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 #ifdef CONFIG_MTK_EXTMEM
31 extern void* extmem_malloc_page_align(size_t bytes
);
32 extern void extmem_free(void* mem
);
35 static void update_pages_handler(struct work_struct
*work
);
38 * The ring buffer header is special. We must manually up keep it.
40 int ring_buffer_print_entry_header(struct trace_seq
*s
)
44 ret
= trace_seq_printf(s
, "# compressed entry header\n");
45 ret
= trace_seq_printf(s
, "\ttype_len : 5 bits\n");
46 ret
= trace_seq_printf(s
, "\ttime_delta : 27 bits\n");
47 ret
= trace_seq_printf(s
, "\tarray : 32 bits\n");
48 ret
= trace_seq_printf(s
, "\n");
49 ret
= trace_seq_printf(s
, "\tpadding : type == %d\n",
50 RINGBUF_TYPE_PADDING
);
51 ret
= trace_seq_printf(s
, "\ttime_extend : type == %d\n",
52 RINGBUF_TYPE_TIME_EXTEND
);
53 ret
= trace_seq_printf(s
, "\tdata max type_len == %d\n",
54 RINGBUF_TYPE_DATA_TYPE_LEN_MAX
);
60 * The ring buffer is made up of a list of pages. A separate list of pages is
61 * allocated for each CPU. A writer may only write to a buffer that is
62 * associated with the CPU it is currently executing on. A reader may read
63 * from any per cpu buffer.
65 * The reader is special. For each per cpu buffer, the reader has its own
66 * reader page. When a reader has read the entire reader page, this reader
67 * page is swapped with another page in the ring buffer.
69 * Now, as long as the writer is off the reader page, the reader can do what
70 * ever it wants with that page. The writer will never write to that page
71 * again (as long as it is out of the ring buffer).
73 * Here's some silly ASCII art.
76 * |reader| RING BUFFER
78 * +------+ +---+ +---+ +---+
87 * |reader| RING BUFFER
88 * |page |------------------v
89 * +------+ +---+ +---+ +---+
98 * |reader| RING BUFFER
99 * |page |------------------v
100 * +------+ +---+ +---+ +---+
102 * | +---+ +---+ +---+
105 * +------------------------------+
109 * |buffer| RING BUFFER
110 * |page |------------------v
111 * +------+ +---+ +---+ +---+
113 * | New +---+ +---+ +---+
116 * +------------------------------+
119 * After we make this swap, the reader can hand this page off to the splice
120 * code and be done with it. It can even allocate a new page if it needs to
121 * and swap that into the ring buffer.
123 * We will be using cmpxchg soon to make all this lockless.
128 * A fast way to enable or disable all ring buffers is to
129 * call tracing_on or tracing_off. Turning off the ring buffers
130 * prevents all ring buffers from being recorded to.
131 * Turning this switch on, makes it OK to write to the
132 * ring buffer, if the ring buffer is enabled itself.
134 * There's three layers that must be on in order to write
135 * to the ring buffer.
137 * 1) This global flag must be set.
138 * 2) The ring buffer must be enabled for recording.
139 * 3) The per cpu buffer must be enabled for recording.
141 * In case of an anomaly, this global flag has a bit set that
142 * will permantly disable all ring buffers.
146 * Global flag to disable all recording to ring buffers
147 * This has two bits: ON, DISABLED
151 * 0 0 : ring buffers are off
152 * 1 0 : ring buffers are on
153 * X 1 : ring buffers are permanently disabled
157 RB_BUFFERS_ON_BIT
= 0,
158 RB_BUFFERS_DISABLED_BIT
= 1,
162 RB_BUFFERS_ON
= 1 << RB_BUFFERS_ON_BIT
,
163 RB_BUFFERS_DISABLED
= 1 << RB_BUFFERS_DISABLED_BIT
,
166 static unsigned long ring_buffer_flags __read_mostly
= RB_BUFFERS_ON
;
168 /* Used for individual buffers (after the counter) */
169 #define RB_BUFFER_OFF (1 << 20)
171 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
174 * tracing_off_permanent - permanently disable ring buffers
176 * This function, once called, will disable all ring buffers
179 void tracing_off_permanent(void)
181 set_bit(RB_BUFFERS_DISABLED_BIT
, &ring_buffer_flags
);
184 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
185 #define RB_ALIGNMENT 4U
186 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
187 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
189 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
190 # define RB_FORCE_8BYTE_ALIGNMENT 0
191 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
193 # define RB_FORCE_8BYTE_ALIGNMENT 1
194 # define RB_ARCH_ALIGNMENT 8U
197 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
199 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
200 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
203 RB_LEN_TIME_EXTEND
= 8,
204 RB_LEN_TIME_STAMP
= 16,
207 #define skip_time_extend(event) \
208 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
210 static inline int rb_null_event(struct ring_buffer_event
*event
)
212 return event
->type_len
== RINGBUF_TYPE_PADDING
&& !event
->time_delta
;
215 static void rb_event_set_padding(struct ring_buffer_event
*event
)
217 /* padding has a NULL time_delta */
218 event
->type_len
= RINGBUF_TYPE_PADDING
;
219 event
->time_delta
= 0;
223 rb_event_data_length(struct ring_buffer_event
*event
)
228 length
= event
->type_len
* RB_ALIGNMENT
;
230 length
= event
->array
[0];
231 return length
+ RB_EVNT_HDR_SIZE
;
235 * Return the length of the given event. Will return
236 * the length of the time extend if the event is a
239 static inline unsigned
240 rb_event_length(struct ring_buffer_event
*event
)
242 switch (event
->type_len
) {
243 case RINGBUF_TYPE_PADDING
:
244 if (rb_null_event(event
))
247 return event
->array
[0] + RB_EVNT_HDR_SIZE
;
249 case RINGBUF_TYPE_TIME_EXTEND
:
250 return RB_LEN_TIME_EXTEND
;
252 case RINGBUF_TYPE_TIME_STAMP
:
253 return RB_LEN_TIME_STAMP
;
255 case RINGBUF_TYPE_DATA
:
256 return rb_event_data_length(event
);
265 * Return total length of time extend and data,
266 * or just the event length for all other events.
268 static inline unsigned
269 rb_event_ts_length(struct ring_buffer_event
*event
)
273 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
) {
274 /* time extends include the data event after it */
275 len
= RB_LEN_TIME_EXTEND
;
276 event
= skip_time_extend(event
);
278 return len
+ rb_event_length(event
);
282 * ring_buffer_event_length - return the length of the event
283 * @event: the event to get the length of
285 * Returns the size of the data load of a data event.
286 * If the event is something other than a data event, it
287 * returns the size of the event itself. With the exception
288 * of a TIME EXTEND, where it still returns the size of the
289 * data load of the data event after it.
291 unsigned ring_buffer_event_length(struct ring_buffer_event
*event
)
295 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
296 event
= skip_time_extend(event
);
298 length
= rb_event_length(event
);
299 if (event
->type_len
> RINGBUF_TYPE_DATA_TYPE_LEN_MAX
)
301 length
-= RB_EVNT_HDR_SIZE
;
302 if (length
> RB_MAX_SMALL_DATA
+ sizeof(event
->array
[0]))
303 length
-= sizeof(event
->array
[0]);
306 EXPORT_SYMBOL_GPL(ring_buffer_event_length
);
308 /* inline for ring buffer fast paths */
310 rb_event_data(struct ring_buffer_event
*event
)
312 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
313 event
= skip_time_extend(event
);
314 BUG_ON(event
->type_len
> RINGBUF_TYPE_DATA_TYPE_LEN_MAX
);
315 /* If length is in len field, then array[0] has the data */
317 return (void *)&event
->array
[0];
318 /* Otherwise length is in array[0] and array[1] has the data */
319 return (void *)&event
->array
[1];
323 * ring_buffer_event_data - return the data of the event
324 * @event: the event to get the data from
326 void *ring_buffer_event_data(struct ring_buffer_event
*event
)
328 return rb_event_data(event
);
330 EXPORT_SYMBOL_GPL(ring_buffer_event_data
);
332 #define for_each_buffer_cpu(buffer, cpu) \
333 for_each_cpu(cpu, buffer->cpumask)
336 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
337 #define TS_DELTA_TEST (~TS_MASK)
339 /* Flag when events were overwritten */
340 #define RB_MISSED_EVENTS (1 << 31)
341 /* Missed count stored at end */
342 #define RB_MISSED_STORED (1 << 30)
344 struct buffer_data_page
{
345 u64 time_stamp
; /* page time stamp */
346 local_t commit
; /* write committed index */
347 unsigned char data
[] RB_ALIGN_DATA
; /* data of buffer page */
351 * Note, the buffer_page list must be first. The buffer pages
352 * are allocated in cache lines, which means that each buffer
353 * page will be at the beginning of a cache line, and thus
354 * the least significant bits will be zero. We use this to
355 * add flags in the list struct pointers, to make the ring buffer
359 struct list_head list
; /* list of buffer pages */
360 local_t write
; /* index for next write */
361 unsigned read
; /* index for next read */
362 local_t entries
; /* entries on this page */
363 unsigned long real_end
; /* real end of data */
364 struct buffer_data_page
*page
; /* Actual data page */
368 * The buffer page counters, write and entries, must be reset
369 * atomically when crossing page boundaries. To synchronize this
370 * update, two counters are inserted into the number. One is
371 * the actual counter for the write position or count on the page.
373 * The other is a counter of updaters. Before an update happens
374 * the update partition of the counter is incremented. This will
375 * allow the updater to update the counter atomically.
377 * The counter is 20 bits, and the state data is 12.
379 #define RB_WRITE_MASK 0xfffff
380 #define RB_WRITE_INTCNT (1 << 20)
382 static void rb_init_page(struct buffer_data_page
*bpage
)
384 local_set(&bpage
->commit
, 0);
388 * ring_buffer_page_len - the size of data on the page.
389 * @page: The page to read
391 * Returns the amount of data on the page, including buffer page header.
393 size_t ring_buffer_page_len(void *page
)
395 return local_read(&((struct buffer_data_page
*)page
)->commit
)
400 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
403 static void free_buffer_page(struct buffer_page
*bpage
)
405 #ifdef CONFIG_MTK_EXTMEM
406 extmem_free((void*) bpage
->page
);
408 free_page((unsigned long)bpage
->page
);
414 * We need to fit the time_stamp delta into 27 bits.
416 static inline int test_time_stamp(u64 delta
)
418 if (delta
& TS_DELTA_TEST
)
423 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
425 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
426 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
428 int ring_buffer_print_page_header(struct trace_seq
*s
)
430 struct buffer_data_page field
;
433 ret
= trace_seq_printf(s
, "\tfield: u64 timestamp;\t"
434 "offset:0;\tsize:%u;\tsigned:%u;\n",
435 (unsigned int)sizeof(field
.time_stamp
),
436 (unsigned int)is_signed_type(u64
));
438 ret
= trace_seq_printf(s
, "\tfield: local_t commit;\t"
439 "offset:%u;\tsize:%u;\tsigned:%u;\n",
440 (unsigned int)offsetof(typeof(field
), commit
),
441 (unsigned int)sizeof(field
.commit
),
442 (unsigned int)is_signed_type(long));
444 ret
= trace_seq_printf(s
, "\tfield: int overwrite;\t"
445 "offset:%u;\tsize:%u;\tsigned:%u;\n",
446 (unsigned int)offsetof(typeof(field
), commit
),
448 (unsigned int)is_signed_type(long));
450 ret
= trace_seq_printf(s
, "\tfield: char data;\t"
451 "offset:%u;\tsize:%u;\tsigned:%u;\n",
452 (unsigned int)offsetof(typeof(field
), data
),
453 (unsigned int)BUF_PAGE_SIZE
,
454 (unsigned int)is_signed_type(char));
460 struct irq_work work
;
461 wait_queue_head_t waiters
;
462 bool waiters_pending
;
466 * head_page == tail_page && head == tail then buffer is empty.
468 struct ring_buffer_per_cpu
{
470 atomic_t record_disabled
;
471 struct ring_buffer
*buffer
;
472 raw_spinlock_t reader_lock
; /* serialize readers */
473 arch_spinlock_t lock
;
474 struct lock_class_key lock_key
;
475 unsigned long nr_pages
;
476 struct list_head
*pages
;
477 struct buffer_page
*head_page
; /* read from head */
478 struct buffer_page
*tail_page
; /* write to tail */
479 struct buffer_page
*commit_page
; /* committed pages */
480 struct buffer_page
*reader_page
;
481 unsigned long lost_events
;
482 unsigned long last_overrun
;
483 local_t entries_bytes
;
486 local_t commit_overrun
;
487 local_t dropped_events
;
491 unsigned long read_bytes
;
494 /* ring buffer pages to update, > 0 to add, < 0 to remove */
495 long nr_pages_to_update
;
496 struct list_head new_pages
; /* new pages to add */
497 struct work_struct update_pages_work
;
498 struct completion update_done
;
500 struct rb_irq_work irq_work
;
506 atomic_t record_disabled
;
507 atomic_t resize_disabled
;
508 cpumask_var_t cpumask
;
510 struct lock_class_key
*reader_lock_key
;
514 struct ring_buffer_per_cpu
**buffers
;
516 #ifdef CONFIG_HOTPLUG_CPU
517 struct notifier_block cpu_notify
;
521 struct rb_irq_work irq_work
;
524 struct ring_buffer_iter
{
525 struct ring_buffer_per_cpu
*cpu_buffer
;
527 struct buffer_page
*head_page
;
528 struct buffer_page
*cache_reader_page
;
529 unsigned long cache_read
;
534 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
536 * Schedules a delayed work to wake up any task that is blocked on the
537 * ring buffer waiters queue.
539 static void rb_wake_up_waiters(struct irq_work
*work
)
541 struct rb_irq_work
*rbwork
= container_of(work
, struct rb_irq_work
, work
);
543 wake_up_all(&rbwork
->waiters
);
547 * ring_buffer_wait - wait for input to the ring buffer
548 * @buffer: buffer to wait on
549 * @cpu: the cpu buffer to wait on
551 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
552 * as data is added to any of the @buffer's cpu buffers. Otherwise
553 * it will wait for data to be added to a specific cpu buffer.
555 int ring_buffer_wait(struct ring_buffer
*buffer
, int cpu
)
557 struct ring_buffer_per_cpu
*cpu_buffer
;
559 struct rb_irq_work
*work
;
562 * Depending on what the caller is waiting for, either any
563 * data in any cpu buffer, or a specific buffer, put the
564 * caller on the appropriate wait queue.
566 if (cpu
== RING_BUFFER_ALL_CPUS
)
567 work
= &buffer
->irq_work
;
569 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
571 cpu_buffer
= buffer
->buffers
[cpu
];
572 work
= &cpu_buffer
->irq_work
;
576 prepare_to_wait(&work
->waiters
, &wait
, TASK_INTERRUPTIBLE
);
579 * The events can happen in critical sections where
580 * checking a work queue can cause deadlocks.
581 * After adding a task to the queue, this flag is set
582 * only to notify events to try to wake up the queue
585 * We don't clear it even if the buffer is no longer
586 * empty. The flag only causes the next event to run
587 * irq_work to do the work queue wake up. The worse
588 * that can happen if we race with !trace_empty() is that
589 * an event will cause an irq_work to try to wake up
592 * There's no reason to protect this flag either, as
593 * the work queue and irq_work logic will do the necessary
594 * synchronization for the wake ups. The only thing
595 * that is necessary is that the wake up happens after
596 * a task has been queued. It's OK for spurious wake ups.
598 work
->waiters_pending
= true;
600 if ((cpu
== RING_BUFFER_ALL_CPUS
&& ring_buffer_empty(buffer
)) ||
601 (cpu
!= RING_BUFFER_ALL_CPUS
&& ring_buffer_empty_cpu(buffer
, cpu
)))
604 finish_wait(&work
->waiters
, &wait
);
609 * ring_buffer_poll_wait - poll on buffer input
610 * @buffer: buffer to wait on
611 * @cpu: the cpu buffer to wait on
612 * @filp: the file descriptor
613 * @poll_table: The poll descriptor
615 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
616 * as data is added to any of the @buffer's cpu buffers. Otherwise
617 * it will wait for data to be added to a specific cpu buffer.
619 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
622 int ring_buffer_poll_wait(struct ring_buffer
*buffer
, int cpu
,
623 struct file
*filp
, poll_table
*poll_table
)
625 struct ring_buffer_per_cpu
*cpu_buffer
;
626 struct rb_irq_work
*work
;
628 if (cpu
== RING_BUFFER_ALL_CPUS
)
629 work
= &buffer
->irq_work
;
631 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
634 cpu_buffer
= buffer
->buffers
[cpu
];
635 work
= &cpu_buffer
->irq_work
;
638 poll_wait(filp
, &work
->waiters
, poll_table
);
639 work
->waiters_pending
= true;
641 * There's a tight race between setting the waiters_pending and
642 * checking if the ring buffer is empty. Once the waiters_pending bit
643 * is set, the next event will wake the task up, but we can get stuck
644 * if there's only a single event in.
646 * FIXME: Ideally, we need a memory barrier on the writer side as well,
647 * but adding a memory barrier to all events will cause too much of a
648 * performance hit in the fast path. We only need a memory barrier when
649 * the buffer goes from empty to having content. But as this race is
650 * extremely small, and it's not a problem if another event comes in, we
655 if ((cpu
== RING_BUFFER_ALL_CPUS
&& !ring_buffer_empty(buffer
)) ||
656 (cpu
!= RING_BUFFER_ALL_CPUS
&& !ring_buffer_empty_cpu(buffer
, cpu
)))
657 return POLLIN
| POLLRDNORM
;
661 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
662 #define RB_WARN_ON(b, cond) \
664 int _____ret = unlikely(cond); \
666 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
667 struct ring_buffer_per_cpu *__b = \
669 atomic_inc(&__b->buffer->record_disabled); \
671 atomic_inc(&b->record_disabled); \
677 /* Up this if you want to test the TIME_EXTENTS and normalization */
678 #define DEBUG_SHIFT 0
680 static inline u64
rb_time_stamp(struct ring_buffer
*buffer
)
682 /* shift to debug/test normalization and TIME_EXTENTS */
683 return buffer
->clock() << DEBUG_SHIFT
;
686 u64
ring_buffer_time_stamp(struct ring_buffer
*buffer
, int cpu
)
690 preempt_disable_notrace();
691 time
= rb_time_stamp(buffer
);
692 preempt_enable_no_resched_notrace();
696 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp
);
698 void ring_buffer_normalize_time_stamp(struct ring_buffer
*buffer
,
701 /* Just stupid testing the normalize function and deltas */
704 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp
);
707 * Making the ring buffer lockless makes things tricky.
708 * Although writes only happen on the CPU that they are on,
709 * and they only need to worry about interrupts. Reads can
712 * The reader page is always off the ring buffer, but when the
713 * reader finishes with a page, it needs to swap its page with
714 * a new one from the buffer. The reader needs to take from
715 * the head (writes go to the tail). But if a writer is in overwrite
716 * mode and wraps, it must push the head page forward.
718 * Here lies the problem.
720 * The reader must be careful to replace only the head page, and
721 * not another one. As described at the top of the file in the
722 * ASCII art, the reader sets its old page to point to the next
723 * page after head. It then sets the page after head to point to
724 * the old reader page. But if the writer moves the head page
725 * during this operation, the reader could end up with the tail.
727 * We use cmpxchg to help prevent this race. We also do something
728 * special with the page before head. We set the LSB to 1.
730 * When the writer must push the page forward, it will clear the
731 * bit that points to the head page, move the head, and then set
732 * the bit that points to the new head page.
734 * We also don't want an interrupt coming in and moving the head
735 * page on another writer. Thus we use the second LSB to catch
738 * head->list->prev->next bit 1 bit 0
741 * Points to head page 0 1
744 * Note we can not trust the prev pointer of the head page, because:
746 * +----+ +-----+ +-----+
747 * | |------>| T |---X--->| N |
749 * +----+ +-----+ +-----+
752 * +----------| R |----------+ |
756 * Key: ---X--> HEAD flag set in pointer
761 * (see __rb_reserve_next() to see where this happens)
763 * What the above shows is that the reader just swapped out
764 * the reader page with a page in the buffer, but before it
765 * could make the new header point back to the new page added
766 * it was preempted by a writer. The writer moved forward onto
767 * the new page added by the reader and is about to move forward
770 * You can see, it is legitimate for the previous pointer of
771 * the head (or any page) not to point back to itself. But only
775 #define RB_PAGE_NORMAL 0UL
776 #define RB_PAGE_HEAD 1UL
777 #define RB_PAGE_UPDATE 2UL
780 #define RB_FLAG_MASK 3UL
782 /* PAGE_MOVED is not part of the mask */
783 #define RB_PAGE_MOVED 4UL
786 * rb_list_head - remove any bit
788 static struct list_head
*rb_list_head(struct list_head
*list
)
790 unsigned long val
= (unsigned long)list
;
792 return (struct list_head
*)(val
& ~RB_FLAG_MASK
);
796 * rb_is_head_page - test if the given page is the head page
798 * Because the reader may move the head_page pointer, we can
799 * not trust what the head page is (it may be pointing to
800 * the reader page). But if the next page is a header page,
801 * its flags will be non zero.
804 rb_is_head_page(struct ring_buffer_per_cpu
*cpu_buffer
,
805 struct buffer_page
*page
, struct list_head
*list
)
809 val
= (unsigned long)list
->next
;
811 if ((val
& ~RB_FLAG_MASK
) != (unsigned long)&page
->list
)
812 return RB_PAGE_MOVED
;
814 return val
& RB_FLAG_MASK
;
820 * The unique thing about the reader page, is that, if the
821 * writer is ever on it, the previous pointer never points
822 * back to the reader page.
824 static int rb_is_reader_page(struct buffer_page
*page
)
826 struct list_head
*list
= page
->list
.prev
;
828 return rb_list_head(list
->next
) != &page
->list
;
832 * rb_set_list_to_head - set a list_head to be pointing to head.
834 static void rb_set_list_to_head(struct ring_buffer_per_cpu
*cpu_buffer
,
835 struct list_head
*list
)
839 ptr
= (unsigned long *)&list
->next
;
840 *ptr
|= RB_PAGE_HEAD
;
841 *ptr
&= ~RB_PAGE_UPDATE
;
845 * rb_head_page_activate - sets up head page
847 static void rb_head_page_activate(struct ring_buffer_per_cpu
*cpu_buffer
)
849 struct buffer_page
*head
;
851 head
= cpu_buffer
->head_page
;
856 * Set the previous list pointer to have the HEAD flag.
858 rb_set_list_to_head(cpu_buffer
, head
->list
.prev
);
861 static void rb_list_head_clear(struct list_head
*list
)
863 unsigned long *ptr
= (unsigned long *)&list
->next
;
865 *ptr
&= ~RB_FLAG_MASK
;
869 * rb_head_page_dactivate - clears head page ptr (for free list)
872 rb_head_page_deactivate(struct ring_buffer_per_cpu
*cpu_buffer
)
874 struct list_head
*hd
;
876 /* Go through the whole list and clear any pointers found. */
877 rb_list_head_clear(cpu_buffer
->pages
);
879 list_for_each(hd
, cpu_buffer
->pages
)
880 rb_list_head_clear(hd
);
883 static int rb_head_page_set(struct ring_buffer_per_cpu
*cpu_buffer
,
884 struct buffer_page
*head
,
885 struct buffer_page
*prev
,
886 int old_flag
, int new_flag
)
888 struct list_head
*list
;
889 unsigned long val
= (unsigned long)&head
->list
;
894 val
&= ~RB_FLAG_MASK
;
896 ret
= cmpxchg((unsigned long *)&list
->next
,
897 val
| old_flag
, val
| new_flag
);
899 /* check if the reader took the page */
900 if ((ret
& ~RB_FLAG_MASK
) != val
)
901 return RB_PAGE_MOVED
;
903 return ret
& RB_FLAG_MASK
;
906 static int rb_head_page_set_update(struct ring_buffer_per_cpu
*cpu_buffer
,
907 struct buffer_page
*head
,
908 struct buffer_page
*prev
,
911 return rb_head_page_set(cpu_buffer
, head
, prev
,
912 old_flag
, RB_PAGE_UPDATE
);
915 static int rb_head_page_set_head(struct ring_buffer_per_cpu
*cpu_buffer
,
916 struct buffer_page
*head
,
917 struct buffer_page
*prev
,
920 return rb_head_page_set(cpu_buffer
, head
, prev
,
921 old_flag
, RB_PAGE_HEAD
);
924 static int rb_head_page_set_normal(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_NORMAL
);
933 static inline void rb_inc_page(struct ring_buffer_per_cpu
*cpu_buffer
,
934 struct buffer_page
**bpage
)
936 struct list_head
*p
= rb_list_head((*bpage
)->list
.next
);
938 *bpage
= list_entry(p
, struct buffer_page
, list
);
941 static struct buffer_page
*
942 rb_set_head_page(struct ring_buffer_per_cpu
*cpu_buffer
)
944 struct buffer_page
*head
;
945 struct buffer_page
*page
;
946 struct list_head
*list
;
949 if (RB_WARN_ON(cpu_buffer
, !cpu_buffer
->head_page
))
953 list
= cpu_buffer
->pages
;
954 if (RB_WARN_ON(cpu_buffer
, rb_list_head(list
->prev
->next
) != list
))
957 page
= head
= cpu_buffer
->head_page
;
959 * It is possible that the writer moves the header behind
960 * where we started, and we miss in one loop.
961 * A second loop should grab the header, but we'll do
962 * three loops just because I'm paranoid.
964 for (i
= 0; i
< 3; i
++) {
966 if (rb_is_head_page(cpu_buffer
, page
, page
->list
.prev
)) {
967 cpu_buffer
->head_page
= page
;
970 rb_inc_page(cpu_buffer
, &page
);
971 } while (page
!= head
);
974 RB_WARN_ON(cpu_buffer
, 1);
979 static int rb_head_page_replace(struct buffer_page
*old
,
980 struct buffer_page
*new)
982 unsigned long *ptr
= (unsigned long *)&old
->list
.prev
->next
;
986 val
= *ptr
& ~RB_FLAG_MASK
;
989 ret
= cmpxchg(ptr
, val
, (unsigned long)&new->list
);
995 * rb_tail_page_update - move the tail page forward
997 * Returns 1 if moved tail page, 0 if someone else did.
999 static int rb_tail_page_update(struct ring_buffer_per_cpu
*cpu_buffer
,
1000 struct buffer_page
*tail_page
,
1001 struct buffer_page
*next_page
)
1003 struct buffer_page
*old_tail
;
1004 unsigned long old_entries
;
1005 unsigned long old_write
;
1009 * The tail page now needs to be moved forward.
1011 * We need to reset the tail page, but without messing
1012 * with possible erasing of data brought in by interrupts
1013 * that have moved the tail page and are currently on it.
1015 * We add a counter to the write field to denote this.
1017 old_write
= local_add_return(RB_WRITE_INTCNT
, &next_page
->write
);
1018 old_entries
= local_add_return(RB_WRITE_INTCNT
, &next_page
->entries
);
1021 * Just make sure we have seen our old_write and synchronize
1022 * with any interrupts that come in.
1027 * If the tail page is still the same as what we think
1028 * it is, then it is up to us to update the tail
1031 if (tail_page
== cpu_buffer
->tail_page
) {
1032 /* Zero the write counter */
1033 unsigned long val
= old_write
& ~RB_WRITE_MASK
;
1034 unsigned long eval
= old_entries
& ~RB_WRITE_MASK
;
1037 * This will only succeed if an interrupt did
1038 * not come in and change it. In which case, we
1039 * do not want to modify it.
1041 * We add (void) to let the compiler know that we do not care
1042 * about the return value of these functions. We use the
1043 * cmpxchg to only update if an interrupt did not already
1044 * do it for us. If the cmpxchg fails, we don't care.
1046 (void)local_cmpxchg(&next_page
->write
, old_write
, val
);
1047 (void)local_cmpxchg(&next_page
->entries
, old_entries
, eval
);
1050 * No need to worry about races with clearing out the commit.
1051 * it only can increment when a commit takes place. But that
1052 * only happens in the outer most nested commit.
1054 local_set(&next_page
->page
->commit
, 0);
1056 old_tail
= cmpxchg(&cpu_buffer
->tail_page
,
1057 tail_page
, next_page
);
1059 if (old_tail
== tail_page
)
1066 static int rb_check_bpage(struct ring_buffer_per_cpu
*cpu_buffer
,
1067 struct buffer_page
*bpage
)
1069 unsigned long val
= (unsigned long)bpage
;
1071 if (RB_WARN_ON(cpu_buffer
, val
& RB_FLAG_MASK
))
1078 * rb_check_list - make sure a pointer to a list has the last bits zero
1080 static int rb_check_list(struct ring_buffer_per_cpu
*cpu_buffer
,
1081 struct list_head
*list
)
1083 if (RB_WARN_ON(cpu_buffer
, rb_list_head(list
->prev
) != list
->prev
))
1085 if (RB_WARN_ON(cpu_buffer
, rb_list_head(list
->next
) != list
->next
))
1091 * check_pages - integrity check of buffer pages
1092 * @cpu_buffer: CPU buffer with pages to test
1094 * As a safety measure we check to make sure the data pages have not
1097 static int rb_check_pages(struct ring_buffer_per_cpu
*cpu_buffer
)
1099 struct list_head
*head
= cpu_buffer
->pages
;
1100 struct buffer_page
*bpage
, *tmp
;
1102 /* Reset the head page if it exists */
1103 if (cpu_buffer
->head_page
)
1104 rb_set_head_page(cpu_buffer
);
1106 rb_head_page_deactivate(cpu_buffer
);
1108 if (RB_WARN_ON(cpu_buffer
, head
->next
->prev
!= head
))
1110 if (RB_WARN_ON(cpu_buffer
, head
->prev
->next
!= head
))
1113 if (rb_check_list(cpu_buffer
, head
))
1116 list_for_each_entry_safe(bpage
, tmp
, head
, list
) {
1117 if (RB_WARN_ON(cpu_buffer
,
1118 bpage
->list
.next
->prev
!= &bpage
->list
))
1120 if (RB_WARN_ON(cpu_buffer
,
1121 bpage
->list
.prev
->next
!= &bpage
->list
))
1123 if (rb_check_list(cpu_buffer
, &bpage
->list
))
1127 rb_head_page_activate(cpu_buffer
);
1132 static int __rb_allocate_pages(long nr_pages
, struct list_head
*pages
, int cpu
)
1134 struct buffer_page
*bpage
, *tmp
;
1137 for (i
= 0; i
< nr_pages
; i
++) {
1138 #if !defined (CONFIG_MTK_EXTMEM)
1142 * __GFP_NORETRY flag makes sure that the allocation fails
1143 * gracefully without invoking oom-killer and the system is
1146 bpage
= kzalloc_node(ALIGN(sizeof(*bpage
), cache_line_size()),
1147 GFP_KERNEL
| __GFP_NORETRY
,
1152 list_add(&bpage
->list
, pages
);
1154 #ifdef CONFIG_MTK_EXTMEM
1155 bpage
->page
= extmem_malloc_page_align(PAGE_SIZE
);
1156 if(bpage
->page
== NULL
) {
1157 pr_err("%s[%s] ext memory alloc failed!!!\n", __FILE__
, __FUNCTION__
);
1161 page
= alloc_pages_node(cpu_to_node(cpu
),
1162 GFP_KERNEL
| __GFP_NORETRY
, 0);
1165 bpage
->page
= page_address(page
);
1167 rb_init_page(bpage
->page
);
1173 list_for_each_entry_safe(bpage
, tmp
, pages
, list
) {
1174 list_del_init(&bpage
->list
);
1175 free_buffer_page(bpage
);
1181 static int rb_allocate_pages(struct ring_buffer_per_cpu
*cpu_buffer
,
1182 unsigned long nr_pages
)
1188 if (__rb_allocate_pages(nr_pages
, &pages
, cpu_buffer
->cpu
))
1192 * The ring buffer page list is a circular list that does not
1193 * start and end with a list head. All page list items point to
1196 cpu_buffer
->pages
= pages
.next
;
1199 cpu_buffer
->nr_pages
= nr_pages
;
1201 rb_check_pages(cpu_buffer
);
1206 static struct ring_buffer_per_cpu
*
1207 rb_allocate_cpu_buffer(struct ring_buffer
*buffer
, long nr_pages
, int cpu
)
1209 struct ring_buffer_per_cpu
*cpu_buffer
;
1210 struct buffer_page
*bpage
;
1211 #if !defined (CONFIG_MTK_EXTMEM)
1216 cpu_buffer
= kzalloc_node(ALIGN(sizeof(*cpu_buffer
), cache_line_size()),
1217 GFP_KERNEL
, cpu_to_node(cpu
));
1221 cpu_buffer
->cpu
= cpu
;
1222 cpu_buffer
->buffer
= buffer
;
1223 raw_spin_lock_init(&cpu_buffer
->reader_lock
);
1224 lockdep_set_class(&cpu_buffer
->reader_lock
, buffer
->reader_lock_key
);
1225 cpu_buffer
->lock
= (arch_spinlock_t
)__ARCH_SPIN_LOCK_UNLOCKED
;
1226 INIT_WORK(&cpu_buffer
->update_pages_work
, update_pages_handler
);
1227 init_completion(&cpu_buffer
->update_done
);
1228 init_irq_work(&cpu_buffer
->irq_work
.work
, rb_wake_up_waiters
);
1229 init_waitqueue_head(&cpu_buffer
->irq_work
.waiters
);
1231 bpage
= kzalloc_node(ALIGN(sizeof(*bpage
), cache_line_size()),
1232 GFP_KERNEL
, cpu_to_node(cpu
));
1234 goto fail_free_buffer
;
1236 rb_check_bpage(cpu_buffer
, bpage
);
1238 cpu_buffer
->reader_page
= bpage
;
1240 #ifdef CONFIG_MTK_EXTMEM
1241 bpage
->page
= extmem_malloc_page_align(PAGE_SIZE
);
1242 if(bpage
->page
== NULL
)
1243 goto fail_free_reader
;
1245 page
= alloc_pages_node(cpu_to_node(cpu
), GFP_KERNEL
, 0);
1247 goto fail_free_reader
;
1248 bpage
->page
= page_address(page
);
1250 rb_init_page(bpage
->page
);
1252 INIT_LIST_HEAD(&cpu_buffer
->reader_page
->list
);
1253 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
1255 ret
= rb_allocate_pages(cpu_buffer
, nr_pages
);
1257 goto fail_free_reader
;
1259 cpu_buffer
->head_page
1260 = list_entry(cpu_buffer
->pages
, struct buffer_page
, list
);
1261 cpu_buffer
->tail_page
= cpu_buffer
->commit_page
= cpu_buffer
->head_page
;
1263 rb_head_page_activate(cpu_buffer
);
1268 free_buffer_page(cpu_buffer
->reader_page
);
1275 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu
*cpu_buffer
)
1277 struct list_head
*head
= cpu_buffer
->pages
;
1278 struct buffer_page
*bpage
, *tmp
;
1280 free_buffer_page(cpu_buffer
->reader_page
);
1282 rb_head_page_deactivate(cpu_buffer
);
1285 list_for_each_entry_safe(bpage
, tmp
, head
, list
) {
1286 list_del_init(&bpage
->list
);
1287 free_buffer_page(bpage
);
1289 bpage
= list_entry(head
, struct buffer_page
, list
);
1290 free_buffer_page(bpage
);
1296 #ifdef CONFIG_HOTPLUG_CPU
1297 static int rb_cpu_notify(struct notifier_block
*self
,
1298 unsigned long action
, void *hcpu
);
1302 * ring_buffer_alloc - allocate a new ring_buffer
1303 * @size: the size in bytes per cpu that is needed.
1304 * @flags: attributes to set for the ring buffer.
1306 * Currently the only flag that is available is the RB_FL_OVERWRITE
1307 * flag. This flag means that the buffer will overwrite old data
1308 * when the buffer wraps. If this flag is not set, the buffer will
1309 * drop data when the tail hits the head.
1311 struct ring_buffer
*__ring_buffer_alloc(unsigned long size
, unsigned flags
,
1312 struct lock_class_key
*key
)
1314 struct ring_buffer
*buffer
;
1319 /* keep it in its own cache line */
1320 buffer
= kzalloc(ALIGN(sizeof(*buffer
), cache_line_size()),
1325 if (!alloc_cpumask_var(&buffer
->cpumask
, GFP_KERNEL
))
1326 goto fail_free_buffer
;
1328 nr_pages
= DIV_ROUND_UP(size
, BUF_PAGE_SIZE
);
1329 buffer
->flags
= flags
;
1330 buffer
->clock
= trace_clock_local
;
1331 buffer
->reader_lock_key
= key
;
1333 init_irq_work(&buffer
->irq_work
.work
, rb_wake_up_waiters
);
1334 init_waitqueue_head(&buffer
->irq_work
.waiters
);
1336 /* need at least two pages */
1341 * In case of non-hotplug cpu, if the ring-buffer is allocated
1342 * in early initcall, it will not be notified of secondary cpus.
1343 * In that off case, we need to allocate for all possible cpus.
1345 #ifdef CONFIG_HOTPLUG_CPU
1347 cpumask_copy(buffer
->cpumask
, cpu_online_mask
);
1349 cpumask_copy(buffer
->cpumask
, cpu_possible_mask
);
1351 buffer
->cpus
= nr_cpu_ids
;
1353 bsize
= sizeof(void *) * nr_cpu_ids
;
1354 buffer
->buffers
= kzalloc(ALIGN(bsize
, cache_line_size()),
1356 if (!buffer
->buffers
)
1357 goto fail_free_cpumask
;
1359 for_each_buffer_cpu(buffer
, cpu
) {
1360 buffer
->buffers
[cpu
] =
1361 rb_allocate_cpu_buffer(buffer
, nr_pages
, cpu
);
1362 if (!buffer
->buffers
[cpu
])
1363 goto fail_free_buffers
;
1366 #ifdef CONFIG_HOTPLUG_CPU
1367 buffer
->cpu_notify
.notifier_call
= rb_cpu_notify
;
1368 buffer
->cpu_notify
.priority
= 0;
1369 register_cpu_notifier(&buffer
->cpu_notify
);
1373 mutex_init(&buffer
->mutex
);
1378 for_each_buffer_cpu(buffer
, cpu
) {
1379 if (buffer
->buffers
[cpu
])
1380 rb_free_cpu_buffer(buffer
->buffers
[cpu
]);
1382 kfree(buffer
->buffers
);
1385 free_cpumask_var(buffer
->cpumask
);
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
)
1405 #ifdef CONFIG_HOTPLUG_CPU
1406 unregister_cpu_notifier(&buffer
->cpu_notify
);
1409 for_each_buffer_cpu(buffer
, cpu
)
1410 rb_free_cpu_buffer(buffer
->buffers
[cpu
]);
1414 kfree(buffer
->buffers
);
1415 free_cpumask_var(buffer
->cpumask
);
1419 EXPORT_SYMBOL_GPL(ring_buffer_free
);
1421 void ring_buffer_set_clock(struct ring_buffer
*buffer
,
1424 buffer
->clock
= clock
;
1427 static void rb_reset_cpu(struct ring_buffer_per_cpu
*cpu_buffer
);
1429 static inline unsigned long rb_page_entries(struct buffer_page
*bpage
)
1431 return local_read(&bpage
->entries
) & RB_WRITE_MASK
;
1434 static inline unsigned long rb_page_write(struct buffer_page
*bpage
)
1436 return local_read(&bpage
->write
) & RB_WRITE_MASK
;
1440 rb_remove_pages(struct ring_buffer_per_cpu
*cpu_buffer
, unsigned long nr_pages
)
1442 struct list_head
*tail_page
, *to_remove
, *next_page
;
1443 struct buffer_page
*to_remove_page
, *tmp_iter_page
;
1444 struct buffer_page
*last_page
, *first_page
;
1445 unsigned long nr_removed
;
1446 unsigned long head_bit
;
1451 raw_spin_lock_irq(&cpu_buffer
->reader_lock
);
1452 atomic_inc(&cpu_buffer
->record_disabled
);
1454 * We don't race with the readers since we have acquired the reader
1455 * lock. We also don't race with writers after disabling recording.
1456 * This makes it easy to figure out the first and the last page to be
1457 * removed from the list. We unlink all the pages in between including
1458 * the first and last pages. This is done in a busy loop so that we
1459 * lose the least number of traces.
1460 * The pages are freed after we restart recording and unlock readers.
1462 tail_page
= &cpu_buffer
->tail_page
->list
;
1465 * tail page might be on reader page, we remove the next page
1466 * from the ring buffer
1468 if (cpu_buffer
->tail_page
== cpu_buffer
->reader_page
)
1469 tail_page
= rb_list_head(tail_page
->next
);
1470 to_remove
= tail_page
;
1472 /* start of pages to remove */
1473 first_page
= list_entry(rb_list_head(to_remove
->next
),
1474 struct buffer_page
, list
);
1476 for (nr_removed
= 0; nr_removed
< nr_pages
; nr_removed
++) {
1477 to_remove
= rb_list_head(to_remove
)->next
;
1478 head_bit
|= (unsigned long)to_remove
& RB_PAGE_HEAD
;
1481 next_page
= rb_list_head(to_remove
)->next
;
1484 * Now we remove all pages between tail_page and next_page.
1485 * Make sure that we have head_bit value preserved for the
1488 tail_page
->next
= (struct list_head
*)((unsigned long)next_page
|
1490 next_page
= rb_list_head(next_page
);
1491 next_page
->prev
= tail_page
;
1493 /* make sure pages points to a valid page in the ring buffer */
1494 cpu_buffer
->pages
= next_page
;
1496 /* update head page */
1498 cpu_buffer
->head_page
= list_entry(next_page
,
1499 struct buffer_page
, list
);
1502 * change read pointer to make sure any read iterators reset
1505 cpu_buffer
->read
= 0;
1507 /* pages are removed, resume tracing and then free the pages */
1508 atomic_dec(&cpu_buffer
->record_disabled
);
1509 raw_spin_unlock_irq(&cpu_buffer
->reader_lock
);
1511 RB_WARN_ON(cpu_buffer
, list_empty(cpu_buffer
->pages
));
1513 /* last buffer page to remove */
1514 last_page
= list_entry(rb_list_head(to_remove
), struct buffer_page
,
1516 tmp_iter_page
= first_page
;
1519 to_remove_page
= tmp_iter_page
;
1520 rb_inc_page(cpu_buffer
, &tmp_iter_page
);
1522 /* update the counters */
1523 page_entries
= rb_page_entries(to_remove_page
);
1526 * If something was added to this page, it was full
1527 * since it is not the tail page. So we deduct the
1528 * bytes consumed in ring buffer from here.
1529 * Increment overrun to account for the lost events.
1531 local_add(page_entries
, &cpu_buffer
->overrun
);
1532 local_sub(BUF_PAGE_SIZE
, &cpu_buffer
->entries_bytes
);
1536 * We have already removed references to this list item, just
1537 * free up the buffer_page and its page
1539 free_buffer_page(to_remove_page
);
1542 } while (to_remove_page
!= last_page
);
1544 RB_WARN_ON(cpu_buffer
, nr_removed
);
1546 return nr_removed
== 0;
1550 rb_insert_pages(struct ring_buffer_per_cpu
*cpu_buffer
)
1552 struct list_head
*pages
= &cpu_buffer
->new_pages
;
1553 int retries
, success
;
1555 raw_spin_lock_irq(&cpu_buffer
->reader_lock
);
1557 * We are holding the reader lock, so the reader page won't be swapped
1558 * in the ring buffer. Now we are racing with the writer trying to
1559 * move head page and the tail page.
1560 * We are going to adapt the reader page update process where:
1561 * 1. We first splice the start and end of list of new pages between
1562 * the head page and its previous page.
1563 * 2. We cmpxchg the prev_page->next to point from head page to the
1564 * start of new pages list.
1565 * 3. Finally, we update the head->prev to the end of new list.
1567 * We will try this process 10 times, to make sure that we don't keep
1573 struct list_head
*head_page
, *prev_page
, *r
;
1574 struct list_head
*last_page
, *first_page
;
1575 struct list_head
*head_page_with_bit
;
1577 head_page
= &rb_set_head_page(cpu_buffer
)->list
;
1580 prev_page
= head_page
->prev
;
1582 first_page
= pages
->next
;
1583 last_page
= pages
->prev
;
1585 head_page_with_bit
= (struct list_head
*)
1586 ((unsigned long)head_page
| RB_PAGE_HEAD
);
1588 last_page
->next
= head_page_with_bit
;
1589 first_page
->prev
= prev_page
;
1591 r
= cmpxchg(&prev_page
->next
, head_page_with_bit
, first_page
);
1593 if (r
== head_page_with_bit
) {
1595 * yay, we replaced the page pointer to our new list,
1596 * now, we just have to update to head page's prev
1597 * pointer to point to end of list
1599 head_page
->prev
= last_page
;
1606 INIT_LIST_HEAD(pages
);
1608 * If we weren't successful in adding in new pages, warn and stop
1611 RB_WARN_ON(cpu_buffer
, !success
);
1612 raw_spin_unlock_irq(&cpu_buffer
->reader_lock
);
1614 /* free pages if they weren't inserted */
1616 struct buffer_page
*bpage
, *tmp
;
1617 list_for_each_entry_safe(bpage
, tmp
, &cpu_buffer
->new_pages
,
1619 list_del_init(&bpage
->list
);
1620 free_buffer_page(bpage
);
1626 static void rb_update_pages(struct ring_buffer_per_cpu
*cpu_buffer
)
1630 if (cpu_buffer
->nr_pages_to_update
> 0)
1631 success
= rb_insert_pages(cpu_buffer
);
1633 success
= rb_remove_pages(cpu_buffer
,
1634 -cpu_buffer
->nr_pages_to_update
);
1637 cpu_buffer
->nr_pages
+= cpu_buffer
->nr_pages_to_update
;
1640 static void update_pages_handler(struct work_struct
*work
)
1642 struct ring_buffer_per_cpu
*cpu_buffer
= container_of(work
,
1643 struct ring_buffer_per_cpu
, update_pages_work
);
1644 rb_update_pages(cpu_buffer
);
1645 complete(&cpu_buffer
->update_done
);
1649 * ring_buffer_resize - resize the ring buffer
1650 * @buffer: the buffer to resize.
1651 * @size: the new size.
1653 * Minimum size is 2 * BUF_PAGE_SIZE.
1655 * Returns 0 on success and < 0 on failure.
1657 int ring_buffer_resize(struct ring_buffer
*buffer
, unsigned long size
,
1660 struct ring_buffer_per_cpu
*cpu_buffer
;
1661 unsigned long nr_pages
;
1665 * Always succeed at resizing a non-existent buffer:
1670 /* Make sure the requested buffer exists */
1671 if (cpu_id
!= RING_BUFFER_ALL_CPUS
&&
1672 !cpumask_test_cpu(cpu_id
, buffer
->cpumask
))
1675 nr_pages
= DIV_ROUND_UP(size
, BUF_PAGE_SIZE
);
1677 /* we need a minimum of two pages */
1681 size
= nr_pages
* BUF_PAGE_SIZE
;
1684 * Don't succeed if resizing is disabled, as a reader might be
1685 * manipulating the ring buffer and is expecting a sane state while
1688 if (atomic_read(&buffer
->resize_disabled
))
1691 /* prevent another thread from changing buffer sizes */
1692 mutex_lock(&buffer
->mutex
);
1694 if (cpu_id
== RING_BUFFER_ALL_CPUS
) {
1695 /* calculate the pages to update */
1696 for_each_buffer_cpu(buffer
, cpu
) {
1697 cpu_buffer
= buffer
->buffers
[cpu
];
1699 cpu_buffer
->nr_pages_to_update
= nr_pages
-
1700 cpu_buffer
->nr_pages
;
1702 * nothing more to do for removing pages or no update
1704 if (cpu_buffer
->nr_pages_to_update
<= 0)
1707 * to add pages, make sure all new pages can be
1708 * allocated without receiving ENOMEM
1710 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
1711 if (__rb_allocate_pages(cpu_buffer
->nr_pages_to_update
,
1712 &cpu_buffer
->new_pages
, cpu
)) {
1713 /* not enough memory for new pages */
1721 * Fire off all the required work handlers
1722 * We can't schedule on offline CPUs, but it's not necessary
1723 * since we can change their buffer sizes without any race.
1725 for_each_buffer_cpu(buffer
, cpu
) {
1726 cpu_buffer
= buffer
->buffers
[cpu
];
1727 if (!cpu_buffer
->nr_pages_to_update
)
1730 /* The update must run on the CPU that is being updated. */
1732 if (cpu
== smp_processor_id() || !cpu_online(cpu
)) {
1733 rb_update_pages(cpu_buffer
);
1734 cpu_buffer
->nr_pages_to_update
= 0;
1737 * Can not disable preemption for schedule_work_on()
1741 schedule_work_on(cpu
,
1742 &cpu_buffer
->update_pages_work
);
1748 /* wait for all the updates to complete */
1749 for_each_buffer_cpu(buffer
, cpu
) {
1750 cpu_buffer
= buffer
->buffers
[cpu
];
1751 if (!cpu_buffer
->nr_pages_to_update
)
1754 if (cpu_online(cpu
))
1755 wait_for_completion(&cpu_buffer
->update_done
);
1756 cpu_buffer
->nr_pages_to_update
= 0;
1761 /* Make sure this CPU has been intitialized */
1762 if (!cpumask_test_cpu(cpu_id
, buffer
->cpumask
))
1765 cpu_buffer
= buffer
->buffers
[cpu_id
];
1767 if (nr_pages
== cpu_buffer
->nr_pages
)
1770 cpu_buffer
->nr_pages_to_update
= nr_pages
-
1771 cpu_buffer
->nr_pages
;
1773 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
1774 if (cpu_buffer
->nr_pages_to_update
> 0 &&
1775 __rb_allocate_pages(cpu_buffer
->nr_pages_to_update
,
1776 &cpu_buffer
->new_pages
, cpu_id
)) {
1784 /* The update must run on the CPU that is being updated. */
1785 if (cpu_id
== smp_processor_id() || !cpu_online(cpu_id
))
1786 rb_update_pages(cpu_buffer
);
1789 * Can not disable preemption for schedule_work_on()
1793 schedule_work_on(cpu_id
,
1794 &cpu_buffer
->update_pages_work
);
1795 wait_for_completion(&cpu_buffer
->update_done
);
1800 cpu_buffer
->nr_pages_to_update
= 0;
1806 * The ring buffer resize can happen with the ring buffer
1807 * enabled, so that the update disturbs the tracing as little
1808 * as possible. But if the buffer is disabled, we do not need
1809 * to worry about that, and we can take the time to verify
1810 * that the buffer is not corrupt.
1812 if (atomic_read(&buffer
->record_disabled
)) {
1813 atomic_inc(&buffer
->record_disabled
);
1815 * Even though the buffer was disabled, we must make sure
1816 * that it is truly disabled before calling rb_check_pages.
1817 * There could have been a race between checking
1818 * record_disable and incrementing it.
1820 synchronize_sched();
1821 for_each_buffer_cpu(buffer
, cpu
) {
1822 cpu_buffer
= buffer
->buffers
[cpu
];
1823 rb_check_pages(cpu_buffer
);
1825 atomic_dec(&buffer
->record_disabled
);
1828 mutex_unlock(&buffer
->mutex
);
1832 for_each_buffer_cpu(buffer
, cpu
) {
1833 struct buffer_page
*bpage
, *tmp
;
1835 cpu_buffer
= buffer
->buffers
[cpu
];
1836 cpu_buffer
->nr_pages_to_update
= 0;
1838 if (list_empty(&cpu_buffer
->new_pages
))
1841 list_for_each_entry_safe(bpage
, tmp
, &cpu_buffer
->new_pages
,
1843 list_del_init(&bpage
->list
);
1844 free_buffer_page(bpage
);
1847 mutex_unlock(&buffer
->mutex
);
1850 EXPORT_SYMBOL_GPL(ring_buffer_resize
);
1852 void ring_buffer_change_overwrite(struct ring_buffer
*buffer
, int val
)
1854 mutex_lock(&buffer
->mutex
);
1856 buffer
->flags
|= RB_FL_OVERWRITE
;
1858 buffer
->flags
&= ~RB_FL_OVERWRITE
;
1859 mutex_unlock(&buffer
->mutex
);
1861 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite
);
1863 static inline void *
1864 __rb_data_page_index(struct buffer_data_page
*bpage
, unsigned index
)
1866 return bpage
->data
+ index
;
1869 static inline void *__rb_page_index(struct buffer_page
*bpage
, unsigned index
)
1871 return bpage
->page
->data
+ index
;
1874 static inline struct ring_buffer_event
*
1875 rb_reader_event(struct ring_buffer_per_cpu
*cpu_buffer
)
1877 return __rb_page_index(cpu_buffer
->reader_page
,
1878 cpu_buffer
->reader_page
->read
);
1881 static inline struct ring_buffer_event
*
1882 rb_iter_head_event(struct ring_buffer_iter
*iter
)
1884 return __rb_page_index(iter
->head_page
, iter
->head
);
1887 static inline unsigned rb_page_commit(struct buffer_page
*bpage
)
1889 return local_read(&bpage
->page
->commit
);
1892 /* Size is determined by what has been committed */
1893 static inline unsigned rb_page_size(struct buffer_page
*bpage
)
1895 return rb_page_commit(bpage
);
1898 static inline unsigned
1899 rb_commit_index(struct ring_buffer_per_cpu
*cpu_buffer
)
1901 return rb_page_commit(cpu_buffer
->commit_page
);
1904 static inline unsigned
1905 rb_event_index(struct ring_buffer_event
*event
)
1907 unsigned long addr
= (unsigned long)event
;
1909 return (addr
& ~PAGE_MASK
) - BUF_PAGE_HDR_SIZE
;
1913 rb_event_is_commit(struct ring_buffer_per_cpu
*cpu_buffer
,
1914 struct ring_buffer_event
*event
)
1916 unsigned long addr
= (unsigned long)event
;
1917 unsigned long index
;
1919 index
= rb_event_index(event
);
1922 return cpu_buffer
->commit_page
->page
== (void *)addr
&&
1923 rb_commit_index(cpu_buffer
) == index
;
1927 rb_set_commit_to_write(struct ring_buffer_per_cpu
*cpu_buffer
)
1929 unsigned long max_count
;
1932 * We only race with interrupts and NMIs on this CPU.
1933 * If we own the commit event, then we can commit
1934 * all others that interrupted us, since the interruptions
1935 * are in stack format (they finish before they come
1936 * back to us). This allows us to do a simple loop to
1937 * assign the commit to the tail.
1940 max_count
= cpu_buffer
->nr_pages
* 100;
1942 while (cpu_buffer
->commit_page
!= cpu_buffer
->tail_page
) {
1943 if (RB_WARN_ON(cpu_buffer
, !(--max_count
)))
1945 if (RB_WARN_ON(cpu_buffer
,
1946 rb_is_reader_page(cpu_buffer
->tail_page
)))
1948 local_set(&cpu_buffer
->commit_page
->page
->commit
,
1949 rb_page_write(cpu_buffer
->commit_page
));
1950 rb_inc_page(cpu_buffer
, &cpu_buffer
->commit_page
);
1951 cpu_buffer
->write_stamp
=
1952 cpu_buffer
->commit_page
->page
->time_stamp
;
1953 /* add barrier to keep gcc from optimizing too much */
1956 while (rb_commit_index(cpu_buffer
) !=
1957 rb_page_write(cpu_buffer
->commit_page
)) {
1959 local_set(&cpu_buffer
->commit_page
->page
->commit
,
1960 rb_page_write(cpu_buffer
->commit_page
));
1961 RB_WARN_ON(cpu_buffer
,
1962 local_read(&cpu_buffer
->commit_page
->page
->commit
) &
1967 /* again, keep gcc from optimizing */
1971 * If an interrupt came in just after the first while loop
1972 * and pushed the tail page forward, we will be left with
1973 * a dangling commit that will never go forward.
1975 if (unlikely(cpu_buffer
->commit_page
!= cpu_buffer
->tail_page
))
1979 static void rb_inc_iter(struct ring_buffer_iter
*iter
)
1981 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
1984 * The iterator could be on the reader page (it starts there).
1985 * But the head could have moved, since the reader was
1986 * found. Check for this case and assign the iterator
1987 * to the head page instead of next.
1989 if (iter
->head_page
== cpu_buffer
->reader_page
)
1990 iter
->head_page
= rb_set_head_page(cpu_buffer
);
1992 rb_inc_page(cpu_buffer
, &iter
->head_page
);
1994 iter
->read_stamp
= iter
->head_page
->page
->time_stamp
;
1998 /* Slow path, do not inline */
1999 static noinline
struct ring_buffer_event
*
2000 rb_add_time_stamp(struct ring_buffer_event
*event
, u64 delta
)
2002 event
->type_len
= RINGBUF_TYPE_TIME_EXTEND
;
2004 /* Not the first event on the page? */
2005 if (rb_event_index(event
)) {
2006 event
->time_delta
= delta
& TS_MASK
;
2007 event
->array
[0] = delta
>> TS_SHIFT
;
2009 /* nope, just zero it */
2010 event
->time_delta
= 0;
2011 event
->array
[0] = 0;
2014 return skip_time_extend(event
);
2018 * rb_update_event - update event type and data
2019 * @event: the event to update
2020 * @type: the type of event
2021 * @length: the size of the event field in the ring buffer
2023 * Update the type and data fields of the event. The length
2024 * is the actual size that is written to the ring buffer,
2025 * and with this, we can determine what to place into the
2029 rb_update_event(struct ring_buffer_per_cpu
*cpu_buffer
,
2030 struct ring_buffer_event
*event
, unsigned length
,
2031 int add_timestamp
, u64 delta
)
2033 /* Only a commit updates the timestamp */
2034 if (unlikely(!rb_event_is_commit(cpu_buffer
, event
)))
2038 * If we need to add a timestamp, then we
2039 * add it to the start of the resevered space.
2041 if (unlikely(add_timestamp
)) {
2042 event
= rb_add_time_stamp(event
, delta
);
2043 length
-= RB_LEN_TIME_EXTEND
;
2047 event
->time_delta
= delta
;
2048 length
-= RB_EVNT_HDR_SIZE
;
2049 if (length
> RB_MAX_SMALL_DATA
|| RB_FORCE_8BYTE_ALIGNMENT
) {
2050 event
->type_len
= 0;
2051 event
->array
[0] = length
;
2053 event
->type_len
= DIV_ROUND_UP(length
, RB_ALIGNMENT
);
2057 * rb_handle_head_page - writer hit the head page
2059 * Returns: +1 to retry page
2064 rb_handle_head_page(struct ring_buffer_per_cpu
*cpu_buffer
,
2065 struct buffer_page
*tail_page
,
2066 struct buffer_page
*next_page
)
2068 struct buffer_page
*new_head
;
2073 entries
= rb_page_entries(next_page
);
2076 * The hard part is here. We need to move the head
2077 * forward, and protect against both readers on
2078 * other CPUs and writers coming in via interrupts.
2080 type
= rb_head_page_set_update(cpu_buffer
, next_page
, tail_page
,
2084 * type can be one of four:
2085 * NORMAL - an interrupt already moved it for us
2086 * HEAD - we are the first to get here.
2087 * UPDATE - we are the interrupt interrupting
2089 * MOVED - a reader on another CPU moved the next
2090 * pointer to its reader page. Give up
2097 * We changed the head to UPDATE, thus
2098 * it is our responsibility to update
2101 local_add(entries
, &cpu_buffer
->overrun
);
2102 local_sub(BUF_PAGE_SIZE
, &cpu_buffer
->entries_bytes
);
2105 * The entries will be zeroed out when we move the
2109 /* still more to do */
2112 case RB_PAGE_UPDATE
:
2114 * This is an interrupt that interrupt the
2115 * previous update. Still more to do.
2118 case RB_PAGE_NORMAL
:
2120 * An interrupt came in before the update
2121 * and processed this for us.
2122 * Nothing left to do.
2127 * The reader is on another CPU and just did
2128 * a swap with our next_page.
2133 RB_WARN_ON(cpu_buffer
, 1); /* WTF??? */
2138 * Now that we are here, the old head pointer is
2139 * set to UPDATE. This will keep the reader from
2140 * swapping the head page with the reader page.
2141 * The reader (on another CPU) will spin till
2144 * We just need to protect against interrupts
2145 * doing the job. We will set the next pointer
2146 * to HEAD. After that, we set the old pointer
2147 * to NORMAL, but only if it was HEAD before.
2148 * otherwise we are an interrupt, and only
2149 * want the outer most commit to reset it.
2151 new_head
= next_page
;
2152 rb_inc_page(cpu_buffer
, &new_head
);
2154 ret
= rb_head_page_set_head(cpu_buffer
, new_head
, next_page
,
2158 * Valid returns are:
2159 * HEAD - an interrupt came in and already set it.
2160 * NORMAL - One of two things:
2161 * 1) We really set it.
2162 * 2) A bunch of interrupts came in and moved
2163 * the page forward again.
2167 case RB_PAGE_NORMAL
:
2171 RB_WARN_ON(cpu_buffer
, 1);
2176 * It is possible that an interrupt came in,
2177 * set the head up, then more interrupts came in
2178 * and moved it again. When we get back here,
2179 * the page would have been set to NORMAL but we
2180 * just set it back to HEAD.
2182 * How do you detect this? Well, if that happened
2183 * the tail page would have moved.
2185 if (ret
== RB_PAGE_NORMAL
) {
2187 * If the tail had moved passed next, then we need
2188 * to reset the pointer.
2190 if (cpu_buffer
->tail_page
!= tail_page
&&
2191 cpu_buffer
->tail_page
!= next_page
)
2192 rb_head_page_set_normal(cpu_buffer
, new_head
,
2198 * If this was the outer most commit (the one that
2199 * changed the original pointer from HEAD to UPDATE),
2200 * then it is up to us to reset it to NORMAL.
2202 if (type
== RB_PAGE_HEAD
) {
2203 ret
= rb_head_page_set_normal(cpu_buffer
, next_page
,
2206 if (RB_WARN_ON(cpu_buffer
,
2207 ret
!= RB_PAGE_UPDATE
))
2214 static unsigned rb_calculate_event_length(unsigned length
)
2216 struct ring_buffer_event event
; /* Used only for sizeof array */
2218 /* zero length can cause confusions */
2222 if (length
> RB_MAX_SMALL_DATA
|| RB_FORCE_8BYTE_ALIGNMENT
)
2223 length
+= sizeof(event
.array
[0]);
2225 length
+= RB_EVNT_HDR_SIZE
;
2226 length
= ALIGN(length
, RB_ARCH_ALIGNMENT
);
2232 rb_reset_tail(struct ring_buffer_per_cpu
*cpu_buffer
,
2233 struct buffer_page
*tail_page
,
2234 unsigned long tail
, unsigned long length
)
2236 struct ring_buffer_event
*event
;
2239 * Only the event that crossed the page boundary
2240 * must fill the old tail_page with padding.
2242 if (tail
>= BUF_PAGE_SIZE
) {
2244 * If the page was filled, then we still need
2245 * to update the real_end. Reset it to zero
2246 * and the reader will ignore it.
2248 if (tail
== BUF_PAGE_SIZE
)
2249 tail_page
->real_end
= 0;
2251 local_sub(length
, &tail_page
->write
);
2255 event
= __rb_page_index(tail_page
, tail
);
2256 kmemcheck_annotate_bitfield(event
, bitfield
);
2258 /* account for padding bytes */
2259 local_add(BUF_PAGE_SIZE
- tail
, &cpu_buffer
->entries_bytes
);
2262 * Save the original length to the meta data.
2263 * This will be used by the reader to add lost event
2266 tail_page
->real_end
= tail
;
2269 * If this event is bigger than the minimum size, then
2270 * we need to be careful that we don't subtract the
2271 * write counter enough to allow another writer to slip
2273 * We put in a discarded commit instead, to make sure
2274 * that this space is not used again.
2276 * If we are less than the minimum size, we don't need to
2279 if (tail
> (BUF_PAGE_SIZE
- RB_EVNT_MIN_SIZE
)) {
2280 /* No room for any events */
2282 /* Mark the rest of the page with padding */
2283 rb_event_set_padding(event
);
2285 /* Set the write back to the previous setting */
2286 local_sub(length
, &tail_page
->write
);
2290 /* Put in a discarded event */
2291 event
->array
[0] = (BUF_PAGE_SIZE
- tail
) - RB_EVNT_HDR_SIZE
;
2292 event
->type_len
= RINGBUF_TYPE_PADDING
;
2293 /* time delta must be non zero */
2294 event
->time_delta
= 1;
2296 /* Set write to end of buffer */
2297 length
= (tail
+ length
) - BUF_PAGE_SIZE
;
2298 local_sub(length
, &tail_page
->write
);
2302 * This is the slow path, force gcc not to inline it.
2304 static noinline
struct ring_buffer_event
*
2305 rb_move_tail(struct ring_buffer_per_cpu
*cpu_buffer
,
2306 unsigned long length
, unsigned long tail
,
2307 struct buffer_page
*tail_page
, u64 ts
)
2309 struct buffer_page
*commit_page
= cpu_buffer
->commit_page
;
2310 struct ring_buffer
*buffer
= cpu_buffer
->buffer
;
2311 struct buffer_page
*next_page
;
2314 next_page
= tail_page
;
2316 rb_inc_page(cpu_buffer
, &next_page
);
2319 * If for some reason, we had an interrupt storm that made
2320 * it all the way around the buffer, bail, and warn
2323 if (unlikely(next_page
== commit_page
)) {
2324 local_inc(&cpu_buffer
->commit_overrun
);
2329 * This is where the fun begins!
2331 * We are fighting against races between a reader that
2332 * could be on another CPU trying to swap its reader
2333 * page with the buffer head.
2335 * We are also fighting against interrupts coming in and
2336 * moving the head or tail on us as well.
2338 * If the next page is the head page then we have filled
2339 * the buffer, unless the commit page is still on the
2342 if (rb_is_head_page(cpu_buffer
, next_page
, &tail_page
->list
)) {
2345 * If the commit is not on the reader page, then
2346 * move the header page.
2348 if (!rb_is_reader_page(cpu_buffer
->commit_page
)) {
2350 * If we are not in overwrite mode,
2351 * this is easy, just stop here.
2353 if (!(buffer
->flags
& RB_FL_OVERWRITE
)) {
2354 local_inc(&cpu_buffer
->dropped_events
);
2358 ret
= rb_handle_head_page(cpu_buffer
,
2367 * We need to be careful here too. The
2368 * commit page could still be on the reader
2369 * page. We could have a small buffer, and
2370 * have filled up the buffer with events
2371 * from interrupts and such, and wrapped.
2373 * Note, if the tail page is also the on the
2374 * reader_page, we let it move out.
2376 if (unlikely((cpu_buffer
->commit_page
!=
2377 cpu_buffer
->tail_page
) &&
2378 (cpu_buffer
->commit_page
==
2379 cpu_buffer
->reader_page
))) {
2380 local_inc(&cpu_buffer
->commit_overrun
);
2386 ret
= rb_tail_page_update(cpu_buffer
, tail_page
, next_page
);
2389 * Nested commits always have zero deltas, so
2390 * just reread the time stamp
2392 ts
= rb_time_stamp(buffer
);
2393 next_page
->page
->time_stamp
= ts
;
2398 rb_reset_tail(cpu_buffer
, tail_page
, tail
, length
);
2400 /* fail and let the caller try again */
2401 return ERR_PTR(-EAGAIN
);
2405 rb_reset_tail(cpu_buffer
, tail_page
, tail
, length
);
2410 static struct ring_buffer_event
*
2411 __rb_reserve_next(struct ring_buffer_per_cpu
*cpu_buffer
,
2412 unsigned long length
, u64 ts
,
2413 u64 delta
, int add_timestamp
)
2415 struct buffer_page
*tail_page
;
2416 struct ring_buffer_event
*event
;
2417 unsigned long tail
, write
;
2420 * If the time delta since the last event is too big to
2421 * hold in the time field of the event, then we append a
2422 * TIME EXTEND event ahead of the data event.
2424 if (unlikely(add_timestamp
))
2425 length
+= RB_LEN_TIME_EXTEND
;
2427 tail_page
= cpu_buffer
->tail_page
;
2428 write
= local_add_return(length
, &tail_page
->write
);
2430 /* set write to only the index of the write */
2431 write
&= RB_WRITE_MASK
;
2432 tail
= write
- length
;
2435 * If this is the first commit on the page, then it has the same
2436 * timestamp as the page itself.
2441 /* See if we shot pass the end of this buffer page */
2442 if (unlikely(write
> BUF_PAGE_SIZE
))
2443 return rb_move_tail(cpu_buffer
, length
, tail
,
2446 /* We reserved something on the buffer */
2448 event
= __rb_page_index(tail_page
, tail
);
2449 kmemcheck_annotate_bitfield(event
, bitfield
);
2450 rb_update_event(cpu_buffer
, event
, length
, add_timestamp
, delta
);
2452 local_inc(&tail_page
->entries
);
2455 * If this is the first commit on the page, then update
2459 tail_page
->page
->time_stamp
= ts
;
2461 /* account for these added bytes */
2462 local_add(length
, &cpu_buffer
->entries_bytes
);
2468 rb_try_to_discard(struct ring_buffer_per_cpu
*cpu_buffer
,
2469 struct ring_buffer_event
*event
)
2471 unsigned long new_index
, old_index
;
2472 struct buffer_page
*bpage
;
2473 unsigned long index
;
2476 new_index
= rb_event_index(event
);
2477 old_index
= new_index
+ rb_event_ts_length(event
);
2478 addr
= (unsigned long)event
;
2481 bpage
= cpu_buffer
->tail_page
;
2483 if (bpage
->page
== (void *)addr
&& rb_page_write(bpage
) == old_index
) {
2484 unsigned long write_mask
=
2485 local_read(&bpage
->write
) & ~RB_WRITE_MASK
;
2486 unsigned long event_length
= rb_event_length(event
);
2488 * This is on the tail page. It is possible that
2489 * a write could come in and move the tail page
2490 * and write to the next page. That is fine
2491 * because we just shorten what is on this page.
2493 old_index
+= write_mask
;
2494 new_index
+= write_mask
;
2495 index
= local_cmpxchg(&bpage
->write
, old_index
, new_index
);
2496 if (index
== old_index
) {
2497 /* update counters */
2498 local_sub(event_length
, &cpu_buffer
->entries_bytes
);
2503 /* could not discard */
2507 static void rb_start_commit(struct ring_buffer_per_cpu
*cpu_buffer
)
2509 local_inc(&cpu_buffer
->committing
);
2510 local_inc(&cpu_buffer
->commits
);
2513 static inline void rb_end_commit(struct ring_buffer_per_cpu
*cpu_buffer
)
2515 unsigned long commits
;
2517 if (RB_WARN_ON(cpu_buffer
,
2518 !local_read(&cpu_buffer
->committing
)))
2522 commits
= local_read(&cpu_buffer
->commits
);
2523 /* synchronize with interrupts */
2525 if (local_read(&cpu_buffer
->committing
) == 1)
2526 rb_set_commit_to_write(cpu_buffer
);
2528 local_dec(&cpu_buffer
->committing
);
2530 /* synchronize with interrupts */
2534 * Need to account for interrupts coming in between the
2535 * updating of the commit page and the clearing of the
2536 * committing counter.
2538 if (unlikely(local_read(&cpu_buffer
->commits
) != commits
) &&
2539 !local_read(&cpu_buffer
->committing
)) {
2540 local_inc(&cpu_buffer
->committing
);
2545 static struct ring_buffer_event
*
2546 rb_reserve_next_event(struct ring_buffer
*buffer
,
2547 struct ring_buffer_per_cpu
*cpu_buffer
,
2548 unsigned long length
)
2550 struct ring_buffer_event
*event
;
2556 rb_start_commit(cpu_buffer
);
2558 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2560 * Due to the ability to swap a cpu buffer from a buffer
2561 * it is possible it was swapped before we committed.
2562 * (committing stops a swap). We check for it here and
2563 * if it happened, we have to fail the write.
2566 if (unlikely(ACCESS_ONCE(cpu_buffer
->buffer
) != buffer
)) {
2567 local_dec(&cpu_buffer
->committing
);
2568 local_dec(&cpu_buffer
->commits
);
2573 length
= rb_calculate_event_length(length
);
2579 * We allow for interrupts to reenter here and do a trace.
2580 * If one does, it will cause this original code to loop
2581 * back here. Even with heavy interrupts happening, this
2582 * should only happen a few times in a row. If this happens
2583 * 1000 times in a row, there must be either an interrupt
2584 * storm or we have something buggy.
2587 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 1000))
2590 ts
= rb_time_stamp(cpu_buffer
->buffer
);
2591 diff
= ts
- cpu_buffer
->write_stamp
;
2593 /* make sure this diff is calculated here */
2596 /* Did the write stamp get updated already? */
2597 if (likely(ts
>= cpu_buffer
->write_stamp
)) {
2599 if (unlikely(test_time_stamp(delta
))) {
2600 int local_clock_stable
= 1;
2601 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2602 local_clock_stable
= sched_clock_stable
;
2604 WARN_ONCE(delta
> (1ULL << 59),
2605 KERN_WARNING
"Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2606 (unsigned long long)delta
,
2607 (unsigned long long)ts
,
2608 (unsigned long long)cpu_buffer
->write_stamp
,
2609 local_clock_stable
? "" :
2610 "If you just came from a suspend/resume,\n"
2611 "please switch to the trace global clock:\n"
2612 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2617 event
= __rb_reserve_next(cpu_buffer
, length
, ts
,
2618 delta
, add_timestamp
);
2619 if (unlikely(PTR_ERR(event
) == -EAGAIN
))
2628 rb_end_commit(cpu_buffer
);
2632 #ifdef CONFIG_TRACING
2635 * The lock and unlock are done within a preempt disable section.
2636 * The current_context per_cpu variable can only be modified
2637 * by the current task between lock and unlock. But it can
2638 * be modified more than once via an interrupt. To pass this
2639 * information from the lock to the unlock without having to
2640 * access the 'in_interrupt()' functions again (which do show
2641 * a bit of overhead in something as critical as function tracing,
2642 * we use a bitmask trick.
2644 * bit 0 = NMI context
2645 * bit 1 = IRQ context
2646 * bit 2 = SoftIRQ context
2647 * bit 3 = normal context.
2649 * This works because this is the order of contexts that can
2650 * preempt other contexts. A SoftIRQ never preempts an IRQ
2653 * When the context is determined, the corresponding bit is
2654 * checked and set (if it was set, then a recursion of that context
2657 * On unlock, we need to clear this bit. To do so, just subtract
2658 * 1 from the current_context and AND it to itself.
2662 * 101 & 100 = 100 (clearing bit zero)
2665 * 1010 & 1001 = 1000 (clearing bit 1)
2667 * The least significant bit can be cleared this way, and it
2668 * just so happens that it is the same bit corresponding to
2669 * the current context.
2671 static DEFINE_PER_CPU(unsigned int, current_context
);
2673 static __always_inline
int trace_recursive_lock(void)
2675 unsigned int val
= __this_cpu_read(current_context
);
2678 if (in_interrupt()) {
2688 if (unlikely(val
& (1 << bit
)))
2692 __this_cpu_write(current_context
, val
);
2697 static __always_inline
void trace_recursive_unlock(void)
2699 unsigned int val
= __this_cpu_read(current_context
);
2701 val
&= val
& (val
- 1);
2702 __this_cpu_write(current_context
, val
);
2707 #define trace_recursive_lock() (0)
2708 #define trace_recursive_unlock() do { } while (0)
2713 * ring_buffer_lock_reserve - reserve a part of the buffer
2714 * @buffer: the ring buffer to reserve from
2715 * @length: the length of the data to reserve (excluding event header)
2717 * Returns a reseverd event on the ring buffer to copy directly to.
2718 * The user of this interface will need to get the body to write into
2719 * and can use the ring_buffer_event_data() interface.
2721 * The length is the length of the data needed, not the event length
2722 * which also includes the event header.
2724 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2725 * If NULL is returned, then nothing has been allocated or locked.
2727 struct ring_buffer_event
*
2728 ring_buffer_lock_reserve(struct ring_buffer
*buffer
, unsigned long length
)
2730 struct ring_buffer_per_cpu
*cpu_buffer
;
2731 struct ring_buffer_event
*event
;
2734 if (ring_buffer_flags
!= RB_BUFFERS_ON
)
2737 /* If we are tracing schedule, we don't want to recurse */
2738 preempt_disable_notrace();
2740 if (atomic_read(&buffer
->record_disabled
))
2743 if (trace_recursive_lock())
2746 cpu
= raw_smp_processor_id();
2748 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
2751 cpu_buffer
= buffer
->buffers
[cpu
];
2753 if (atomic_read(&cpu_buffer
->record_disabled
))
2756 if (length
> BUF_MAX_DATA_SIZE
)
2759 event
= rb_reserve_next_event(buffer
, cpu_buffer
, length
);
2766 trace_recursive_unlock();
2769 preempt_enable_notrace();
2772 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve
);
2775 rb_update_write_stamp(struct ring_buffer_per_cpu
*cpu_buffer
,
2776 struct ring_buffer_event
*event
)
2781 * The event first in the commit queue updates the
2784 if (rb_event_is_commit(cpu_buffer
, event
)) {
2786 * A commit event that is first on a page
2787 * updates the write timestamp with the page stamp
2789 if (!rb_event_index(event
))
2790 cpu_buffer
->write_stamp
=
2791 cpu_buffer
->commit_page
->page
->time_stamp
;
2792 else if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
) {
2793 delta
= event
->array
[0];
2795 delta
+= event
->time_delta
;
2796 cpu_buffer
->write_stamp
+= delta
;
2798 cpu_buffer
->write_stamp
+= event
->time_delta
;
2802 static void rb_commit(struct ring_buffer_per_cpu
*cpu_buffer
,
2803 struct ring_buffer_event
*event
)
2805 local_inc(&cpu_buffer
->entries
);
2806 rb_update_write_stamp(cpu_buffer
, event
);
2807 rb_end_commit(cpu_buffer
);
2810 static __always_inline
void
2811 rb_wakeups(struct ring_buffer
*buffer
, struct ring_buffer_per_cpu
*cpu_buffer
)
2813 if (buffer
->irq_work
.waiters_pending
) {
2814 buffer
->irq_work
.waiters_pending
= false;
2815 /* irq_work_queue() supplies it's own memory barriers */
2816 irq_work_queue(&buffer
->irq_work
.work
);
2819 if (cpu_buffer
->irq_work
.waiters_pending
) {
2820 cpu_buffer
->irq_work
.waiters_pending
= false;
2821 /* irq_work_queue() supplies it's own memory barriers */
2822 irq_work_queue(&cpu_buffer
->irq_work
.work
);
2827 * ring_buffer_unlock_commit - commit a reserved
2828 * @buffer: The buffer to commit to
2829 * @event: The event pointer to commit.
2831 * This commits the data to the ring buffer, and releases any locks held.
2833 * Must be paired with ring_buffer_lock_reserve.
2835 int ring_buffer_unlock_commit(struct ring_buffer
*buffer
,
2836 struct ring_buffer_event
*event
)
2838 struct ring_buffer_per_cpu
*cpu_buffer
;
2839 int cpu
= raw_smp_processor_id();
2841 cpu_buffer
= buffer
->buffers
[cpu
];
2843 rb_commit(cpu_buffer
, event
);
2845 rb_wakeups(buffer
, cpu_buffer
);
2847 trace_recursive_unlock();
2849 preempt_enable_notrace();
2853 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit
);
2855 static inline void rb_event_discard(struct ring_buffer_event
*event
)
2857 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
2858 event
= skip_time_extend(event
);
2860 /* array[0] holds the actual length for the discarded event */
2861 event
->array
[0] = rb_event_data_length(event
) - RB_EVNT_HDR_SIZE
;
2862 event
->type_len
= RINGBUF_TYPE_PADDING
;
2863 /* time delta must be non zero */
2864 if (!event
->time_delta
)
2865 event
->time_delta
= 1;
2869 * Decrement the entries to the page that an event is on.
2870 * The event does not even need to exist, only the pointer
2871 * to the page it is on. This may only be called before the commit
2875 rb_decrement_entry(struct ring_buffer_per_cpu
*cpu_buffer
,
2876 struct ring_buffer_event
*event
)
2878 unsigned long addr
= (unsigned long)event
;
2879 struct buffer_page
*bpage
= cpu_buffer
->commit_page
;
2880 struct buffer_page
*start
;
2884 /* Do the likely case first */
2885 if (likely(bpage
->page
== (void *)addr
)) {
2886 local_dec(&bpage
->entries
);
2891 * Because the commit page may be on the reader page we
2892 * start with the next page and check the end loop there.
2894 rb_inc_page(cpu_buffer
, &bpage
);
2897 if (bpage
->page
== (void *)addr
) {
2898 local_dec(&bpage
->entries
);
2901 rb_inc_page(cpu_buffer
, &bpage
);
2902 } while (bpage
!= start
);
2904 /* commit not part of this buffer?? */
2905 RB_WARN_ON(cpu_buffer
, 1);
2909 * ring_buffer_commit_discard - discard an event that has not been committed
2910 * @buffer: the ring buffer
2911 * @event: non committed event to discard
2913 * Sometimes an event that is in the ring buffer needs to be ignored.
2914 * This function lets the user discard an event in the ring buffer
2915 * and then that event will not be read later.
2917 * This function only works if it is called before the the item has been
2918 * committed. It will try to free the event from the ring buffer
2919 * if another event has not been added behind it.
2921 * If another event has been added behind it, it will set the event
2922 * up as discarded, and perform the commit.
2924 * If this function is called, do not call ring_buffer_unlock_commit on
2927 void ring_buffer_discard_commit(struct ring_buffer
*buffer
,
2928 struct ring_buffer_event
*event
)
2930 struct ring_buffer_per_cpu
*cpu_buffer
;
2933 /* The event is discarded regardless */
2934 rb_event_discard(event
);
2936 cpu
= smp_processor_id();
2937 cpu_buffer
= buffer
->buffers
[cpu
];
2940 * This must only be called if the event has not been
2941 * committed yet. Thus we can assume that preemption
2942 * is still disabled.
2944 RB_WARN_ON(buffer
, !local_read(&cpu_buffer
->committing
));
2946 rb_decrement_entry(cpu_buffer
, event
);
2947 if (rb_try_to_discard(cpu_buffer
, event
))
2951 * The commit is still visible by the reader, so we
2952 * must still update the timestamp.
2954 rb_update_write_stamp(cpu_buffer
, event
);
2956 rb_end_commit(cpu_buffer
);
2958 trace_recursive_unlock();
2960 preempt_enable_notrace();
2963 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit
);
2966 * ring_buffer_write - write data to the buffer without reserving
2967 * @buffer: The ring buffer to write to.
2968 * @length: The length of the data being written (excluding the event header)
2969 * @data: The data to write to the buffer.
2971 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2972 * one function. If you already have the data to write to the buffer, it
2973 * may be easier to simply call this function.
2975 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2976 * and not the length of the event which would hold the header.
2978 int ring_buffer_write(struct ring_buffer
*buffer
,
2979 unsigned long length
,
2982 struct ring_buffer_per_cpu
*cpu_buffer
;
2983 struct ring_buffer_event
*event
;
2988 if (ring_buffer_flags
!= RB_BUFFERS_ON
)
2991 preempt_disable_notrace();
2993 if (atomic_read(&buffer
->record_disabled
))
2996 cpu
= raw_smp_processor_id();
2998 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3001 cpu_buffer
= buffer
->buffers
[cpu
];
3003 if (atomic_read(&cpu_buffer
->record_disabled
))
3006 if (length
> BUF_MAX_DATA_SIZE
)
3009 event
= rb_reserve_next_event(buffer
, cpu_buffer
, length
);
3013 body
= rb_event_data(event
);
3015 memcpy(body
, data
, length
);
3017 rb_commit(cpu_buffer
, event
);
3019 rb_wakeups(buffer
, cpu_buffer
);
3023 preempt_enable_notrace();
3027 EXPORT_SYMBOL_GPL(ring_buffer_write
);
3029 static int rb_per_cpu_empty(struct ring_buffer_per_cpu
*cpu_buffer
)
3031 struct buffer_page
*reader
= cpu_buffer
->reader_page
;
3032 struct buffer_page
*head
= rb_set_head_page(cpu_buffer
);
3033 struct buffer_page
*commit
= cpu_buffer
->commit_page
;
3035 /* In case of error, head will be NULL */
3036 if (unlikely(!head
))
3039 return reader
->read
== rb_page_commit(reader
) &&
3040 (commit
== reader
||
3042 head
->read
== rb_page_commit(commit
)));
3046 * ring_buffer_record_disable - stop all writes into the buffer
3047 * @buffer: The ring buffer to stop writes to.
3049 * This prevents all writes to the buffer. Any attempt to write
3050 * to the buffer after this will fail and return NULL.
3052 * The caller should call synchronize_sched() after this.
3054 void ring_buffer_record_disable(struct ring_buffer
*buffer
)
3056 atomic_inc(&buffer
->record_disabled
);
3058 EXPORT_SYMBOL_GPL(ring_buffer_record_disable
);
3061 * ring_buffer_record_enable - enable writes to the buffer
3062 * @buffer: The ring buffer to enable writes
3064 * Note, multiple disables will need the same number of enables
3065 * to truly enable the writing (much like preempt_disable).
3067 void ring_buffer_record_enable(struct ring_buffer
*buffer
)
3069 atomic_dec(&buffer
->record_disabled
);
3071 EXPORT_SYMBOL_GPL(ring_buffer_record_enable
);
3074 * ring_buffer_record_off - stop all writes into the buffer
3075 * @buffer: The ring buffer to stop writes to.
3077 * This prevents all writes to the buffer. Any attempt to write
3078 * to the buffer after this will fail and return NULL.
3080 * This is different than ring_buffer_record_disable() as
3081 * it works like an on/off switch, where as the disable() version
3082 * must be paired with a enable().
3084 void ring_buffer_record_off(struct ring_buffer
*buffer
)
3087 unsigned int new_rd
;
3090 rd
= atomic_read(&buffer
->record_disabled
);
3091 new_rd
= rd
| RB_BUFFER_OFF
;
3092 } while (atomic_cmpxchg(&buffer
->record_disabled
, rd
, new_rd
) != rd
);
3094 EXPORT_SYMBOL_GPL(ring_buffer_record_off
);
3097 * ring_buffer_record_on - restart writes into the buffer
3098 * @buffer: The ring buffer to start writes to.
3100 * This enables all writes to the buffer that was disabled by
3101 * ring_buffer_record_off().
3103 * This is different than ring_buffer_record_enable() as
3104 * it works like an on/off switch, where as the enable() version
3105 * must be paired with a disable().
3107 void ring_buffer_record_on(struct ring_buffer
*buffer
)
3110 unsigned int new_rd
;
3113 rd
= atomic_read(&buffer
->record_disabled
);
3114 new_rd
= rd
& ~RB_BUFFER_OFF
;
3115 } while (atomic_cmpxchg(&buffer
->record_disabled
, rd
, new_rd
) != rd
);
3117 EXPORT_SYMBOL_GPL(ring_buffer_record_on
);
3120 * ring_buffer_record_is_on - return true if the ring buffer can write
3121 * @buffer: The ring buffer to see if write is enabled
3123 * Returns true if the ring buffer is in a state that it accepts writes.
3125 int ring_buffer_record_is_on(struct ring_buffer
*buffer
)
3127 return !atomic_read(&buffer
->record_disabled
);
3131 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3132 * @buffer: The ring buffer to stop writes to.
3133 * @cpu: The CPU buffer to stop
3135 * This prevents all writes to the buffer. Any attempt to write
3136 * to the buffer after this will fail and return NULL.
3138 * The caller should call synchronize_sched() after this.
3140 void ring_buffer_record_disable_cpu(struct ring_buffer
*buffer
, int cpu
)
3142 struct ring_buffer_per_cpu
*cpu_buffer
;
3144 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3147 cpu_buffer
= buffer
->buffers
[cpu
];
3148 atomic_inc(&cpu_buffer
->record_disabled
);
3150 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu
);
3153 * ring_buffer_record_enable_cpu - enable writes to the buffer
3154 * @buffer: The ring buffer to enable writes
3155 * @cpu: The CPU to enable.
3157 * Note, multiple disables will need the same number of enables
3158 * to truly enable the writing (much like preempt_disable).
3160 void ring_buffer_record_enable_cpu(struct ring_buffer
*buffer
, int cpu
)
3162 struct ring_buffer_per_cpu
*cpu_buffer
;
3164 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3167 cpu_buffer
= buffer
->buffers
[cpu
];
3168 atomic_dec(&cpu_buffer
->record_disabled
);
3170 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu
);
3173 * The total entries in the ring buffer is the running counter
3174 * of entries entered into the ring buffer, minus the sum of
3175 * the entries read from the ring buffer and the number of
3176 * entries that were overwritten.
3178 static inline unsigned long
3179 rb_num_of_entries(struct ring_buffer_per_cpu
*cpu_buffer
)
3181 return local_read(&cpu_buffer
->entries
) -
3182 (local_read(&cpu_buffer
->overrun
) + cpu_buffer
->read
);
3186 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3187 * @buffer: The ring buffer
3188 * @cpu: The per CPU buffer to read from.
3190 u64
ring_buffer_oldest_event_ts(struct ring_buffer
*buffer
, int cpu
)
3192 unsigned long flags
;
3193 struct ring_buffer_per_cpu
*cpu_buffer
;
3194 struct buffer_page
*bpage
;
3197 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3200 cpu_buffer
= buffer
->buffers
[cpu
];
3201 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3203 * if the tail is on reader_page, oldest time stamp is on the reader
3206 if (cpu_buffer
->tail_page
== cpu_buffer
->reader_page
)
3207 bpage
= cpu_buffer
->reader_page
;
3209 bpage
= rb_set_head_page(cpu_buffer
);
3211 ret
= bpage
->page
->time_stamp
;
3212 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3216 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts
);
3219 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3220 * @buffer: The ring buffer
3221 * @cpu: The per CPU buffer to read from.
3223 unsigned long ring_buffer_bytes_cpu(struct ring_buffer
*buffer
, int cpu
)
3225 struct ring_buffer_per_cpu
*cpu_buffer
;
3228 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3231 cpu_buffer
= buffer
->buffers
[cpu
];
3232 ret
= local_read(&cpu_buffer
->entries_bytes
) - cpu_buffer
->read_bytes
;
3236 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu
);
3239 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3240 * @buffer: The ring buffer
3241 * @cpu: The per CPU buffer to get the entries from.
3243 unsigned long ring_buffer_entries_cpu(struct ring_buffer
*buffer
, int cpu
)
3245 struct ring_buffer_per_cpu
*cpu_buffer
;
3247 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3250 cpu_buffer
= buffer
->buffers
[cpu
];
3252 return rb_num_of_entries(cpu_buffer
);
3254 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu
);
3257 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3258 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3259 * @buffer: The ring buffer
3260 * @cpu: The per CPU buffer to get the number of overruns from
3262 unsigned long ring_buffer_overrun_cpu(struct ring_buffer
*buffer
, int cpu
)
3264 struct ring_buffer_per_cpu
*cpu_buffer
;
3267 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3270 cpu_buffer
= buffer
->buffers
[cpu
];
3271 ret
= local_read(&cpu_buffer
->overrun
);
3275 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu
);
3278 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3279 * commits failing due to the buffer wrapping around while there are uncommitted
3280 * events, such as during an interrupt storm.
3281 * @buffer: The ring buffer
3282 * @cpu: The per CPU buffer to get the number of overruns from
3285 ring_buffer_commit_overrun_cpu(struct ring_buffer
*buffer
, int cpu
)
3287 struct ring_buffer_per_cpu
*cpu_buffer
;
3290 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3293 cpu_buffer
= buffer
->buffers
[cpu
];
3294 ret
= local_read(&cpu_buffer
->commit_overrun
);
3298 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu
);
3301 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3302 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3303 * @buffer: The ring buffer
3304 * @cpu: The per CPU buffer to get the number of overruns from
3307 ring_buffer_dropped_events_cpu(struct ring_buffer
*buffer
, int cpu
)
3309 struct ring_buffer_per_cpu
*cpu_buffer
;
3312 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3315 cpu_buffer
= buffer
->buffers
[cpu
];
3316 ret
= local_read(&cpu_buffer
->dropped_events
);
3320 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu
);
3323 * ring_buffer_read_events_cpu - get the number of events successfully read
3324 * @buffer: The ring buffer
3325 * @cpu: The per CPU buffer to get the number of events read
3328 ring_buffer_read_events_cpu(struct ring_buffer
*buffer
, int cpu
)
3330 struct ring_buffer_per_cpu
*cpu_buffer
;
3332 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3335 cpu_buffer
= buffer
->buffers
[cpu
];
3336 return cpu_buffer
->read
;
3338 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu
);
3341 * ring_buffer_entries - get the number of entries in a buffer
3342 * @buffer: The ring buffer
3344 * Returns the total number of entries in the ring buffer
3347 unsigned long ring_buffer_entries(struct ring_buffer
*buffer
)
3349 struct ring_buffer_per_cpu
*cpu_buffer
;
3350 unsigned long entries
= 0;
3353 /* if you care about this being correct, lock the buffer */
3354 for_each_buffer_cpu(buffer
, cpu
) {
3355 cpu_buffer
= buffer
->buffers
[cpu
];
3356 entries
+= rb_num_of_entries(cpu_buffer
);
3361 EXPORT_SYMBOL_GPL(ring_buffer_entries
);
3364 * ring_buffer_overruns - get the number of overruns in buffer
3365 * @buffer: The ring buffer
3367 * Returns the total number of overruns in the ring buffer
3370 unsigned long ring_buffer_overruns(struct ring_buffer
*buffer
)
3372 struct ring_buffer_per_cpu
*cpu_buffer
;
3373 unsigned long overruns
= 0;
3376 /* if you care about this being correct, lock the buffer */
3377 for_each_buffer_cpu(buffer
, cpu
) {
3378 cpu_buffer
= buffer
->buffers
[cpu
];
3379 overruns
+= local_read(&cpu_buffer
->overrun
);
3384 EXPORT_SYMBOL_GPL(ring_buffer_overruns
);
3386 static void rb_iter_reset(struct ring_buffer_iter
*iter
)
3388 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
3390 /* Iterator usage is expected to have record disabled */
3391 iter
->head_page
= cpu_buffer
->reader_page
;
3392 iter
->head
= cpu_buffer
->reader_page
->read
;
3394 iter
->cache_reader_page
= iter
->head_page
;
3395 iter
->cache_read
= cpu_buffer
->read
;
3398 iter
->read_stamp
= cpu_buffer
->read_stamp
;
3400 iter
->read_stamp
= iter
->head_page
->page
->time_stamp
;
3404 * ring_buffer_iter_reset - reset an iterator
3405 * @iter: The iterator to reset
3407 * Resets the iterator, so that it will start from the beginning
3410 void ring_buffer_iter_reset(struct ring_buffer_iter
*iter
)
3412 struct ring_buffer_per_cpu
*cpu_buffer
;
3413 unsigned long flags
;
3418 cpu_buffer
= iter
->cpu_buffer
;
3420 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3421 rb_iter_reset(iter
);
3422 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3424 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset
);
3427 * ring_buffer_iter_empty - check if an iterator has no more to read
3428 * @iter: The iterator to check
3430 int ring_buffer_iter_empty(struct ring_buffer_iter
*iter
)
3432 struct ring_buffer_per_cpu
*cpu_buffer
;
3433 struct buffer_page
*reader
;
3434 struct buffer_page
*head_page
;
3435 struct buffer_page
*commit_page
;
3438 cpu_buffer
= iter
->cpu_buffer
;
3440 /* Remember, trace recording is off when iterator is in use */
3441 reader
= cpu_buffer
->reader_page
;
3442 head_page
= cpu_buffer
->head_page
;
3443 commit_page
= cpu_buffer
->commit_page
;
3444 commit
= rb_page_commit(commit_page
);
3446 return ((iter
->head_page
== commit_page
&& iter
->head
== commit
) ||
3447 (iter
->head_page
== reader
&& commit_page
== head_page
&&
3448 head_page
->read
== commit
&&
3449 iter
->head
== rb_page_commit(cpu_buffer
->reader_page
)));
3451 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty
);
3454 rb_update_read_stamp(struct ring_buffer_per_cpu
*cpu_buffer
,
3455 struct ring_buffer_event
*event
)
3459 switch (event
->type_len
) {
3460 case RINGBUF_TYPE_PADDING
:
3463 case RINGBUF_TYPE_TIME_EXTEND
:
3464 delta
= event
->array
[0];
3466 delta
+= event
->time_delta
;
3467 cpu_buffer
->read_stamp
+= delta
;
3470 case RINGBUF_TYPE_TIME_STAMP
:
3471 /* FIXME: not implemented */
3474 case RINGBUF_TYPE_DATA
:
3475 cpu_buffer
->read_stamp
+= event
->time_delta
;
3485 rb_update_iter_read_stamp(struct ring_buffer_iter
*iter
,
3486 struct ring_buffer_event
*event
)
3490 switch (event
->type_len
) {
3491 case RINGBUF_TYPE_PADDING
:
3494 case RINGBUF_TYPE_TIME_EXTEND
:
3495 delta
= event
->array
[0];
3497 delta
+= event
->time_delta
;
3498 iter
->read_stamp
+= delta
;
3501 case RINGBUF_TYPE_TIME_STAMP
:
3502 /* FIXME: not implemented */
3505 case RINGBUF_TYPE_DATA
:
3506 iter
->read_stamp
+= event
->time_delta
;
3515 static struct buffer_page
*
3516 rb_get_reader_page(struct ring_buffer_per_cpu
*cpu_buffer
)
3518 struct buffer_page
*reader
= NULL
;
3519 unsigned long overwrite
;
3520 unsigned long flags
;
3524 local_irq_save(flags
);
3525 arch_spin_lock(&cpu_buffer
->lock
);
3529 * This should normally only loop twice. But because the
3530 * start of the reader inserts an empty page, it causes
3531 * a case where we will loop three times. There should be no
3532 * reason to loop four times (that I know of).
3534 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 3)) {
3539 reader
= cpu_buffer
->reader_page
;
3541 /* If there's more to read, return this page */
3542 if (cpu_buffer
->reader_page
->read
< rb_page_size(reader
))
3545 /* Never should we have an index greater than the size */
3546 if (RB_WARN_ON(cpu_buffer
,
3547 cpu_buffer
->reader_page
->read
> rb_page_size(reader
)))
3550 /* check if we caught up to the tail */
3552 if (cpu_buffer
->commit_page
== cpu_buffer
->reader_page
)
3555 /* Don't bother swapping if the ring buffer is empty */
3556 if (rb_num_of_entries(cpu_buffer
) == 0)
3560 * Reset the reader page to size zero.
3562 local_set(&cpu_buffer
->reader_page
->write
, 0);
3563 local_set(&cpu_buffer
->reader_page
->entries
, 0);
3564 local_set(&cpu_buffer
->reader_page
->page
->commit
, 0);
3565 cpu_buffer
->reader_page
->real_end
= 0;
3569 * Splice the empty reader page into the list around the head.
3571 reader
= rb_set_head_page(cpu_buffer
);
3574 cpu_buffer
->reader_page
->list
.next
= rb_list_head(reader
->list
.next
);
3575 cpu_buffer
->reader_page
->list
.prev
= reader
->list
.prev
;
3578 * cpu_buffer->pages just needs to point to the buffer, it
3579 * has no specific buffer page to point to. Lets move it out
3580 * of our way so we don't accidentally swap it.
3582 cpu_buffer
->pages
= reader
->list
.prev
;
3584 /* The reader page will be pointing to the new head */
3585 rb_set_list_to_head(cpu_buffer
, &cpu_buffer
->reader_page
->list
);
3588 * We want to make sure we read the overruns after we set up our
3589 * pointers to the next object. The writer side does a
3590 * cmpxchg to cross pages which acts as the mb on the writer
3591 * side. Note, the reader will constantly fail the swap
3592 * while the writer is updating the pointers, so this
3593 * guarantees that the overwrite recorded here is the one we
3594 * want to compare with the last_overrun.
3597 overwrite
= local_read(&(cpu_buffer
->overrun
));
3600 * Here's the tricky part.
3602 * We need to move the pointer past the header page.
3603 * But we can only do that if a writer is not currently
3604 * moving it. The page before the header page has the
3605 * flag bit '1' set if it is pointing to the page we want.
3606 * but if the writer is in the process of moving it
3607 * than it will be '2' or already moved '0'.
3610 ret
= rb_head_page_replace(reader
, cpu_buffer
->reader_page
);
3613 * If we did not convert it, then we must try again.
3619 * Yeah! We succeeded in replacing the page.
3621 * Now make the new head point back to the reader page.
3623 rb_list_head(reader
->list
.next
)->prev
= &cpu_buffer
->reader_page
->list
;
3624 rb_inc_page(cpu_buffer
, &cpu_buffer
->head_page
);
3626 /* Finally update the reader page to the new head */
3627 cpu_buffer
->reader_page
= reader
;
3628 cpu_buffer
->reader_page
->read
= 0;
3630 if (overwrite
!= cpu_buffer
->last_overrun
) {
3631 cpu_buffer
->lost_events
= overwrite
- cpu_buffer
->last_overrun
;
3632 cpu_buffer
->last_overrun
= overwrite
;
3638 /* Update the read_stamp on the first event */
3639 if (reader
&& reader
->read
== 0)
3640 cpu_buffer
->read_stamp
= reader
->page
->time_stamp
;
3642 arch_spin_unlock(&cpu_buffer
->lock
);
3643 local_irq_restore(flags
);
3648 static void rb_advance_reader(struct ring_buffer_per_cpu
*cpu_buffer
)
3650 struct ring_buffer_event
*event
;
3651 struct buffer_page
*reader
;
3654 reader
= rb_get_reader_page(cpu_buffer
);
3656 /* This function should not be called when buffer is empty */
3657 if (RB_WARN_ON(cpu_buffer
, !reader
))
3660 event
= rb_reader_event(cpu_buffer
);
3662 if (event
->type_len
<= RINGBUF_TYPE_DATA_TYPE_LEN_MAX
)
3665 rb_update_read_stamp(cpu_buffer
, event
);
3667 length
= rb_event_length(event
);
3668 cpu_buffer
->reader_page
->read
+= length
;
3671 static void rb_advance_iter(struct ring_buffer_iter
*iter
)
3673 struct ring_buffer_per_cpu
*cpu_buffer
;
3674 struct ring_buffer_event
*event
;
3677 cpu_buffer
= iter
->cpu_buffer
;
3680 * Check if we are at the end of the buffer.
3682 if (iter
->head
>= rb_page_size(iter
->head_page
)) {
3683 /* discarded commits can make the page empty */
3684 if (iter
->head_page
== cpu_buffer
->commit_page
)
3690 event
= rb_iter_head_event(iter
);
3692 length
= rb_event_length(event
);
3695 * This should not be called to advance the header if we are
3696 * at the tail of the buffer.
3698 if (RB_WARN_ON(cpu_buffer
,
3699 (iter
->head_page
== cpu_buffer
->commit_page
) &&
3700 (iter
->head
+ length
> rb_commit_index(cpu_buffer
))))
3703 rb_update_iter_read_stamp(iter
, event
);
3705 iter
->head
+= length
;
3707 /* check for end of page padding */
3708 if ((iter
->head
>= rb_page_size(iter
->head_page
)) &&
3709 (iter
->head_page
!= cpu_buffer
->commit_page
))
3713 static int rb_lost_events(struct ring_buffer_per_cpu
*cpu_buffer
)
3715 return cpu_buffer
->lost_events
;
3718 static struct ring_buffer_event
*
3719 rb_buffer_peek(struct ring_buffer_per_cpu
*cpu_buffer
, u64
*ts
,
3720 unsigned long *lost_events
)
3722 struct ring_buffer_event
*event
;
3723 struct buffer_page
*reader
;
3728 * We repeat when a time extend is encountered.
3729 * Since the time extend is always attached to a data event,
3730 * we should never loop more than once.
3731 * (We never hit the following condition more than twice).
3733 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 2))
3736 reader
= rb_get_reader_page(cpu_buffer
);
3740 event
= rb_reader_event(cpu_buffer
);
3742 switch (event
->type_len
) {
3743 case RINGBUF_TYPE_PADDING
:
3744 if (rb_null_event(event
))
3745 RB_WARN_ON(cpu_buffer
, 1);
3747 * Because the writer could be discarding every
3748 * event it creates (which would probably be bad)
3749 * if we were to go back to "again" then we may never
3750 * catch up, and will trigger the warn on, or lock
3751 * the box. Return the padding, and we will release
3752 * the current locks, and try again.
3756 case RINGBUF_TYPE_TIME_EXTEND
:
3757 /* Internal data, OK to advance */
3758 rb_advance_reader(cpu_buffer
);
3761 case RINGBUF_TYPE_TIME_STAMP
:
3762 /* FIXME: not implemented */
3763 rb_advance_reader(cpu_buffer
);
3766 case RINGBUF_TYPE_DATA
:
3768 *ts
= cpu_buffer
->read_stamp
+ event
->time_delta
;
3769 ring_buffer_normalize_time_stamp(cpu_buffer
->buffer
,
3770 cpu_buffer
->cpu
, ts
);
3773 *lost_events
= rb_lost_events(cpu_buffer
);
3782 EXPORT_SYMBOL_GPL(ring_buffer_peek
);
3784 static struct ring_buffer_event
*
3785 rb_iter_peek(struct ring_buffer_iter
*iter
, u64
*ts
)
3787 struct ring_buffer
*buffer
;
3788 struct ring_buffer_per_cpu
*cpu_buffer
;
3789 struct ring_buffer_event
*event
;
3792 cpu_buffer
= iter
->cpu_buffer
;
3793 buffer
= cpu_buffer
->buffer
;
3796 * Check if someone performed a consuming read to
3797 * the buffer. A consuming read invalidates the iterator
3798 * and we need to reset the iterator in this case.
3800 if (unlikely(iter
->cache_read
!= cpu_buffer
->read
||
3801 iter
->cache_reader_page
!= cpu_buffer
->reader_page
))
3802 rb_iter_reset(iter
);
3805 if (ring_buffer_iter_empty(iter
))
3809 * We repeat when a time extend is encountered or we hit
3810 * the end of the page. Since the time extend is always attached
3811 * to a data event, we should never loop more than three times.
3812 * Once for going to next page, once on time extend, and
3813 * finally once to get the event.
3814 * (We never hit the following condition more than thrice).
3816 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 3))
3819 if (rb_per_cpu_empty(cpu_buffer
))
3822 if (iter
->head
>= local_read(&iter
->head_page
->page
->commit
)) {
3827 event
= rb_iter_head_event(iter
);
3829 switch (event
->type_len
) {
3830 case RINGBUF_TYPE_PADDING
:
3831 if (rb_null_event(event
)) {
3835 rb_advance_iter(iter
);
3838 case RINGBUF_TYPE_TIME_EXTEND
:
3839 /* Internal data, OK to advance */
3840 rb_advance_iter(iter
);
3843 case RINGBUF_TYPE_TIME_STAMP
:
3844 /* FIXME: not implemented */
3845 rb_advance_iter(iter
);
3848 case RINGBUF_TYPE_DATA
:
3850 *ts
= iter
->read_stamp
+ event
->time_delta
;
3851 ring_buffer_normalize_time_stamp(buffer
,
3852 cpu_buffer
->cpu
, ts
);
3862 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek
);
3864 static inline int rb_ok_to_lock(void)
3867 * If an NMI die dumps out the content of the ring buffer
3868 * do not grab locks. We also permanently disable the ring
3869 * buffer too. A one time deal is all you get from reading
3870 * the ring buffer from an NMI.
3872 if (likely(!in_nmi()))
3875 tracing_off_permanent();
3880 * ring_buffer_peek - peek at the next event to be read
3881 * @buffer: The ring buffer to read
3882 * @cpu: The cpu to peak at
3883 * @ts: The timestamp counter of this event.
3884 * @lost_events: a variable to store if events were lost (may be NULL)
3886 * This will return the event that will be read next, but does
3887 * not consume the data.
3889 struct ring_buffer_event
*
3890 ring_buffer_peek(struct ring_buffer
*buffer
, int cpu
, u64
*ts
,
3891 unsigned long *lost_events
)
3893 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
3894 struct ring_buffer_event
*event
;
3895 unsigned long flags
;
3898 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3901 dolock
= rb_ok_to_lock();
3903 local_irq_save(flags
);
3905 raw_spin_lock(&cpu_buffer
->reader_lock
);
3906 event
= rb_buffer_peek(cpu_buffer
, ts
, lost_events
);
3907 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3908 rb_advance_reader(cpu_buffer
);
3910 raw_spin_unlock(&cpu_buffer
->reader_lock
);
3911 local_irq_restore(flags
);
3913 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3920 * ring_buffer_iter_peek - peek at the next event to be read
3921 * @iter: The ring buffer iterator
3922 * @ts: The timestamp counter of this event.
3924 * This will return the event that will be read next, but does
3925 * not increment the iterator.
3927 struct ring_buffer_event
*
3928 ring_buffer_iter_peek(struct ring_buffer_iter
*iter
, u64
*ts
)
3930 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
3931 struct ring_buffer_event
*event
;
3932 unsigned long flags
;
3935 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3936 event
= rb_iter_peek(iter
, ts
);
3937 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3939 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3946 * ring_buffer_consume - return an event and consume it
3947 * @buffer: The ring buffer to get the next event from
3948 * @cpu: the cpu to read the buffer from
3949 * @ts: a variable to store the timestamp (may be NULL)
3950 * @lost_events: a variable to store if events were lost (may be NULL)
3952 * Returns the next event in the ring buffer, and that event is consumed.
3953 * Meaning, that sequential reads will keep returning a different event,
3954 * and eventually empty the ring buffer if the producer is slower.
3956 struct ring_buffer_event
*
3957 ring_buffer_consume(struct ring_buffer
*buffer
, int cpu
, u64
*ts
,
3958 unsigned long *lost_events
)
3960 struct ring_buffer_per_cpu
*cpu_buffer
;
3961 struct ring_buffer_event
*event
= NULL
;
3962 unsigned long flags
;
3965 dolock
= rb_ok_to_lock();
3968 /* might be called in atomic */
3971 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3974 cpu_buffer
= buffer
->buffers
[cpu
];
3975 local_irq_save(flags
);
3977 raw_spin_lock(&cpu_buffer
->reader_lock
);
3979 event
= rb_buffer_peek(cpu_buffer
, ts
, lost_events
);
3981 cpu_buffer
->lost_events
= 0;
3982 rb_advance_reader(cpu_buffer
);
3986 raw_spin_unlock(&cpu_buffer
->reader_lock
);
3987 local_irq_restore(flags
);
3992 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3997 EXPORT_SYMBOL_GPL(ring_buffer_consume
);
4000 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4001 * @buffer: The ring buffer to read from
4002 * @cpu: The cpu buffer to iterate over
4004 * This performs the initial preparations necessary to iterate
4005 * through the buffer. Memory is allocated, buffer recording
4006 * is disabled, and the iterator pointer is returned to the caller.
4008 * Disabling buffer recordng prevents the reading from being
4009 * corrupted. This is not a consuming read, so a producer is not
4012 * After a sequence of ring_buffer_read_prepare calls, the user is
4013 * expected to make at least one call to ring_buffer_prepare_sync.
4014 * Afterwards, ring_buffer_read_start is invoked to get things going
4017 * This overall must be paired with ring_buffer_finish.
4019 struct ring_buffer_iter
*
4020 ring_buffer_read_prepare(struct ring_buffer
*buffer
, int cpu
)
4022 struct ring_buffer_per_cpu
*cpu_buffer
;
4023 struct ring_buffer_iter
*iter
;
4025 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4028 iter
= kmalloc(sizeof(*iter
), GFP_KERNEL
);
4032 cpu_buffer
= buffer
->buffers
[cpu
];
4034 iter
->cpu_buffer
= cpu_buffer
;
4036 atomic_inc(&buffer
->resize_disabled
);
4037 atomic_inc(&cpu_buffer
->record_disabled
);
4041 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare
);
4044 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4046 * All previously invoked ring_buffer_read_prepare calls to prepare
4047 * iterators will be synchronized. Afterwards, read_buffer_read_start
4048 * calls on those iterators are allowed.
4051 ring_buffer_read_prepare_sync(void)
4053 synchronize_sched();
4055 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync
);
4058 * ring_buffer_read_start - start a non consuming read of the buffer
4059 * @iter: The iterator returned by ring_buffer_read_prepare
4061 * This finalizes the startup of an iteration through the buffer.
4062 * The iterator comes from a call to ring_buffer_read_prepare and
4063 * an intervening ring_buffer_read_prepare_sync must have been
4066 * Must be paired with ring_buffer_finish.
4069 ring_buffer_read_start(struct ring_buffer_iter
*iter
)
4071 struct ring_buffer_per_cpu
*cpu_buffer
;
4072 unsigned long flags
;
4077 cpu_buffer
= iter
->cpu_buffer
;
4079 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4080 arch_spin_lock(&cpu_buffer
->lock
);
4081 rb_iter_reset(iter
);
4082 arch_spin_unlock(&cpu_buffer
->lock
);
4083 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4085 EXPORT_SYMBOL_GPL(ring_buffer_read_start
);
4088 * ring_buffer_finish - finish reading the iterator of the buffer
4089 * @iter: The iterator retrieved by ring_buffer_start
4091 * This re-enables the recording to the buffer, and frees the
4095 ring_buffer_read_finish(struct ring_buffer_iter
*iter
)
4097 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
4098 unsigned long flags
;
4101 * Ring buffer is disabled from recording, here's a good place
4102 * to check the integrity of the ring buffer.
4103 * Must prevent readers from trying to read, as the check
4104 * clears the HEAD page and readers require it.
4106 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4107 rb_check_pages(cpu_buffer
);
4108 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4110 atomic_dec(&cpu_buffer
->record_disabled
);
4111 atomic_dec(&cpu_buffer
->buffer
->resize_disabled
);
4114 EXPORT_SYMBOL_GPL(ring_buffer_read_finish
);
4117 * ring_buffer_read - read the next item in the ring buffer by the iterator
4118 * @iter: The ring buffer iterator
4119 * @ts: The time stamp of the event read.
4121 * This reads the next event in the ring buffer and increments the iterator.
4123 struct ring_buffer_event
*
4124 ring_buffer_read(struct ring_buffer_iter
*iter
, u64
*ts
)
4126 struct ring_buffer_event
*event
;
4127 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
4128 unsigned long flags
;
4130 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4132 event
= rb_iter_peek(iter
, ts
);
4136 if (event
->type_len
== RINGBUF_TYPE_PADDING
)
4139 rb_advance_iter(iter
);
4141 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4145 EXPORT_SYMBOL_GPL(ring_buffer_read
);
4148 * ring_buffer_size - return the size of the ring buffer (in bytes)
4149 * @buffer: The ring buffer.
4151 unsigned long ring_buffer_size(struct ring_buffer
*buffer
, int cpu
)
4154 * Earlier, this method returned
4155 * BUF_PAGE_SIZE * buffer->nr_pages
4156 * Since the nr_pages field is now removed, we have converted this to
4157 * return the per cpu buffer value.
4159 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4162 return BUF_PAGE_SIZE
* buffer
->buffers
[cpu
]->nr_pages
;
4164 EXPORT_SYMBOL_GPL(ring_buffer_size
);
4167 rb_reset_cpu(struct ring_buffer_per_cpu
*cpu_buffer
)
4169 rb_head_page_deactivate(cpu_buffer
);
4171 cpu_buffer
->head_page
4172 = list_entry(cpu_buffer
->pages
, struct buffer_page
, list
);
4173 local_set(&cpu_buffer
->head_page
->write
, 0);
4174 local_set(&cpu_buffer
->head_page
->entries
, 0);
4175 local_set(&cpu_buffer
->head_page
->page
->commit
, 0);
4177 cpu_buffer
->head_page
->read
= 0;
4179 cpu_buffer
->tail_page
= cpu_buffer
->head_page
;
4180 cpu_buffer
->commit_page
= cpu_buffer
->head_page
;
4182 INIT_LIST_HEAD(&cpu_buffer
->reader_page
->list
);
4183 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
4184 local_set(&cpu_buffer
->reader_page
->write
, 0);
4185 local_set(&cpu_buffer
->reader_page
->entries
, 0);
4186 local_set(&cpu_buffer
->reader_page
->page
->commit
, 0);
4187 cpu_buffer
->reader_page
->read
= 0;
4189 local_set(&cpu_buffer
->entries_bytes
, 0);
4190 local_set(&cpu_buffer
->overrun
, 0);
4191 local_set(&cpu_buffer
->commit_overrun
, 0);
4192 local_set(&cpu_buffer
->dropped_events
, 0);
4193 local_set(&cpu_buffer
->entries
, 0);
4194 local_set(&cpu_buffer
->committing
, 0);
4195 local_set(&cpu_buffer
->commits
, 0);
4196 cpu_buffer
->read
= 0;
4197 cpu_buffer
->read_bytes
= 0;
4199 cpu_buffer
->write_stamp
= 0;
4200 cpu_buffer
->read_stamp
= 0;
4202 cpu_buffer
->lost_events
= 0;
4203 cpu_buffer
->last_overrun
= 0;
4205 rb_head_page_activate(cpu_buffer
);
4209 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4210 * @buffer: The ring buffer to reset a per cpu buffer of
4211 * @cpu: The CPU buffer to be reset
4213 void ring_buffer_reset_cpu(struct ring_buffer
*buffer
, int cpu
)
4215 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
4216 unsigned long flags
;
4218 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4221 atomic_inc(&buffer
->resize_disabled
);
4222 atomic_inc(&cpu_buffer
->record_disabled
);
4224 /* Make sure all commits have finished */
4225 synchronize_sched();
4227 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4229 if (RB_WARN_ON(cpu_buffer
, local_read(&cpu_buffer
->committing
)))
4232 arch_spin_lock(&cpu_buffer
->lock
);
4234 rb_reset_cpu(cpu_buffer
);
4236 arch_spin_unlock(&cpu_buffer
->lock
);
4239 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4241 atomic_dec(&cpu_buffer
->record_disabled
);
4242 atomic_dec(&buffer
->resize_disabled
);
4244 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu
);
4247 * ring_buffer_reset - reset a ring buffer
4248 * @buffer: The ring buffer to reset all cpu buffers
4250 void ring_buffer_reset(struct ring_buffer
*buffer
)
4254 for_each_buffer_cpu(buffer
, cpu
)
4255 ring_buffer_reset_cpu(buffer
, cpu
);
4257 EXPORT_SYMBOL_GPL(ring_buffer_reset
);
4260 * rind_buffer_empty - is the ring buffer empty?
4261 * @buffer: The ring buffer to test
4263 int ring_buffer_empty(struct ring_buffer
*buffer
)
4265 struct ring_buffer_per_cpu
*cpu_buffer
;
4266 unsigned long flags
;
4271 dolock
= rb_ok_to_lock();
4273 /* yes this is racy, but if you don't like the race, lock the buffer */
4274 for_each_buffer_cpu(buffer
, cpu
) {
4275 cpu_buffer
= buffer
->buffers
[cpu
];
4276 local_irq_save(flags
);
4278 raw_spin_lock(&cpu_buffer
->reader_lock
);
4279 ret
= rb_per_cpu_empty(cpu_buffer
);
4281 raw_spin_unlock(&cpu_buffer
->reader_lock
);
4282 local_irq_restore(flags
);
4290 EXPORT_SYMBOL_GPL(ring_buffer_empty
);
4293 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4294 * @buffer: The ring buffer
4295 * @cpu: The CPU buffer to test
4297 int ring_buffer_empty_cpu(struct ring_buffer
*buffer
, int cpu
)
4299 struct ring_buffer_per_cpu
*cpu_buffer
;
4300 unsigned long flags
;
4304 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4307 dolock
= rb_ok_to_lock();
4309 cpu_buffer
= buffer
->buffers
[cpu
];
4310 local_irq_save(flags
);
4312 raw_spin_lock(&cpu_buffer
->reader_lock
);
4313 ret
= rb_per_cpu_empty(cpu_buffer
);
4315 raw_spin_unlock(&cpu_buffer
->reader_lock
);
4316 local_irq_restore(flags
);
4320 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu
);
4322 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4324 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4325 * @buffer_a: One buffer to swap with
4326 * @buffer_b: The other buffer to swap with
4328 * This function is useful for tracers that want to take a "snapshot"
4329 * of a CPU buffer and has another back up buffer lying around.
4330 * it is expected that the tracer handles the cpu buffer not being
4331 * used at the moment.
4333 int ring_buffer_swap_cpu(struct ring_buffer
*buffer_a
,
4334 struct ring_buffer
*buffer_b
, int cpu
)
4336 struct ring_buffer_per_cpu
*cpu_buffer_a
;
4337 struct ring_buffer_per_cpu
*cpu_buffer_b
;
4340 if (!cpumask_test_cpu(cpu
, buffer_a
->cpumask
) ||
4341 !cpumask_test_cpu(cpu
, buffer_b
->cpumask
))
4344 cpu_buffer_a
= buffer_a
->buffers
[cpu
];
4345 cpu_buffer_b
= buffer_b
->buffers
[cpu
];
4347 /* At least make sure the two buffers are somewhat the same */
4348 if (cpu_buffer_a
->nr_pages
!= cpu_buffer_b
->nr_pages
)
4353 if (ring_buffer_flags
!= RB_BUFFERS_ON
)
4356 if (atomic_read(&buffer_a
->record_disabled
))
4359 if (atomic_read(&buffer_b
->record_disabled
))
4362 if (atomic_read(&cpu_buffer_a
->record_disabled
))
4365 if (atomic_read(&cpu_buffer_b
->record_disabled
))
4369 * We can't do a synchronize_sched here because this
4370 * function can be called in atomic context.
4371 * Normally this will be called from the same CPU as cpu.
4372 * If not it's up to the caller to protect this.
4374 atomic_inc(&cpu_buffer_a
->record_disabled
);
4375 atomic_inc(&cpu_buffer_b
->record_disabled
);
4378 if (local_read(&cpu_buffer_a
->committing
))
4380 if (local_read(&cpu_buffer_b
->committing
))
4383 buffer_a
->buffers
[cpu
] = cpu_buffer_b
;
4384 buffer_b
->buffers
[cpu
] = cpu_buffer_a
;
4386 cpu_buffer_b
->buffer
= buffer_a
;
4387 cpu_buffer_a
->buffer
= buffer_b
;
4392 atomic_dec(&cpu_buffer_a
->record_disabled
);
4393 atomic_dec(&cpu_buffer_b
->record_disabled
);
4397 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu
);
4398 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4401 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4402 * @buffer: the buffer to allocate for.
4404 * This function is used in conjunction with ring_buffer_read_page.
4405 * When reading a full page from the ring buffer, these functions
4406 * can be used to speed up the process. The calling function should
4407 * allocate a few pages first with this function. Then when it
4408 * needs to get pages from the ring buffer, it passes the result
4409 * of this function into ring_buffer_read_page, which will swap
4410 * the page that was allocated, with the read page of the buffer.
4413 * The page allocated, or NULL on error.
4415 void *ring_buffer_alloc_read_page(struct ring_buffer
*buffer
, int cpu
)
4417 struct buffer_data_page
*bpage
;
4420 page
= alloc_pages_node(cpu_to_node(cpu
),
4421 GFP_KERNEL
| __GFP_NORETRY
, 0);
4425 bpage
= page_address(page
);
4427 rb_init_page(bpage
);
4431 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page
);
4434 * ring_buffer_free_read_page - free an allocated read page
4435 * @buffer: the buffer the page was allocate for
4436 * @data: the page to free
4438 * Free a page allocated from ring_buffer_alloc_read_page.
4440 void ring_buffer_free_read_page(struct ring_buffer
*buffer
, void *data
)
4442 free_page((unsigned long)data
);
4444 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page
);
4447 * ring_buffer_read_page - extract a page from the ring buffer
4448 * @buffer: buffer to extract from
4449 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4450 * @len: amount to extract
4451 * @cpu: the cpu of the buffer to extract
4452 * @full: should the extraction only happen when the page is full.
4454 * This function will pull out a page from the ring buffer and consume it.
4455 * @data_page must be the address of the variable that was returned
4456 * from ring_buffer_alloc_read_page. This is because the page might be used
4457 * to swap with a page in the ring buffer.
4460 * rpage = ring_buffer_alloc_read_page(buffer);
4463 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4465 * process_page(rpage, ret);
4467 * When @full is set, the function will not return true unless
4468 * the writer is off the reader page.
4470 * Note: it is up to the calling functions to handle sleeps and wakeups.
4471 * The ring buffer can be used anywhere in the kernel and can not
4472 * blindly call wake_up. The layer that uses the ring buffer must be
4473 * responsible for that.
4476 * >=0 if data has been transferred, returns the offset of consumed data.
4477 * <0 if no data has been transferred.
4479 int ring_buffer_read_page(struct ring_buffer
*buffer
,
4480 void **data_page
, size_t len
, int cpu
, int full
)
4482 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
4483 struct ring_buffer_event
*event
;
4484 struct buffer_data_page
*bpage
;
4485 struct buffer_page
*reader
;
4486 unsigned long missed_events
;
4487 unsigned long flags
;
4488 unsigned int commit
;
4493 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4497 * If len is not big enough to hold the page header, then
4498 * we can not copy anything.
4500 if (len
<= BUF_PAGE_HDR_SIZE
)
4503 len
-= BUF_PAGE_HDR_SIZE
;
4512 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4514 reader
= rb_get_reader_page(cpu_buffer
);
4518 event
= rb_reader_event(cpu_buffer
);
4520 read
= reader
->read
;
4521 commit
= rb_page_commit(reader
);
4523 /* Check if any events were dropped */
4524 missed_events
= cpu_buffer
->lost_events
;
4527 * If this page has been partially read or
4528 * if len is not big enough to read the rest of the page or
4529 * a writer is still on the page, then
4530 * we must copy the data from the page to the buffer.
4531 * Otherwise, we can simply swap the page with the one passed in.
4533 if (read
|| (len
< (commit
- read
)) ||
4534 cpu_buffer
->reader_page
== cpu_buffer
->commit_page
) {
4535 struct buffer_data_page
*rpage
= cpu_buffer
->reader_page
->page
;
4536 unsigned int rpos
= read
;
4537 unsigned int pos
= 0;
4543 if (len
> (commit
- read
))
4544 len
= (commit
- read
);
4546 /* Always keep the time extend and data together */
4547 size
= rb_event_ts_length(event
);
4552 /* save the current timestamp, since the user will need it */
4553 save_timestamp
= cpu_buffer
->read_stamp
;
4555 /* Need to copy one event at a time */
4557 /* We need the size of one event, because
4558 * rb_advance_reader only advances by one event,
4559 * whereas rb_event_ts_length may include the size of
4560 * one or two events.
4561 * We have already ensured there's enough space if this
4562 * is a time extend. */
4563 size
= rb_event_length(event
);
4564 memcpy(bpage
->data
+ pos
, rpage
->data
+ rpos
, size
);
4568 rb_advance_reader(cpu_buffer
);
4569 rpos
= reader
->read
;
4575 event
= rb_reader_event(cpu_buffer
);
4576 /* Always keep the time extend and data together */
4577 size
= rb_event_ts_length(event
);
4578 } while (len
>= size
);
4581 local_set(&bpage
->commit
, pos
);
4582 bpage
->time_stamp
= save_timestamp
;
4584 /* we copied everything to the beginning */
4587 /* update the entry counter */
4588 cpu_buffer
->read
+= rb_page_entries(reader
);
4589 cpu_buffer
->read_bytes
+= BUF_PAGE_SIZE
;
4591 /* swap the pages */
4592 rb_init_page(bpage
);
4593 bpage
= reader
->page
;
4594 reader
->page
= *data_page
;
4595 local_set(&reader
->write
, 0);
4596 local_set(&reader
->entries
, 0);
4601 * Use the real_end for the data size,
4602 * This gives us a chance to store the lost events
4605 if (reader
->real_end
)
4606 local_set(&bpage
->commit
, reader
->real_end
);
4610 cpu_buffer
->lost_events
= 0;
4612 commit
= local_read(&bpage
->commit
);
4614 * Set a flag in the commit field if we lost events
4616 if (missed_events
) {
4617 /* If there is room at the end of the page to save the
4618 * missed events, then record it there.
4620 if (BUF_PAGE_SIZE
- commit
>= sizeof(missed_events
)) {
4621 memcpy(&bpage
->data
[commit
], &missed_events
,
4622 sizeof(missed_events
));
4623 local_add(RB_MISSED_STORED
, &bpage
->commit
);
4624 commit
+= sizeof(missed_events
);
4626 local_add(RB_MISSED_EVENTS
, &bpage
->commit
);
4630 * This page may be off to user land. Zero it out here.
4632 if (commit
< BUF_PAGE_SIZE
)
4633 memset(&bpage
->data
[commit
], 0, BUF_PAGE_SIZE
- commit
);
4636 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4641 EXPORT_SYMBOL_GPL(ring_buffer_read_page
);
4643 #ifdef CONFIG_HOTPLUG_CPU
4644 static int rb_cpu_notify(struct notifier_block
*self
,
4645 unsigned long action
, void *hcpu
)
4647 struct ring_buffer
*buffer
=
4648 container_of(self
, struct ring_buffer
, cpu_notify
);
4649 long cpu
= (long)hcpu
;
4652 unsigned long nr_pages
;
4655 case CPU_UP_PREPARE
:
4656 case CPU_UP_PREPARE_FROZEN
:
4657 if (cpumask_test_cpu(cpu
, buffer
->cpumask
))
4662 /* check if all cpu sizes are same */
4663 for_each_buffer_cpu(buffer
, cpu_i
) {
4664 /* fill in the size from first enabled cpu */
4666 nr_pages
= buffer
->buffers
[cpu_i
]->nr_pages
;
4667 if (nr_pages
!= buffer
->buffers
[cpu_i
]->nr_pages
) {
4672 /* allocate minimum pages, user can later expand it */
4675 buffer
->buffers
[cpu
] =
4676 rb_allocate_cpu_buffer(buffer
, nr_pages
, cpu
);
4677 if (!buffer
->buffers
[cpu
]) {
4678 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4683 cpumask_set_cpu(cpu
, buffer
->cpumask
);
4685 case CPU_DOWN_PREPARE
:
4686 case CPU_DOWN_PREPARE_FROZEN
:
4689 * If we were to free the buffer, then the user would
4690 * lose any trace that was in the buffer.
4700 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4702 * This is a basic integrity check of the ring buffer.
4703 * Late in the boot cycle this test will run when configured in.
4704 * It will kick off a thread per CPU that will go into a loop
4705 * writing to the per cpu ring buffer various sizes of data.
4706 * Some of the data will be large items, some small.
4708 * Another thread is created that goes into a spin, sending out
4709 * IPIs to the other CPUs to also write into the ring buffer.
4710 * this is to test the nesting ability of the buffer.
4712 * Basic stats are recorded and reported. If something in the
4713 * ring buffer should happen that's not expected, a big warning
4714 * is displayed and all ring buffers are disabled.
4716 static struct task_struct
*rb_threads
[NR_CPUS
] __initdata
;
4718 struct rb_test_data
{
4719 struct ring_buffer
*buffer
;
4720 unsigned long events
;
4721 unsigned long bytes_written
;
4722 unsigned long bytes_alloc
;
4723 unsigned long bytes_dropped
;
4724 unsigned long events_nested
;
4725 unsigned long bytes_written_nested
;
4726 unsigned long bytes_alloc_nested
;
4727 unsigned long bytes_dropped_nested
;
4728 int min_size_nested
;
4729 int max_size_nested
;
4736 static struct rb_test_data rb_data
[NR_CPUS
] __initdata
;
4739 #define RB_TEST_BUFFER_SIZE 1048576
4741 static char rb_string
[] __initdata
=
4742 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4743 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4744 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4746 static bool rb_test_started __initdata
;
4753 static __init
int rb_write_something(struct rb_test_data
*data
, bool nested
)
4755 struct ring_buffer_event
*event
;
4756 struct rb_item
*item
;
4763 /* Have nested writes different that what is written */
4764 cnt
= data
->cnt
+ (nested
? 27 : 0);
4766 /* Multiply cnt by ~e, to make some unique increment */
4767 size
= (data
->cnt
* 68 / 25) % (sizeof(rb_string
) - 1);
4769 len
= size
+ sizeof(struct rb_item
);
4771 started
= rb_test_started
;
4772 /* read rb_test_started before checking buffer enabled */
4775 event
= ring_buffer_lock_reserve(data
->buffer
, len
);
4777 /* Ignore dropped events before test starts. */
4780 data
->bytes_dropped
+= len
;
4782 data
->bytes_dropped_nested
+= len
;
4787 event_len
= ring_buffer_event_length(event
);
4789 if (RB_WARN_ON(data
->buffer
, event_len
< len
))
4792 item
= ring_buffer_event_data(event
);
4794 memcpy(item
->str
, rb_string
, size
);
4797 data
->bytes_alloc_nested
+= event_len
;
4798 data
->bytes_written_nested
+= len
;
4799 data
->events_nested
++;
4800 if (!data
->min_size_nested
|| len
< data
->min_size_nested
)
4801 data
->min_size_nested
= len
;
4802 if (len
> data
->max_size_nested
)
4803 data
->max_size_nested
= len
;
4805 data
->bytes_alloc
+= event_len
;
4806 data
->bytes_written
+= len
;
4808 if (!data
->min_size
|| len
< data
->min_size
)
4809 data
->max_size
= len
;
4810 if (len
> data
->max_size
)
4811 data
->max_size
= len
;
4815 ring_buffer_unlock_commit(data
->buffer
, event
);
4820 static __init
int rb_test(void *arg
)
4822 struct rb_test_data
*data
= arg
;
4824 while (!kthread_should_stop()) {
4825 rb_write_something(data
, false);
4828 set_current_state(TASK_INTERRUPTIBLE
);
4829 /* Now sleep between a min of 100-300us and a max of 1ms */
4830 usleep_range(((data
->cnt
% 3) + 1) * 100, 1000);
4836 static __init
void rb_ipi(void *ignore
)
4838 struct rb_test_data
*data
;
4839 int cpu
= smp_processor_id();
4841 data
= &rb_data
[cpu
];
4842 rb_write_something(data
, true);
4845 static __init
int rb_hammer_test(void *arg
)
4847 while (!kthread_should_stop()) {
4849 /* Send an IPI to all cpus to write data! */
4850 smp_call_function(rb_ipi
, NULL
, 1);
4851 /* No sleep, but for non preempt, let others run */
4858 static __init
int test_ringbuffer(void)
4860 struct task_struct
*rb_hammer
;
4861 struct ring_buffer
*buffer
;
4865 pr_info("Running ring buffer tests...\n");
4867 buffer
= ring_buffer_alloc(RB_TEST_BUFFER_SIZE
, RB_FL_OVERWRITE
);
4868 if (WARN_ON(!buffer
))
4871 /* Disable buffer so that threads can't write to it yet */
4872 ring_buffer_record_off(buffer
);
4874 for_each_online_cpu(cpu
) {
4875 rb_data
[cpu
].buffer
= buffer
;
4876 rb_data
[cpu
].cpu
= cpu
;
4877 rb_data
[cpu
].cnt
= cpu
;
4878 rb_threads
[cpu
] = kthread_create(rb_test
, &rb_data
[cpu
],
4879 "rbtester/%d", cpu
);
4880 if (WARN_ON(IS_ERR(rb_threads
[cpu
]))) {
4881 pr_cont("FAILED\n");
4882 ret
= PTR_ERR(rb_threads
[cpu
]);
4886 kthread_bind(rb_threads
[cpu
], cpu
);
4887 wake_up_process(rb_threads
[cpu
]);
4890 /* Now create the rb hammer! */
4891 rb_hammer
= kthread_run(rb_hammer_test
, NULL
, "rbhammer");
4892 if (WARN_ON(IS_ERR(rb_hammer
))) {
4893 pr_cont("FAILED\n");
4894 ret
= PTR_ERR(rb_hammer
);
4898 ring_buffer_record_on(buffer
);
4900 * Show buffer is enabled before setting rb_test_started.
4901 * Yes there's a small race window where events could be
4902 * dropped and the thread wont catch it. But when a ring
4903 * buffer gets enabled, there will always be some kind of
4904 * delay before other CPUs see it. Thus, we don't care about
4905 * those dropped events. We care about events dropped after
4906 * the threads see that the buffer is active.
4909 rb_test_started
= true;
4911 set_current_state(TASK_INTERRUPTIBLE
);
4912 /* Just run for 10 seconds */;
4913 schedule_timeout(10 * HZ
);
4915 kthread_stop(rb_hammer
);
4918 for_each_online_cpu(cpu
) {
4919 if (!rb_threads
[cpu
])
4921 kthread_stop(rb_threads
[cpu
]);
4924 ring_buffer_free(buffer
);
4929 pr_info("finished\n");
4930 for_each_online_cpu(cpu
) {
4931 struct ring_buffer_event
*event
;
4932 struct rb_test_data
*data
= &rb_data
[cpu
];
4933 struct rb_item
*item
;
4934 unsigned long total_events
;
4935 unsigned long total_dropped
;
4936 unsigned long total_written
;
4937 unsigned long total_alloc
;
4938 unsigned long total_read
= 0;
4939 unsigned long total_size
= 0;
4940 unsigned long total_len
= 0;
4941 unsigned long total_lost
= 0;
4944 int small_event_size
;
4948 total_events
= data
->events
+ data
->events_nested
;
4949 total_written
= data
->bytes_written
+ data
->bytes_written_nested
;
4950 total_alloc
= data
->bytes_alloc
+ data
->bytes_alloc_nested
;
4951 total_dropped
= data
->bytes_dropped
+ data
->bytes_dropped_nested
;
4953 big_event_size
= data
->max_size
+ data
->max_size_nested
;
4954 small_event_size
= data
->min_size
+ data
->min_size_nested
;
4956 pr_info("CPU %d:\n", cpu
);
4957 pr_info(" events: %ld\n", total_events
);
4958 pr_info(" dropped bytes: %ld\n", total_dropped
);
4959 pr_info(" alloced bytes: %ld\n", total_alloc
);
4960 pr_info(" written bytes: %ld\n", total_written
);
4961 pr_info(" biggest event: %d\n", big_event_size
);
4962 pr_info(" smallest event: %d\n", small_event_size
);
4964 if (RB_WARN_ON(buffer
, total_dropped
))
4969 while ((event
= ring_buffer_consume(buffer
, cpu
, NULL
, &lost
))) {
4971 item
= ring_buffer_event_data(event
);
4972 total_len
+= ring_buffer_event_length(event
);
4973 total_size
+= item
->size
+ sizeof(struct rb_item
);
4974 if (memcmp(&item
->str
[0], rb_string
, item
->size
) != 0) {
4975 pr_info("FAILED!\n");
4976 pr_info("buffer had: %.*s\n", item
->size
, item
->str
);
4977 pr_info("expected: %.*s\n", item
->size
, rb_string
);
4978 RB_WARN_ON(buffer
, 1);
4989 pr_info(" read events: %ld\n", total_read
);
4990 pr_info(" lost events: %ld\n", total_lost
);
4991 pr_info(" total events: %ld\n", total_lost
+ total_read
);
4992 pr_info(" recorded len bytes: %ld\n", total_len
);
4993 pr_info(" recorded size bytes: %ld\n", total_size
);
4995 pr_info(" With dropped events, record len and size may not match\n"
4996 " alloced and written from above\n");
4998 if (RB_WARN_ON(buffer
, total_len
!= total_alloc
||
4999 total_size
!= total_written
))
5002 if (RB_WARN_ON(buffer
, total_lost
+ total_read
!= total_events
))
5008 pr_info("Ring buffer PASSED!\n");
5010 ring_buffer_free(buffer
);
5014 late_initcall(test_ringbuffer
);
5015 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */