. Even with heavy interrupts happening, this
2590 * should only happen a few times in a row. If this happens
2591 * 1000 times in a row, there must be either an interrupt
2592 * storm or we have something buggy.
2595 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 1000))
2598 ts
= rb_time_stamp(cpu_buffer
->buffer
);
2599 diff
= ts
- cpu_buffer
->write_stamp
;
2601 /* make sure this diff is calculated here */
2604 /* Did the write stamp get updated already? */
2605 if (likely(ts
>= cpu_buffer
->write_stamp
)) {
2607 if (unlikely(test_time_stamp(delta
))) {
2608 int local_clock_stable
= 1;
2609 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2610 local_clock_stable
= sched_clock_stable();
2612 WARN_ONCE(delta
> (1ULL << 59),
2613 KERN_WARNING
"Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2614 (unsigned long long)delta
,
2615 (unsigned long long)ts
,
2616 (unsigned long long)cpu_buffer
->write_stamp
,
2617 local_clock_stable
? "" :
2618 "If you just came from a suspend/resume,\n"
2619 "please switch to the trace global clock:\n"
2620 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2625 event
= __rb_reserve_next(cpu_buffer
, length
, ts
,
2626 delta
, add_timestamp
);
2627 if (unlikely(PTR_ERR(event
) == -EAGAIN
))
2636 rb_end_commit(cpu_buffer
);
2641 * The lock and unlock are done within a preempt disable section.
2642 * The current_context per_cpu variable can only be modified
2643 * by the current task between lock and unlock. But it can
2644 * be modified more than once via an interrupt. To pass this
2645 * information from the lock to the unlock without having to
2646 * access the 'in_interrupt()' functions again (which do show
2647 * a bit of overhead in something as critical as function tracing,
2648 * we use a bitmask trick.
2650 * bit 0 = NMI context
2651 * bit 1 = IRQ context
2652 * bit 2 = SoftIRQ context
2653 * bit 3 = normal context.
2655 * This works because this is the order of contexts that can
2656 * preempt other contexts. A SoftIRQ never preempts an IRQ
2659 * When the context is determined, the corresponding bit is
2660 * checked and set (if it was set, then a recursion of that context
2663 * On unlock, we need to clear this bit. To do so, just subtract
2664 * 1 from the current_context and AND it to itself.
2668 * 101 & 100 = 100 (clearing bit zero)
2671 * 1010 & 1001 = 1000 (clearing bit 1)
2673 * The least significant bit can be cleared this way, and it
2674 * just so happens that it is the same bit corresponding to
2675 * the current context.
2678 static __always_inline
int
2679 trace_recursive_lock(struct ring_buffer_per_cpu
*cpu_buffer
)
2681 unsigned int val
= cpu_buffer
->current_context
;
2684 if (in_interrupt()) {
2694 if (unlikely(val
& (1 << bit
)))
2698 cpu_buffer
->current_context
= val
;
2703 static __always_inline
void
2704 trace_recursive_unlock(struct ring_buffer_per_cpu
*cpu_buffer
)
2706 cpu_buffer
->current_context
&= cpu_buffer
->current_context
- 1;
2710 * ring_buffer_lock_reserve - reserve a part of the buffer
2711 * @buffer: the ring buffer to reserve from
2712 * @length: the length of the data to reserve (excluding event header)
2714 * Returns a reseverd event on the ring buffer to copy directly to.
2715 * The user of this interface will need to get the body to write into
2716 * and can use the ring_buffer_event_data() interface.
2718 * The length is the length of the data needed, not the event length
2719 * which also includes the event header.
2721 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2722 * If NULL is returned, then nothing has been allocated or locked.
2724 struct ring_buffer_event
*
2725 ring_buffer_lock_reserve(struct ring_buffer
*buffer
, unsigned long length
)
2727 struct ring_buffer_per_cpu
*cpu_buffer
;
2728 struct ring_buffer_event
*event
;
2731 if (ring_buffer_flags
!= RB_BUFFERS_ON
)
2734 /* If we are tracing schedule, we don't want to recurse */
2735 preempt_disable_notrace();
2737 if (unlikely(atomic_read(&buffer
->record_disabled
)))
2740 cpu
= raw_smp_processor_id();
2742 if (unlikely(!cpumask_test_cpu(cpu
, buffer
->cpumask
)))
2745 cpu_buffer
= buffer
->buffers
[cpu
];
2747 if (unlikely(atomic_read(&cpu_buffer
->record_disabled
)))
2750 if (unlikely(length
> BUF_MAX_DATA_SIZE
))
2753 if (unlikely(trace_recursive_lock(cpu_buffer
)))
2756 event
= rb_reserve_next_event(buffer
, cpu_buffer
, length
);
2763 trace_recursive_unlock(cpu_buffer
);
2765 preempt_enable_notrace();
2768 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve
);
2771 rb_update_write_stamp(struct ring_buffer_per_cpu
*cpu_buffer
,
2772 struct ring_buffer_event
*event
)
2777 * The event first in the commit queue updates the
2780 if (rb_event_is_commit(cpu_buffer
, event
)) {
2782 * A commit event that is first on a page
2783 * updates the write timestamp with the page stamp
2785 if (!rb_event_index(event
))
2786 cpu_buffer
->write_stamp
=
2787 cpu_buffer
->commit_page
->page
->time_stamp
;
2788 else if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
) {
2789 delta
= event
->array
[0];
2791 delta
+= event
->time_delta
;
2792 cpu_buffer
->write_stamp
+= delta
;
2794 cpu_buffer
->write_stamp
+= event
->time_delta
;
2798 static void rb_commit(struct ring_buffer_per_cpu
*cpu_buffer
,
2799 struct ring_buffer_event
*event
)
2801 local_inc(&cpu_buffer
->entries
);
2802 rb_update_write_stamp(cpu_buffer
, event
);
2803 rb_end_commit(cpu_buffer
);
2806 static __always_inline
void
2807 rb_wakeups(struct ring_buffer
*buffer
, struct ring_buffer_per_cpu
*cpu_buffer
)
2811 if (buffer
->irq_work
.waiters_pending
) {
2812 buffer
->irq_work
.waiters_pending
= false;
2813 /* irq_work_queue() supplies it's own memory barriers */
2814 irq_work_queue(&buffer
->irq_work
.work
);
2817 if (cpu_buffer
->irq_work
.waiters_pending
) {
2818 cpu_buffer
->irq_work
.waiters_pending
= false;
2819 /* irq_work_queue() supplies it's own memory barriers */
2820 irq_work_queue(&cpu_buffer
->irq_work
.work
);
2823 pagebusy
= cpu_buffer
->reader_page
== cpu_buffer
->commit_page
;
2825 if (!pagebusy
&& cpu_buffer
->irq_work
.full_waiters_pending
) {
2826 cpu_buffer
->irq_work
.wakeup_full
= true;
2827 cpu_buffer
->irq_work
.full_waiters_pending
= false;
2828 /* irq_work_queue() supplies it's own memory barriers */
2829 irq_work_queue(&cpu_buffer
->irq_work
.work
);
2834 * ring_buffer_unlock_commit - commit a reserved
2835 * @buffer: The buffer to commit to
2836 * @event: The event pointer to commit.
2838 * This commits the data to the ring buffer, and releases any locks held.
2840 * Must be paired with ring_buffer_lock_reserve.
2842 int ring_buffer_unlock_commit(struct ring_buffer
*buffer
,
2843 struct ring_buffer_event
*event
)
2845 struct ring_buffer_per_cpu
*cpu_buffer
;
2846 int cpu
= raw_smp_processor_id();
2848 cpu_buffer
= buffer
->buffers
[cpu
];
2850 rb_commit(cpu_buffer
, event
);
2852 rb_wakeups(buffer
, cpu_buffer
);
2854 trace_recursive_unlock(cpu_buffer
);
2856 preempt_enable_notrace();
2860 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit
);
2862 static inline void rb_event_discard(struct ring_buffer_event
*event
)
2864 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
2865 event
= skip_time_extend(event
);
2867 /* array[0] holds the actual length for the discarded event */
2868 event
->array
[0] = rb_event_data_length(event
) - RB_EVNT_HDR_SIZE
;
2869 event
->type_len
= RINGBUF_TYPE_PADDING
;
2870 /* time delta must be non zero */
2871 if (!event
->time_delta
)
2872 event
->time_delta
= 1;
2876 * Decrement the entries to the page that an event is on.
2877 * The event does not even need to exist, only the pointer
2878 * to the page it is on. This may only be called before the commit
2882 rb_decrement_entry(struct ring_buffer_per_cpu
*cpu_buffer
,
2883 struct ring_buffer_event
*event
)
2885 unsigned long addr
= (unsigned long)event
;
2886 struct buffer_page
*bpage
= cpu_buffer
->commit_page
;
2887 struct buffer_page
*start
;
2891 /* Do the likely case first */
2892 if (likely(bpage
->page
== (void *)addr
)) {
2893 local_dec(&bpage
->entries
);
2898 * Because the commit page may be on the reader page we
2899 * start with the next page and check the end loop there.
2901 rb_inc_page(cpu_buffer
, &bpage
);
2904 if (bpage
->page
== (void *)addr
) {
2905 local_dec(&bpage
->entries
);
2908 rb_inc_page(cpu_buffer
, &bpage
);
2909 } while (bpage
!= start
);
2911 /* commit not part of this buffer?? */
2912 RB_WARN_ON(cpu_buffer
, 1);
2916 * ring_buffer_commit_discard - discard an event that has not been committed
2917 * @buffer: the ring buffer
2918 * @event: non committed event to discard
2920 * Sometimes an event that is in the ring buffer needs to be ignored.
2921 * This function lets the user discard an event in the ring buffer
2922 * and then that event will not be read later.
2924 * This function only works if it is called before the the item has been
2925 * committed. It will try to free the event from the ring buffer
2926 * if another event has not been added behind it.
2928 * If another event has been added behind it, it will set the event
2929 * up as discarded, and perform the commit.
2931 * If this function is called, do not call ring_buffer_unlock_commit on
2934 void ring_buffer_discard_commit(struct ring_buffer
*buffer
,
2935 struct ring_buffer_event
*event
)
2937 struct ring_buffer_per_cpu
*cpu_buffer
;
2940 /* The event is discarded regardless */
2941 rb_event_discard(event
);
2943 cpu
= smp_processor_id();
2944 cpu_buffer
= buffer
->buffers
[cpu
];
2947 * This must only be called if the event has not been
2948 * committed yet. Thus we can assume that preemption
2949 * is still disabled.
2951 RB_WARN_ON(buffer
, !local_read(&cpu_buffer
->committing
));
2953 rb_decrement_entry(cpu_buffer
, event
);
2954 if (rb_try_to_discard(cpu_buffer
, event
))
2958 * The commit is still visible by the reader, so we
2959 * must still update the timestamp.
2961 rb_update_write_stamp(cpu_buffer
, event
);
2963 rb_end_commit(cpu_buffer
);
2965 trace_recursive_unlock(cpu_buffer
);
2967 preempt_enable_notrace();
2970 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit
);
2973 * ring_buffer_write - write data to the buffer without reserving
2974 * @buffer: The ring buffer to write to.
2975 * @length: The length of the data being written (excluding the event header)
2976 * @data: The data to write to the buffer.
2978 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2979 * one function. If you already have the data to write to the buffer, it
2980 * may be easier to simply call this function.
2982 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2983 * and not the length of the event which would hold the header.
2985 int ring_buffer_write(struct ring_buffer
*buffer
,
2986 unsigned long length
,
2989 struct ring_buffer_per_cpu
*cpu_buffer
;
2990 struct ring_buffer_event
*event
;
2995 if (ring_buffer_flags
!= RB_BUFFERS_ON
)
2998 preempt_disable_notrace();
3000 if (atomic_read(&buffer
->record_disabled
))
3003 cpu
= raw_smp_processor_id();
3005 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3008 cpu_buffer
= buffer
->buffers
[cpu
];
3010 if (atomic_read(&cpu_buffer
->record_disabled
))
3013 if (length
> BUF_MAX_DATA_SIZE
)
3016 if (unlikely(trace_recursive_lock(cpu_buffer
)))
3019 event
= rb_reserve_next_event(buffer
, cpu_buffer
, length
);
3023 body
= rb_event_data(event
);
3025 memcpy(body
, data
, length
);
3027 rb_commit(cpu_buffer
, event
);
3029 rb_wakeups(buffer
, cpu_buffer
);
3034 trace_recursive_unlock(cpu_buffer
);
3037 preempt_enable_notrace();
3041 EXPORT_SYMBOL_GPL(ring_buffer_write
);
3043 static int rb_per_cpu_empty(struct ring_buffer_per_cpu
*cpu_buffer
)
3045 struct buffer_page
*reader
= cpu_buffer
->reader_page
;
3046 struct buffer_page
*head
= rb_set_head_page(cpu_buffer
);
3047 struct buffer_page
*commit
= cpu_buffer
->commit_page
;
3049 /* In case of error, head will be NULL */
3050 if (unlikely(!head
))
3053 return reader
->read
== rb_page_commit(reader
) &&
3054 (commit
== reader
||
3056 head
->read
== rb_page_commit(commit
)));
3060 * ring_buffer_record_disable - stop all writes into the buffer
3061 * @buffer: The ring buffer to stop writes to.
3063 * This prevents all writes to the buffer. Any attempt to write
3064 * to the buffer after this will fail and return NULL.
3066 * The caller should call synchronize_sched() after this.
3068 void ring_buffer_record_disable(struct ring_buffer
*buffer
)
3070 atomic_inc(&buffer
->record_disabled
);
3072 EXPORT_SYMBOL_GPL(ring_buffer_record_disable
);
3075 * ring_buffer_record_enable - enable writes to the buffer
3076 * @buffer: The ring buffer to enable writes
3078 * Note, multiple disables will need the same number of enables
3079 * to truly enable the writing (much like preempt_disable).
3081 void ring_buffer_record_enable(struct ring_buffer
*buffer
)
3083 atomic_dec(&buffer
->record_disabled
);
3085 EXPORT_SYMBOL_GPL(ring_buffer_record_enable
);
3088 * ring_buffer_record_off - stop all writes into the buffer
3089 * @buffer: The ring buffer to stop writes to.
3091 * This prevents all writes to the buffer. Any attempt to write
3092 * to the buffer after this will fail and return NULL.
3094 * This is different than ring_buffer_record_disable() as
3095 * it works like an on/off switch, where as the disable() version
3096 * must be paired with a enable().
3098 void ring_buffer_record_off(struct ring_buffer
*buffer
)
3101 unsigned int new_rd
;
3104 rd
= atomic_read(&buffer
->record_disabled
);
3105 new_rd
= rd
| RB_BUFFER_OFF
;
3106 } while (atomic_cmpxchg(&buffer
->record_disabled
, rd
, new_rd
) != rd
);
3108 EXPORT_SYMBOL_GPL(ring_buffer_record_off
);
3111 * ring_buffer_record_on - restart writes into the buffer
3112 * @buffer: The ring buffer to start writes to.
3114 * This enables all writes to the buffer that was disabled by
3115 * ring_buffer_record_off().
3117 * This is different than ring_buffer_record_enable() as
3118 * it works like an on/off switch, where as the enable() version
3119 * must be paired with a disable().
3121 void ring_buffer_record_on(struct ring_buffer
*buffer
)
3124 unsigned int new_rd
;
3127 rd
= atomic_read(&buffer
->record_disabled
);
3128 new_rd
= rd
& ~RB_BUFFER_OFF
;
3129 } while (atomic_cmpxchg(&buffer
->record_disabled
, rd
, new_rd
) != rd
);
3131 EXPORT_SYMBOL_GPL(ring_buffer_record_on
);
3134 * ring_buffer_record_is_on - return true if the ring buffer can write
3135 * @buffer: The ring buffer to see if write is enabled
3137 * Returns true if the ring buffer is in a state that it accepts writes.
3139 int ring_buffer_record_is_on(struct ring_buffer
*buffer
)
3141 return !atomic_read(&buffer
->record_disabled
);
3145 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3146 * @buffer: The ring buffer to stop writes to.
3147 * @cpu: The CPU buffer to stop
3149 * This prevents all writes to the buffer. Any attempt to write
3150 * to the buffer after this will fail and return NULL.
3152 * The caller should call synchronize_sched() after this.
3154 void ring_buffer_record_disable_cpu(struct ring_buffer
*buffer
, int cpu
)
3156 struct ring_buffer_per_cpu
*cpu_buffer
;
3158 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3161 cpu_buffer
= buffer
->buffers
[cpu
];
3162 atomic_inc(&cpu_buffer
->record_disabled
);
3164 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu
);
3167 * ring_buffer_record_enable_cpu - enable writes to the buffer
3168 * @buffer: The ring buffer to enable writes
3169 * @cpu: The CPU to enable.
3171 * Note, multiple disables will need the same number of enables
3172 * to truly enable the writing (much like preempt_disable).
3174 void ring_buffer_record_enable_cpu(struct ring_buffer
*buffer
, int cpu
)
3176 struct ring_buffer_per_cpu
*cpu_buffer
;
3178 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3181 cpu_buffer
= buffer
->buffers
[cpu
];
3182 atomic_dec(&cpu_buffer
->record_disabled
);
3184 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu
);
3187 * The total entries in the ring buffer is the running counter
3188 * of entries entered into the ring buffer, minus the sum of
3189 * the entries read from the ring buffer and the number of
3190 * entries that were overwritten.
3192 static inline unsigned long
3193 rb_num_of_entries(struct ring_buffer_per_cpu
*cpu_buffer
)
3195 return local_read(&cpu_buffer
->entries
) -
3196 (local_read(&cpu_buffer
->overrun
) + cpu_buffer
->read
);
3200 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3201 * @buffer: The ring buffer
3202 * @cpu: The per CPU buffer to read from.
3204 u64
ring_buffer_oldest_event_ts(struct ring_buffer
*buffer
, int cpu
)
3206 unsigned long flags
;
3207 struct ring_buffer_per_cpu
*cpu_buffer
;
3208 struct buffer_page
*bpage
;
3211 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3214 cpu_buffer
= buffer
->buffers
[cpu
];
3215 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3217 * if the tail is on reader_page, oldest time stamp is on the reader
3220 if (cpu_buffer
->tail_page
== cpu_buffer
->reader_page
)
3221 bpage
= cpu_buffer
->reader_page
;
3223 bpage
= rb_set_head_page(cpu_buffer
);
3225 ret
= bpage
->page
->time_stamp
;
3226 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3230 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts
);
3233 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3234 * @buffer: The ring buffer
3235 * @cpu: The per CPU buffer to read from.
3237 unsigned long ring_buffer_bytes_cpu(struct ring_buffer
*buffer
, int cpu
)
3239 struct ring_buffer_per_cpu
*cpu_buffer
;
3242 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3245 cpu_buffer
= buffer
->buffers
[cpu
];
3246 ret
= local_read(&cpu_buffer
->entries_bytes
) - cpu_buffer
->read_bytes
;
3250 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu
);
3253 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3254 * @buffer: The ring buffer
3255 * @cpu: The per CPU buffer to get the entries from.
3257 unsigned long ring_buffer_entries_cpu(struct ring_buffer
*buffer
, int cpu
)
3259 struct ring_buffer_per_cpu
*cpu_buffer
;
3261 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3264 cpu_buffer
= buffer
->buffers
[cpu
];
3266 return rb_num_of_entries(cpu_buffer
);
3268 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu
);
3271 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3272 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3273 * @buffer: The ring buffer
3274 * @cpu: The per CPU buffer to get the number of overruns from
3276 unsigned long ring_buffer_overrun_cpu(struct ring_buffer
*buffer
, int cpu
)
3278 struct ring_buffer_per_cpu
*cpu_buffer
;
3281 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3284 cpu_buffer
= buffer
->buffers
[cpu
];
3285 ret
= local_read(&cpu_buffer
->overrun
);
3289 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu
);
3292 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3293 * commits failing due to the buffer wrapping around while there are uncommitted
3294 * events, such as during an interrupt storm.
3295 * @buffer: The ring buffer
3296 * @cpu: The per CPU buffer to get the number of overruns from
3299 ring_buffer_commit_overrun_cpu(struct ring_buffer
*buffer
, int cpu
)
3301 struct ring_buffer_per_cpu
*cpu_buffer
;
3304 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3307 cpu_buffer
= buffer
->buffers
[cpu
];
3308 ret
= local_read(&cpu_buffer
->commit_overrun
);
3312 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu
);
3315 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3316 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3317 * @buffer: The ring buffer
3318 * @cpu: The per CPU buffer to get the number of overruns from
3321 ring_buffer_dropped_events_cpu(struct ring_buffer
*buffer
, int cpu
)
3323 struct ring_buffer_per_cpu
*cpu_buffer
;
3326 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3329 cpu_buffer
= buffer
->buffers
[cpu
];
3330 ret
= local_read(&cpu_buffer
->dropped_events
);
3334 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu
);
3337 * ring_buffer_read_events_cpu - get the number of events successfully read
3338 * @buffer: The ring buffer
3339 * @cpu: The per CPU buffer to get the number of events read
3342 ring_buffer_read_events_cpu(struct ring_buffer
*buffer
, int cpu
)
3344 struct ring_buffer_per_cpu
*cpu_buffer
;
3346 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3349 cpu_buffer
= buffer
->buffers
[cpu
];
3350 return cpu_buffer
->read
;
3352 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu
);
3355 * ring_buffer_entries - get the number of entries in a buffer
3356 * @buffer: The ring buffer
3358 * Returns the total number of entries in the ring buffer
3361 unsigned long ring_buffer_entries(struct ring_buffer
*buffer
)
3363 struct ring_buffer_per_cpu
*cpu_buffer
;
3364 unsigned long entries
= 0;
3367 /* if you care about this being correct, lock the buffer */
3368 for_each_buffer_cpu(buffer
, cpu
) {
3369 cpu_buffer
= buffer
->buffers
[cpu
];
3370 entries
+= rb_num_of_entries(cpu_buffer
);
3375 EXPORT_SYMBOL_GPL(ring_buffer_entries
);
3378 * ring_buffer_overruns - get the number of overruns in buffer
3379 * @buffer: The ring buffer
3381 * Returns the total number of overruns in the ring buffer
3384 unsigned long ring_buffer_overruns(struct ring_buffer
*buffer
)
3386 struct ring_buffer_per_cpu
*cpu_buffer
;
3387 unsigned long overruns
= 0;
3390 /* if you care about this being correct, lock the buffer */
3391 for_each_buffer_cpu(buffer
, cpu
) {
3392 cpu_buffer
= buffer
->buffers
[cpu
];
3393 overruns
+= local_read(&cpu_buffer
->overrun
);
3398 EXPORT_SYMBOL_GPL(ring_buffer_overruns
);
3400 static void rb_iter_reset(struct ring_buffer_iter
*iter
)
3402 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
3404 /* Iterator usage is expected to have record disabled */
3405 iter
->head_page
= cpu_buffer
->reader_page
;
3406 iter
->head
= cpu_buffer
->reader_page
->read
;
3408 iter
->cache_reader_page
= iter
->head_page
;
3409 iter
->cache_read
= cpu_buffer
->read
;
3412 iter
->read_stamp
= cpu_buffer
->read_stamp
;
3414 iter
->read_stamp
= iter
->head_page
->page
->time_stamp
;
3418 * ring_buffer_iter_reset - reset an iterator
3419 * @iter: The iterator to reset
3421 * Resets the iterator, so that it will start from the beginning
3424 void ring_buffer_iter_reset(struct ring_buffer_iter
*iter
)
3426 struct ring_buffer_per_cpu
*cpu_buffer
;
3427 unsigned long flags
;
3432 cpu_buffer
= iter
->cpu_buffer
;
3434 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3435 rb_iter_reset(iter
);
3436 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3438 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset
);
3441 * ring_buffer_iter_empty - check if an iterator has no more to read
3442 * @iter: The iterator to check
3444 int ring_buffer_iter_empty(struct ring_buffer_iter
*iter
)
3446 struct ring_buffer_per_cpu
*cpu_buffer
;
3448 cpu_buffer
= iter
->cpu_buffer
;
3450 return iter
->head_page
== cpu_buffer
->commit_page
&&
3451 iter
->head
== rb_commit_index(cpu_buffer
);
3453 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty
);
3456 rb_update_read_stamp(struct ring_buffer_per_cpu
*cpu_buffer
,
3457 struct ring_buffer_event
*event
)
3461 switch (event
->type_len
) {
3462 case RINGBUF_TYPE_PADDING
:
3465 case RINGBUF_TYPE_TIME_EXTEND
:
3466 delta
= event
->array
[0];
3468 delta
+= event
->time_delta
;
3469 cpu_buffer
->read_stamp
+= delta
;
3472 case RINGBUF_TYPE_TIME_STAMP
:
3473 /* FIXME: not implemented */
3476 case RINGBUF_TYPE_DATA
:
3477 cpu_buffer
->read_stamp
+= event
->time_delta
;
3487 rb_update_iter_read_stamp(struct ring_buffer_iter
*iter
,
3488 struct ring_buffer_event
*event
)
3492 switch (event
->type_len
) {
3493 case RINGBUF_TYPE_PADDING
:
3496 case RINGBUF_TYPE_TIME_EXTEND
:
3497 delta
= event
->array
[0];
3499 delta
+= event
->time_delta
;
3500 iter
->read_stamp
+= delta
;
3503 case RINGBUF_TYPE_TIME_STAMP
:
3504 /* FIXME: not implemented */
3507 case RINGBUF_TYPE_DATA
:
3508 iter
->read_stamp
+= event
->time_delta
;
3517 static struct buffer_page
*
3518 rb_get_reader_page(struct ring_buffer_per_cpu
*cpu_buffer
)
3520 struct buffer_page
*reader
= NULL
;
3521 unsigned long overwrite
;
3522 unsigned long flags
;
3526 local_irq_save(flags
);
3527 arch_spin_lock(&cpu_buffer
->lock
);
3531 * This should normally only loop twice. But because the
3532 * start of the reader inserts an empty page, it causes
3533 * a case where we will loop three times. There should be no
3534 * reason to loop four times (that I know of).
3536 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 3)) {
3541 reader
= cpu_buffer
->reader_page
;
3543 /* If there's more to read, return this page */
3544 if (cpu_buffer
->reader_page
->read
< rb_page_size(reader
))
3547 /* Never should we have an index greater than the size */
3548 if (RB_WARN_ON(cpu_buffer
,
3549 cpu_buffer
->reader_page
->read
> rb_page_size(reader
)))
3552 /* check if we caught up to the tail */
3554 if (cpu_buffer
->commit_page
== cpu_buffer
->reader_page
)
3557 /* Don't bother swapping if the ring buffer is empty */
3558 if (rb_num_of_entries(cpu_buffer
) == 0)
3562 * Reset the reader page to size zero.
3564 local_set(&cpu_buffer
->reader_page
->write
, 0);
3565 local_set(&cpu_buffer
->reader_page
->entries
, 0);
3566 local_set(&cpu_buffer
->reader_page
->page
->commit
, 0);
3567 cpu_buffer
->reader_page
->real_end
= 0;
3571 * Splice the empty reader page into the list around the head.
3573 reader
= rb_set_head_page(cpu_buffer
);
3576 cpu_buffer
->reader_page
->list
.next
= rb_list_head(reader
->list
.next
);
3577 cpu_buffer
->reader_page
->list
.prev
= reader
->list
.prev
;
3580 * cpu_buffer->pages just needs to point to the buffer, it
3581 * has no specific buffer page to point to. Lets move it out
3582 * of our way so we don't accidentally swap it.
3584 cpu_buffer
->pages
= reader
->list
.prev
;
3586 /* The reader page will be pointing to the new head */
3587 rb_set_list_to_head(cpu_buffer
, &cpu_buffer
->reader_page
->list
);
3590 * We want to make sure we read the overruns after we set up our
3591 * pointers to the next object. The writer side does a
3592 * cmpxchg to cross pages which acts as the mb on the writer
3593 * side. Note, the reader will constantly fail the swap
3594 * while the writer is updating the pointers, so this
3595 * guarantees that the overwrite recorded here is the one we
3596 * want to compare with the last_overrun.
3599 overwrite
= local_read(&(cpu_buffer
->overrun
));
3602 * Here's the tricky part.
3604 * We need to move the pointer past the header page.
3605 * But we can only do that if a writer is not currently
3606 * moving it. The page before the header page has the
3607 * flag bit '1' set if it is pointing to the page we want.
3608 * but if the writer is in the process of moving it
3609 * than it will be '2' or already moved '0'.
3612 ret
= rb_head_page_replace(reader
, cpu_buffer
->reader_page
);
3615 * If we did not convert it, then we must try again.
3621 * Yeah! We succeeded in replacing the page.
3623 * Now make the new head point back to the reader page.
3625 rb_list_head(reader
->list
.next
)->prev
= &cpu_buffer
->reader_page
->list
;
3626 rb_inc_page(cpu_buffer
, &cpu_buffer
->head_page
);
3628 /* Finally update the reader page to the new head */
3629 cpu_buffer
->reader_page
= reader
;
3630 rb_reset_reader_page(cpu_buffer
);
3632 if (overwrite
!= cpu_buffer
->last_overrun
) {
3633 cpu_buffer
->lost_events
= overwrite
- cpu_buffer
->last_overrun
;
3634 cpu_buffer
->last_overrun
= overwrite
;
3640 arch_spin_unlock(&cpu_buffer
->lock
);
3641 local_irq_restore(flags
);
3646 static void rb_advance_reader(struct ring_buffer_per_cpu
*cpu_buffer
)
3648 struct ring_buffer_event
*event
;
3649 struct buffer_page
*reader
;
3652 reader
= rb_get_reader_page(cpu_buffer
);
3654 /* This function should not be called when buffer is empty */
3655 if (RB_WARN_ON(cpu_buffer
, !reader
))
3658 event
= rb_reader_event(cpu_buffer
);
3660 if (event
->type_len
<= RINGBUF_TYPE_DATA_TYPE_LEN_MAX
)
3663 rb_update_read_stamp(cpu_buffer
, event
);
3665 length
= rb_event_length(event
);
3666 cpu_buffer
->reader_page
->read
+= length
;
3669 static void rb_advance_iter(struct ring_buffer_iter
*iter
)
3671 struct ring_buffer_per_cpu
*cpu_buffer
;
3672 struct ring_buffer_event
*event
;
3675 cpu_buffer
= iter
->cpu_buffer
;
3678 * Check if we are at the end of the buffer.
3680 if (iter
->head
>= rb_page_size(iter
->head_page
)) {
3681 /* discarded commits can make the page empty */
3682 if (iter
->head_page
== cpu_buffer
->commit_page
)
3688 event
= rb_iter_head_event(iter
);
3690 length
= rb_event_length(event
);
3693 * This should not be called to advance the header if we are
3694 * at the tail of the buffer.
3696 if (RB_WARN_ON(cpu_buffer
,
3697 (iter
->head_page
== cpu_buffer
->commit_page
) &&
3698 (iter
->head
+ length
> rb_commit_index(cpu_buffer
))))
3701 rb_update_iter_read_stamp(iter
, event
);
3703 iter
->head
+= length
;
3705 /* check for end of page padding */
3706 if ((iter
->head
>= rb_page_size(iter
->head_page
)) &&
3707 (iter
->head_page
!= cpu_buffer
->commit_page
))
3711 static int rb_lost_events(struct ring_buffer_per_cpu
*cpu_buffer
)
3713 return cpu_buffer
->lost_events
;
3716 static struct ring_buffer_event
*
3717 rb_buffer_peek(struct ring_buffer_per_cpu
*cpu_buffer
, u64
*ts
,
3718 unsigned long *lost_events
)
3720 struct ring_buffer_event
*event
;
3721 struct buffer_page
*reader
;
3726 * We repeat when a time extend is encountered.
3727 * Since the time extend is always attached to a data event,
3728 * we should never loop more than once.
3729 * (We never hit the following condition more than twice).
3731 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 2))
3734 reader
= rb_get_reader_page(cpu_buffer
);
3738 event
= rb_reader_event(cpu_buffer
);
3740 switch (event
->type_len
) {
3741 case RINGBUF_TYPE_PADDING
:
3742 if (rb_null_event(event
))
3743 RB_WARN_ON(cpu_buffer
, 1);
3745 * Because the writer could be discarding every
3746 * event it creates (which would probably be bad)
3747 * if we were to go back to "again" then we may never
3748 * catch up, and will trigger the warn on, or lock
3749 * the box. Return the padding, and we will release
3750 * the current locks, and try again.
3754 case RINGBUF_TYPE_TIME_EXTEND
:
3755 /* Internal data, OK to advance */
3756 rb_advance_reader(cpu_buffer
);
3759 case RINGBUF_TYPE_TIME_STAMP
:
3760 /* FIXME: not implemented */
3761 rb_advance_reader(cpu_buffer
);
3764 case RINGBUF_TYPE_DATA
:
3766 *ts
= cpu_buffer
->read_stamp
+ event
->time_delta
;
3767 ring_buffer_normalize_time_stamp(cpu_buffer
->buffer
,
3768 cpu_buffer
->cpu
, ts
);
3771 *lost_events
= rb_lost_events(cpu_buffer
);
3780 EXPORT_SYMBOL_GPL(ring_buffer_peek
);
3782 static struct ring_buffer_event
*
3783 rb_iter_peek(struct ring_buffer_iter
*iter
, u64
*ts
)
3785 struct ring_buffer
*buffer
;
3786 struct ring_buffer_per_cpu
*cpu_buffer
;
3787 struct ring_buffer_event
*event
;
3790 cpu_buffer
= iter
->cpu_buffer
;
3791 buffer
= cpu_buffer
->buffer
;
3794 * Check if someone performed a consuming read to
3795 * the buffer. A consuming read invalidates the iterator
3796 * and we need to reset the iterator in this case.
3798 if (unlikely(iter
->cache_read
!= cpu_buffer
->read
||
3799 iter
->cache_reader_page
!= cpu_buffer
->reader_page
))
3800 rb_iter_reset(iter
);
3803 if (ring_buffer_iter_empty(iter
))
3807 * We repeat when a time extend is encountered or we hit
3808 * the end of the page. Since the time extend is always attached
3809 * to a data event, we should never loop more than three times.
3810 * Once for going to next page, once on time extend, and
3811 * finally once to get the event.
3812 * (We never hit the following condition more than thrice).
3814 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 3))
3817 if (rb_per_cpu_empty(cpu_buffer
))
3820 if (iter
->head
>= rb_page_size(iter
->head_page
)) {
3825 event
= rb_iter_head_event(iter
);
3827 switch (event
->type_len
) {
3828 case RINGBUF_TYPE_PADDING
:
3829 if (rb_null_event(event
)) {
3833 rb_advance_iter(iter
);
3836 case RINGBUF_TYPE_TIME_EXTEND
:
3837 /* Internal data, OK to advance */
3838 rb_advance_iter(iter
);
3841 case RINGBUF_TYPE_TIME_STAMP
:
3842 /* FIXME: not implemented */
3843 rb_advance_iter(iter
);
3846 case RINGBUF_TYPE_DATA
:
3848 *ts
= iter
->read_stamp
+ event
->time_delta
;
3849 ring_buffer_normalize_time_stamp(buffer
,
3850 cpu_buffer
->cpu
, ts
);
3860 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek
);
3862 static inline bool rb_reader_lock(struct ring_buffer_per_cpu
*cpu_buffer
)
3864 if (likely(!in_nmi())) {
3865 raw_spin_lock(&cpu_buffer
->reader_lock
);
3870 * If an NMI die dumps out the content of the ring buffer
3871 * trylock must be used to prevent a deadlock if the NMI
3872 * preempted a task that holds the ring buffer locks. If
3873 * we get the lock then all is fine, if not, then continue
3874 * to do the read, but this can corrupt the ring buffer,
3875 * so it must be permanently disabled from future writes.
3876 * Reading from NMI is a oneshot deal.
3878 if (raw_spin_trylock(&cpu_buffer
->reader_lock
))
3881 /* Continue without locking, but disable the ring buffer */
3882 atomic_inc(&cpu_buffer
->record_disabled
);
3887 rb_reader_unlock(struct ring_buffer_per_cpu
*cpu_buffer
, bool locked
)
3890 raw_spin_unlock(&cpu_buffer
->reader_lock
);
3895 * ring_buffer_peek - peek at the next event to be read
3896 * @buffer: The ring buffer to read
3897 * @cpu: The cpu to peak at
3898 * @ts: The timestamp counter of this event.
3899 * @lost_events: a variable to store if events were lost (may be NULL)
3901 * This will return the event that will be read next, but does
3902 * not consume the data.
3904 struct ring_buffer_event
*
3905 ring_buffer_peek(struct ring_buffer
*buffer
, int cpu
, u64
*ts
,
3906 unsigned long *lost_events
)
3908 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
3909 struct ring_buffer_event
*event
;
3910 unsigned long flags
;
3913 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3917 local_irq_save(flags
);
3918 dolock
= rb_reader_lock(cpu_buffer
);
3919 event
= rb_buffer_peek(cpu_buffer
, ts
, lost_events
);
3920 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3921 rb_advance_reader(cpu_buffer
);
3922 rb_reader_unlock(cpu_buffer
, dolock
);
3923 local_irq_restore(flags
);
3925 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3932 * ring_buffer_iter_peek - peek at the next event to be read
3933 * @iter: The ring buffer iterator
3934 * @ts: The timestamp counter of this event.
3936 * This will return the event that will be read next, but does
3937 * not increment the iterator.
3939 struct ring_buffer_event
*
3940 ring_buffer_iter_peek(struct ring_buffer_iter
*iter
, u64
*ts
)
3942 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
3943 struct ring_buffer_event
*event
;
3944 unsigned long flags
;
3947 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3948 event
= rb_iter_peek(iter
, ts
);
3949 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3951 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3958 * ring_buffer_consume - return an event and consume it
3959 * @buffer: The ring buffer to get the next event from
3960 * @cpu: the cpu to read the buffer from
3961 * @ts: a variable to store the timestamp (may be NULL)
3962 * @lost_events: a variable to store if events were lost (may be NULL)
3964 * Returns the next event in the ring buffer, and that event is consumed.
3965 * Meaning, that sequential reads will keep returning a different event,
3966 * and eventually empty the ring buffer if the producer is slower.
3968 struct ring_buffer_event
*
3969 ring_buffer_consume(struct ring_buffer
*buffer
, int cpu
, u64
*ts
,
3970 unsigned long *lost_events
)
3972 struct ring_buffer_per_cpu
*cpu_buffer
;
3973 struct ring_buffer_event
*event
= NULL
;
3974 unsigned long flags
;
3978 /* might be called in atomic */
3981 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3984 cpu_buffer
= buffer
->buffers
[cpu
];
3985 local_irq_save(flags
);
3986 dolock
= rb_reader_lock(cpu_buffer
);
3988 event
= rb_buffer_peek(cpu_buffer
, ts
, lost_events
);
3990 cpu_buffer
->lost_events
= 0;
3991 rb_advance_reader(cpu_buffer
);
3994 rb_reader_unlock(cpu_buffer
, dolock
);
3995 local_irq_restore(flags
);
4000 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
4005 EXPORT_SYMBOL_GPL(ring_buffer_consume
);
4008 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4009 * @buffer: The ring buffer to read from
4010 * @cpu: The cpu buffer to iterate over
4012 * This performs the initial preparations necessary to iterate
4013 * through the buffer. Memory is allocated, buffer recording
4014 * is disabled, and the iterator pointer is returned to the caller.
4016 * Disabling buffer recordng prevents the reading from being
4017 * corrupted. This is not a consuming read, so a producer is not
4020 * After a sequence of ring_buffer_read_prepare calls, the user is
4021 * expected to make at least one call to ring_buffer_read_prepare_sync.
4022 * Afterwards, ring_buffer_read_start is invoked to get things going
4025 * This overall must be paired with ring_buffer_read_finish.
4027 struct ring_buffer_iter
*
4028 ring_buffer_read_prepare(struct ring_buffer
*buffer
, int cpu
)
4030 struct ring_buffer_per_cpu
*cpu_buffer
;
4031 struct ring_buffer_iter
*iter
;
4033 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4036 iter
= kmalloc(sizeof(*iter
), GFP_KERNEL
);
4040 cpu_buffer
= buffer
->buffers
[cpu
];
4042 iter
->cpu_buffer
= cpu_buffer
;
4044 atomic_inc(&buffer
->resize_disabled
);
4045 atomic_inc(&cpu_buffer
->record_disabled
);
4049 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare
);
4052 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4054 * All previously invoked ring_buffer_read_prepare calls to prepare
4055 * iterators will be synchronized. Afterwards, read_buffer_read_start
4056 * calls on those iterators are allowed.
4059 ring_buffer_read_prepare_sync(void)
4061 synchronize_sched();
4063 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync
);
4066 * ring_buffer_read_start - start a non consuming read of the buffer
4067 * @iter: The iterator returned by ring_buffer_read_prepare
4069 * This finalizes the startup of an iteration through the buffer.
4070 * The iterator comes from a call to ring_buffer_read_prepare and
4071 * an intervening ring_buffer_read_prepare_sync must have been
4074 * Must be paired with ring_buffer_read_finish.
4077 ring_buffer_read_start(struct ring_buffer_iter
*iter
)
4079 struct ring_buffer_per_cpu
*cpu_buffer
;
4080 unsigned long flags
;
4085 cpu_buffer
= iter
->cpu_buffer
;
4087 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4088 arch_spin_lock(&cpu_buffer
->lock
);
4089 rb_iter_reset(iter
);
4090 arch_spin_unlock(&cpu_buffer
->lock
);
4091 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4093 EXPORT_SYMBOL_GPL(ring_buffer_read_start
);
4096 * ring_buffer_read_finish - finish reading the iterator of the buffer
4097 * @iter: The iterator retrieved by ring_buffer_start
4099 * This re-enables the recording to the buffer, and frees the
4103 ring_buffer_read_finish(struct ring_buffer_iter
*iter
)
4105 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
4106 unsigned long flags
;
4109 * Ring buffer is disabled from recording, here's a good place
4110 * to check the integrity of the ring buffer.
4111 * Must prevent readers from trying to read, as the check
4112 * clears the HEAD page and readers require it.
4114 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4115 rb_check_pages(cpu_buffer
);
4116 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4118 atomic_dec(&cpu_buffer
->record_disabled
);
4119 atomic_dec(&cpu_buffer
->buffer
->resize_disabled
);
4122 EXPORT_SYMBOL_GPL(ring_buffer_read_finish
);
4125 * ring_buffer_read - read the next item in the ring buffer by the iterator
4126 * @iter: The ring buffer iterator
4127 * @ts: The time stamp of the event read.
4129 * This reads the next event in the ring buffer and increments the iterator.
4131 struct ring_buffer_event
*
4132 ring_buffer_read(struct ring_buffer_iter
*iter
, u64
*ts
)
4134 struct ring_buffer_event
*event
;
4135 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
4136 unsigned long flags
;
4138 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4140 event
= rb_iter_peek(iter
, ts
);
4144 if (event
->type_len
== RINGBUF_TYPE_PADDING
)
4147 rb_advance_iter(iter
);
4149 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4153 EXPORT_SYMBOL_GPL(ring_buffer_read
);
4156 * ring_buffer_size - return the size of the ring buffer (in bytes)
4157 * @buffer: The ring buffer.
4159 unsigned long ring_buffer_size(struct ring_buffer
*buffer
, int cpu
)
4162 * Earlier, this method returned
4163 * BUF_PAGE_SIZE * buffer->nr_pages
4164 * Since the nr_pages field is now removed, we have converted this to
4165 * return the per cpu buffer value.
4167 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4170 return BUF_PAGE_SIZE
* buffer
->buffers
[cpu
]->nr_pages
;
4172 EXPORT_SYMBOL_GPL(ring_buffer_size
);
4175 rb_reset_cpu(struct ring_buffer_per_cpu
*cpu_buffer
)
4177 rb_head_page_deactivate(cpu_buffer
);
4179 cpu_buffer
->head_page
4180 = list_entry(cpu_buffer
->pages
, struct buffer_page
, list
);
4181 local_set(&cpu_buffer
->head_page
->write
, 0);
4182 local_set(&cpu_buffer
->head_page
->entries
, 0);
4183 local_set(&cpu_buffer
->head_page
->page
->commit
, 0);
4185 cpu_buffer
->head_page
->read
= 0;
4187 cpu_buffer
->tail_page
= cpu_buffer
->head_page
;
4188 cpu_buffer
->commit_page
= cpu_buffer
->head_page
;
4190 INIT_LIST_HEAD(&cpu_buffer
->reader_page
->list
);
4191 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
4192 local_set(&cpu_buffer
->reader_page
->write
, 0);
4193 local_set(&cpu_buffer
->reader_page
->entries
, 0);
4194 local_set(&cpu_buffer
->reader_page
->page
->commit
, 0);
4195 cpu_buffer
->reader_page
->read
= 0;
4197 local_set(&cpu_buffer
->entries_bytes
, 0);
4198 local_set(&cpu_buffer
->overrun
, 0);
4199 local_set(&cpu_buffer
->commit_overrun
, 0);
4200 local_set(&cpu_buffer
->dropped_events
, 0);
4201 local_set(&cpu_buffer
->entries
, 0);
4202 local_set(&cpu_buffer
->committing
, 0);
4203 local_set(&cpu_buffer
->commits
, 0);
4204 cpu_buffer
->read
= 0;
4205 cpu_buffer
->read_bytes
= 0;
4207 cpu_buffer
->write_stamp
= 0;
4208 cpu_buffer
->read_stamp
= 0;
4210 cpu_buffer
->lost_events
= 0;
4211 cpu_buffer
->last_overrun
= 0;
4213 rb_head_page_activate(cpu_buffer
);
4217 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4218 * @buffer: The ring buffer to reset a per cpu buffer of
4219 * @cpu: The CPU buffer to be reset
4221 void ring_buffer_reset_cpu(struct ring_buffer
*buffer
, int cpu
)
4223 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
4224 unsigned long flags
;
4226 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4229 atomic_inc(&buffer
->resize_disabled
);
4230 atomic_inc(&cpu_buffer
->record_disabled
);
4232 /* Make sure all commits have finished */
4233 synchronize_sched();
4235 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4237 if (RB_WARN_ON(cpu_buffer
, local_read(&cpu_buffer
->committing
)))
4240 arch_spin_lock(&cpu_buffer
->lock
);
4242 rb_reset_cpu(cpu_buffer
);
4244 arch_spin_unlock(&cpu_buffer
->lock
);
4247 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4249 atomic_dec(&cpu_buffer
->record_disabled
);
4250 atomic_dec(&buffer
->resize_disabled
);
4252 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu
);
4255 * ring_buffer_reset - reset a ring buffer
4256 * @buffer: The ring buffer to reset all cpu buffers
4258 void ring_buffer_reset(struct ring_buffer
*buffer
)
4262 for_each_buffer_cpu(buffer
, cpu
)
4263 ring_buffer_reset_cpu(buffer
, cpu
);
4265 EXPORT_SYMBOL_GPL(ring_buffer_reset
);
4268 * rind_buffer_empty - is the ring buffer empty?
4269 * @buffer: The ring buffer to test
4271 int ring_buffer_empty(struct ring_buffer
*buffer
)
4273 struct ring_buffer_per_cpu
*cpu_buffer
;
4274 unsigned long flags
;
4279 /* yes this is racy, but if you don't like the race, lock the buffer */
4280 for_each_buffer_cpu(buffer
, cpu
) {
4281 cpu_buffer
= buffer
->buffers
[cpu
];
4282 local_irq_save(flags
);
4283 dolock
= rb_reader_lock(cpu_buffer
);
4284 ret
= rb_per_cpu_empty(cpu_buffer
);
4285 rb_reader_unlock(cpu_buffer
, dolock
);
4286 local_irq_restore(flags
);
4294 EXPORT_SYMBOL_GPL(ring_buffer_empty
);
4297 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4298 * @buffer: The ring buffer
4299 * @cpu: The CPU buffer to test
4301 int ring_buffer_empty_cpu(struct ring_buffer
*buffer
, int cpu
)
4303 struct ring_buffer_per_cpu
*cpu_buffer
;
4304 unsigned long flags
;
4308 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4311 cpu_buffer
= buffer
->buffers
[cpu
];
4312 local_irq_save(flags
);
4313 dolock
= rb_reader_lock(cpu_buffer
);
4314 ret
= rb_per_cpu_empty(cpu_buffer
);
4315 rb_reader_unlock(cpu_buffer
, dolock
);
4316 local_irq_restore(flags
);
4320 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu
);
4322 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4324 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4325 * @buffer_a: One buffer to swap with
4326 * @buffer_b: The other buffer to swap with
4328 * This function is useful for tracers that want to take a "snapshot"
4329 * of a CPU buffer and has another back up buffer lying around.
4330 * it is expected that the tracer handles the cpu buffer not being
4331 * used at the moment.
4333 int ring_buffer_swap_cpu(struct ring_buffer
*buffer_a
,
4334 struct ring_buffer
*buffer_b
, int cpu
)
4336 struct ring_buffer_per_cpu
*cpu_buffer_a
;
4337 struct ring_buffer_per_cpu
*cpu_buffer_b
;
4340 if (!cpumask_test_cpu(cpu
, buffer_a
->cpumask
) ||
4341 !cpumask_test_cpu(cpu
, buffer_b
->cpumask
))
4344 cpu_buffer_a
= buffer_a
->buffers
[cpu
];
4345 cpu_buffer_b
= buffer_b
->buffers
[cpu
];
4347 /* At least make sure the two buffers are somewhat the same */
4348 if (cpu_buffer_a
->nr_pages
!= cpu_buffer_b
->nr_pages
)
4353 if (ring_buffer_flags
!= RB_BUFFERS_ON
)
4356 if (atomic_read(&buffer_a
->record_disabled
))
4359 if (atomic_read(&buffer_b
->record_disabled
))
4362 if (atomic_read(&cpu_buffer_a
->record_disabled
))
4365 if (atomic_read(&cpu_buffer_b
->record_disabled
))
4369 * We can't do a synchronize_sched here because this
4370 * function can be called in atomic context.
4371 * Normally this will be called from the same CPU as cpu.
4372 * If not it's up to the caller to protect this.
4374 atomic_inc(&cpu_buffer_a
->record_disabled
);
4375 atomic_inc(&cpu_buffer_b
->record_disabled
);
4378 if (local_read(&cpu_buffer_a
->committing
))
4380 if (local_read(&cpu_buffer_b
->committing
))
4383 buffer_a
->buffers
[cpu
] = cpu_buffer_b
;
4384 buffer_b
->buffers
[cpu
] = cpu_buffer_a
;
4386 cpu_buffer_b
->buffer
= buffer_a
;
4387 cpu_buffer_a
->buffer
= buffer_b
;
4392 atomic_dec(&cpu_buffer_a
->record_disabled
);
4393 atomic_dec(&cpu_buffer_b
->record_disabled
);
4397 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu
);
4398 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4401 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4402 * @buffer: the buffer to allocate for.
4403 * @cpu: the cpu buffer to allocate.
4405 * This function is used in conjunction with ring_buffer_read_page.
4406 * When reading a full page from the ring buffer, these functions
4407 * can be used to speed up the process. The calling function should
4408 * allocate a few pages first with this function. Then when it
4409 * needs to get pages from the ring buffer, it passes the result
4410 * of this function into ring_buffer_read_page, which will swap
4411 * the page that was allocated, with the read page of the buffer.
4414 * The page allocated, or NULL on error.
4416 void *ring_buffer_alloc_read_page(struct ring_buffer
*buffer
, int cpu
)
4418 struct buffer_data_page
*bpage
;
4421 page
= alloc_pages_node(cpu_to_node(cpu
),
4422 GFP_KERNEL
| __GFP_NORETRY
, 0);
4426 bpage
= page_address(page
);
4428 rb_init_page(bpage
);
4432 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page
);
4435 * ring_buffer_free_read_page - free an allocated read page
4436 * @buffer: the buffer the page was allocate for
4437 * @data: the page to free
4439 * Free a page allocated from ring_buffer_alloc_read_page.
4441 void ring_buffer_free_read_page(struct ring_buffer
*buffer
, void *data
)
4443 free_page((unsigned long)data
);
4445 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page
);
4448 * ring_buffer_read_page - extract a page from the ring buffer
4449 * @buffer: buffer to extract from
4450 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4451 * @len: amount to extract
4452 * @cpu: the cpu of the buffer to extract
4453 * @full: should the extraction only happen when the page is full.
4455 * This function will pull out a page from the ring buffer and consume it.
4456 * @data_page must be the address of the variable that was returned
4457 * from ring_buffer_alloc_read_page. This is because the page might be used
4458 * to swap with a page in the ring buffer.
4461 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4464 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4466 * process_page(rpage, ret);
4468 * When @full is set, the function will not return true unless
4469 * the writer is off the reader page.
4471 * Note: it is up to the calling functions to handle sleeps and wakeups.
4472 * The ring buffer can be used anywhere in the kernel and can not
4473 * blindly call wake_up. The layer that uses the ring buffer must be
4474 * responsible for that.
4477 * >=0 if data has been transferred, returns the offset of consumed data.
4478 * <0 if no data has been transferred.
4480 int ring_buffer_read_page(struct ring_buffer
*buffer
,
4481 void **data_page
, size_t len
, int cpu
, int full
)
4483 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
4484 struct ring_buffer_event
*event
;
4485 struct buffer_data_page
*bpage
;
4486 struct buffer_page
*reader
;
4487 unsigned long missed_events
;
4488 unsigned long flags
;
4489 unsigned int commit
;
4494 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4498 * If len is not big enough to hold the page header, then
4499 * we can not copy anything.
4501 if (len
<= BUF_PAGE_HDR_SIZE
)
4504 len
-= BUF_PAGE_HDR_SIZE
;
4513 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4515 reader
= rb_get_reader_page(cpu_buffer
);
4519 event
= rb_reader_event(cpu_buffer
);
4521 read
= reader
->read
;
4522 commit
= rb_page_commit(reader
);
4524 /* Check if any events were dropped */
4525 missed_events
= cpu_buffer
->lost_events
;
4528 * If this page has been partially read or
4529 * if len is not big enough to read the rest of the page or
4530 * a writer is still on the page, then
4531 * we must copy the data from the page to the buffer.
4532 * Otherwise, we can simply swap the page with the one passed in.
4534 if (read
|| (len
< (commit
- read
)) ||
4535 cpu_buffer
->reader_page
== cpu_buffer
->commit_page
) {
4536 struct buffer_data_page
*rpage
= cpu_buffer
->reader_page
->page
;
4537 unsigned int rpos
= read
;
4538 unsigned int pos
= 0;
4544 if (len
> (commit
- read
))
4545 len
= (commit
- read
);
4547 /* Always keep the time extend and data together */
4548 size
= rb_event_ts_length(event
);
4553 /* save the current timestamp, since the user will need it */
4554 save_timestamp
= cpu_buffer
->read_stamp
;
4556 /* Need to copy one event at a time */
4558 /* We need the size of one event, because
4559 * rb_advance_reader only advances by one event,
4560 * whereas rb_event_ts_length may include the size of
4561 * one or two events.
4562 * We have already ensured there's enough space if this
4563 * is a time extend. */
4564 size
= rb_event_length(event
);
4565 memcpy(bpage
->data
+ pos
, rpage
->data
+ rpos
, size
);
4569 rb_advance_reader(cpu_buffer
);
4570 rpos
= reader
->read
;
4576 event
= rb_reader_event(cpu_buffer
);
4577 /* Always keep the time extend and data together */
4578 size
= rb_event_ts_length(event
);
4579 } while (len
>= size
);
4582 local_set(&bpage
->commit
, pos
);
4583 bpage
->time_stamp
= save_timestamp
;
4585 /* we copied everything to the beginning */
4588 /* update the entry counter */
4589 cpu_buffer
->read
+= rb_page_entries(reader
);
4590 cpu_buffer
->read_bytes
+= BUF_PAGE_SIZE
;
4592 /* swap the pages */
4593 rb_init_page(bpage
);
4594 bpage
= reader
->page
;
4595 reader
->page
= *data_page
;
4596 local_set(&reader
->write
, 0);
4597 local_set(&reader
->entries
, 0);
4602 * Use the real_end for the data size,
4603 * This gives us a chance to store the lost events
4606 if (reader
->real_end
)
4607 local_set(&bpage
->commit
, reader
->real_end
);
4611 cpu_buffer
->lost_events
= 0;
4613 commit
= local_read(&bpage
->commit
);
4615 * Set a flag in the commit field if we lost events
4617 if (missed_events
) {
4618 /* If there is room at the end of the page to save the
4619 * missed events, then record it there.
4621 if (BUF_PAGE_SIZE
- commit
>= sizeof(missed_events
)) {
4622 memcpy(&bpage
->data
[commit
], &missed_events
,
4623 sizeof(missed_events
));
4624 local_add(RB_MISSED_STORED
, &bpage
->commit
);
4625 commit
+= sizeof(missed_events
);
4627 local_add(RB_MISSED_EVENTS
, &bpage
->commit
);
4631 * This page may be off to user land. Zero it out here.
4633 if (commit
< BUF_PAGE_SIZE
)
4634 memset(&bpage
->data
[commit
], 0, BUF_PAGE_SIZE
- commit
);
4637 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4642 EXPORT_SYMBOL_GPL(ring_buffer_read_page
);
4644 #ifdef CONFIG_HOTPLUG_CPU
4645 static int rb_cpu_notify(struct notifier_block
*self
,
4646 unsigned long action
, void *hcpu
)
4648 struct ring_buffer
*buffer
=
4649 container_of(self
, struct ring_buffer
, cpu_notify
);
4650 long cpu
= (long)hcpu
;
4651 int cpu_i
, nr_pages_same
;
4652 unsigned int nr_pages
;
4655 case CPU_UP_PREPARE
:
4656 case CPU_UP_PREPARE_FROZEN
:
4657 if (cpumask_test_cpu(cpu
, buffer
->cpumask
))
4662 /* check if all cpu sizes are same */
4663 for_each_buffer_cpu(buffer
, cpu_i
) {
4664 /* fill in the size from first enabled cpu */
4666 nr_pages
= buffer
->buffers
[cpu_i
]->nr_pages
;
4667 if (nr_pages
!= buffer
->buffers
[cpu_i
]->nr_pages
) {
4672 /* allocate minimum pages, user can later expand it */
4675 buffer
->buffers
[cpu
] =
4676 rb_allocate_cpu_buffer(buffer
, nr_pages
, cpu
);
4677 if (!buffer
->buffers
[cpu
]) {
4678 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4683 cpumask_set_cpu(cpu
, buffer
->cpumask
);
4685 case CPU_DOWN_PREPARE
:
4686 case CPU_DOWN_PREPARE_FROZEN
:
4689 * If we were to free the buffer, then the user would
4690 * lose any trace that was in the buffer.
4700 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4702 * This is a basic integrity check of the ring buffer.
4703 * Late in the boot cycle this test will run when configured in.
4704 * It will kick off a thread per CPU that will go into a loop
4705 * writing to the per cpu ring buffer various sizes of data.
4706 * Some of the data will be large items, some small.
4708 * Another thread is created that goes into a spin, sending out
4709 * IPIs to the other CPUs to also write into the ring buffer.
4710 * this is to test the nesting ability of the buffer.
4712 * Basic stats are recorded and reported. If something in the
4713 * ring buffer should happen that's not expected, a big warning
4714 * is displayed and all ring buffers are disabled.
4716 static struct task_struct
*rb_threads
[NR_CPUS
] __initdata
;
4718 struct rb_test_data
{
4719 struct ring_buffer
*buffer
;
4720 unsigned long events
;
4721 unsigned long bytes_written
;
4722 unsigned long bytes_alloc
;
4723 unsigned long bytes_dropped
;
4724 unsigned long events_nested
;
4725 unsigned long bytes_written_nested
;
4726 unsigned long bytes_alloc_nested
;
4727 unsigned long bytes_dropped_nested
;
4728 int min_size_nested
;
4729 int max_size_nested
;
4736 static struct rb_test_data rb_data
[NR_CPUS
] __initdata
;
4739 #define RB_TEST_BUFFER_SIZE 1048576
4741 static char rb_string
[] __initdata
=
4742 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4743 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4744 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4746 static bool rb_test_started __initdata
;
4753 static __init
int rb_write_something(struct rb_test_data
*data
, bool nested
)
4755 struct ring_buffer_event
*event
;
4756 struct rb_item
*item
;
4763 /* Have nested writes different that what is written */
4764 cnt
= data
->cnt
+ (nested
? 27 : 0);
4766 /* Multiply cnt by ~e, to make some unique increment */
4767 size
= (data
->cnt
* 68 / 25) % (sizeof(rb_string
) - 1);
4769 len
= size
+ sizeof(struct rb_item
);
4771 started
= rb_test_started
;
4772 /* read rb_test_started before checking buffer enabled */
4775 event
= ring_buffer_lock_reserve(data
->buffer
, len
);
4777 /* Ignore dropped events before test starts. */
4780 data
->bytes_dropped
+= len
;
4782 data
->bytes_dropped_nested
+= len
;
4787 event_len
= ring_buffer_event_length(event
);
4789 if (RB_WARN_ON(data
->buffer
, event_len
< len
))
4792 item
= ring_buffer_event_data(event
);
4794 memcpy(item
->str
, rb_string
, size
);
4797 data
->bytes_alloc_nested
+= event_len
;
4798 data
->bytes_written_nested
+= len
;
4799 data
->events_nested
++;
4800 if (!data
->min_size_nested
|| len
< data
->min_size_nested
)
4801 data
->min_size_nested
= len
;
4802 if (len
> data
->max_size_nested
)
4803 data
->max_size_nested
= len
;
4805 data
->bytes_alloc
+= event_len
;
4806 data
->bytes_written
+= len
;
4808 if (!data
->min_size
|| len
< data
->min_size
)
4809 data
->max_size
= len
;
4810 if (len
> data
->max_size
)
4811 data
->max_size
= len
;
4815 ring_buffer_unlock_commit(data
->buffer
, event
);
4820 static __init
int rb_test(void *arg
)
4822 struct rb_test_data
*data
= arg
;
4824 while (!kthread_should_stop()) {
4825 rb_write_something(data
, false);
4828 set_current_state(TASK_INTERRUPTIBLE
);
4829 /* Now sleep between a min of 100-300us and a max of 1ms */
4830 usleep_range(((data
->cnt
% 3) + 1) * 100, 1000);
4836 static __init
void rb_ipi(void *ignore
)
4838 struct rb_test_data
*data
;
4839 int cpu
= smp_processor_id();
4841 data
= &rb_data
[cpu
];
4842 rb_write_something(data
, true);
4845 static __init
int rb_hammer_test(void *arg
)
4847 while (!kthread_should_stop()) {
4849 /* Send an IPI to all cpus to write data! */
4850 smp_call_function(rb_ipi
, NULL
, 1);
4851 /* No sleep, but for non preempt, let others run */
4858 static __init
int test_ringbuffer(void)
4860 struct task_struct
*rb_hammer
;
4861 struct ring_buffer
*buffer
;
4865 pr_info("Running ring buffer tests...\n");
4867 buffer
= ring_buffer_alloc(RB_TEST_BUFFER_SIZE
, RB_FL_OVERWRITE
);
4868 if (WARN_ON(!buffer
))
4871 /* Disable buffer so that threads can't write to it yet */
4872 ring_buffer_record_off(buffer
);
4874 for_each_online_cpu(cpu
) {
4875 rb_data
[cpu
].buffer
= buffer
;
4876 rb_data
[cpu
].cpu
= cpu
;
4877 rb_data
[cpu
].cnt
= cpu
;
4878 rb_threads
[cpu
] = kthread_create(rb_test
, &rb_data
[cpu
],
4879 "rbtester/%d", cpu
);
4880 if (WARN_ON(!rb_threads
[cpu
])) {
4881 pr_cont("FAILED\n");
4886 kthread_bind(rb_threads
[cpu
], cpu
);
4887 wake_up_process(rb_threads
[cpu
]);
4890 /* Now create the rb hammer! */
4891 rb_hammer
= kthread_run(rb_hammer_test
, NULL
, "rbhammer");
4892 if (WARN_ON(!rb_hammer
)) {
4893 pr_cont("FAILED\n");
4898 ring_buffer_record_on(buffer
);
4900 * Show buffer is enabled before setting rb_test_started.
4901 * Yes there's a small race window where events could be
4902 * dropped and the thread wont catch it. But when a ring
4903 * buffer gets enabled, there will always be some kind of
4904 * delay before other CPUs see it. Thus, we don't care about
4905 * those dropped events. We care about events dropped after
4906 * the threads see that the buffer is active.
4909 rb_test_started
= true;
4911 set_current_state(TASK_INTERRUPTIBLE
);
4912 /* Just run for 10 seconds */;
4913 schedule_timeout(10 * HZ
);
4915 kthread_stop(rb_hammer
);
4918 for_each_online_cpu(cpu
) {
4919 if (!rb_threads
[cpu
])
4921 kthread_stop(rb_threads
[cpu
]);
4924 ring_buffer_free(buffer
);
4929 pr_info("finished\n");
4930 for_each_online_cpu(cpu
) {
4931 struct ring_buffer_event
*event
;
4932 struct rb_test_data
*data
= &rb_data
[cpu
];
4933 struct rb_item
*item
;
4934 unsigned long total_events
;
4935 unsigned long total_dropped
;
4936 unsigned long total_written
;
4937 unsigned long total_alloc
;
4938 unsigned long total_read
= 0;
4939 unsigned long total_size
= 0;
4940 unsigned long total_len
= 0;
4941 unsigned long total_lost
= 0;
4944 int small_event_size
;
4948 total_events
= data
->events
+ data
->events_nested
;
4949 total_written
= data
->bytes_written
+ data
->bytes_written_nested
;
4950 total_alloc
= data
->bytes_alloc
+ data
->bytes_alloc_nested
;
4951 total_dropped
= data
->bytes_dropped
+ data
->bytes_dropped_nested
;
4953 big_event_size
= data
->max_size
+ data
->max_size_nested
;
4954 small_event_size
= data
->min_size
+ data
->min_size_nested
;
4956 pr_info("CPU %d:\n", cpu
);
4957 pr_info(" events: %ld\n", total_events
);
4958 pr_info(" dropped bytes: %ld\n", total_dropped
);
4959 pr_info(" alloced bytes: %ld\n", total_alloc
);
4960 pr_info(" written bytes: %ld\n", total_written
);
4961 pr_info(" biggest event: %d\n", big_event_size
);
4962 pr_info(" smallest event: %d\n", small_event_size
);
4964 if (RB_WARN_ON(buffer
, total_dropped
))
4969 while ((event
= ring_buffer_consume(buffer
, cpu
, NULL
, &lost
))) {
4971 item
= ring_buffer_event_data(event
);
4972 total_len
+= ring_buffer_event_length(event
);
4973 total_size
+= item
->size
+ sizeof(struct rb_item
);
4974 if (memcmp(&item
->str
[0], rb_string
, item
->size
) != 0) {
4975 pr_info("FAILED!\n");
4976 pr_info("buffer had: %.*s\n", item
->size
, item
->str
);
4977 pr_info("expected: %.*s\n", item
->size
, rb_string
);
4978 RB_WARN_ON(buffer
, 1);
4989 pr_info(" read events: %ld\n", total_read
);
4990 pr_info(" lost events: %ld\n", total_lost
);
4991 pr_info(" total events: %ld\n", total_lost
+ total_read
);
4992 pr_info(" recorded len bytes: %ld\n", total_len
);
4993 pr_info(" recorded size bytes: %ld\n", total_size
);
4995 pr_info(" With dropped events, record len and size may not match\n"
4996 " alloced and written from above\n");
4998 if (RB_WARN_ON(buffer
, total_len
!= total_alloc
||
4999 total_size
!= total_written
))
5002 if (RB_WARN_ON(buffer
, total_lost
+ total_read
!= total_events
))
5008 pr_info("Ring buffer PASSED!\n");
5010 ring_buffer_free(buffer
);
5014 late_initcall(test_ringbuffer
);
5015 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */