4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 #include <linux/ftrace_event.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/debugfs.h>
13 #include <linux/uaccess.h>
14 #include <linux/hardirq.h>
15 #include <linux/kthread.h> /* for self test */
16 #include <linux/kmemcheck.h>
17 #include <linux/module.h>
18 #include <linux/percpu.h>
19 #include <linux/mutex.h>
20 #include <linux/delay.h>
21 #include <linux/slab.h>
22 #include <linux/init.h>
23 #include <linux/hash.h>
24 #include <linux/list.h>
25 #include <linux/cpu.h>
28 #include <asm/local.h>
30 static void update_pages_handler(struct work_struct
*work
);
33 * The ring buffer header is special. We must manually up keep it.
35 int ring_buffer_print_entry_header(struct trace_seq
*s
)
39 ret
= trace_seq_puts(s
, "# compressed entry header\n");
40 ret
= trace_seq_puts(s
, "\ttype_len : 5 bits\n");
41 ret
= trace_seq_puts(s
, "\ttime_delta : 27 bits\n");
42 ret
= trace_seq_puts(s
, "\tarray : 32 bits\n");
43 ret
= trace_seq_putc(s
, '\n');
44 ret
= trace_seq_printf(s
, "\tpadding : type == %d\n",
45 RINGBUF_TYPE_PADDING
);
46 ret
= trace_seq_printf(s
, "\ttime_extend : type == %d\n",
47 RINGBUF_TYPE_TIME_EXTEND
);
48 ret
= trace_seq_printf(s
, "\tdata max type_len == %d\n",
49 RINGBUF_TYPE_DATA_TYPE_LEN_MAX
);
55 * The ring buffer is made up of a list of pages. A separate list of pages is
56 * allocated for each CPU. A writer may only write to a buffer that is
57 * associated with the CPU it is currently executing on. A reader may read
58 * from any per cpu buffer.
60 * The reader is special. For each per cpu buffer, the reader has its own
61 * reader page. When a reader has read the entire reader page, this reader
62 * page is swapped with another page in the ring buffer.
64 * Now, as long as the writer is off the reader page, the reader can do what
65 * ever it wants with that page. The writer will never write to that page
66 * again (as long as it is out of the ring buffer).
68 * Here's some silly ASCII art.
71 * |reader| RING BUFFER
73 * +------+ +---+ +---+ +---+
82 * |reader| RING BUFFER
83 * |page |------------------v
84 * +------+ +---+ +---+ +---+
93 * |reader| RING BUFFER
94 * |page |------------------v
95 * +------+ +---+ +---+ +---+
100 * +------------------------------+
104 * |buffer| RING BUFFER
105 * |page |------------------v
106 * +------+ +---+ +---+ +---+
108 * | New +---+ +---+ +---+
111 * +------------------------------+
114 * After we make this swap, the reader can hand this page off to the splice
115 * code and be done with it. It can even allocate a new page if it needs to
116 * and swap that into the ring buffer.
118 * We will be using cmpxchg soon to make all this lockless.
123 * A fast way to enable or disable all ring buffers is to
124 * call tracing_on or tracing_off. Turning off the ring buffers
125 * prevents all ring buffers from being recorded to.
126 * Turning this switch on, makes it OK to write to the
127 * ring buffer, if the ring buffer is enabled itself.
129 * There's three layers that must be on in order to write
130 * to the ring buffer.
132 * 1) This global flag must be set.
133 * 2) The ring buffer must be enabled for recording.
134 * 3) The per cpu buffer must be enabled for recording.
136 * In case of an anomaly, this global flag has a bit set that
137 * will permantly disable all ring buffers.
141 * Global flag to disable all recording to ring buffers
142 * This has two bits: ON, DISABLED
146 * 0 0 : ring buffers are off
147 * 1 0 : ring buffers are on
148 * X 1 : ring buffers are permanently disabled
152 RB_BUFFERS_ON_BIT
= 0,
153 RB_BUFFERS_DISABLED_BIT
= 1,
157 RB_BUFFERS_ON
= 1 << RB_BUFFERS_ON_BIT
,
158 RB_BUFFERS_DISABLED
= 1 << RB_BUFFERS_DISABLED_BIT
,
161 static unsigned long ring_buffer_flags __read_mostly
= RB_BUFFERS_ON
;
163 /* Used for individual buffers (after the counter) */
164 #define RB_BUFFER_OFF (1 << 20)
166 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
169 * tracing_off_permanent - permanently disable ring buffers
171 * This function, once called, will disable all ring buffers
174 void tracing_off_permanent(void)
176 set_bit(RB_BUFFERS_DISABLED_BIT
, &ring_buffer_flags
);
179 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
180 #define RB_ALIGNMENT 4U
181 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
182 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
184 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
185 # define RB_FORCE_8BYTE_ALIGNMENT 0
186 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
188 # define RB_FORCE_8BYTE_ALIGNMENT 1
189 # define RB_ARCH_ALIGNMENT 8U
192 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
194 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
195 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
198 RB_LEN_TIME_EXTEND
= 8,
199 RB_LEN_TIME_STAMP
= 16,
202 #define skip_time_extend(event) \
203 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
205 static inline int rb_null_event(struct ring_buffer_event
*event
)
207 return event
->type_len
== RINGBUF_TYPE_PADDING
&& !event
->time_delta
;
210 static void rb_event_set_padding(struct ring_buffer_event
*event
)
212 /* padding has a NULL time_delta */
213 event
->type_len
= RINGBUF_TYPE_PADDING
;
214 event
->time_delta
= 0;
218 rb_event_data_length(struct ring_buffer_event
*event
)
223 length
= event
->type_len
* RB_ALIGNMENT
;
225 length
= event
->array
[0];
226 return length
+ RB_EVNT_HDR_SIZE
;
230 * Return the length of the given event. Will return
231 * the length of the time extend if the event is a
234 static inline unsigned
235 rb_event_length(struct ring_buffer_event
*event
)
237 switch (event
->type_len
) {
238 case RINGBUF_TYPE_PADDING
:
239 if (rb_null_event(event
))
242 return event
->array
[0] + RB_EVNT_HDR_SIZE
;
244 case RINGBUF_TYPE_TIME_EXTEND
:
245 return RB_LEN_TIME_EXTEND
;
247 case RINGBUF_TYPE_TIME_STAMP
:
248 return RB_LEN_TIME_STAMP
;
250 case RINGBUF_TYPE_DATA
:
251 return rb_event_data_length(event
);
260 * Return total length of time extend and data,
261 * or just the event length for all other events.
263 static inline unsigned
264 rb_event_ts_length(struct ring_buffer_event
*event
)
268 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
) {
269 /* time extends include the data event after it */
270 len
= RB_LEN_TIME_EXTEND
;
271 event
= skip_time_extend(event
);
273 return len
+ rb_event_length(event
);
277 * ring_buffer_event_length - return the length of the event
278 * @event: the event to get the length of
280 * Returns the size of the data load of a data event.
281 * If the event is something other than a data event, it
282 * returns the size of the event itself. With the exception
283 * of a TIME EXTEND, where it still returns the size of the
284 * data load of the data event after it.
286 unsigned ring_buffer_event_length(struct ring_buffer_event
*event
)
290 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
291 event
= skip_time_extend(event
);
293 length
= rb_event_length(event
);
294 if (event
->type_len
> RINGBUF_TYPE_DATA_TYPE_LEN_MAX
)
296 length
-= RB_EVNT_HDR_SIZE
;
297 if (length
> RB_MAX_SMALL_DATA
+ sizeof(event
->array
[0]))
298 length
-= sizeof(event
->array
[0]);
301 EXPORT_SYMBOL_GPL(ring_buffer_event_length
);
303 /* inline for ring buffer fast paths */
305 rb_event_data(struct ring_buffer_event
*event
)
307 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
308 event
= skip_time_extend(event
);
309 BUG_ON(event
->type_len
> RINGBUF_TYPE_DATA_TYPE_LEN_MAX
);
310 /* If length is in len field, then array[0] has the data */
312 return (void *)&event
->array
[0];
313 /* Otherwise length is in array[0] and array[1] has the data */
314 return (void *)&event
->array
[1];
318 * ring_buffer_event_data - return the data of the event
319 * @event: the event to get the data from
321 void *ring_buffer_event_data(struct ring_buffer_event
*event
)
323 return rb_event_data(event
);
325 EXPORT_SYMBOL_GPL(ring_buffer_event_data
);
327 #define for_each_buffer_cpu(buffer, cpu) \
328 for_each_cpu(cpu, buffer->cpumask)
331 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
332 #define TS_DELTA_TEST (~TS_MASK)
334 /* Flag when events were overwritten */
335 #define RB_MISSED_EVENTS (1 << 31)
336 /* Missed count stored at end */
337 #define RB_MISSED_STORED (1 << 30)
339 struct buffer_data_page
{
340 u64 time_stamp
; /* page time stamp */
341 local_t commit
; /* write committed index */
342 unsigned char data
[] RB_ALIGN_DATA
; /* data of buffer page */
346 * Note, the buffer_page list must be first. The buffer pages
347 * are allocated in cache lines, which means that each buffer
348 * page will be at the beginning of a cache line, and thus
349 * the least significant bits will be zero. We use this to
350 * add flags in the list struct pointers, to make the ring buffer
354 struct list_head list
; /* list of buffer pages */
355 local_t write
; /* index for next write */
356 unsigned read
; /* index for next read */
357 local_t entries
; /* entries on this page */
358 unsigned long real_end
; /* real end of data */
359 struct buffer_data_page
*page
; /* Actual data page */
363 * The buffer page counters, write and entries, must be reset
364 * atomically when crossing page boundaries. To synchronize this
365 * update, two counters are inserted into the number. One is
366 * the actual counter for the write position or count on the page.
368 * The other is a counter of updaters. Before an update happens
369 * the update partition of the counter is incremented. This will
370 * allow the updater to update the counter atomically.
372 * The counter is 20 bits, and the state data is 12.
374 #define RB_WRITE_MASK 0xfffff
375 #define RB_WRITE_INTCNT (1 << 20)
377 static void rb_init_page(struct buffer_data_page
*bpage
)
379 local_set(&bpage
->commit
, 0);
383 * ring_buffer_page_len - the size of data on the page.
384 * @page: The page to read
386 * Returns the amount of data on the page, including buffer page header.
388 size_t ring_buffer_page_len(void *page
)
390 return local_read(&((struct buffer_data_page
*)page
)->commit
)
395 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
398 static void free_buffer_page(struct buffer_page
*bpage
)
400 free_page((unsigned long)bpage
->page
);
405 * We need to fit the time_stamp delta into 27 bits.
407 static inline int test_time_stamp(u64 delta
)
409 if (delta
& TS_DELTA_TEST
)
414 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
416 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
417 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
419 int ring_buffer_print_page_header(struct trace_seq
*s
)
421 struct buffer_data_page field
;
424 ret
= trace_seq_printf(s
, "\tfield: u64 timestamp;\t"
425 "offset:0;\tsize:%u;\tsigned:%u;\n",
426 (unsigned int)sizeof(field
.time_stamp
),
427 (unsigned int)is_signed_type(u64
));
429 ret
= trace_seq_printf(s
, "\tfield: local_t commit;\t"
430 "offset:%u;\tsize:%u;\tsigned:%u;\n",
431 (unsigned int)offsetof(typeof(field
), commit
),
432 (unsigned int)sizeof(field
.commit
),
433 (unsigned int)is_signed_type(long));
435 ret
= trace_seq_printf(s
, "\tfield: int overwrite;\t"
436 "offset:%u;\tsize:%u;\tsigned:%u;\n",
437 (unsigned int)offsetof(typeof(field
), commit
),
439 (unsigned int)is_signed_type(long));
441 ret
= trace_seq_printf(s
, "\tfield: char data;\t"
442 "offset:%u;\tsize:%u;\tsigned:%u;\n",
443 (unsigned int)offsetof(typeof(field
), data
),
444 (unsigned int)BUF_PAGE_SIZE
,
445 (unsigned int)is_signed_type(char));
451 struct irq_work work
;
452 wait_queue_head_t waiters
;
453 bool waiters_pending
;
457 * head_page == tail_page && head == tail then buffer is empty.
459 struct ring_buffer_per_cpu
{
461 atomic_t record_disabled
;
462 struct ring_buffer
*buffer
;
463 raw_spinlock_t reader_lock
; /* serialize readers */
464 arch_spinlock_t lock
;
465 struct lock_class_key lock_key
;
466 unsigned int nr_pages
;
467 struct list_head
*pages
;
468 struct buffer_page
*head_page
; /* read from head */
469 struct buffer_page
*tail_page
; /* write to tail */
470 struct buffer_page
*commit_page
; /* committed pages */
471 struct buffer_page
*reader_page
;
472 unsigned long lost_events
;
473 unsigned long last_overrun
;
474 local_t entries_bytes
;
477 local_t commit_overrun
;
478 local_t dropped_events
;
482 unsigned long read_bytes
;
485 /* ring buffer pages to update, > 0 to add, < 0 to remove */
486 int nr_pages_to_update
;
487 struct list_head new_pages
; /* new pages to add */
488 struct work_struct update_pages_work
;
489 struct completion update_done
;
491 struct rb_irq_work irq_work
;
497 atomic_t record_disabled
;
498 atomic_t resize_disabled
;
499 cpumask_var_t cpumask
;
501 struct lock_class_key
*reader_lock_key
;
505 struct ring_buffer_per_cpu
**buffers
;
507 #ifdef CONFIG_HOTPLUG_CPU
508 struct notifier_block cpu_notify
;
512 struct rb_irq_work irq_work
;
515 struct ring_buffer_iter
{
516 struct ring_buffer_per_cpu
*cpu_buffer
;
518 struct buffer_page
*head_page
;
519 struct buffer_page
*cache_reader_page
;
520 unsigned long cache_read
;
525 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
527 * Schedules a delayed work to wake up any task that is blocked on the
528 * ring buffer waiters queue.
530 static void rb_wake_up_waiters(struct irq_work
*work
)
532 struct rb_irq_work
*rbwork
= container_of(work
, struct rb_irq_work
, work
);
534 wake_up_all(&rbwork
->waiters
);
538 * ring_buffer_wait - wait for input to the ring buffer
539 * @buffer: buffer to wait on
540 * @cpu: the cpu buffer to wait on
541 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
543 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
544 * as data is added to any of the @buffer's cpu buffers. Otherwise
545 * it will wait for data to be added to a specific cpu buffer.
547 int ring_buffer_wait(struct ring_buffer
*buffer
, int cpu
, bool full
)
549 struct ring_buffer_per_cpu
*uninitialized_var(cpu_buffer
);
551 struct rb_irq_work
*work
;
555 * Depending on what the caller is waiting for, either any
556 * data in any cpu buffer, or a specific buffer, put the
557 * caller on the appropriate wait queue.
559 if (cpu
== RING_BUFFER_ALL_CPUS
)
560 work
= &buffer
->irq_work
;
562 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
564 cpu_buffer
= buffer
->buffers
[cpu
];
565 work
= &cpu_buffer
->irq_work
;
570 prepare_to_wait(&work
->waiters
, &wait
, TASK_INTERRUPTIBLE
);
573 * The events can happen in critical sections where
574 * checking a work queue can cause deadlocks.
575 * After adding a task to the queue, this flag is set
576 * only to notify events to try to wake up the queue
579 * We don't clear it even if the buffer is no longer
580 * empty. The flag only causes the next event to run
581 * irq_work to do the work queue wake up. The worse
582 * that can happen if we race with !trace_empty() is that
583 * an event will cause an irq_work to try to wake up
586 * There's no reason to protect this flag either, as
587 * the work queue and irq_work logic will do the necessary
588 * synchronization for the wake ups. The only thing
589 * that is necessary is that the wake up happens after
590 * a task has been queued. It's OK for spurious wake ups.
592 work
->waiters_pending
= true;
594 if (signal_pending(current
)) {
599 if (cpu
== RING_BUFFER_ALL_CPUS
&& !ring_buffer_empty(buffer
))
602 if (cpu
!= RING_BUFFER_ALL_CPUS
&&
603 !ring_buffer_empty_cpu(buffer
, cpu
)) {
610 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
611 pagebusy
= cpu_buffer
->reader_page
== cpu_buffer
->commit_page
;
612 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
621 finish_wait(&work
->waiters
, &wait
);
627 * ring_buffer_poll_wait - poll on buffer input
628 * @buffer: buffer to wait on
629 * @cpu: the cpu buffer to wait on
630 * @filp: the file descriptor
631 * @poll_table: The poll descriptor
633 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
634 * as data is added to any of the @buffer's cpu buffers. Otherwise
635 * it will wait for data to be added to a specific cpu buffer.
637 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
640 int ring_buffer_poll_wait(struct ring_buffer
*buffer
, int cpu
,
641 struct file
*filp
, poll_table
*poll_table
)
643 struct ring_buffer_per_cpu
*cpu_buffer
;
644 struct rb_irq_work
*work
;
646 if (cpu
== RING_BUFFER_ALL_CPUS
)
647 work
= &buffer
->irq_work
;
649 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
652 cpu_buffer
= buffer
->buffers
[cpu
];
653 work
= &cpu_buffer
->irq_work
;
656 poll_wait(filp
, &work
->waiters
, poll_table
);
657 work
->waiters_pending
= true;
659 * There's a tight race between setting the waiters_pending and
660 * checking if the ring buffer is empty. Once the waiters_pending bit
661 * is set, the next event will wake the task up, but we can get stuck
662 * if there's only a single event in.
664 * FIXME: Ideally, we need a memory barrier on the writer side as well,
665 * but adding a memory barrier to all events will cause too much of a
666 * performance hit in the fast path. We only need a memory barrier when
667 * the buffer goes from empty to having content. But as this race is
668 * extremely small, and it's not a problem if another event comes in, we
673 if ((cpu
== RING_BUFFER_ALL_CPUS
&& !ring_buffer_empty(buffer
)) ||
674 (cpu
!= RING_BUFFER_ALL_CPUS
&& !ring_buffer_empty_cpu(buffer
, cpu
)))
675 return POLLIN
| POLLRDNORM
;
679 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
680 #define RB_WARN_ON(b, cond) \
682 int _____ret = unlikely(cond); \
684 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
685 struct ring_buffer_per_cpu *__b = \
687 atomic_inc(&__b->buffer->record_disabled); \
689 atomic_inc(&b->record_disabled); \
695 /* Up this if you want to test the TIME_EXTENTS and normalization */
696 #define DEBUG_SHIFT 0
698 static inline u64
rb_time_stamp(struct ring_buffer
*buffer
)
700 /* shift to debug/test normalization and TIME_EXTENTS */
701 return buffer
->clock() << DEBUG_SHIFT
;
704 u64
ring_buffer_time_stamp(struct ring_buffer
*buffer
, int cpu
)
708 preempt_disable_notrace();
709 time
= rb_time_stamp(buffer
);
710 preempt_enable_no_resched_notrace();
714 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp
);
716 void ring_buffer_normalize_time_stamp(struct ring_buffer
*buffer
,
719 /* Just stupid testing the normalize function and deltas */
722 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp
);
725 * Making the ring buffer lockless makes things tricky.
726 * Although writes only happen on the CPU that they are on,
727 * and they only need to worry about interrupts. Reads can
730 * The reader page is always off the ring buffer, but when the
731 * reader finishes with a page, it needs to swap its page with
732 * a new one from the buffer. The reader needs to take from
733 * the head (writes go to the tail). But if a writer is in overwrite
734 * mode and wraps, it must push the head page forward.
736 * Here lies the problem.
738 * The reader must be careful to replace only the head page, and
739 * not another one. As described at the top of the file in the
740 * ASCII art, the reader sets its old page to point to the next
741 * page after head. It then sets the page after head to point to
742 * the old reader page. But if the writer moves the head page
743 * during this operation, the reader could end up with the tail.
745 * We use cmpxchg to help prevent this race. We also do something
746 * special with the page before head. We set the LSB to 1.
748 * When the writer must push the page forward, it will clear the
749 * bit that points to the head page, move the head, and then set
750 * the bit that points to the new head page.
752 * We also don't want an interrupt coming in and moving the head
753 * page on another writer. Thus we use the second LSB to catch
756 * head->list->prev->next bit 1 bit 0
759 * Points to head page 0 1
762 * Note we can not trust the prev pointer of the head page, because:
764 * +----+ +-----+ +-----+
765 * | |------>| T |---X--->| N |
767 * +----+ +-----+ +-----+
770 * +----------| R |----------+ |
774 * Key: ---X--> HEAD flag set in pointer
779 * (see __rb_reserve_next() to see where this happens)
781 * What the above shows is that the reader just swapped out
782 * the reader page with a page in the buffer, but before it
783 * could make the new header point back to the new page added
784 * it was preempted by a writer. The writer moved forward onto
785 * the new page added by the reader and is about to move forward
788 * You can see, it is legitimate for the previous pointer of
789 * the head (or any page) not to point back to itself. But only
793 #define RB_PAGE_NORMAL 0UL
794 #define RB_PAGE_HEAD 1UL
795 #define RB_PAGE_UPDATE 2UL
798 #define RB_FLAG_MASK 3UL
800 /* PAGE_MOVED is not part of the mask */
801 #define RB_PAGE_MOVED 4UL
804 * rb_list_head - remove any bit
806 static struct list_head
*rb_list_head(struct list_head
*list
)
808 unsigned long val
= (unsigned long)list
;
810 return (struct list_head
*)(val
& ~RB_FLAG_MASK
);
814 * rb_is_head_page - test if the given page is the head page
816 * Because the reader may move the head_page pointer, we can
817 * not trust what the head page is (it may be pointing to
818 * the reader page). But if the next page is a header page,
819 * its flags will be non zero.
822 rb_is_head_page(struct ring_buffer_per_cpu
*cpu_buffer
,
823 struct buffer_page
*page
, struct list_head
*list
)
827 val
= (unsigned long)list
->next
;
829 if ((val
& ~RB_FLAG_MASK
) != (unsigned long)&page
->list
)
830 return RB_PAGE_MOVED
;
832 return val
& RB_FLAG_MASK
;
838 * The unique thing about the reader page, is that, if the
839 * writer is ever on it, the previous pointer never points
840 * back to the reader page.
842 static int rb_is_reader_page(struct buffer_page
*page
)
844 struct list_head
*list
= page
->list
.prev
;
846 return rb_list_head(list
->next
) != &page
->list
;
850 * rb_set_list_to_head - set a list_head to be pointing to head.
852 static void rb_set_list_to_head(struct ring_buffer_per_cpu
*cpu_buffer
,
853 struct list_head
*list
)
857 ptr
= (unsigned long *)&list
->next
;
858 *ptr
|= RB_PAGE_HEAD
;
859 *ptr
&= ~RB_PAGE_UPDATE
;
863 * rb_head_page_activate - sets up head page
865 static void rb_head_page_activate(struct ring_buffer_per_cpu
*cpu_buffer
)
867 struct buffer_page
*head
;
869 head
= cpu_buffer
->head_page
;
874 * Set the previous list pointer to have the HEAD flag.
876 rb_set_list_to_head(cpu_buffer
, head
->list
.prev
);
879 static void rb_list_head_clear(struct list_head
*list
)
881 unsigned long *ptr
= (unsigned long *)&list
->next
;
883 *ptr
&= ~RB_FLAG_MASK
;
887 * rb_head_page_dactivate - clears head page ptr (for free list)
890 rb_head_page_deactivate(struct ring_buffer_per_cpu
*cpu_buffer
)
892 struct list_head
*hd
;
894 /* Go through the whole list and clear any pointers found. */
895 rb_list_head_clear(cpu_buffer
->pages
);
897 list_for_each(hd
, cpu_buffer
->pages
)
898 rb_list_head_clear(hd
);
901 static int rb_head_page_set(struct ring_buffer_per_cpu
*cpu_buffer
,
902 struct buffer_page
*head
,
903 struct buffer_page
*prev
,
904 int old_flag
, int new_flag
)
906 struct list_head
*list
;
907 unsigned long val
= (unsigned long)&head
->list
;
912 val
&= ~RB_FLAG_MASK
;
914 ret
= cmpxchg((unsigned long *)&list
->next
,
915 val
| old_flag
, val
| new_flag
);
917 /* check if the reader took the page */
918 if ((ret
& ~RB_FLAG_MASK
) != val
)
919 return RB_PAGE_MOVED
;
921 return ret
& RB_FLAG_MASK
;
924 static int rb_head_page_set_update(struct ring_buffer_per_cpu
*cpu_buffer
,
925 struct buffer_page
*head
,
926 struct buffer_page
*prev
,
929 return rb_head_page_set(cpu_buffer
, head
, prev
,
930 old_flag
, RB_PAGE_UPDATE
);
933 static int rb_head_page_set_head(struct ring_buffer_per_cpu
*cpu_buffer
,
934 struct buffer_page
*head
,
935 struct buffer_page
*prev
,
938 return rb_head_page_set(cpu_buffer
, head
, prev
,
939 old_flag
, RB_PAGE_HEAD
);
942 static int rb_head_page_set_normal(struct ring_buffer_per_cpu
*cpu_buffer
,
943 struct buffer_page
*head
,
944 struct buffer_page
*prev
,
947 return rb_head_page_set(cpu_buffer
, head
, prev
,
948 old_flag
, RB_PAGE_NORMAL
);
951 static inline void rb_inc_page(struct ring_buffer_per_cpu
*cpu_buffer
,
952 struct buffer_page
**bpage
)
954 struct list_head
*p
= rb_list_head((*bpage
)->list
.next
);
956 *bpage
= list_entry(p
, struct buffer_page
, list
);
959 static struct buffer_page
*
960 rb_set_head_page(struct ring_buffer_per_cpu
*cpu_buffer
)
962 struct buffer_page
*head
;
963 struct buffer_page
*page
;
964 struct list_head
*list
;
967 if (RB_WARN_ON(cpu_buffer
, !cpu_buffer
->head_page
))
971 list
= cpu_buffer
->pages
;
972 if (RB_WARN_ON(cpu_buffer
, rb_list_head(list
->prev
->next
) != list
))
975 page
= head
= cpu_buffer
->head_page
;
977 * It is possible that the writer moves the header behind
978 * where we started, and we miss in one loop.
979 * A second loop should grab the header, but we'll do
980 * three loops just because I'm paranoid.
982 for (i
= 0; i
< 3; i
++) {
984 if (rb_is_head_page(cpu_buffer
, page
, page
->list
.prev
)) {
985 cpu_buffer
->head_page
= page
;
988 rb_inc_page(cpu_buffer
, &page
);
989 } while (page
!= head
);
992 RB_WARN_ON(cpu_buffer
, 1);
997 static int rb_head_page_replace(struct buffer_page
*old
,
998 struct buffer_page
*new)
1000 unsigned long *ptr
= (unsigned long *)&old
->list
.prev
->next
;
1004 val
= *ptr
& ~RB_FLAG_MASK
;
1005 val
|= RB_PAGE_HEAD
;
1007 ret
= cmpxchg(ptr
, val
, (unsigned long)&new->list
);
1013 * rb_tail_page_update - move the tail page forward
1015 * Returns 1 if moved tail page, 0 if someone else did.
1017 static int rb_tail_page_update(struct ring_buffer_per_cpu
*cpu_buffer
,
1018 struct buffer_page
*tail_page
,
1019 struct buffer_page
*next_page
)
1021 struct buffer_page
*old_tail
;
1022 unsigned long old_entries
;
1023 unsigned long old_write
;
1027 * The tail page now needs to be moved forward.
1029 * We need to reset the tail page, but without messing
1030 * with possible erasing of data brought in by interrupts
1031 * that have moved the tail page and are currently on it.
1033 * We add a counter to the write field to denote this.
1035 old_write
= local_add_return(RB_WRITE_INTCNT
, &next_page
->write
);
1036 old_entries
= local_add_return(RB_WRITE_INTCNT
, &next_page
->entries
);
1039 * Just make sure we have seen our old_write and synchronize
1040 * with any interrupts that come in.
1045 * If the tail page is still the same as what we think
1046 * it is, then it is up to us to update the tail
1049 if (tail_page
== cpu_buffer
->tail_page
) {
1050 /* Zero the write counter */
1051 unsigned long val
= old_write
& ~RB_WRITE_MASK
;
1052 unsigned long eval
= old_entries
& ~RB_WRITE_MASK
;
1055 * This will only succeed if an interrupt did
1056 * not come in and change it. In which case, we
1057 * do not want to modify it.
1059 * We add (void) to let the compiler know that we do not care
1060 * about the return value of these functions. We use the
1061 * cmpxchg to only update if an interrupt did not already
1062 * do it for us. If the cmpxchg fails, we don't care.
1064 (void)local_cmpxchg(&next_page
->write
, old_write
, val
);
1065 (void)local_cmpxchg(&next_page
->entries
, old_entries
, eval
);
1068 * No need to worry about races with clearing out the commit.
1069 * it only can increment when a commit takes place. But that
1070 * only happens in the outer most nested commit.
1072 local_set(&next_page
->page
->commit
, 0);
1074 old_tail
= cmpxchg(&cpu_buffer
->tail_page
,
1075 tail_page
, next_page
);
1077 if (old_tail
== tail_page
)
1084 static int rb_check_bpage(struct ring_buffer_per_cpu
*cpu_buffer
,
1085 struct buffer_page
*bpage
)
1087 unsigned long val
= (unsigned long)bpage
;
1089 if (RB_WARN_ON(cpu_buffer
, val
& RB_FLAG_MASK
))
1096 * rb_check_list - make sure a pointer to a list has the last bits zero
1098 static int rb_check_list(struct ring_buffer_per_cpu
*cpu_buffer
,
1099 struct list_head
*list
)
1101 if (RB_WARN_ON(cpu_buffer
, rb_list_head(list
->prev
) != list
->prev
))
1103 if (RB_WARN_ON(cpu_buffer
, rb_list_head(list
->next
) != list
->next
))
1109 * rb_check_pages - integrity check of buffer pages
1110 * @cpu_buffer: CPU buffer with pages to test
1112 * As a safety measure we check to make sure the data pages have not
1115 static int rb_check_pages(struct ring_buffer_per_cpu
*cpu_buffer
)
1117 struct list_head
*head
= cpu_buffer
->pages
;
1118 struct buffer_page
*bpage
, *tmp
;
1120 /* Reset the head page if it exists */
1121 if (cpu_buffer
->head_page
)
1122 rb_set_head_page(cpu_buffer
);
1124 rb_head_page_deactivate(cpu_buffer
);
1126 if (RB_WARN_ON(cpu_buffer
, head
->next
->prev
!= head
))
1128 if (RB_WARN_ON(cpu_buffer
, head
->prev
->next
!= head
))
1131 if (rb_check_list(cpu_buffer
, head
))
1134 list_for_each_entry_safe(bpage
, tmp
, head
, list
) {
1135 if (RB_WARN_ON(cpu_buffer
,
1136 bpage
->list
.next
->prev
!= &bpage
->list
))
1138 if (RB_WARN_ON(cpu_buffer
,
1139 bpage
->list
.prev
->next
!= &bpage
->list
))
1141 if (rb_check_list(cpu_buffer
, &bpage
->list
))
1145 rb_head_page_activate(cpu_buffer
);
1150 static int __rb_allocate_pages(int nr_pages
, struct list_head
*pages
, int cpu
)
1153 struct buffer_page
*bpage
, *tmp
;
1155 for (i
= 0; i
< nr_pages
; i
++) {
1158 * __GFP_NORETRY flag makes sure that the allocation fails
1159 * gracefully without invoking oom-killer and the system is
1162 bpage
= kzalloc_node(ALIGN(sizeof(*bpage
), cache_line_size()),
1163 GFP_KERNEL
| __GFP_NORETRY
,
1168 list_add(&bpage
->list
, pages
);
1170 page
= alloc_pages_node(cpu_to_node(cpu
),
1171 GFP_KERNEL
| __GFP_NORETRY
, 0);
1174 bpage
->page
= page_address(page
);
1175 rb_init_page(bpage
->page
);
1181 list_for_each_entry_safe(bpage
, tmp
, pages
, list
) {
1182 list_del_init(&bpage
->list
);
1183 free_buffer_page(bpage
);
1189 static int rb_allocate_pages(struct ring_buffer_per_cpu
*cpu_buffer
,
1196 if (__rb_allocate_pages(nr_pages
, &pages
, cpu_buffer
->cpu
))
1200 * The ring buffer page list is a circular list that does not
1201 * start and end with a list head. All page list items point to
1204 cpu_buffer
->pages
= pages
.next
;
1207 cpu_buffer
->nr_pages
= nr_pages
;
1209 rb_check_pages(cpu_buffer
);
1214 static struct ring_buffer_per_cpu
*
1215 rb_allocate_cpu_buffer(struct ring_buffer
*buffer
, int nr_pages
, int cpu
)
1217 struct ring_buffer_per_cpu
*cpu_buffer
;
1218 struct buffer_page
*bpage
;
1222 cpu_buffer
= kzalloc_node(ALIGN(sizeof(*cpu_buffer
), cache_line_size()),
1223 GFP_KERNEL
, cpu_to_node(cpu
));
1227 cpu_buffer
->cpu
= cpu
;
1228 cpu_buffer
->buffer
= buffer
;
1229 raw_spin_lock_init(&cpu_buffer
->reader_lock
);
1230 lockdep_set_class(&cpu_buffer
->reader_lock
, buffer
->reader_lock_key
);
1231 cpu_buffer
->lock
= (arch_spinlock_t
)__ARCH_SPIN_LOCK_UNLOCKED
;
1232 INIT_WORK(&cpu_buffer
->update_pages_work
, update_pages_handler
);
1233 init_completion(&cpu_buffer
->update_done
);
1234 init_irq_work(&cpu_buffer
->irq_work
.work
, rb_wake_up_waiters
);
1235 init_waitqueue_head(&cpu_buffer
->irq_work
.waiters
);
1237 bpage
= kzalloc_node(ALIGN(sizeof(*bpage
), cache_line_size()),
1238 GFP_KERNEL
, cpu_to_node(cpu
));
1240 goto fail_free_buffer
;
1242 rb_check_bpage(cpu_buffer
, bpage
);
1244 cpu_buffer
->reader_page
= bpage
;
1245 page
= alloc_pages_node(cpu_to_node(cpu
), GFP_KERNEL
, 0);
1247 goto fail_free_reader
;
1248 bpage
->page
= page_address(page
);
1249 rb_init_page(bpage
->page
);
1251 INIT_LIST_HEAD(&cpu_buffer
->reader_page
->list
);
1252 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
1254 ret
= rb_allocate_pages(cpu_buffer
, nr_pages
);
1256 goto fail_free_reader
;
1258 cpu_buffer
->head_page
1259 = list_entry(cpu_buffer
->pages
, struct buffer_page
, list
);
1260 cpu_buffer
->tail_page
= cpu_buffer
->commit_page
= cpu_buffer
->head_page
;
1262 rb_head_page_activate(cpu_buffer
);
1267 free_buffer_page(cpu_buffer
->reader_page
);
1274 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu
*cpu_buffer
)
1276 struct list_head
*head
= cpu_buffer
->pages
;
1277 struct buffer_page
*bpage
, *tmp
;
1279 free_buffer_page(cpu_buffer
->reader_page
);
1281 rb_head_page_deactivate(cpu_buffer
);
1284 list_for_each_entry_safe(bpage
, tmp
, head
, list
) {
1285 list_del_init(&bpage
->list
);
1286 free_buffer_page(bpage
);
1288 bpage
= list_entry(head
, struct buffer_page
, list
);
1289 free_buffer_page(bpage
);
1295 #ifdef CONFIG_HOTPLUG_CPU
1296 static int rb_cpu_notify(struct notifier_block
*self
,
1297 unsigned long action
, void *hcpu
);
1301 * __ring_buffer_alloc - allocate a new ring_buffer
1302 * @size: the size in bytes per cpu that is needed.
1303 * @flags: attributes to set for the ring buffer.
1305 * Currently the only flag that is available is the RB_FL_OVERWRITE
1306 * flag. This flag means that the buffer will overwrite old data
1307 * when the buffer wraps. If this flag is not set, the buffer will
1308 * drop data when the tail hits the head.
1310 struct ring_buffer
*__ring_buffer_alloc(unsigned long size
, unsigned flags
,
1311 struct lock_class_key
*key
)
1313 struct ring_buffer
*buffer
;
1317 /* keep it in its own cache line */
1318 buffer
= kzalloc(ALIGN(sizeof(*buffer
), cache_line_size()),
1323 if (!alloc_cpumask_var(&buffer
->cpumask
, GFP_KERNEL
))
1324 goto fail_free_buffer
;
1326 nr_pages
= DIV_ROUND_UP(size
, BUF_PAGE_SIZE
);
1327 buffer
->flags
= flags
;
1328 buffer
->clock
= trace_clock_local
;
1329 buffer
->reader_lock_key
= key
;
1331 init_irq_work(&buffer
->irq_work
.work
, rb_wake_up_waiters
);
1332 init_waitqueue_head(&buffer
->irq_work
.waiters
);
1334 /* need at least two pages */
1339 * In case of non-hotplug cpu, if the ring-buffer is allocated
1340 * in early initcall, it will not be notified of secondary cpus.
1341 * In that off case, we need to allocate for all possible cpus.
1343 #ifdef CONFIG_HOTPLUG_CPU
1344 cpu_notifier_register_begin();
1345 cpumask_copy(buffer
->cpumask
, cpu_online_mask
);
1347 cpumask_copy(buffer
->cpumask
, cpu_possible_mask
);
1349 buffer
->cpus
= nr_cpu_ids
;
1351 bsize
= sizeof(void *) * nr_cpu_ids
;
1352 buffer
->buffers
= kzalloc(ALIGN(bsize
, cache_line_size()),
1354 if (!buffer
->buffers
)
1355 goto fail_free_cpumask
;
1357 for_each_buffer_cpu(buffer
, cpu
) {
1358 buffer
->buffers
[cpu
] =
1359 rb_allocate_cpu_buffer(buffer
, nr_pages
, cpu
);
1360 if (!buffer
->buffers
[cpu
])
1361 goto fail_free_buffers
;
1364 #ifdef CONFIG_HOTPLUG_CPU
1365 buffer
->cpu_notify
.notifier_call
= rb_cpu_notify
;
1366 buffer
->cpu_notify
.priority
= 0;
1367 __register_cpu_notifier(&buffer
->cpu_notify
);
1368 cpu_notifier_register_done();
1371 mutex_init(&buffer
->mutex
);
1376 for_each_buffer_cpu(buffer
, cpu
) {
1377 if (buffer
->buffers
[cpu
])
1378 rb_free_cpu_buffer(buffer
->buffers
[cpu
]);
1380 kfree(buffer
->buffers
);
1383 free_cpumask_var(buffer
->cpumask
);
1384 #ifdef CONFIG_HOTPLUG_CPU
1385 cpu_notifier_register_done();
1392 EXPORT_SYMBOL_GPL(__ring_buffer_alloc
);
1395 * ring_buffer_free - free a ring buffer.
1396 * @buffer: the buffer to free.
1399 ring_buffer_free(struct ring_buffer
*buffer
)
1403 #ifdef CONFIG_HOTPLUG_CPU
1404 cpu_notifier_register_begin();
1405 __unregister_cpu_notifier(&buffer
->cpu_notify
);
1408 for_each_buffer_cpu(buffer
, cpu
)
1409 rb_free_cpu_buffer(buffer
->buffers
[cpu
]);
1411 #ifdef CONFIG_HOTPLUG_CPU
1412 cpu_notifier_register_done();
1415 kfree(buffer
->buffers
);
1416 free_cpumask_var(buffer
->cpumask
);
1420 EXPORT_SYMBOL_GPL(ring_buffer_free
);
1422 void ring_buffer_set_clock(struct ring_buffer
*buffer
,
1425 buffer
->clock
= clock
;
1428 static void rb_reset_cpu(struct ring_buffer_per_cpu
*cpu_buffer
);
1430 static inline unsigned long rb_page_entries(struct buffer_page
*bpage
)
1432 return local_read(&bpage
->entries
) & RB_WRITE_MASK
;
1435 static inline unsigned long rb_page_write(struct buffer_page
*bpage
)
1437 return local_read(&bpage
->write
) & RB_WRITE_MASK
;
1441 rb_remove_pages(struct ring_buffer_per_cpu
*cpu_buffer
, unsigned int nr_pages
)
1443 struct list_head
*tail_page
, *to_remove
, *next_page
;
1444 struct buffer_page
*to_remove_page
, *tmp_iter_page
;
1445 struct buffer_page
*last_page
, *first_page
;
1446 unsigned int nr_removed
;
1447 unsigned long head_bit
;
1452 raw_spin_lock_irq(&cpu_buffer
->reader_lock
);
1453 atomic_inc(&cpu_buffer
->record_disabled
);
1455 * We don't race with the readers since we have acquired the reader
1456 * lock. We also don't race with writers after disabling recording.
1457 * This makes it easy to figure out the first and the last page to be
1458 * removed from the list. We unlink all the pages in between including
1459 * the first and last pages. This is done in a busy loop so that we
1460 * lose the least number of traces.
1461 * The pages are freed after we restart recording and unlock readers.
1463 tail_page
= &cpu_buffer
->tail_page
->list
;
1466 * tail page might be on reader page, we remove the next page
1467 * from the ring buffer
1469 if (cpu_buffer
->tail_page
== cpu_buffer
->reader_page
)
1470 tail_page
= rb_list_head(tail_page
->next
);
1471 to_remove
= tail_page
;
1473 /* start of pages to remove */
1474 first_page
= list_entry(rb_list_head(to_remove
->next
),
1475 struct buffer_page
, list
);
1477 for (nr_removed
= 0; nr_removed
< nr_pages
; nr_removed
++) {
1478 to_remove
= rb_list_head(to_remove
)->next
;
1479 head_bit
|= (unsigned long)to_remove
& RB_PAGE_HEAD
;
1482 next_page
= rb_list_head(to_remove
)->next
;
1485 * Now we remove all pages between tail_page and next_page.
1486 * Make sure that we have head_bit value preserved for the
1489 tail_page
->next
= (struct list_head
*)((unsigned long)next_page
|
1491 next_page
= rb_list_head(next_page
);
1492 next_page
->prev
= tail_page
;
1494 /* make sure pages points to a valid page in the ring buffer */
1495 cpu_buffer
->pages
= next_page
;
1497 /* update head page */
1499 cpu_buffer
->head_page
= list_entry(next_page
,
1500 struct buffer_page
, list
);
1503 * change read pointer to make sure any read iterators reset
1506 cpu_buffer
->read
= 0;
1508 /* pages are removed, resume tracing and then free the pages */
1509 atomic_dec(&cpu_buffer
->record_disabled
);
1510 raw_spin_unlock_irq(&cpu_buffer
->reader_lock
);
1512 RB_WARN_ON(cpu_buffer
, list_empty(cpu_buffer
->pages
));
1514 /* last buffer page to remove */
1515 last_page
= list_entry(rb_list_head(to_remove
), struct buffer_page
,
1517 tmp_iter_page
= first_page
;
1520 to_remove_page
= tmp_iter_page
;
1521 rb_inc_page(cpu_buffer
, &tmp_iter_page
);
1523 /* update the counters */
1524 page_entries
= rb_page_entries(to_remove_page
);
1527 * If something was added to this page, it was full
1528 * since it is not the tail page. So we deduct the
1529 * bytes consumed in ring buffer from here.
1530 * Increment overrun to account for the lost events.
1532 local_add(page_entries
, &cpu_buffer
->overrun
);
1533 local_sub(BUF_PAGE_SIZE
, &cpu_buffer
->entries_bytes
);
1537 * We have already removed references to this list item, just
1538 * free up the buffer_page and its page
1540 free_buffer_page(to_remove_page
);
1543 } while (to_remove_page
!= last_page
);
1545 RB_WARN_ON(cpu_buffer
, nr_removed
);
1547 return nr_removed
== 0;
1551 rb_insert_pages(struct ring_buffer_per_cpu
*cpu_buffer
)
1553 struct list_head
*pages
= &cpu_buffer
->new_pages
;
1554 int retries
, success
;
1556 raw_spin_lock_irq(&cpu_buffer
->reader_lock
);
1558 * We are holding the reader lock, so the reader page won't be swapped
1559 * in the ring buffer. Now we are racing with the writer trying to
1560 * move head page and the tail page.
1561 * We are going to adapt the reader page update process where:
1562 * 1. We first splice the start and end of list of new pages between
1563 * the head page and its previous page.
1564 * 2. We cmpxchg the prev_page->next to point from head page to the
1565 * start of new pages list.
1566 * 3. Finally, we update the head->prev to the end of new list.
1568 * We will try this process 10 times, to make sure that we don't keep
1574 struct list_head
*head_page
, *prev_page
, *r
;
1575 struct list_head
*last_page
, *first_page
;
1576 struct list_head
*head_page_with_bit
;
1578 head_page
= &rb_set_head_page(cpu_buffer
)->list
;
1581 prev_page
= head_page
->prev
;
1583 first_page
= pages
->next
;
1584 last_page
= pages
->prev
;
1586 head_page_with_bit
= (struct list_head
*)
1587 ((unsigned long)head_page
| RB_PAGE_HEAD
);
1589 last_page
->next
= head_page_with_bit
;
1590 first_page
->prev
= prev_page
;
1592 r
= cmpxchg(&prev_page
->next
, head_page_with_bit
, first_page
);
1594 if (r
== head_page_with_bit
) {
1596 * yay, we replaced the page pointer to our new list,
1597 * now, we just have to update to head page's prev
1598 * pointer to point to end of list
1600 head_page
->prev
= last_page
;
1607 INIT_LIST_HEAD(pages
);
1609 * If we weren't successful in adding in new pages, warn and stop
1612 RB_WARN_ON(cpu_buffer
, !success
);
1613 raw_spin_unlock_irq(&cpu_buffer
->reader_lock
);
1615 /* free pages if they weren't inserted */
1617 struct buffer_page
*bpage
, *tmp
;
1618 list_for_each_entry_safe(bpage
, tmp
, &cpu_buffer
->new_pages
,
1620 list_del_init(&bpage
->list
);
1621 free_buffer_page(bpage
);
1627 static void rb_update_pages(struct ring_buffer_per_cpu
*cpu_buffer
)
1631 if (cpu_buffer
->nr_pages_to_update
> 0)
1632 success
= rb_insert_pages(cpu_buffer
);
1634 success
= rb_remove_pages(cpu_buffer
,
1635 -cpu_buffer
->nr_pages_to_update
);
1638 cpu_buffer
->nr_pages
+= cpu_buffer
->nr_pages_to_update
;
1641 static void update_pages_handler(struct work_struct
*work
)
1643 struct ring_buffer_per_cpu
*cpu_buffer
= container_of(work
,
1644 struct ring_buffer_per_cpu
, update_pages_work
);
1645 rb_update_pages(cpu_buffer
);
1646 complete(&cpu_buffer
->update_done
);
1650 * ring_buffer_resize - resize the ring buffer
1651 * @buffer: the buffer to resize.
1652 * @size: the new size.
1653 * @cpu_id: the cpu buffer to resize
1655 * Minimum size is 2 * BUF_PAGE_SIZE.
1657 * Returns 0 on success and < 0 on failure.
1659 int ring_buffer_resize(struct ring_buffer
*buffer
, unsigned long size
,
1662 struct ring_buffer_per_cpu
*cpu_buffer
;
1667 * Always succeed at resizing a non-existent buffer:
1672 /* Make sure the requested buffer exists */
1673 if (cpu_id
!= RING_BUFFER_ALL_CPUS
&&
1674 !cpumask_test_cpu(cpu_id
, buffer
->cpumask
))
1677 size
= DIV_ROUND_UP(size
, BUF_PAGE_SIZE
);
1678 size
*= BUF_PAGE_SIZE
;
1680 /* we need a minimum of two pages */
1681 if (size
< BUF_PAGE_SIZE
* 2)
1682 size
= BUF_PAGE_SIZE
* 2;
1684 nr_pages
= DIV_ROUND_UP(size
, BUF_PAGE_SIZE
);
1687 * Don't succeed if resizing is disabled, as a reader might be
1688 * manipulating the ring buffer and is expecting a sane state while
1691 if (atomic_read(&buffer
->resize_disabled
))
1694 /* prevent another thread from changing buffer sizes */
1695 mutex_lock(&buffer
->mutex
);
1697 if (cpu_id
== RING_BUFFER_ALL_CPUS
) {
1698 /* calculate the pages to update */
1699 for_each_buffer_cpu(buffer
, cpu
) {
1700 cpu_buffer
= buffer
->buffers
[cpu
];
1702 cpu_buffer
->nr_pages_to_update
= nr_pages
-
1703 cpu_buffer
->nr_pages
;
1705 * nothing more to do for removing pages or no update
1707 if (cpu_buffer
->nr_pages_to_update
<= 0)
1710 * to add pages, make sure all new pages can be
1711 * allocated without receiving ENOMEM
1713 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
1714 if (__rb_allocate_pages(cpu_buffer
->nr_pages_to_update
,
1715 &cpu_buffer
->new_pages
, cpu
)) {
1716 /* not enough memory for new pages */
1724 * Fire off all the required work handlers
1725 * We can't schedule on offline CPUs, but it's not necessary
1726 * since we can change their buffer sizes without any race.
1728 for_each_buffer_cpu(buffer
, cpu
) {
1729 cpu_buffer
= buffer
->buffers
[cpu
];
1730 if (!cpu_buffer
->nr_pages_to_update
)
1733 /* Can't run something on an offline CPU. */
1734 if (!cpu_online(cpu
)) {
1735 rb_update_pages(cpu_buffer
);
1736 cpu_buffer
->nr_pages_to_update
= 0;
1738 schedule_work_on(cpu
,
1739 &cpu_buffer
->update_pages_work
);
1743 /* wait for all the updates to complete */
1744 for_each_buffer_cpu(buffer
, cpu
) {
1745 cpu_buffer
= buffer
->buffers
[cpu
];
1746 if (!cpu_buffer
->nr_pages_to_update
)
1749 if (cpu_online(cpu
))
1750 wait_for_completion(&cpu_buffer
->update_done
);
1751 cpu_buffer
->nr_pages_to_update
= 0;
1756 /* Make sure this CPU has been intitialized */
1757 if (!cpumask_test_cpu(cpu_id
, buffer
->cpumask
))
1760 cpu_buffer
= buffer
->buffers
[cpu_id
];
1762 if (nr_pages
== cpu_buffer
->nr_pages
)
1765 cpu_buffer
->nr_pages_to_update
= nr_pages
-
1766 cpu_buffer
->nr_pages
;
1768 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
1769 if (cpu_buffer
->nr_pages_to_update
> 0 &&
1770 __rb_allocate_pages(cpu_buffer
->nr_pages_to_update
,
1771 &cpu_buffer
->new_pages
, cpu_id
)) {
1778 /* Can't run something on an offline CPU. */
1779 if (!cpu_online(cpu_id
))
1780 rb_update_pages(cpu_buffer
);
1782 schedule_work_on(cpu_id
,
1783 &cpu_buffer
->update_pages_work
);
1784 wait_for_completion(&cpu_buffer
->update_done
);
1787 cpu_buffer
->nr_pages_to_update
= 0;
1793 * The ring buffer resize can happen with the ring buffer
1794 * enabled, so that the update disturbs the tracing as little
1795 * as possible. But if the buffer is disabled, we do not need
1796 * to worry about that, and we can take the time to verify
1797 * that the buffer is not corrupt.
1799 if (atomic_read(&buffer
->record_disabled
)) {
1800 atomic_inc(&buffer
->record_disabled
);
1802 * Even though the buffer was disabled, we must make sure
1803 * that it is truly disabled before calling rb_check_pages.
1804 * There could have been a race between checking
1805 * record_disable and incrementing it.
1807 synchronize_sched();
1808 for_each_buffer_cpu(buffer
, cpu
) {
1809 cpu_buffer
= buffer
->buffers
[cpu
];
1810 rb_check_pages(cpu_buffer
);
1812 atomic_dec(&buffer
->record_disabled
);
1815 mutex_unlock(&buffer
->mutex
);
1819 for_each_buffer_cpu(buffer
, cpu
) {
1820 struct buffer_page
*bpage
, *tmp
;
1822 cpu_buffer
= buffer
->buffers
[cpu
];
1823 cpu_buffer
->nr_pages_to_update
= 0;
1825 if (list_empty(&cpu_buffer
->new_pages
))
1828 list_for_each_entry_safe(bpage
, tmp
, &cpu_buffer
->new_pages
,
1830 list_del_init(&bpage
->list
);
1831 free_buffer_page(bpage
);
1834 mutex_unlock(&buffer
->mutex
);
1837 EXPORT_SYMBOL_GPL(ring_buffer_resize
);
1839 void ring_buffer_change_overwrite(struct ring_buffer
*buffer
, int val
)
1841 mutex_lock(&buffer
->mutex
);
1843 buffer
->flags
|= RB_FL_OVERWRITE
;
1845 buffer
->flags
&= ~RB_FL_OVERWRITE
;
1846 mutex_unlock(&buffer
->mutex
);
1848 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite
);
1850 static inline void *
1851 __rb_data_page_index(struct buffer_data_page
*bpage
, unsigned index
)
1853 return bpage
->data
+ index
;
1856 static inline void *__rb_page_index(struct buffer_page
*bpage
, unsigned index
)
1858 return bpage
->page
->data
+ index
;
1861 static inline struct ring_buffer_event
*
1862 rb_reader_event(struct ring_buffer_per_cpu
*cpu_buffer
)
1864 return __rb_page_index(cpu_buffer
->reader_page
,
1865 cpu_buffer
->reader_page
->read
);
1868 static inline struct ring_buffer_event
*
1869 rb_iter_head_event(struct ring_buffer_iter
*iter
)
1871 return __rb_page_index(iter
->head_page
, iter
->head
);
1874 static inline unsigned rb_page_commit(struct buffer_page
*bpage
)
1876 return local_read(&bpage
->page
->commit
);
1879 /* Size is determined by what has been committed */
1880 static inline unsigned rb_page_size(struct buffer_page
*bpage
)
1882 return rb_page_commit(bpage
);
1885 static inline unsigned
1886 rb_commit_index(struct ring_buffer_per_cpu
*cpu_buffer
)
1888 return rb_page_commit(cpu_buffer
->commit_page
);
1891 static inline unsigned
1892 rb_event_index(struct ring_buffer_event
*event
)
1894 unsigned long addr
= (unsigned long)event
;
1896 return (addr
& ~PAGE_MASK
) - BUF_PAGE_HDR_SIZE
;
1900 rb_event_is_commit(struct ring_buffer_per_cpu
*cpu_buffer
,
1901 struct ring_buffer_event
*event
)
1903 unsigned long addr
= (unsigned long)event
;
1904 unsigned long index
;
1906 index
= rb_event_index(event
);
1909 return cpu_buffer
->commit_page
->page
== (void *)addr
&&
1910 rb_commit_index(cpu_buffer
) == index
;
1914 rb_set_commit_to_write(struct ring_buffer_per_cpu
*cpu_buffer
)
1916 unsigned long max_count
;
1919 * We only race with interrupts and NMIs on this CPU.
1920 * If we own the commit event, then we can commit
1921 * all others that interrupted us, since the interruptions
1922 * are in stack format (they finish before they come
1923 * back to us). This allows us to do a simple loop to
1924 * assign the commit to the tail.
1927 max_count
= cpu_buffer
->nr_pages
* 100;
1929 while (cpu_buffer
->commit_page
!= cpu_buffer
->tail_page
) {
1930 if (RB_WARN_ON(cpu_buffer
, !(--max_count
)))
1932 if (RB_WARN_ON(cpu_buffer
,
1933 rb_is_reader_page(cpu_buffer
->tail_page
)))
1935 local_set(&cpu_buffer
->commit_page
->page
->commit
,
1936 rb_page_write(cpu_buffer
->commit_page
));
1937 rb_inc_page(cpu_buffer
, &cpu_buffer
->commit_page
);
1938 cpu_buffer
->write_stamp
=
1939 cpu_buffer
->commit_page
->page
->time_stamp
;
1940 /* add barrier to keep gcc from optimizing too much */
1943 while (rb_commit_index(cpu_buffer
) !=
1944 rb_page_write(cpu_buffer
->commit_page
)) {
1946 local_set(&cpu_buffer
->commit_page
->page
->commit
,
1947 rb_page_write(cpu_buffer
->commit_page
));
1948 RB_WARN_ON(cpu_buffer
,
1949 local_read(&cpu_buffer
->commit_page
->page
->commit
) &
1954 /* again, keep gcc from optimizing */
1958 * If an interrupt came in just after the first while loop
1959 * and pushed the tail page forward, we will be left with
1960 * a dangling commit that will never go forward.
1962 if (unlikely(cpu_buffer
->commit_page
!= cpu_buffer
->tail_page
))
1966 static void rb_reset_reader_page(struct ring_buffer_per_cpu
*cpu_buffer
)
1968 cpu_buffer
->read_stamp
= cpu_buffer
->reader_page
->page
->time_stamp
;
1969 cpu_buffer
->reader_page
->read
= 0;
1972 static void rb_inc_iter(struct ring_buffer_iter
*iter
)
1974 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
1977 * The iterator could be on the reader page (it starts there).
1978 * But the head could have moved, since the reader was
1979 * found. Check for this case and assign the iterator
1980 * to the head page instead of next.
1982 if (iter
->head_page
== cpu_buffer
->reader_page
)
1983 iter
->head_page
= rb_set_head_page(cpu_buffer
);
1985 rb_inc_page(cpu_buffer
, &iter
->head_page
);
1987 iter
->read_stamp
= iter
->head_page
->page
->time_stamp
;
1991 /* Slow path, do not inline */
1992 static noinline
struct ring_buffer_event
*
1993 rb_add_time_stamp(struct ring_buffer_event
*event
, u64 delta
)
1995 event
->type_len
= RINGBUF_TYPE_TIME_EXTEND
;
1997 /* Not the first event on the page? */
1998 if (rb_event_index(event
)) {
1999 event
->time_delta
= delta
& TS_MASK
;
2000 event
->array
[0] = delta
>> TS_SHIFT
;
2002 /* nope, just zero it */
2003 event
->time_delta
= 0;
2004 event
->array
[0] = 0;
2007 return skip_time_extend(event
);
2011 * rb_update_event - update event type and data
2012 * @event: the event to update
2013 * @type: the type of event
2014 * @length: the size of the event field in the ring buffer
2016 * Update the type and data fields of the event. The length
2017 * is the actual size that is written to the ring buffer,
2018 * and with this, we can determine what to place into the
2022 rb_update_event(struct ring_buffer_per_cpu
*cpu_buffer
,
2023 struct ring_buffer_event
*event
, unsigned length
,
2024 int add_timestamp
, u64 delta
)
2026 /* Only a commit updates the timestamp */
2027 if (unlikely(!rb_event_is_commit(cpu_buffer
, event
)))
2031 * If we need to add a timestamp, then we
2032 * add it to the start of the resevered space.
2034 if (unlikely(add_timestamp
)) {
2035 event
= rb_add_time_stamp(event
, delta
);
2036 length
-= RB_LEN_TIME_EXTEND
;
2040 event
->time_delta
= delta
;
2041 length
-= RB_EVNT_HDR_SIZE
;
2042 if (length
> RB_MAX_SMALL_DATA
|| RB_FORCE_8BYTE_ALIGNMENT
) {
2043 event
->type_len
= 0;
2044 event
->array
[0] = length
;
2046 event
->type_len
= DIV_ROUND_UP(length
, RB_ALIGNMENT
);
2050 * rb_handle_head_page - writer hit the head page
2052 * Returns: +1 to retry page
2057 rb_handle_head_page(struct ring_buffer_per_cpu
*cpu_buffer
,
2058 struct buffer_page
*tail_page
,
2059 struct buffer_page
*next_page
)
2061 struct buffer_page
*new_head
;
2066 entries
= rb_page_entries(next_page
);
2069 * The hard part is here. We need to move the head
2070 * forward, and protect against both readers on
2071 * other CPUs and writers coming in via interrupts.
2073 type
= rb_head_page_set_update(cpu_buffer
, next_page
, tail_page
,
2077 * type can be one of four:
2078 * NORMAL - an interrupt already moved it for us
2079 * HEAD - we are the first to get here.
2080 * UPDATE - we are the interrupt interrupting
2082 * MOVED - a reader on another CPU moved the next
2083 * pointer to its reader page. Give up
2090 * We changed the head to UPDATE, thus
2091 * it is our responsibility to update
2094 local_add(entries
, &cpu_buffer
->overrun
);
2095 local_sub(BUF_PAGE_SIZE
, &cpu_buffer
->entries_bytes
);
2098 * The entries will be zeroed out when we move the
2102 /* still more to do */
2105 case RB_PAGE_UPDATE
:
2107 * This is an interrupt that interrupt the
2108 * previous update. Still more to do.
2111 case RB_PAGE_NORMAL
:
2113 * An interrupt came in before the update
2114 * and processed this for us.
2115 * Nothing left to do.
2120 * The reader is on another CPU and just did
2121 * a swap with our next_page.
2126 RB_WARN_ON(cpu_buffer
, 1); /* WTF??? */
2131 * Now that we are here, the old head pointer is
2132 * set to UPDATE. This will keep the reader from
2133 * swapping the head page with the reader page.
2134 * The reader (on another CPU) will spin till
2137 * We just need to protect against interrupts
2138 * doing the job. We will set the next pointer
2139 * to HEAD. After that, we set the old pointer
2140 * to NORMAL, but only if it was HEAD before.
2141 * otherwise we are an interrupt, and only
2142 * want the outer most commit to reset it.
2144 new_head
= next_page
;
2145 rb_inc_page(cpu_buffer
, &new_head
);
2147 ret
= rb_head_page_set_head(cpu_buffer
, new_head
, next_page
,
2151 * Valid returns are:
2152 * HEAD - an interrupt came in and already set it.
2153 * NORMAL - One of two things:
2154 * 1) We really set it.
2155 * 2) A bunch of interrupts came in and moved
2156 * the page forward again.
2160 case RB_PAGE_NORMAL
:
2164 RB_WARN_ON(cpu_buffer
, 1);
2169 * It is possible that an interrupt came in,
2170 * set the head up, then more interrupts came in
2171 * and moved it again. When we get back here,
2172 * the page would have been set to NORMAL but we
2173 * just set it back to HEAD.
2175 * How do you detect this? Well, if that happened
2176 * the tail page would have moved.
2178 if (ret
== RB_PAGE_NORMAL
) {
2180 * If the tail had moved passed next, then we need
2181 * to reset the pointer.
2183 if (cpu_buffer
->tail_page
!= tail_page
&&
2184 cpu_buffer
->tail_page
!= next_page
)
2185 rb_head_page_set_normal(cpu_buffer
, new_head
,
2191 * If this was the outer most commit (the one that
2192 * changed the original pointer from HEAD to UPDATE),
2193 * then it is up to us to reset it to NORMAL.
2195 if (type
== RB_PAGE_HEAD
) {
2196 ret
= rb_head_page_set_normal(cpu_buffer
, next_page
,
2199 if (RB_WARN_ON(cpu_buffer
,
2200 ret
!= RB_PAGE_UPDATE
))
2207 static unsigned rb_calculate_event_length(unsigned length
)
2209 struct ring_buffer_event event
; /* Used only for sizeof array */
2211 /* zero length can cause confusions */
2215 if (length
> RB_MAX_SMALL_DATA
|| RB_FORCE_8BYTE_ALIGNMENT
)
2216 length
+= sizeof(event
.array
[0]);
2218 length
+= RB_EVNT_HDR_SIZE
;
2219 length
= ALIGN(length
, RB_ARCH_ALIGNMENT
);
2225 rb_reset_tail(struct ring_buffer_per_cpu
*cpu_buffer
,
2226 struct buffer_page
*tail_page
,
2227 unsigned long tail
, unsigned long length
)
2229 struct ring_buffer_event
*event
;
2232 * Only the event that crossed the page boundary
2233 * must fill the old tail_page with padding.
2235 if (tail
>= BUF_PAGE_SIZE
) {
2237 * If the page was filled, then we still need
2238 * to update the real_end. Reset it to zero
2239 * and the reader will ignore it.
2241 if (tail
== BUF_PAGE_SIZE
)
2242 tail_page
->real_end
= 0;
2244 local_sub(length
, &tail_page
->write
);
2248 event
= __rb_page_index(tail_page
, tail
);
2249 kmemcheck_annotate_bitfield(event
, bitfield
);
2251 /* account for padding bytes */
2252 local_add(BUF_PAGE_SIZE
- tail
, &cpu_buffer
->entries_bytes
);
2255 * Save the original length to the meta data.
2256 * This will be used by the reader to add lost event
2259 tail_page
->real_end
= tail
;
2262 * If this event is bigger than the minimum size, then
2263 * we need to be careful that we don't subtract the
2264 * write counter enough to allow another writer to slip
2266 * We put in a discarded commit instead, to make sure
2267 * that this space is not used again.
2269 * If we are less than the minimum size, we don't need to
2272 if (tail
> (BUF_PAGE_SIZE
- RB_EVNT_MIN_SIZE
)) {
2273 /* No room for any events */
2275 /* Mark the rest of the page with padding */
2276 rb_event_set_padding(event
);
2278 /* Set the write back to the previous setting */
2279 local_sub(length
, &tail_page
->write
);
2283 /* Put in a discarded event */
2284 event
->array
[0] = (BUF_PAGE_SIZE
- tail
) - RB_EVNT_HDR_SIZE
;
2285 event
->type_len
= RINGBUF_TYPE_PADDING
;
2286 /* time delta must be non zero */
2287 event
->time_delta
= 1;
2289 /* Set write to end of buffer */
2290 length
= (tail
+ length
) - BUF_PAGE_SIZE
;
2291 local_sub(length
, &tail_page
->write
);
2295 * This is the slow path, force gcc not to inline it.
2297 static noinline
struct ring_buffer_event
*
2298 rb_move_tail(struct ring_buffer_per_cpu
*cpu_buffer
,
2299 unsigned long length
, unsigned long tail
,
2300 struct buffer_page
*tail_page
, u64 ts
)
2302 struct buffer_page
*commit_page
= cpu_buffer
->commit_page
;
2303 struct ring_buffer
*buffer
= cpu_buffer
->buffer
;
2304 struct buffer_page
*next_page
;
2307 next_page
= tail_page
;
2309 rb_inc_page(cpu_buffer
, &next_page
);
2312 * If for some reason, we had an interrupt storm that made
2313 * it all the way around the buffer, bail, and warn
2316 if (unlikely(next_page
== commit_page
)) {
2317 local_inc(&cpu_buffer
->commit_overrun
);
2322 * This is where the fun begins!
2324 * We are fighting against races between a reader that
2325 * could be on another CPU trying to swap its reader
2326 * page with the buffer head.
2328 * We are also fighting against interrupts coming in and
2329 * moving the head or tail on us as well.
2331 * If the next page is the head page then we have filled
2332 * the buffer, unless the commit page is still on the
2335 if (rb_is_head_page(cpu_buffer
, next_page
, &tail_page
->list
)) {
2338 * If the commit is not on the reader page, then
2339 * move the header page.
2341 if (!rb_is_reader_page(cpu_buffer
->commit_page
)) {
2343 * If we are not in overwrite mode,
2344 * this is easy, just stop here.
2346 if (!(buffer
->flags
& RB_FL_OVERWRITE
)) {
2347 local_inc(&cpu_buffer
->dropped_events
);
2351 ret
= rb_handle_head_page(cpu_buffer
,
2360 * We need to be careful here too. The
2361 * commit page could still be on the reader
2362 * page. We could have a small buffer, and
2363 * have filled up the buffer with events
2364 * from interrupts and such, and wrapped.
2366 * Note, if the tail page is also the on the
2367 * reader_page, we let it move out.
2369 if (unlikely((cpu_buffer
->commit_page
!=
2370 cpu_buffer
->tail_page
) &&
2371 (cpu_buffer
->commit_page
==
2372 cpu_buffer
->reader_page
))) {
2373 local_inc(&cpu_buffer
->commit_overrun
);
2379 ret
= rb_tail_page_update(cpu_buffer
, tail_page
, next_page
);
2382 * Nested commits always have zero deltas, so
2383 * just reread the time stamp
2385 ts
= rb_time_stamp(buffer
);
2386 next_page
->page
->time_stamp
= ts
;
2391 rb_reset_tail(cpu_buffer
, tail_page
, tail
, length
);
2393 /* fail and let the caller try again */
2394 return ERR_PTR(-EAGAIN
);
2398 rb_reset_tail(cpu_buffer
, tail_page
, tail
, length
);
2403 static struct ring_buffer_event
*
2404 __rb_reserve_next(struct ring_buffer_per_cpu
*cpu_buffer
,
2405 unsigned long length
, u64 ts
,
2406 u64 delta
, int add_timestamp
)
2408 struct buffer_page
*tail_page
;
2409 struct ring_buffer_event
*event
;
2410 unsigned long tail
, write
;
2413 * If the time delta since the last event is too big to
2414 * hold in the time field of the event, then we append a
2415 * TIME EXTEND event ahead of the data event.
2417 if (unlikely(add_timestamp
))
2418 length
+= RB_LEN_TIME_EXTEND
;
2420 tail_page
= cpu_buffer
->tail_page
;
2421 write
= local_add_return(length
, &tail_page
->write
);
2423 /* set write to only the index of the write */
2424 write
&= RB_WRITE_MASK
;
2425 tail
= write
- length
;
2428 * If this is the first commit on the page, then it has the same
2429 * timestamp as the page itself.
2434 /* See if we shot pass the end of this buffer page */
2435 if (unlikely(write
> BUF_PAGE_SIZE
))
2436 return rb_move_tail(cpu_buffer
, length
, tail
,
2439 /* We reserved something on the buffer */
2441 event
= __rb_page_index(tail_page
, tail
);
2442 kmemcheck_annotate_bitfield(event
, bitfield
);
2443 rb_update_event(cpu_buffer
, event
, length
, add_timestamp
, delta
);
2445 local_inc(&tail_page
->entries
);
2448 * If this is the first commit on the page, then update
2452 tail_page
->page
->time_stamp
= ts
;
2454 /* account for these added bytes */
2455 local_add(length
, &cpu_buffer
->entries_bytes
);
2461 rb_try_to_discard(struct ring_buffer_per_cpu
*cpu_buffer
,
2462 struct ring_buffer_event
*event
)
2464 unsigned long new_index
, old_index
;
2465 struct buffer_page
*bpage
;
2466 unsigned long index
;
2469 new_index
= rb_event_index(event
);
2470 old_index
= new_index
+ rb_event_ts_length(event
);
2471 addr
= (unsigned long)event
;
2474 bpage
= cpu_buffer
->tail_page
;
2476 if (bpage
->page
== (void *)addr
&& rb_page_write(bpage
) == old_index
) {
2477 unsigned long write_mask
=
2478 local_read(&bpage
->write
) & ~RB_WRITE_MASK
;
2479 unsigned long event_length
= rb_event_length(event
);
2481 * This is on the tail page. It is possible that
2482 * a write could come in and move the tail page
2483 * and write to the next page. That is fine
2484 * because we just shorten what is on this page.
2486 old_index
+= write_mask
;
2487 new_index
+= write_mask
;
2488 index
= local_cmpxchg(&bpage
->write
, old_index
, new_index
);
2489 if (index
== old_index
) {
2490 /* update counters */
2491 local_sub(event_length
, &cpu_buffer
->entries_bytes
);
2496 /* could not discard */
2500 static void rb_start_commit(struct ring_buffer_per_cpu
*cpu_buffer
)
2502 local_inc(&cpu_buffer
->committing
);
2503 local_inc(&cpu_buffer
->commits
);
2506 static inline void rb_end_commit(struct ring_buffer_per_cpu
*cpu_buffer
)
2508 unsigned long commits
;
2510 if (RB_WARN_ON(cpu_buffer
,
2511 !local_read(&cpu_buffer
->committing
)))
2515 commits
= local_read(&cpu_buffer
->commits
);
2516 /* synchronize with interrupts */
2518 if (local_read(&cpu_buffer
->committing
) == 1)
2519 rb_set_commit_to_write(cpu_buffer
);
2521 local_dec(&cpu_buffer
->committing
);
2523 /* synchronize with interrupts */
2527 * Need to account for interrupts coming in between the
2528 * updating of the commit page and the clearing of the
2529 * committing counter.
2531 if (unlikely(local_read(&cpu_buffer
->commits
) != commits
) &&
2532 !local_read(&cpu_buffer
->committing
)) {
2533 local_inc(&cpu_buffer
->committing
);
2538 static struct ring_buffer_event
*
2539 rb_reserve_next_event(struct ring_buffer
*buffer
,
2540 struct ring_buffer_per_cpu
*cpu_buffer
,
2541 unsigned long length
)
2543 struct ring_buffer_event
*event
;
2549 rb_start_commit(cpu_buffer
);
2551 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2553 * Due to the ability to swap a cpu buffer from a buffer
2554 * it is possible it was swapped before we committed.
2555 * (committing stops a swap). We check for it here and
2556 * if it happened, we have to fail the write.
2559 if (unlikely(ACCESS_ONCE(cpu_buffer
->buffer
) != buffer
)) {
2560 local_dec(&cpu_buffer
->committing
);
2561 local_dec(&cpu_buffer
->commits
);
2566 length
= rb_calculate_event_length(length
);
2572 * We allow for interrupts to reenter here and do a trace.
2573 * If one does, it will cause this original code to loop
2574 * back here. Even with heavy interrupts happening, this
2575 * should only happen a few times in a row. If this happens
2576 * 1000 times in a row, there must be either an interrupt
2577 * storm or we have something buggy.
2580 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 1000))
2583 ts
= rb_time_stamp(cpu_buffer
->buffer
);
2584 diff
= ts
- cpu_buffer
->write_stamp
;
2586 /* make sure this diff is calculated here */
2589 /* Did the write stamp get updated already? */
2590 if (likely(ts
>= cpu_buffer
->write_stamp
)) {
2592 if (unlikely(test_time_stamp(delta
))) {
2593 int local_clock_stable
= 1;
2594 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2595 local_clock_stable
= sched_clock_stable();
2597 WARN_ONCE(delta
> (1ULL << 59),
2598 KERN_WARNING
"Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2599 (unsigned long long)delta
,
2600 (unsigned long long)ts
,
2601 (unsigned long long)cpu_buffer
->write_stamp
,
2602 local_clock_stable
? "" :
2603 "If you just came from a suspend/resume,\n"
2604 "please switch to the trace global clock:\n"
2605 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2610 event
= __rb_reserve_next(cpu_buffer
, length
, ts
,
2611 delta
, add_timestamp
);
2612 if (unlikely(PTR_ERR(event
) == -EAGAIN
))
2621 rb_end_commit(cpu_buffer
);
2625 #ifdef CONFIG_TRACING
2628 * The lock and unlock are done within a preempt disable section.
2629 * The current_context per_cpu variable can only be modified
2630 * by the current task between lock and unlock. But it can
2631 * be modified more than once via an interrupt. To pass this
2632 * information from the lock to the unlock without having to
2633 * access the 'in_interrupt()' functions again (which do show
2634 * a bit of overhead in something as critical as function tracing,
2635 * we use a bitmask trick.
2637 * bit 0 = NMI context
2638 * bit 1 = IRQ context
2639 * bit 2 = SoftIRQ context
2640 * bit 3 = normal context.
2642 * This works because this is the order of contexts that can
2643 * preempt other contexts. A SoftIRQ never preempts an IRQ
2646 * When the context is determined, the corresponding bit is
2647 * checked and set (if it was set, then a recursion of that context
2650 * On unlock, we need to clear this bit. To do so, just subtract
2651 * 1 from the current_context and AND it to itself.
2655 * 101 & 100 = 100 (clearing bit zero)
2658 * 1010 & 1001 = 1000 (clearing bit 1)
2660 * The least significant bit can be cleared this way, and it
2661 * just so happens that it is the same bit corresponding to
2662 * the current context.
2664 static DEFINE_PER_CPU(unsigned int, current_context
);
2666 static __always_inline
int trace_recursive_lock(void)
2668 unsigned int val
= this_cpu_read(current_context
);
2671 if (in_interrupt()) {
2681 if (unlikely(val
& (1 << bit
)))
2685 this_cpu_write(current_context
, val
);
2690 static __always_inline
void trace_recursive_unlock(void)
2692 unsigned int val
= this_cpu_read(current_context
);
2695 val
&= this_cpu_read(current_context
);
2696 this_cpu_write(current_context
, val
);
2701 #define trace_recursive_lock() (0)
2702 #define trace_recursive_unlock() do { } while (0)
2707 * ring_buffer_lock_reserve - reserve a part of the buffer
2708 * @buffer: the ring buffer to reserve from
2709 * @length: the length of the data to reserve (excluding event header)
2711 * Returns a reseverd event on the ring buffer to copy directly to.
2712 * The user of this interface will need to get the body to write into
2713 * and can use the ring_buffer_event_data() interface.
2715 * The length is the length of the data needed, not the event length
2716 * which also includes the event header.
2718 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2719 * If NULL is returned, then nothing has been allocated or locked.
2721 struct ring_buffer_event
*
2722 ring_buffer_lock_reserve(struct ring_buffer
*buffer
, unsigned long length
)
2724 struct ring_buffer_per_cpu
*cpu_buffer
;
2725 struct ring_buffer_event
*event
;
2728 if (ring_buffer_flags
!= RB_BUFFERS_ON
)
2731 /* If we are tracing schedule, we don't want to recurse */
2732 preempt_disable_notrace();
2734 if (atomic_read(&buffer
->record_disabled
))
2737 if (trace_recursive_lock())
2740 cpu
= raw_smp_processor_id();
2742 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
2745 cpu_buffer
= buffer
->buffers
[cpu
];
2747 if (atomic_read(&cpu_buffer
->record_disabled
))
2750 if (length
> BUF_MAX_DATA_SIZE
)
2753 event
= rb_reserve_next_event(buffer
, cpu_buffer
, length
);
2760 trace_recursive_unlock();
2763 preempt_enable_notrace();
2766 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve
);
2769 rb_update_write_stamp(struct ring_buffer_per_cpu
*cpu_buffer
,
2770 struct ring_buffer_event
*event
)
2775 * The event first in the commit queue updates the
2778 if (rb_event_is_commit(cpu_buffer
, event
)) {
2780 * A commit event that is first on a page
2781 * updates the write timestamp with the page stamp
2783 if (!rb_event_index(event
))
2784 cpu_buffer
->write_stamp
=
2785 cpu_buffer
->commit_page
->page
->time_stamp
;
2786 else if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
) {
2787 delta
= event
->array
[0];
2789 delta
+= event
->time_delta
;
2790 cpu_buffer
->write_stamp
+= delta
;
2792 cpu_buffer
->write_stamp
+= event
->time_delta
;
2796 static void rb_commit(struct ring_buffer_per_cpu
*cpu_buffer
,
2797 struct ring_buffer_event
*event
)
2799 local_inc(&cpu_buffer
->entries
);
2800 rb_update_write_stamp(cpu_buffer
, event
);
2801 rb_end_commit(cpu_buffer
);
2804 static __always_inline
void
2805 rb_wakeups(struct ring_buffer
*buffer
, struct ring_buffer_per_cpu
*cpu_buffer
)
2807 if (buffer
->irq_work
.waiters_pending
) {
2808 buffer
->irq_work
.waiters_pending
= false;
2809 /* irq_work_queue() supplies it's own memory barriers */
2810 irq_work_queue(&buffer
->irq_work
.work
);
2813 if (cpu_buffer
->irq_work
.waiters_pending
) {
2814 cpu_buffer
->irq_work
.waiters_pending
= false;
2815 /* irq_work_queue() supplies it's own memory barriers */
2816 irq_work_queue(&cpu_buffer
->irq_work
.work
);
2821 * ring_buffer_unlock_commit - commit a reserved
2822 * @buffer: The buffer to commit to
2823 * @event: The event pointer to commit.
2825 * This commits the data to the ring buffer, and releases any locks held.
2827 * Must be paired with ring_buffer_lock_reserve.
2829 int ring_buffer_unlock_commit(struct ring_buffer
*buffer
,
2830 struct ring_buffer_event
*event
)
2832 struct ring_buffer_per_cpu
*cpu_buffer
;
2833 int cpu
= raw_smp_processor_id();
2835 cpu_buffer
= buffer
->buffers
[cpu
];
2837 rb_commit(cpu_buffer
, event
);
2839 rb_wakeups(buffer
, cpu_buffer
);
2841 trace_recursive_unlock();
2843 preempt_enable_notrace();
2847 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit
);
2849 static inline void rb_event_discard(struct ring_buffer_event
*event
)
2851 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
2852 event
= skip_time_extend(event
);
2854 /* array[0] holds the actual length for the discarded event */
2855 event
->array
[0] = rb_event_data_length(event
) - RB_EVNT_HDR_SIZE
;
2856 event
->type_len
= RINGBUF_TYPE_PADDING
;
2857 /* time delta must be non zero */
2858 if (!event
->time_delta
)
2859 event
->time_delta
= 1;
2863 * Decrement the entries to the page that an event is on.
2864 * The event does not even need to exist, only the pointer
2865 * to the page it is on. This may only be called before the commit
2869 rb_decrement_entry(struct ring_buffer_per_cpu
*cpu_buffer
,
2870 struct ring_buffer_event
*event
)
2872 unsigned long addr
= (unsigned long)event
;
2873 struct buffer_page
*bpage
= cpu_buffer
->commit_page
;
2874 struct buffer_page
*start
;
2878 /* Do the likely case first */
2879 if (likely(bpage
->page
== (void *)addr
)) {
2880 local_dec(&bpage
->entries
);
2885 * Because the commit page may be on the reader page we
2886 * start with the next page and check the end loop there.
2888 rb_inc_page(cpu_buffer
, &bpage
);
2891 if (bpage
->page
== (void *)addr
) {
2892 local_dec(&bpage
->entries
);
2895 rb_inc_page(cpu_buffer
, &bpage
);
2896 } while (bpage
!= start
);
2898 /* commit not part of this buffer?? */
2899 RB_WARN_ON(cpu_buffer
, 1);
2903 * ring_buffer_commit_discard - discard an event that has not been committed
2904 * @buffer: the ring buffer
2905 * @event: non committed event to discard
2907 * Sometimes an event that is in the ring buffer needs to be ignored.
2908 * This function lets the user discard an event in the ring buffer
2909 * and then that event will not be read later.
2911 * This function only works if it is called before the the item has been
2912 * committed. It will try to free the event from the ring buffer
2913 * if another event has not been added behind it.
2915 * If another event has been added behind it, it will set the event
2916 * up as discarded, and perform the commit.
2918 * If this function is called, do not call ring_buffer_unlock_commit on
2921 void ring_buffer_discard_commit(struct ring_buffer
*buffer
,
2922 struct ring_buffer_event
*event
)
2924 struct ring_buffer_per_cpu
*cpu_buffer
;
2927 /* The event is discarded regardless */
2928 rb_event_discard(event
);
2930 cpu
= smp_processor_id();
2931 cpu_buffer
= buffer
->buffers
[cpu
];
2934 * This must only be called if the event has not been
2935 * committed yet. Thus we can assume that preemption
2936 * is still disabled.
2938 RB_WARN_ON(buffer
, !local_read(&cpu_buffer
->committing
));
2940 rb_decrement_entry(cpu_buffer
, event
);
2941 if (rb_try_to_discard(cpu_buffer
, event
))
2945 * The commit is still visible by the reader, so we
2946 * must still update the timestamp.
2948 rb_update_write_stamp(cpu_buffer
, event
);
2950 rb_end_commit(cpu_buffer
);
2952 trace_recursive_unlock();
2954 preempt_enable_notrace();
2957 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit
);
2960 * ring_buffer_write - write data to the buffer without reserving
2961 * @buffer: The ring buffer to write to.
2962 * @length: The length of the data being written (excluding the event header)
2963 * @data: The data to write to the buffer.
2965 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2966 * one function. If you already have the data to write to the buffer, it
2967 * may be easier to simply call this function.
2969 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2970 * and not the length of the event which would hold the header.
2972 int ring_buffer_write(struct ring_buffer
*buffer
,
2973 unsigned long length
,
2976 struct ring_buffer_per_cpu
*cpu_buffer
;
2977 struct ring_buffer_event
*event
;
2982 if (ring_buffer_flags
!= RB_BUFFERS_ON
)
2985 preempt_disable_notrace();
2987 if (atomic_read(&buffer
->record_disabled
))
2990 cpu
= raw_smp_processor_id();
2992 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
2995 cpu_buffer
= buffer
->buffers
[cpu
];
2997 if (atomic_read(&cpu_buffer
->record_disabled
))
3000 if (length
> BUF_MAX_DATA_SIZE
)
3003 event
= rb_reserve_next_event(buffer
, cpu_buffer
, length
);
3007 body
= rb_event_data(event
);
3009 memcpy(body
, data
, length
);
3011 rb_commit(cpu_buffer
, event
);
3013 rb_wakeups(buffer
, cpu_buffer
);
3017 preempt_enable_notrace();
3021 EXPORT_SYMBOL_GPL(ring_buffer_write
);
3023 static int rb_per_cpu_empty(struct ring_buffer_per_cpu
*cpu_buffer
)
3025 struct buffer_page
*reader
= cpu_buffer
->reader_page
;
3026 struct buffer_page
*head
= rb_set_head_page(cpu_buffer
);
3027 struct buffer_page
*commit
= cpu_buffer
->commit_page
;
3029 /* In case of error, head will be NULL */
3030 if (unlikely(!head
))
3033 return reader
->read
== rb_page_commit(reader
) &&
3034 (commit
== reader
||
3036 head
->read
== rb_page_commit(commit
)));
3040 * ring_buffer_record_disable - stop all writes into the buffer
3041 * @buffer: The ring buffer to stop writes to.
3043 * This prevents all writes to the buffer. Any attempt to write
3044 * to the buffer after this will fail and return NULL.
3046 * The caller should call synchronize_sched() after this.
3048 void ring_buffer_record_disable(struct ring_buffer
*buffer
)
3050 atomic_inc(&buffer
->record_disabled
);
3052 EXPORT_SYMBOL_GPL(ring_buffer_record_disable
);
3055 * ring_buffer_record_enable - enable writes to the buffer
3056 * @buffer: The ring buffer to enable writes
3058 * Note, multiple disables will need the same number of enables
3059 * to truly enable the writing (much like preempt_disable).
3061 void ring_buffer_record_enable(struct ring_buffer
*buffer
)
3063 atomic_dec(&buffer
->record_disabled
);
3065 EXPORT_SYMBOL_GPL(ring_buffer_record_enable
);
3068 * ring_buffer_record_off - stop all writes into the buffer
3069 * @buffer: The ring buffer to stop writes to.
3071 * This prevents all writes to the buffer. Any attempt to write
3072 * to the buffer after this will fail and return NULL.
3074 * This is different than ring_buffer_record_disable() as
3075 * it works like an on/off switch, where as the disable() version
3076 * must be paired with a enable().
3078 void ring_buffer_record_off(struct ring_buffer
*buffer
)
3081 unsigned int new_rd
;
3084 rd
= atomic_read(&buffer
->record_disabled
);
3085 new_rd
= rd
| RB_BUFFER_OFF
;
3086 } while (atomic_cmpxchg(&buffer
->record_disabled
, rd
, new_rd
) != rd
);
3088 EXPORT_SYMBOL_GPL(ring_buffer_record_off
);
3091 * ring_buffer_record_on - restart writes into the buffer
3092 * @buffer: The ring buffer to start writes to.
3094 * This enables all writes to the buffer that was disabled by
3095 * ring_buffer_record_off().
3097 * This is different than ring_buffer_record_enable() as
3098 * it works like an on/off switch, where as the enable() version
3099 * must be paired with a disable().
3101 void ring_buffer_record_on(struct ring_buffer
*buffer
)
3104 unsigned int new_rd
;
3107 rd
= atomic_read(&buffer
->record_disabled
);
3108 new_rd
= rd
& ~RB_BUFFER_OFF
;
3109 } while (atomic_cmpxchg(&buffer
->record_disabled
, rd
, new_rd
) != rd
);
3111 EXPORT_SYMBOL_GPL(ring_buffer_record_on
);
3114 * ring_buffer_record_is_on - return true if the ring buffer can write
3115 * @buffer: The ring buffer to see if write is enabled
3117 * Returns true if the ring buffer is in a state that it accepts writes.
3119 int ring_buffer_record_is_on(struct ring_buffer
*buffer
)
3121 return !atomic_read(&buffer
->record_disabled
);
3125 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3126 * @buffer: The ring buffer to stop writes to.
3127 * @cpu: The CPU buffer to stop
3129 * This prevents all writes to the buffer. Any attempt to write
3130 * to the buffer after this will fail and return NULL.
3132 * The caller should call synchronize_sched() after this.
3134 void ring_buffer_record_disable_cpu(struct ring_buffer
*buffer
, int cpu
)
3136 struct ring_buffer_per_cpu
*cpu_buffer
;
3138 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3141 cpu_buffer
= buffer
->buffers
[cpu
];
3142 atomic_inc(&cpu_buffer
->record_disabled
);
3144 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu
);
3147 * ring_buffer_record_enable_cpu - enable writes to the buffer
3148 * @buffer: The ring buffer to enable writes
3149 * @cpu: The CPU to enable.
3151 * Note, multiple disables will need the same number of enables
3152 * to truly enable the writing (much like preempt_disable).
3154 void ring_buffer_record_enable_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_dec(&cpu_buffer
->record_disabled
);
3164 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu
);
3167 * The total entries in the ring buffer is the running counter
3168 * of entries entered into the ring buffer, minus the sum of
3169 * the entries read from the ring buffer and the number of
3170 * entries that were overwritten.
3172 static inline unsigned long
3173 rb_num_of_entries(struct ring_buffer_per_cpu
*cpu_buffer
)
3175 return local_read(&cpu_buffer
->entries
) -
3176 (local_read(&cpu_buffer
->overrun
) + cpu_buffer
->read
);
3180 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3181 * @buffer: The ring buffer
3182 * @cpu: The per CPU buffer to read from.
3184 u64
ring_buffer_oldest_event_ts(struct ring_buffer
*buffer
, int cpu
)
3186 unsigned long flags
;
3187 struct ring_buffer_per_cpu
*cpu_buffer
;
3188 struct buffer_page
*bpage
;
3191 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3194 cpu_buffer
= buffer
->buffers
[cpu
];
3195 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3197 * if the tail is on reader_page, oldest time stamp is on the reader
3200 if (cpu_buffer
->tail_page
== cpu_buffer
->reader_page
)
3201 bpage
= cpu_buffer
->reader_page
;
3203 bpage
= rb_set_head_page(cpu_buffer
);
3205 ret
= bpage
->page
->time_stamp
;
3206 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3210 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts
);
3213 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3214 * @buffer: The ring buffer
3215 * @cpu: The per CPU buffer to read from.
3217 unsigned long ring_buffer_bytes_cpu(struct ring_buffer
*buffer
, int cpu
)
3219 struct ring_buffer_per_cpu
*cpu_buffer
;
3222 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3225 cpu_buffer
= buffer
->buffers
[cpu
];
3226 ret
= local_read(&cpu_buffer
->entries_bytes
) - cpu_buffer
->read_bytes
;
3230 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu
);
3233 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3234 * @buffer: The ring buffer
3235 * @cpu: The per CPU buffer to get the entries from.
3237 unsigned long ring_buffer_entries_cpu(struct ring_buffer
*buffer
, int cpu
)
3239 struct ring_buffer_per_cpu
*cpu_buffer
;
3241 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3244 cpu_buffer
= buffer
->buffers
[cpu
];
3246 return rb_num_of_entries(cpu_buffer
);
3248 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu
);
3251 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3252 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3253 * @buffer: The ring buffer
3254 * @cpu: The per CPU buffer to get the number of overruns from
3256 unsigned long ring_buffer_overrun_cpu(struct ring_buffer
*buffer
, int cpu
)
3258 struct ring_buffer_per_cpu
*cpu_buffer
;
3261 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3264 cpu_buffer
= buffer
->buffers
[cpu
];
3265 ret
= local_read(&cpu_buffer
->overrun
);
3269 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu
);
3272 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3273 * commits failing due to the buffer wrapping around while there are uncommitted
3274 * events, such as during an interrupt storm.
3275 * @buffer: The ring buffer
3276 * @cpu: The per CPU buffer to get the number of overruns from
3279 ring_buffer_commit_overrun_cpu(struct ring_buffer
*buffer
, int cpu
)
3281 struct ring_buffer_per_cpu
*cpu_buffer
;
3284 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3287 cpu_buffer
= buffer
->buffers
[cpu
];
3288 ret
= local_read(&cpu_buffer
->commit_overrun
);
3292 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu
);
3295 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3296 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3297 * @buffer: The ring buffer
3298 * @cpu: The per CPU buffer to get the number of overruns from
3301 ring_buffer_dropped_events_cpu(struct ring_buffer
*buffer
, int cpu
)
3303 struct ring_buffer_per_cpu
*cpu_buffer
;
3306 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3309 cpu_buffer
= buffer
->buffers
[cpu
];
3310 ret
= local_read(&cpu_buffer
->dropped_events
);
3314 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu
);
3317 * ring_buffer_read_events_cpu - get the number of events successfully read
3318 * @buffer: The ring buffer
3319 * @cpu: The per CPU buffer to get the number of events read
3322 ring_buffer_read_events_cpu(struct ring_buffer
*buffer
, int cpu
)
3324 struct ring_buffer_per_cpu
*cpu_buffer
;
3326 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3329 cpu_buffer
= buffer
->buffers
[cpu
];
3330 return cpu_buffer
->read
;
3332 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu
);
3335 * ring_buffer_entries - get the number of entries in a buffer
3336 * @buffer: The ring buffer
3338 * Returns the total number of entries in the ring buffer
3341 unsigned long ring_buffer_entries(struct ring_buffer
*buffer
)
3343 struct ring_buffer_per_cpu
*cpu_buffer
;
3344 unsigned long entries
= 0;
3347 /* if you care about this being correct, lock the buffer */
3348 for_each_buffer_cpu(buffer
, cpu
) {
3349 cpu_buffer
= buffer
->buffers
[cpu
];
3350 entries
+= rb_num_of_entries(cpu_buffer
);
3355 EXPORT_SYMBOL_GPL(ring_buffer_entries
);
3358 * ring_buffer_overruns - get the number of overruns in buffer
3359 * @buffer: The ring buffer
3361 * Returns the total number of overruns in the ring buffer
3364 unsigned long ring_buffer_overruns(struct ring_buffer
*buffer
)
3366 struct ring_buffer_per_cpu
*cpu_buffer
;
3367 unsigned long overruns
= 0;
3370 /* if you care about this being correct, lock the buffer */
3371 for_each_buffer_cpu(buffer
, cpu
) {
3372 cpu_buffer
= buffer
->buffers
[cpu
];
3373 overruns
+= local_read(&cpu_buffer
->overrun
);
3378 EXPORT_SYMBOL_GPL(ring_buffer_overruns
);
3380 static void rb_iter_reset(struct ring_buffer_iter
*iter
)
3382 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
3384 /* Iterator usage is expected to have record disabled */
3385 iter
->head_page
= cpu_buffer
->reader_page
;
3386 iter
->head
= cpu_buffer
->reader_page
->read
;
3388 iter
->cache_reader_page
= iter
->head_page
;
3389 iter
->cache_read
= cpu_buffer
->read
;
3392 iter
->read_stamp
= cpu_buffer
->read_stamp
;
3394 iter
->read_stamp
= iter
->head_page
->page
->time_stamp
;
3398 * ring_buffer_iter_reset - reset an iterator
3399 * @iter: The iterator to reset
3401 * Resets the iterator, so that it will start from the beginning
3404 void ring_buffer_iter_reset(struct ring_buffer_iter
*iter
)
3406 struct ring_buffer_per_cpu
*cpu_buffer
;
3407 unsigned long flags
;
3412 cpu_buffer
= iter
->cpu_buffer
;
3414 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3415 rb_iter_reset(iter
);
3416 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3418 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset
);
3421 * ring_buffer_iter_empty - check if an iterator has no more to read
3422 * @iter: The iterator to check
3424 int ring_buffer_iter_empty(struct ring_buffer_iter
*iter
)
3426 struct ring_buffer_per_cpu
*cpu_buffer
;
3428 cpu_buffer
= iter
->cpu_buffer
;
3430 return iter
->head_page
== cpu_buffer
->commit_page
&&
3431 iter
->head
== rb_commit_index(cpu_buffer
);
3433 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty
);
3436 rb_update_read_stamp(struct ring_buffer_per_cpu
*cpu_buffer
,
3437 struct ring_buffer_event
*event
)
3441 switch (event
->type_len
) {
3442 case RINGBUF_TYPE_PADDING
:
3445 case RINGBUF_TYPE_TIME_EXTEND
:
3446 delta
= event
->array
[0];
3448 delta
+= event
->time_delta
;
3449 cpu_buffer
->read_stamp
+= delta
;
3452 case RINGBUF_TYPE_TIME_STAMP
:
3453 /* FIXME: not implemented */
3456 case RINGBUF_TYPE_DATA
:
3457 cpu_buffer
->read_stamp
+= event
->time_delta
;
3467 rb_update_iter_read_stamp(struct ring_buffer_iter
*iter
,
3468 struct ring_buffer_event
*event
)
3472 switch (event
->type_len
) {
3473 case RINGBUF_TYPE_PADDING
:
3476 case RINGBUF_TYPE_TIME_EXTEND
:
3477 delta
= event
->array
[0];
3479 delta
+= event
->time_delta
;
3480 iter
->read_stamp
+= delta
;
3483 case RINGBUF_TYPE_TIME_STAMP
:
3484 /* FIXME: not implemented */
3487 case RINGBUF_TYPE_DATA
:
3488 iter
->read_stamp
+= event
->time_delta
;
3497 static struct buffer_page
*
3498 rb_get_reader_page(struct ring_buffer_per_cpu
*cpu_buffer
)
3500 struct buffer_page
*reader
= NULL
;
3501 unsigned long overwrite
;
3502 unsigned long flags
;
3506 local_irq_save(flags
);
3507 arch_spin_lock(&cpu_buffer
->lock
);
3511 * This should normally only loop twice. But because the
3512 * start of the reader inserts an empty page, it causes
3513 * a case where we will loop three times. There should be no
3514 * reason to loop four times (that I know of).
3516 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 3)) {
3521 reader
= cpu_buffer
->reader_page
;
3523 /* If there's more to read, return this page */
3524 if (cpu_buffer
->reader_page
->read
< rb_page_size(reader
))
3527 /* Never should we have an index greater than the size */
3528 if (RB_WARN_ON(cpu_buffer
,
3529 cpu_buffer
->reader_page
->read
> rb_page_size(reader
)))
3532 /* check if we caught up to the tail */
3534 if (cpu_buffer
->commit_page
== cpu_buffer
->reader_page
)
3537 /* Don't bother swapping if the ring buffer is empty */
3538 if (rb_num_of_entries(cpu_buffer
) == 0)
3542 * Reset the reader page to size zero.
3544 local_set(&cpu_buffer
->reader_page
->write
, 0);
3545 local_set(&cpu_buffer
->reader_page
->entries
, 0);
3546 local_set(&cpu_buffer
->reader_page
->page
->commit
, 0);
3547 cpu_buffer
->reader_page
->real_end
= 0;
3551 * Splice the empty reader page into the list around the head.
3553 reader
= rb_set_head_page(cpu_buffer
);
3556 cpu_buffer
->reader_page
->list
.next
= rb_list_head(reader
->list
.next
);
3557 cpu_buffer
->reader_page
->list
.prev
= reader
->list
.prev
;
3560 * cpu_buffer->pages just needs to point to the buffer, it
3561 * has no specific buffer page to point to. Lets move it out
3562 * of our way so we don't accidentally swap it.
3564 cpu_buffer
->pages
= reader
->list
.prev
;
3566 /* The reader page will be pointing to the new head */
3567 rb_set_list_to_head(cpu_buffer
, &cpu_buffer
->reader_page
->list
);
3570 * We want to make sure we read the overruns after we set up our
3571 * pointers to the next object. The writer side does a
3572 * cmpxchg to cross pages which acts as the mb on the writer
3573 * side. Note, the reader will constantly fail the swap
3574 * while the writer is updating the pointers, so this
3575 * guarantees that the overwrite recorded here is the one we
3576 * want to compare with the last_overrun.
3579 overwrite
= local_read(&(cpu_buffer
->overrun
));
3582 * Here's the tricky part.
3584 * We need to move the pointer past the header page.
3585 * But we can only do that if a writer is not currently
3586 * moving it. The page before the header page has the
3587 * flag bit '1' set if it is pointing to the page we want.
3588 * but if the writer is in the process of moving it
3589 * than it will be '2' or already moved '0'.
3592 ret
= rb_head_page_replace(reader
, cpu_buffer
->reader_page
);
3595 * If we did not convert it, then we must try again.
3601 * Yeah! We succeeded in replacing the page.
3603 * Now make the new head point back to the reader page.
3605 rb_list_head(reader
->list
.next
)->prev
= &cpu_buffer
->reader_page
->list
;
3606 rb_inc_page(cpu_buffer
, &cpu_buffer
->head_page
);
3608 /* Finally update the reader page to the new head */
3609 cpu_buffer
->reader_page
= reader
;
3610 rb_reset_reader_page(cpu_buffer
);
3612 if (overwrite
!= cpu_buffer
->last_overrun
) {
3613 cpu_buffer
->lost_events
= overwrite
- cpu_buffer
->last_overrun
;
3614 cpu_buffer
->last_overrun
= overwrite
;
3620 arch_spin_unlock(&cpu_buffer
->lock
);
3621 local_irq_restore(flags
);
3626 static void rb_advance_reader(struct ring_buffer_per_cpu
*cpu_buffer
)
3628 struct ring_buffer_event
*event
;
3629 struct buffer_page
*reader
;
3632 reader
= rb_get_reader_page(cpu_buffer
);
3634 /* This function should not be called when buffer is empty */
3635 if (RB_WARN_ON(cpu_buffer
, !reader
))
3638 event
= rb_reader_event(cpu_buffer
);
3640 if (event
->type_len
<= RINGBUF_TYPE_DATA_TYPE_LEN_MAX
)
3643 rb_update_read_stamp(cpu_buffer
, event
);
3645 length
= rb_event_length(event
);
3646 cpu_buffer
->reader_page
->read
+= length
;
3649 static void rb_advance_iter(struct ring_buffer_iter
*iter
)
3651 struct ring_buffer_per_cpu
*cpu_buffer
;
3652 struct ring_buffer_event
*event
;
3655 cpu_buffer
= iter
->cpu_buffer
;
3658 * Check if we are at the end of the buffer.
3660 if (iter
->head
>= rb_page_size(iter
->head_page
)) {
3661 /* discarded commits can make the page empty */
3662 if (iter
->head_page
== cpu_buffer
->commit_page
)
3668 event
= rb_iter_head_event(iter
);
3670 length
= rb_event_length(event
);
3673 * This should not be called to advance the header if we are
3674 * at the tail of the buffer.
3676 if (RB_WARN_ON(cpu_buffer
,
3677 (iter
->head_page
== cpu_buffer
->commit_page
) &&
3678 (iter
->head
+ length
> rb_commit_index(cpu_buffer
))))
3681 rb_update_iter_read_stamp(iter
, event
);
3683 iter
->head
+= length
;
3685 /* check for end of page padding */
3686 if ((iter
->head
>= rb_page_size(iter
->head_page
)) &&
3687 (iter
->head_page
!= cpu_buffer
->commit_page
))
3691 static int rb_lost_events(struct ring_buffer_per_cpu
*cpu_buffer
)
3693 return cpu_buffer
->lost_events
;
3696 static struct ring_buffer_event
*
3697 rb_buffer_peek(struct ring_buffer_per_cpu
*cpu_buffer
, u64
*ts
,
3698 unsigned long *lost_events
)
3700 struct ring_buffer_event
*event
;
3701 struct buffer_page
*reader
;
3706 * We repeat when a time extend is encountered.
3707 * Since the time extend is always attached to a data event,
3708 * we should never loop more than once.
3709 * (We never hit the following condition more than twice).
3711 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 2))
3714 reader
= rb_get_reader_page(cpu_buffer
);
3718 event
= rb_reader_event(cpu_buffer
);
3720 switch (event
->type_len
) {
3721 case RINGBUF_TYPE_PADDING
:
3722 if (rb_null_event(event
))
3723 RB_WARN_ON(cpu_buffer
, 1);
3725 * Because the writer could be discarding every
3726 * event it creates (which would probably be bad)
3727 * if we were to go back to "again" then we may never
3728 * catch up, and will trigger the warn on, or lock
3729 * the box. Return the padding, and we will release
3730 * the current locks, and try again.
3734 case RINGBUF_TYPE_TIME_EXTEND
:
3735 /* Internal data, OK to advance */
3736 rb_advance_reader(cpu_buffer
);
3739 case RINGBUF_TYPE_TIME_STAMP
:
3740 /* FIXME: not implemented */
3741 rb_advance_reader(cpu_buffer
);
3744 case RINGBUF_TYPE_DATA
:
3746 *ts
= cpu_buffer
->read_stamp
+ event
->time_delta
;
3747 ring_buffer_normalize_time_stamp(cpu_buffer
->buffer
,
3748 cpu_buffer
->cpu
, ts
);
3751 *lost_events
= rb_lost_events(cpu_buffer
);
3760 EXPORT_SYMBOL_GPL(ring_buffer_peek
);
3762 static struct ring_buffer_event
*
3763 rb_iter_peek(struct ring_buffer_iter
*iter
, u64
*ts
)
3765 struct ring_buffer
*buffer
;
3766 struct ring_buffer_per_cpu
*cpu_buffer
;
3767 struct ring_buffer_event
*event
;
3770 cpu_buffer
= iter
->cpu_buffer
;
3771 buffer
= cpu_buffer
->buffer
;
3774 * Check if someone performed a consuming read to
3775 * the buffer. A consuming read invalidates the iterator
3776 * and we need to reset the iterator in this case.
3778 if (unlikely(iter
->cache_read
!= cpu_buffer
->read
||
3779 iter
->cache_reader_page
!= cpu_buffer
->reader_page
))
3780 rb_iter_reset(iter
);
3783 if (ring_buffer_iter_empty(iter
))
3787 * We repeat when a time extend is encountered or we hit
3788 * the end of the page. Since the time extend is always attached
3789 * to a data event, we should never loop more than three times.
3790 * Once for going to next page, once on time extend, and
3791 * finally once to get the event.
3792 * (We never hit the following condition more than thrice).
3794 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 3))
3797 if (rb_per_cpu_empty(cpu_buffer
))
3800 if (iter
->head
>= rb_page_size(iter
->head_page
)) {
3805 event
= rb_iter_head_event(iter
);
3807 switch (event
->type_len
) {
3808 case RINGBUF_TYPE_PADDING
:
3809 if (rb_null_event(event
)) {
3813 rb_advance_iter(iter
);
3816 case RINGBUF_TYPE_TIME_EXTEND
:
3817 /* Internal data, OK to advance */
3818 rb_advance_iter(iter
);
3821 case RINGBUF_TYPE_TIME_STAMP
:
3822 /* FIXME: not implemented */
3823 rb_advance_iter(iter
);
3826 case RINGBUF_TYPE_DATA
:
3828 *ts
= iter
->read_stamp
+ event
->time_delta
;
3829 ring_buffer_normalize_time_stamp(buffer
,
3830 cpu_buffer
->cpu
, ts
);
3840 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek
);
3842 static inline int rb_ok_to_lock(void)
3845 * If an NMI die dumps out the content of the ring buffer
3846 * do not grab locks. We also permanently disable the ring
3847 * buffer too. A one time deal is all you get from reading
3848 * the ring buffer from an NMI.
3850 if (likely(!in_nmi()))
3853 tracing_off_permanent();
3858 * ring_buffer_peek - peek at the next event to be read
3859 * @buffer: The ring buffer to read
3860 * @cpu: The cpu to peak at
3861 * @ts: The timestamp counter of this event.
3862 * @lost_events: a variable to store if events were lost (may be NULL)
3864 * This will return the event that will be read next, but does
3865 * not consume the data.
3867 struct ring_buffer_event
*
3868 ring_buffer_peek(struct ring_buffer
*buffer
, int cpu
, u64
*ts
,
3869 unsigned long *lost_events
)
3871 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
3872 struct ring_buffer_event
*event
;
3873 unsigned long flags
;
3876 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3879 dolock
= rb_ok_to_lock();
3881 local_irq_save(flags
);
3883 raw_spin_lock(&cpu_buffer
->reader_lock
);
3884 event
= rb_buffer_peek(cpu_buffer
, ts
, lost_events
);
3885 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3886 rb_advance_reader(cpu_buffer
);
3888 raw_spin_unlock(&cpu_buffer
->reader_lock
);
3889 local_irq_restore(flags
);
3891 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3898 * ring_buffer_iter_peek - peek at the next event to be read
3899 * @iter: The ring buffer iterator
3900 * @ts: The timestamp counter of this event.
3902 * This will return the event that will be read next, but does
3903 * not increment the iterator.
3905 struct ring_buffer_event
*
3906 ring_buffer_iter_peek(struct ring_buffer_iter
*iter
, u64
*ts
)
3908 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
3909 struct ring_buffer_event
*event
;
3910 unsigned long flags
;
3913 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3914 event
= rb_iter_peek(iter
, ts
);
3915 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3917 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3924 * ring_buffer_consume - return an event and consume it
3925 * @buffer: The ring buffer to get the next event from
3926 * @cpu: the cpu to read the buffer from
3927 * @ts: a variable to store the timestamp (may be NULL)
3928 * @lost_events: a variable to store if events were lost (may be NULL)
3930 * Returns the next event in the ring buffer, and that event is consumed.
3931 * Meaning, that sequential reads will keep returning a different event,
3932 * and eventually empty the ring buffer if the producer is slower.
3934 struct ring_buffer_event
*
3935 ring_buffer_consume(struct ring_buffer
*buffer
, int cpu
, u64
*ts
,
3936 unsigned long *lost_events
)
3938 struct ring_buffer_per_cpu
*cpu_buffer
;
3939 struct ring_buffer_event
*event
= NULL
;
3940 unsigned long flags
;
3943 dolock
= rb_ok_to_lock();
3946 /* might be called in atomic */
3949 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3952 cpu_buffer
= buffer
->buffers
[cpu
];
3953 local_irq_save(flags
);
3955 raw_spin_lock(&cpu_buffer
->reader_lock
);
3957 event
= rb_buffer_peek(cpu_buffer
, ts
, lost_events
);
3959 cpu_buffer
->lost_events
= 0;
3960 rb_advance_reader(cpu_buffer
);
3964 raw_spin_unlock(&cpu_buffer
->reader_lock
);
3965 local_irq_restore(flags
);
3970 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3975 EXPORT_SYMBOL_GPL(ring_buffer_consume
);
3978 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3979 * @buffer: The ring buffer to read from
3980 * @cpu: The cpu buffer to iterate over
3982 * This performs the initial preparations necessary to iterate
3983 * through the buffer. Memory is allocated, buffer recording
3984 * is disabled, and the iterator pointer is returned to the caller.
3986 * Disabling buffer recordng prevents the reading from being
3987 * corrupted. This is not a consuming read, so a producer is not
3990 * After a sequence of ring_buffer_read_prepare calls, the user is
3991 * expected to make at least one call to ring_buffer_read_prepare_sync.
3992 * Afterwards, ring_buffer_read_start is invoked to get things going
3995 * This overall must be paired with ring_buffer_read_finish.
3997 struct ring_buffer_iter
*
3998 ring_buffer_read_prepare(struct ring_buffer
*buffer
, int cpu
)
4000 struct ring_buffer_per_cpu
*cpu_buffer
;
4001 struct ring_buffer_iter
*iter
;
4003 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4006 iter
= kmalloc(sizeof(*iter
), GFP_KERNEL
);
4010 cpu_buffer
= buffer
->buffers
[cpu
];
4012 iter
->cpu_buffer
= cpu_buffer
;
4014 atomic_inc(&buffer
->resize_disabled
);
4015 atomic_inc(&cpu_buffer
->record_disabled
);
4019 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare
);
4022 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4024 * All previously invoked ring_buffer_read_prepare calls to prepare
4025 * iterators will be synchronized. Afterwards, read_buffer_read_start
4026 * calls on those iterators are allowed.
4029 ring_buffer_read_prepare_sync(void)
4031 synchronize_sched();
4033 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync
);
4036 * ring_buffer_read_start - start a non consuming read of the buffer
4037 * @iter: The iterator returned by ring_buffer_read_prepare
4039 * This finalizes the startup of an iteration through the buffer.
4040 * The iterator comes from a call to ring_buffer_read_prepare and
4041 * an intervening ring_buffer_read_prepare_sync must have been
4044 * Must be paired with ring_buffer_read_finish.
4047 ring_buffer_read_start(struct ring_buffer_iter
*iter
)
4049 struct ring_buffer_per_cpu
*cpu_buffer
;
4050 unsigned long flags
;
4055 cpu_buffer
= iter
->cpu_buffer
;
4057 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4058 arch_spin_lock(&cpu_buffer
->lock
);
4059 rb_iter_reset(iter
);
4060 arch_spin_unlock(&cpu_buffer
->lock
);
4061 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4063 EXPORT_SYMBOL_GPL(ring_buffer_read_start
);
4066 * ring_buffer_read_finish - finish reading the iterator of the buffer
4067 * @iter: The iterator retrieved by ring_buffer_start
4069 * This re-enables the recording to the buffer, and frees the
4073 ring_buffer_read_finish(struct ring_buffer_iter
*iter
)
4075 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
4076 unsigned long flags
;
4079 * Ring buffer is disabled from recording, here's a good place
4080 * to check the integrity of the ring buffer.
4081 * Must prevent readers from trying to read, as the check
4082 * clears the HEAD page and readers require it.
4084 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4085 rb_check_pages(cpu_buffer
);
4086 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4088 atomic_dec(&cpu_buffer
->record_disabled
);
4089 atomic_dec(&cpu_buffer
->buffer
->resize_disabled
);
4092 EXPORT_SYMBOL_GPL(ring_buffer_read_finish
);
4095 * ring_buffer_read - read the next item in the ring buffer by the iterator
4096 * @iter: The ring buffer iterator
4097 * @ts: The time stamp of the event read.
4099 * This reads the next event in the ring buffer and increments the iterator.
4101 struct ring_buffer_event
*
4102 ring_buffer_read(struct ring_buffer_iter
*iter
, u64
*ts
)
4104 struct ring_buffer_event
*event
;
4105 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
4106 unsigned long flags
;
4108 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4110 event
= rb_iter_peek(iter
, ts
);
4114 if (event
->type_len
== RINGBUF_TYPE_PADDING
)
4117 rb_advance_iter(iter
);
4119 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4123 EXPORT_SYMBOL_GPL(ring_buffer_read
);
4126 * ring_buffer_size - return the size of the ring buffer (in bytes)
4127 * @buffer: The ring buffer.
4129 unsigned long ring_buffer_size(struct ring_buffer
*buffer
, int cpu
)
4132 * Earlier, this method returned
4133 * BUF_PAGE_SIZE * buffer->nr_pages
4134 * Since the nr_pages field is now removed, we have converted this to
4135 * return the per cpu buffer value.
4137 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4140 return BUF_PAGE_SIZE
* buffer
->buffers
[cpu
]->nr_pages
;
4142 EXPORT_SYMBOL_GPL(ring_buffer_size
);
4145 rb_reset_cpu(struct ring_buffer_per_cpu
*cpu_buffer
)
4147 rb_head_page_deactivate(cpu_buffer
);
4149 cpu_buffer
->head_page
4150 = list_entry(cpu_buffer
->pages
, struct buffer_page
, list
);
4151 local_set(&cpu_buffer
->head_page
->write
, 0);
4152 local_set(&cpu_buffer
->head_page
->entries
, 0);
4153 local_set(&cpu_buffer
->head_page
->page
->commit
, 0);
4155 cpu_buffer
->head_page
->read
= 0;
4157 cpu_buffer
->tail_page
= cpu_buffer
->head_page
;
4158 cpu_buffer
->commit_page
= cpu_buffer
->head_page
;
4160 INIT_LIST_HEAD(&cpu_buffer
->reader_page
->list
);
4161 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
4162 local_set(&cpu_buffer
->reader_page
->write
, 0);
4163 local_set(&cpu_buffer
->reader_page
->entries
, 0);
4164 local_set(&cpu_buffer
->reader_page
->page
->commit
, 0);
4165 cpu_buffer
->reader_page
->read
= 0;
4167 local_set(&cpu_buffer
->entries_bytes
, 0);
4168 local_set(&cpu_buffer
->overrun
, 0);
4169 local_set(&cpu_buffer
->commit_overrun
, 0);
4170 local_set(&cpu_buffer
->dropped_events
, 0);
4171 local_set(&cpu_buffer
->entries
, 0);
4172 local_set(&cpu_buffer
->committing
, 0);
4173 local_set(&cpu_buffer
->commits
, 0);
4174 cpu_buffer
->read
= 0;
4175 cpu_buffer
->read_bytes
= 0;
4177 cpu_buffer
->write_stamp
= 0;
4178 cpu_buffer
->read_stamp
= 0;
4180 cpu_buffer
->lost_events
= 0;
4181 cpu_buffer
->last_overrun
= 0;
4183 rb_head_page_activate(cpu_buffer
);
4187 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4188 * @buffer: The ring buffer to reset a per cpu buffer of
4189 * @cpu: The CPU buffer to be reset
4191 void ring_buffer_reset_cpu(struct ring_buffer
*buffer
, int cpu
)
4193 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
4194 unsigned long flags
;
4196 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4199 atomic_inc(&buffer
->resize_disabled
);
4200 atomic_inc(&cpu_buffer
->record_disabled
);
4202 /* Make sure all commits have finished */
4203 synchronize_sched();
4205 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4207 if (RB_WARN_ON(cpu_buffer
, local_read(&cpu_buffer
->committing
)))
4210 arch_spin_lock(&cpu_buffer
->lock
);
4212 rb_reset_cpu(cpu_buffer
);
4214 arch_spin_unlock(&cpu_buffer
->lock
);
4217 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4219 atomic_dec(&cpu_buffer
->record_disabled
);
4220 atomic_dec(&buffer
->resize_disabled
);
4222 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu
);
4225 * ring_buffer_reset - reset a ring buffer
4226 * @buffer: The ring buffer to reset all cpu buffers
4228 void ring_buffer_reset(struct ring_buffer
*buffer
)
4232 for_each_buffer_cpu(buffer
, cpu
)
4233 ring_buffer_reset_cpu(buffer
, cpu
);
4235 EXPORT_SYMBOL_GPL(ring_buffer_reset
);
4238 * rind_buffer_empty - is the ring buffer empty?
4239 * @buffer: The ring buffer to test
4241 int ring_buffer_empty(struct ring_buffer
*buffer
)
4243 struct ring_buffer_per_cpu
*cpu_buffer
;
4244 unsigned long flags
;
4249 dolock
= rb_ok_to_lock();
4251 /* yes this is racy, but if you don't like the race, lock the buffer */
4252 for_each_buffer_cpu(buffer
, cpu
) {
4253 cpu_buffer
= buffer
->buffers
[cpu
];
4254 local_irq_save(flags
);
4256 raw_spin_lock(&cpu_buffer
->reader_lock
);
4257 ret
= rb_per_cpu_empty(cpu_buffer
);
4259 raw_spin_unlock(&cpu_buffer
->reader_lock
);
4260 local_irq_restore(flags
);
4268 EXPORT_SYMBOL_GPL(ring_buffer_empty
);
4271 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4272 * @buffer: The ring buffer
4273 * @cpu: The CPU buffer to test
4275 int ring_buffer_empty_cpu(struct ring_buffer
*buffer
, int cpu
)
4277 struct ring_buffer_per_cpu
*cpu_buffer
;
4278 unsigned long flags
;
4282 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4285 dolock
= rb_ok_to_lock();
4287 cpu_buffer
= buffer
->buffers
[cpu
];
4288 local_irq_save(flags
);
4290 raw_spin_lock(&cpu_buffer
->reader_lock
);
4291 ret
= rb_per_cpu_empty(cpu_buffer
);
4293 raw_spin_unlock(&cpu_buffer
->reader_lock
);
4294 local_irq_restore(flags
);
4298 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu
);
4300 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4302 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4303 * @buffer_a: One buffer to swap with
4304 * @buffer_b: The other buffer to swap with
4306 * This function is useful for tracers that want to take a "snapshot"
4307 * of a CPU buffer and has another back up buffer lying around.
4308 * it is expected that the tracer handles the cpu buffer not being
4309 * used at the moment.
4311 int ring_buffer_swap_cpu(struct ring_buffer
*buffer_a
,
4312 struct ring_buffer
*buffer_b
, int cpu
)
4314 struct ring_buffer_per_cpu
*cpu_buffer_a
;
4315 struct ring_buffer_per_cpu
*cpu_buffer_b
;
4318 if (!cpumask_test_cpu(cpu
, buffer_a
->cpumask
) ||
4319 !cpumask_test_cpu(cpu
, buffer_b
->cpumask
))
4322 cpu_buffer_a
= buffer_a
->buffers
[cpu
];
4323 cpu_buffer_b
= buffer_b
->buffers
[cpu
];
4325 /* At least make sure the two buffers are somewhat the same */
4326 if (cpu_buffer_a
->nr_pages
!= cpu_buffer_b
->nr_pages
)
4331 if (ring_buffer_flags
!= RB_BUFFERS_ON
)
4334 if (atomic_read(&buffer_a
->record_disabled
))
4337 if (atomic_read(&buffer_b
->record_disabled
))
4340 if (atomic_read(&cpu_buffer_a
->record_disabled
))
4343 if (atomic_read(&cpu_buffer_b
->record_disabled
))
4347 * We can't do a synchronize_sched here because this
4348 * function can be called in atomic context.
4349 * Normally this will be called from the same CPU as cpu.
4350 * If not it's up to the caller to protect this.
4352 atomic_inc(&cpu_buffer_a
->record_disabled
);
4353 atomic_inc(&cpu_buffer_b
->record_disabled
);
4356 if (local_read(&cpu_buffer_a
->committing
))
4358 if (local_read(&cpu_buffer_b
->committing
))
4361 buffer_a
->buffers
[cpu
] = cpu_buffer_b
;
4362 buffer_b
->buffers
[cpu
] = cpu_buffer_a
;
4364 cpu_buffer_b
->buffer
= buffer_a
;
4365 cpu_buffer_a
->buffer
= buffer_b
;
4370 atomic_dec(&cpu_buffer_a
->record_disabled
);
4371 atomic_dec(&cpu_buffer_b
->record_disabled
);
4375 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu
);
4376 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4379 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4380 * @buffer: the buffer to allocate for.
4381 * @cpu: the cpu buffer to allocate.
4383 * This function is used in conjunction with ring_buffer_read_page.
4384 * When reading a full page from the ring buffer, these functions
4385 * can be used to speed up the process. The calling function should
4386 * allocate a few pages first with this function. Then when it
4387 * needs to get pages from the ring buffer, it passes the result
4388 * of this function into ring_buffer_read_page, which will swap
4389 * the page that was allocated, with the read page of the buffer.
4392 * The page allocated, or NULL on error.
4394 void *ring_buffer_alloc_read_page(struct ring_buffer
*buffer
, int cpu
)
4396 struct buffer_data_page
*bpage
;
4399 page
= alloc_pages_node(cpu_to_node(cpu
),
4400 GFP_KERNEL
| __GFP_NORETRY
, 0);
4404 bpage
= page_address(page
);
4406 rb_init_page(bpage
);
4410 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page
);
4413 * ring_buffer_free_read_page - free an allocated read page
4414 * @buffer: the buffer the page was allocate for
4415 * @data: the page to free
4417 * Free a page allocated from ring_buffer_alloc_read_page.
4419 void ring_buffer_free_read_page(struct ring_buffer
*buffer
, void *data
)
4421 free_page((unsigned long)data
);
4423 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page
);
4426 * ring_buffer_read_page - extract a page from the ring buffer
4427 * @buffer: buffer to extract from
4428 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4429 * @len: amount to extract
4430 * @cpu: the cpu of the buffer to extract
4431 * @full: should the extraction only happen when the page is full.
4433 * This function will pull out a page from the ring buffer and consume it.
4434 * @data_page must be the address of the variable that was returned
4435 * from ring_buffer_alloc_read_page. This is because the page might be used
4436 * to swap with a page in the ring buffer.
4439 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4442 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4444 * process_page(rpage, ret);
4446 * When @full is set, the function will not return true unless
4447 * the writer is off the reader page.
4449 * Note: it is up to the calling functions to handle sleeps and wakeups.
4450 * The ring buffer can be used anywhere in the kernel and can not
4451 * blindly call wake_up. The layer that uses the ring buffer must be
4452 * responsible for that.
4455 * >=0 if data has been transferred, returns the offset of consumed data.
4456 * <0 if no data has been transferred.
4458 int ring_buffer_read_page(struct ring_buffer
*buffer
,
4459 void **data_page
, size_t len
, int cpu
, int full
)
4461 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
4462 struct ring_buffer_event
*event
;
4463 struct buffer_data_page
*bpage
;
4464 struct buffer_page
*reader
;
4465 unsigned long missed_events
;
4466 unsigned long flags
;
4467 unsigned int commit
;
4472 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4476 * If len is not big enough to hold the page header, then
4477 * we can not copy anything.
4479 if (len
<= BUF_PAGE_HDR_SIZE
)
4482 len
-= BUF_PAGE_HDR_SIZE
;
4491 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4493 reader
= rb_get_reader_page(cpu_buffer
);
4497 event
= rb_reader_event(cpu_buffer
);
4499 read
= reader
->read
;
4500 commit
= rb_page_commit(reader
);
4502 /* Check if any events were dropped */
4503 missed_events
= cpu_buffer
->lost_events
;
4506 * If this page has been partially read or
4507 * if len is not big enough to read the rest of the page or
4508 * a writer is still on the page, then
4509 * we must copy the data from the page to the buffer.
4510 * Otherwise, we can simply swap the page with the one passed in.
4512 if (read
|| (len
< (commit
- read
)) ||
4513 cpu_buffer
->reader_page
== cpu_buffer
->commit_page
) {
4514 struct buffer_data_page
*rpage
= cpu_buffer
->reader_page
->page
;
4515 unsigned int rpos
= read
;
4516 unsigned int pos
= 0;
4522 if (len
> (commit
- read
))
4523 len
= (commit
- read
);
4525 /* Always keep the time extend and data together */
4526 size
= rb_event_ts_length(event
);
4531 /* save the current timestamp, since the user will need it */
4532 save_timestamp
= cpu_buffer
->read_stamp
;
4534 /* Need to copy one event at a time */
4536 /* We need the size of one event, because
4537 * rb_advance_reader only advances by one event,
4538 * whereas rb_event_ts_length may include the size of
4539 * one or two events.
4540 * We have already ensured there's enough space if this
4541 * is a time extend. */
4542 size
= rb_event_length(event
);
4543 memcpy(bpage
->data
+ pos
, rpage
->data
+ rpos
, size
);
4547 rb_advance_reader(cpu_buffer
);
4548 rpos
= reader
->read
;
4554 event
= rb_reader_event(cpu_buffer
);
4555 /* Always keep the time extend and data together */
4556 size
= rb_event_ts_length(event
);
4557 } while (len
>= size
);
4560 local_set(&bpage
->commit
, pos
);
4561 bpage
->time_stamp
= save_timestamp
;
4563 /* we copied everything to the beginning */
4566 /* update the entry counter */
4567 cpu_buffer
->read
+= rb_page_entries(reader
);
4568 cpu_buffer
->read_bytes
+= BUF_PAGE_SIZE
;
4570 /* swap the pages */
4571 rb_init_page(bpage
);
4572 bpage
= reader
->page
;
4573 reader
->page
= *data_page
;
4574 local_set(&reader
->write
, 0);
4575 local_set(&reader
->entries
, 0);
4580 * Use the real_end for the data size,
4581 * This gives us a chance to store the lost events
4584 if (reader
->real_end
)
4585 local_set(&bpage
->commit
, reader
->real_end
);
4589 cpu_buffer
->lost_events
= 0;
4591 commit
= local_read(&bpage
->commit
);
4593 * Set a flag in the commit field if we lost events
4595 if (missed_events
) {
4596 /* If there is room at the end of the page to save the
4597 * missed events, then record it there.
4599 if (BUF_PAGE_SIZE
- commit
>= sizeof(missed_events
)) {
4600 memcpy(&bpage
->data
[commit
], &missed_events
,
4601 sizeof(missed_events
));
4602 local_add(RB_MISSED_STORED
, &bpage
->commit
);
4603 commit
+= sizeof(missed_events
);
4605 local_add(RB_MISSED_EVENTS
, &bpage
->commit
);
4609 * This page may be off to user land. Zero it out here.
4611 if (commit
< BUF_PAGE_SIZE
)
4612 memset(&bpage
->data
[commit
], 0, BUF_PAGE_SIZE
- commit
);
4615 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4620 EXPORT_SYMBOL_GPL(ring_buffer_read_page
);
4622 #ifdef CONFIG_HOTPLUG_CPU
4623 static int rb_cpu_notify(struct notifier_block
*self
,
4624 unsigned long action
, void *hcpu
)
4626 struct ring_buffer
*buffer
=
4627 container_of(self
, struct ring_buffer
, cpu_notify
);
4628 long cpu
= (long)hcpu
;
4629 int cpu_i
, nr_pages_same
;
4630 unsigned int nr_pages
;
4633 case CPU_UP_PREPARE
:
4634 case CPU_UP_PREPARE_FROZEN
:
4635 if (cpumask_test_cpu(cpu
, buffer
->cpumask
))
4640 /* check if all cpu sizes are same */
4641 for_each_buffer_cpu(buffer
, cpu_i
) {
4642 /* fill in the size from first enabled cpu */
4644 nr_pages
= buffer
->buffers
[cpu_i
]->nr_pages
;
4645 if (nr_pages
!= buffer
->buffers
[cpu_i
]->nr_pages
) {
4650 /* allocate minimum pages, user can later expand it */
4653 buffer
->buffers
[cpu
] =
4654 rb_allocate_cpu_buffer(buffer
, nr_pages
, cpu
);
4655 if (!buffer
->buffers
[cpu
]) {
4656 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4661 cpumask_set_cpu(cpu
, buffer
->cpumask
);
4663 case CPU_DOWN_PREPARE
:
4664 case CPU_DOWN_PREPARE_FROZEN
:
4667 * If we were to free the buffer, then the user would
4668 * lose any trace that was in the buffer.
4678 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4680 * This is a basic integrity check of the ring buffer.
4681 * Late in the boot cycle this test will run when configured in.
4682 * It will kick off a thread per CPU that will go into a loop
4683 * writing to the per cpu ring buffer various sizes of data.
4684 * Some of the data will be large items, some small.
4686 * Another thread is created that goes into a spin, sending out
4687 * IPIs to the other CPUs to also write into the ring buffer.
4688 * this is to test the nesting ability of the buffer.
4690 * Basic stats are recorded and reported. If something in the
4691 * ring buffer should happen that's not expected, a big warning
4692 * is displayed and all ring buffers are disabled.
4694 static struct task_struct
*rb_threads
[NR_CPUS
] __initdata
;
4696 struct rb_test_data
{
4697 struct ring_buffer
*buffer
;
4698 unsigned long events
;
4699 unsigned long bytes_written
;
4700 unsigned long bytes_alloc
;
4701 unsigned long bytes_dropped
;
4702 unsigned long events_nested
;
4703 unsigned long bytes_written_nested
;
4704 unsigned long bytes_alloc_nested
;
4705 unsigned long bytes_dropped_nested
;
4706 int min_size_nested
;
4707 int max_size_nested
;
4714 static struct rb_test_data rb_data
[NR_CPUS
] __initdata
;
4717 #define RB_TEST_BUFFER_SIZE 1048576
4719 static char rb_string
[] __initdata
=
4720 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4721 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4722 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4724 static bool rb_test_started __initdata
;
4731 static __init
int rb_write_something(struct rb_test_data
*data
, bool nested
)
4733 struct ring_buffer_event
*event
;
4734 struct rb_item
*item
;
4741 /* Have nested writes different that what is written */
4742 cnt
= data
->cnt
+ (nested
? 27 : 0);
4744 /* Multiply cnt by ~e, to make some unique increment */
4745 size
= (data
->cnt
* 68 / 25) % (sizeof(rb_string
) - 1);
4747 len
= size
+ sizeof(struct rb_item
);
4749 started
= rb_test_started
;
4750 /* read rb_test_started before checking buffer enabled */
4753 event
= ring_buffer_lock_reserve(data
->buffer
, len
);
4755 /* Ignore dropped events before test starts. */
4758 data
->bytes_dropped
+= len
;
4760 data
->bytes_dropped_nested
+= len
;
4765 event_len
= ring_buffer_event_length(event
);
4767 if (RB_WARN_ON(data
->buffer
, event_len
< len
))
4770 item
= ring_buffer_event_data(event
);
4772 memcpy(item
->str
, rb_string
, size
);
4775 data
->bytes_alloc_nested
+= event_len
;
4776 data
->bytes_written_nested
+= len
;
4777 data
->events_nested
++;
4778 if (!data
->min_size_nested
|| len
< data
->min_size_nested
)
4779 data
->min_size_nested
= len
;
4780 if (len
> data
->max_size_nested
)
4781 data
->max_size_nested
= len
;
4783 data
->bytes_alloc
+= event_len
;
4784 data
->bytes_written
+= len
;
4786 if (!data
->min_size
|| len
< data
->min_size
)
4787 data
->max_size
= len
;
4788 if (len
> data
->max_size
)
4789 data
->max_size
= len
;
4793 ring_buffer_unlock_commit(data
->buffer
, event
);
4798 static __init
int rb_test(void *arg
)
4800 struct rb_test_data
*data
= arg
;
4802 while (!kthread_should_stop()) {
4803 rb_write_something(data
, false);
4806 set_current_state(TASK_INTERRUPTIBLE
);
4807 /* Now sleep between a min of 100-300us and a max of 1ms */
4808 usleep_range(((data
->cnt
% 3) + 1) * 100, 1000);
4814 static __init
void rb_ipi(void *ignore
)
4816 struct rb_test_data
*data
;
4817 int cpu
= smp_processor_id();
4819 data
= &rb_data
[cpu
];
4820 rb_write_something(data
, true);
4823 static __init
int rb_hammer_test(void *arg
)
4825 while (!kthread_should_stop()) {
4827 /* Send an IPI to all cpus to write data! */
4828 smp_call_function(rb_ipi
, NULL
, 1);
4829 /* No sleep, but for non preempt, let others run */
4836 static __init
int test_ringbuffer(void)
4838 struct task_struct
*rb_hammer
;
4839 struct ring_buffer
*buffer
;
4843 pr_info("Running ring buffer tests...\n");
4845 buffer
= ring_buffer_alloc(RB_TEST_BUFFER_SIZE
, RB_FL_OVERWRITE
);
4846 if (WARN_ON(!buffer
))
4849 /* Disable buffer so that threads can't write to it yet */
4850 ring_buffer_record_off(buffer
);
4852 for_each_online_cpu(cpu
) {
4853 rb_data
[cpu
].buffer
= buffer
;
4854 rb_data
[cpu
].cpu
= cpu
;
4855 rb_data
[cpu
].cnt
= cpu
;
4856 rb_threads
[cpu
] = kthread_create(rb_test
, &rb_data
[cpu
],
4857 "rbtester/%d", cpu
);
4858 if (WARN_ON(!rb_threads
[cpu
])) {
4859 pr_cont("FAILED\n");
4864 kthread_bind(rb_threads
[cpu
], cpu
);
4865 wake_up_process(rb_threads
[cpu
]);
4868 /* Now create the rb hammer! */
4869 rb_hammer
= kthread_run(rb_hammer_test
, NULL
, "rbhammer");
4870 if (WARN_ON(!rb_hammer
)) {
4871 pr_cont("FAILED\n");
4876 ring_buffer_record_on(buffer
);
4878 * Show buffer is enabled before setting rb_test_started.
4879 * Yes there's a small race window where events could be
4880 * dropped and the thread wont catch it. But when a ring
4881 * buffer gets enabled, there will always be some kind of
4882 * delay before other CPUs see it. Thus, we don't care about
4883 * those dropped events. We care about events dropped after
4884 * the threads see that the buffer is active.
4887 rb_test_started
= true;
4889 set_current_state(TASK_INTERRUPTIBLE
);
4890 /* Just run for 10 seconds */;
4891 schedule_timeout(10 * HZ
);
4893 kthread_stop(rb_hammer
);
4896 for_each_online_cpu(cpu
) {
4897 if (!rb_threads
[cpu
])
4899 kthread_stop(rb_threads
[cpu
]);
4902 ring_buffer_free(buffer
);
4907 pr_info("finished\n");
4908 for_each_online_cpu(cpu
) {
4909 struct ring_buffer_event
*event
;
4910 struct rb_test_data
*data
= &rb_data
[cpu
];
4911 struct rb_item
*item
;
4912 unsigned long total_events
;
4913 unsigned long total_dropped
;
4914 unsigned long total_written
;
4915 unsigned long total_alloc
;
4916 unsigned long total_read
= 0;
4917 unsigned long total_size
= 0;
4918 unsigned long total_len
= 0;
4919 unsigned long total_lost
= 0;
4922 int small_event_size
;
4926 total_events
= data
->events
+ data
->events_nested
;
4927 total_written
= data
->bytes_written
+ data
->bytes_written_nested
;
4928 total_alloc
= data
->bytes_alloc
+ data
->bytes_alloc_nested
;
4929 total_dropped
= data
->bytes_dropped
+ data
->bytes_dropped_nested
;
4931 big_event_size
= data
->max_size
+ data
->max_size_nested
;
4932 small_event_size
= data
->min_size
+ data
->min_size_nested
;
4934 pr_info("CPU %d:\n", cpu
);
4935 pr_info(" events: %ld\n", total_events
);
4936 pr_info(" dropped bytes: %ld\n", total_dropped
);
4937 pr_info(" alloced bytes: %ld\n", total_alloc
);
4938 pr_info(" written bytes: %ld\n", total_written
);
4939 pr_info(" biggest event: %d\n", big_event_size
);
4940 pr_info(" smallest event: %d\n", small_event_size
);
4942 if (RB_WARN_ON(buffer
, total_dropped
))
4947 while ((event
= ring_buffer_consume(buffer
, cpu
, NULL
, &lost
))) {
4949 item
= ring_buffer_event_data(event
);
4950 total_len
+= ring_buffer_event_length(event
);
4951 total_size
+= item
->size
+ sizeof(struct rb_item
);
4952 if (memcmp(&item
->str
[0], rb_string
, item
->size
) != 0) {
4953 pr_info("FAILED!\n");
4954 pr_info("buffer had: %.*s\n", item
->size
, item
->str
);
4955 pr_info("expected: %.*s\n", item
->size
, rb_string
);
4956 RB_WARN_ON(buffer
, 1);
4967 pr_info(" read events: %ld\n", total_read
);
4968 pr_info(" lost events: %ld\n", total_lost
);
4969 pr_info(" total events: %ld\n", total_lost
+ total_read
);
4970 pr_info(" recorded len bytes: %ld\n", total_len
);
4971 pr_info(" recorded size bytes: %ld\n", total_size
);
4973 pr_info(" With dropped events, record len and size may not match\n"
4974 " alloced and written from above\n");
4976 if (RB_WARN_ON(buffer
, total_len
!= total_alloc
||
4977 total_size
!= total_written
))
4980 if (RB_WARN_ON(buffer
, total_lost
+ total_read
!= total_events
))
4986 pr_info("Ring buffer PASSED!\n");
4988 ring_buffer_free(buffer
);
4992 late_initcall(test_ringbuffer
);
4993 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */