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 for (i
= 0; i
< nr_pages
; i
++) {
1142 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1143 * gracefully without invoking oom-killer and the system is not
1146 bpage
= kzalloc_node(ALIGN(sizeof(*bpage
), cache_line_size()),
1147 GFP_KERNEL
| __GFP_RETRY_MAYFAIL
,
1152 list_add(&bpage
->list
, pages
);
1154 page
= alloc_pages_node(cpu_to_node(cpu
),
1155 GFP_KERNEL
| __GFP_RETRY_MAYFAIL
, 0);
1158 bpage
->page
= page_address(page
);
1159 rb_init_page(bpage
->page
);
1165 list_for_each_entry_safe(bpage
, tmp
, pages
, list
) {
1166 list_del_init(&bpage
->list
);
1167 free_buffer_page(bpage
);
1173 static int rb_allocate_pages(struct ring_buffer_per_cpu
*cpu_buffer
,
1174 unsigned long nr_pages
)
1180 if (__rb_allocate_pages(nr_pages
, &pages
, cpu_buffer
->cpu
))
1184 * The ring buffer page list is a circular list that does not
1185 * start and end with a list head. All page list items point to
1188 cpu_buffer
->pages
= pages
.next
;
1191 cpu_buffer
->nr_pages
= nr_pages
;
1193 rb_check_pages(cpu_buffer
);
1198 static struct ring_buffer_per_cpu
*
1199 rb_allocate_cpu_buffer(struct ring_buffer
*buffer
, long nr_pages
, int cpu
)
1201 struct ring_buffer_per_cpu
*cpu_buffer
;
1202 struct buffer_page
*bpage
;
1206 cpu_buffer
= kzalloc_node(ALIGN(sizeof(*cpu_buffer
), cache_line_size()),
1207 GFP_KERNEL
, cpu_to_node(cpu
));
1211 cpu_buffer
->cpu
= cpu
;
1212 cpu_buffer
->buffer
= buffer
;
1213 raw_spin_lock_init(&cpu_buffer
->reader_lock
);
1214 lockdep_set_class(&cpu_buffer
->reader_lock
, buffer
->reader_lock_key
);
1215 cpu_buffer
->lock
= (arch_spinlock_t
)__ARCH_SPIN_LOCK_UNLOCKED
;
1216 INIT_WORK(&cpu_buffer
->update_pages_work
, update_pages_handler
);
1217 init_completion(&cpu_buffer
->update_done
);
1218 init_irq_work(&cpu_buffer
->irq_work
.work
, rb_wake_up_waiters
);
1219 init_waitqueue_head(&cpu_buffer
->irq_work
.waiters
);
1220 init_waitqueue_head(&cpu_buffer
->irq_work
.full_waiters
);
1222 bpage
= kzalloc_node(ALIGN(sizeof(*bpage
), cache_line_size()),
1223 GFP_KERNEL
, cpu_to_node(cpu
));
1225 goto fail_free_buffer
;
1227 rb_check_bpage(cpu_buffer
, bpage
);
1229 cpu_buffer
->reader_page
= bpage
;
1230 page
= alloc_pages_node(cpu_to_node(cpu
), GFP_KERNEL
, 0);
1232 goto fail_free_reader
;
1233 bpage
->page
= page_address(page
);
1234 rb_init_page(bpage
->page
);
1236 INIT_LIST_HEAD(&cpu_buffer
->reader_page
->list
);
1237 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
1239 ret
= rb_allocate_pages(cpu_buffer
, nr_pages
);
1241 goto fail_free_reader
;
1243 cpu_buffer
->head_page
1244 = list_entry(cpu_buffer
->pages
, struct buffer_page
, list
);
1245 cpu_buffer
->tail_page
= cpu_buffer
->commit_page
= cpu_buffer
->head_page
;
1247 rb_head_page_activate(cpu_buffer
);
1252 free_buffer_page(cpu_buffer
->reader_page
);
1259 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu
*cpu_buffer
)
1261 struct list_head
*head
= cpu_buffer
->pages
;
1262 struct buffer_page
*bpage
, *tmp
;
1264 free_buffer_page(cpu_buffer
->reader_page
);
1266 rb_head_page_deactivate(cpu_buffer
);
1269 list_for_each_entry_safe(bpage
, tmp
, head
, list
) {
1270 list_del_init(&bpage
->list
);
1271 free_buffer_page(bpage
);
1273 bpage
= list_entry(head
, struct buffer_page
, list
);
1274 free_buffer_page(bpage
);
1281 * __ring_buffer_alloc - allocate a new ring_buffer
1282 * @size: the size in bytes per cpu that is needed.
1283 * @flags: attributes to set for the ring buffer.
1285 * Currently the only flag that is available is the RB_FL_OVERWRITE
1286 * flag. This flag means that the buffer will overwrite old data
1287 * when the buffer wraps. If this flag is not set, the buffer will
1288 * drop data when the tail hits the head.
1290 struct ring_buffer
*__ring_buffer_alloc(unsigned long size
, unsigned flags
,
1291 struct lock_class_key
*key
)
1293 struct ring_buffer
*buffer
;
1299 /* keep it in its own cache line */
1300 buffer
= kzalloc(ALIGN(sizeof(*buffer
), cache_line_size()),
1305 if (!zalloc_cpumask_var(&buffer
->cpumask
, GFP_KERNEL
))
1306 goto fail_free_buffer
;
1308 nr_pages
= DIV_ROUND_UP(size
, BUF_PAGE_SIZE
);
1309 buffer
->flags
= flags
;
1310 buffer
->clock
= trace_clock_local
;
1311 buffer
->reader_lock_key
= key
;
1313 init_irq_work(&buffer
->irq_work
.work
, rb_wake_up_waiters
);
1314 init_waitqueue_head(&buffer
->irq_work
.waiters
);
1316 /* need at least two pages */
1320 buffer
->cpus
= nr_cpu_ids
;
1322 bsize
= sizeof(void *) * nr_cpu_ids
;
1323 buffer
->buffers
= kzalloc(ALIGN(bsize
, cache_line_size()),
1325 if (!buffer
->buffers
)
1326 goto fail_free_cpumask
;
1328 cpu
= raw_smp_processor_id();
1329 cpumask_set_cpu(cpu
, buffer
->cpumask
);
1330 buffer
->buffers
[cpu
] = rb_allocate_cpu_buffer(buffer
, nr_pages
, cpu
);
1331 if (!buffer
->buffers
[cpu
])
1332 goto fail_free_buffers
;
1334 ret
= cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE
, &buffer
->node
);
1336 goto fail_free_buffers
;
1338 mutex_init(&buffer
->mutex
);
1343 for_each_buffer_cpu(buffer
, cpu
) {
1344 if (buffer
->buffers
[cpu
])
1345 rb_free_cpu_buffer(buffer
->buffers
[cpu
]);
1347 kfree(buffer
->buffers
);
1350 free_cpumask_var(buffer
->cpumask
);
1356 EXPORT_SYMBOL_GPL(__ring_buffer_alloc
);
1359 * ring_buffer_free - free a ring buffer.
1360 * @buffer: the buffer to free.
1363 ring_buffer_free(struct ring_buffer
*buffer
)
1367 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE
, &buffer
->node
);
1369 for_each_buffer_cpu(buffer
, cpu
)
1370 rb_free_cpu_buffer(buffer
->buffers
[cpu
]);
1372 kfree(buffer
->buffers
);
1373 free_cpumask_var(buffer
->cpumask
);
1377 EXPORT_SYMBOL_GPL(ring_buffer_free
);
1379 void ring_buffer_set_clock(struct ring_buffer
*buffer
,
1382 buffer
->clock
= clock
;
1385 static void rb_reset_cpu(struct ring_buffer_per_cpu
*cpu_buffer
);
1387 static inline unsigned long rb_page_entries(struct buffer_page
*bpage
)
1389 return local_read(&bpage
->entries
) & RB_WRITE_MASK
;
1392 static inline unsigned long rb_page_write(struct buffer_page
*bpage
)
1394 return local_read(&bpage
->write
) & RB_WRITE_MASK
;
1398 rb_remove_pages(struct ring_buffer_per_cpu
*cpu_buffer
, unsigned long nr_pages
)
1400 struct list_head
*tail_page
, *to_remove
, *next_page
;
1401 struct buffer_page
*to_remove_page
, *tmp_iter_page
;
1402 struct buffer_page
*last_page
, *first_page
;
1403 unsigned long nr_removed
;
1404 unsigned long head_bit
;
1409 raw_spin_lock_irq(&cpu_buffer
->reader_lock
);
1410 atomic_inc(&cpu_buffer
->record_disabled
);
1412 * We don't race with the readers since we have acquired the reader
1413 * lock. We also don't race with writers after disabling recording.
1414 * This makes it easy to figure out the first and the last page to be
1415 * removed from the list. We unlink all the pages in between including
1416 * the first and last pages. This is done in a busy loop so that we
1417 * lose the least number of traces.
1418 * The pages are freed after we restart recording and unlock readers.
1420 tail_page
= &cpu_buffer
->tail_page
->list
;
1423 * tail page might be on reader page, we remove the next page
1424 * from the ring buffer
1426 if (cpu_buffer
->tail_page
== cpu_buffer
->reader_page
)
1427 tail_page
= rb_list_head(tail_page
->next
);
1428 to_remove
= tail_page
;
1430 /* start of pages to remove */
1431 first_page
= list_entry(rb_list_head(to_remove
->next
),
1432 struct buffer_page
, list
);
1434 for (nr_removed
= 0; nr_removed
< nr_pages
; nr_removed
++) {
1435 to_remove
= rb_list_head(to_remove
)->next
;
1436 head_bit
|= (unsigned long)to_remove
& RB_PAGE_HEAD
;
1439 next_page
= rb_list_head(to_remove
)->next
;
1442 * Now we remove all pages between tail_page and next_page.
1443 * Make sure that we have head_bit value preserved for the
1446 tail_page
->next
= (struct list_head
*)((unsigned long)next_page
|
1448 next_page
= rb_list_head(next_page
);
1449 next_page
->prev
= tail_page
;
1451 /* make sure pages points to a valid page in the ring buffer */
1452 cpu_buffer
->pages
= next_page
;
1454 /* update head page */
1456 cpu_buffer
->head_page
= list_entry(next_page
,
1457 struct buffer_page
, list
);
1460 * change read pointer to make sure any read iterators reset
1463 cpu_buffer
->read
= 0;
1465 /* pages are removed, resume tracing and then free the pages */
1466 atomic_dec(&cpu_buffer
->record_disabled
);
1467 raw_spin_unlock_irq(&cpu_buffer
->reader_lock
);
1469 RB_WARN_ON(cpu_buffer
, list_empty(cpu_buffer
->pages
));
1471 /* last buffer page to remove */
1472 last_page
= list_entry(rb_list_head(to_remove
), struct buffer_page
,
1474 tmp_iter_page
= first_page
;
1477 to_remove_page
= tmp_iter_page
;
1478 rb_inc_page(cpu_buffer
, &tmp_iter_page
);
1480 /* update the counters */
1481 page_entries
= rb_page_entries(to_remove_page
);
1484 * If something was added to this page, it was full
1485 * since it is not the tail page. So we deduct the
1486 * bytes consumed in ring buffer from here.
1487 * Increment overrun to account for the lost events.
1489 local_add(page_entries
, &cpu_buffer
->overrun
);
1490 local_sub(BUF_PAGE_SIZE
, &cpu_buffer
->entries_bytes
);
1494 * We have already removed references to this list item, just
1495 * free up the buffer_page and its page
1497 free_buffer_page(to_remove_page
);
1500 } while (to_remove_page
!= last_page
);
1502 RB_WARN_ON(cpu_buffer
, nr_removed
);
1504 return nr_removed
== 0;
1508 rb_insert_pages(struct ring_buffer_per_cpu
*cpu_buffer
)
1510 struct list_head
*pages
= &cpu_buffer
->new_pages
;
1511 int retries
, success
;
1513 raw_spin_lock_irq(&cpu_buffer
->reader_lock
);
1515 * We are holding the reader lock, so the reader page won't be swapped
1516 * in the ring buffer. Now we are racing with the writer trying to
1517 * move head page and the tail page.
1518 * We are going to adapt the reader page update process where:
1519 * 1. We first splice the start and end of list of new pages between
1520 * the head page and its previous page.
1521 * 2. We cmpxchg the prev_page->next to point from head page to the
1522 * start of new pages list.
1523 * 3. Finally, we update the head->prev to the end of new list.
1525 * We will try this process 10 times, to make sure that we don't keep
1531 struct list_head
*head_page
, *prev_page
, *r
;
1532 struct list_head
*last_page
, *first_page
;
1533 struct list_head
*head_page_with_bit
;
1535 head_page
= &rb_set_head_page(cpu_buffer
)->list
;
1538 prev_page
= head_page
->prev
;
1540 first_page
= pages
->next
;
1541 last_page
= pages
->prev
;
1543 head_page_with_bit
= (struct list_head
*)
1544 ((unsigned long)head_page
| RB_PAGE_HEAD
);
1546 last_page
->next
= head_page_with_bit
;
1547 first_page
->prev
= prev_page
;
1549 r
= cmpxchg(&prev_page
->next
, head_page_with_bit
, first_page
);
1551 if (r
== head_page_with_bit
) {
1553 * yay, we replaced the page pointer to our new list,
1554 * now, we just have to update to head page's prev
1555 * pointer to point to end of list
1557 head_page
->prev
= last_page
;
1564 INIT_LIST_HEAD(pages
);
1566 * If we weren't successful in adding in new pages, warn and stop
1569 RB_WARN_ON(cpu_buffer
, !success
);
1570 raw_spin_unlock_irq(&cpu_buffer
->reader_lock
);
1572 /* free pages if they weren't inserted */
1574 struct buffer_page
*bpage
, *tmp
;
1575 list_for_each_entry_safe(bpage
, tmp
, &cpu_buffer
->new_pages
,
1577 list_del_init(&bpage
->list
);
1578 free_buffer_page(bpage
);
1584 static void rb_update_pages(struct ring_buffer_per_cpu
*cpu_buffer
)
1588 if (cpu_buffer
->nr_pages_to_update
> 0)
1589 success
= rb_insert_pages(cpu_buffer
);
1591 success
= rb_remove_pages(cpu_buffer
,
1592 -cpu_buffer
->nr_pages_to_update
);
1595 cpu_buffer
->nr_pages
+= cpu_buffer
->nr_pages_to_update
;
1598 static void update_pages_handler(struct work_struct
*work
)
1600 struct ring_buffer_per_cpu
*cpu_buffer
= container_of(work
,
1601 struct ring_buffer_per_cpu
, update_pages_work
);
1602 rb_update_pages(cpu_buffer
);
1603 complete(&cpu_buffer
->update_done
);
1607 * ring_buffer_resize - resize the ring buffer
1608 * @buffer: the buffer to resize.
1609 * @size: the new size.
1610 * @cpu_id: the cpu buffer to resize
1612 * Minimum size is 2 * BUF_PAGE_SIZE.
1614 * Returns 0 on success and < 0 on failure.
1616 int ring_buffer_resize(struct ring_buffer
*buffer
, unsigned long size
,
1619 struct ring_buffer_per_cpu
*cpu_buffer
;
1620 unsigned long nr_pages
;
1624 * Always succeed at resizing a non-existent buffer:
1629 /* Make sure the requested buffer exists */
1630 if (cpu_id
!= RING_BUFFER_ALL_CPUS
&&
1631 !cpumask_test_cpu(cpu_id
, buffer
->cpumask
))
1634 nr_pages
= DIV_ROUND_UP(size
, BUF_PAGE_SIZE
);
1636 /* we need a minimum of two pages */
1640 size
= nr_pages
* BUF_PAGE_SIZE
;
1643 * Don't succeed if resizing is disabled, as a reader might be
1644 * manipulating the ring buffer and is expecting a sane state while
1647 if (atomic_read(&buffer
->resize_disabled
))
1650 /* prevent another thread from changing buffer sizes */
1651 mutex_lock(&buffer
->mutex
);
1653 if (cpu_id
== RING_BUFFER_ALL_CPUS
) {
1654 /* calculate the pages to update */
1655 for_each_buffer_cpu(buffer
, cpu
) {
1656 cpu_buffer
= buffer
->buffers
[cpu
];
1658 cpu_buffer
->nr_pages_to_update
= nr_pages
-
1659 cpu_buffer
->nr_pages
;
1661 * nothing more to do for removing pages or no update
1663 if (cpu_buffer
->nr_pages_to_update
<= 0)
1666 * to add pages, make sure all new pages can be
1667 * allocated without receiving ENOMEM
1669 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
1670 if (__rb_allocate_pages(cpu_buffer
->nr_pages_to_update
,
1671 &cpu_buffer
->new_pages
, cpu
)) {
1672 /* not enough memory for new pages */
1680 * Fire off all the required work handlers
1681 * We can't schedule on offline CPUs, but it's not necessary
1682 * since we can change their buffer sizes without any race.
1684 for_each_buffer_cpu(buffer
, cpu
) {
1685 cpu_buffer
= buffer
->buffers
[cpu
];
1686 if (!cpu_buffer
->nr_pages_to_update
)
1689 /* Can't run something on an offline CPU. */
1690 if (!cpu_online(cpu
)) {
1691 rb_update_pages(cpu_buffer
);
1692 cpu_buffer
->nr_pages_to_update
= 0;
1694 schedule_work_on(cpu
,
1695 &cpu_buffer
->update_pages_work
);
1699 /* wait for all the updates to complete */
1700 for_each_buffer_cpu(buffer
, cpu
) {
1701 cpu_buffer
= buffer
->buffers
[cpu
];
1702 if (!cpu_buffer
->nr_pages_to_update
)
1705 if (cpu_online(cpu
))
1706 wait_for_completion(&cpu_buffer
->update_done
);
1707 cpu_buffer
->nr_pages_to_update
= 0;
1712 /* Make sure this CPU has been intitialized */
1713 if (!cpumask_test_cpu(cpu_id
, buffer
->cpumask
))
1716 cpu_buffer
= buffer
->buffers
[cpu_id
];
1718 if (nr_pages
== cpu_buffer
->nr_pages
)
1721 cpu_buffer
->nr_pages_to_update
= nr_pages
-
1722 cpu_buffer
->nr_pages
;
1724 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
1725 if (cpu_buffer
->nr_pages_to_update
> 0 &&
1726 __rb_allocate_pages(cpu_buffer
->nr_pages_to_update
,
1727 &cpu_buffer
->new_pages
, cpu_id
)) {
1734 /* Can't run something on an offline CPU. */
1735 if (!cpu_online(cpu_id
))
1736 rb_update_pages(cpu_buffer
);
1738 schedule_work_on(cpu_id
,
1739 &cpu_buffer
->update_pages_work
);
1740 wait_for_completion(&cpu_buffer
->update_done
);
1743 cpu_buffer
->nr_pages_to_update
= 0;
1749 * The ring buffer resize can happen with the ring buffer
1750 * enabled, so that the update disturbs the tracing as little
1751 * as possible. But if the buffer is disabled, we do not need
1752 * to worry about that, and we can take the time to verify
1753 * that the buffer is not corrupt.
1755 if (atomic_read(&buffer
->record_disabled
)) {
1756 atomic_inc(&buffer
->record_disabled
);
1758 * Even though the buffer was disabled, we must make sure
1759 * that it is truly disabled before calling rb_check_pages.
1760 * There could have been a race between checking
1761 * record_disable and incrementing it.
1763 synchronize_sched();
1764 for_each_buffer_cpu(buffer
, cpu
) {
1765 cpu_buffer
= buffer
->buffers
[cpu
];
1766 rb_check_pages(cpu_buffer
);
1768 atomic_dec(&buffer
->record_disabled
);
1771 mutex_unlock(&buffer
->mutex
);
1775 for_each_buffer_cpu(buffer
, cpu
) {
1776 struct buffer_page
*bpage
, *tmp
;
1778 cpu_buffer
= buffer
->buffers
[cpu
];
1779 cpu_buffer
->nr_pages_to_update
= 0;
1781 if (list_empty(&cpu_buffer
->new_pages
))
1784 list_for_each_entry_safe(bpage
, tmp
, &cpu_buffer
->new_pages
,
1786 list_del_init(&bpage
->list
);
1787 free_buffer_page(bpage
);
1790 mutex_unlock(&buffer
->mutex
);
1793 EXPORT_SYMBOL_GPL(ring_buffer_resize
);
1795 void ring_buffer_change_overwrite(struct ring_buffer
*buffer
, int val
)
1797 mutex_lock(&buffer
->mutex
);
1799 buffer
->flags
|= RB_FL_OVERWRITE
;
1801 buffer
->flags
&= ~RB_FL_OVERWRITE
;
1802 mutex_unlock(&buffer
->mutex
);
1804 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite
);
1806 static __always_inline
void *__rb_page_index(struct buffer_page
*bpage
, unsigned index
)
1808 return bpage
->page
->data
+ index
;
1811 static __always_inline
struct ring_buffer_event
*
1812 rb_reader_event(struct ring_buffer_per_cpu
*cpu_buffer
)
1814 return __rb_page_index(cpu_buffer
->reader_page
,
1815 cpu_buffer
->reader_page
->read
);
1818 static __always_inline
struct ring_buffer_event
*
1819 rb_iter_head_event(struct ring_buffer_iter
*iter
)
1821 return __rb_page_index(iter
->head_page
, iter
->head
);
1824 static __always_inline
unsigned rb_page_commit(struct buffer_page
*bpage
)
1826 return local_read(&bpage
->page
->commit
);
1829 /* Size is determined by what has been committed */
1830 static __always_inline
unsigned rb_page_size(struct buffer_page
*bpage
)
1832 return rb_page_commit(bpage
);
1835 static __always_inline
unsigned
1836 rb_commit_index(struct ring_buffer_per_cpu
*cpu_buffer
)
1838 return rb_page_commit(cpu_buffer
->commit_page
);
1841 static __always_inline
unsigned
1842 rb_event_index(struct ring_buffer_event
*event
)
1844 unsigned long addr
= (unsigned long)event
;
1846 return (addr
& ~PAGE_MASK
) - BUF_PAGE_HDR_SIZE
;
1849 static void rb_inc_iter(struct ring_buffer_iter
*iter
)
1851 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
1854 * The iterator could be on the reader page (it starts there).
1855 * But the head could have moved, since the reader was
1856 * found. Check for this case and assign the iterator
1857 * to the head page instead of next.
1859 if (iter
->head_page
== cpu_buffer
->reader_page
)
1860 iter
->head_page
= rb_set_head_page(cpu_buffer
);
1862 rb_inc_page(cpu_buffer
, &iter
->head_page
);
1864 iter
->read_stamp
= iter
->head_page
->page
->time_stamp
;
1869 * rb_handle_head_page - writer hit the head page
1871 * Returns: +1 to retry page
1876 rb_handle_head_page(struct ring_buffer_per_cpu
*cpu_buffer
,
1877 struct buffer_page
*tail_page
,
1878 struct buffer_page
*next_page
)
1880 struct buffer_page
*new_head
;
1885 entries
= rb_page_entries(next_page
);
1888 * The hard part is here. We need to move the head
1889 * forward, and protect against both readers on
1890 * other CPUs and writers coming in via interrupts.
1892 type
= rb_head_page_set_update(cpu_buffer
, next_page
, tail_page
,
1896 * type can be one of four:
1897 * NORMAL - an interrupt already moved it for us
1898 * HEAD - we are the first to get here.
1899 * UPDATE - we are the interrupt interrupting
1901 * MOVED - a reader on another CPU moved the next
1902 * pointer to its reader page. Give up
1909 * We changed the head to UPDATE, thus
1910 * it is our responsibility to update
1913 local_add(entries
, &cpu_buffer
->overrun
);
1914 local_sub(BUF_PAGE_SIZE
, &cpu_buffer
->entries_bytes
);
1917 * The entries will be zeroed out when we move the
1921 /* still more to do */
1924 case RB_PAGE_UPDATE
:
1926 * This is an interrupt that interrupt the
1927 * previous update. Still more to do.
1930 case RB_PAGE_NORMAL
:
1932 * An interrupt came in before the update
1933 * and processed this for us.
1934 * Nothing left to do.
1939 * The reader is on another CPU and just did
1940 * a swap with our next_page.
1945 RB_WARN_ON(cpu_buffer
, 1); /* WTF??? */
1950 * Now that we are here, the old head pointer is
1951 * set to UPDATE. This will keep the reader from
1952 * swapping the head page with the reader page.
1953 * The reader (on another CPU) will spin till
1956 * We just need to protect against interrupts
1957 * doing the job. We will set the next pointer
1958 * to HEAD. After that, we set the old pointer
1959 * to NORMAL, but only if it was HEAD before.
1960 * otherwise we are an interrupt, and only
1961 * want the outer most commit to reset it.
1963 new_head
= next_page
;
1964 rb_inc_page(cpu_buffer
, &new_head
);
1966 ret
= rb_head_page_set_head(cpu_buffer
, new_head
, next_page
,
1970 * Valid returns are:
1971 * HEAD - an interrupt came in and already set it.
1972 * NORMAL - One of two things:
1973 * 1) We really set it.
1974 * 2) A bunch of interrupts came in and moved
1975 * the page forward again.
1979 case RB_PAGE_NORMAL
:
1983 RB_WARN_ON(cpu_buffer
, 1);
1988 * It is possible that an interrupt came in,
1989 * set the head up, then more interrupts came in
1990 * and moved it again. When we get back here,
1991 * the page would have been set to NORMAL but we
1992 * just set it back to HEAD.
1994 * How do you detect this? Well, if that happened
1995 * the tail page would have moved.
1997 if (ret
== RB_PAGE_NORMAL
) {
1998 struct buffer_page
*buffer_tail_page
;
2000 buffer_tail_page
= READ_ONCE(cpu_buffer
->tail_page
);
2002 * If the tail had moved passed next, then we need
2003 * to reset the pointer.
2005 if (buffer_tail_page
!= tail_page
&&
2006 buffer_tail_page
!= next_page
)
2007 rb_head_page_set_normal(cpu_buffer
, new_head
,
2013 * If this was the outer most commit (the one that
2014 * changed the original pointer from HEAD to UPDATE),
2015 * then it is up to us to reset it to NORMAL.
2017 if (type
== RB_PAGE_HEAD
) {
2018 ret
= rb_head_page_set_normal(cpu_buffer
, next_page
,
2021 if (RB_WARN_ON(cpu_buffer
,
2022 ret
!= RB_PAGE_UPDATE
))
2030 rb_reset_tail(struct ring_buffer_per_cpu
*cpu_buffer
,
2031 unsigned long tail
, struct rb_event_info
*info
)
2033 struct buffer_page
*tail_page
= info
->tail_page
;
2034 struct ring_buffer_event
*event
;
2035 unsigned long length
= info
->length
;
2038 * Only the event that crossed the page boundary
2039 * must fill the old tail_page with padding.
2041 if (tail
>= BUF_PAGE_SIZE
) {
2043 * If the page was filled, then we still need
2044 * to update the real_end. Reset it to zero
2045 * and the reader will ignore it.
2047 if (tail
== BUF_PAGE_SIZE
)
2048 tail_page
->real_end
= 0;
2050 local_sub(length
, &tail_page
->write
);
2054 event
= __rb_page_index(tail_page
, tail
);
2056 /* account for padding bytes */
2057 local_add(BUF_PAGE_SIZE
- tail
, &cpu_buffer
->entries_bytes
);
2060 * Save the original length to the meta data.
2061 * This will be used by the reader to add lost event
2064 tail_page
->real_end
= tail
;
2067 * If this event is bigger than the minimum size, then
2068 * we need to be careful that we don't subtract the
2069 * write counter enough to allow another writer to slip
2071 * We put in a discarded commit instead, to make sure
2072 * that this space is not used again.
2074 * If we are less than the minimum size, we don't need to
2077 if (tail
> (BUF_PAGE_SIZE
- RB_EVNT_MIN_SIZE
)) {
2078 /* No room for any events */
2080 /* Mark the rest of the page with padding */
2081 rb_event_set_padding(event
);
2083 /* Set the write back to the previous setting */
2084 local_sub(length
, &tail_page
->write
);
2088 /* Put in a discarded event */
2089 event
->array
[0] = (BUF_PAGE_SIZE
- tail
) - RB_EVNT_HDR_SIZE
;
2090 event
->type_len
= RINGBUF_TYPE_PADDING
;
2091 /* time delta must be non zero */
2092 event
->time_delta
= 1;
2094 /* Set write to end of buffer */
2095 length
= (tail
+ length
) - BUF_PAGE_SIZE
;
2096 local_sub(length
, &tail_page
->write
);
2099 static inline void rb_end_commit(struct ring_buffer_per_cpu
*cpu_buffer
);
2102 * This is the slow path, force gcc not to inline it.
2104 static noinline
struct ring_buffer_event
*
2105 rb_move_tail(struct ring_buffer_per_cpu
*cpu_buffer
,
2106 unsigned long tail
, struct rb_event_info
*info
)
2108 struct buffer_page
*tail_page
= info
->tail_page
;
2109 struct buffer_page
*commit_page
= cpu_buffer
->commit_page
;
2110 struct ring_buffer
*buffer
= cpu_buffer
->buffer
;
2111 struct buffer_page
*next_page
;
2114 next_page
= tail_page
;
2116 rb_inc_page(cpu_buffer
, &next_page
);
2119 * If for some reason, we had an interrupt storm that made
2120 * it all the way around the buffer, bail, and warn
2123 if (unlikely(next_page
== commit_page
)) {
2124 local_inc(&cpu_buffer
->commit_overrun
);
2129 * This is where the fun begins!
2131 * We are fighting against races between a reader that
2132 * could be on another CPU trying to swap its reader
2133 * page with the buffer head.
2135 * We are also fighting against interrupts coming in and
2136 * moving the head or tail on us as well.
2138 * If the next page is the head page then we have filled
2139 * the buffer, unless the commit page is still on the
2142 if (rb_is_head_page(cpu_buffer
, next_page
, &tail_page
->list
)) {
2145 * If the commit is not on the reader page, then
2146 * move the header page.
2148 if (!rb_is_reader_page(cpu_buffer
->commit_page
)) {
2150 * If we are not in overwrite mode,
2151 * this is easy, just stop here.
2153 if (!(buffer
->flags
& RB_FL_OVERWRITE
)) {
2154 local_inc(&cpu_buffer
->dropped_events
);
2158 ret
= rb_handle_head_page(cpu_buffer
,
2167 * We need to be careful here too. The
2168 * commit page could still be on the reader
2169 * page. We could have a small buffer, and
2170 * have filled up the buffer with events
2171 * from interrupts and such, and wrapped.
2173 * Note, if the tail page is also the on the
2174 * reader_page, we let it move out.
2176 if (unlikely((cpu_buffer
->commit_page
!=
2177 cpu_buffer
->tail_page
) &&
2178 (cpu_buffer
->commit_page
==
2179 cpu_buffer
->reader_page
))) {
2180 local_inc(&cpu_buffer
->commit_overrun
);
2186 rb_tail_page_update(cpu_buffer
, tail_page
, next_page
);
2190 rb_reset_tail(cpu_buffer
, tail
, info
);
2192 /* Commit what we have for now. */
2193 rb_end_commit(cpu_buffer
);
2194 /* rb_end_commit() decs committing */
2195 local_inc(&cpu_buffer
->committing
);
2197 /* fail and let the caller try again */
2198 return ERR_PTR(-EAGAIN
);
2202 rb_reset_tail(cpu_buffer
, tail
, info
);
2207 /* Slow path, do not inline */
2208 static noinline
struct ring_buffer_event
*
2209 rb_add_time_stamp(struct ring_buffer_event
*event
, u64 delta
)
2211 event
->type_len
= RINGBUF_TYPE_TIME_EXTEND
;
2213 /* Not the first event on the page? */
2214 if (rb_event_index(event
)) {
2215 event
->time_delta
= delta
& TS_MASK
;
2216 event
->array
[0] = delta
>> TS_SHIFT
;
2218 /* nope, just zero it */
2219 event
->time_delta
= 0;
2220 event
->array
[0] = 0;
2223 return skip_time_extend(event
);
2226 static inline bool rb_event_is_commit(struct ring_buffer_per_cpu
*cpu_buffer
,
2227 struct ring_buffer_event
*event
);
2230 * rb_update_event - update event type and data
2231 * @event: the event to update
2232 * @type: the type of event
2233 * @length: the size of the event field in the ring buffer
2235 * Update the type and data fields of the event. The length
2236 * is the actual size that is written to the ring buffer,
2237 * and with this, we can determine what to place into the
2241 rb_update_event(struct ring_buffer_per_cpu
*cpu_buffer
,
2242 struct ring_buffer_event
*event
,
2243 struct rb_event_info
*info
)
2245 unsigned length
= info
->length
;
2246 u64 delta
= info
->delta
;
2248 /* Only a commit updates the timestamp */
2249 if (unlikely(!rb_event_is_commit(cpu_buffer
, event
)))
2253 * If we need to add a timestamp, then we
2254 * add it to the start of the resevered space.
2256 if (unlikely(info
->add_timestamp
)) {
2257 event
= rb_add_time_stamp(event
, delta
);
2258 length
-= RB_LEN_TIME_EXTEND
;
2262 event
->time_delta
= delta
;
2263 length
-= RB_EVNT_HDR_SIZE
;
2264 if (length
> RB_MAX_SMALL_DATA
|| RB_FORCE_8BYTE_ALIGNMENT
) {
2265 event
->type_len
= 0;
2266 event
->array
[0] = length
;
2268 event
->type_len
= DIV_ROUND_UP(length
, RB_ALIGNMENT
);
2271 static unsigned rb_calculate_event_length(unsigned length
)
2273 struct ring_buffer_event event
; /* Used only for sizeof array */
2275 /* zero length can cause confusions */
2279 if (length
> RB_MAX_SMALL_DATA
|| RB_FORCE_8BYTE_ALIGNMENT
)
2280 length
+= sizeof(event
.array
[0]);
2282 length
+= RB_EVNT_HDR_SIZE
;
2283 length
= ALIGN(length
, RB_ARCH_ALIGNMENT
);
2286 * In case the time delta is larger than the 27 bits for it
2287 * in the header, we need to add a timestamp. If another
2288 * event comes in when trying to discard this one to increase
2289 * the length, then the timestamp will be added in the allocated
2290 * space of this event. If length is bigger than the size needed
2291 * for the TIME_EXTEND, then padding has to be used. The events
2292 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2293 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2294 * As length is a multiple of 4, we only need to worry if it
2295 * is 12 (RB_LEN_TIME_EXTEND + 4).
2297 if (length
== RB_LEN_TIME_EXTEND
+ RB_ALIGNMENT
)
2298 length
+= RB_ALIGNMENT
;
2303 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2304 static inline bool sched_clock_stable(void)
2311 rb_try_to_discard(struct ring_buffer_per_cpu
*cpu_buffer
,
2312 struct ring_buffer_event
*event
)
2314 unsigned long new_index
, old_index
;
2315 struct buffer_page
*bpage
;
2316 unsigned long index
;
2319 new_index
= rb_event_index(event
);
2320 old_index
= new_index
+ rb_event_ts_length(event
);
2321 addr
= (unsigned long)event
;
2324 bpage
= READ_ONCE(cpu_buffer
->tail_page
);
2326 if (bpage
->page
== (void *)addr
&& rb_page_write(bpage
) == old_index
) {
2327 unsigned long write_mask
=
2328 local_read(&bpage
->write
) & ~RB_WRITE_MASK
;
2329 unsigned long event_length
= rb_event_length(event
);
2331 * This is on the tail page. It is possible that
2332 * a write could come in and move the tail page
2333 * and write to the next page. That is fine
2334 * because we just shorten what is on this page.
2336 old_index
+= write_mask
;
2337 new_index
+= write_mask
;
2338 index
= local_cmpxchg(&bpage
->write
, old_index
, new_index
);
2339 if (index
== old_index
) {
2340 /* update counters */
2341 local_sub(event_length
, &cpu_buffer
->entries_bytes
);
2346 /* could not discard */
2350 static void rb_start_commit(struct ring_buffer_per_cpu
*cpu_buffer
)
2352 local_inc(&cpu_buffer
->committing
);
2353 local_inc(&cpu_buffer
->commits
);
2356 static __always_inline
void
2357 rb_set_commit_to_write(struct ring_buffer_per_cpu
*cpu_buffer
)
2359 unsigned long max_count
;
2362 * We only race with interrupts and NMIs on this CPU.
2363 * If we own the commit event, then we can commit
2364 * all others that interrupted us, since the interruptions
2365 * are in stack format (they finish before they come
2366 * back to us). This allows us to do a simple loop to
2367 * assign the commit to the tail.
2370 max_count
= cpu_buffer
->nr_pages
* 100;
2372 while (cpu_buffer
->commit_page
!= READ_ONCE(cpu_buffer
->tail_page
)) {
2373 if (RB_WARN_ON(cpu_buffer
, !(--max_count
)))
2375 if (RB_WARN_ON(cpu_buffer
,
2376 rb_is_reader_page(cpu_buffer
->tail_page
)))
2378 local_set(&cpu_buffer
->commit_page
->page
->commit
,
2379 rb_page_write(cpu_buffer
->commit_page
));
2380 rb_inc_page(cpu_buffer
, &cpu_buffer
->commit_page
);
2381 /* Only update the write stamp if the page has an event */
2382 if (rb_page_write(cpu_buffer
->commit_page
))
2383 cpu_buffer
->write_stamp
=
2384 cpu_buffer
->commit_page
->page
->time_stamp
;
2385 /* add barrier to keep gcc from optimizing too much */
2388 while (rb_commit_index(cpu_buffer
) !=
2389 rb_page_write(cpu_buffer
->commit_page
)) {
2391 local_set(&cpu_buffer
->commit_page
->page
->commit
,
2392 rb_page_write(cpu_buffer
->commit_page
));
2393 RB_WARN_ON(cpu_buffer
,
2394 local_read(&cpu_buffer
->commit_page
->page
->commit
) &
2399 /* again, keep gcc from optimizing */
2403 * If an interrupt came in just after the first while loop
2404 * and pushed the tail page forward, we will be left with
2405 * a dangling commit that will never go forward.
2407 if (unlikely(cpu_buffer
->commit_page
!= READ_ONCE(cpu_buffer
->tail_page
)))
2411 static __always_inline
void rb_end_commit(struct ring_buffer_per_cpu
*cpu_buffer
)
2413 unsigned long commits
;
2415 if (RB_WARN_ON(cpu_buffer
,
2416 !local_read(&cpu_buffer
->committing
)))
2420 commits
= local_read(&cpu_buffer
->commits
);
2421 /* synchronize with interrupts */
2423 if (local_read(&cpu_buffer
->committing
) == 1)
2424 rb_set_commit_to_write(cpu_buffer
);
2426 local_dec(&cpu_buffer
->committing
);
2428 /* synchronize with interrupts */
2432 * Need to account for interrupts coming in between the
2433 * updating of the commit page and the clearing of the
2434 * committing counter.
2436 if (unlikely(local_read(&cpu_buffer
->commits
) != commits
) &&
2437 !local_read(&cpu_buffer
->committing
)) {
2438 local_inc(&cpu_buffer
->committing
);
2443 static inline void rb_event_discard(struct ring_buffer_event
*event
)
2445 if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
)
2446 event
= skip_time_extend(event
);
2448 /* array[0] holds the actual length for the discarded event */
2449 event
->array
[0] = rb_event_data_length(event
) - RB_EVNT_HDR_SIZE
;
2450 event
->type_len
= RINGBUF_TYPE_PADDING
;
2451 /* time delta must be non zero */
2452 if (!event
->time_delta
)
2453 event
->time_delta
= 1;
2456 static __always_inline
bool
2457 rb_event_is_commit(struct ring_buffer_per_cpu
*cpu_buffer
,
2458 struct ring_buffer_event
*event
)
2460 unsigned long addr
= (unsigned long)event
;
2461 unsigned long index
;
2463 index
= rb_event_index(event
);
2466 return cpu_buffer
->commit_page
->page
== (void *)addr
&&
2467 rb_commit_index(cpu_buffer
) == index
;
2470 static __always_inline
void
2471 rb_update_write_stamp(struct ring_buffer_per_cpu
*cpu_buffer
,
2472 struct ring_buffer_event
*event
)
2477 * The event first in the commit queue updates the
2480 if (rb_event_is_commit(cpu_buffer
, event
)) {
2482 * A commit event that is first on a page
2483 * updates the write timestamp with the page stamp
2485 if (!rb_event_index(event
))
2486 cpu_buffer
->write_stamp
=
2487 cpu_buffer
->commit_page
->page
->time_stamp
;
2488 else if (event
->type_len
== RINGBUF_TYPE_TIME_EXTEND
) {
2489 delta
= event
->array
[0];
2491 delta
+= event
->time_delta
;
2492 cpu_buffer
->write_stamp
+= delta
;
2494 cpu_buffer
->write_stamp
+= event
->time_delta
;
2498 static void rb_commit(struct ring_buffer_per_cpu
*cpu_buffer
,
2499 struct ring_buffer_event
*event
)
2501 local_inc(&cpu_buffer
->entries
);
2502 rb_update_write_stamp(cpu_buffer
, event
);
2503 rb_end_commit(cpu_buffer
);
2506 static __always_inline
void
2507 rb_wakeups(struct ring_buffer
*buffer
, struct ring_buffer_per_cpu
*cpu_buffer
)
2511 if (buffer
->irq_work
.waiters_pending
) {
2512 buffer
->irq_work
.waiters_pending
= false;
2513 /* irq_work_queue() supplies it's own memory barriers */
2514 irq_work_queue(&buffer
->irq_work
.work
);
2517 if (cpu_buffer
->irq_work
.waiters_pending
) {
2518 cpu_buffer
->irq_work
.waiters_pending
= false;
2519 /* irq_work_queue() supplies it's own memory barriers */
2520 irq_work_queue(&cpu_buffer
->irq_work
.work
);
2523 pagebusy
= cpu_buffer
->reader_page
== cpu_buffer
->commit_page
;
2525 if (!pagebusy
&& cpu_buffer
->irq_work
.full_waiters_pending
) {
2526 cpu_buffer
->irq_work
.wakeup_full
= true;
2527 cpu_buffer
->irq_work
.full_waiters_pending
= false;
2528 /* irq_work_queue() supplies it's own memory barriers */
2529 irq_work_queue(&cpu_buffer
->irq_work
.work
);
2534 * The lock and unlock are done within a preempt disable section.
2535 * The current_context per_cpu variable can only be modified
2536 * by the current task between lock and unlock. But it can
2537 * be modified more than once via an interrupt. To pass this
2538 * information from the lock to the unlock without having to
2539 * access the 'in_interrupt()' functions again (which do show
2540 * a bit of overhead in something as critical as function tracing,
2541 * we use a bitmask trick.
2543 * bit 0 = NMI context
2544 * bit 1 = IRQ context
2545 * bit 2 = SoftIRQ context
2546 * bit 3 = normal context.
2548 * This works because this is the order of contexts that can
2549 * preempt other contexts. A SoftIRQ never preempts an IRQ
2552 * When the context is determined, the corresponding bit is
2553 * checked and set (if it was set, then a recursion of that context
2556 * On unlock, we need to clear this bit. To do so, just subtract
2557 * 1 from the current_context and AND it to itself.
2561 * 101 & 100 = 100 (clearing bit zero)
2564 * 1010 & 1001 = 1000 (clearing bit 1)
2566 * The least significant bit can be cleared this way, and it
2567 * just so happens that it is the same bit corresponding to
2568 * the current context.
2571 static __always_inline
int
2572 trace_recursive_lock(struct ring_buffer_per_cpu
*cpu_buffer
)
2574 unsigned int val
= cpu_buffer
->current_context
;
2575 unsigned long pc
= preempt_count();
2578 if (!(pc
& (NMI_MASK
| HARDIRQ_MASK
| SOFTIRQ_OFFSET
)))
2579 bit
= RB_CTX_NORMAL
;
2581 bit
= pc
& NMI_MASK
? RB_CTX_NMI
:
2582 pc
& HARDIRQ_MASK
? RB_CTX_IRQ
:
2583 pc
& SOFTIRQ_OFFSET
? 2 : RB_CTX_SOFTIRQ
;
2585 if (unlikely(val
& (1 << bit
)))
2589 cpu_buffer
->current_context
= val
;
2594 static __always_inline
void
2595 trace_recursive_unlock(struct ring_buffer_per_cpu
*cpu_buffer
)
2597 cpu_buffer
->current_context
&= cpu_buffer
->current_context
- 1;
2601 * ring_buffer_unlock_commit - commit a reserved
2602 * @buffer: The buffer to commit to
2603 * @event: The event pointer to commit.
2605 * This commits the data to the ring buffer, and releases any locks held.
2607 * Must be paired with ring_buffer_lock_reserve.
2609 int ring_buffer_unlock_commit(struct ring_buffer
*buffer
,
2610 struct ring_buffer_event
*event
)
2612 struct ring_buffer_per_cpu
*cpu_buffer
;
2613 int cpu
= raw_smp_processor_id();
2615 cpu_buffer
= buffer
->buffers
[cpu
];
2617 rb_commit(cpu_buffer
, event
);
2619 rb_wakeups(buffer
, cpu_buffer
);
2621 trace_recursive_unlock(cpu_buffer
);
2623 preempt_enable_notrace();
2627 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit
);
2629 static noinline
void
2630 rb_handle_timestamp(struct ring_buffer_per_cpu
*cpu_buffer
,
2631 struct rb_event_info
*info
)
2633 WARN_ONCE(info
->delta
> (1ULL << 59),
2634 KERN_WARNING
"Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2635 (unsigned long long)info
->delta
,
2636 (unsigned long long)info
->ts
,
2637 (unsigned long long)cpu_buffer
->write_stamp
,
2638 sched_clock_stable() ? "" :
2639 "If you just came from a suspend/resume,\n"
2640 "please switch to the trace global clock:\n"
2641 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2642 info
->add_timestamp
= 1;
2645 static struct ring_buffer_event
*
2646 __rb_reserve_next(struct ring_buffer_per_cpu
*cpu_buffer
,
2647 struct rb_event_info
*info
)
2649 struct ring_buffer_event
*event
;
2650 struct buffer_page
*tail_page
;
2651 unsigned long tail
, write
;
2654 * If the time delta since the last event is too big to
2655 * hold in the time field of the event, then we append a
2656 * TIME EXTEND event ahead of the data event.
2658 if (unlikely(info
->add_timestamp
))
2659 info
->length
+= RB_LEN_TIME_EXTEND
;
2661 /* Don't let the compiler play games with cpu_buffer->tail_page */
2662 tail_page
= info
->tail_page
= READ_ONCE(cpu_buffer
->tail_page
);
2663 write
= local_add_return(info
->length
, &tail_page
->write
);
2665 /* set write to only the index of the write */
2666 write
&= RB_WRITE_MASK
;
2667 tail
= write
- info
->length
;
2670 * If this is the first commit on the page, then it has the same
2671 * timestamp as the page itself.
2676 /* See if we shot pass the end of this buffer page */
2677 if (unlikely(write
> BUF_PAGE_SIZE
))
2678 return rb_move_tail(cpu_buffer
, tail
, info
);
2680 /* We reserved something on the buffer */
2682 event
= __rb_page_index(tail_page
, tail
);
2683 rb_update_event(cpu_buffer
, event
, info
);
2685 local_inc(&tail_page
->entries
);
2688 * If this is the first commit on the page, then update
2692 tail_page
->page
->time_stamp
= info
->ts
;
2694 /* account for these added bytes */
2695 local_add(info
->length
, &cpu_buffer
->entries_bytes
);
2700 static __always_inline
struct ring_buffer_event
*
2701 rb_reserve_next_event(struct ring_buffer
*buffer
,
2702 struct ring_buffer_per_cpu
*cpu_buffer
,
2703 unsigned long length
)
2705 struct ring_buffer_event
*event
;
2706 struct rb_event_info info
;
2710 rb_start_commit(cpu_buffer
);
2712 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2714 * Due to the ability to swap a cpu buffer from a buffer
2715 * it is possible it was swapped before we committed.
2716 * (committing stops a swap). We check for it here and
2717 * if it happened, we have to fail the write.
2720 if (unlikely(READ_ONCE(cpu_buffer
->buffer
) != buffer
)) {
2721 local_dec(&cpu_buffer
->committing
);
2722 local_dec(&cpu_buffer
->commits
);
2727 info
.length
= rb_calculate_event_length(length
);
2729 info
.add_timestamp
= 0;
2733 * We allow for interrupts to reenter here and do a trace.
2734 * If one does, it will cause this original code to loop
2735 * back here. Even with heavy interrupts happening, this
2736 * should only happen a few times in a row. If this happens
2737 * 1000 times in a row, there must be either an interrupt
2738 * storm or we have something buggy.
2741 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 1000))
2744 info
.ts
= rb_time_stamp(cpu_buffer
->buffer
);
2745 diff
= info
.ts
- cpu_buffer
->write_stamp
;
2747 /* make sure this diff is calculated here */
2750 /* Did the write stamp get updated already? */
2751 if (likely(info
.ts
>= cpu_buffer
->write_stamp
)) {
2753 if (unlikely(test_time_stamp(info
.delta
)))
2754 rb_handle_timestamp(cpu_buffer
, &info
);
2757 event
= __rb_reserve_next(cpu_buffer
, &info
);
2759 if (unlikely(PTR_ERR(event
) == -EAGAIN
)) {
2760 if (info
.add_timestamp
)
2761 info
.length
-= RB_LEN_TIME_EXTEND
;
2771 rb_end_commit(cpu_buffer
);
2776 * ring_buffer_lock_reserve - reserve a part of the buffer
2777 * @buffer: the ring buffer to reserve from
2778 * @length: the length of the data to reserve (excluding event header)
2780 * Returns a reseverd event on the ring buffer to copy directly to.
2781 * The user of this interface will need to get the body to write into
2782 * and can use the ring_buffer_event_data() interface.
2784 * The length is the length of the data needed, not the event length
2785 * which also includes the event header.
2787 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2788 * If NULL is returned, then nothing has been allocated or locked.
2790 struct ring_buffer_event
*
2791 ring_buffer_lock_reserve(struct ring_buffer
*buffer
, unsigned long length
)
2793 struct ring_buffer_per_cpu
*cpu_buffer
;
2794 struct ring_buffer_event
*event
;
2797 /* If we are tracing schedule, we don't want to recurse */
2798 preempt_disable_notrace();
2800 if (unlikely(atomic_read(&buffer
->record_disabled
)))
2803 cpu
= raw_smp_processor_id();
2805 if (unlikely(!cpumask_test_cpu(cpu
, buffer
->cpumask
)))
2808 cpu_buffer
= buffer
->buffers
[cpu
];
2810 if (unlikely(atomic_read(&cpu_buffer
->record_disabled
)))
2813 if (unlikely(length
> BUF_MAX_DATA_SIZE
))
2816 if (unlikely(trace_recursive_lock(cpu_buffer
)))
2819 event
= rb_reserve_next_event(buffer
, cpu_buffer
, length
);
2826 trace_recursive_unlock(cpu_buffer
);
2828 preempt_enable_notrace();
2831 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve
);
2834 * Decrement the entries to the page that an event is on.
2835 * The event does not even need to exist, only the pointer
2836 * to the page it is on. This may only be called before the commit
2840 rb_decrement_entry(struct ring_buffer_per_cpu
*cpu_buffer
,
2841 struct ring_buffer_event
*event
)
2843 unsigned long addr
= (unsigned long)event
;
2844 struct buffer_page
*bpage
= cpu_buffer
->commit_page
;
2845 struct buffer_page
*start
;
2849 /* Do the likely case first */
2850 if (likely(bpage
->page
== (void *)addr
)) {
2851 local_dec(&bpage
->entries
);
2856 * Because the commit page may be on the reader page we
2857 * start with the next page and check the end loop there.
2859 rb_inc_page(cpu_buffer
, &bpage
);
2862 if (bpage
->page
== (void *)addr
) {
2863 local_dec(&bpage
->entries
);
2866 rb_inc_page(cpu_buffer
, &bpage
);
2867 } while (bpage
!= start
);
2869 /* commit not part of this buffer?? */
2870 RB_WARN_ON(cpu_buffer
, 1);
2874 * ring_buffer_commit_discard - discard an event that has not been committed
2875 * @buffer: the ring buffer
2876 * @event: non committed event to discard
2878 * Sometimes an event that is in the ring buffer needs to be ignored.
2879 * This function lets the user discard an event in the ring buffer
2880 * and then that event will not be read later.
2882 * This function only works if it is called before the the item has been
2883 * committed. It will try to free the event from the ring buffer
2884 * if another event has not been added behind it.
2886 * If another event has been added behind it, it will set the event
2887 * up as discarded, and perform the commit.
2889 * If this function is called, do not call ring_buffer_unlock_commit on
2892 void ring_buffer_discard_commit(struct ring_buffer
*buffer
,
2893 struct ring_buffer_event
*event
)
2895 struct ring_buffer_per_cpu
*cpu_buffer
;
2898 /* The event is discarded regardless */
2899 rb_event_discard(event
);
2901 cpu
= smp_processor_id();
2902 cpu_buffer
= buffer
->buffers
[cpu
];
2905 * This must only be called if the event has not been
2906 * committed yet. Thus we can assume that preemption
2907 * is still disabled.
2909 RB_WARN_ON(buffer
, !local_read(&cpu_buffer
->committing
));
2911 rb_decrement_entry(cpu_buffer
, event
);
2912 if (rb_try_to_discard(cpu_buffer
, event
))
2916 * The commit is still visible by the reader, so we
2917 * must still update the timestamp.
2919 rb_update_write_stamp(cpu_buffer
, event
);
2921 rb_end_commit(cpu_buffer
);
2923 trace_recursive_unlock(cpu_buffer
);
2925 preempt_enable_notrace();
2928 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit
);
2931 * ring_buffer_write - write data to the buffer without reserving
2932 * @buffer: The ring buffer to write to.
2933 * @length: The length of the data being written (excluding the event header)
2934 * @data: The data to write to the buffer.
2936 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2937 * one function. If you already have the data to write to the buffer, it
2938 * may be easier to simply call this function.
2940 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2941 * and not the length of the event which would hold the header.
2943 int ring_buffer_write(struct ring_buffer
*buffer
,
2944 unsigned long length
,
2947 struct ring_buffer_per_cpu
*cpu_buffer
;
2948 struct ring_buffer_event
*event
;
2953 preempt_disable_notrace();
2955 if (atomic_read(&buffer
->record_disabled
))
2958 cpu
= raw_smp_processor_id();
2960 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
2963 cpu_buffer
= buffer
->buffers
[cpu
];
2965 if (atomic_read(&cpu_buffer
->record_disabled
))
2968 if (length
> BUF_MAX_DATA_SIZE
)
2971 if (unlikely(trace_recursive_lock(cpu_buffer
)))
2974 event
= rb_reserve_next_event(buffer
, cpu_buffer
, length
);
2978 body
= rb_event_data(event
);
2980 memcpy(body
, data
, length
);
2982 rb_commit(cpu_buffer
, event
);
2984 rb_wakeups(buffer
, cpu_buffer
);
2989 trace_recursive_unlock(cpu_buffer
);
2992 preempt_enable_notrace();
2996 EXPORT_SYMBOL_GPL(ring_buffer_write
);
2998 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu
*cpu_buffer
)
3000 struct buffer_page
*reader
= cpu_buffer
->reader_page
;
3001 struct buffer_page
*head
= rb_set_head_page(cpu_buffer
);
3002 struct buffer_page
*commit
= cpu_buffer
->commit_page
;
3004 /* In case of error, head will be NULL */
3005 if (unlikely(!head
))
3008 return reader
->read
== rb_page_commit(reader
) &&
3009 (commit
== reader
||
3011 head
->read
== rb_page_commit(commit
)));
3015 * ring_buffer_record_disable - stop all writes into the buffer
3016 * @buffer: The ring buffer to stop writes to.
3018 * This prevents all writes to the buffer. Any attempt to write
3019 * to the buffer after this will fail and return NULL.
3021 * The caller should call synchronize_sched() after this.
3023 void ring_buffer_record_disable(struct ring_buffer
*buffer
)
3025 atomic_inc(&buffer
->record_disabled
);
3027 EXPORT_SYMBOL_GPL(ring_buffer_record_disable
);
3030 * ring_buffer_record_enable - enable writes to the buffer
3031 * @buffer: The ring buffer to enable writes
3033 * Note, multiple disables will need the same number of enables
3034 * to truly enable the writing (much like preempt_disable).
3036 void ring_buffer_record_enable(struct ring_buffer
*buffer
)
3038 atomic_dec(&buffer
->record_disabled
);
3040 EXPORT_SYMBOL_GPL(ring_buffer_record_enable
);
3043 * ring_buffer_record_off - stop all writes into the buffer
3044 * @buffer: The ring buffer to stop writes to.
3046 * This prevents all writes to the buffer. Any attempt to write
3047 * to the buffer after this will fail and return NULL.
3049 * This is different than ring_buffer_record_disable() as
3050 * it works like an on/off switch, where as the disable() version
3051 * must be paired with a enable().
3053 void ring_buffer_record_off(struct ring_buffer
*buffer
)
3056 unsigned int new_rd
;
3059 rd
= atomic_read(&buffer
->record_disabled
);
3060 new_rd
= rd
| RB_BUFFER_OFF
;
3061 } while (atomic_cmpxchg(&buffer
->record_disabled
, rd
, new_rd
) != rd
);
3063 EXPORT_SYMBOL_GPL(ring_buffer_record_off
);
3066 * ring_buffer_record_on - restart writes into the buffer
3067 * @buffer: The ring buffer to start writes to.
3069 * This enables all writes to the buffer that was disabled by
3070 * ring_buffer_record_off().
3072 * This is different than ring_buffer_record_enable() as
3073 * it works like an on/off switch, where as the enable() version
3074 * must be paired with a disable().
3076 void ring_buffer_record_on(struct ring_buffer
*buffer
)
3079 unsigned int new_rd
;
3082 rd
= atomic_read(&buffer
->record_disabled
);
3083 new_rd
= rd
& ~RB_BUFFER_OFF
;
3084 } while (atomic_cmpxchg(&buffer
->record_disabled
, rd
, new_rd
) != rd
);
3086 EXPORT_SYMBOL_GPL(ring_buffer_record_on
);
3089 * ring_buffer_record_is_on - return true if the ring buffer can write
3090 * @buffer: The ring buffer to see if write is enabled
3092 * Returns true if the ring buffer is in a state that it accepts writes.
3094 int ring_buffer_record_is_on(struct ring_buffer
*buffer
)
3096 return !atomic_read(&buffer
->record_disabled
);
3100 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3101 * @buffer: The ring buffer to stop writes to.
3102 * @cpu: The CPU buffer to stop
3104 * This prevents all writes to the buffer. Any attempt to write
3105 * to the buffer after this will fail and return NULL.
3107 * The caller should call synchronize_sched() after this.
3109 void ring_buffer_record_disable_cpu(struct ring_buffer
*buffer
, int cpu
)
3111 struct ring_buffer_per_cpu
*cpu_buffer
;
3113 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3116 cpu_buffer
= buffer
->buffers
[cpu
];
3117 atomic_inc(&cpu_buffer
->record_disabled
);
3119 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu
);
3122 * ring_buffer_record_enable_cpu - enable writes to the buffer
3123 * @buffer: The ring buffer to enable writes
3124 * @cpu: The CPU to enable.
3126 * Note, multiple disables will need the same number of enables
3127 * to truly enable the writing (much like preempt_disable).
3129 void ring_buffer_record_enable_cpu(struct ring_buffer
*buffer
, int cpu
)
3131 struct ring_buffer_per_cpu
*cpu_buffer
;
3133 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3136 cpu_buffer
= buffer
->buffers
[cpu
];
3137 atomic_dec(&cpu_buffer
->record_disabled
);
3139 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu
);
3142 * The total entries in the ring buffer is the running counter
3143 * of entries entered into the ring buffer, minus the sum of
3144 * the entries read from the ring buffer and the number of
3145 * entries that were overwritten.
3147 static inline unsigned long
3148 rb_num_of_entries(struct ring_buffer_per_cpu
*cpu_buffer
)
3150 return local_read(&cpu_buffer
->entries
) -
3151 (local_read(&cpu_buffer
->overrun
) + cpu_buffer
->read
);
3155 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3156 * @buffer: The ring buffer
3157 * @cpu: The per CPU buffer to read from.
3159 u64
ring_buffer_oldest_event_ts(struct ring_buffer
*buffer
, int cpu
)
3161 unsigned long flags
;
3162 struct ring_buffer_per_cpu
*cpu_buffer
;
3163 struct buffer_page
*bpage
;
3166 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3169 cpu_buffer
= buffer
->buffers
[cpu
];
3170 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3172 * if the tail is on reader_page, oldest time stamp is on the reader
3175 if (cpu_buffer
->tail_page
== cpu_buffer
->reader_page
)
3176 bpage
= cpu_buffer
->reader_page
;
3178 bpage
= rb_set_head_page(cpu_buffer
);
3180 ret
= bpage
->page
->time_stamp
;
3181 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3185 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts
);
3188 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3189 * @buffer: The ring buffer
3190 * @cpu: The per CPU buffer to read from.
3192 unsigned long ring_buffer_bytes_cpu(struct ring_buffer
*buffer
, int cpu
)
3194 struct ring_buffer_per_cpu
*cpu_buffer
;
3197 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3200 cpu_buffer
= buffer
->buffers
[cpu
];
3201 ret
= local_read(&cpu_buffer
->entries_bytes
) - cpu_buffer
->read_bytes
;
3205 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu
);
3208 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3209 * @buffer: The ring buffer
3210 * @cpu: The per CPU buffer to get the entries from.
3212 unsigned long ring_buffer_entries_cpu(struct ring_buffer
*buffer
, int cpu
)
3214 struct ring_buffer_per_cpu
*cpu_buffer
;
3216 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3219 cpu_buffer
= buffer
->buffers
[cpu
];
3221 return rb_num_of_entries(cpu_buffer
);
3223 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu
);
3226 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3227 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3228 * @buffer: The ring buffer
3229 * @cpu: The per CPU buffer to get the number of overruns from
3231 unsigned long ring_buffer_overrun_cpu(struct ring_buffer
*buffer
, int cpu
)
3233 struct ring_buffer_per_cpu
*cpu_buffer
;
3236 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3239 cpu_buffer
= buffer
->buffers
[cpu
];
3240 ret
= local_read(&cpu_buffer
->overrun
);
3244 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu
);
3247 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3248 * commits failing due to the buffer wrapping around while there are uncommitted
3249 * events, such as during an interrupt storm.
3250 * @buffer: The ring buffer
3251 * @cpu: The per CPU buffer to get the number of overruns from
3254 ring_buffer_commit_overrun_cpu(struct ring_buffer
*buffer
, int cpu
)
3256 struct ring_buffer_per_cpu
*cpu_buffer
;
3259 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3262 cpu_buffer
= buffer
->buffers
[cpu
];
3263 ret
= local_read(&cpu_buffer
->commit_overrun
);
3267 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu
);
3270 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3271 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3272 * @buffer: The ring buffer
3273 * @cpu: The per CPU buffer to get the number of overruns from
3276 ring_buffer_dropped_events_cpu(struct ring_buffer
*buffer
, int cpu
)
3278 struct ring_buffer_per_cpu
*cpu_buffer
;
3281 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3284 cpu_buffer
= buffer
->buffers
[cpu
];
3285 ret
= local_read(&cpu_buffer
->dropped_events
);
3289 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu
);
3292 * ring_buffer_read_events_cpu - get the number of events successfully read
3293 * @buffer: The ring buffer
3294 * @cpu: The per CPU buffer to get the number of events read
3297 ring_buffer_read_events_cpu(struct ring_buffer
*buffer
, int cpu
)
3299 struct ring_buffer_per_cpu
*cpu_buffer
;
3301 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3304 cpu_buffer
= buffer
->buffers
[cpu
];
3305 return cpu_buffer
->read
;
3307 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu
);
3310 * ring_buffer_entries - get the number of entries in a buffer
3311 * @buffer: The ring buffer
3313 * Returns the total number of entries in the ring buffer
3316 unsigned long ring_buffer_entries(struct ring_buffer
*buffer
)
3318 struct ring_buffer_per_cpu
*cpu_buffer
;
3319 unsigned long entries
= 0;
3322 /* if you care about this being correct, lock the buffer */
3323 for_each_buffer_cpu(buffer
, cpu
) {
3324 cpu_buffer
= buffer
->buffers
[cpu
];
3325 entries
+= rb_num_of_entries(cpu_buffer
);
3330 EXPORT_SYMBOL_GPL(ring_buffer_entries
);
3333 * ring_buffer_overruns - get the number of overruns in buffer
3334 * @buffer: The ring buffer
3336 * Returns the total number of overruns in the ring buffer
3339 unsigned long ring_buffer_overruns(struct ring_buffer
*buffer
)
3341 struct ring_buffer_per_cpu
*cpu_buffer
;
3342 unsigned long overruns
= 0;
3345 /* if you care about this being correct, lock the buffer */
3346 for_each_buffer_cpu(buffer
, cpu
) {
3347 cpu_buffer
= buffer
->buffers
[cpu
];
3348 overruns
+= local_read(&cpu_buffer
->overrun
);
3353 EXPORT_SYMBOL_GPL(ring_buffer_overruns
);
3355 static void rb_iter_reset(struct ring_buffer_iter
*iter
)
3357 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
3359 /* Iterator usage is expected to have record disabled */
3360 iter
->head_page
= cpu_buffer
->reader_page
;
3361 iter
->head
= cpu_buffer
->reader_page
->read
;
3363 iter
->cache_reader_page
= iter
->head_page
;
3364 iter
->cache_read
= cpu_buffer
->read
;
3367 iter
->read_stamp
= cpu_buffer
->read_stamp
;
3369 iter
->read_stamp
= iter
->head_page
->page
->time_stamp
;
3373 * ring_buffer_iter_reset - reset an iterator
3374 * @iter: The iterator to reset
3376 * Resets the iterator, so that it will start from the beginning
3379 void ring_buffer_iter_reset(struct ring_buffer_iter
*iter
)
3381 struct ring_buffer_per_cpu
*cpu_buffer
;
3382 unsigned long flags
;
3387 cpu_buffer
= iter
->cpu_buffer
;
3389 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3390 rb_iter_reset(iter
);
3391 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3393 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset
);
3396 * ring_buffer_iter_empty - check if an iterator has no more to read
3397 * @iter: The iterator to check
3399 int ring_buffer_iter_empty(struct ring_buffer_iter
*iter
)
3401 struct ring_buffer_per_cpu
*cpu_buffer
;
3402 struct buffer_page
*reader
;
3403 struct buffer_page
*head_page
;
3404 struct buffer_page
*commit_page
;
3407 cpu_buffer
= iter
->cpu_buffer
;
3409 /* Remember, trace recording is off when iterator is in use */
3410 reader
= cpu_buffer
->reader_page
;
3411 head_page
= cpu_buffer
->head_page
;
3412 commit_page
= cpu_buffer
->commit_page
;
3413 commit
= rb_page_commit(commit_page
);
3415 return ((iter
->head_page
== commit_page
&& iter
->head
== commit
) ||
3416 (iter
->head_page
== reader
&& commit_page
== head_page
&&
3417 head_page
->read
== commit
&&
3418 iter
->head
== rb_page_commit(cpu_buffer
->reader_page
)));
3420 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty
);
3423 rb_update_read_stamp(struct ring_buffer_per_cpu
*cpu_buffer
,
3424 struct ring_buffer_event
*event
)
3428 switch (event
->type_len
) {
3429 case RINGBUF_TYPE_PADDING
:
3432 case RINGBUF_TYPE_TIME_EXTEND
:
3433 delta
= event
->array
[0];
3435 delta
+= event
->time_delta
;
3436 cpu_buffer
->read_stamp
+= delta
;
3439 case RINGBUF_TYPE_TIME_STAMP
:
3440 /* FIXME: not implemented */
3443 case RINGBUF_TYPE_DATA
:
3444 cpu_buffer
->read_stamp
+= event
->time_delta
;
3454 rb_update_iter_read_stamp(struct ring_buffer_iter
*iter
,
3455 struct ring_buffer_event
*event
)
3459 switch (event
->type_len
) {
3460 case RINGBUF_TYPE_PADDING
:
3463 case RINGBUF_TYPE_TIME_EXTEND
:
3464 delta
= event
->array
[0];
3466 delta
+= event
->time_delta
;
3467 iter
->read_stamp
+= delta
;
3470 case RINGBUF_TYPE_TIME_STAMP
:
3471 /* FIXME: not implemented */
3474 case RINGBUF_TYPE_DATA
:
3475 iter
->read_stamp
+= event
->time_delta
;
3484 static struct buffer_page
*
3485 rb_get_reader_page(struct ring_buffer_per_cpu
*cpu_buffer
)
3487 struct buffer_page
*reader
= NULL
;
3488 unsigned long overwrite
;
3489 unsigned long flags
;
3493 local_irq_save(flags
);
3494 arch_spin_lock(&cpu_buffer
->lock
);
3498 * This should normally only loop twice. But because the
3499 * start of the reader inserts an empty page, it causes
3500 * a case where we will loop three times. There should be no
3501 * reason to loop four times (that I know of).
3503 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 3)) {
3508 reader
= cpu_buffer
->reader_page
;
3510 /* If there's more to read, return this page */
3511 if (cpu_buffer
->reader_page
->read
< rb_page_size(reader
))
3514 /* Never should we have an index greater than the size */
3515 if (RB_WARN_ON(cpu_buffer
,
3516 cpu_buffer
->reader_page
->read
> rb_page_size(reader
)))
3519 /* check if we caught up to the tail */
3521 if (cpu_buffer
->commit_page
== cpu_buffer
->reader_page
)
3524 /* Don't bother swapping if the ring buffer is empty */
3525 if (rb_num_of_entries(cpu_buffer
) == 0)
3529 * Reset the reader page to size zero.
3531 local_set(&cpu_buffer
->reader_page
->write
, 0);
3532 local_set(&cpu_buffer
->reader_page
->entries
, 0);
3533 local_set(&cpu_buffer
->reader_page
->page
->commit
, 0);
3534 cpu_buffer
->reader_page
->real_end
= 0;
3538 * Splice the empty reader page into the list around the head.
3540 reader
= rb_set_head_page(cpu_buffer
);
3543 cpu_buffer
->reader_page
->list
.next
= rb_list_head(reader
->list
.next
);
3544 cpu_buffer
->reader_page
->list
.prev
= reader
->list
.prev
;
3547 * cpu_buffer->pages just needs to point to the buffer, it
3548 * has no specific buffer page to point to. Lets move it out
3549 * of our way so we don't accidentally swap it.
3551 cpu_buffer
->pages
= reader
->list
.prev
;
3553 /* The reader page will be pointing to the new head */
3554 rb_set_list_to_head(cpu_buffer
, &cpu_buffer
->reader_page
->list
);
3557 * We want to make sure we read the overruns after we set up our
3558 * pointers to the next object. The writer side does a
3559 * cmpxchg to cross pages which acts as the mb on the writer
3560 * side. Note, the reader will constantly fail the swap
3561 * while the writer is updating the pointers, so this
3562 * guarantees that the overwrite recorded here is the one we
3563 * want to compare with the last_overrun.
3566 overwrite
= local_read(&(cpu_buffer
->overrun
));
3569 * Here's the tricky part.
3571 * We need to move the pointer past the header page.
3572 * But we can only do that if a writer is not currently
3573 * moving it. The page before the header page has the
3574 * flag bit '1' set if it is pointing to the page we want.
3575 * but if the writer is in the process of moving it
3576 * than it will be '2' or already moved '0'.
3579 ret
= rb_head_page_replace(reader
, cpu_buffer
->reader_page
);
3582 * If we did not convert it, then we must try again.
3588 * Yeah! We succeeded in replacing the page.
3590 * Now make the new head point back to the reader page.
3592 rb_list_head(reader
->list
.next
)->prev
= &cpu_buffer
->reader_page
->list
;
3593 rb_inc_page(cpu_buffer
, &cpu_buffer
->head_page
);
3595 /* Finally update the reader page to the new head */
3596 cpu_buffer
->reader_page
= reader
;
3597 cpu_buffer
->reader_page
->read
= 0;
3599 if (overwrite
!= cpu_buffer
->last_overrun
) {
3600 cpu_buffer
->lost_events
= overwrite
- cpu_buffer
->last_overrun
;
3601 cpu_buffer
->last_overrun
= overwrite
;
3607 /* Update the read_stamp on the first event */
3608 if (reader
&& reader
->read
== 0)
3609 cpu_buffer
->read_stamp
= reader
->page
->time_stamp
;
3611 arch_spin_unlock(&cpu_buffer
->lock
);
3612 local_irq_restore(flags
);
3617 static void rb_advance_reader(struct ring_buffer_per_cpu
*cpu_buffer
)
3619 struct ring_buffer_event
*event
;
3620 struct buffer_page
*reader
;
3623 reader
= rb_get_reader_page(cpu_buffer
);
3625 /* This function should not be called when buffer is empty */
3626 if (RB_WARN_ON(cpu_buffer
, !reader
))
3629 event
= rb_reader_event(cpu_buffer
);
3631 if (event
->type_len
<= RINGBUF_TYPE_DATA_TYPE_LEN_MAX
)
3634 rb_update_read_stamp(cpu_buffer
, event
);
3636 length
= rb_event_length(event
);
3637 cpu_buffer
->reader_page
->read
+= length
;
3640 static void rb_advance_iter(struct ring_buffer_iter
*iter
)
3642 struct ring_buffer_per_cpu
*cpu_buffer
;
3643 struct ring_buffer_event
*event
;
3646 cpu_buffer
= iter
->cpu_buffer
;
3649 * Check if we are at the end of the buffer.
3651 if (iter
->head
>= rb_page_size(iter
->head_page
)) {
3652 /* discarded commits can make the page empty */
3653 if (iter
->head_page
== cpu_buffer
->commit_page
)
3659 event
= rb_iter_head_event(iter
);
3661 length
= rb_event_length(event
);
3664 * This should not be called to advance the header if we are
3665 * at the tail of the buffer.
3667 if (RB_WARN_ON(cpu_buffer
,
3668 (iter
->head_page
== cpu_buffer
->commit_page
) &&
3669 (iter
->head
+ length
> rb_commit_index(cpu_buffer
))))
3672 rb_update_iter_read_stamp(iter
, event
);
3674 iter
->head
+= length
;
3676 /* check for end of page padding */
3677 if ((iter
->head
>= rb_page_size(iter
->head_page
)) &&
3678 (iter
->head_page
!= cpu_buffer
->commit_page
))
3682 static int rb_lost_events(struct ring_buffer_per_cpu
*cpu_buffer
)
3684 return cpu_buffer
->lost_events
;
3687 static struct ring_buffer_event
*
3688 rb_buffer_peek(struct ring_buffer_per_cpu
*cpu_buffer
, u64
*ts
,
3689 unsigned long *lost_events
)
3691 struct ring_buffer_event
*event
;
3692 struct buffer_page
*reader
;
3697 * We repeat when a time extend is encountered.
3698 * Since the time extend is always attached to a data event,
3699 * we should never loop more than once.
3700 * (We never hit the following condition more than twice).
3702 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 2))
3705 reader
= rb_get_reader_page(cpu_buffer
);
3709 event
= rb_reader_event(cpu_buffer
);
3711 switch (event
->type_len
) {
3712 case RINGBUF_TYPE_PADDING
:
3713 if (rb_null_event(event
))
3714 RB_WARN_ON(cpu_buffer
, 1);
3716 * Because the writer could be discarding every
3717 * event it creates (which would probably be bad)
3718 * if we were to go back to "again" then we may never
3719 * catch up, and will trigger the warn on, or lock
3720 * the box. Return the padding, and we will release
3721 * the current locks, and try again.
3725 case RINGBUF_TYPE_TIME_EXTEND
:
3726 /* Internal data, OK to advance */
3727 rb_advance_reader(cpu_buffer
);
3730 case RINGBUF_TYPE_TIME_STAMP
:
3731 /* FIXME: not implemented */
3732 rb_advance_reader(cpu_buffer
);
3735 case RINGBUF_TYPE_DATA
:
3737 *ts
= cpu_buffer
->read_stamp
+ event
->time_delta
;
3738 ring_buffer_normalize_time_stamp(cpu_buffer
->buffer
,
3739 cpu_buffer
->cpu
, ts
);
3742 *lost_events
= rb_lost_events(cpu_buffer
);
3751 EXPORT_SYMBOL_GPL(ring_buffer_peek
);
3753 static struct ring_buffer_event
*
3754 rb_iter_peek(struct ring_buffer_iter
*iter
, u64
*ts
)
3756 struct ring_buffer
*buffer
;
3757 struct ring_buffer_per_cpu
*cpu_buffer
;
3758 struct ring_buffer_event
*event
;
3761 cpu_buffer
= iter
->cpu_buffer
;
3762 buffer
= cpu_buffer
->buffer
;
3765 * Check if someone performed a consuming read to
3766 * the buffer. A consuming read invalidates the iterator
3767 * and we need to reset the iterator in this case.
3769 if (unlikely(iter
->cache_read
!= cpu_buffer
->read
||
3770 iter
->cache_reader_page
!= cpu_buffer
->reader_page
))
3771 rb_iter_reset(iter
);
3774 if (ring_buffer_iter_empty(iter
))
3778 * We repeat when a time extend is encountered or we hit
3779 * the end of the page. Since the time extend is always attached
3780 * to a data event, we should never loop more than three times.
3781 * Once for going to next page, once on time extend, and
3782 * finally once to get the event.
3783 * (We never hit the following condition more than thrice).
3785 if (RB_WARN_ON(cpu_buffer
, ++nr_loops
> 3))
3788 if (rb_per_cpu_empty(cpu_buffer
))
3791 if (iter
->head
>= rb_page_size(iter
->head_page
)) {
3796 event
= rb_iter_head_event(iter
);
3798 switch (event
->type_len
) {
3799 case RINGBUF_TYPE_PADDING
:
3800 if (rb_null_event(event
)) {
3804 rb_advance_iter(iter
);
3807 case RINGBUF_TYPE_TIME_EXTEND
:
3808 /* Internal data, OK to advance */
3809 rb_advance_iter(iter
);
3812 case RINGBUF_TYPE_TIME_STAMP
:
3813 /* FIXME: not implemented */
3814 rb_advance_iter(iter
);
3817 case RINGBUF_TYPE_DATA
:
3819 *ts
= iter
->read_stamp
+ event
->time_delta
;
3820 ring_buffer_normalize_time_stamp(buffer
,
3821 cpu_buffer
->cpu
, ts
);
3831 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek
);
3833 static inline bool rb_reader_lock(struct ring_buffer_per_cpu
*cpu_buffer
)
3835 if (likely(!in_nmi())) {
3836 raw_spin_lock(&cpu_buffer
->reader_lock
);
3841 * If an NMI die dumps out the content of the ring buffer
3842 * trylock must be used to prevent a deadlock if the NMI
3843 * preempted a task that holds the ring buffer locks. If
3844 * we get the lock then all is fine, if not, then continue
3845 * to do the read, but this can corrupt the ring buffer,
3846 * so it must be permanently disabled from future writes.
3847 * Reading from NMI is a oneshot deal.
3849 if (raw_spin_trylock(&cpu_buffer
->reader_lock
))
3852 /* Continue without locking, but disable the ring buffer */
3853 atomic_inc(&cpu_buffer
->record_disabled
);
3858 rb_reader_unlock(struct ring_buffer_per_cpu
*cpu_buffer
, bool locked
)
3861 raw_spin_unlock(&cpu_buffer
->reader_lock
);
3866 * ring_buffer_peek - peek at the next event to be read
3867 * @buffer: The ring buffer to read
3868 * @cpu: The cpu to peak at
3869 * @ts: The timestamp counter of this event.
3870 * @lost_events: a variable to store if events were lost (may be NULL)
3872 * This will return the event that will be read next, but does
3873 * not consume the data.
3875 struct ring_buffer_event
*
3876 ring_buffer_peek(struct ring_buffer
*buffer
, int cpu
, u64
*ts
,
3877 unsigned long *lost_events
)
3879 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
3880 struct ring_buffer_event
*event
;
3881 unsigned long flags
;
3884 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3888 local_irq_save(flags
);
3889 dolock
= rb_reader_lock(cpu_buffer
);
3890 event
= rb_buffer_peek(cpu_buffer
, ts
, lost_events
);
3891 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3892 rb_advance_reader(cpu_buffer
);
3893 rb_reader_unlock(cpu_buffer
, dolock
);
3894 local_irq_restore(flags
);
3896 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3903 * ring_buffer_iter_peek - peek at the next event to be read
3904 * @iter: The ring buffer iterator
3905 * @ts: The timestamp counter of this event.
3907 * This will return the event that will be read next, but does
3908 * not increment the iterator.
3910 struct ring_buffer_event
*
3911 ring_buffer_iter_peek(struct ring_buffer_iter
*iter
, u64
*ts
)
3913 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
3914 struct ring_buffer_event
*event
;
3915 unsigned long flags
;
3918 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
3919 event
= rb_iter_peek(iter
, ts
);
3920 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
3922 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3929 * ring_buffer_consume - return an event and consume it
3930 * @buffer: The ring buffer to get the next event from
3931 * @cpu: the cpu to read the buffer from
3932 * @ts: a variable to store the timestamp (may be NULL)
3933 * @lost_events: a variable to store if events were lost (may be NULL)
3935 * Returns the next event in the ring buffer, and that event is consumed.
3936 * Meaning, that sequential reads will keep returning a different event,
3937 * and eventually empty the ring buffer if the producer is slower.
3939 struct ring_buffer_event
*
3940 ring_buffer_consume(struct ring_buffer
*buffer
, int cpu
, u64
*ts
,
3941 unsigned long *lost_events
)
3943 struct ring_buffer_per_cpu
*cpu_buffer
;
3944 struct ring_buffer_event
*event
= NULL
;
3945 unsigned long flags
;
3949 /* might be called in atomic */
3952 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
3955 cpu_buffer
= buffer
->buffers
[cpu
];
3956 local_irq_save(flags
);
3957 dolock
= rb_reader_lock(cpu_buffer
);
3959 event
= rb_buffer_peek(cpu_buffer
, ts
, lost_events
);
3961 cpu_buffer
->lost_events
= 0;
3962 rb_advance_reader(cpu_buffer
);
3965 rb_reader_unlock(cpu_buffer
, dolock
);
3966 local_irq_restore(flags
);
3971 if (event
&& event
->type_len
== RINGBUF_TYPE_PADDING
)
3976 EXPORT_SYMBOL_GPL(ring_buffer_consume
);
3979 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3980 * @buffer: The ring buffer to read from
3981 * @cpu: The cpu buffer to iterate over
3983 * This performs the initial preparations necessary to iterate
3984 * through the buffer. Memory is allocated, buffer recording
3985 * is disabled, and the iterator pointer is returned to the caller.
3987 * Disabling buffer recordng prevents the reading from being
3988 * corrupted. This is not a consuming read, so a producer is not
3991 * After a sequence of ring_buffer_read_prepare calls, the user is
3992 * expected to make at least one call to ring_buffer_read_prepare_sync.
3993 * Afterwards, ring_buffer_read_start is invoked to get things going
3996 * This overall must be paired with ring_buffer_read_finish.
3998 struct ring_buffer_iter
*
3999 ring_buffer_read_prepare(struct ring_buffer
*buffer
, int cpu
)
4001 struct ring_buffer_per_cpu
*cpu_buffer
;
4002 struct ring_buffer_iter
*iter
;
4004 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4007 iter
= kmalloc(sizeof(*iter
), GFP_KERNEL
);
4011 cpu_buffer
= buffer
->buffers
[cpu
];
4013 iter
->cpu_buffer
= cpu_buffer
;
4015 atomic_inc(&buffer
->resize_disabled
);
4016 atomic_inc(&cpu_buffer
->record_disabled
);
4020 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare
);
4023 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4025 * All previously invoked ring_buffer_read_prepare calls to prepare
4026 * iterators will be synchronized. Afterwards, read_buffer_read_start
4027 * calls on those iterators are allowed.
4030 ring_buffer_read_prepare_sync(void)
4032 synchronize_sched();
4034 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync
);
4037 * ring_buffer_read_start - start a non consuming read of the buffer
4038 * @iter: The iterator returned by ring_buffer_read_prepare
4040 * This finalizes the startup of an iteration through the buffer.
4041 * The iterator comes from a call to ring_buffer_read_prepare and
4042 * an intervening ring_buffer_read_prepare_sync must have been
4045 * Must be paired with ring_buffer_read_finish.
4048 ring_buffer_read_start(struct ring_buffer_iter
*iter
)
4050 struct ring_buffer_per_cpu
*cpu_buffer
;
4051 unsigned long flags
;
4056 cpu_buffer
= iter
->cpu_buffer
;
4058 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4059 arch_spin_lock(&cpu_buffer
->lock
);
4060 rb_iter_reset(iter
);
4061 arch_spin_unlock(&cpu_buffer
->lock
);
4062 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4064 EXPORT_SYMBOL_GPL(ring_buffer_read_start
);
4067 * ring_buffer_read_finish - finish reading the iterator of the buffer
4068 * @iter: The iterator retrieved by ring_buffer_start
4070 * This re-enables the recording to the buffer, and frees the
4074 ring_buffer_read_finish(struct ring_buffer_iter
*iter
)
4076 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
4077 unsigned long flags
;
4080 * Ring buffer is disabled from recording, here's a good place
4081 * to check the integrity of the ring buffer.
4082 * Must prevent readers from trying to read, as the check
4083 * clears the HEAD page and readers require it.
4085 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4086 rb_check_pages(cpu_buffer
);
4087 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4089 atomic_dec(&cpu_buffer
->record_disabled
);
4090 atomic_dec(&cpu_buffer
->buffer
->resize_disabled
);
4093 EXPORT_SYMBOL_GPL(ring_buffer_read_finish
);
4096 * ring_buffer_read - read the next item in the ring buffer by the iterator
4097 * @iter: The ring buffer iterator
4098 * @ts: The time stamp of the event read.
4100 * This reads the next event in the ring buffer and increments the iterator.
4102 struct ring_buffer_event
*
4103 ring_buffer_read(struct ring_buffer_iter
*iter
, u64
*ts
)
4105 struct ring_buffer_event
*event
;
4106 struct ring_buffer_per_cpu
*cpu_buffer
= iter
->cpu_buffer
;
4107 unsigned long flags
;
4109 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4111 event
= rb_iter_peek(iter
, ts
);
4115 if (event
->type_len
== RINGBUF_TYPE_PADDING
)
4118 rb_advance_iter(iter
);
4120 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4124 EXPORT_SYMBOL_GPL(ring_buffer_read
);
4127 * ring_buffer_size - return the size of the ring buffer (in bytes)
4128 * @buffer: The ring buffer.
4130 unsigned long ring_buffer_size(struct ring_buffer
*buffer
, int cpu
)
4133 * Earlier, this method returned
4134 * BUF_PAGE_SIZE * buffer->nr_pages
4135 * Since the nr_pages field is now removed, we have converted this to
4136 * return the per cpu buffer value.
4138 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4141 return BUF_PAGE_SIZE
* buffer
->buffers
[cpu
]->nr_pages
;
4143 EXPORT_SYMBOL_GPL(ring_buffer_size
);
4146 rb_reset_cpu(struct ring_buffer_per_cpu
*cpu_buffer
)
4148 rb_head_page_deactivate(cpu_buffer
);
4150 cpu_buffer
->head_page
4151 = list_entry(cpu_buffer
->pages
, struct buffer_page
, list
);
4152 local_set(&cpu_buffer
->head_page
->write
, 0);
4153 local_set(&cpu_buffer
->head_page
->entries
, 0);
4154 local_set(&cpu_buffer
->head_page
->page
->commit
, 0);
4156 cpu_buffer
->head_page
->read
= 0;
4158 cpu_buffer
->tail_page
= cpu_buffer
->head_page
;
4159 cpu_buffer
->commit_page
= cpu_buffer
->head_page
;
4161 INIT_LIST_HEAD(&cpu_buffer
->reader_page
->list
);
4162 INIT_LIST_HEAD(&cpu_buffer
->new_pages
);
4163 local_set(&cpu_buffer
->reader_page
->write
, 0);
4164 local_set(&cpu_buffer
->reader_page
->entries
, 0);
4165 local_set(&cpu_buffer
->reader_page
->page
->commit
, 0);
4166 cpu_buffer
->reader_page
->read
= 0;
4168 local_set(&cpu_buffer
->entries_bytes
, 0);
4169 local_set(&cpu_buffer
->overrun
, 0);
4170 local_set(&cpu_buffer
->commit_overrun
, 0);
4171 local_set(&cpu_buffer
->dropped_events
, 0);
4172 local_set(&cpu_buffer
->entries
, 0);
4173 local_set(&cpu_buffer
->committing
, 0);
4174 local_set(&cpu_buffer
->commits
, 0);
4175 cpu_buffer
->read
= 0;
4176 cpu_buffer
->read_bytes
= 0;
4178 cpu_buffer
->write_stamp
= 0;
4179 cpu_buffer
->read_stamp
= 0;
4181 cpu_buffer
->lost_events
= 0;
4182 cpu_buffer
->last_overrun
= 0;
4184 rb_head_page_activate(cpu_buffer
);
4188 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4189 * @buffer: The ring buffer to reset a per cpu buffer of
4190 * @cpu: The CPU buffer to be reset
4192 void ring_buffer_reset_cpu(struct ring_buffer
*buffer
, int cpu
)
4194 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
4195 unsigned long flags
;
4197 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4200 atomic_inc(&buffer
->resize_disabled
);
4201 atomic_inc(&cpu_buffer
->record_disabled
);
4203 /* Make sure all commits have finished */
4204 synchronize_sched();
4206 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4208 if (RB_WARN_ON(cpu_buffer
, local_read(&cpu_buffer
->committing
)))
4211 arch_spin_lock(&cpu_buffer
->lock
);
4213 rb_reset_cpu(cpu_buffer
);
4215 arch_spin_unlock(&cpu_buffer
->lock
);
4218 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4220 atomic_dec(&cpu_buffer
->record_disabled
);
4221 atomic_dec(&buffer
->resize_disabled
);
4223 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu
);
4226 * ring_buffer_reset - reset a ring buffer
4227 * @buffer: The ring buffer to reset all cpu buffers
4229 void ring_buffer_reset(struct ring_buffer
*buffer
)
4233 for_each_buffer_cpu(buffer
, cpu
)
4234 ring_buffer_reset_cpu(buffer
, cpu
);
4236 EXPORT_SYMBOL_GPL(ring_buffer_reset
);
4239 * rind_buffer_empty - is the ring buffer empty?
4240 * @buffer: The ring buffer to test
4242 bool ring_buffer_empty(struct ring_buffer
*buffer
)
4244 struct ring_buffer_per_cpu
*cpu_buffer
;
4245 unsigned long flags
;
4250 /* yes this is racy, but if you don't like the race, lock the buffer */
4251 for_each_buffer_cpu(buffer
, cpu
) {
4252 cpu_buffer
= buffer
->buffers
[cpu
];
4253 local_irq_save(flags
);
4254 dolock
= rb_reader_lock(cpu_buffer
);
4255 ret
= rb_per_cpu_empty(cpu_buffer
);
4256 rb_reader_unlock(cpu_buffer
, dolock
);
4257 local_irq_restore(flags
);
4265 EXPORT_SYMBOL_GPL(ring_buffer_empty
);
4268 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4269 * @buffer: The ring buffer
4270 * @cpu: The CPU buffer to test
4272 bool ring_buffer_empty_cpu(struct ring_buffer
*buffer
, int cpu
)
4274 struct ring_buffer_per_cpu
*cpu_buffer
;
4275 unsigned long flags
;
4279 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4282 cpu_buffer
= buffer
->buffers
[cpu
];
4283 local_irq_save(flags
);
4284 dolock
= rb_reader_lock(cpu_buffer
);
4285 ret
= rb_per_cpu_empty(cpu_buffer
);
4286 rb_reader_unlock(cpu_buffer
, dolock
);
4287 local_irq_restore(flags
);
4291 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu
);
4293 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4295 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4296 * @buffer_a: One buffer to swap with
4297 * @buffer_b: The other buffer to swap with
4299 * This function is useful for tracers that want to take a "snapshot"
4300 * of a CPU buffer and has another back up buffer lying around.
4301 * it is expected that the tracer handles the cpu buffer not being
4302 * used at the moment.
4304 int ring_buffer_swap_cpu(struct ring_buffer
*buffer_a
,
4305 struct ring_buffer
*buffer_b
, int cpu
)
4307 struct ring_buffer_per_cpu
*cpu_buffer_a
;
4308 struct ring_buffer_per_cpu
*cpu_buffer_b
;
4311 if (!cpumask_test_cpu(cpu
, buffer_a
->cpumask
) ||
4312 !cpumask_test_cpu(cpu
, buffer_b
->cpumask
))
4315 cpu_buffer_a
= buffer_a
->buffers
[cpu
];
4316 cpu_buffer_b
= buffer_b
->buffers
[cpu
];
4318 /* At least make sure the two buffers are somewhat the same */
4319 if (cpu_buffer_a
->nr_pages
!= cpu_buffer_b
->nr_pages
)
4324 if (atomic_read(&buffer_a
->record_disabled
))
4327 if (atomic_read(&buffer_b
->record_disabled
))
4330 if (atomic_read(&cpu_buffer_a
->record_disabled
))
4333 if (atomic_read(&cpu_buffer_b
->record_disabled
))
4337 * We can't do a synchronize_sched here because this
4338 * function can be called in atomic context.
4339 * Normally this will be called from the same CPU as cpu.
4340 * If not it's up to the caller to protect this.
4342 atomic_inc(&cpu_buffer_a
->record_disabled
);
4343 atomic_inc(&cpu_buffer_b
->record_disabled
);
4346 if (local_read(&cpu_buffer_a
->committing
))
4348 if (local_read(&cpu_buffer_b
->committing
))
4351 buffer_a
->buffers
[cpu
] = cpu_buffer_b
;
4352 buffer_b
->buffers
[cpu
] = cpu_buffer_a
;
4354 cpu_buffer_b
->buffer
= buffer_a
;
4355 cpu_buffer_a
->buffer
= buffer_b
;
4360 atomic_dec(&cpu_buffer_a
->record_disabled
);
4361 atomic_dec(&cpu_buffer_b
->record_disabled
);
4365 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu
);
4366 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4369 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4370 * @buffer: the buffer to allocate for.
4371 * @cpu: the cpu buffer to allocate.
4373 * This function is used in conjunction with ring_buffer_read_page.
4374 * When reading a full page from the ring buffer, these functions
4375 * can be used to speed up the process. The calling function should
4376 * allocate a few pages first with this function. Then when it
4377 * needs to get pages from the ring buffer, it passes the result
4378 * of this function into ring_buffer_read_page, which will swap
4379 * the page that was allocated, with the read page of the buffer.
4382 * The page allocated, or ERR_PTR
4384 void *ring_buffer_alloc_read_page(struct ring_buffer
*buffer
, int cpu
)
4386 struct ring_buffer_per_cpu
*cpu_buffer
;
4387 struct buffer_data_page
*bpage
= NULL
;
4388 unsigned long flags
;
4391 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4392 return ERR_PTR(-ENODEV
);
4394 cpu_buffer
= buffer
->buffers
[cpu
];
4395 local_irq_save(flags
);
4396 arch_spin_lock(&cpu_buffer
->lock
);
4398 if (cpu_buffer
->free_page
) {
4399 bpage
= cpu_buffer
->free_page
;
4400 cpu_buffer
->free_page
= NULL
;
4403 arch_spin_unlock(&cpu_buffer
->lock
);
4404 local_irq_restore(flags
);
4409 page
= alloc_pages_node(cpu_to_node(cpu
),
4410 GFP_KERNEL
| __GFP_NORETRY
, 0);
4412 return ERR_PTR(-ENOMEM
);
4414 bpage
= page_address(page
);
4417 rb_init_page(bpage
);
4421 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page
);
4424 * ring_buffer_free_read_page - free an allocated read page
4425 * @buffer: the buffer the page was allocate for
4426 * @cpu: the cpu buffer the page came from
4427 * @data: the page to free
4429 * Free a page allocated from ring_buffer_alloc_read_page.
4431 void ring_buffer_free_read_page(struct ring_buffer
*buffer
, int cpu
, void *data
)
4433 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
4434 struct buffer_data_page
*bpage
= data
;
4435 struct page
*page
= virt_to_page(bpage
);
4436 unsigned long flags
;
4438 /* If the page is still in use someplace else, we can't reuse it */
4439 if (page_ref_count(page
) > 1)
4442 local_irq_save(flags
);
4443 arch_spin_lock(&cpu_buffer
->lock
);
4445 if (!cpu_buffer
->free_page
) {
4446 cpu_buffer
->free_page
= bpage
;
4450 arch_spin_unlock(&cpu_buffer
->lock
);
4451 local_irq_restore(flags
);
4454 free_page((unsigned long)bpage
);
4456 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page
);
4459 * ring_buffer_read_page - extract a page from the ring buffer
4460 * @buffer: buffer to extract from
4461 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4462 * @len: amount to extract
4463 * @cpu: the cpu of the buffer to extract
4464 * @full: should the extraction only happen when the page is full.
4466 * This function will pull out a page from the ring buffer and consume it.
4467 * @data_page must be the address of the variable that was returned
4468 * from ring_buffer_alloc_read_page. This is because the page might be used
4469 * to swap with a page in the ring buffer.
4472 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4473 * if (IS_ERR(rpage))
4474 * return PTR_ERR(rpage);
4475 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4477 * process_page(rpage, ret);
4479 * When @full is set, the function will not return true unless
4480 * the writer is off the reader page.
4482 * Note: it is up to the calling functions to handle sleeps and wakeups.
4483 * The ring buffer can be used anywhere in the kernel and can not
4484 * blindly call wake_up. The layer that uses the ring buffer must be
4485 * responsible for that.
4488 * >=0 if data has been transferred, returns the offset of consumed data.
4489 * <0 if no data has been transferred.
4491 int ring_buffer_read_page(struct ring_buffer
*buffer
,
4492 void **data_page
, size_t len
, int cpu
, int full
)
4494 struct ring_buffer_per_cpu
*cpu_buffer
= buffer
->buffers
[cpu
];
4495 struct ring_buffer_event
*event
;
4496 struct buffer_data_page
*bpage
;
4497 struct buffer_page
*reader
;
4498 unsigned long missed_events
;
4499 unsigned long flags
;
4500 unsigned int commit
;
4505 if (!cpumask_test_cpu(cpu
, buffer
->cpumask
))
4509 * If len is not big enough to hold the page header, then
4510 * we can not copy anything.
4512 if (len
<= BUF_PAGE_HDR_SIZE
)
4515 len
-= BUF_PAGE_HDR_SIZE
;
4524 raw_spin_lock_irqsave(&cpu_buffer
->reader_lock
, flags
);
4526 reader
= rb_get_reader_page(cpu_buffer
);
4530 event
= rb_reader_event(cpu_buffer
);
4532 read
= reader
->read
;
4533 commit
= rb_page_commit(reader
);
4535 /* Check if any events were dropped */
4536 missed_events
= cpu_buffer
->lost_events
;
4539 * If this page has been partially read or
4540 * if len is not big enough to read the rest of the page or
4541 * a writer is still on the page, then
4542 * we must copy the data from the page to the buffer.
4543 * Otherwise, we can simply swap the page with the one passed in.
4545 if (read
|| (len
< (commit
- read
)) ||
4546 cpu_buffer
->reader_page
== cpu_buffer
->commit_page
) {
4547 struct buffer_data_page
*rpage
= cpu_buffer
->reader_page
->page
;
4548 unsigned int rpos
= read
;
4549 unsigned int pos
= 0;
4555 if (len
> (commit
- read
))
4556 len
= (commit
- read
);
4558 /* Always keep the time extend and data together */
4559 size
= rb_event_ts_length(event
);
4564 /* save the current timestamp, since the user will need it */
4565 save_timestamp
= cpu_buffer
->read_stamp
;
4567 /* Need to copy one event at a time */
4569 /* We need the size of one event, because
4570 * rb_advance_reader only advances by one event,
4571 * whereas rb_event_ts_length may include the size of
4572 * one or two events.
4573 * We have already ensured there's enough space if this
4574 * is a time extend. */
4575 size
= rb_event_length(event
);
4576 memcpy(bpage
->data
+ pos
, rpage
->data
+ rpos
, size
);
4580 rb_advance_reader(cpu_buffer
);
4581 rpos
= reader
->read
;
4587 event
= rb_reader_event(cpu_buffer
);
4588 /* Always keep the time extend and data together */
4589 size
= rb_event_ts_length(event
);
4590 } while (len
>= size
);
4593 local_set(&bpage
->commit
, pos
);
4594 bpage
->time_stamp
= save_timestamp
;
4596 /* we copied everything to the beginning */
4599 /* update the entry counter */
4600 cpu_buffer
->read
+= rb_page_entries(reader
);
4601 cpu_buffer
->read_bytes
+= BUF_PAGE_SIZE
;
4603 /* swap the pages */
4604 rb_init_page(bpage
);
4605 bpage
= reader
->page
;
4606 reader
->page
= *data_page
;
4607 local_set(&reader
->write
, 0);
4608 local_set(&reader
->entries
, 0);
4613 * Use the real_end for the data size,
4614 * This gives us a chance to store the lost events
4617 if (reader
->real_end
)
4618 local_set(&bpage
->commit
, reader
->real_end
);
4622 cpu_buffer
->lost_events
= 0;
4624 commit
= local_read(&bpage
->commit
);
4626 * Set a flag in the commit field if we lost events
4628 if (missed_events
) {
4629 /* If there is room at the end of the page to save the
4630 * missed events, then record it there.
4632 if (BUF_PAGE_SIZE
- commit
>= sizeof(missed_events
)) {
4633 memcpy(&bpage
->data
[commit
], &missed_events
,
4634 sizeof(missed_events
));
4635 local_add(RB_MISSED_STORED
, &bpage
->commit
);
4636 commit
+= sizeof(missed_events
);
4638 local_add(RB_MISSED_EVENTS
, &bpage
->commit
);
4642 * This page may be off to user land. Zero it out here.
4644 if (commit
< BUF_PAGE_SIZE
)
4645 memset(&bpage
->data
[commit
], 0, BUF_PAGE_SIZE
- commit
);
4648 raw_spin_unlock_irqrestore(&cpu_buffer
->reader_lock
, flags
);
4653 EXPORT_SYMBOL_GPL(ring_buffer_read_page
);
4656 * We only allocate new buffers, never free them if the CPU goes down.
4657 * If we were to free the buffer, then the user would lose any trace that was in
4660 int trace_rb_cpu_prepare(unsigned int cpu
, struct hlist_node
*node
)
4662 struct ring_buffer
*buffer
;
4665 unsigned long nr_pages
;
4667 buffer
= container_of(node
, struct ring_buffer
, node
);
4668 if (cpumask_test_cpu(cpu
, buffer
->cpumask
))
4673 /* check if all cpu sizes are same */
4674 for_each_buffer_cpu(buffer
, cpu_i
) {
4675 /* fill in the size from first enabled cpu */
4677 nr_pages
= buffer
->buffers
[cpu_i
]->nr_pages
;
4678 if (nr_pages
!= buffer
->buffers
[cpu_i
]->nr_pages
) {
4683 /* allocate minimum pages, user can later expand it */
4686 buffer
->buffers
[cpu
] =
4687 rb_allocate_cpu_buffer(buffer
, nr_pages
, cpu
);
4688 if (!buffer
->buffers
[cpu
]) {
4689 WARN(1, "failed to allocate ring buffer on CPU %u\n",
4694 cpumask_set_cpu(cpu
, buffer
->cpumask
);
4698 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4700 * This is a basic integrity check of the ring buffer.
4701 * Late in the boot cycle this test will run when configured in.
4702 * It will kick off a thread per CPU that will go into a loop
4703 * writing to the per cpu ring buffer various sizes of data.
4704 * Some of the data will be large items, some small.
4706 * Another thread is created that goes into a spin, sending out
4707 * IPIs to the other CPUs to also write into the ring buffer.
4708 * this is to test the nesting ability of the buffer.
4710 * Basic stats are recorded and reported. If something in the
4711 * ring buffer should happen that's not expected, a big warning
4712 * is displayed and all ring buffers are disabled.
4714 static struct task_struct
*rb_threads
[NR_CPUS
] __initdata
;
4716 struct rb_test_data
{
4717 struct ring_buffer
*buffer
;
4718 unsigned long events
;
4719 unsigned long bytes_written
;
4720 unsigned long bytes_alloc
;
4721 unsigned long bytes_dropped
;
4722 unsigned long events_nested
;
4723 unsigned long bytes_written_nested
;
4724 unsigned long bytes_alloc_nested
;
4725 unsigned long bytes_dropped_nested
;
4726 int min_size_nested
;
4727 int max_size_nested
;
4734 static struct rb_test_data rb_data
[NR_CPUS
] __initdata
;
4737 #define RB_TEST_BUFFER_SIZE 1048576
4739 static char rb_string
[] __initdata
=
4740 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4741 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4742 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4744 static bool rb_test_started __initdata
;
4751 static __init
int rb_write_something(struct rb_test_data
*data
, bool nested
)
4753 struct ring_buffer_event
*event
;
4754 struct rb_item
*item
;
4761 /* Have nested writes different that what is written */
4762 cnt
= data
->cnt
+ (nested
? 27 : 0);
4764 /* Multiply cnt by ~e, to make some unique increment */
4765 size
= (data
->cnt
* 68 / 25) % (sizeof(rb_string
) - 1);
4767 len
= size
+ sizeof(struct rb_item
);
4769 started
= rb_test_started
;
4770 /* read rb_test_started before checking buffer enabled */
4773 event
= ring_buffer_lock_reserve(data
->buffer
, len
);
4775 /* Ignore dropped events before test starts. */
4778 data
->bytes_dropped
+= len
;
4780 data
->bytes_dropped_nested
+= len
;
4785 event_len
= ring_buffer_event_length(event
);
4787 if (RB_WARN_ON(data
->buffer
, event_len
< len
))
4790 item
= ring_buffer_event_data(event
);
4792 memcpy(item
->str
, rb_string
, size
);
4795 data
->bytes_alloc_nested
+= event_len
;
4796 data
->bytes_written_nested
+= len
;
4797 data
->events_nested
++;
4798 if (!data
->min_size_nested
|| len
< data
->min_size_nested
)
4799 data
->min_size_nested
= len
;
4800 if (len
> data
->max_size_nested
)
4801 data
->max_size_nested
= len
;
4803 data
->bytes_alloc
+= event_len
;
4804 data
->bytes_written
+= len
;
4806 if (!data
->min_size
|| len
< data
->min_size
)
4807 data
->max_size
= len
;
4808 if (len
> data
->max_size
)
4809 data
->max_size
= len
;
4813 ring_buffer_unlock_commit(data
->buffer
, event
);
4818 static __init
int rb_test(void *arg
)
4820 struct rb_test_data
*data
= arg
;
4822 while (!kthread_should_stop()) {
4823 rb_write_something(data
, false);
4826 set_current_state(TASK_INTERRUPTIBLE
);
4827 /* Now sleep between a min of 100-300us and a max of 1ms */
4828 usleep_range(((data
->cnt
% 3) + 1) * 100, 1000);
4834 static __init
void rb_ipi(void *ignore
)
4836 struct rb_test_data
*data
;
4837 int cpu
= smp_processor_id();
4839 data
= &rb_data
[cpu
];
4840 rb_write_something(data
, true);
4843 static __init
int rb_hammer_test(void *arg
)
4845 while (!kthread_should_stop()) {
4847 /* Send an IPI to all cpus to write data! */
4848 smp_call_function(rb_ipi
, NULL
, 1);
4849 /* No sleep, but for non preempt, let others run */
4856 static __init
int test_ringbuffer(void)
4858 struct task_struct
*rb_hammer
;
4859 struct ring_buffer
*buffer
;
4863 pr_info("Running ring buffer tests...\n");
4865 buffer
= ring_buffer_alloc(RB_TEST_BUFFER_SIZE
, RB_FL_OVERWRITE
);
4866 if (WARN_ON(!buffer
))
4869 /* Disable buffer so that threads can't write to it yet */
4870 ring_buffer_record_off(buffer
);
4872 for_each_online_cpu(cpu
) {
4873 rb_data
[cpu
].buffer
= buffer
;
4874 rb_data
[cpu
].cpu
= cpu
;
4875 rb_data
[cpu
].cnt
= cpu
;
4876 rb_threads
[cpu
] = kthread_create(rb_test
, &rb_data
[cpu
],
4877 "rbtester/%d", cpu
);
4878 if (WARN_ON(IS_ERR(rb_threads
[cpu
]))) {
4879 pr_cont("FAILED\n");
4880 ret
= PTR_ERR(rb_threads
[cpu
]);
4884 kthread_bind(rb_threads
[cpu
], cpu
);
4885 wake_up_process(rb_threads
[cpu
]);
4888 /* Now create the rb hammer! */
4889 rb_hammer
= kthread_run(rb_hammer_test
, NULL
, "rbhammer");
4890 if (WARN_ON(IS_ERR(rb_hammer
))) {
4891 pr_cont("FAILED\n");
4892 ret
= PTR_ERR(rb_hammer
);
4896 ring_buffer_record_on(buffer
);
4898 * Show buffer is enabled before setting rb_test_started.
4899 * Yes there's a small race window where events could be
4900 * dropped and the thread wont catch it. But when a ring
4901 * buffer gets enabled, there will always be some kind of
4902 * delay before other CPUs see it. Thus, we don't care about
4903 * those dropped events. We care about events dropped after
4904 * the threads see that the buffer is active.
4907 rb_test_started
= true;
4909 set_current_state(TASK_INTERRUPTIBLE
);
4910 /* Just run for 10 seconds */;
4911 schedule_timeout(10 * HZ
);
4913 kthread_stop(rb_hammer
);
4916 for_each_online_cpu(cpu
) {
4917 if (!rb_threads
[cpu
])
4919 kthread_stop(rb_threads
[cpu
]);
4922 ring_buffer_free(buffer
);
4927 pr_info("finished\n");
4928 for_each_online_cpu(cpu
) {
4929 struct ring_buffer_event
*event
;
4930 struct rb_test_data
*data
= &rb_data
[cpu
];
4931 struct rb_item
*item
;
4932 unsigned long total_events
;
4933 unsigned long total_dropped
;
4934 unsigned long total_written
;
4935 unsigned long total_alloc
;
4936 unsigned long total_read
= 0;
4937 unsigned long total_size
= 0;
4938 unsigned long total_len
= 0;
4939 unsigned long total_lost
= 0;
4942 int small_event_size
;
4946 total_events
= data
->events
+ data
->events_nested
;
4947 total_written
= data
->bytes_written
+ data
->bytes_written_nested
;
4948 total_alloc
= data
->bytes_alloc
+ data
->bytes_alloc_nested
;
4949 total_dropped
= data
->bytes_dropped
+ data
->bytes_dropped_nested
;
4951 big_event_size
= data
->max_size
+ data
->max_size_nested
;
4952 small_event_size
= data
->min_size
+ data
->min_size_nested
;
4954 pr_info("CPU %d:\n", cpu
);
4955 pr_info(" events: %ld\n", total_events
);
4956 pr_info(" dropped bytes: %ld\n", total_dropped
);
4957 pr_info(" alloced bytes: %ld\n", total_alloc
);
4958 pr_info(" written bytes: %ld\n", total_written
);
4959 pr_info(" biggest event: %d\n", big_event_size
);
4960 pr_info(" smallest event: %d\n", small_event_size
);
4962 if (RB_WARN_ON(buffer
, total_dropped
))
4967 while ((event
= ring_buffer_consume(buffer
, cpu
, NULL
, &lost
))) {
4969 item
= ring_buffer_event_data(event
);
4970 total_len
+= ring_buffer_event_length(event
);
4971 total_size
+= item
->size
+ sizeof(struct rb_item
);
4972 if (memcmp(&item
->str
[0], rb_string
, item
->size
) != 0) {
4973 pr_info("FAILED!\n");
4974 pr_info("buffer had: %.*s\n", item
->size
, item
->str
);
4975 pr_info("expected: %.*s\n", item
->size
, rb_string
);
4976 RB_WARN_ON(buffer
, 1);
4987 pr_info(" read events: %ld\n", total_read
);
4988 pr_info(" lost events: %ld\n", total_lost
);
4989 pr_info(" total events: %ld\n", total_lost
+ total_read
);
4990 pr_info(" recorded len bytes: %ld\n", total_len
);
4991 pr_info(" recorded size bytes: %ld\n", total_size
);
4993 pr_info(" With dropped events, record len and size may not match\n"
4994 " alloced and written from above\n");
4996 if (RB_WARN_ON(buffer
, total_len
!= total_alloc
||
4997 total_size
!= total_written
))
5000 if (RB_WARN_ON(buffer
, total_lost
+ total_read
!= total_events
))
5006 pr_info("Ring buffer PASSED!\n");
5008 ring_buffer_free(buffer
);
5012 late_initcall(test_ringbuffer
);
5013 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */