4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 #include <linux/trace_events.h>
7 #include <linux/ring_buffer.h>
8 #include <linux/trace_clock.h>
9 #include <linux/sched/clock.h>
10 #include <linux/trace_seq.h>
11 #include <linux/spinlock.h>
12 #include <linux/irq_work.h>
13 #include <linux/uaccess.h>
14 #include <linux/hardirq.h>
15 #include <linux/kthread.h> /* for self test */
16 #include <linux/module.h>
17 #include <linux/percpu.h>
18 #include <linux/mutex.h>
19 #include <linux/delay.h>
20 #include <linux/slab.h>
21 #include <linux/init.h>
22 #include <linux/hash.h>
23 #include <linux/list.h>
24 #include <linux/cpu.h>
26 #include <asm/local.h>
28 static void update_pages_handler(struct work_struct
*work
);
31 * The ring buffer header is special. We must manually up keep it.
33 int ring_buffer_print_entry_header(struct trace_seq
*s
)
35 trace_seq_puts(s
, "# compressed entry header\n");
36 trace_seq_puts(s
, "\ttype_len : 5 bits\n");
37 trace_seq_puts(s
, "\ttime_delta : 27 bits\n");
38 trace_seq_puts(s
, "\tarray : 32 bits\n");
39 trace_seq_putc(s
, '\n');
40 trace_seq_printf(s
, "\tpadding : type == %d\n",
41 RINGBUF_TYPE_PADDING
);
42 trace_seq_printf(s
, "\ttime_extend : type == %d\n",
43 RINGBUF_TYPE_TIME_EXTEND
);
44 trace_seq_printf(s
, "\tdata max type_len == %d\n",
45 RINGBUF_TYPE_DATA_TYPE_LEN_MAX
);
47 return !trace_seq_has_overflowed(s
);
51 * The ring buffer is made up of a list of pages. A separate list of pages is
52 * allocated for each CPU. A writer may only write to a buffer that is
53 * associated with the CPU it is currently executing on. A reader may read
54 * from any per cpu buffer.
56 * The reader is special. For each per cpu buffer, the reader has its own
57 * reader page. When a reader has read the entire reader page, this reader
58 * page is swapped with another page in the ring buffer.
60 * Now, as long as the writer is off the reader page, the reader can do what
61 * ever it wants with that page. The writer will never write to that page
62 * again (as long as it is out of the ring buffer).
64 * Here's some silly ASCII art.
67 * |reader| RING BUFFER
69 * +------+ +---+ +---+ +---+
78 * |reader| RING BUFFER
79 * |page |------------------v
80 * +------+ +---+ +---+ +---+
89 * |reader| RING BUFFER
90 * |page |------------------v
91 * +------+ +---+ +---+ +---+
96 * +------------------------------+
100 * |buffer| RING BUFFER
101 * |page |------------------v
102 * +------+ +---+ +---+ +---+
104 * | New +---+ +---+ +---+
107 * +------------------------------+
110 * After we make this swap, the reader can hand this page off to the splice
111 * code and be done with it. It can even allocate a new page if it needs to
112 * and swap that into the ring buffer.
114 * We will be using cmpxchg soon to make all this lockless.
118 /* Used for individual buffers (after the counter) */
119 #define RB_BUFFER_OFF (1 << 20)
121 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
123 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
124 #define RB_ALIGNMENT 4U
125 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
126 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
128 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
129 # define RB_FORCE_8BYTE_ALIGNMENT 0
130 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
132 # define RB_FORCE_8BYTE_ALIGNMENT 1
133 # define RB_ARCH_ALIGNMENT 8U
136 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
138 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
139 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
142 RB_LEN_TIME_EXTEND
= 8,
143 RB_LEN_TIME_STAMP
= 16,
146 #define skip_time_extend(event) \
147 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
149 static inline int rb_null_event(struct ring_buffer_event
*event
)
151 return event
->type_len
== RINGBUF_TYPE_PADDING
&& !event
->time_delta
;
154 static void rb_event_set_padding(struct ring_buffer_event
*event
)
156 /* padding has a NULL time_delta */
157 event
->type_len
= RINGBUF_TYPE_PADDING
;
158 event
->time_delta
= 0;
162 rb_event_data_length(struct ring_buffer_event
*event
)
167 length
= event
->type_len
* RB_ALIGNMENT
;
169 length
= event
->array
[0];
170 return length
+ RB_EVNT_HDR_SIZE
;
174 * Return the length of the given event. Will return
175 * the length of the time extend if the event is a
178 static inline unsigned
179 rb_event_length(struct ring_buffer_event
*event
)
181 switch (event
->type_len
) {
182 case RINGBUF_TYPE_PADDING
:
183 if (rb_null_event(event
))
186 return event
->array
[0] + RB_EVNT_HDR_SIZE
;
188 case RINGBUF_TYPE_TIME_EXTEND
:
189 return RB_LEN_TIME_EXTEND
;
191 case RINGBUF_TYPE_TIME_STAMP
:
192 return RB_LEN_TIME_STAMP
;
194 case RINGBUF_TYPE_DATA
:
195 return rb_event_data_length(event
);
204 * Return total length of time extend and data,
205 * or just the event length for all other events.
207 static inline unsigned
208 rb_event_ts_length(struct ring_buffer_event
*event
)
212 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
) {
213 /* time extends include the data event after it */
214 len
= RB_LEN_TIME_EXTEND
;
215 event
= skip_time_extend(event
);
217 return len
+ rb_event_length(event
);
221 * ring_buffer_event_length - return the length of the event
222 * @event: the event to get the length of
224 * Returns the size of the data load of a data event.
225 * If the event is something other than a data event, it
226 * returns the size of the event itself. With the exception
227 * of a TIME EXTEND, where it still returns the size of the
228 * data load of the data event after it.
230 unsigned ring_buffer_event_length(struct ring_buffer_event
*event
)
234 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
235 event
= skip_time_extend(event
);
237 length
= rb_event_length(event
);
238 if (event
->type_len
> RINGBUF_TYPE_DATA_TYPE_LEN_MAX
)
240 length
-= RB_EVNT_HDR_SIZE
;
241 if (length
> RB_MAX_SMALL_DATA
+ sizeof(event
->array
[0]))
242 length
-= sizeof(event
->array
[0]);
245 EXPORT_SYMBOL_GPL(ring_buffer_event_length
);
247 /* inline for ring buffer fast paths */
248 static __always_inline
void *
249 rb_event_data(struct ring_buffer_event
*event
)
251 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
252 event
= skip_time_extend(event
);
253 BUG_ON(event
->type_len
> RINGBUF_TYPE_DATA_TYPE_LEN_MAX
);
254 /* If length is in len field, then array[0] has the data */
256 return (void *)&event
->array
[0];
257 /* Otherwise length is in array[0] and array[1] has the data */
258 return (void *)&event
->array
[1];
262 * ring_buffer_event_data - return the data of the event
263 * @event: the event to get the data from
265 void *ring_buffer_event_data(struct ring_buffer_event
*event
)
267 return rb_event_data(event
);
269 EXPORT_SYMBOL_GPL(ring_buffer_event_data
);
271 #define for_each_buffer_cpu(buffer, cpu) \
272 for_each_cpu(cpu, buffer->cpumask)
275 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
276 #define TS_DELTA_TEST (~TS_MASK)
278 /* Flag when events were overwritten */
279 #define RB_MISSED_EVENTS (1 << 31)
280 /* Missed count stored at end */
281 #define RB_MISSED_STORED (1 << 30)
283 #define RB_MISSED_FLAGS (RB_MISSED_EVENTS|RB_MISSED_STORED)
285 struct buffer_data_page
{
286 u64 time_stamp
; /* page time stamp */
287 local_t commit
; /* write committed index */
288 unsigned char data
[] RB_ALIGN_DATA
; /* data of buffer page */
292 * Note, the buffer_page list must be first. The buffer pages
293 * are allocated in cache lines, which means that each buffer
294 * page will be at the beginning of a cache line, and thus
295 * the least significant bits will be zero. We use this to
296 * add flags in the list struct pointers, to make the ring buffer
300 struct list_head list
; /* list of buffer pages */
301 local_t write
; /* index for next write */
302 unsigned read
; /* index for next read */
303 local_t entries
; /* entries on this page */
304 unsigned long real_end
; /* real end of data */
305 struct buffer_data_page
*page
; /* Actual data page */
309 * The buffer page counters, write and entries, must be reset
310 * atomically when crossing page boundaries. To synchronize this
311 * update, two counters are inserted into the number. One is
312 * the actual counter for the write position or count on the page.
314 * The other is a counter of updaters. Before an update happens
315 * the update partition of the counter is incremented. This will
316 * allow the updater to update the counter atomically.
318 * The counter is 20 bits, and the state data is 12.
320 #define RB_WRITE_MASK 0xfffff
321 #define RB_WRITE_INTCNT (1 << 20)
323 static void rb_init_page(struct buffer_data_page
*bpage
)
325 local_set(&bpage
->commit
, 0);
329 * ring_buffer_page_len - the size of data on the page.
330 * @page: The page to read
332 * Returns the amount of data on the page, including buffer page header.
334 size_t ring_buffer_page_len(void *page
)
336 struct buffer_data_page
*bpage
= page
;
338 return (local_read(&bpage
->commit
) & ~RB_MISSED_FLAGS
)
343 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
346 static void free_buffer_page(struct buffer_page
*bpage
)
348 free_page((unsigned long)bpage
->page
);
353 * We need to fit the time_stamp delta into 27 bits.
355 static inline int test_time_stamp(u64 delta
)
357 if (delta
& TS_DELTA_TEST
)
362 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
364 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
365 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
367 int ring_buffer_print_page_header(struct trace_seq
*s
)
369 struct buffer_data_page field
;
371 trace_seq_printf(s
, "\tfield: u64 timestamp;\t"
372 "offset:0;\tsize:%u;\tsigned:%u;\n",
373 (unsigned int)sizeof(field
.time_stamp
),
374 (unsigned int)is_signed_type(u64
));
376 trace_seq_printf(s
, "\tfield: local_t commit;\t"
377 "offset:%u;\tsize:%u;\tsigned:%u;\n",
378 (unsigned int)offsetof(typeof(field
), commit
),
379 (unsigned int)sizeof(field
.commit
),
380 (unsigned int)is_signed_type(long));
382 trace_seq_printf(s
, "\tfield: int overwrite;\t"
383 "offset:%u;\tsize:%u;\tsigned:%u;\n",
384 (unsigned int)offsetof(typeof(field
), commit
),
386 (unsigned int)is_signed_type(long));
388 trace_seq_printf(s
, "\tfield: char data;\t"
389 "offset:%u;\tsize:%u;\tsigned:%u;\n",
390 (unsigned int)offsetof(typeof(field
), data
),
391 (unsigned int)BUF_PAGE_SIZE
,
392 (unsigned int)is_signed_type(char));
394 return !trace_seq_has_overflowed(s
);
398 struct irq_work work
;
399 wait_queue_head_t waiters
;
400 wait_queue_head_t full_waiters
;
401 bool waiters_pending
;
402 bool full_waiters_pending
;
407 * Structure to hold event state and handle nested events.
409 struct rb_event_info
{
412 unsigned long length
;
413 struct buffer_page
*tail_page
;
418 * Used for which event context the event is in.
424 * See trace_recursive_lock() comment below for more details.
435 * head_page == tail_page && head == tail then buffer is empty.
437 struct ring_buffer_per_cpu
{
439 atomic_t record_disabled
;
440 struct ring_buffer
*buffer
;
441 raw_spinlock_t reader_lock
; /* serialize readers */
442 arch_spinlock_t lock
;
443 struct lock_class_key lock_key
;
444 struct buffer_data_page
*free_page
;
445 unsigned long nr_pages
;
446 unsigned int current_context
;
447 struct list_head
*pages
;
448 struct buffer_page
*head_page
; /* read from head */
449 struct buffer_page
*tail_page
; /* write to tail */
450 struct buffer_page
*commit_page
; /* committed pages */
451 struct buffer_page
*reader_page
;
452 unsigned long lost_events
;
453 unsigned long last_overrun
;
454 local_t entries_bytes
;
457 local_t commit_overrun
;
458 local_t dropped_events
;
462 unsigned long read_bytes
;
465 /* ring buffer pages to update, > 0 to add, < 0 to remove */
466 long nr_pages_to_update
;
467 struct list_head new_pages
; /* new pages to add */
468 struct work_struct update_pages_work
;
469 struct completion update_done
;
471 struct rb_irq_work irq_work
;
477 atomic_t record_disabled
;
478 atomic_t resize_disabled
;
479 cpumask_var_t cpumask
;
481 struct lock_class_key
*reader_lock_key
;
485 struct ring_buffer_per_cpu
**buffers
;
487 struct hlist_node node
;
490 struct rb_irq_work irq_work
;
493 struct ring_buffer_iter
{
494 struct ring_buffer_per_cpu
*cpu_buffer
;
496 struct buffer_page
*head_page
;
497 struct buffer_page
*cache_reader_page
;
498 unsigned long cache_read
;
503 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
505 * Schedules a delayed work to wake up any task that is blocked on the
506 * ring buffer waiters queue.
508 static void rb_wake_up_waiters(struct irq_work
*work
)
510 struct rb_irq_work
*rbwork
= container_of(work
, struct rb_irq_work
, work
);
512 wake_up_all(&rbwork
->waiters
);
513 if (rbwork
->wakeup_full
) {
514 rbwork
->wakeup_full
= false;
515 wake_up_all(&rbwork
->full_waiters
);
520 * ring_buffer_wait - wait for input to the ring buffer
521 * @buffer: buffer to wait on
522 * @cpu: the cpu buffer to wait on
523 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
525 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
526 * as data is added to any of the @buffer's cpu buffers. Otherwise
527 * it will wait for data to be added to a specific cpu buffer.
529 int ring_buffer_wait(struct ring_buffer
*buffer
, int cpu
, bool full
)
531 struct ring_buffer_per_cpu
*uninitialized_var(cpu_buffer
);
533 struct rb_irq_work
*work
;
537 * Depending on what the caller is waiting for, either any
538 * data in any cpu buffer, or a specific buffer, put the
539 * caller on the appropriate wait queue.
541 if (cpu
== RING_BUFFER_ALL_CPUS
) {
542 work
= &buffer
->irq_work
;
543 /* Full only makes sense on per cpu reads */
546 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
548 cpu_buffer
= buffer
->buffers
[cpu
];
549 work
= &cpu_buffer
->irq_work
;
555 prepare_to_wait(&work
->full_waiters
, &wait
, TASK_INTERRUPTIBLE
);
557 prepare_to_wait(&work
->waiters
, &wait
, TASK_INTERRUPTIBLE
);
560 * The events can happen in critical sections where
561 * checking a work queue can cause deadlocks.
562 * After adding a task to the queue, this flag is set
563 * only to notify events to try to wake up the queue
566 * We don't clear it even if the buffer is no longer
567 * empty. The flag only causes the next event to run
568 * irq_work to do the work queue wake up. The worse
569 * that can happen if we race with !trace_empty() is that
570 * an event will cause an irq_work to try to wake up
573 * There's no reason to protect this flag either, as
574 * the work queue and irq_work logic will do the necessary
575 * synchronization for the wake ups. The only thing
576 * that is necessary is that the wake up happens after
577 * a task has been queued. It's OK for spurious wake ups.
580 work
->full_waiters_pending
= true;
582 work
->waiters_pending
= true;
584 if (signal_pending(current
)) {
589 if (cpu
== RING_BUFFER_ALL_CPUS
&& !ring_buffer_empty(buffer
))
592 if (cpu
!= RING_BUFFER_ALL_CPUS
&&
593 !ring_buffer_empty_cpu(buffer
, cpu
)) {
600 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
601 pagebusy
= cpu_buffer
->reader_page
== cpu_buffer
->commit_page
;
602 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
612 finish_wait(&work
->full_waiters
, &wait
);
614 finish_wait(&work
->waiters
, &wait
);
620 * ring_buffer_poll_wait - poll on buffer input
621 * @buffer: buffer to wait on
622 * @cpu: the cpu buffer to wait on
623 * @filp: the file descriptor
624 * @poll_table: The poll descriptor
626 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
627 * as data is added to any of the @buffer's cpu buffers. Otherwise
628 * it will wait for data to be added to a specific cpu buffer.
630 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
633 int ring_buffer_poll_wait(struct ring_buffer
*buffer
, int cpu
,
634 struct file
*filp
, poll_table
*poll_table
)
636 struct ring_buffer_per_cpu
*cpu_buffer
;
637 struct rb_irq_work
*work
;
639 if (cpu
== RING_BUFFER_ALL_CPUS
)
640 work
= &buffer
->irq_work
;
642 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
645 cpu_buffer
= buffer
->buffers
[cpu
];
646 work
= &cpu_buffer
->irq_work
;
649 poll_wait(filp
, &work
->waiters
, poll_table
);
650 work
->waiters_pending
= true;
652 * There's a tight race between setting the waiters_pending and
653 * checking if the ring buffer is empty. Once the waiters_pending bit
654 * is set, the next event will wake the task up, but we can get stuck
655 * if there's only a single event in.
657 * FIXME: Ideally, we need a memory barrier on the writer side as well,
658 * but adding a memory barrier to all events will cause too much of a
659 * performance hit in the fast path. We only need a memory barrier when
660 * the buffer goes from empty to having content. But as this race is
661 * extremely small, and it's not a problem if another event comes in, we
666 if ((cpu
== RING_BUFFER_ALL_CPUS
&& !ring_buffer_empty(buffer
)) ||
667 (cpu
!= RING_BUFFER_ALL_CPUS
&& !ring_buffer_empty_cpu(buffer
, cpu
)))
668 return POLLIN
| POLLRDNORM
;
672 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
673 #define RB_WARN_ON(b, cond) \
675 int _____ret = unlikely(cond); \
677 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
678 struct ring_buffer_per_cpu *__b = \
680 atomic_inc(&__b->buffer->record_disabled); \
682 atomic_inc(&b->record_disabled); \
688 /* Up this if you want to test the TIME_EXTENTS and normalization */
689 #define DEBUG_SHIFT 0
691 static inline u64
rb_time_stamp(struct ring_buffer
*buffer
)
693 /* shift to debug/test normalization and TIME_EXTENTS */
694 return buffer
->clock() << DEBUG_SHIFT
;
697 u64
ring_buffer_time_stamp(struct ring_buffer
*buffer
, int cpu
)
701 preempt_disable_notrace();
702 time
= rb_time_stamp(buffer
);
703 preempt_enable_no_resched_notrace();
707 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp
);
709 void ring_buffer_normalize_time_stamp(struct ring_buffer
*buffer
,
712 /* Just stupid testing the normalize function and deltas */
715 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp
);
718 * Making the ring buffer lockless makes things tricky.
719 * Although writes only happen on the CPU that they are on,
720 * and they only need to worry about interrupts. Reads can
723 * The reader page is always off the ring buffer, but when the
724 * reader finishes with a page, it needs to swap its page with
725 * a new one from the buffer. The reader needs to take from
726 * the head (writes go to the tail). But if a writer is in overwrite
727 * mode and wraps, it must push the head page forward.
729 * Here lies the problem.
731 * The reader must be careful to replace only the head page, and
732 * not another one. As described at the top of the file in the
733 * ASCII art, the reader sets its old page to point to the next
734 * page after head. It then sets the page after head to point to
735 * the old reader page. But if the writer moves the head page
736 * during this operation, the reader could end up with the tail.
738 * We use cmpxchg to help prevent this race. We also do something
739 * special with the page before head. We set the LSB to 1.
741 * When the writer must push the page forward, it will clear the
742 * bit that points to the head page, move the head, and then set
743 * the bit that points to the new head page.
745 * We also don't want an interrupt coming in and moving the head
746 * page on another writer. Thus we use the second LSB to catch
749 * head->list->prev->next bit 1 bit 0
752 * Points to head page 0 1
755 * Note we can not trust the prev pointer of the head page, because:
757 * +----+ +-----+ +-----+
758 * | |------>| T |---X--->| N |
760 * +----+ +-----+ +-----+
763 * +----------| R |----------+ |
767 * Key: ---X--> HEAD flag set in pointer
772 * (see __rb_reserve_next() to see where this happens)
774 * What the above shows is that the reader just swapped out
775 * the reader page with a page in the buffer, but before it
776 * could make the new header point back to the new page added
777 * it was preempted by a writer. The writer moved forward onto
778 * the new page added by the reader and is about to move forward
781 * You can see, it is legitimate for the previous pointer of
782 * the head (or any page) not to point back to itself. But only
786 #define RB_PAGE_NORMAL 0UL
787 #define RB_PAGE_HEAD 1UL
788 #define RB_PAGE_UPDATE 2UL
791 #define RB_FLAG_MASK 3UL
793 /* PAGE_MOVED is not part of the mask */
794 #define RB_PAGE_MOVED 4UL
797 * rb_list_head - remove any bit
799 static struct list_head
*rb_list_head(struct list_head
*list
)
801 unsigned long val
= (unsigned long)list
;
803 return (struct list_head
*)(val
& ~RB_FLAG_MASK
);
807 * rb_is_head_page - test if the given page is the head page
809 * Because the reader may move the head_page pointer, we can
810 * not trust what the head page is (it may be pointing to
811 * the reader page). But if the next page is a header page,
812 * its flags will be non zero.
815 rb_is_head_page(struct ring_buffer_per_cpu
*cpu_buffer
,
816 struct buffer_page
*page
, struct list_head
*list
)
820 val
= (unsigned long)list
->next
;
822 if ((val
& ~RB_FLAG_MASK
) != (unsigned long)&page
->list
)
823 return RB_PAGE_MOVED
;
825 return val
& RB_FLAG_MASK
;
831 * The unique thing about the reader page, is that, if the
832 * writer is ever on it, the previous pointer never points
833 * back to the reader page.
835 static bool rb_is_reader_page(struct buffer_page
*page
)
837 struct list_head
*list
= page
->list
.prev
;
839 return rb_list_head(list
->next
) != &page
->list
;
843 * rb_set_list_to_head - set a list_head to be pointing to head.
845 static void rb_set_list_to_head(struct ring_buffer_per_cpu
*cpu_buffer
,
846 struct list_head
*list
)
850 ptr
= (unsigned long *)&list
->next
;
851 *ptr
|= RB_PAGE_HEAD
;
852 *ptr
&= ~RB_PAGE_UPDATE
;
856 * rb_head_page_activate - sets up head page
858 static void rb_head_page_activate(struct ring_buffer_per_cpu
*cpu_buffer
)
860 struct buffer_page
*head
;
862 head
= cpu_buffer
->head_page
;
867 * Set the previous list pointer to have the HEAD flag.
869 rb_set_list_to_head(cpu_buffer
, head
->list
.prev
);
872 static void rb_list_head_clear(struct list_head
*list
)
874 unsigned long *ptr
= (unsigned long *)&list
->next
;
876 *ptr
&= ~RB_FLAG_MASK
;
880 * rb_head_page_dactivate - clears head page ptr (for free list)
883 rb_head_page_deactivate(struct ring_buffer_per_cpu
*cpu_buffer
)
885 struct list_head
*hd
;
887 /* Go through the whole list and clear any pointers found. */
888 rb_list_head_clear(cpu_buffer
->pages
);
890 list_for_each(hd
, cpu_buffer
->pages
)
891 rb_list_head_clear(hd
);
894 static int rb_head_page_set(struct ring_buffer_per_cpu
*cpu_buffer
,
895 struct buffer_page
*head
,
896 struct buffer_page
*prev
,
897 int old_flag
, int new_flag
)
899 struct list_head
*list
;
900 unsigned long val
= (unsigned long)&head
->list
;
905 val
&= ~RB_FLAG_MASK
;
907 ret
= cmpxchg((unsigned long *)&list
->next
,
908 val
| old_flag
, val
| new_flag
);
910 /* check if the reader took the page */
911 if ((ret
& ~RB_FLAG_MASK
) != val
)
912 return RB_PAGE_MOVED
;
914 return ret
& RB_FLAG_MASK
;
917 static int rb_head_page_set_update(struct ring_buffer_per_cpu
*cpu_buffer
,
918 struct buffer_page
*head
,
919 struct buffer_page
*prev
,
922 return rb_head_page_set(cpu_buffer
, head
, prev
,
923 old_flag
, RB_PAGE_UPDATE
);
926 static int rb_head_page_set_head(struct ring_buffer_per_cpu
*cpu_buffer
,
927 struct buffer_page
*head
,
928 struct buffer_page
*prev
,
931 return rb_head_page_set(cpu_buffer
, head
, prev
,
932 old_flag
, RB_PAGE_HEAD
);
935 static int rb_head_page_set_normal(struct ring_buffer_per_cpu
*cpu_buffer
,
936 struct buffer_page
*head
,
937 struct buffer_page
*prev
,
940 return rb_head_page_set(cpu_buffer
, head
, prev
,
941 old_flag
, RB_PAGE_NORMAL
);
944 static inline void rb_inc_page(struct ring_buffer_per_cpu
*cpu_buffer
,
945 struct buffer_page
**bpage
)
947 struct list_head
*p
= rb_list_head((*bpage
)->list
.next
);
949 *bpage
= list_entry(p
, struct buffer_page
, list
);
952 static struct buffer_page
*
953 rb_set_head_page(struct ring_buffer_per_cpu
*cpu_buffer
)
955 struct buffer_page
*head
;
956 struct buffer_page
*page
;
957 struct list_head
*list
;
960 if (RB_WARN_ON(cpu_buffer
, !cpu_buffer
->head_page
))
964 list
= cpu_buffer
->pages
;
965 if (RB_WARN_ON(cpu_buffer
, rb_list_head(list
->prev
->next
) != list
))
968 page
= head
= cpu_buffer
->head_page
;
970 * It is possible that the writer moves the header behind
971 * where we started, and we miss in one loop.
972 * A second loop should grab the header, but we'll do
973 * three loops just because I'm paranoid.
975 for (i
= 0; i
< 3; i
++) {
977 if (rb_is_head_page(cpu_buffer
, page
, page
->list
.prev
)) {
978 cpu_buffer
->head_page
= page
;
981 rb_inc_page(cpu_buffer
, &page
);
982 } while (page
!= head
);
985 RB_WARN_ON(cpu_buffer
, 1);
990 static int rb_head_page_replace(struct buffer_page
*old
,
991 struct buffer_page
*new)
993 unsigned long *ptr
= (unsigned long *)&old
->list
.prev
->next
;
997 val
= *ptr
& ~RB_FLAG_MASK
;
1000 ret
= cmpxchg(ptr
, val
, (unsigned long)&new->list
);
1006 * rb_tail_page_update - move the tail page forward
1008 static void rb_tail_page_update(struct ring_buffer_per_cpu
*cpu_buffer
,
1009 struct buffer_page
*tail_page
,
1010 struct buffer_page
*next_page
)
1012 unsigned long old_entries
;
1013 unsigned long old_write
;
1016 * The tail page now needs to be moved forward.
1018 * We need to reset the tail page, but without messing
1019 * with possible erasing of data brought in by interrupts
1020 * that have moved the tail page and are currently on it.
1022 * We add a counter to the write field to denote this.
1024 old_write
= local_add_return(RB_WRITE_INTCNT
, &next_page
->write
);
1025 old_entries
= local_add_return(RB_WRITE_INTCNT
, &next_page
->entries
);
1028 * Just make sure we have seen our old_write and synchronize
1029 * with any interrupts that come in.
1034 * If the tail page is still the same as what we think
1035 * it is, then it is up to us to update the tail
1038 if (tail_page
== READ_ONCE(cpu_buffer
->tail_page
)) {
1039 /* Zero the write counter */
1040 unsigned long val
= old_write
& ~RB_WRITE_MASK
;
1041 unsigned long eval
= old_entries
& ~RB_WRITE_MASK
;
1044 * This will only succeed if an interrupt did
1045 * not come in and change it. In which case, we
1046 * do not want to modify it.
1048 * We add (void) to let the compiler know that we do not care
1049 * about the return value of these functions. We use the
1050 * cmpxchg to only update if an interrupt did not already
1051 * do it for us. If the cmpxchg fails, we don't care.
1053 (void)local_cmpxchg(&next_page
->write
, old_write
, val
);
1054 (void)local_cmpxchg(&next_page
->entries
, old_entries
, eval
);
1057 * No need to worry about races with clearing out the commit.
1058 * it only can increment when a commit takes place. But that
1059 * only happens in the outer most nested commit.
1061 local_set(&next_page
->page
->commit
, 0);
1063 /* Again, either we update tail_page or an interrupt does */
1064 (void)cmpxchg(&cpu_buffer
->tail_page
, tail_page
, next_page
);
1068 static int rb_check_bpage(struct ring_buffer_per_cpu
*cpu_buffer
,
1069 struct buffer_page
*bpage
)
1071 unsigned long val
= (unsigned long)bpage
;
1073 if (RB_WARN_ON(cpu_buffer
, val
& RB_FLAG_MASK
))
1080 * rb_check_list - make sure a pointer to a list has the last bits zero
1082 static int rb_check_list(struct ring_buffer_per_cpu
*cpu_buffer
,
1083 struct list_head
*list
)
1085 if (RB_WARN_ON(cpu_buffer
, rb_list_head(list
->prev
) != list
->prev
))
1087 if (RB_WARN_ON(cpu_buffer
, rb_list_head(list
->next
) != list
->next
))
1093 * rb_check_pages - integrity check of buffer pages
1094 * @cpu_buffer: CPU buffer with pages to test
1096 * As a safety measure we check to make sure the data pages have not
1099 static int rb_check_pages(struct ring_buffer_per_cpu
*cpu_buffer
)
1101 struct list_head
*head
= cpu_buffer
->pages
;
1102 struct buffer_page
*bpage
, *tmp
;
1104 /* Reset the head page if it exists */
1105 if (cpu_buffer
->head_page
)
1106 rb_set_head_page(cpu_buffer
);
1108 rb_head_page_deactivate(cpu_buffer
);
1110 if (RB_WARN_ON(cpu_buffer
, head
->next
->prev
!= head
))
1112 if (RB_WARN_ON(cpu_buffer
, head
->prev
->next
!= head
))
1115 if (rb_check_list(cpu_buffer
, head
))
1118 list_for_each_entry_safe(bpage
, tmp
, head
, list
) {
1119 if (RB_WARN_ON(cpu_buffer
,
1120 bpage
->list
.next
->prev
!= &bpage
->list
))
1122 if (RB_WARN_ON(cpu_buffer
,
1123 bpage
->list
.prev
->next
!= &bpage
->list
))
1125 if (rb_check_list(cpu_buffer
, &bpage
->list
))
1129 rb_head_page_activate(cpu_buffer
);
1134 static int __rb_allocate_pages(long nr_pages
, struct list_head
*pages
, int cpu
)
1136 struct buffer_page
*bpage
, *tmp
;
1139 /* Check if the available memory is there first */
1140 i
= si_mem_available();
1144 for (i
= 0; i
< nr_pages
; i
++) {
1147 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1148 * gracefully without invoking oom-killer and the system is not
1151 bpage
= kzalloc_node(ALIGN(sizeof(*bpage
), cache_line_size()),
1152 GFP_KERNEL
| __GFP_RETRY_MAYFAIL
,
1157 list_add(&bpage
->list
, pages
);
1159 page
= alloc_pages_node(cpu_to_node(cpu
),
1160 GFP_KERNEL
| __GFP_RETRY_MAYFAIL
, 0);
1163 bpage
->page
= page_address(page
);
1164 rb_init_page(bpage
->page
);
1170 list_for_each_entry_safe(bpage
, tmp
, pages
, list
) {
1171 list_del_init(&bpage
->list
);
1172 free_buffer_page(bpage
);
1178 static int rb_allocate_pages(struct ring_buffer_per_cpu
*cpu_buffer
,
1179 unsigned long nr_pages
)
1185 if (__rb_allocate_pages(nr_pages
, &pages
, cpu_buffer
->cpu
))
1189 * The ring buffer page list is a circular list that does not
1190 * start and end with a list head. All page list items point to
1193 cpu_buffer
->pages
= pages
.next
;
1196 cpu_buffer
->nr_pages
= nr_pages
;
1198 rb_check_pages(cpu_buffer
);
1203 static struct ring_buffer_per_cpu
*
1204 rb_allocate_cpu_buffer(struct ring_buffer
*buffer
, long nr_pages
, int cpu
)
1206 struct ring_buffer_per_cpu
*cpu_buffer
;
1207 struct buffer_page
*bpage
;
1211 cpu_buffer
= kzalloc_node(ALIGN(sizeof(*cpu_buffer
), cache_line_size()),
1212 GFP_KERNEL
, cpu_to_node(cpu
));
1216 cpu_buffer
->cpu
= cpu
;
1217 cpu_buffer
->buffer
= buffer
;
1218 raw_spin_lock_init(&cpu_buffer
->reader_lock
);
1219 lockdep_set_class(&cpu_buffer
->reader_lock
, buffer
->reader_lock_key
);
1220 cpu_buffer
->lock
= (arch_spinlock_t
)__ARCH_SPIN_LOCK_UNLOCKED
;
1221 INIT_WORK(&cpu_buffer
->update_pages_work
, update_pages_handler
);
1222 init_completion(&cpu_buffer
->update_done
);
1223 init_irq_work(&cpu_buffer
->irq_work
.work
, rb_wake_up_waiters
);
1224 init_waitqueue_head(&cpu_buffer
->irq_work
.waiters
);
1225 init_waitqueue_head(&cpu_buffer
->irq_work
.full_waiters
);
1227 bpage
= kzalloc_node(ALIGN(sizeof(*bpage
), cache_line_size()),
1228 GFP_KERNEL
, cpu_to_node(cpu
));
1230 goto fail_free_buffer
;
1232 rb_check_bpage(cpu_buffer
, bpage
);
1234 cpu_buffer
->reader_page
= bpage
;
1235 page
= alloc_pages_node(cpu_to_node(cpu
), GFP_KERNEL
, 0);
1237 goto fail_free_reader
;
1238 bpage
->page
= page_address(page
);
1239 rb_init_page(bpage
->page
);
1241 INIT_LIST_HEAD(&cpu_buffer
->reader_page
->list
);
1242 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
1244 ret
= rb_allocate_pages(cpu_buffer
, nr_pages
);
1246 goto fail_free_reader
;
1248 cpu_buffer
->head_page
1249 = list_entry(cpu_buffer
->pages
, struct buffer_page
, list
);
1250 cpu_buffer
->tail_page
= cpu_buffer
->commit_page
= cpu_buffer
->head_page
;
1252 rb_head_page_activate(cpu_buffer
);
1257 free_buffer_page(cpu_buffer
->reader_page
);
1264 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu
*cpu_buffer
)
1266 struct list_head
*head
= cpu_buffer
->pages
;
1267 struct buffer_page
*bpage
, *tmp
;
1269 free_buffer_page(cpu_buffer
->reader_page
);
1271 rb_head_page_deactivate(cpu_buffer
);
1274 list_for_each_entry_safe(bpage
, tmp
, head
, list
) {
1275 list_del_init(&bpage
->list
);
1276 free_buffer_page(bpage
);
1278 bpage
= list_entry(head
, struct buffer_page
, list
);
1279 free_buffer_page(bpage
);
1286 * __ring_buffer_alloc - allocate a new ring_buffer
1287 * @size: the size in bytes per cpu that is needed.
1288 * @flags: attributes to set for the ring buffer.
1290 * Currently the only flag that is available is the RB_FL_OVERWRITE
1291 * flag. This flag means that the buffer will overwrite old data
1292 * when the buffer wraps. If this flag is not set, the buffer will
1293 * drop data when the tail hits the head.
1295 struct ring_buffer
*__ring_buffer_alloc(unsigned long size
, unsigned flags
,
1296 struct lock_class_key
*key
)
1298 struct ring_buffer
*buffer
;
1304 /* keep it in its own cache line */
1305 buffer
= kzalloc(ALIGN(sizeof(*buffer
), cache_line_size()),
1310 if (!zalloc_cpumask_var(&buffer
->cpumask
, GFP_KERNEL
))
1311 goto fail_free_buffer
;
1313 nr_pages
= DIV_ROUND_UP(size
, BUF_PAGE_SIZE
);
1314 buffer
->flags
= flags
;
1315 buffer
->clock
= trace_clock_local
;
1316 buffer
->reader_lock_key
= key
;
1318 init_irq_work(&buffer
->irq_work
.work
, rb_wake_up_waiters
);
1319 init_waitqueue_head(&buffer
->irq_work
.waiters
);
1321 /* need at least two pages */
1325 buffer
->cpus
= nr_cpu_ids
;
1327 bsize
= sizeof(void *) * nr_cpu_ids
;
1328 buffer
->buffers
= kzalloc(ALIGN(bsize
, cache_line_size()),
1330 if (!buffer
->buffers
)
1331 goto fail_free_cpumask
;
1333 cpu
= raw_smp_processor_id();
1334 cpumask_set_cpu(cpu
, buffer
->cpumask
);
1335 buffer
->buffers
[cpu
] = rb_allocate_cpu_buffer(buffer
, nr_pages
, cpu
);
1336 if (!buffer
->buffers
[cpu
])
1337 goto fail_free_buffers
;
1339 ret
= cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE
, &buffer
->node
);
1341 goto fail_free_buffers
;
1343 mutex_init(&buffer
->mutex
);
1348 for_each_buffer_cpu(buffer
, cpu
) {
1349 if (buffer
->buffers
[cpu
])
1350 rb_free_cpu_buffer(buffer
->buffers
[cpu
]);
1352 kfree(buffer
->buffers
);
1355 free_cpumask_var(buffer
->cpumask
);
1361 EXPORT_SYMBOL_GPL(__ring_buffer_alloc
);
1364 * ring_buffer_free - free a ring buffer.
1365 * @buffer: the buffer to free.
1368 ring_buffer_free(struct ring_buffer
*buffer
)
1372 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE
, &buffer
->node
);
1374 for_each_buffer_cpu(buffer
, cpu
)
1375 rb_free_cpu_buffer(buffer
->buffers
[cpu
]);
1377 kfree(buffer
->buffers
);
1378 free_cpumask_var(buffer
->cpumask
);
1382 EXPORT_SYMBOL_GPL(ring_buffer_free
);
1384 void ring_buffer_set_clock(struct ring_buffer
*buffer
,
1387 buffer
->clock
= clock
;
1390 static void rb_reset_cpu(struct ring_buffer_per_cpu
*cpu_buffer
);
1392 static inline unsigned long rb_page_entries(struct buffer_page
*bpage
)
1394 return local_read(&bpage
->entries
) & RB_WRITE_MASK
;
1397 static inline unsigned long rb_page_write(struct buffer_page
*bpage
)
1399 return local_read(&bpage
->write
) & RB_WRITE_MASK
;
1403 rb_remove_pages(struct ring_buffer_per_cpu
*cpu_buffer
, unsigned long nr_pages
)
1405 struct list_head
*tail_page
, *to_remove
, *next_page
;
1406 struct buffer_page
*to_remove_page
, *tmp_iter_page
;
1407 struct buffer_page
*last_page
, *first_page
;
1408 unsigned long nr_removed
;
1409 unsigned long head_bit
;
1414 raw_spin_lock_irq(&cpu_buffer
->reader_lock
);
1415 atomic_inc(&cpu_buffer
->record_disabled
);
1417 * We don't race with the readers since we have acquired the reader
1418 * lock. We also don't race with writers after disabling recording.
1419 * This makes it easy to figure out the first and the last page to be
1420 * removed from the list. We unlink all the pages in between including
1421 * the first and last pages. This is done in a busy loop so that we
1422 * lose the least number of traces.
1423 * The pages are freed after we restart recording and unlock readers.
1425 tail_page
= &cpu_buffer
->tail_page
->list
;
1428 * tail page might be on reader page, we remove the next page
1429 * from the ring buffer
1431 if (cpu_buffer
->tail_page
== cpu_buffer
->reader_page
)
1432 tail_page
= rb_list_head(tail_page
->next
);
1433 to_remove
= tail_page
;
1435 /* start of pages to remove */
1436 first_page
= list_entry(rb_list_head(to_remove
->next
),
1437 struct buffer_page
, list
);
1439 for (nr_removed
= 0; nr_removed
< nr_pages
; nr_removed
++) {
1440 to_remove
= rb_list_head(to_remove
)->next
;
1441 head_bit
|= (unsigned long)to_remove
& RB_PAGE_HEAD
;
1444 next_page
= rb_list_head(to_remove
)->next
;
1447 * Now we remove all pages between tail_page and next_page.
1448 * Make sure that we have head_bit value preserved for the
1451 tail_page
->next
= (struct list_head
*)((unsigned long)next_page
|
1453 next_page
= rb_list_head(next_page
);
1454 next_page
->prev
= tail_page
;
1456 /* make sure pages points to a valid page in the ring buffer */
1457 cpu_buffer
->pages
= next_page
;
1459 /* update head page */
1461 cpu_buffer
->head_page
= list_entry(next_page
,
1462 struct buffer_page
, list
);
1465 * change read pointer to make sure any read iterators reset
1468 cpu_buffer
->read
= 0;
1470 /* pages are removed, resume tracing and then free the pages */
1471 atomic_dec(&cpu_buffer
->record_disabled
);
1472 raw_spin_unlock_irq(&cpu_buffer
->reader_lock
);
1474 RB_WARN_ON(cpu_buffer
, list_empty(cpu_buffer
->pages
));
1476 /* last buffer page to remove */
1477 last_page
= list_entry(rb_list_head(to_remove
), struct buffer_page
,
1479 tmp_iter_page
= first_page
;
1482 to_remove_page
= tmp_iter_page
;
1483 rb_inc_page(cpu_buffer
, &tmp_iter_page
);
1485 /* update the counters */
1486 page_entries
= rb_page_entries(to_remove_page
);
1489 * If something was added to this page, it was full
1490 * since it is not the tail page. So we deduct the
1491 * bytes consumed in ring buffer from here.
1492 * Increment overrun to account for the lost events.
1494 local_add(page_entries
, &cpu_buffer
->overrun
);
1495 local_sub(BUF_PAGE_SIZE
, &cpu_buffer
->entries_bytes
);
1499 * We have already removed references to this list item, just
1500 * free up the buffer_page and its page
1502 free_buffer_page(to_remove_page
);
1505 } while (to_remove_page
!= last_page
);
1507 RB_WARN_ON(cpu_buffer
, nr_removed
);
1509 return nr_removed
== 0;
1513 rb_insert_pages(struct ring_buffer_per_cpu
*cpu_buffer
)
1515 struct list_head
*pages
= &cpu_buffer
->new_pages
;
1516 int retries
, success
;
1518 raw_spin_lock_irq(&cpu_buffer
->reader_lock
);
1520 * We are holding the reader lock, so the reader page won't be swapped
1521 * in the ring buffer. Now we are racing with the writer trying to
1522 * move head page and the tail page.
1523 * We are going to adapt the reader page update process where:
1524 * 1. We first splice the start and end of list of new pages between
1525 * the head page and its previous page.
1526 * 2. We cmpxchg the prev_page->next to point from head page to the
1527 * start of new pages list.
1528 * 3. Finally, we update the head->prev to the end of new list.
1530 * We will try this process 10 times, to make sure that we don't keep
1536 struct list_head
*head_page
, *prev_page
, *r
;
1537 struct list_head
*last_page
, *first_page
;
1538 struct list_head
*head_page_with_bit
;
1540 head_page
= &rb_set_head_page(cpu_buffer
)->list
;
1543 prev_page
= head_page
->prev
;
1545 first_page
= pages
->next
;
1546 last_page
= pages
->prev
;
1548 head_page_with_bit
= (struct list_head
*)
1549 ((unsigned long)head_page
| RB_PAGE_HEAD
);
1551 last_page
->next
= head_page_with_bit
;
1552 first_page
->prev
= prev_page
;
1554 r
= cmpxchg(&prev_page
->next
, head_page_with_bit
, first_page
);
1556 if (r
== head_page_with_bit
) {
1558 * yay, we replaced the page pointer to our new list,
1559 * now, we just have to update to head page's prev
1560 * pointer to point to end of list
1562 head_page
->prev
= last_page
;
1569 INIT_LIST_HEAD(pages
);
1571 * If we weren't successful in adding in new pages, warn and stop
1574 RB_WARN_ON(cpu_buffer
, !success
);
1575 raw_spin_unlock_irq(&cpu_buffer
->reader_lock
);
1577 /* free pages if they weren't inserted */
1579 struct buffer_page
*bpage
, *tmp
;
1580 list_for_each_entry_safe(bpage
, tmp
, &cpu_buffer
->new_pages
,
1582 list_del_init(&bpage
->list
);
1583 free_buffer_page(bpage
);
1589 static void rb_update_pages(struct ring_buffer_per_cpu
*cpu_buffer
)
1593 if (cpu_buffer
->nr_pages_to_update
> 0)
1594 success
= rb_insert_pages(cpu_buffer
);
1596 success
= rb_remove_pages(cpu_buffer
,
1597 -cpu_buffer
->nr_pages_to_update
);
1600 cpu_buffer
->nr_pages
+= cpu_buffer
->nr_pages_to_update
;
1603 static void update_pages_handler(struct work_struct
*work
)
1605 struct ring_buffer_per_cpu
*cpu_buffer
= container_of(work
,
1606 struct ring_buffer_per_cpu
, update_pages_work
);
1607 rb_update_pages(cpu_buffer
);
1608 complete(&cpu_buffer
->update_done
);
1612 * ring_buffer_resize - resize the ring buffer
1613 * @buffer: the buffer to resize.
1614 * @size: the new size.
1615 * @cpu_id: the cpu buffer to resize
1617 * Minimum size is 2 * BUF_PAGE_SIZE.
1619 * Returns 0 on success and < 0 on failure.
1621 int ring_buffer_resize(struct ring_buffer
*buffer
, unsigned long size
,
1624 struct ring_buffer_per_cpu
*cpu_buffer
;
1625 unsigned long nr_pages
;
1629 * Always succeed at resizing a non-existent buffer:
1634 /* Make sure the requested buffer exists */
1635 if (cpu_id
!= RING_BUFFER_ALL_CPUS
&&
1636 !cpumask_test_cpu(cpu_id
, buffer
->cpumask
))
1639 nr_pages
= DIV_ROUND_UP(size
, BUF_PAGE_SIZE
);
1641 /* we need a minimum of two pages */
1645 size
= nr_pages
* BUF_PAGE_SIZE
;
1648 * Don't succeed if resizing is disabled, as a reader might be
1649 * manipulating the ring buffer and is expecting a sane state while
1652 if (atomic_read(&buffer
->resize_disabled
))
1655 /* prevent another thread from changing buffer sizes */
1656 mutex_lock(&buffer
->mutex
);
1658 if (cpu_id
== RING_BUFFER_ALL_CPUS
) {
1659 /* calculate the pages to update */
1660 for_each_buffer_cpu(buffer
, cpu
) {
1661 cpu_buffer
= buffer
->buffers
[cpu
];
1663 cpu_buffer
->nr_pages_to_update
= nr_pages
-
1664 cpu_buffer
->nr_pages
;
1666 * nothing more to do for removing pages or no update
1668 if (cpu_buffer
->nr_pages_to_update
<= 0)
1671 * to add pages, make sure all new pages can be
1672 * allocated without receiving ENOMEM
1674 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
1675 if (__rb_allocate_pages(cpu_buffer
->nr_pages_to_update
,
1676 &cpu_buffer
->new_pages
, cpu
)) {
1677 /* not enough memory for new pages */
1685 * Fire off all the required work handlers
1686 * We can't schedule on offline CPUs, but it's not necessary
1687 * since we can change their buffer sizes without any race.
1689 for_each_buffer_cpu(buffer
, cpu
) {
1690 cpu_buffer
= buffer
->buffers
[cpu
];
1691 if (!cpu_buffer
->nr_pages_to_update
)
1694 /* Can't run something on an offline CPU. */
1695 if (!cpu_online(cpu
)) {
1696 rb_update_pages(cpu_buffer
);
1697 cpu_buffer
->nr_pages_to_update
= 0;
1699 schedule_work_on(cpu
,
1700 &cpu_buffer
->update_pages_work
);
1704 /* wait for all the updates to complete */
1705 for_each_buffer_cpu(buffer
, cpu
) {
1706 cpu_buffer
= buffer
->buffers
[cpu
];
1707 if (!cpu_buffer
->nr_pages_to_update
)
1710 if (cpu_online(cpu
))
1711 wait_for_completion(&cpu_buffer
->update_done
);
1712 cpu_buffer
->nr_pages_to_update
= 0;
1717 /* Make sure this CPU has been intitialized */
1718 if (!cpumask_test_cpu(cpu_id
, buffer
->cpumask
))
1721 cpu_buffer
= buffer
->buffers
[cpu_id
];
1723 if (nr_pages
== cpu_buffer
->nr_pages
)
1726 cpu_buffer
->nr_pages_to_update
= nr_pages
-
1727 cpu_buffer
->nr_pages
;
1729 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
1730 if (cpu_buffer
->nr_pages_to_update
> 0 &&
1731 __rb_allocate_pages(cpu_buffer
->nr_pages_to_update
,
1732 &cpu_buffer
->new_pages
, cpu_id
)) {
1739 /* Can't run something on an offline CPU. */
1740 if (!cpu_online(cpu_id
))
1741 rb_update_pages(cpu_buffer
);
1743 schedule_work_on(cpu_id
,
1744 &cpu_buffer
->update_pages_work
);
1745 wait_for_completion(&cpu_buffer
->update_done
);
1748 cpu_buffer
->nr_pages_to_update
= 0;
1754 * The ring buffer resize can happen with the ring buffer
1755 * enabled, so that the update disturbs the tracing as little
1756 * as possible. But if the buffer is disabled, we do not need
1757 * to worry about that, and we can take the time to verify
1758 * that the buffer is not corrupt.
1760 if (atomic_read(&buffer
->record_disabled
)) {
1761 atomic_inc(&buffer
->record_disabled
);
1763 * Even though the buffer was disabled, we must make sure
1764 * that it is truly disabled before calling rb_check_pages.
1765 * There could have been a race between checking
1766 * record_disable and incrementing it.
1768 synchronize_sched();
1769 for_each_buffer_cpu(buffer
, cpu
) {
1770 cpu_buffer
= buffer
->buffers
[cpu
];
1771 rb_check_pages(cpu_buffer
);
1773 atomic_dec(&buffer
->record_disabled
);
1776 mutex_unlock(&buffer
->mutex
);
1780 for_each_buffer_cpu(buffer
, cpu
) {
1781 struct buffer_page
*bpage
, *tmp
;
1783 cpu_buffer
= buffer
->buffers
[cpu
];
1784 cpu_buffer
->nr_pages_to_update
= 0;
1786 if (list_empty(&cpu_buffer
->new_pages
))
1789 list_for_each_entry_safe(bpage
, tmp
, &cpu_buffer
->new_pages
,
1791 list_del_init(&bpage
->list
);
1792 free_buffer_page(bpage
);
1795 mutex_unlock(&buffer
->mutex
);
1798 EXPORT_SYMBOL_GPL(ring_buffer_resize
);
1800 void ring_buffer_change_overwrite(struct ring_buffer
*buffer
, int val
)
1802 mutex_lock(&buffer
->mutex
);
1804 buffer
->flags
|= RB_FL_OVERWRITE
;
1806 buffer
->flags
&= ~RB_FL_OVERWRITE
;
1807 mutex_unlock(&buffer
->mutex
);
1809 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite
);
1811 static __always_inline
void *
1812 __rb_data_page_index(struct buffer_data_page
*bpage
, unsigned index
)
1814 return bpage
->data
+ index
;
1817 static __always_inline
void *__rb_page_index(struct buffer_page
*bpage
, unsigned index
)
1819 return bpage
->page
->data
+ index
;
1822 static __always_inline
struct ring_buffer_event
*
1823 rb_reader_event(struct ring_buffer_per_cpu
*cpu_buffer
)
1825 return __rb_page_index(cpu_buffer
->reader_page
,
1826 cpu_buffer
->reader_page
->read
);
1829 static __always_inline
struct ring_buffer_event
*
1830 rb_iter_head_event(struct ring_buffer_iter
*iter
)
1832 return __rb_page_index(iter
->head_page
, iter
->head
);
1835 static __always_inline
unsigned rb_page_commit(struct buffer_page
*bpage
)
1837 return local_read(&bpage
->page
->commit
);
1840 /* Size is determined by what has been committed */
1841 static __always_inline
unsigned rb_page_size(struct buffer_page
*bpage
)
1843 return rb_page_commit(bpage
);
1846 static __always_inline
unsigned
1847 rb_commit_index(struct ring_buffer_per_cpu
*cpu_buffer
)
1849 return rb_page_commit(cpu_buffer
->commit_page
);
1852 static __always_inline
unsigned
1853 rb_event_index(struct ring_buffer_event
*event
)
1855 unsigned long addr
= (unsigned long)event
;
1857 return (addr
& ~PAGE_MASK
) - BUF_PAGE_HDR_SIZE
;
1860 static void rb_inc_iter(struct ring_buffer_iter
*iter
)
1862 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
1865 * The iterator could be on the reader page (it starts there).
1866 * But the head could have moved, since the reader was
1867 * found. Check for this case and assign the iterator
1868 * to the head page instead of next.
1870 if (iter
->head_page
== cpu_buffer
->reader_page
)
1871 iter
->head_page
= rb_set_head_page(cpu_buffer
);
1873 rb_inc_page(cpu_buffer
, &iter
->head_page
);
1875 iter
->read_stamp
= iter
->head_page
->page
->time_stamp
;
1880 * rb_handle_head_page - writer hit the head page
1882 * Returns: +1 to retry page
1887 rb_handle_head_page(struct ring_buffer_per_cpu
*cpu_buffer
,
1888 struct buffer_page
*tail_page
,
1889 struct buffer_page
*next_page
)
1891 struct buffer_page
*new_head
;
1896 entries
= rb_page_entries(next_page
);
1899 * The hard part is here. We need to move the head
1900 * forward, and protect against both readers on
1901 * other CPUs and writers coming in via interrupts.
1903 type
= rb_head_page_set_update(cpu_buffer
, next_page
, tail_page
,
1907 * type can be one of four:
1908 * NORMAL - an interrupt already moved it for us
1909 * HEAD - we are the first to get here.
1910 * UPDATE - we are the interrupt interrupting
1912 * MOVED - a reader on another CPU moved the next
1913 * pointer to its reader page. Give up
1920 * We changed the head to UPDATE, thus
1921 * it is our responsibility to update
1924 local_add(entries
, &cpu_buffer
->overrun
);
1925 local_sub(BUF_PAGE_SIZE
, &cpu_buffer
->entries_bytes
);
1928 * The entries will be zeroed out when we move the
1932 /* still more to do */
1935 case RB_PAGE_UPDATE
:
1937 * This is an interrupt that interrupt the
1938 * previous update. Still more to do.
1941 case RB_PAGE_NORMAL
:
1943 * An interrupt came in before the update
1944 * and processed this for us.
1945 * Nothing left to do.
1950 * The reader is on another CPU and just did
1951 * a swap with our next_page.
1956 RB_WARN_ON(cpu_buffer
, 1); /* WTF??? */
1961 * Now that we are here, the old head pointer is
1962 * set to UPDATE. This will keep the reader from
1963 * swapping the head page with the reader page.
1964 * The reader (on another CPU) will spin till
1967 * We just need to protect against interrupts
1968 * doing the job. We will set the next pointer
1969 * to HEAD. After that, we set the old pointer
1970 * to NORMAL, but only if it was HEAD before.
1971 * otherwise we are an interrupt, and only
1972 * want the outer most commit to reset it.
1974 new_head
= next_page
;
1975 rb_inc_page(cpu_buffer
, &new_head
);
1977 ret
= rb_head_page_set_head(cpu_buffer
, new_head
, next_page
,
1981 * Valid returns are:
1982 * HEAD - an interrupt came in and already set it.
1983 * NORMAL - One of two things:
1984 * 1) We really set it.
1985 * 2) A bunch of interrupts came in and moved
1986 * the page forward again.
1990 case RB_PAGE_NORMAL
:
1994 RB_WARN_ON(cpu_buffer
, 1);
1999 * It is possible that an interrupt came in,
2000 * set the head up, then more interrupts came in
2001 * and moved it again. When we get back here,
2002 * the page would have been set to NORMAL but we
2003 * just set it back to HEAD.
2005 * How do you detect this? Well, if that happened
2006 * the tail page would have moved.
2008 if (ret
== RB_PAGE_NORMAL
) {
2009 struct buffer_page
*buffer_tail_page
;
2011 buffer_tail_page
= READ_ONCE(cpu_buffer
->tail_page
);
2013 * If the tail had moved passed next, then we need
2014 * to reset the pointer.
2016 if (buffer_tail_page
!= tail_page
&&
2017 buffer_tail_page
!= next_page
)
2018 rb_head_page_set_normal(cpu_buffer
, new_head
,
2024 * If this was the outer most commit (the one that
2025 * changed the original pointer from HEAD to UPDATE),
2026 * then it is up to us to reset it to NORMAL.
2028 if (type
== RB_PAGE_HEAD
) {
2029 ret
= rb_head_page_set_normal(cpu_buffer
, next_page
,
2032 if (RB_WARN_ON(cpu_buffer
,
2033 ret
!= RB_PAGE_UPDATE
))
2041 rb_reset_tail(struct ring_buffer_per_cpu
*cpu_buffer
,
2042 unsigned long tail
, struct rb_event_info
*info
)
2044 struct buffer_page
*tail_page
= info
->tail_page
;
2045 struct ring_buffer_event
*event
;
2046 unsigned long length
= info
->length
;
2049 * Only the event that crossed the page boundary
2050 * must fill the old tail_page with padding.
2052 if (tail
>= BUF_PAGE_SIZE
) {
2054 * If the page was filled, then we still need
2055 * to update the real_end. Reset it to zero
2056 * and the reader will ignore it.
2058 if (tail
== BUF_PAGE_SIZE
)
2059 tail_page
->real_end
= 0;
2061 local_sub(length
, &tail_page
->write
);
2065 event
= __rb_page_index(tail_page
, tail
);
2067 /* account for padding bytes */
2068 local_add(BUF_PAGE_SIZE
- tail
, &cpu_buffer
->entries_bytes
);
2071 * Save the original length to the meta data.
2072 * This will be used by the reader to add lost event
2075 tail_page
->real_end
= tail
;
2078 * If this event is bigger than the minimum size, then
2079 * we need to be careful that we don't subtract the
2080 * write counter enough to allow another writer to slip
2082 * We put in a discarded commit instead, to make sure
2083 * that this space is not used again.
2085 * If we are less than the minimum size, we don't need to
2088 if (tail
> (BUF_PAGE_SIZE
- RB_EVNT_MIN_SIZE
)) {
2089 /* No room for any events */
2091 /* Mark the rest of the page with padding */
2092 rb_event_set_padding(event
);
2094 /* Set the write back to the previous setting */
2095 local_sub(length
, &tail_page
->write
);
2099 /* Put in a discarded event */
2100 event
->array
[0] = (BUF_PAGE_SIZE
- tail
) - RB_EVNT_HDR_SIZE
;
2101 event
->type_len
= RINGBUF_TYPE_PADDING
;
2102 /* time delta must be non zero */
2103 event
->time_delta
= 1;
2105 /* Set write to end of buffer */
2106 length
= (tail
+ length
) - BUF_PAGE_SIZE
;
2107 local_sub(length
, &tail_page
->write
);
2110 static inline void rb_end_commit(struct ring_buffer_per_cpu
*cpu_buffer
);
2113 * This is the slow path, force gcc not to inline it.
2115 static noinline
struct ring_buffer_event
*
2116 rb_move_tail(struct ring_buffer_per_cpu
*cpu_buffer
,
2117 unsigned long tail
, struct rb_event_info
*info
)
2119 struct buffer_page
*tail_page
= info
->tail_page
;
2120 struct buffer_page
*commit_page
= cpu_buffer
->commit_page
;
2121 struct ring_buffer
*buffer
= cpu_buffer
->buffer
;
2122 struct buffer_page
*next_page
;
2125 next_page
= tail_page
;
2127 rb_inc_page(cpu_buffer
, &next_page
);
2130 * If for some reason, we had an interrupt storm that made
2131 * it all the way around the buffer, bail, and warn
2134 if (unlikely(next_page
== commit_page
)) {
2135 local_inc(&cpu_buffer
->commit_overrun
);
2140 * This is where the fun begins!
2142 * We are fighting against races between a reader that
2143 * could be on another CPU trying to swap its reader
2144 * page with the buffer head.
2146 * We are also fighting against interrupts coming in and
2147 * moving the head or tail on us as well.
2149 * If the next page is the head page then we have filled
2150 * the buffer, unless the commit page is still on the
2153 if (rb_is_head_page(cpu_buffer
, next_page
, &tail_page
->list
)) {
2156 * If the commit is not on the reader page, then
2157 * move the header page.
2159 if (!rb_is_reader_page(cpu_buffer
->commit_page
)) {
2161 * If we are not in overwrite mode,
2162 * this is easy, just stop here.
2164 if (!(buffer
->flags
& RB_FL_OVERWRITE
)) {
2165 local_inc(&cpu_buffer
->dropped_events
);
2169 ret
= rb_handle_head_page(cpu_buffer
,
2178 * We need to be careful here too. The
2179 * commit page could still be on the reader
2180 * page. We could have a small buffer, and
2181 * have filled up the buffer with events
2182 * from interrupts and such, and wrapped.
2184 * Note, if the tail page is also the on the
2185 * reader_page, we let it move out.
2187 if (unlikely((cpu_buffer
->commit_page
!=
2188 cpu_buffer
->tail_page
) &&
2189 (cpu_buffer
->commit_page
==
2190 cpu_buffer
->reader_page
))) {
2191 local_inc(&cpu_buffer
->commit_overrun
);
2197 rb_tail_page_update(cpu_buffer
, tail_page
, next_page
);
2201 rb_reset_tail(cpu_buffer
, tail
, info
);
2203 /* Commit what we have for now. */
2204 rb_end_commit(cpu_buffer
);
2205 /* rb_end_commit() decs committing */
2206 local_inc(&cpu_buffer
->committing
);
2208 /* fail and let the caller try again */
2209 return ERR_PTR(-EAGAIN
);
2213 rb_reset_tail(cpu_buffer
, tail
, info
);
2218 /* Slow path, do not inline */
2219 static noinline
struct ring_buffer_event
*
2220 rb_add_time_stamp(struct ring_buffer_event
*event
, u64 delta
)
2222 event
->type_len
= RINGBUF_TYPE_TIME_EXTEND
;
2224 /* Not the first event on the page? */
2225 if (rb_event_index(event
)) {
2226 event
->time_delta
= delta
& TS_MASK
;
2227 event
->array
[0] = delta
>> TS_SHIFT
;
2229 /* nope, just zero it */
2230 event
->time_delta
= 0;
2231 event
->array
[0] = 0;
2234 return skip_time_extend(event
);
2237 static inline bool rb_event_is_commit(struct ring_buffer_per_cpu
*cpu_buffer
,
2238 struct ring_buffer_event
*event
);
2241 * rb_update_event - update event type and data
2242 * @event: the event to update
2243 * @type: the type of event
2244 * @length: the size of the event field in the ring buffer
2246 * Update the type and data fields of the event. The length
2247 * is the actual size that is written to the ring buffer,
2248 * and with this, we can determine what to place into the
2252 rb_update_event(struct ring_buffer_per_cpu
*cpu_buffer
,
2253 struct ring_buffer_event
*event
,
2254 struct rb_event_info
*info
)
2256 unsigned length
= info
->length
;
2257 u64 delta
= info
->delta
;
2259 /* Only a commit updates the timestamp */
2260 if (unlikely(!rb_event_is_commit(cpu_buffer
, event
)))
2264 * If we need to add a timestamp, then we
2265 * add it to the start of the resevered space.
2267 if (unlikely(info
->add_timestamp
)) {
2268 event
= rb_add_time_stamp(event
, delta
);
2269 length
-= RB_LEN_TIME_EXTEND
;
2273 event
->time_delta
= delta
;
2274 length
-= RB_EVNT_HDR_SIZE
;
2275 if (length
> RB_MAX_SMALL_DATA
|| RB_FORCE_8BYTE_ALIGNMENT
) {
2276 event
->type_len
= 0;
2277 event
->array
[0] = length
;
2279 event
->type_len
= DIV_ROUND_UP(length
, RB_ALIGNMENT
);
2282 static unsigned rb_calculate_event_length(unsigned length
)
2284 struct ring_buffer_event event
; /* Used only for sizeof array */
2286 /* zero length can cause confusions */
2290 if (length
> RB_MAX_SMALL_DATA
|| RB_FORCE_8BYTE_ALIGNMENT
)
2291 length
+= sizeof(event
.array
[0]);
2293 length
+= RB_EVNT_HDR_SIZE
;
2294 length
= ALIGN(length
, RB_ARCH_ALIGNMENT
);
2297 * In case the time delta is larger than the 27 bits for it
2298 * in the header, we need to add a timestamp. If another
2299 * event comes in when trying to discard this one to increase
2300 * the length, then the timestamp will be added in the allocated
2301 * space of this event. If length is bigger than the size needed
2302 * for the TIME_EXTEND, then padding has to be used. The events
2303 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2304 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2305 * As length is a multiple of 4, we only need to worry if it
2306 * is 12 (RB_LEN_TIME_EXTEND + 4).
2308 if (length
== RB_LEN_TIME_EXTEND
+ RB_ALIGNMENT
)
2309 length
+= RB_ALIGNMENT
;
2314 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2315 static inline bool sched_clock_stable(void)
2322 rb_try_to_discard(struct ring_buffer_per_cpu
*cpu_buffer
,
2323 struct ring_buffer_event
*event
)
2325 unsigned long new_index
, old_index
;
2326 struct buffer_page
*bpage
;
2327 unsigned long index
;
2330 new_index
= rb_event_index(event
);
2331 old_index
= new_index
+ rb_event_ts_length(event
);
2332 addr
= (unsigned long)event
;
2335 bpage
= READ_ONCE(cpu_buffer
->tail_page
);
2337 if (bpage
->page
== (void *)addr
&& rb_page_write(bpage
) == old_index
) {
2338 unsigned long write_mask
=
2339 local_read(&bpage
->write
) & ~RB_WRITE_MASK
;
2340 unsigned long event_length
= rb_event_length(event
);
2342 * This is on the tail page. It is possible that
2343 * a write could come in and move the tail page
2344 * and write to the next page. That is fine
2345 * because we just shorten what is on this page.
2347 old_index
+= write_mask
;
2348 new_index
+= write_mask
;
2349 index
= local_cmpxchg(&bpage
->write
, old_index
, new_index
);
2350 if (index
== old_index
) {
2351 /* update counters */
2352 local_sub(event_length
, &cpu_buffer
->entries_bytes
);
2357 /* could not discard */
2361 static void rb_start_commit(struct ring_buffer_per_cpu
*cpu_buffer
)
2363 local_inc(&cpu_buffer
->committing
);
2364 local_inc(&cpu_buffer
->commits
);
2367 static __always_inline
void
2368 rb_set_commit_to_write(struct ring_buffer_per_cpu
*cpu_buffer
)
2370 unsigned long max_count
;
2373 * We only race with interrupts and NMIs on this CPU.
2374 * If we own the commit event, then we can commit
2375 * all others that interrupted us, since the interruptions
2376 * are in stack format (they finish before they come
2377 * back to us). This allows us to do a simple loop to
2378 * assign the commit to the tail.
2381 max_count
= cpu_buffer
->nr_pages
* 100;
2383 while (cpu_buffer
->commit_page
!= READ_ONCE(cpu_buffer
->tail_page
)) {
2384 if (RB_WARN_ON(cpu_buffer
, !(--max_count
)))
2386 if (RB_WARN_ON(cpu_buffer
,
2387 rb_is_reader_page(cpu_buffer
->tail_page
)))
2389 local_set(&cpu_buffer
->commit_page
->page
->commit
,
2390 rb_page_write(cpu_buffer
->commit_page
));
2391 rb_inc_page(cpu_buffer
, &cpu_buffer
->commit_page
);
2392 /* Only update the write stamp if the page has an event */
2393 if (rb_page_write(cpu_buffer
->commit_page
))
2394 cpu_buffer
->write_stamp
=
2395 cpu_buffer
->commit_page
->page
->time_stamp
;
2396 /* add barrier to keep gcc from optimizing too much */
2399 while (rb_commit_index(cpu_buffer
) !=
2400 rb_page_write(cpu_buffer
->commit_page
)) {
2402 local_set(&cpu_buffer
->commit_page
->page
->commit
,
2403 rb_page_write(cpu_buffer
->commit_page
));
2404 RB_WARN_ON(cpu_buffer
,
2405 local_read(&cpu_buffer
->commit_page
->page
->commit
) &
2410 /* again, keep gcc from optimizing */
2414 * If an interrupt came in just after the first while loop
2415 * and pushed the tail page forward, we will be left with
2416 * a dangling commit that will never go forward.
2418 if (unlikely(cpu_buffer
->commit_page
!= READ_ONCE(cpu_buffer
->tail_page
)))
2422 static __always_inline
void rb_end_commit(struct ring_buffer_per_cpu
*cpu_buffer
)
2424 unsigned long commits
;
2426 if (RB_WARN_ON(cpu_buffer
,
2427 !local_read(&cpu_buffer
->committing
)))
2431 commits
= local_read(&cpu_buffer
->commits
);
2432 /* synchronize with interrupts */
2434 if (local_read(&cpu_buffer
->committing
) == 1)
2435 rb_set_commit_to_write(cpu_buffer
);
2437 local_dec(&cpu_buffer
->committing
);
2439 /* synchronize with interrupts */
2443 * Need to account for interrupts coming in between the
2444 * updating of the commit page and the clearing of the
2445 * committing counter.
2447 if (unlikely(local_read(&cpu_buffer
->commits
) != commits
) &&
2448 !local_read(&cpu_buffer
->committing
)) {
2449 local_inc(&cpu_buffer
->committing
);
2454 static inline void rb_event_discard(struct ring_buffer_event
*event
)
2456 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
2457 event
= skip_time_extend(event
);
2459 /* array[0] holds the actual length for the discarded event */
2460 event
->array
[0] = rb_event_data_length(event
) - RB_EVNT_HDR_SIZE
;
2461 event
->type_len
= RINGBUF_TYPE_PADDING
;
2462 /* time delta must be non zero */
2463 if (!event
->time_delta
)
2464 event
->time_delta
= 1;
2467 static __always_inline
bool
2468 rb_event_is_commit(struct ring_buffer_per_cpu
*cpu_buffer
,
2469 struct ring_buffer_event
*event
)
2471 unsigned long addr
= (unsigned long)event
;
2472 unsigned long index
;
2474 index
= rb_event_index(event
);
2477 return cpu_buffer
->commit_page
->page
== (void *)addr
&&
2478 rb_commit_index(cpu_buffer
) == index
;
2481 static __always_inline
void
2482 rb_update_write_stamp(struct ring_buffer_per_cpu
*cpu_buffer
,
2483 struct ring_buffer_event
*event
)
2488 * The event first in the commit queue updates the
2491 if (rb_event_is_commit(cpu_buffer
, event
)) {
2493 * A commit event that is first on a page
2494 * updates the write timestamp with the page stamp
2496 if (!rb_event_index(event
))
2497 cpu_buffer
->write_stamp
=
2498 cpu_buffer
->commit_page
->page
->time_stamp
;
2499 else if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
) {
2500 delta
= event
->array
[0];
2502 delta
+= event
->time_delta
;
2503 cpu_buffer
->write_stamp
+= delta
;
2505 cpu_buffer
->write_stamp
+= event
->time_delta
;
2509 static void rb_commit(struct ring_buffer_per_cpu
*cpu_buffer
,
2510 struct ring_buffer_event
*event
)
2512 local_inc(&cpu_buffer
->entries
);
2513 rb_update_write_stamp(cpu_buffer
, event
);
2514 rb_end_commit(cpu_buffer
);
2517 static __always_inline
void
2518 rb_wakeups(struct ring_buffer
*buffer
, struct ring_buffer_per_cpu
*cpu_buffer
)
2522 if (buffer
->irq_work
.waiters_pending
) {
2523 buffer
->irq_work
.waiters_pending
= false;
2524 /* irq_work_queue() supplies it's own memory barriers */
2525 irq_work_queue(&buffer
->irq_work
.work
);
2528 if (cpu_buffer
->irq_work
.waiters_pending
) {
2529 cpu_buffer
->irq_work
.waiters_pending
= false;
2530 /* irq_work_queue() supplies it's own memory barriers */
2531 irq_work_queue(&cpu_buffer
->irq_work
.work
);
2534 pagebusy
= cpu_buffer
->reader_page
== cpu_buffer
->commit_page
;
2536 if (!pagebusy
&& cpu_buffer
->irq_work
.full_waiters_pending
) {
2537 cpu_buffer
->irq_work
.wakeup_full
= true;
2538 cpu_buffer
->irq_work
.full_waiters_pending
= false;
2539 /* irq_work_queue() supplies it's own memory barriers */
2540 irq_work_queue(&cpu_buffer
->irq_work
.work
);
2545 * The lock and unlock are done within a preempt disable section.
2546 * The current_context per_cpu variable can only be modified
2547 * by the current task between lock and unlock. But it can
2548 * be modified more than once via an interrupt. To pass this
2549 * information from the lock to the unlock without having to
2550 * access the 'in_interrupt()' functions again (which do show
2551 * a bit of overhead in something as critical as function tracing,
2552 * we use a bitmask trick.
2554 * bit 0 = NMI context
2555 * bit 1 = IRQ context
2556 * bit 2 = SoftIRQ context
2557 * bit 3 = normal context.
2559 * This works because this is the order of contexts that can
2560 * preempt other contexts. A SoftIRQ never preempts an IRQ
2563 * When the context is determined, the corresponding bit is
2564 * checked and set (if it was set, then a recursion of that context
2567 * On unlock, we need to clear this bit. To do so, just subtract
2568 * 1 from the current_context and AND it to itself.
2572 * 101 & 100 = 100 (clearing bit zero)
2575 * 1010 & 1001 = 1000 (clearing bit 1)
2577 * The least significant bit can be cleared this way, and it
2578 * just so happens that it is the same bit corresponding to
2579 * the current context.
2582 static __always_inline
int
2583 trace_recursive_lock(struct ring_buffer_per_cpu
*cpu_buffer
)
2585 unsigned int val
= cpu_buffer
->current_context
;
2588 if (in_interrupt()) {
2594 bit
= RB_CTX_SOFTIRQ
;
2596 bit
= RB_CTX_NORMAL
;
2598 if (unlikely(val
& (1 << bit
)))
2602 cpu_buffer
->current_context
= val
;
2607 static __always_inline
void
2608 trace_recursive_unlock(struct ring_buffer_per_cpu
*cpu_buffer
)
2610 cpu_buffer
->current_context
&= cpu_buffer
->current_context
- 1;
2614 * ring_buffer_unlock_commit - commit a reserved
2615 * @buffer: The buffer to commit to
2616 * @event: The event pointer to commit.
2618 * This commits the data to the ring buffer, and releases any locks held.
2620 * Must be paired with ring_buffer_lock_reserve.
2622 int ring_buffer_unlock_commit(struct ring_buffer
*buffer
,
2623 struct ring_buffer_event
*event
)
2625 struct ring_buffer_per_cpu
*cpu_buffer
;
2626 int cpu
= raw_smp_processor_id();
2628 cpu_buffer
= buffer
->buffers
[cpu
];
2630 rb_commit(cpu_buffer
, event
);
2632 rb_wakeups(buffer
, cpu_buffer
);
2634 trace_recursive_unlock(cpu_buffer
);
2636 preempt_enable_notrace();
2640 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit
);
2642 static noinline
void
2643 rb_handle_timestamp(struct ring_buffer_per_cpu
*cpu_buffer
,
2644 struct rb_event_info
*info
)
2646 WARN_ONCE(info
->delta
> (1ULL << 59),
2647 KERN_WARNING
"Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2648 (unsigned long long)info
->delta
,
2649 (unsigned long long)info
->ts
,
2650 (unsigned long long)cpu_buffer
->write_stamp
,
2651 sched_clock_stable() ? "" :
2652 "If you just came from a suspend/resume,\n"
2653 "please switch to the trace global clock:\n"
2654 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2655 info
->add_timestamp
= 1;
2658 static struct ring_buffer_event
*
2659 __rb_reserve_next(struct ring_buffer_per_cpu
*cpu_buffer
,
2660 struct rb_event_info
*info
)
2662 struct ring_buffer_event
*event
;
2663 struct buffer_page
*tail_page
;
2664 unsigned long tail
, write
;
2667 * If the time delta since the last event is too big to
2668 * hold in the time field of the event, then we append a
2669 * TIME EXTEND event ahead of the data event.
2671 if (unlikely(info
->add_timestamp
))
2672 info
->length
+= RB_LEN_TIME_EXTEND
;
2674 /* Don't let the compiler play games with cpu_buffer->tail_page */
2675 tail_page
= info
->tail_page
= READ_ONCE(cpu_buffer
->tail_page
);
2676 write
= local_add_return(info
->length
, &tail_page
->write
);
2678 /* set write to only the index of the write */
2679 write
&= RB_WRITE_MASK
;
2680 tail
= write
- info
->length
;
2683 * If this is the first commit on the page, then it has the same
2684 * timestamp as the page itself.
2689 /* See if we shot pass the end of this buffer page */
2690 if (unlikely(write
> BUF_PAGE_SIZE
))
2691 return rb_move_tail(cpu_buffer
, tail
, info
);
2693 /* We reserved something on the buffer */
2695 event
= __rb_page_index(tail_page
, tail
);
2696 rb_update_event(cpu_buffer
, event
, info
);
2698 local_inc(&tail_page
->entries
);
2701 * If this is the first commit on the page, then update
2705 tail_page
->page
->time_stamp
= info
->ts
;
2707 /* account for these added bytes */
2708 local_add(info
->length
, &cpu_buffer
->entries_bytes
);
2713 static __always_inline
struct ring_buffer_event
*
2714 rb_reserve_next_event(struct ring_buffer
*buffer
,
2715 struct ring_buffer_per_cpu
*cpu_buffer
,
2716 unsigned long length
)
2718 struct ring_buffer_event
*event
;
2719 struct rb_event_info info
;
2723 rb_start_commit(cpu_buffer
);
2725 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2727 * Due to the ability to swap a cpu buffer from a buffer
2728 * it is possible it was swapped before we committed.
2729 * (committing stops a swap). We check for it here and
2730 * if it happened, we have to fail the write.
2733 if (unlikely(ACCESS_ONCE(cpu_buffer
->buffer
) != buffer
)) {
2734 local_dec(&cpu_buffer
->committing
);
2735 local_dec(&cpu_buffer
->commits
);
2740 info
.length
= rb_calculate_event_length(length
);
2742 info
.add_timestamp
= 0;
2746 * We allow for interrupts to reenter here and do a trace.
2747 * If one does, it will cause this original code to loop
2748 * back here. Even with heavy interrupts happening, this
2749 * should only happen a few times in a row. If this happens
2750 * 1000 times in a row, there must be either an interrupt
2751 * storm or we have something buggy.
2754 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 1000))
2757 info
.ts
= rb_time_stamp(cpu_buffer
->buffer
);
2758 diff
= info
.ts
- cpu_buffer
->write_stamp
;
2760 /* make sure this diff is calculated here */
2763 /* Did the write stamp get updated already? */
2764 if (likely(info
.ts
>= cpu_buffer
->write_stamp
)) {
2766 if (unlikely(test_time_stamp(info
.delta
)))
2767 rb_handle_timestamp(cpu_buffer
, &info
);
2770 event
= __rb_reserve_next(cpu_buffer
, &info
);
2772 if (unlikely(PTR_ERR(event
) == -EAGAIN
)) {
2773 if (info
.add_timestamp
)
2774 info
.length
-= RB_LEN_TIME_EXTEND
;
2784 rb_end_commit(cpu_buffer
);
2789 * ring_buffer_lock_reserve - reserve a part of the buffer
2790 * @buffer: the ring buffer to reserve from
2791 * @length: the length of the data to reserve (excluding event header)
2793 * Returns a reseverd event on the ring buffer to copy directly to.
2794 * The user of this interface will need to get the body to write into
2795 * and can use the ring_buffer_event_data() interface.
2797 * The length is the length of the data needed, not the event length
2798 * which also includes the event header.
2800 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2801 * If NULL is returned, then nothing has been allocated or locked.
2803 struct ring_buffer_event
*
2804 ring_buffer_lock_reserve(struct ring_buffer
*buffer
, unsigned long length
)
2806 struct ring_buffer_per_cpu
*cpu_buffer
;
2807 struct ring_buffer_event
*event
;
2810 /* If we are tracing schedule, we don't want to recurse */
2811 preempt_disable_notrace();
2813 if (unlikely(atomic_read(&buffer
->record_disabled
)))
2816 cpu
= raw_smp_processor_id();
2818 if (unlikely(!cpumask_test_cpu(cpu
, buffer
->cpumask
)))
2821 cpu_buffer
= buffer
->buffers
[cpu
];
2823 if (unlikely(atomic_read(&cpu_buffer
->record_disabled
)))
2826 if (unlikely(length
> BUF_MAX_DATA_SIZE
))
2829 if (unlikely(trace_recursive_lock(cpu_buffer
)))
2832 event
= rb_reserve_next_event(buffer
, cpu_buffer
, length
);
2839 trace_recursive_unlock(cpu_buffer
);
2841 preempt_enable_notrace();
2844 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve
);
2847 * Decrement the entries to the page that an event is on.
2848 * The event does not even need to exist, only the pointer
2849 * to the page it is on. This may only be called before the commit
2853 rb_decrement_entry(struct ring_buffer_per_cpu
*cpu_buffer
,
2854 struct ring_buffer_event
*event
)
2856 unsigned long addr
= (unsigned long)event
;
2857 struct buffer_page
*bpage
= cpu_buffer
->commit_page
;
2858 struct buffer_page
*start
;
2862 /* Do the likely case first */
2863 if (likely(bpage
->page
== (void *)addr
)) {
2864 local_dec(&bpage
->entries
);
2869 * Because the commit page may be on the reader page we
2870 * start with the next page and check the end loop there.
2872 rb_inc_page(cpu_buffer
, &bpage
);
2875 if (bpage
->page
== (void *)addr
) {
2876 local_dec(&bpage
->entries
);
2879 rb_inc_page(cpu_buffer
, &bpage
);
2880 } while (bpage
!= start
);
2882 /* commit not part of this buffer?? */
2883 RB_WARN_ON(cpu_buffer
, 1);
2887 * ring_buffer_commit_discard - discard an event that has not been committed
2888 * @buffer: the ring buffer
2889 * @event: non committed event to discard
2891 * Sometimes an event that is in the ring buffer needs to be ignored.
2892 * This function lets the user discard an event in the ring buffer
2893 * and then that event will not be read later.
2895 * This function only works if it is called before the the item has been
2896 * committed. It will try to free the event from the ring buffer
2897 * if another event has not been added behind it.
2899 * If another event has been added behind it, it will set the event
2900 * up as discarded, and perform the commit.
2902 * If this function is called, do not call ring_buffer_unlock_commit on
2905 void ring_buffer_discard_commit(struct ring_buffer
*buffer
,
2906 struct ring_buffer_event
*event
)
2908 struct ring_buffer_per_cpu
*cpu_buffer
;
2911 /* The event is discarded regardless */
2912 rb_event_discard(event
);
2914 cpu
= smp_processor_id();
2915 cpu_buffer
= buffer
->buffers
[cpu
];
2918 * This must only be called if the event has not been
2919 * committed yet. Thus we can assume that preemption
2920 * is still disabled.
2922 RB_WARN_ON(buffer
, !local_read(&cpu_buffer
->committing
));
2924 rb_decrement_entry(cpu_buffer
, event
);
2925 if (rb_try_to_discard(cpu_buffer
, event
))
2929 * The commit is still visible by the reader, so we
2930 * must still update the timestamp.
2932 rb_update_write_stamp(cpu_buffer
, event
);
2934 rb_end_commit(cpu_buffer
);
2936 trace_recursive_unlock(cpu_buffer
);
2938 preempt_enable_notrace();
2941 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit
);
2944 * ring_buffer_write - write data to the buffer without reserving
2945 * @buffer: The ring buffer to write to.
2946 * @length: The length of the data being written (excluding the event header)
2947 * @data: The data to write to the buffer.
2949 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2950 * one function. If you already have the data to write to the buffer, it
2951 * may be easier to simply call this function.
2953 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2954 * and not the length of the event which would hold the header.
2956 int ring_buffer_write(struct ring_buffer
*buffer
,
2957 unsigned long length
,
2960 struct ring_buffer_per_cpu
*cpu_buffer
;
2961 struct ring_buffer_event
*event
;
2966 preempt_disable_notrace();
2968 if (atomic_read(&buffer
->record_disabled
))
2971 cpu
= raw_smp_processor_id();
2973 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
2976 cpu_buffer
= buffer
->buffers
[cpu
];
2978 if (atomic_read(&cpu_buffer
->record_disabled
))
2981 if (length
> BUF_MAX_DATA_SIZE
)
2984 if (unlikely(trace_recursive_lock(cpu_buffer
)))
2987 event
= rb_reserve_next_event(buffer
, cpu_buffer
, length
);
2991 body
= rb_event_data(event
);
2993 memcpy(body
, data
, length
);
2995 rb_commit(cpu_buffer
, event
);
2997 rb_wakeups(buffer
, cpu_buffer
);
3002 trace_recursive_unlock(cpu_buffer
);
3005 preempt_enable_notrace();
3009 EXPORT_SYMBOL_GPL(ring_buffer_write
);
3011 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu
*cpu_buffer
)
3013 struct buffer_page
*reader
= cpu_buffer
->reader_page
;
3014 struct buffer_page
*head
= rb_set_head_page(cpu_buffer
);
3015 struct buffer_page
*commit
= cpu_buffer
->commit_page
;
3017 /* In case of error, head will be NULL */
3018 if (unlikely(!head
))
3021 return reader
->read
== rb_page_commit(reader
) &&
3022 (commit
== reader
||
3024 head
->read
== rb_page_commit(commit
)));
3028 * ring_buffer_record_disable - stop all writes into the buffer
3029 * @buffer: The ring buffer to stop writes to.
3031 * This prevents all writes to the buffer. Any attempt to write
3032 * to the buffer after this will fail and return NULL.
3034 * The caller should call synchronize_sched() after this.
3036 void ring_buffer_record_disable(struct ring_buffer
*buffer
)
3038 atomic_inc(&buffer
->record_disabled
);
3040 EXPORT_SYMBOL_GPL(ring_buffer_record_disable
);
3043 * ring_buffer_record_enable - enable writes to the buffer
3044 * @buffer: The ring buffer to enable writes
3046 * Note, multiple disables will need the same number of enables
3047 * to truly enable the writing (much like preempt_disable).
3049 void ring_buffer_record_enable(struct ring_buffer
*buffer
)
3051 atomic_dec(&buffer
->record_disabled
);
3053 EXPORT_SYMBOL_GPL(ring_buffer_record_enable
);
3056 * ring_buffer_record_off - stop all writes into the buffer
3057 * @buffer: The ring buffer to stop writes to.
3059 * This prevents all writes to the buffer. Any attempt to write
3060 * to the buffer after this will fail and return NULL.
3062 * This is different than ring_buffer_record_disable() as
3063 * it works like an on/off switch, where as the disable() version
3064 * must be paired with a enable().
3066 void ring_buffer_record_off(struct ring_buffer
*buffer
)
3069 unsigned int new_rd
;
3072 rd
= atomic_read(&buffer
->record_disabled
);
3073 new_rd
= rd
| RB_BUFFER_OFF
;
3074 } while (atomic_cmpxchg(&buffer
->record_disabled
, rd
, new_rd
) != rd
);
3076 EXPORT_SYMBOL_GPL(ring_buffer_record_off
);
3079 * ring_buffer_record_on - restart writes into the buffer
3080 * @buffer: The ring buffer to start writes to.
3082 * This enables all writes to the buffer that was disabled by
3083 * ring_buffer_record_off().
3085 * This is different than ring_buffer_record_enable() as
3086 * it works like an on/off switch, where as the enable() version
3087 * must be paired with a disable().
3089 void ring_buffer_record_on(struct ring_buffer
*buffer
)
3092 unsigned int new_rd
;
3095 rd
= atomic_read(&buffer
->record_disabled
);
3096 new_rd
= rd
& ~RB_BUFFER_OFF
;
3097 } while (atomic_cmpxchg(&buffer
->record_disabled
, rd
, new_rd
) != rd
);
3099 EXPORT_SYMBOL_GPL(ring_buffer_record_on
);
3102 * ring_buffer_record_is_on - return true if the ring buffer can write
3103 * @buffer: The ring buffer to see if write is enabled
3105 * Returns true if the ring buffer is in a state that it accepts writes.
3107 int ring_buffer_record_is_on(struct ring_buffer
*buffer
)
3109 return !atomic_read(&buffer
->record_disabled
);
3113 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
3114 * @buffer: The ring buffer to see if write is set enabled
3116 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
3117 * Note that this does NOT mean it is in a writable state.
3119 * It may return true when the ring buffer has been disabled by
3120 * ring_buffer_record_disable(), as that is a temporary disabling of
3123 int ring_buffer_record_is_set_on(struct ring_buffer
*buffer
)
3125 return !(atomic_read(&buffer
->record_disabled
) & RB_BUFFER_OFF
);
3129 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3130 * @buffer: The ring buffer to stop writes to.
3131 * @cpu: The CPU buffer to stop
3133 * This prevents all writes to the buffer. Any attempt to write
3134 * to the buffer after this will fail and return NULL.
3136 * The caller should call synchronize_sched() after this.
3138 void ring_buffer_record_disable_cpu(struct ring_buffer
*buffer
, int cpu
)
3140 struct ring_buffer_per_cpu
*cpu_buffer
;
3142 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3145 cpu_buffer
= buffer
->buffers
[cpu
];
3146 atomic_inc(&cpu_buffer
->record_disabled
);
3148 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu
);
3151 * ring_buffer_record_enable_cpu - enable writes to the buffer
3152 * @buffer: The ring buffer to enable writes
3153 * @cpu: The CPU to enable.
3155 * Note, multiple disables will need the same number of enables
3156 * to truly enable the writing (much like preempt_disable).
3158 void ring_buffer_record_enable_cpu(struct ring_buffer
*buffer
, int cpu
)
3160 struct ring_buffer_per_cpu
*cpu_buffer
;
3162 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3165 cpu_buffer
= buffer
->buffers
[cpu
];
3166 atomic_dec(&cpu_buffer
->record_disabled
);
3168 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu
);
3171 * The total entries in the ring buffer is the running counter
3172 * of entries entered into the ring buffer, minus the sum of
3173 * the entries read from the ring buffer and the number of
3174 * entries that were overwritten.
3176 static inline unsigned long
3177 rb_num_of_entries(struct ring_buffer_per_cpu
*cpu_buffer
)
3179 return local_read(&cpu_buffer
->entries
) -
3180 (local_read(&cpu_buffer
->overrun
) + cpu_buffer
->read
);
3184 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3185 * @buffer: The ring buffer
3186 * @cpu: The per CPU buffer to read from.
3188 u64
ring_buffer_oldest_event_ts(struct ring_buffer
*buffer
, int cpu
)
3190 unsigned long flags
;
3191 struct ring_buffer_per_cpu
*cpu_buffer
;
3192 struct buffer_page
*bpage
;
3195 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3198 cpu_buffer
= buffer
->buffers
[cpu
];
3199 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3201 * if the tail is on reader_page, oldest time stamp is on the reader
3204 if (cpu_buffer
->tail_page
== cpu_buffer
->reader_page
)
3205 bpage
= cpu_buffer
->reader_page
;
3207 bpage
= rb_set_head_page(cpu_buffer
);
3209 ret
= bpage
->page
->time_stamp
;
3210 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3214 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts
);
3217 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3218 * @buffer: The ring buffer
3219 * @cpu: The per CPU buffer to read from.
3221 unsigned long ring_buffer_bytes_cpu(struct ring_buffer
*buffer
, int cpu
)
3223 struct ring_buffer_per_cpu
*cpu_buffer
;
3226 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3229 cpu_buffer
= buffer
->buffers
[cpu
];
3230 ret
= local_read(&cpu_buffer
->entries_bytes
) - cpu_buffer
->read_bytes
;
3234 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu
);
3237 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3238 * @buffer: The ring buffer
3239 * @cpu: The per CPU buffer to get the entries from.
3241 unsigned long ring_buffer_entries_cpu(struct ring_buffer
*buffer
, int cpu
)
3243 struct ring_buffer_per_cpu
*cpu_buffer
;
3245 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3248 cpu_buffer
= buffer
->buffers
[cpu
];
3250 return rb_num_of_entries(cpu_buffer
);
3252 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu
);
3255 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3256 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3257 * @buffer: The ring buffer
3258 * @cpu: The per CPU buffer to get the number of overruns from
3260 unsigned long ring_buffer_overrun_cpu(struct ring_buffer
*buffer
, int cpu
)
3262 struct ring_buffer_per_cpu
*cpu_buffer
;
3265 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3268 cpu_buffer
= buffer
->buffers
[cpu
];
3269 ret
= local_read(&cpu_buffer
->overrun
);
3273 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu
);
3276 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3277 * commits failing due to the buffer wrapping around while there are uncommitted
3278 * events, such as during an interrupt storm.
3279 * @buffer: The ring buffer
3280 * @cpu: The per CPU buffer to get the number of overruns from
3283 ring_buffer_commit_overrun_cpu(struct ring_buffer
*buffer
, int cpu
)
3285 struct ring_buffer_per_cpu
*cpu_buffer
;
3288 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3291 cpu_buffer
= buffer
->buffers
[cpu
];
3292 ret
= local_read(&cpu_buffer
->commit_overrun
);
3296 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu
);
3299 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3300 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3301 * @buffer: The ring buffer
3302 * @cpu: The per CPU buffer to get the number of overruns from
3305 ring_buffer_dropped_events_cpu(struct ring_buffer
*buffer
, int cpu
)
3307 struct ring_buffer_per_cpu
*cpu_buffer
;
3310 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3313 cpu_buffer
= buffer
->buffers
[cpu
];
3314 ret
= local_read(&cpu_buffer
->dropped_events
);
3318 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu
);
3321 * ring_buffer_read_events_cpu - get the number of events successfully read
3322 * @buffer: The ring buffer
3323 * @cpu: The per CPU buffer to get the number of events read
3326 ring_buffer_read_events_cpu(struct ring_buffer
*buffer
, int cpu
)
3328 struct ring_buffer_per_cpu
*cpu_buffer
;
3330 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3333 cpu_buffer
= buffer
->buffers
[cpu
];
3334 return cpu_buffer
->read
;
3336 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu
);
3339 * ring_buffer_entries - get the number of entries in a buffer
3340 * @buffer: The ring buffer
3342 * Returns the total number of entries in the ring buffer
3345 unsigned long ring_buffer_entries(struct ring_buffer
*buffer
)
3347 struct ring_buffer_per_cpu
*cpu_buffer
;
3348 unsigned long entries
= 0;
3351 /* if you care about this being correct, lock the buffer */
3352 for_each_buffer_cpu(buffer
, cpu
) {
3353 cpu_buffer
= buffer
->buffers
[cpu
];
3354 entries
+= rb_num_of_entries(cpu_buffer
);
3359 EXPORT_SYMBOL_GPL(ring_buffer_entries
);
3362 * ring_buffer_overruns - get the number of overruns in buffer
3363 * @buffer: The ring buffer
3365 * Returns the total number of overruns in the ring buffer
3368 unsigned long ring_buffer_overruns(struct ring_buffer
*buffer
)
3370 struct ring_buffer_per_cpu
*cpu_buffer
;
3371 unsigned long overruns
= 0;
3374 /* if you care about this being correct, lock the buffer */
3375 for_each_buffer_cpu(buffer
, cpu
) {
3376 cpu_buffer
= buffer
->buffers
[cpu
];
3377 overruns
+= local_read(&cpu_buffer
->overrun
);
3382 EXPORT_SYMBOL_GPL(ring_buffer_overruns
);
3384 static void rb_iter_reset(struct ring_buffer_iter
*iter
)
3386 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
3388 /* Iterator usage is expected to have record disabled */
3389 iter
->head_page
= cpu_buffer
->reader_page
;
3390 iter
->head
= cpu_buffer
->reader_page
->read
;
3392 iter
->cache_reader_page
= iter
->head_page
;
3393 iter
->cache_read
= cpu_buffer
->read
;
3396 iter
->read_stamp
= cpu_buffer
->read_stamp
;
3398 iter
->read_stamp
= iter
->head_page
->page
->time_stamp
;
3402 * ring_buffer_iter_reset - reset an iterator
3403 * @iter: The iterator to reset
3405 * Resets the iterator, so that it will start from the beginning
3408 void ring_buffer_iter_reset(struct ring_buffer_iter
*iter
)
3410 struct ring_buffer_per_cpu
*cpu_buffer
;
3411 unsigned long flags
;
3416 cpu_buffer
= iter
->cpu_buffer
;
3418 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3419 rb_iter_reset(iter
);
3420 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3422 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset
);
3425 * ring_buffer_iter_empty - check if an iterator has no more to read
3426 * @iter: The iterator to check
3428 int ring_buffer_iter_empty(struct ring_buffer_iter
*iter
)
3430 struct ring_buffer_per_cpu
*cpu_buffer
;
3431 struct buffer_page
*reader
;
3432 struct buffer_page
*head_page
;
3433 struct buffer_page
*commit_page
;
3436 cpu_buffer
= iter
->cpu_buffer
;
3438 /* Remember, trace recording is off when iterator is in use */
3439 reader
= cpu_buffer
->reader_page
;
3440 head_page
= cpu_buffer
->head_page
;
3441 commit_page
= cpu_buffer
->commit_page
;
3442 commit
= rb_page_commit(commit_page
);
3444 return ((iter
->head_page
== commit_page
&& iter
->head
== commit
) ||
3445 (iter
->head_page
== reader
&& commit_page
== head_page
&&
3446 head_page
->read
== commit
&&
3447 iter
->head
== rb_page_commit(cpu_buffer
->reader_page
)));
3449 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty
);
3452 rb_update_read_stamp(struct ring_buffer_per_cpu
*cpu_buffer
,
3453 struct ring_buffer_event
*event
)
3457 switch (event
->type_len
) {
3458 case RINGBUF_TYPE_PADDING
:
3461 case RINGBUF_TYPE_TIME_EXTEND
:
3462 delta
= event
->array
[0];
3464 delta
+= event
->time_delta
;
3465 cpu_buffer
->read_stamp
+= delta
;
3468 case RINGBUF_TYPE_TIME_STAMP
:
3469 /* FIXME: not implemented */
3472 case RINGBUF_TYPE_DATA
:
3473 cpu_buffer
->read_stamp
+= event
->time_delta
;
3483 rb_update_iter_read_stamp(struct ring_buffer_iter
*iter
,
3484 struct ring_buffer_event
*event
)
3488 switch (event
->type_len
) {
3489 case RINGBUF_TYPE_PADDING
:
3492 case RINGBUF_TYPE_TIME_EXTEND
:
3493 delta
= event
->array
[0];
3495 delta
+= event
->time_delta
;
3496 iter
->read_stamp
+= delta
;
3499 case RINGBUF_TYPE_TIME_STAMP
:
3500 /* FIXME: not implemented */
3503 case RINGBUF_TYPE_DATA
:
3504 iter
->read_stamp
+= event
->time_delta
;
3513 static struct buffer_page
*
3514 rb_get_reader_page(struct ring_buffer_per_cpu
*cpu_buffer
)
3516 struct buffer_page
*reader
= NULL
;
3517 unsigned long overwrite
;
3518 unsigned long flags
;
3522 local_irq_save(flags
);
3523 arch_spin_lock(&cpu_buffer
->lock
);
3527 * This should normally only loop twice. But because the
3528 * start of the reader inserts an empty page, it causes
3529 * a case where we will loop three times. There should be no
3530 * reason to loop four times (that I know of).
3532 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 3)) {
3537 reader
= cpu_buffer
->reader_page
;
3539 /* If there's more to read, return this page */
3540 if (cpu_buffer
->reader_page
->read
< rb_page_size(reader
))
3543 /* Never should we have an index greater than the size */
3544 if (RB_WARN_ON(cpu_buffer
,
3545 cpu_buffer
->reader_page
->read
> rb_page_size(reader
)))
3548 /* check if we caught up to the tail */
3550 if (cpu_buffer
->commit_page
== cpu_buffer
->reader_page
)
3553 /* Don't bother swapping if the ring buffer is empty */
3554 if (rb_num_of_entries(cpu_buffer
) == 0)
3558 * Reset the reader page to size zero.
3560 local_set(&cpu_buffer
->reader_page
->write
, 0);
3561 local_set(&cpu_buffer
->reader_page
->entries
, 0);
3562 local_set(&cpu_buffer
->reader_page
->page
->commit
, 0);
3563 cpu_buffer
->reader_page
->real_end
= 0;
3567 * Splice the empty reader page into the list around the head.
3569 reader
= rb_set_head_page(cpu_buffer
);
3572 cpu_buffer
->reader_page
->list
.next
= rb_list_head(reader
->list
.next
);
3573 cpu_buffer
->reader_page
->list
.prev
= reader
->list
.prev
;
3576 * cpu_buffer->pages just needs to point to the buffer, it
3577 * has no specific buffer page to point to. Lets move it out
3578 * of our way so we don't accidentally swap it.
3580 cpu_buffer
->pages
= reader
->list
.prev
;
3582 /* The reader page will be pointing to the new head */
3583 rb_set_list_to_head(cpu_buffer
, &cpu_buffer
->reader_page
->list
);
3586 * We want to make sure we read the overruns after we set up our
3587 * pointers to the next object. The writer side does a
3588 * cmpxchg to cross pages which acts as the mb on the writer
3589 * side. Note, the reader will constantly fail the swap
3590 * while the writer is updating the pointers, so this
3591 * guarantees that the overwrite recorded here is the one we
3592 * want to compare with the last_overrun.
3595 overwrite
= local_read(&(cpu_buffer
->overrun
));
3598 * Here's the tricky part.
3600 * We need to move the pointer past the header page.
3601 * But we can only do that if a writer is not currently
3602 * moving it. The page before the header page has the
3603 * flag bit '1' set if it is pointing to the page we want.
3604 * but if the writer is in the process of moving it
3605 * than it will be '2' or already moved '0'.
3608 ret
= rb_head_page_replace(reader
, cpu_buffer
->reader_page
);
3611 * If we did not convert it, then we must try again.
3617 * Yeah! We succeeded in replacing the page.
3619 * Now make the new head point back to the reader page.
3621 rb_list_head(reader
->list
.next
)->prev
= &cpu_buffer
->reader_page
->list
;
3622 rb_inc_page(cpu_buffer
, &cpu_buffer
->head_page
);
3624 /* Finally update the reader page to the new head */
3625 cpu_buffer
->reader_page
= reader
;
3626 cpu_buffer
->reader_page
->read
= 0;
3628 if (overwrite
!= cpu_buffer
->last_overrun
) {
3629 cpu_buffer
->lost_events
= overwrite
- cpu_buffer
->last_overrun
;
3630 cpu_buffer
->last_overrun
= overwrite
;
3636 /* Update the read_stamp on the first event */
3637 if (reader
&& reader
->read
== 0)
3638 cpu_buffer
->read_stamp
= reader
->page
->time_stamp
;
3640 arch_spin_unlock(&cpu_buffer
->lock
);
3641 local_irq_restore(flags
);
3646 static void rb_advance_reader(struct ring_buffer_per_cpu
*cpu_buffer
)
3648 struct ring_buffer_event
*event
;
3649 struct buffer_page
*reader
;
3652 reader
= rb_get_reader_page(cpu_buffer
);
3654 /* This function should not be called when buffer is empty */
3655 if (RB_WARN_ON(cpu_buffer
, !reader
))
3658 event
= rb_reader_event(cpu_buffer
);
3660 if (event
->type_len
<= RINGBUF_TYPE_DATA_TYPE_LEN_MAX
)
3663 rb_update_read_stamp(cpu_buffer
, event
);
3665 length
= rb_event_length(event
);
3666 cpu_buffer
->reader_page
->read
+= length
;
3669 static void rb_advance_iter(struct ring_buffer_iter
*iter
)
3671 struct ring_buffer_per_cpu
*cpu_buffer
;
3672 struct ring_buffer_event
*event
;
3675 cpu_buffer
= iter
->cpu_buffer
;
3678 * Check if we are at the end of the buffer.
3680 if (iter
->head
>= rb_page_size(iter
->head_page
)) {
3681 /* discarded commits can make the page empty */
3682 if (iter
->head_page
== cpu_buffer
->commit_page
)
3688 event
= rb_iter_head_event(iter
);
3690 length
= rb_event_length(event
);
3693 * This should not be called to advance the header if we are
3694 * at the tail of the buffer.
3696 if (RB_WARN_ON(cpu_buffer
,
3697 (iter
->head_page
== cpu_buffer
->commit_page
) &&
3698 (iter
->head
+ length
> rb_commit_index(cpu_buffer
))))
3701 rb_update_iter_read_stamp(iter
, event
);
3703 iter
->head
+= length
;
3705 /* check for end of page padding */
3706 if ((iter
->head
>= rb_page_size(iter
->head_page
)) &&
3707 (iter
->head_page
!= cpu_buffer
->commit_page
))
3711 static int rb_lost_events(struct ring_buffer_per_cpu
*cpu_buffer
)
3713 return cpu_buffer
->lost_events
;
3716 static struct ring_buffer_event
*
3717 rb_buffer_peek(struct ring_buffer_per_cpu
*cpu_buffer
, u64
*ts
,
3718 unsigned long *lost_events
)
3720 struct ring_buffer_event
*event
;
3721 struct buffer_page
*reader
;
3726 * We repeat when a time extend is encountered.
3727 * Since the time extend is always attached to a data event,
3728 * we should never loop more than once.
3729 * (We never hit the following condition more than twice).
3731 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 2))
3734 reader
= rb_get_reader_page(cpu_buffer
);
3738 event
= rb_reader_event(cpu_buffer
);
3740 switch (event
->type_len
) {
3741 case RINGBUF_TYPE_PADDING
:
3742 if (rb_null_event(event
))
3743 RB_WARN_ON(cpu_buffer
, 1);
3745 * Because the writer could be discarding every
3746 * event it creates (which would probably be bad)
3747 * if we were to go back to "again" then we may never
3748 * catch up, and will trigger the warn on, or lock
3749 * the box. Return the padding, and we will release
3750 * the current locks, and try again.
3754 case RINGBUF_TYPE_TIME_EXTEND
:
3755 /* Internal data, OK to advance */
3756 rb_advance_reader(cpu_buffer
);
3759 case RINGBUF_TYPE_TIME_STAMP
:
3760 /* FIXME: not implemented */
3761 rb_advance_reader(cpu_buffer
);
3764 case RINGBUF_TYPE_DATA
:
3766 *ts
= cpu_buffer
->read_stamp
+ event
->time_delta
;
3767 ring_buffer_normalize_time_stamp(cpu_buffer
->buffer
,
3768 cpu_buffer
->cpu
, ts
);
3771 *lost_events
= rb_lost_events(cpu_buffer
);
3780 EXPORT_SYMBOL_GPL(ring_buffer_peek
);
3782 static struct ring_buffer_event
*
3783 rb_iter_peek(struct ring_buffer_iter
*iter
, u64
*ts
)
3785 struct ring_buffer
*buffer
;
3786 struct ring_buffer_per_cpu
*cpu_buffer
;
3787 struct ring_buffer_event
*event
;
3790 cpu_buffer
= iter
->cpu_buffer
;
3791 buffer
= cpu_buffer
->buffer
;
3794 * Check if someone performed a consuming read to
3795 * the buffer. A consuming read invalidates the iterator
3796 * and we need to reset the iterator in this case.
3798 if (unlikely(iter
->cache_read
!= cpu_buffer
->read
||
3799 iter
->cache_reader_page
!= cpu_buffer
->reader_page
))
3800 rb_iter_reset(iter
);
3803 if (ring_buffer_iter_empty(iter
))
3807 * We repeat when a time extend is encountered or we hit
3808 * the end of the page. Since the time extend is always attached
3809 * to a data event, we should never loop more than three times.
3810 * Once for going to next page, once on time extend, and
3811 * finally once to get the event.
3812 * (We never hit the following condition more than thrice).
3814 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 3))
3817 if (rb_per_cpu_empty(cpu_buffer
))
3820 if (iter
->head
>= rb_page_size(iter
->head_page
)) {
3825 event
= rb_iter_head_event(iter
);
3827 switch (event
->type_len
) {
3828 case RINGBUF_TYPE_PADDING
:
3829 if (rb_null_event(event
)) {
3833 rb_advance_iter(iter
);
3836 case RINGBUF_TYPE_TIME_EXTEND
:
3837 /* Internal data, OK to advance */
3838 rb_advance_iter(iter
);
3841 case RINGBUF_TYPE_TIME_STAMP
:
3842 /* FIXME: not implemented */
3843 rb_advance_iter(iter
);
3846 case RINGBUF_TYPE_DATA
:
3848 *ts
= iter
->read_stamp
+ event
->time_delta
;
3849 ring_buffer_normalize_time_stamp(buffer
,
3850 cpu_buffer
->cpu
, ts
);
3860 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek
);
3862 static inline bool rb_reader_lock(struct ring_buffer_per_cpu
*cpu_buffer
)
3864 if (likely(!in_nmi())) {
3865 raw_spin_lock(&cpu_buffer
->reader_lock
);
3870 * If an NMI die dumps out the content of the ring buffer
3871 * trylock must be used to prevent a deadlock if the NMI
3872 * preempted a task that holds the ring buffer locks. If
3873 * we get the lock then all is fine, if not, then continue
3874 * to do the read, but this can corrupt the ring buffer,
3875 * so it must be permanently disabled from future writes.
3876 * Reading from NMI is a oneshot deal.
3878 if (raw_spin_trylock(&cpu_buffer
->reader_lock
))
3881 /* Continue without locking, but disable the ring buffer */
3882 atomic_inc(&cpu_buffer
->record_disabled
);
3887 rb_reader_unlock(struct ring_buffer_per_cpu
*cpu_buffer
, bool locked
)
3890 raw_spin_unlock(&cpu_buffer
->reader_lock
);
3895 * ring_buffer_peek - peek at the next event to be read
3896 * @buffer: The ring buffer to read
3897 * @cpu: The cpu to peak at
3898 * @ts: The timestamp counter of this event.
3899 * @lost_events: a variable to store if events were lost (may be NULL)
3901 * This will return the event that will be read next, but does
3902 * not consume the data.
3904 struct ring_buffer_event
*
3905 ring_buffer_peek(struct ring_buffer
*buffer
, int cpu
, u64
*ts
,
3906 unsigned long *lost_events
)
3908 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
3909 struct ring_buffer_event
*event
;
3910 unsigned long flags
;
3913 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3917 local_irq_save(flags
);
3918 dolock
= rb_reader_lock(cpu_buffer
);
3919 event
= rb_buffer_peek(cpu_buffer
, ts
, lost_events
);
3920 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3921 rb_advance_reader(cpu_buffer
);
3922 rb_reader_unlock(cpu_buffer
, dolock
);
3923 local_irq_restore(flags
);
3925 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3932 * ring_buffer_iter_peek - peek at the next event to be read
3933 * @iter: The ring buffer iterator
3934 * @ts: The timestamp counter of this event.
3936 * This will return the event that will be read next, but does
3937 * not increment the iterator.
3939 struct ring_buffer_event
*
3940 ring_buffer_iter_peek(struct ring_buffer_iter
*iter
, u64
*ts
)
3942 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
3943 struct ring_buffer_event
*event
;
3944 unsigned long flags
;
3947 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3948 event
= rb_iter_peek(iter
, ts
);
3949 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3951 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3958 * ring_buffer_consume - return an event and consume it
3959 * @buffer: The ring buffer to get the next event from
3960 * @cpu: the cpu to read the buffer from
3961 * @ts: a variable to store the timestamp (may be NULL)
3962 * @lost_events: a variable to store if events were lost (may be NULL)
3964 * Returns the next event in the ring buffer, and that event is consumed.
3965 * Meaning, that sequential reads will keep returning a different event,
3966 * and eventually empty the ring buffer if the producer is slower.
3968 struct ring_buffer_event
*
3969 ring_buffer_consume(struct ring_buffer
*buffer
, int cpu
, u64
*ts
,
3970 unsigned long *lost_events
)
3972 struct ring_buffer_per_cpu
*cpu_buffer
;
3973 struct ring_buffer_event
*event
= NULL
;
3974 unsigned long flags
;
3978 /* might be called in atomic */
3981 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3984 cpu_buffer
= buffer
->buffers
[cpu
];
3985 local_irq_save(flags
);
3986 dolock
= rb_reader_lock(cpu_buffer
);
3988 event
= rb_buffer_peek(cpu_buffer
, ts
, lost_events
);
3990 cpu_buffer
->lost_events
= 0;
3991 rb_advance_reader(cpu_buffer
);
3994 rb_reader_unlock(cpu_buffer
, dolock
);
3995 local_irq_restore(flags
);
4000 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
4005 EXPORT_SYMBOL_GPL(ring_buffer_consume
);
4008 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4009 * @buffer: The ring buffer to read from
4010 * @cpu: The cpu buffer to iterate over
4012 * This performs the initial preparations necessary to iterate
4013 * through the buffer. Memory is allocated, buffer recording
4014 * is disabled, and the iterator pointer is returned to the caller.
4016 * Disabling buffer recordng prevents the reading from being
4017 * corrupted. This is not a consuming read, so a producer is not
4020 * After a sequence of ring_buffer_read_prepare calls, the user is
4021 * expected to make at least one call to ring_buffer_read_prepare_sync.
4022 * Afterwards, ring_buffer_read_start is invoked to get things going
4025 * This overall must be paired with ring_buffer_read_finish.
4027 struct ring_buffer_iter
*
4028 ring_buffer_read_prepare(struct ring_buffer
*buffer
, int cpu
)
4030 struct ring_buffer_per_cpu
*cpu_buffer
;
4031 struct ring_buffer_iter
*iter
;
4033 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4036 iter
= kmalloc(sizeof(*iter
), GFP_KERNEL
);
4040 cpu_buffer
= buffer
->buffers
[cpu
];
4042 iter
->cpu_buffer
= cpu_buffer
;
4044 atomic_inc(&buffer
->resize_disabled
);
4045 atomic_inc(&cpu_buffer
->record_disabled
);
4049 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare
);
4052 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4054 * All previously invoked ring_buffer_read_prepare calls to prepare
4055 * iterators will be synchronized. Afterwards, read_buffer_read_start
4056 * calls on those iterators are allowed.
4059 ring_buffer_read_prepare_sync(void)
4061 synchronize_sched();
4063 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync
);
4066 * ring_buffer_read_start - start a non consuming read of the buffer
4067 * @iter: The iterator returned by ring_buffer_read_prepare
4069 * This finalizes the startup of an iteration through the buffer.
4070 * The iterator comes from a call to ring_buffer_read_prepare and
4071 * an intervening ring_buffer_read_prepare_sync must have been
4074 * Must be paired with ring_buffer_read_finish.
4077 ring_buffer_read_start(struct ring_buffer_iter
*iter
)
4079 struct ring_buffer_per_cpu
*cpu_buffer
;
4080 unsigned long flags
;
4085 cpu_buffer
= iter
->cpu_buffer
;
4087 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4088 arch_spin_lock(&cpu_buffer
->lock
);
4089 rb_iter_reset(iter
);
4090 arch_spin_unlock(&cpu_buffer
->lock
);
4091 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4093 EXPORT_SYMBOL_GPL(ring_buffer_read_start
);
4096 * ring_buffer_read_finish - finish reading the iterator of the buffer
4097 * @iter: The iterator retrieved by ring_buffer_start
4099 * This re-enables the recording to the buffer, and frees the
4103 ring_buffer_read_finish(struct ring_buffer_iter
*iter
)
4105 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
4106 unsigned long flags
;
4109 * Ring buffer is disabled from recording, here's a good place
4110 * to check the integrity of the ring buffer.
4111 * Must prevent readers from trying to read, as the check
4112 * clears the HEAD page and readers require it.
4114 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4115 rb_check_pages(cpu_buffer
);
4116 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4118 atomic_dec(&cpu_buffer
->record_disabled
);
4119 atomic_dec(&cpu_buffer
->buffer
->resize_disabled
);
4122 EXPORT_SYMBOL_GPL(ring_buffer_read_finish
);
4125 * ring_buffer_read - read the next item in the ring buffer by the iterator
4126 * @iter: The ring buffer iterator
4127 * @ts: The time stamp of the event read.
4129 * This reads the next event in the ring buffer and increments the iterator.
4131 struct ring_buffer_event
*
4132 ring_buffer_read(struct ring_buffer_iter
*iter
, u64
*ts
)
4134 struct ring_buffer_event
*event
;
4135 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
4136 unsigned long flags
;
4138 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4140 event
= rb_iter_peek(iter
, ts
);
4144 if (event
->type_len
== RINGBUF_TYPE_PADDING
)
4147 rb_advance_iter(iter
);
4149 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4153 EXPORT_SYMBOL_GPL(ring_buffer_read
);
4156 * ring_buffer_size - return the size of the ring buffer (in bytes)
4157 * @buffer: The ring buffer.
4159 unsigned long ring_buffer_size(struct ring_buffer
*buffer
, int cpu
)
4162 * Earlier, this method returned
4163 * BUF_PAGE_SIZE * buffer->nr_pages
4164 * Since the nr_pages field is now removed, we have converted this to
4165 * return the per cpu buffer value.
4167 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4170 return BUF_PAGE_SIZE
* buffer
->buffers
[cpu
]->nr_pages
;
4172 EXPORT_SYMBOL_GPL(ring_buffer_size
);
4175 rb_reset_cpu(struct ring_buffer_per_cpu
*cpu_buffer
)
4177 rb_head_page_deactivate(cpu_buffer
);
4179 cpu_buffer
->head_page
4180 = list_entry(cpu_buffer
->pages
, struct buffer_page
, list
);
4181 local_set(&cpu_buffer
->head_page
->write
, 0);
4182 local_set(&cpu_buffer
->head_page
->entries
, 0);
4183 local_set(&cpu_buffer
->head_page
->page
->commit
, 0);
4185 cpu_buffer
->head_page
->read
= 0;
4187 cpu_buffer
->tail_page
= cpu_buffer
->head_page
;
4188 cpu_buffer
->commit_page
= cpu_buffer
->head_page
;
4190 INIT_LIST_HEAD(&cpu_buffer
->reader_page
->list
);
4191 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
4192 local_set(&cpu_buffer
->reader_page
->write
, 0);
4193 local_set(&cpu_buffer
->reader_page
->entries
, 0);
4194 local_set(&cpu_buffer
->reader_page
->page
->commit
, 0);
4195 cpu_buffer
->reader_page
->read
= 0;
4197 local_set(&cpu_buffer
->entries_bytes
, 0);
4198 local_set(&cpu_buffer
->overrun
, 0);
4199 local_set(&cpu_buffer
->commit_overrun
, 0);
4200 local_set(&cpu_buffer
->dropped_events
, 0);
4201 local_set(&cpu_buffer
->entries
, 0);
4202 local_set(&cpu_buffer
->committing
, 0);
4203 local_set(&cpu_buffer
->commits
, 0);
4204 cpu_buffer
->read
= 0;
4205 cpu_buffer
->read_bytes
= 0;
4207 cpu_buffer
->write_stamp
= 0;
4208 cpu_buffer
->read_stamp
= 0;
4210 cpu_buffer
->lost_events
= 0;
4211 cpu_buffer
->last_overrun
= 0;
4213 rb_head_page_activate(cpu_buffer
);
4217 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4218 * @buffer: The ring buffer to reset a per cpu buffer of
4219 * @cpu: The CPU buffer to be reset
4221 void ring_buffer_reset_cpu(struct ring_buffer
*buffer
, int cpu
)
4223 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
4224 unsigned long flags
;
4226 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4229 atomic_inc(&buffer
->resize_disabled
);
4230 atomic_inc(&cpu_buffer
->record_disabled
);
4232 /* Make sure all commits have finished */
4233 synchronize_sched();
4235 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4237 if (RB_WARN_ON(cpu_buffer
, local_read(&cpu_buffer
->committing
)))
4240 arch_spin_lock(&cpu_buffer
->lock
);
4242 rb_reset_cpu(cpu_buffer
);
4244 arch_spin_unlock(&cpu_buffer
->lock
);
4247 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4249 atomic_dec(&cpu_buffer
->record_disabled
);
4250 atomic_dec(&buffer
->resize_disabled
);
4252 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu
);
4255 * ring_buffer_reset - reset a ring buffer
4256 * @buffer: The ring buffer to reset all cpu buffers
4258 void ring_buffer_reset(struct ring_buffer
*buffer
)
4262 for_each_buffer_cpu(buffer
, cpu
)
4263 ring_buffer_reset_cpu(buffer
, cpu
);
4265 EXPORT_SYMBOL_GPL(ring_buffer_reset
);
4268 * rind_buffer_empty - is the ring buffer empty?
4269 * @buffer: The ring buffer to test
4271 bool ring_buffer_empty(struct ring_buffer
*buffer
)
4273 struct ring_buffer_per_cpu
*cpu_buffer
;
4274 unsigned long flags
;
4279 /* yes this is racy, but if you don't like the race, lock the buffer */
4280 for_each_buffer_cpu(buffer
, cpu
) {
4281 cpu_buffer
= buffer
->buffers
[cpu
];
4282 local_irq_save(flags
);
4283 dolock
= rb_reader_lock(cpu_buffer
);
4284 ret
= rb_per_cpu_empty(cpu_buffer
);
4285 rb_reader_unlock(cpu_buffer
, dolock
);
4286 local_irq_restore(flags
);
4294 EXPORT_SYMBOL_GPL(ring_buffer_empty
);
4297 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4298 * @buffer: The ring buffer
4299 * @cpu: The CPU buffer to test
4301 bool ring_buffer_empty_cpu(struct ring_buffer
*buffer
, int cpu
)
4303 struct ring_buffer_per_cpu
*cpu_buffer
;
4304 unsigned long flags
;
4308 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4311 cpu_buffer
= buffer
->buffers
[cpu
];
4312 local_irq_save(flags
);
4313 dolock
= rb_reader_lock(cpu_buffer
);
4314 ret
= rb_per_cpu_empty(cpu_buffer
);
4315 rb_reader_unlock(cpu_buffer
, dolock
);
4316 local_irq_restore(flags
);
4320 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu
);
4322 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4324 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4325 * @buffer_a: One buffer to swap with
4326 * @buffer_b: The other buffer to swap with
4328 * This function is useful for tracers that want to take a "snapshot"
4329 * of a CPU buffer and has another back up buffer lying around.
4330 * it is expected that the tracer handles the cpu buffer not being
4331 * used at the moment.
4333 int ring_buffer_swap_cpu(struct ring_buffer
*buffer_a
,
4334 struct ring_buffer
*buffer_b
, int cpu
)
4336 struct ring_buffer_per_cpu
*cpu_buffer_a
;
4337 struct ring_buffer_per_cpu
*cpu_buffer_b
;
4340 if (!cpumask_test_cpu(cpu
, buffer_a
->cpumask
) ||
4341 !cpumask_test_cpu(cpu
, buffer_b
->cpumask
))
4344 cpu_buffer_a
= buffer_a
->buffers
[cpu
];
4345 cpu_buffer_b
= buffer_b
->buffers
[cpu
];
4347 /* At least make sure the two buffers are somewhat the same */
4348 if (cpu_buffer_a
->nr_pages
!= cpu_buffer_b
->nr_pages
)
4353 if (atomic_read(&buffer_a
->record_disabled
))
4356 if (atomic_read(&buffer_b
->record_disabled
))
4359 if (atomic_read(&cpu_buffer_a
->record_disabled
))
4362 if (atomic_read(&cpu_buffer_b
->record_disabled
))
4366 * We can't do a synchronize_sched here because this
4367 * function can be called in atomic context.
4368 * Normally this will be called from the same CPU as cpu.
4369 * If not it's up to the caller to protect this.
4371 atomic_inc(&cpu_buffer_a
->record_disabled
);
4372 atomic_inc(&cpu_buffer_b
->record_disabled
);
4375 if (local_read(&cpu_buffer_a
->committing
))
4377 if (local_read(&cpu_buffer_b
->committing
))
4380 buffer_a
->buffers
[cpu
] = cpu_buffer_b
;
4381 buffer_b
->buffers
[cpu
] = cpu_buffer_a
;
4383 cpu_buffer_b
->buffer
= buffer_a
;
4384 cpu_buffer_a
->buffer
= buffer_b
;
4389 atomic_dec(&cpu_buffer_a
->record_disabled
);
4390 atomic_dec(&cpu_buffer_b
->record_disabled
);
4394 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu
);
4395 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4398 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4399 * @buffer: the buffer to allocate for.
4400 * @cpu: the cpu buffer to allocate.
4402 * This function is used in conjunction with ring_buffer_read_page.
4403 * When reading a full page from the ring buffer, these functions
4404 * can be used to speed up the process. The calling function should
4405 * allocate a few pages first with this function. Then when it
4406 * needs to get pages from the ring buffer, it passes the result
4407 * of this function into ring_buffer_read_page, which will swap
4408 * the page that was allocated, with the read page of the buffer.
4411 * The page allocated, or ERR_PTR
4413 void *ring_buffer_alloc_read_page(struct ring_buffer
*buffer
, int cpu
)
4415 struct ring_buffer_per_cpu
*cpu_buffer
;
4416 struct buffer_data_page
*bpage
= NULL
;
4417 unsigned long flags
;
4420 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4421 return ERR_PTR(-ENODEV
);
4423 cpu_buffer
= buffer
->buffers
[cpu
];
4424 local_irq_save(flags
);
4425 arch_spin_lock(&cpu_buffer
->lock
);
4427 if (cpu_buffer
->free_page
) {
4428 bpage
= cpu_buffer
->free_page
;
4429 cpu_buffer
->free_page
= NULL
;
4432 arch_spin_unlock(&cpu_buffer
->lock
);
4433 local_irq_restore(flags
);
4438 page
= alloc_pages_node(cpu_to_node(cpu
),
4439 GFP_KERNEL
| __GFP_NORETRY
, 0);
4441 return ERR_PTR(-ENOMEM
);
4443 bpage
= page_address(page
);
4446 rb_init_page(bpage
);
4450 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page
);
4453 * ring_buffer_free_read_page - free an allocated read page
4454 * @buffer: the buffer the page was allocate for
4455 * @cpu: the cpu buffer the page came from
4456 * @data: the page to free
4458 * Free a page allocated from ring_buffer_alloc_read_page.
4460 void ring_buffer_free_read_page(struct ring_buffer
*buffer
, int cpu
, void *data
)
4462 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
4463 struct buffer_data_page
*bpage
= data
;
4464 struct page
*page
= virt_to_page(bpage
);
4465 unsigned long flags
;
4467 /* If the page is still in use someplace else, we can't reuse it */
4468 if (page_ref_count(page
) > 1)
4471 local_irq_save(flags
);
4472 arch_spin_lock(&cpu_buffer
->lock
);
4474 if (!cpu_buffer
->free_page
) {
4475 cpu_buffer
->free_page
= bpage
;
4479 arch_spin_unlock(&cpu_buffer
->lock
);
4480 local_irq_restore(flags
);
4483 free_page((unsigned long)bpage
);
4485 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page
);
4488 * ring_buffer_read_page - extract a page from the ring buffer
4489 * @buffer: buffer to extract from
4490 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4491 * @len: amount to extract
4492 * @cpu: the cpu of the buffer to extract
4493 * @full: should the extraction only happen when the page is full.
4495 * This function will pull out a page from the ring buffer and consume it.
4496 * @data_page must be the address of the variable that was returned
4497 * from ring_buffer_alloc_read_page. This is because the page might be used
4498 * to swap with a page in the ring buffer.
4501 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4502 * if (IS_ERR(rpage))
4503 * return PTR_ERR(rpage);
4504 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4506 * process_page(rpage, ret);
4508 * When @full is set, the function will not return true unless
4509 * the writer is off the reader page.
4511 * Note: it is up to the calling functions to handle sleeps and wakeups.
4512 * The ring buffer can be used anywhere in the kernel and can not
4513 * blindly call wake_up. The layer that uses the ring buffer must be
4514 * responsible for that.
4517 * >=0 if data has been transferred, returns the offset of consumed data.
4518 * <0 if no data has been transferred.
4520 int ring_buffer_read_page(struct ring_buffer
*buffer
,
4521 void **data_page
, size_t len
, int cpu
, int full
)
4523 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
4524 struct ring_buffer_event
*event
;
4525 struct buffer_data_page
*bpage
;
4526 struct buffer_page
*reader
;
4527 unsigned long missed_events
;
4528 unsigned long flags
;
4529 unsigned int commit
;
4534 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4538 * If len is not big enough to hold the page header, then
4539 * we can not copy anything.
4541 if (len
<= BUF_PAGE_HDR_SIZE
)
4544 len
-= BUF_PAGE_HDR_SIZE
;
4553 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4555 reader
= rb_get_reader_page(cpu_buffer
);
4559 event
= rb_reader_event(cpu_buffer
);
4561 read
= reader
->read
;
4562 commit
= rb_page_commit(reader
);
4564 /* Check if any events were dropped */
4565 missed_events
= cpu_buffer
->lost_events
;
4568 * If this page has been partially read or
4569 * if len is not big enough to read the rest of the page or
4570 * a writer is still on the page, then
4571 * we must copy the data from the page to the buffer.
4572 * Otherwise, we can simply swap the page with the one passed in.
4574 if (read
|| (len
< (commit
- read
)) ||
4575 cpu_buffer
->reader_page
== cpu_buffer
->commit_page
) {
4576 struct buffer_data_page
*rpage
= cpu_buffer
->reader_page
->page
;
4577 unsigned int rpos
= read
;
4578 unsigned int pos
= 0;
4584 if (len
> (commit
- read
))
4585 len
= (commit
- read
);
4587 /* Always keep the time extend and data together */
4588 size
= rb_event_ts_length(event
);
4593 /* save the current timestamp, since the user will need it */
4594 save_timestamp
= cpu_buffer
->read_stamp
;
4596 /* Need to copy one event at a time */
4598 /* We need the size of one event, because
4599 * rb_advance_reader only advances by one event,
4600 * whereas rb_event_ts_length may include the size of
4601 * one or two events.
4602 * We have already ensured there's enough space if this
4603 * is a time extend. */
4604 size
= rb_event_length(event
);
4605 memcpy(bpage
->data
+ pos
, rpage
->data
+ rpos
, size
);
4609 rb_advance_reader(cpu_buffer
);
4610 rpos
= reader
->read
;
4616 event
= rb_reader_event(cpu_buffer
);
4617 /* Always keep the time extend and data together */
4618 size
= rb_event_ts_length(event
);
4619 } while (len
>= size
);
4622 local_set(&bpage
->commit
, pos
);
4623 bpage
->time_stamp
= save_timestamp
;
4625 /* we copied everything to the beginning */
4628 /* update the entry counter */
4629 cpu_buffer
->read
+= rb_page_entries(reader
);
4630 cpu_buffer
->read_bytes
+= BUF_PAGE_SIZE
;
4632 /* swap the pages */
4633 rb_init_page(bpage
);
4634 bpage
= reader
->page
;
4635 reader
->page
= *data_page
;
4636 local_set(&reader
->write
, 0);
4637 local_set(&reader
->entries
, 0);
4642 * Use the real_end for the data size,
4643 * This gives us a chance to store the lost events
4646 if (reader
->real_end
)
4647 local_set(&bpage
->commit
, reader
->real_end
);
4651 cpu_buffer
->lost_events
= 0;
4653 commit
= local_read(&bpage
->commit
);
4655 * Set a flag in the commit field if we lost events
4657 if (missed_events
) {
4658 /* If there is room at the end of the page to save the
4659 * missed events, then record it there.
4661 if (BUF_PAGE_SIZE
- commit
>= sizeof(missed_events
)) {
4662 memcpy(&bpage
->data
[commit
], &missed_events
,
4663 sizeof(missed_events
));
4664 local_add(RB_MISSED_STORED
, &bpage
->commit
);
4665 commit
+= sizeof(missed_events
);
4667 local_add(RB_MISSED_EVENTS
, &bpage
->commit
);
4671 * This page may be off to user land. Zero it out here.
4673 if (commit
< BUF_PAGE_SIZE
)
4674 memset(&bpage
->data
[commit
], 0, BUF_PAGE_SIZE
- commit
);
4677 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4682 EXPORT_SYMBOL_GPL(ring_buffer_read_page
);
4685 * We only allocate new buffers, never free them if the CPU goes down.
4686 * If we were to free the buffer, then the user would lose any trace that was in
4689 int trace_rb_cpu_prepare(unsigned int cpu
, struct hlist_node
*node
)
4691 struct ring_buffer
*buffer
;
4694 unsigned long nr_pages
;
4696 buffer
= container_of(node
, struct ring_buffer
, node
);
4697 if (cpumask_test_cpu(cpu
, buffer
->cpumask
))
4702 /* check if all cpu sizes are same */
4703 for_each_buffer_cpu(buffer
, cpu_i
) {
4704 /* fill in the size from first enabled cpu */
4706 nr_pages
= buffer
->buffers
[cpu_i
]->nr_pages
;
4707 if (nr_pages
!= buffer
->buffers
[cpu_i
]->nr_pages
) {
4712 /* allocate minimum pages, user can later expand it */
4715 buffer
->buffers
[cpu
] =
4716 rb_allocate_cpu_buffer(buffer
, nr_pages
, cpu
);
4717 if (!buffer
->buffers
[cpu
]) {
4718 WARN(1, "failed to allocate ring buffer on CPU %u\n",
4723 cpumask_set_cpu(cpu
, buffer
->cpumask
);
4727 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4729 * This is a basic integrity check of the ring buffer.
4730 * Late in the boot cycle this test will run when configured in.
4731 * It will kick off a thread per CPU that will go into a loop
4732 * writing to the per cpu ring buffer various sizes of data.
4733 * Some of the data will be large items, some small.
4735 * Another thread is created that goes into a spin, sending out
4736 * IPIs to the other CPUs to also write into the ring buffer.
4737 * this is to test the nesting ability of the buffer.
4739 * Basic stats are recorded and reported. If something in the
4740 * ring buffer should happen that's not expected, a big warning
4741 * is displayed and all ring buffers are disabled.
4743 static struct task_struct
*rb_threads
[NR_CPUS
] __initdata
;
4745 struct rb_test_data
{
4746 struct ring_buffer
*buffer
;
4747 unsigned long events
;
4748 unsigned long bytes_written
;
4749 unsigned long bytes_alloc
;
4750 unsigned long bytes_dropped
;
4751 unsigned long events_nested
;
4752 unsigned long bytes_written_nested
;
4753 unsigned long bytes_alloc_nested
;
4754 unsigned long bytes_dropped_nested
;
4755 int min_size_nested
;
4756 int max_size_nested
;
4763 static struct rb_test_data rb_data
[NR_CPUS
] __initdata
;
4766 #define RB_TEST_BUFFER_SIZE 1048576
4768 static char rb_string
[] __initdata
=
4769 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4770 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4771 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4773 static bool rb_test_started __initdata
;
4780 static __init
int rb_write_something(struct rb_test_data
*data
, bool nested
)
4782 struct ring_buffer_event
*event
;
4783 struct rb_item
*item
;
4790 /* Have nested writes different that what is written */
4791 cnt
= data
->cnt
+ (nested
? 27 : 0);
4793 /* Multiply cnt by ~e, to make some unique increment */
4794 size
= (data
->cnt
* 68 / 25) % (sizeof(rb_string
) - 1);
4796 len
= size
+ sizeof(struct rb_item
);
4798 started
= rb_test_started
;
4799 /* read rb_test_started before checking buffer enabled */
4802 event
= ring_buffer_lock_reserve(data
->buffer
, len
);
4804 /* Ignore dropped events before test starts. */
4807 data
->bytes_dropped
+= len
;
4809 data
->bytes_dropped_nested
+= len
;
4814 event_len
= ring_buffer_event_length(event
);
4816 if (RB_WARN_ON(data
->buffer
, event_len
< len
))
4819 item
= ring_buffer_event_data(event
);
4821 memcpy(item
->str
, rb_string
, size
);
4824 data
->bytes_alloc_nested
+= event_len
;
4825 data
->bytes_written_nested
+= len
;
4826 data
->events_nested
++;
4827 if (!data
->min_size_nested
|| len
< data
->min_size_nested
)
4828 data
->min_size_nested
= len
;
4829 if (len
> data
->max_size_nested
)
4830 data
->max_size_nested
= len
;
4832 data
->bytes_alloc
+= event_len
;
4833 data
->bytes_written
+= len
;
4835 if (!data
->min_size
|| len
< data
->min_size
)
4836 data
->max_size
= len
;
4837 if (len
> data
->max_size
)
4838 data
->max_size
= len
;
4842 ring_buffer_unlock_commit(data
->buffer
, event
);
4847 static __init
int rb_test(void *arg
)
4849 struct rb_test_data
*data
= arg
;
4851 while (!kthread_should_stop()) {
4852 rb_write_something(data
, false);
4855 set_current_state(TASK_INTERRUPTIBLE
);
4856 /* Now sleep between a min of 100-300us and a max of 1ms */
4857 usleep_range(((data
->cnt
% 3) + 1) * 100, 1000);
4863 static __init
void rb_ipi(void *ignore
)
4865 struct rb_test_data
*data
;
4866 int cpu
= smp_processor_id();
4868 data
= &rb_data
[cpu
];
4869 rb_write_something(data
, true);
4872 static __init
int rb_hammer_test(void *arg
)
4874 while (!kthread_should_stop()) {
4876 /* Send an IPI to all cpus to write data! */
4877 smp_call_function(rb_ipi
, NULL
, 1);
4878 /* No sleep, but for non preempt, let others run */
4885 static __init
int test_ringbuffer(void)
4887 struct task_struct
*rb_hammer
;
4888 struct ring_buffer
*buffer
;
4892 pr_info("Running ring buffer tests...\n");
4894 buffer
= ring_buffer_alloc(RB_TEST_BUFFER_SIZE
, RB_FL_OVERWRITE
);
4895 if (WARN_ON(!buffer
))
4898 /* Disable buffer so that threads can't write to it yet */
4899 ring_buffer_record_off(buffer
);
4901 for_each_online_cpu(cpu
) {
4902 rb_data
[cpu
].buffer
= buffer
;
4903 rb_data
[cpu
].cpu
= cpu
;
4904 rb_data
[cpu
].cnt
= cpu
;
4905 rb_threads
[cpu
] = kthread_create(rb_test
, &rb_data
[cpu
],
4906 "rbtester/%d", cpu
);
4907 if (WARN_ON(IS_ERR(rb_threads
[cpu
]))) {
4908 pr_cont("FAILED\n");
4909 ret
= PTR_ERR(rb_threads
[cpu
]);
4913 kthread_bind(rb_threads
[cpu
], cpu
);
4914 wake_up_process(rb_threads
[cpu
]);
4917 /* Now create the rb hammer! */
4918 rb_hammer
= kthread_run(rb_hammer_test
, NULL
, "rbhammer");
4919 if (WARN_ON(IS_ERR(rb_hammer
))) {
4920 pr_cont("FAILED\n");
4921 ret
= PTR_ERR(rb_hammer
);
4925 ring_buffer_record_on(buffer
);
4927 * Show buffer is enabled before setting rb_test_started.
4928 * Yes there's a small race window where events could be
4929 * dropped and the thread wont catch it. But when a ring
4930 * buffer gets enabled, there will always be some kind of
4931 * delay before other CPUs see it. Thus, we don't care about
4932 * those dropped events. We care about events dropped after
4933 * the threads see that the buffer is active.
4936 rb_test_started
= true;
4938 set_current_state(TASK_INTERRUPTIBLE
);
4939 /* Just run for 10 seconds */;
4940 schedule_timeout(10 * HZ
);
4942 kthread_stop(rb_hammer
);
4945 for_each_online_cpu(cpu
) {
4946 if (!rb_threads
[cpu
])
4948 kthread_stop(rb_threads
[cpu
]);
4951 ring_buffer_free(buffer
);
4956 pr_info("finished\n");
4957 for_each_online_cpu(cpu
) {
4958 struct ring_buffer_event
*event
;
4959 struct rb_test_data
*data
= &rb_data
[cpu
];
4960 struct rb_item
*item
;
4961 unsigned long total_events
;
4962 unsigned long total_dropped
;
4963 unsigned long total_written
;
4964 unsigned long total_alloc
;
4965 unsigned long total_read
= 0;
4966 unsigned long total_size
= 0;
4967 unsigned long total_len
= 0;
4968 unsigned long total_lost
= 0;
4971 int small_event_size
;
4975 total_events
= data
->events
+ data
->events_nested
;
4976 total_written
= data
->bytes_written
+ data
->bytes_written_nested
;
4977 total_alloc
= data
->bytes_alloc
+ data
->bytes_alloc_nested
;
4978 total_dropped
= data
->bytes_dropped
+ data
->bytes_dropped_nested
;
4980 big_event_size
= data
->max_size
+ data
->max_size_nested
;
4981 small_event_size
= data
->min_size
+ data
->min_size_nested
;
4983 pr_info("CPU %d:\n", cpu
);
4984 pr_info(" events: %ld\n", total_events
);
4985 pr_info(" dropped bytes: %ld\n", total_dropped
);
4986 pr_info(" alloced bytes: %ld\n", total_alloc
);
4987 pr_info(" written bytes: %ld\n", total_written
);
4988 pr_info(" biggest event: %d\n", big_event_size
);
4989 pr_info(" smallest event: %d\n", small_event_size
);
4991 if (RB_WARN_ON(buffer
, total_dropped
))
4996 while ((event
= ring_buffer_consume(buffer
, cpu
, NULL
, &lost
))) {
4998 item
= ring_buffer_event_data(event
);
4999 total_len
+= ring_buffer_event_length(event
);
5000 total_size
+= item
->size
+ sizeof(struct rb_item
);
5001 if (memcmp(&item
->str
[0], rb_string
, item
->size
) != 0) {
5002 pr_info("FAILED!\n");
5003 pr_info("buffer had: %.*s\n", item
->size
, item
->str
);
5004 pr_info("expected: %.*s\n", item
->size
, rb_string
);
5005 RB_WARN_ON(buffer
, 1);
5016 pr_info(" read events: %ld\n", total_read
);
5017 pr_info(" lost events: %ld\n", total_lost
);
5018 pr_info(" total events: %ld\n", total_lost
+ total_read
);
5019 pr_info(" recorded len bytes: %ld\n", total_len
);
5020 pr_info(" recorded size bytes: %ld\n", total_size
);
5022 pr_info(" With dropped events, record len and size may not match\n"
5023 " alloced and written from above\n");
5025 if (RB_WARN_ON(buffer
, total_len
!= total_alloc
||
5026 total_size
!= total_written
))
5029 if (RB_WARN_ON(buffer
, total_lost
+ total_read
!= total_events
))
5035 pr_info("Ring buffer PASSED!\n");
5037 ring_buffer_free(buffer
);
5041 late_initcall(test_ringbuffer
);
5042 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */