4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 #include <linux/ftrace_event.h>
7 #include <linux/ring_buffer.h>
8 #include <linux/trace_clock.h>
9 #include <linux/trace_seq.h>
10 #include <linux/spinlock.h>
11 #include <linux/irq_work.h>
12 #include <linux/debugfs.h>
13 #include <linux/uaccess.h>
14 #include <linux/hardirq.h>
15 #include <linux/kmemcheck.h>
16 #include <linux/module.h>
17 #include <linux/percpu.h>
18 #include <linux/mutex.h>
19 #include <linux/slab.h>
20 #include <linux/init.h>
21 #include <linux/hash.h>
22 #include <linux/list.h>
23 #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)
37 ret = trace_seq_printf(s, "# compressed entry header\n");
38 ret = trace_seq_printf(s, "\ttype_len : 5 bits\n");
39 ret = trace_seq_printf(s, "\ttime_delta : 27 bits\n");
40 ret = trace_seq_printf(s, "\tarray : 32 bits\n");
41 ret = trace_seq_printf(s, "\n");
42 ret = trace_seq_printf(s, "\tpadding : type == %d\n",
43 RINGBUF_TYPE_PADDING);
44 ret = trace_seq_printf(s, "\ttime_extend : type == %d\n",
45 RINGBUF_TYPE_TIME_EXTEND);
46 ret = trace_seq_printf(s, "\tdata max type_len == %d\n",
47 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
53 * The ring buffer is made up of a list of pages. A separate list of pages is
54 * allocated for each CPU. A writer may only write to a buffer that is
55 * associated with the CPU it is currently executing on. A reader may read
56 * from any per cpu buffer.
58 * The reader is special. For each per cpu buffer, the reader has its own
59 * reader page. When a reader has read the entire reader page, this reader
60 * page is swapped with another page in the ring buffer.
62 * Now, as long as the writer is off the reader page, the reader can do what
63 * ever it wants with that page. The writer will never write to that page
64 * again (as long as it is out of the ring buffer).
66 * Here's some silly ASCII art.
69 * |reader| RING BUFFER
71 * +------+ +---+ +---+ +---+
80 * |reader| RING BUFFER
81 * |page |------------------v
82 * +------+ +---+ +---+ +---+
91 * |reader| RING BUFFER
92 * |page |------------------v
93 * +------+ +---+ +---+ +---+
98 * +------------------------------+
102 * |buffer| RING BUFFER
103 * |page |------------------v
104 * +------+ +---+ +---+ +---+
106 * | New +---+ +---+ +---+
109 * +------------------------------+
112 * After we make this swap, the reader can hand this page off to the splice
113 * code and be done with it. It can even allocate a new page if it needs to
114 * and swap that into the ring buffer.
116 * We will be using cmpxchg soon to make all this lockless.
121 * A fast way to enable or disable all ring buffers is to
122 * call tracing_on or tracing_off. Turning off the ring buffers
123 * prevents all ring buffers from being recorded to.
124 * Turning this switch on, makes it OK to write to the
125 * ring buffer, if the ring buffer is enabled itself.
127 * There's three layers that must be on in order to write
128 * to the ring buffer.
130 * 1) This global flag must be set.
131 * 2) The ring buffer must be enabled for recording.
132 * 3) The per cpu buffer must be enabled for recording.
134 * In case of an anomaly, this global flag has a bit set that
135 * will permantly disable all ring buffers.
139 * Global flag to disable all recording to ring buffers
140 * This has two bits: ON, DISABLED
144 * 0 0 : ring buffers are off
145 * 1 0 : ring buffers are on
146 * X 1 : ring buffers are permanently disabled
150 RB_BUFFERS_ON_BIT = 0,
151 RB_BUFFERS_DISABLED_BIT = 1,
155 RB_BUFFERS_ON = 1 << RB_BUFFERS_ON_BIT,
156 RB_BUFFERS_DISABLED = 1 << RB_BUFFERS_DISABLED_BIT,
159 static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON;
161 /* Used for individual buffers (after the counter) */
162 #define RB_BUFFER_OFF (1 << 20)
164 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
167 * tracing_off_permanent - permanently disable ring buffers
169 * This function, once called, will disable all ring buffers
172 void tracing_off_permanent(void)
174 set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags);
177 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
178 #define RB_ALIGNMENT 4U
179 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
180 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
182 #if !defined(CONFIG_64BIT) || defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
183 # define RB_FORCE_8BYTE_ALIGNMENT 0
184 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
186 # define RB_FORCE_8BYTE_ALIGNMENT 1
187 # define RB_ARCH_ALIGNMENT 8U
190 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
191 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
194 RB_LEN_TIME_EXTEND = 8,
195 RB_LEN_TIME_STAMP = 16,
198 #define skip_time_extend(event) \
199 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
201 static inline int rb_null_event(struct ring_buffer_event *event)
203 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
206 static void rb_event_set_padding(struct ring_buffer_event *event)
208 /* padding has a NULL time_delta */
209 event->type_len = RINGBUF_TYPE_PADDING;
210 event->time_delta = 0;
214 rb_event_data_length(struct ring_buffer_event *event)
219 length = event->type_len * RB_ALIGNMENT;
221 length = event->array[0];
222 return length + RB_EVNT_HDR_SIZE;
226 * Return the length of the given event. Will return
227 * the length of the time extend if the event is a
230 static inline unsigned
231 rb_event_length(struct ring_buffer_event *event)
233 switch (event->type_len) {
234 case RINGBUF_TYPE_PADDING:
235 if (rb_null_event(event))
238 return event->array[0] + RB_EVNT_HDR_SIZE;
240 case RINGBUF_TYPE_TIME_EXTEND:
241 return RB_LEN_TIME_EXTEND;
243 case RINGBUF_TYPE_TIME_STAMP:
244 return RB_LEN_TIME_STAMP;
246 case RINGBUF_TYPE_DATA:
247 return rb_event_data_length(event);
256 * Return total length of time extend and data,
257 * or just the event length for all other events.
259 static inline unsigned
260 rb_event_ts_length(struct ring_buffer_event *event)
264 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
265 /* time extends include the data event after it */
266 len = RB_LEN_TIME_EXTEND;
267 event = skip_time_extend(event);
269 return len + rb_event_length(event);
273 * ring_buffer_event_length - return the length of the event
274 * @event: the event to get the length of
276 * Returns the size of the data load of a data event.
277 * If the event is something other than a data event, it
278 * returns the size of the event itself. With the exception
279 * of a TIME EXTEND, where it still returns the size of the
280 * data load of the data event after it.
282 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
286 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
287 event = skip_time_extend(event);
289 length = rb_event_length(event);
290 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
292 length -= RB_EVNT_HDR_SIZE;
293 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
294 length -= sizeof(event->array[0]);
297 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
299 /* inline for ring buffer fast paths */
301 rb_event_data(struct ring_buffer_event *event)
303 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
304 event = skip_time_extend(event);
305 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
306 /* If length is in len field, then array[0] has the data */
308 return (void *)&event->array[0];
309 /* Otherwise length is in array[0] and array[1] has the data */
310 return (void *)&event->array[1];
314 * ring_buffer_event_data - return the data of the event
315 * @event: the event to get the data from
317 void *ring_buffer_event_data(struct ring_buffer_event *event)
319 return rb_event_data(event);
321 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
323 #define for_each_buffer_cpu(buffer, cpu) \
324 for_each_cpu(cpu, buffer->cpumask)
327 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
328 #define TS_DELTA_TEST (~TS_MASK)
330 /* Flag when events were overwritten */
331 #define RB_MISSED_EVENTS (1 << 31)
332 /* Missed count stored at end */
333 #define RB_MISSED_STORED (1 << 30)
335 struct buffer_data_page {
336 u64 time_stamp; /* page time stamp */
337 local_t commit; /* write committed index */
338 unsigned char data[]; /* data of buffer page */
342 * Note, the buffer_page list must be first. The buffer pages
343 * are allocated in cache lines, which means that each buffer
344 * page will be at the beginning of a cache line, and thus
345 * the least significant bits will be zero. We use this to
346 * add flags in the list struct pointers, to make the ring buffer
350 struct list_head list; /* list of buffer pages */
351 local_t write; /* index for next write */
352 unsigned read; /* index for next read */
353 local_t entries; /* entries on this page */
354 unsigned long real_end; /* real end of data */
355 struct buffer_data_page *page; /* Actual data page */
359 * The buffer page counters, write and entries, must be reset
360 * atomically when crossing page boundaries. To synchronize this
361 * update, two counters are inserted into the number. One is
362 * the actual counter for the write position or count on the page.
364 * The other is a counter of updaters. Before an update happens
365 * the update partition of the counter is incremented. This will
366 * allow the updater to update the counter atomically.
368 * The counter is 20 bits, and the state data is 12.
370 #define RB_WRITE_MASK 0xfffff
371 #define RB_WRITE_INTCNT (1 << 20)
373 static void rb_init_page(struct buffer_data_page *bpage)
375 local_set(&bpage->commit, 0);
379 * ring_buffer_page_len - the size of data on the page.
380 * @page: The page to read
382 * Returns the amount of data on the page, including buffer page header.
384 size_t ring_buffer_page_len(void *page)
386 return local_read(&((struct buffer_data_page *)page)->commit)
391 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
394 static void free_buffer_page(struct buffer_page *bpage)
396 free_page((unsigned long)bpage->page);
401 * We need to fit the time_stamp delta into 27 bits.
403 static inline int test_time_stamp(u64 delta)
405 if (delta & TS_DELTA_TEST)
410 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
412 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
413 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
415 int ring_buffer_print_page_header(struct trace_seq *s)
417 struct buffer_data_page field;
420 ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t"
421 "offset:0;\tsize:%u;\tsigned:%u;\n",
422 (unsigned int)sizeof(field.time_stamp),
423 (unsigned int)is_signed_type(u64));
425 ret = trace_seq_printf(s, "\tfield: local_t commit;\t"
426 "offset:%u;\tsize:%u;\tsigned:%u;\n",
427 (unsigned int)offsetof(typeof(field), commit),
428 (unsigned int)sizeof(field.commit),
429 (unsigned int)is_signed_type(long));
431 ret = trace_seq_printf(s, "\tfield: int overwrite;\t"
432 "offset:%u;\tsize:%u;\tsigned:%u;\n",
433 (unsigned int)offsetof(typeof(field), commit),
435 (unsigned int)is_signed_type(long));
437 ret = trace_seq_printf(s, "\tfield: char data;\t"
438 "offset:%u;\tsize:%u;\tsigned:%u;\n",
439 (unsigned int)offsetof(typeof(field), data),
440 (unsigned int)BUF_PAGE_SIZE,
441 (unsigned int)is_signed_type(char));
447 struct irq_work work;
448 wait_queue_head_t waiters;
449 bool waiters_pending;
453 * head_page == tail_page && head == tail then buffer is empty.
455 struct ring_buffer_per_cpu {
457 atomic_t record_disabled;
458 struct ring_buffer *buffer;
459 raw_spinlock_t reader_lock; /* serialize readers */
460 arch_spinlock_t lock;
461 struct lock_class_key lock_key;
462 unsigned int nr_pages;
463 struct list_head *pages;
464 struct buffer_page *head_page; /* read from head */
465 struct buffer_page *tail_page; /* write to tail */
466 struct buffer_page *commit_page; /* committed pages */
467 struct buffer_page *reader_page;
468 unsigned long lost_events;
469 unsigned long last_overrun;
470 local_t entries_bytes;
473 local_t commit_overrun;
474 local_t dropped_events;
478 unsigned long read_bytes;
481 /* ring buffer pages to update, > 0 to add, < 0 to remove */
482 int nr_pages_to_update;
483 struct list_head new_pages; /* new pages to add */
484 struct work_struct update_pages_work;
485 struct completion update_done;
487 struct rb_irq_work irq_work;
493 atomic_t record_disabled;
494 atomic_t resize_disabled;
495 cpumask_var_t cpumask;
497 struct lock_class_key *reader_lock_key;
501 struct ring_buffer_per_cpu **buffers;
503 #ifdef CONFIG_HOTPLUG_CPU
504 struct notifier_block cpu_notify;
508 struct rb_irq_work irq_work;
511 struct ring_buffer_iter {
512 struct ring_buffer_per_cpu *cpu_buffer;
514 struct buffer_page *head_page;
515 struct buffer_page *cache_reader_page;
516 unsigned long cache_read;
521 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
523 * Schedules a delayed work to wake up any task that is blocked on the
524 * ring buffer waiters queue.
526 static void rb_wake_up_waiters(struct irq_work *work)
528 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
530 wake_up_all(&rbwork->waiters);
534 * ring_buffer_wait - wait for input to the ring buffer
535 * @buffer: buffer to wait on
536 * @cpu: the cpu buffer to wait on
538 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
539 * as data is added to any of the @buffer's cpu buffers. Otherwise
540 * it will wait for data to be added to a specific cpu buffer.
542 void ring_buffer_wait(struct ring_buffer *buffer, int cpu)
544 struct ring_buffer_per_cpu *cpu_buffer;
546 struct rb_irq_work *work;
549 * Depending on what the caller is waiting for, either any
550 * data in any cpu buffer, or a specific buffer, put the
551 * caller on the appropriate wait queue.
553 if (cpu == RING_BUFFER_ALL_CPUS)
554 work = &buffer->irq_work;
556 cpu_buffer = buffer->buffers[cpu];
557 work = &cpu_buffer->irq_work;
561 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
564 * The events can happen in critical sections where
565 * checking a work queue can cause deadlocks.
566 * After adding a task to the queue, this flag is set
567 * only to notify events to try to wake up the queue
570 * We don't clear it even if the buffer is no longer
571 * empty. The flag only causes the next event to run
572 * irq_work to do the work queue wake up. The worse
573 * that can happen if we race with !trace_empty() is that
574 * an event will cause an irq_work to try to wake up
577 * There's no reason to protect this flag either, as
578 * the work queue and irq_work logic will do the necessary
579 * synchronization for the wake ups. The only thing
580 * that is necessary is that the wake up happens after
581 * a task has been queued. It's OK for spurious wake ups.
583 work->waiters_pending = true;
585 if ((cpu == RING_BUFFER_ALL_CPUS && ring_buffer_empty(buffer)) ||
586 (cpu != RING_BUFFER_ALL_CPUS && ring_buffer_empty_cpu(buffer, cpu)))
589 finish_wait(&work->waiters, &wait);
593 * ring_buffer_poll_wait - poll on buffer input
594 * @buffer: buffer to wait on
595 * @cpu: the cpu buffer to wait on
596 * @filp: the file descriptor
597 * @poll_table: The poll descriptor
599 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
600 * as data is added to any of the @buffer's cpu buffers. Otherwise
601 * it will wait for data to be added to a specific cpu buffer.
603 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
606 int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
607 struct file *filp, poll_table *poll_table)
609 struct ring_buffer_per_cpu *cpu_buffer;
610 struct rb_irq_work *work;
612 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
613 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
614 return POLLIN | POLLRDNORM;
616 if (cpu == RING_BUFFER_ALL_CPUS)
617 work = &buffer->irq_work;
619 cpu_buffer = buffer->buffers[cpu];
620 work = &cpu_buffer->irq_work;
623 work->waiters_pending = true;
624 poll_wait(filp, &work->waiters, poll_table);
626 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
627 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
628 return POLLIN | POLLRDNORM;
632 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
633 #define RB_WARN_ON(b, cond) \
635 int _____ret = unlikely(cond); \
637 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
638 struct ring_buffer_per_cpu *__b = \
640 atomic_inc(&__b->buffer->record_disabled); \
642 atomic_inc(&b->record_disabled); \
648 /* Up this if you want to test the TIME_EXTENTS and normalization */
649 #define DEBUG_SHIFT 0
651 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
653 /* shift to debug/test normalization and TIME_EXTENTS */
654 return buffer->clock() << DEBUG_SHIFT;
657 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
661 preempt_disable_notrace();
662 time = rb_time_stamp(buffer);
663 preempt_enable_no_resched_notrace();
667 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
669 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
672 /* Just stupid testing the normalize function and deltas */
675 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
678 * Making the ring buffer lockless makes things tricky.
679 * Although writes only happen on the CPU that they are on,
680 * and they only need to worry about interrupts. Reads can
683 * The reader page is always off the ring buffer, but when the
684 * reader finishes with a page, it needs to swap its page with
685 * a new one from the buffer. The reader needs to take from
686 * the head (writes go to the tail). But if a writer is in overwrite
687 * mode and wraps, it must push the head page forward.
689 * Here lies the problem.
691 * The reader must be careful to replace only the head page, and
692 * not another one. As described at the top of the file in the
693 * ASCII art, the reader sets its old page to point to the next
694 * page after head. It then sets the page after head to point to
695 * the old reader page. But if the writer moves the head page
696 * during this operation, the reader could end up with the tail.
698 * We use cmpxchg to help prevent this race. We also do something
699 * special with the page before head. We set the LSB to 1.
701 * When the writer must push the page forward, it will clear the
702 * bit that points to the head page, move the head, and then set
703 * the bit that points to the new head page.
705 * We also don't want an interrupt coming in and moving the head
706 * page on another writer. Thus we use the second LSB to catch
709 * head->list->prev->next bit 1 bit 0
712 * Points to head page 0 1
715 * Note we can not trust the prev pointer of the head page, because:
717 * +----+ +-----+ +-----+
718 * | |------>| T |---X--->| N |
720 * +----+ +-----+ +-----+
723 * +----------| R |----------+ |
727 * Key: ---X--> HEAD flag set in pointer
732 * (see __rb_reserve_next() to see where this happens)
734 * What the above shows is that the reader just swapped out
735 * the reader page with a page in the buffer, but before it
736 * could make the new header point back to the new page added
737 * it was preempted by a writer. The writer moved forward onto
738 * the new page added by the reader and is about to move forward
741 * You can see, it is legitimate for the previous pointer of
742 * the head (or any page) not to point back to itself. But only
746 #define RB_PAGE_NORMAL 0UL
747 #define RB_PAGE_HEAD 1UL
748 #define RB_PAGE_UPDATE 2UL
751 #define RB_FLAG_MASK 3UL
753 /* PAGE_MOVED is not part of the mask */
754 #define RB_PAGE_MOVED 4UL
757 * rb_list_head - remove any bit
759 static struct list_head *rb_list_head(struct list_head *list)
761 unsigned long val = (unsigned long)list;
763 return (struct list_head *)(val & ~RB_FLAG_MASK);
767 * rb_is_head_page - test if the given page is the head page
769 * Because the reader may move the head_page pointer, we can
770 * not trust what the head page is (it may be pointing to
771 * the reader page). But if the next page is a header page,
772 * its flags will be non zero.
775 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
776 struct buffer_page *page, struct list_head *list)
780 val = (unsigned long)list->next;
782 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
783 return RB_PAGE_MOVED;
785 return val & RB_FLAG_MASK;
791 * The unique thing about the reader page, is that, if the
792 * writer is ever on it, the previous pointer never points
793 * back to the reader page.
795 static int rb_is_reader_page(struct buffer_page *page)
797 struct list_head *list = page->list.prev;
799 return rb_list_head(list->next) != &page->list;
803 * rb_set_list_to_head - set a list_head to be pointing to head.
805 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
806 struct list_head *list)
810 ptr = (unsigned long *)&list->next;
811 *ptr |= RB_PAGE_HEAD;
812 *ptr &= ~RB_PAGE_UPDATE;
816 * rb_head_page_activate - sets up head page
818 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
820 struct buffer_page *head;
822 head = cpu_buffer->head_page;
827 * Set the previous list pointer to have the HEAD flag.
829 rb_set_list_to_head(cpu_buffer, head->list.prev);
832 static void rb_list_head_clear(struct list_head *list)
834 unsigned long *ptr = (unsigned long *)&list->next;
836 *ptr &= ~RB_FLAG_MASK;
840 * rb_head_page_dactivate - clears head page ptr (for free list)
843 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
845 struct list_head *hd;
847 /* Go through the whole list and clear any pointers found. */
848 rb_list_head_clear(cpu_buffer->pages);
850 list_for_each(hd, cpu_buffer->pages)
851 rb_list_head_clear(hd);
854 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
855 struct buffer_page *head,
856 struct buffer_page *prev,
857 int old_flag, int new_flag)
859 struct list_head *list;
860 unsigned long val = (unsigned long)&head->list;
865 val &= ~RB_FLAG_MASK;
867 ret = cmpxchg((unsigned long *)&list->next,
868 val | old_flag, val | new_flag);
870 /* check if the reader took the page */
871 if ((ret & ~RB_FLAG_MASK) != val)
872 return RB_PAGE_MOVED;
874 return ret & RB_FLAG_MASK;
877 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
878 struct buffer_page *head,
879 struct buffer_page *prev,
882 return rb_head_page_set(cpu_buffer, head, prev,
883 old_flag, RB_PAGE_UPDATE);
886 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
887 struct buffer_page *head,
888 struct buffer_page *prev,
891 return rb_head_page_set(cpu_buffer, head, prev,
892 old_flag, RB_PAGE_HEAD);
895 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
896 struct buffer_page *head,
897 struct buffer_page *prev,
900 return rb_head_page_set(cpu_buffer, head, prev,
901 old_flag, RB_PAGE_NORMAL);
904 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
905 struct buffer_page **bpage)
907 struct list_head *p = rb_list_head((*bpage)->list.next);
909 *bpage = list_entry(p, struct buffer_page, list);
912 static struct buffer_page *
913 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
915 struct buffer_page *head;
916 struct buffer_page *page;
917 struct list_head *list;
920 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
924 list = cpu_buffer->pages;
925 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
928 page = head = cpu_buffer->head_page;
930 * It is possible that the writer moves the header behind
931 * where we started, and we miss in one loop.
932 * A second loop should grab the header, but we'll do
933 * three loops just because I'm paranoid.
935 for (i = 0; i < 3; i++) {
937 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
938 cpu_buffer->head_page = page;
941 rb_inc_page(cpu_buffer, &page);
942 } while (page != head);
945 RB_WARN_ON(cpu_buffer, 1);
950 static int rb_head_page_replace(struct buffer_page *old,
951 struct buffer_page *new)
953 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
957 val = *ptr & ~RB_FLAG_MASK;
960 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
966 * rb_tail_page_update - move the tail page forward
968 * Returns 1 if moved tail page, 0 if someone else did.
970 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
971 struct buffer_page *tail_page,
972 struct buffer_page *next_page)
974 struct buffer_page *old_tail;
975 unsigned long old_entries;
976 unsigned long old_write;
980 * The tail page now needs to be moved forward.
982 * We need to reset the tail page, but without messing
983 * with possible erasing of data brought in by interrupts
984 * that have moved the tail page and are currently on it.
986 * We add a counter to the write field to denote this.
988 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
989 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
992 * Just make sure we have seen our old_write and synchronize
993 * with any interrupts that come in.
998 * If the tail page is still the same as what we think
999 * it is, then it is up to us to update the tail
1002 if (tail_page == cpu_buffer->tail_page) {
1003 /* Zero the write counter */
1004 unsigned long val = old_write & ~RB_WRITE_MASK;
1005 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1008 * This will only succeed if an interrupt did
1009 * not come in and change it. In which case, we
1010 * do not want to modify it.
1012 * We add (void) to let the compiler know that we do not care
1013 * about the return value of these functions. We use the
1014 * cmpxchg to only update if an interrupt did not already
1015 * do it for us. If the cmpxchg fails, we don't care.
1017 (void)local_cmpxchg(&next_page->write, old_write, val);
1018 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1021 * No need to worry about races with clearing out the commit.
1022 * it only can increment when a commit takes place. But that
1023 * only happens in the outer most nested commit.
1025 local_set(&next_page->page->commit, 0);
1027 old_tail = cmpxchg(&cpu_buffer->tail_page,
1028 tail_page, next_page);
1030 if (old_tail == tail_page)
1037 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1038 struct buffer_page *bpage)
1040 unsigned long val = (unsigned long)bpage;
1042 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1049 * rb_check_list - make sure a pointer to a list has the last bits zero
1051 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1052 struct list_head *list)
1054 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1056 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1062 * check_pages - integrity check of buffer pages
1063 * @cpu_buffer: CPU buffer with pages to test
1065 * As a safety measure we check to make sure the data pages have not
1068 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1070 struct list_head *head = cpu_buffer->pages;
1071 struct buffer_page *bpage, *tmp;
1073 /* Reset the head page if it exists */
1074 if (cpu_buffer->head_page)
1075 rb_set_head_page(cpu_buffer);
1077 rb_head_page_deactivate(cpu_buffer);
1079 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1081 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1084 if (rb_check_list(cpu_buffer, head))
1087 list_for_each_entry_safe(bpage, tmp, head, list) {
1088 if (RB_WARN_ON(cpu_buffer,
1089 bpage->list.next->prev != &bpage->list))
1091 if (RB_WARN_ON(cpu_buffer,
1092 bpage->list.prev->next != &bpage->list))
1094 if (rb_check_list(cpu_buffer, &bpage->list))
1098 rb_head_page_activate(cpu_buffer);
1103 static int __rb_allocate_pages(int nr_pages, struct list_head *pages, int cpu)
1106 struct buffer_page *bpage, *tmp;
1108 for (i = 0; i < nr_pages; i++) {
1111 * __GFP_NORETRY flag makes sure that the allocation fails
1112 * gracefully without invoking oom-killer and the system is
1115 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1116 GFP_KERNEL | __GFP_NORETRY,
1121 list_add(&bpage->list, pages);
1123 page = alloc_pages_node(cpu_to_node(cpu),
1124 GFP_KERNEL | __GFP_NORETRY, 0);
1127 bpage->page = page_address(page);
1128 rb_init_page(bpage->page);
1134 list_for_each_entry_safe(bpage, tmp, pages, list) {
1135 list_del_init(&bpage->list);
1136 free_buffer_page(bpage);
1142 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1149 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1153 * The ring buffer page list is a circular list that does not
1154 * start and end with a list head. All page list items point to
1157 cpu_buffer->pages = pages.next;
1160 cpu_buffer->nr_pages = nr_pages;
1162 rb_check_pages(cpu_buffer);
1167 static struct ring_buffer_per_cpu *
1168 rb_allocate_cpu_buffer(struct ring_buffer *buffer, int nr_pages, int cpu)
1170 struct ring_buffer_per_cpu *cpu_buffer;
1171 struct buffer_page *bpage;
1175 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1176 GFP_KERNEL, cpu_to_node(cpu));
1180 cpu_buffer->cpu = cpu;
1181 cpu_buffer->buffer = buffer;
1182 raw_spin_lock_init(&cpu_buffer->reader_lock);
1183 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1184 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1185 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1186 init_completion(&cpu_buffer->update_done);
1187 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1188 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1190 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1191 GFP_KERNEL, cpu_to_node(cpu));
1193 goto fail_free_buffer;
1195 rb_check_bpage(cpu_buffer, bpage);
1197 cpu_buffer->reader_page = bpage;
1198 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1200 goto fail_free_reader;
1201 bpage->page = page_address(page);
1202 rb_init_page(bpage->page);
1204 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1205 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1207 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1209 goto fail_free_reader;
1211 cpu_buffer->head_page
1212 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1213 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1215 rb_head_page_activate(cpu_buffer);
1220 free_buffer_page(cpu_buffer->reader_page);
1227 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1229 struct list_head *head = cpu_buffer->pages;
1230 struct buffer_page *bpage, *tmp;
1232 free_buffer_page(cpu_buffer->reader_page);
1234 rb_head_page_deactivate(cpu_buffer);
1237 list_for_each_entry_safe(bpage, tmp, head, list) {
1238 list_del_init(&bpage->list);
1239 free_buffer_page(bpage);
1241 bpage = list_entry(head, struct buffer_page, list);
1242 free_buffer_page(bpage);
1248 #ifdef CONFIG_HOTPLUG_CPU
1249 static int rb_cpu_notify(struct notifier_block *self,
1250 unsigned long action, void *hcpu);
1254 * ring_buffer_alloc - allocate a new ring_buffer
1255 * @size: the size in bytes per cpu that is needed.
1256 * @flags: attributes to set for the ring buffer.
1258 * Currently the only flag that is available is the RB_FL_OVERWRITE
1259 * flag. This flag means that the buffer will overwrite old data
1260 * when the buffer wraps. If this flag is not set, the buffer will
1261 * drop data when the tail hits the head.
1263 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1264 struct lock_class_key *key)
1266 struct ring_buffer *buffer;
1270 /* keep it in its own cache line */
1271 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1276 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1277 goto fail_free_buffer;
1279 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1280 buffer->flags = flags;
1281 buffer->clock = trace_clock_local;
1282 buffer->reader_lock_key = key;
1284 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1285 init_waitqueue_head(&buffer->irq_work.waiters);
1287 /* need at least two pages */
1292 * In case of non-hotplug cpu, if the ring-buffer is allocated
1293 * in early initcall, it will not be notified of secondary cpus.
1294 * In that off case, we need to allocate for all possible cpus.
1296 #ifdef CONFIG_HOTPLUG_CPU
1298 cpumask_copy(buffer->cpumask, cpu_online_mask);
1300 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1302 buffer->cpus = nr_cpu_ids;
1304 bsize = sizeof(void *) * nr_cpu_ids;
1305 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1307 if (!buffer->buffers)
1308 goto fail_free_cpumask;
1310 for_each_buffer_cpu(buffer, cpu) {
1311 buffer->buffers[cpu] =
1312 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1313 if (!buffer->buffers[cpu])
1314 goto fail_free_buffers;
1317 #ifdef CONFIG_HOTPLUG_CPU
1318 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1319 buffer->cpu_notify.priority = 0;
1320 register_cpu_notifier(&buffer->cpu_notify);
1324 mutex_init(&buffer->mutex);
1329 for_each_buffer_cpu(buffer, cpu) {
1330 if (buffer->buffers[cpu])
1331 rb_free_cpu_buffer(buffer->buffers[cpu]);
1333 kfree(buffer->buffers);
1336 free_cpumask_var(buffer->cpumask);
1343 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1346 * ring_buffer_free - free a ring buffer.
1347 * @buffer: the buffer to free.
1350 ring_buffer_free(struct ring_buffer *buffer)
1356 #ifdef CONFIG_HOTPLUG_CPU
1357 unregister_cpu_notifier(&buffer->cpu_notify);
1360 for_each_buffer_cpu(buffer, cpu)
1361 rb_free_cpu_buffer(buffer->buffers[cpu]);
1365 kfree(buffer->buffers);
1366 free_cpumask_var(buffer->cpumask);
1370 EXPORT_SYMBOL_GPL(ring_buffer_free);
1372 void ring_buffer_set_clock(struct ring_buffer *buffer,
1375 buffer->clock = clock;
1378 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1380 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1382 return local_read(&bpage->entries) & RB_WRITE_MASK;
1385 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1387 return local_read(&bpage->write) & RB_WRITE_MASK;
1391 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned int nr_pages)
1393 struct list_head *tail_page, *to_remove, *next_page;
1394 struct buffer_page *to_remove_page, *tmp_iter_page;
1395 struct buffer_page *last_page, *first_page;
1396 unsigned int nr_removed;
1397 unsigned long head_bit;
1402 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1403 atomic_inc(&cpu_buffer->record_disabled);
1405 * We don't race with the readers since we have acquired the reader
1406 * lock. We also don't race with writers after disabling recording.
1407 * This makes it easy to figure out the first and the last page to be
1408 * removed from the list. We unlink all the pages in between including
1409 * the first and last pages. This is done in a busy loop so that we
1410 * lose the least number of traces.
1411 * The pages are freed after we restart recording and unlock readers.
1413 tail_page = &cpu_buffer->tail_page->list;
1416 * tail page might be on reader page, we remove the next page
1417 * from the ring buffer
1419 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1420 tail_page = rb_list_head(tail_page->next);
1421 to_remove = tail_page;
1423 /* start of pages to remove */
1424 first_page = list_entry(rb_list_head(to_remove->next),
1425 struct buffer_page, list);
1427 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1428 to_remove = rb_list_head(to_remove)->next;
1429 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1432 next_page = rb_list_head(to_remove)->next;
1435 * Now we remove all pages between tail_page and next_page.
1436 * Make sure that we have head_bit value preserved for the
1439 tail_page->next = (struct list_head *)((unsigned long)next_page |
1441 next_page = rb_list_head(next_page);
1442 next_page->prev = tail_page;
1444 /* make sure pages points to a valid page in the ring buffer */
1445 cpu_buffer->pages = next_page;
1447 /* update head page */
1449 cpu_buffer->head_page = list_entry(next_page,
1450 struct buffer_page, list);
1453 * change read pointer to make sure any read iterators reset
1456 cpu_buffer->read = 0;
1458 /* pages are removed, resume tracing and then free the pages */
1459 atomic_dec(&cpu_buffer->record_disabled);
1460 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1462 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1464 /* last buffer page to remove */
1465 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1467 tmp_iter_page = first_page;
1470 to_remove_page = tmp_iter_page;
1471 rb_inc_page(cpu_buffer, &tmp_iter_page);
1473 /* update the counters */
1474 page_entries = rb_page_entries(to_remove_page);
1477 * If something was added to this page, it was full
1478 * since it is not the tail page. So we deduct the
1479 * bytes consumed in ring buffer from here.
1480 * Increment overrun to account for the lost events.
1482 local_add(page_entries, &cpu_buffer->overrun);
1483 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1487 * We have already removed references to this list item, just
1488 * free up the buffer_page and its page
1490 free_buffer_page(to_remove_page);
1493 } while (to_remove_page != last_page);
1495 RB_WARN_ON(cpu_buffer, nr_removed);
1497 return nr_removed == 0;
1501 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1503 struct list_head *pages = &cpu_buffer->new_pages;
1504 int retries, success;
1506 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1508 * We are holding the reader lock, so the reader page won't be swapped
1509 * in the ring buffer. Now we are racing with the writer trying to
1510 * move head page and the tail page.
1511 * We are going to adapt the reader page update process where:
1512 * 1. We first splice the start and end of list of new pages between
1513 * the head page and its previous page.
1514 * 2. We cmpxchg the prev_page->next to point from head page to the
1515 * start of new pages list.
1516 * 3. Finally, we update the head->prev to the end of new list.
1518 * We will try this process 10 times, to make sure that we don't keep
1524 struct list_head *head_page, *prev_page, *r;
1525 struct list_head *last_page, *first_page;
1526 struct list_head *head_page_with_bit;
1528 head_page = &rb_set_head_page(cpu_buffer)->list;
1531 prev_page = head_page->prev;
1533 first_page = pages->next;
1534 last_page = pages->prev;
1536 head_page_with_bit = (struct list_head *)
1537 ((unsigned long)head_page | RB_PAGE_HEAD);
1539 last_page->next = head_page_with_bit;
1540 first_page->prev = prev_page;
1542 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1544 if (r == head_page_with_bit) {
1546 * yay, we replaced the page pointer to our new list,
1547 * now, we just have to update to head page's prev
1548 * pointer to point to end of list
1550 head_page->prev = last_page;
1557 INIT_LIST_HEAD(pages);
1559 * If we weren't successful in adding in new pages, warn and stop
1562 RB_WARN_ON(cpu_buffer, !success);
1563 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1565 /* free pages if they weren't inserted */
1567 struct buffer_page *bpage, *tmp;
1568 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1570 list_del_init(&bpage->list);
1571 free_buffer_page(bpage);
1577 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1581 if (cpu_buffer->nr_pages_to_update > 0)
1582 success = rb_insert_pages(cpu_buffer);
1584 success = rb_remove_pages(cpu_buffer,
1585 -cpu_buffer->nr_pages_to_update);
1588 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1591 static void update_pages_handler(struct work_struct *work)
1593 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1594 struct ring_buffer_per_cpu, update_pages_work);
1595 rb_update_pages(cpu_buffer);
1596 complete(&cpu_buffer->update_done);
1600 * ring_buffer_resize - resize the ring buffer
1601 * @buffer: the buffer to resize.
1602 * @size: the new size.
1604 * Minimum size is 2 * BUF_PAGE_SIZE.
1606 * Returns 0 on success and < 0 on failure.
1608 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1611 struct ring_buffer_per_cpu *cpu_buffer;
1616 * Always succeed at resizing a non-existent buffer:
1621 /* Make sure the requested buffer exists */
1622 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1623 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1626 size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1627 size *= BUF_PAGE_SIZE;
1629 /* we need a minimum of two pages */
1630 if (size < BUF_PAGE_SIZE * 2)
1631 size = BUF_PAGE_SIZE * 2;
1633 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1636 * Don't succeed if resizing is disabled, as a reader might be
1637 * manipulating the ring buffer and is expecting a sane state while
1640 if (atomic_read(&buffer->resize_disabled))
1643 /* prevent another thread from changing buffer sizes */
1644 mutex_lock(&buffer->mutex);
1646 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1647 /* calculate the pages to update */
1648 for_each_buffer_cpu(buffer, cpu) {
1649 cpu_buffer = buffer->buffers[cpu];
1651 cpu_buffer->nr_pages_to_update = nr_pages -
1652 cpu_buffer->nr_pages;
1654 * nothing more to do for removing pages or no update
1656 if (cpu_buffer->nr_pages_to_update <= 0)
1659 * to add pages, make sure all new pages can be
1660 * allocated without receiving ENOMEM
1662 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1663 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1664 &cpu_buffer->new_pages, cpu)) {
1665 /* not enough memory for new pages */
1673 * Fire off all the required work handlers
1674 * We can't schedule on offline CPUs, but it's not necessary
1675 * since we can change their buffer sizes without any race.
1677 for_each_buffer_cpu(buffer, cpu) {
1678 cpu_buffer = buffer->buffers[cpu];
1679 if (!cpu_buffer->nr_pages_to_update)
1682 /* The update must run on the CPU that is being updated. */
1684 if (cpu == smp_processor_id() || !cpu_online(cpu)) {
1685 rb_update_pages(cpu_buffer);
1686 cpu_buffer->nr_pages_to_update = 0;
1689 * Can not disable preemption for schedule_work_on()
1693 schedule_work_on(cpu,
1694 &cpu_buffer->update_pages_work);
1700 /* wait for all the updates to complete */
1701 for_each_buffer_cpu(buffer, cpu) {
1702 cpu_buffer = buffer->buffers[cpu];
1703 if (!cpu_buffer->nr_pages_to_update)
1706 if (cpu_online(cpu))
1707 wait_for_completion(&cpu_buffer->update_done);
1708 cpu_buffer->nr_pages_to_update = 0;
1713 /* Make sure this CPU has been intitialized */
1714 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1717 cpu_buffer = buffer->buffers[cpu_id];
1719 if (nr_pages == cpu_buffer->nr_pages)
1722 cpu_buffer->nr_pages_to_update = nr_pages -
1723 cpu_buffer->nr_pages;
1725 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1726 if (cpu_buffer->nr_pages_to_update > 0 &&
1727 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1728 &cpu_buffer->new_pages, cpu_id)) {
1736 /* The update must run on the CPU that is being updated. */
1737 if (cpu_id == smp_processor_id() || !cpu_online(cpu_id))
1738 rb_update_pages(cpu_buffer);
1741 * Can not disable preemption for schedule_work_on()
1745 schedule_work_on(cpu_id,
1746 &cpu_buffer->update_pages_work);
1747 wait_for_completion(&cpu_buffer->update_done);
1752 cpu_buffer->nr_pages_to_update = 0;
1758 * The ring buffer resize can happen with the ring buffer
1759 * enabled, so that the update disturbs the tracing as little
1760 * as possible. But if the buffer is disabled, we do not need
1761 * to worry about that, and we can take the time to verify
1762 * that the buffer is not corrupt.
1764 if (atomic_read(&buffer->record_disabled)) {
1765 atomic_inc(&buffer->record_disabled);
1767 * Even though the buffer was disabled, we must make sure
1768 * that it is truly disabled before calling rb_check_pages.
1769 * There could have been a race between checking
1770 * record_disable and incrementing it.
1772 synchronize_sched();
1773 for_each_buffer_cpu(buffer, cpu) {
1774 cpu_buffer = buffer->buffers[cpu];
1775 rb_check_pages(cpu_buffer);
1777 atomic_dec(&buffer->record_disabled);
1780 mutex_unlock(&buffer->mutex);
1784 for_each_buffer_cpu(buffer, cpu) {
1785 struct buffer_page *bpage, *tmp;
1787 cpu_buffer = buffer->buffers[cpu];
1788 cpu_buffer->nr_pages_to_update = 0;
1790 if (list_empty(&cpu_buffer->new_pages))
1793 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1795 list_del_init(&bpage->list);
1796 free_buffer_page(bpage);
1799 mutex_unlock(&buffer->mutex);
1802 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1804 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1806 mutex_lock(&buffer->mutex);
1808 buffer->flags |= RB_FL_OVERWRITE;
1810 buffer->flags &= ~RB_FL_OVERWRITE;
1811 mutex_unlock(&buffer->mutex);
1813 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1815 static inline void *
1816 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1818 return bpage->data + index;
1821 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1823 return bpage->page->data + index;
1826 static inline struct ring_buffer_event *
1827 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1829 return __rb_page_index(cpu_buffer->reader_page,
1830 cpu_buffer->reader_page->read);
1833 static inline struct ring_buffer_event *
1834 rb_iter_head_event(struct ring_buffer_iter *iter)
1836 return __rb_page_index(iter->head_page, iter->head);
1839 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1841 return local_read(&bpage->page->commit);
1844 /* Size is determined by what has been committed */
1845 static inline unsigned rb_page_size(struct buffer_page *bpage)
1847 return rb_page_commit(bpage);
1850 static inline unsigned
1851 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1853 return rb_page_commit(cpu_buffer->commit_page);
1856 static inline unsigned
1857 rb_event_index(struct ring_buffer_event *event)
1859 unsigned long addr = (unsigned long)event;
1861 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1865 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1866 struct ring_buffer_event *event)
1868 unsigned long addr = (unsigned long)event;
1869 unsigned long index;
1871 index = rb_event_index(event);
1874 return cpu_buffer->commit_page->page == (void *)addr &&
1875 rb_commit_index(cpu_buffer) == index;
1879 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1881 unsigned long max_count;
1884 * We only race with interrupts and NMIs on this CPU.
1885 * If we own the commit event, then we can commit
1886 * all others that interrupted us, since the interruptions
1887 * are in stack format (they finish before they come
1888 * back to us). This allows us to do a simple loop to
1889 * assign the commit to the tail.
1892 max_count = cpu_buffer->nr_pages * 100;
1894 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1895 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1897 if (RB_WARN_ON(cpu_buffer,
1898 rb_is_reader_page(cpu_buffer->tail_page)))
1900 local_set(&cpu_buffer->commit_page->page->commit,
1901 rb_page_write(cpu_buffer->commit_page));
1902 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1903 cpu_buffer->write_stamp =
1904 cpu_buffer->commit_page->page->time_stamp;
1905 /* add barrier to keep gcc from optimizing too much */
1908 while (rb_commit_index(cpu_buffer) !=
1909 rb_page_write(cpu_buffer->commit_page)) {
1911 local_set(&cpu_buffer->commit_page->page->commit,
1912 rb_page_write(cpu_buffer->commit_page));
1913 RB_WARN_ON(cpu_buffer,
1914 local_read(&cpu_buffer->commit_page->page->commit) &
1919 /* again, keep gcc from optimizing */
1923 * If an interrupt came in just after the first while loop
1924 * and pushed the tail page forward, we will be left with
1925 * a dangling commit that will never go forward.
1927 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1931 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1933 cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1934 cpu_buffer->reader_page->read = 0;
1937 static void rb_inc_iter(struct ring_buffer_iter *iter)
1939 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1942 * The iterator could be on the reader page (it starts there).
1943 * But the head could have moved, since the reader was
1944 * found. Check for this case and assign the iterator
1945 * to the head page instead of next.
1947 if (iter->head_page == cpu_buffer->reader_page)
1948 iter->head_page = rb_set_head_page(cpu_buffer);
1950 rb_inc_page(cpu_buffer, &iter->head_page);
1952 iter->read_stamp = iter->head_page->page->time_stamp;
1956 /* Slow path, do not inline */
1957 static noinline struct ring_buffer_event *
1958 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
1960 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
1962 /* Not the first event on the page? */
1963 if (rb_event_index(event)) {
1964 event->time_delta = delta & TS_MASK;
1965 event->array[0] = delta >> TS_SHIFT;
1967 /* nope, just zero it */
1968 event->time_delta = 0;
1969 event->array[0] = 0;
1972 return skip_time_extend(event);
1976 * rb_update_event - update event type and data
1977 * @event: the even to update
1978 * @type: the type of event
1979 * @length: the size of the event field in the ring buffer
1981 * Update the type and data fields of the event. The length
1982 * is the actual size that is written to the ring buffer,
1983 * and with this, we can determine what to place into the
1987 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
1988 struct ring_buffer_event *event, unsigned length,
1989 int add_timestamp, u64 delta)
1991 /* Only a commit updates the timestamp */
1992 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
1996 * If we need to add a timestamp, then we
1997 * add it to the start of the resevered space.
1999 if (unlikely(add_timestamp)) {
2000 event = rb_add_time_stamp(event, delta);
2001 length -= RB_LEN_TIME_EXTEND;
2005 event->time_delta = delta;
2006 length -= RB_EVNT_HDR_SIZE;
2007 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2008 event->type_len = 0;
2009 event->array[0] = length;
2011 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2015 * rb_handle_head_page - writer hit the head page
2017 * Returns: +1 to retry page
2022 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2023 struct buffer_page *tail_page,
2024 struct buffer_page *next_page)
2026 struct buffer_page *new_head;
2031 entries = rb_page_entries(next_page);
2034 * The hard part is here. We need to move the head
2035 * forward, and protect against both readers on
2036 * other CPUs and writers coming in via interrupts.
2038 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2042 * type can be one of four:
2043 * NORMAL - an interrupt already moved it for us
2044 * HEAD - we are the first to get here.
2045 * UPDATE - we are the interrupt interrupting
2047 * MOVED - a reader on another CPU moved the next
2048 * pointer to its reader page. Give up
2055 * We changed the head to UPDATE, thus
2056 * it is our responsibility to update
2059 local_add(entries, &cpu_buffer->overrun);
2060 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2063 * The entries will be zeroed out when we move the
2067 /* still more to do */
2070 case RB_PAGE_UPDATE:
2072 * This is an interrupt that interrupt the
2073 * previous update. Still more to do.
2076 case RB_PAGE_NORMAL:
2078 * An interrupt came in before the update
2079 * and processed this for us.
2080 * Nothing left to do.
2085 * The reader is on another CPU and just did
2086 * a swap with our next_page.
2091 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2096 * Now that we are here, the old head pointer is
2097 * set to UPDATE. This will keep the reader from
2098 * swapping the head page with the reader page.
2099 * The reader (on another CPU) will spin till
2102 * We just need to protect against interrupts
2103 * doing the job. We will set the next pointer
2104 * to HEAD. After that, we set the old pointer
2105 * to NORMAL, but only if it was HEAD before.
2106 * otherwise we are an interrupt, and only
2107 * want the outer most commit to reset it.
2109 new_head = next_page;
2110 rb_inc_page(cpu_buffer, &new_head);
2112 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2116 * Valid returns are:
2117 * HEAD - an interrupt came in and already set it.
2118 * NORMAL - One of two things:
2119 * 1) We really set it.
2120 * 2) A bunch of interrupts came in and moved
2121 * the page forward again.
2125 case RB_PAGE_NORMAL:
2129 RB_WARN_ON(cpu_buffer, 1);
2134 * It is possible that an interrupt came in,
2135 * set the head up, then more interrupts came in
2136 * and moved it again. When we get back here,
2137 * the page would have been set to NORMAL but we
2138 * just set it back to HEAD.
2140 * How do you detect this? Well, if that happened
2141 * the tail page would have moved.
2143 if (ret == RB_PAGE_NORMAL) {
2145 * If the tail had moved passed next, then we need
2146 * to reset the pointer.
2148 if (cpu_buffer->tail_page != tail_page &&
2149 cpu_buffer->tail_page != next_page)
2150 rb_head_page_set_normal(cpu_buffer, new_head,
2156 * If this was the outer most commit (the one that
2157 * changed the original pointer from HEAD to UPDATE),
2158 * then it is up to us to reset it to NORMAL.
2160 if (type == RB_PAGE_HEAD) {
2161 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2164 if (RB_WARN_ON(cpu_buffer,
2165 ret != RB_PAGE_UPDATE))
2172 static unsigned rb_calculate_event_length(unsigned length)
2174 struct ring_buffer_event event; /* Used only for sizeof array */
2176 /* zero length can cause confusions */
2180 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2181 length += sizeof(event.array[0]);
2183 length += RB_EVNT_HDR_SIZE;
2184 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2190 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2191 struct buffer_page *tail_page,
2192 unsigned long tail, unsigned long length)
2194 struct ring_buffer_event *event;
2197 * Only the event that crossed the page boundary
2198 * must fill the old tail_page with padding.
2200 if (tail >= BUF_PAGE_SIZE) {
2202 * If the page was filled, then we still need
2203 * to update the real_end. Reset it to zero
2204 * and the reader will ignore it.
2206 if (tail == BUF_PAGE_SIZE)
2207 tail_page->real_end = 0;
2209 local_sub(length, &tail_page->write);
2213 event = __rb_page_index(tail_page, tail);
2214 kmemcheck_annotate_bitfield(event, bitfield);
2216 /* account for padding bytes */
2217 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2220 * Save the original length to the meta data.
2221 * This will be used by the reader to add lost event
2224 tail_page->real_end = tail;
2227 * If this event is bigger than the minimum size, then
2228 * we need to be careful that we don't subtract the
2229 * write counter enough to allow another writer to slip
2231 * We put in a discarded commit instead, to make sure
2232 * that this space is not used again.
2234 * If we are less than the minimum size, we don't need to
2237 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2238 /* No room for any events */
2240 /* Mark the rest of the page with padding */
2241 rb_event_set_padding(event);
2243 /* Set the write back to the previous setting */
2244 local_sub(length, &tail_page->write);
2248 /* Put in a discarded event */
2249 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2250 event->type_len = RINGBUF_TYPE_PADDING;
2251 /* time delta must be non zero */
2252 event->time_delta = 1;
2254 /* Set write to end of buffer */
2255 length = (tail + length) - BUF_PAGE_SIZE;
2256 local_sub(length, &tail_page->write);
2260 * This is the slow path, force gcc not to inline it.
2262 static noinline struct ring_buffer_event *
2263 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2264 unsigned long length, unsigned long tail,
2265 struct buffer_page *tail_page, u64 ts)
2267 struct buffer_page *commit_page = cpu_buffer->commit_page;
2268 struct ring_buffer *buffer = cpu_buffer->buffer;
2269 struct buffer_page *next_page;
2272 next_page = tail_page;
2274 rb_inc_page(cpu_buffer, &next_page);
2277 * If for some reason, we had an interrupt storm that made
2278 * it all the way around the buffer, bail, and warn
2281 if (unlikely(next_page == commit_page)) {
2282 local_inc(&cpu_buffer->commit_overrun);
2287 * This is where the fun begins!
2289 * We are fighting against races between a reader that
2290 * could be on another CPU trying to swap its reader
2291 * page with the buffer head.
2293 * We are also fighting against interrupts coming in and
2294 * moving the head or tail on us as well.
2296 * If the next page is the head page then we have filled
2297 * the buffer, unless the commit page is still on the
2300 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2303 * If the commit is not on the reader page, then
2304 * move the header page.
2306 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2308 * If we are not in overwrite mode,
2309 * this is easy, just stop here.
2311 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2312 local_inc(&cpu_buffer->dropped_events);
2316 ret = rb_handle_head_page(cpu_buffer,
2325 * We need to be careful here too. The
2326 * commit page could still be on the reader
2327 * page. We could have a small buffer, and
2328 * have filled up the buffer with events
2329 * from interrupts and such, and wrapped.
2331 * Note, if the tail page is also the on the
2332 * reader_page, we let it move out.
2334 if (unlikely((cpu_buffer->commit_page !=
2335 cpu_buffer->tail_page) &&
2336 (cpu_buffer->commit_page ==
2337 cpu_buffer->reader_page))) {
2338 local_inc(&cpu_buffer->commit_overrun);
2344 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
2347 * Nested commits always have zero deltas, so
2348 * just reread the time stamp
2350 ts = rb_time_stamp(buffer);
2351 next_page->page->time_stamp = ts;
2356 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2358 /* fail and let the caller try again */
2359 return ERR_PTR(-EAGAIN);
2363 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2368 static struct ring_buffer_event *
2369 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2370 unsigned long length, u64 ts,
2371 u64 delta, int add_timestamp)
2373 struct buffer_page *tail_page;
2374 struct ring_buffer_event *event;
2375 unsigned long tail, write;
2378 * If the time delta since the last event is too big to
2379 * hold in the time field of the event, then we append a
2380 * TIME EXTEND event ahead of the data event.
2382 if (unlikely(add_timestamp))
2383 length += RB_LEN_TIME_EXTEND;
2385 tail_page = cpu_buffer->tail_page;
2386 write = local_add_return(length, &tail_page->write);
2388 /* set write to only the index of the write */
2389 write &= RB_WRITE_MASK;
2390 tail = write - length;
2392 /* See if we shot pass the end of this buffer page */
2393 if (unlikely(write > BUF_PAGE_SIZE))
2394 return rb_move_tail(cpu_buffer, length, tail,
2397 /* We reserved something on the buffer */
2399 event = __rb_page_index(tail_page, tail);
2400 kmemcheck_annotate_bitfield(event, bitfield);
2401 rb_update_event(cpu_buffer, event, length, add_timestamp, delta);
2403 local_inc(&tail_page->entries);
2406 * If this is the first commit on the page, then update
2410 tail_page->page->time_stamp = ts;
2412 /* account for these added bytes */
2413 local_add(length, &cpu_buffer->entries_bytes);
2419 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2420 struct ring_buffer_event *event)
2422 unsigned long new_index, old_index;
2423 struct buffer_page *bpage;
2424 unsigned long index;
2427 new_index = rb_event_index(event);
2428 old_index = new_index + rb_event_ts_length(event);
2429 addr = (unsigned long)event;
2432 bpage = cpu_buffer->tail_page;
2434 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2435 unsigned long write_mask =
2436 local_read(&bpage->write) & ~RB_WRITE_MASK;
2437 unsigned long event_length = rb_event_length(event);
2439 * This is on the tail page. It is possible that
2440 * a write could come in and move the tail page
2441 * and write to the next page. That is fine
2442 * because we just shorten what is on this page.
2444 old_index += write_mask;
2445 new_index += write_mask;
2446 index = local_cmpxchg(&bpage->write, old_index, new_index);
2447 if (index == old_index) {
2448 /* update counters */
2449 local_sub(event_length, &cpu_buffer->entries_bytes);
2454 /* could not discard */
2458 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2460 local_inc(&cpu_buffer->committing);
2461 local_inc(&cpu_buffer->commits);
2464 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2466 unsigned long commits;
2468 if (RB_WARN_ON(cpu_buffer,
2469 !local_read(&cpu_buffer->committing)))
2473 commits = local_read(&cpu_buffer->commits);
2474 /* synchronize with interrupts */
2476 if (local_read(&cpu_buffer->committing) == 1)
2477 rb_set_commit_to_write(cpu_buffer);
2479 local_dec(&cpu_buffer->committing);
2481 /* synchronize with interrupts */
2485 * Need to account for interrupts coming in between the
2486 * updating of the commit page and the clearing of the
2487 * committing counter.
2489 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2490 !local_read(&cpu_buffer->committing)) {
2491 local_inc(&cpu_buffer->committing);
2496 static struct ring_buffer_event *
2497 rb_reserve_next_event(struct ring_buffer *buffer,
2498 struct ring_buffer_per_cpu *cpu_buffer,
2499 unsigned long length)
2501 struct ring_buffer_event *event;
2507 rb_start_commit(cpu_buffer);
2509 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2511 * Due to the ability to swap a cpu buffer from a buffer
2512 * it is possible it was swapped before we committed.
2513 * (committing stops a swap). We check for it here and
2514 * if it happened, we have to fail the write.
2517 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2518 local_dec(&cpu_buffer->committing);
2519 local_dec(&cpu_buffer->commits);
2524 length = rb_calculate_event_length(length);
2530 * We allow for interrupts to reenter here and do a trace.
2531 * If one does, it will cause this original code to loop
2532 * back here. Even with heavy interrupts happening, this
2533 * should only happen a few times in a row. If this happens
2534 * 1000 times in a row, there must be either an interrupt
2535 * storm or we have something buggy.
2538 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2541 ts = rb_time_stamp(cpu_buffer->buffer);
2542 diff = ts - cpu_buffer->write_stamp;
2544 /* make sure this diff is calculated here */
2547 /* Did the write stamp get updated already? */
2548 if (likely(ts >= cpu_buffer->write_stamp)) {
2550 if (unlikely(test_time_stamp(delta))) {
2551 int local_clock_stable = 1;
2552 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2553 local_clock_stable = sched_clock_stable;
2555 WARN_ONCE(delta > (1ULL << 59),
2556 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2557 (unsigned long long)delta,
2558 (unsigned long long)ts,
2559 (unsigned long long)cpu_buffer->write_stamp,
2560 local_clock_stable ? "" :
2561 "If you just came from a suspend/resume,\n"
2562 "please switch to the trace global clock:\n"
2563 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2568 event = __rb_reserve_next(cpu_buffer, length, ts,
2569 delta, add_timestamp);
2570 if (unlikely(PTR_ERR(event) == -EAGAIN))
2579 rb_end_commit(cpu_buffer);
2583 #ifdef CONFIG_TRACING
2586 * The lock and unlock are done within a preempt disable section.
2587 * The current_context per_cpu variable can only be modified
2588 * by the current task between lock and unlock. But it can
2589 * be modified more than once via an interrupt. To pass this
2590 * information from the lock to the unlock without having to
2591 * access the 'in_interrupt()' functions again (which do show
2592 * a bit of overhead in something as critical as function tracing,
2593 * we use a bitmask trick.
2595 * bit 0 = NMI context
2596 * bit 1 = IRQ context
2597 * bit 2 = SoftIRQ context
2598 * bit 3 = normal context.
2600 * This works because this is the order of contexts that can
2601 * preempt other contexts. A SoftIRQ never preempts an IRQ
2604 * When the context is determined, the corresponding bit is
2605 * checked and set (if it was set, then a recursion of that context
2608 * On unlock, we need to clear this bit. To do so, just subtract
2609 * 1 from the current_context and AND it to itself.
2613 * 101 & 100 = 100 (clearing bit zero)
2616 * 1010 & 1001 = 1000 (clearing bit 1)
2618 * The least significant bit can be cleared this way, and it
2619 * just so happens that it is the same bit corresponding to
2620 * the current context.
2622 static DEFINE_PER_CPU(unsigned int, current_context);
2624 static __always_inline int trace_recursive_lock(void)
2626 unsigned int val = this_cpu_read(current_context);
2629 if (in_interrupt()) {
2639 if (unlikely(val & (1 << bit)))
2643 this_cpu_write(current_context, val);
2648 static __always_inline void trace_recursive_unlock(void)
2650 unsigned int val = this_cpu_read(current_context);
2653 val &= this_cpu_read(current_context);
2654 this_cpu_write(current_context, val);
2659 #define trace_recursive_lock() (0)
2660 #define trace_recursive_unlock() do { } while (0)
2665 * ring_buffer_lock_reserve - reserve a part of the buffer
2666 * @buffer: the ring buffer to reserve from
2667 * @length: the length of the data to reserve (excluding event header)
2669 * Returns a reseverd event on the ring buffer to copy directly to.
2670 * The user of this interface will need to get the body to write into
2671 * and can use the ring_buffer_event_data() interface.
2673 * The length is the length of the data needed, not the event length
2674 * which also includes the event header.
2676 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2677 * If NULL is returned, then nothing has been allocated or locked.
2679 struct ring_buffer_event *
2680 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2682 struct ring_buffer_per_cpu *cpu_buffer;
2683 struct ring_buffer_event *event;
2686 if (ring_buffer_flags != RB_BUFFERS_ON)
2689 /* If we are tracing schedule, we don't want to recurse */
2690 preempt_disable_notrace();
2692 if (atomic_read(&buffer->record_disabled))
2695 if (trace_recursive_lock())
2698 cpu = raw_smp_processor_id();
2700 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2703 cpu_buffer = buffer->buffers[cpu];
2705 if (atomic_read(&cpu_buffer->record_disabled))
2708 if (length > BUF_MAX_DATA_SIZE)
2711 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2718 trace_recursive_unlock();
2721 preempt_enable_notrace();
2724 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2727 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2728 struct ring_buffer_event *event)
2733 * The event first in the commit queue updates the
2736 if (rb_event_is_commit(cpu_buffer, event)) {
2738 * A commit event that is first on a page
2739 * updates the write timestamp with the page stamp
2741 if (!rb_event_index(event))
2742 cpu_buffer->write_stamp =
2743 cpu_buffer->commit_page->page->time_stamp;
2744 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2745 delta = event->array[0];
2747 delta += event->time_delta;
2748 cpu_buffer->write_stamp += delta;
2750 cpu_buffer->write_stamp += event->time_delta;
2754 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2755 struct ring_buffer_event *event)
2757 local_inc(&cpu_buffer->entries);
2758 rb_update_write_stamp(cpu_buffer, event);
2759 rb_end_commit(cpu_buffer);
2762 static __always_inline void
2763 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2765 if (buffer->irq_work.waiters_pending) {
2766 buffer->irq_work.waiters_pending = false;
2767 /* irq_work_queue() supplies it's own memory barriers */
2768 irq_work_queue(&buffer->irq_work.work);
2771 if (cpu_buffer->irq_work.waiters_pending) {
2772 cpu_buffer->irq_work.waiters_pending = false;
2773 /* irq_work_queue() supplies it's own memory barriers */
2774 irq_work_queue(&cpu_buffer->irq_work.work);
2779 * ring_buffer_unlock_commit - commit a reserved
2780 * @buffer: The buffer to commit to
2781 * @event: The event pointer to commit.
2783 * This commits the data to the ring buffer, and releases any locks held.
2785 * Must be paired with ring_buffer_lock_reserve.
2787 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2788 struct ring_buffer_event *event)
2790 struct ring_buffer_per_cpu *cpu_buffer;
2791 int cpu = raw_smp_processor_id();
2793 cpu_buffer = buffer->buffers[cpu];
2795 rb_commit(cpu_buffer, event);
2797 rb_wakeups(buffer, cpu_buffer);
2799 trace_recursive_unlock();
2801 preempt_enable_notrace();
2805 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2807 static inline void rb_event_discard(struct ring_buffer_event *event)
2809 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2810 event = skip_time_extend(event);
2812 /* array[0] holds the actual length for the discarded event */
2813 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2814 event->type_len = RINGBUF_TYPE_PADDING;
2815 /* time delta must be non zero */
2816 if (!event->time_delta)
2817 event->time_delta = 1;
2821 * Decrement the entries to the page that an event is on.
2822 * The event does not even need to exist, only the pointer
2823 * to the page it is on. This may only be called before the commit
2827 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2828 struct ring_buffer_event *event)
2830 unsigned long addr = (unsigned long)event;
2831 struct buffer_page *bpage = cpu_buffer->commit_page;
2832 struct buffer_page *start;
2836 /* Do the likely case first */
2837 if (likely(bpage->page == (void *)addr)) {
2838 local_dec(&bpage->entries);
2843 * Because the commit page may be on the reader page we
2844 * start with the next page and check the end loop there.
2846 rb_inc_page(cpu_buffer, &bpage);
2849 if (bpage->page == (void *)addr) {
2850 local_dec(&bpage->entries);
2853 rb_inc_page(cpu_buffer, &bpage);
2854 } while (bpage != start);
2856 /* commit not part of this buffer?? */
2857 RB_WARN_ON(cpu_buffer, 1);
2861 * ring_buffer_commit_discard - discard an event that has not been committed
2862 * @buffer: the ring buffer
2863 * @event: non committed event to discard
2865 * Sometimes an event that is in the ring buffer needs to be ignored.
2866 * This function lets the user discard an event in the ring buffer
2867 * and then that event will not be read later.
2869 * This function only works if it is called before the the item has been
2870 * committed. It will try to free the event from the ring buffer
2871 * if another event has not been added behind it.
2873 * If another event has been added behind it, it will set the event
2874 * up as discarded, and perform the commit.
2876 * If this function is called, do not call ring_buffer_unlock_commit on
2879 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2880 struct ring_buffer_event *event)
2882 struct ring_buffer_per_cpu *cpu_buffer;
2885 /* The event is discarded regardless */
2886 rb_event_discard(event);
2888 cpu = smp_processor_id();
2889 cpu_buffer = buffer->buffers[cpu];
2892 * This must only be called if the event has not been
2893 * committed yet. Thus we can assume that preemption
2894 * is still disabled.
2896 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2898 rb_decrement_entry(cpu_buffer, event);
2899 if (rb_try_to_discard(cpu_buffer, event))
2903 * The commit is still visible by the reader, so we
2904 * must still update the timestamp.
2906 rb_update_write_stamp(cpu_buffer, event);
2908 rb_end_commit(cpu_buffer);
2910 trace_recursive_unlock();
2912 preempt_enable_notrace();
2915 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2918 * ring_buffer_write - write data to the buffer without reserving
2919 * @buffer: The ring buffer to write to.
2920 * @length: The length of the data being written (excluding the event header)
2921 * @data: The data to write to the buffer.
2923 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2924 * one function. If you already have the data to write to the buffer, it
2925 * may be easier to simply call this function.
2927 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2928 * and not the length of the event which would hold the header.
2930 int ring_buffer_write(struct ring_buffer *buffer,
2931 unsigned long length,
2934 struct ring_buffer_per_cpu *cpu_buffer;
2935 struct ring_buffer_event *event;
2940 if (ring_buffer_flags != RB_BUFFERS_ON)
2943 preempt_disable_notrace();
2945 if (atomic_read(&buffer->record_disabled))
2948 cpu = raw_smp_processor_id();
2950 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2953 cpu_buffer = buffer->buffers[cpu];
2955 if (atomic_read(&cpu_buffer->record_disabled))
2958 if (length > BUF_MAX_DATA_SIZE)
2961 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2965 body = rb_event_data(event);
2967 memcpy(body, data, length);
2969 rb_commit(cpu_buffer, event);
2971 rb_wakeups(buffer, cpu_buffer);
2975 preempt_enable_notrace();
2979 EXPORT_SYMBOL_GPL(ring_buffer_write);
2981 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
2983 struct buffer_page *reader = cpu_buffer->reader_page;
2984 struct buffer_page *head = rb_set_head_page(cpu_buffer);
2985 struct buffer_page *commit = cpu_buffer->commit_page;
2987 /* In case of error, head will be NULL */
2988 if (unlikely(!head))
2991 return reader->read == rb_page_commit(reader) &&
2992 (commit == reader ||
2994 head->read == rb_page_commit(commit)));
2998 * ring_buffer_record_disable - stop all writes into the buffer
2999 * @buffer: The ring buffer to stop writes to.
3001 * This prevents all writes to the buffer. Any attempt to write
3002 * to the buffer after this will fail and return NULL.
3004 * The caller should call synchronize_sched() after this.
3006 void ring_buffer_record_disable(struct ring_buffer *buffer)
3008 atomic_inc(&buffer->record_disabled);
3010 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3013 * ring_buffer_record_enable - enable writes to the buffer
3014 * @buffer: The ring buffer to enable writes
3016 * Note, multiple disables will need the same number of enables
3017 * to truly enable the writing (much like preempt_disable).
3019 void ring_buffer_record_enable(struct ring_buffer *buffer)
3021 atomic_dec(&buffer->record_disabled);
3023 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3026 * ring_buffer_record_off - stop all writes into the buffer
3027 * @buffer: The ring buffer to stop writes to.
3029 * This prevents all writes to the buffer. Any attempt to write
3030 * to the buffer after this will fail and return NULL.
3032 * This is different than ring_buffer_record_disable() as
3033 * it works like an on/off switch, where as the disable() version
3034 * must be paired with a enable().
3036 void ring_buffer_record_off(struct ring_buffer *buffer)
3039 unsigned int new_rd;
3042 rd = atomic_read(&buffer->record_disabled);
3043 new_rd = rd | RB_BUFFER_OFF;
3044 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3046 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3049 * ring_buffer_record_on - restart writes into the buffer
3050 * @buffer: The ring buffer to start writes to.
3052 * This enables all writes to the buffer that was disabled by
3053 * ring_buffer_record_off().
3055 * This is different than ring_buffer_record_enable() as
3056 * it works like an on/off switch, where as the enable() version
3057 * must be paired with a disable().
3059 void ring_buffer_record_on(struct ring_buffer *buffer)
3062 unsigned int new_rd;
3065 rd = atomic_read(&buffer->record_disabled);
3066 new_rd = rd & ~RB_BUFFER_OFF;
3067 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3069 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3072 * ring_buffer_record_is_on - return true if the ring buffer can write
3073 * @buffer: The ring buffer to see if write is enabled
3075 * Returns true if the ring buffer is in a state that it accepts writes.
3077 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3079 return !atomic_read(&buffer->record_disabled);
3083 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3084 * @buffer: The ring buffer to stop writes to.
3085 * @cpu: The CPU buffer to stop
3087 * This prevents all writes to the buffer. Any attempt to write
3088 * to the buffer after this will fail and return NULL.
3090 * The caller should call synchronize_sched() after this.
3092 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3094 struct ring_buffer_per_cpu *cpu_buffer;
3096 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3099 cpu_buffer = buffer->buffers[cpu];
3100 atomic_inc(&cpu_buffer->record_disabled);
3102 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3105 * ring_buffer_record_enable_cpu - enable writes to the buffer
3106 * @buffer: The ring buffer to enable writes
3107 * @cpu: The CPU to enable.
3109 * Note, multiple disables will need the same number of enables
3110 * to truly enable the writing (much like preempt_disable).
3112 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3114 struct ring_buffer_per_cpu *cpu_buffer;
3116 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3119 cpu_buffer = buffer->buffers[cpu];
3120 atomic_dec(&cpu_buffer->record_disabled);
3122 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3125 * The total entries in the ring buffer is the running counter
3126 * of entries entered into the ring buffer, minus the sum of
3127 * the entries read from the ring buffer and the number of
3128 * entries that were overwritten.
3130 static inline unsigned long
3131 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3133 return local_read(&cpu_buffer->entries) -
3134 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3138 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3139 * @buffer: The ring buffer
3140 * @cpu: The per CPU buffer to read from.
3142 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3144 unsigned long flags;
3145 struct ring_buffer_per_cpu *cpu_buffer;
3146 struct buffer_page *bpage;
3149 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3152 cpu_buffer = buffer->buffers[cpu];
3153 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3155 * if the tail is on reader_page, oldest time stamp is on the reader
3158 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3159 bpage = cpu_buffer->reader_page;
3161 bpage = rb_set_head_page(cpu_buffer);
3163 ret = bpage->page->time_stamp;
3164 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3168 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3171 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3172 * @buffer: The ring buffer
3173 * @cpu: The per CPU buffer to read from.
3175 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3177 struct ring_buffer_per_cpu *cpu_buffer;
3180 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3183 cpu_buffer = buffer->buffers[cpu];
3184 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3188 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3191 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3192 * @buffer: The ring buffer
3193 * @cpu: The per CPU buffer to get the entries from.
3195 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3197 struct ring_buffer_per_cpu *cpu_buffer;
3199 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3202 cpu_buffer = buffer->buffers[cpu];
3204 return rb_num_of_entries(cpu_buffer);
3206 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3209 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3210 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3211 * @buffer: The ring buffer
3212 * @cpu: The per CPU buffer to get the number of overruns from
3214 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3216 struct ring_buffer_per_cpu *cpu_buffer;
3219 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3222 cpu_buffer = buffer->buffers[cpu];
3223 ret = local_read(&cpu_buffer->overrun);
3227 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3230 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3231 * commits failing due to the buffer wrapping around while there are uncommitted
3232 * events, such as during an interrupt storm.
3233 * @buffer: The ring buffer
3234 * @cpu: The per CPU buffer to get the number of overruns from
3237 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3239 struct ring_buffer_per_cpu *cpu_buffer;
3242 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3245 cpu_buffer = buffer->buffers[cpu];
3246 ret = local_read(&cpu_buffer->commit_overrun);
3250 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3253 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3254 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3255 * @buffer: The ring buffer
3256 * @cpu: The per CPU buffer to get the number of overruns from
3259 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3261 struct ring_buffer_per_cpu *cpu_buffer;
3264 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3267 cpu_buffer = buffer->buffers[cpu];
3268 ret = local_read(&cpu_buffer->dropped_events);
3272 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3275 * ring_buffer_read_events_cpu - get the number of events successfully read
3276 * @buffer: The ring buffer
3277 * @cpu: The per CPU buffer to get the number of events read
3280 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3282 struct ring_buffer_per_cpu *cpu_buffer;
3284 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3287 cpu_buffer = buffer->buffers[cpu];
3288 return cpu_buffer->read;
3290 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3293 * ring_buffer_entries - get the number of entries in a buffer
3294 * @buffer: The ring buffer
3296 * Returns the total number of entries in the ring buffer
3299 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3301 struct ring_buffer_per_cpu *cpu_buffer;
3302 unsigned long entries = 0;
3305 /* if you care about this being correct, lock the buffer */
3306 for_each_buffer_cpu(buffer, cpu) {
3307 cpu_buffer = buffer->buffers[cpu];
3308 entries += rb_num_of_entries(cpu_buffer);
3313 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3316 * ring_buffer_overruns - get the number of overruns in buffer
3317 * @buffer: The ring buffer
3319 * Returns the total number of overruns in the ring buffer
3322 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3324 struct ring_buffer_per_cpu *cpu_buffer;
3325 unsigned long overruns = 0;
3328 /* if you care about this being correct, lock the buffer */
3329 for_each_buffer_cpu(buffer, cpu) {
3330 cpu_buffer = buffer->buffers[cpu];
3331 overruns += local_read(&cpu_buffer->overrun);
3336 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3338 static void rb_iter_reset(struct ring_buffer_iter *iter)
3340 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3342 /* Iterator usage is expected to have record disabled */
3343 if (list_empty(&cpu_buffer->reader_page->list)) {
3344 iter->head_page = rb_set_head_page(cpu_buffer);
3345 if (unlikely(!iter->head_page))
3347 iter->head = iter->head_page->read;
3349 iter->head_page = cpu_buffer->reader_page;
3350 iter->head = cpu_buffer->reader_page->read;
3353 iter->read_stamp = cpu_buffer->read_stamp;
3355 iter->read_stamp = iter->head_page->page->time_stamp;
3356 iter->cache_reader_page = cpu_buffer->reader_page;
3357 iter->cache_read = cpu_buffer->read;
3361 * ring_buffer_iter_reset - reset an iterator
3362 * @iter: The iterator to reset
3364 * Resets the iterator, so that it will start from the beginning
3367 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3369 struct ring_buffer_per_cpu *cpu_buffer;
3370 unsigned long flags;
3375 cpu_buffer = iter->cpu_buffer;
3377 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3378 rb_iter_reset(iter);
3379 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3381 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3384 * ring_buffer_iter_empty - check if an iterator has no more to read
3385 * @iter: The iterator to check
3387 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3389 struct ring_buffer_per_cpu *cpu_buffer;
3391 cpu_buffer = iter->cpu_buffer;
3393 return iter->head_page == cpu_buffer->commit_page &&
3394 iter->head == rb_commit_index(cpu_buffer);
3396 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3399 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3400 struct ring_buffer_event *event)
3404 switch (event->type_len) {
3405 case RINGBUF_TYPE_PADDING:
3408 case RINGBUF_TYPE_TIME_EXTEND:
3409 delta = event->array[0];
3411 delta += event->time_delta;
3412 cpu_buffer->read_stamp += delta;
3415 case RINGBUF_TYPE_TIME_STAMP:
3416 /* FIXME: not implemented */
3419 case RINGBUF_TYPE_DATA:
3420 cpu_buffer->read_stamp += event->time_delta;
3430 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3431 struct ring_buffer_event *event)
3435 switch (event->type_len) {
3436 case RINGBUF_TYPE_PADDING:
3439 case RINGBUF_TYPE_TIME_EXTEND:
3440 delta = event->array[0];
3442 delta += event->time_delta;
3443 iter->read_stamp += delta;
3446 case RINGBUF_TYPE_TIME_STAMP:
3447 /* FIXME: not implemented */
3450 case RINGBUF_TYPE_DATA:
3451 iter->read_stamp += event->time_delta;
3460 static struct buffer_page *
3461 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3463 struct buffer_page *reader = NULL;
3464 unsigned long overwrite;
3465 unsigned long flags;
3469 local_irq_save(flags);
3470 arch_spin_lock(&cpu_buffer->lock);
3474 * This should normally only loop twice. But because the
3475 * start of the reader inserts an empty page, it causes
3476 * a case where we will loop three times. There should be no
3477 * reason to loop four times (that I know of).
3479 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3484 reader = cpu_buffer->reader_page;
3486 /* If there's more to read, return this page */
3487 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3490 /* Never should we have an index greater than the size */
3491 if (RB_WARN_ON(cpu_buffer,
3492 cpu_buffer->reader_page->read > rb_page_size(reader)))
3495 /* check if we caught up to the tail */
3497 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3500 /* Don't bother swapping if the ring buffer is empty */
3501 if (rb_num_of_entries(cpu_buffer) == 0)
3505 * Reset the reader page to size zero.
3507 local_set(&cpu_buffer->reader_page->write, 0);
3508 local_set(&cpu_buffer->reader_page->entries, 0);
3509 local_set(&cpu_buffer->reader_page->page->commit, 0);
3510 cpu_buffer->reader_page->real_end = 0;
3514 * Splice the empty reader page into the list around the head.
3516 reader = rb_set_head_page(cpu_buffer);
3519 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3520 cpu_buffer->reader_page->list.prev = reader->list.prev;
3523 * cpu_buffer->pages just needs to point to the buffer, it
3524 * has no specific buffer page to point to. Lets move it out
3525 * of our way so we don't accidentally swap it.
3527 cpu_buffer->pages = reader->list.prev;
3529 /* The reader page will be pointing to the new head */
3530 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3533 * We want to make sure we read the overruns after we set up our
3534 * pointers to the next object. The writer side does a
3535 * cmpxchg to cross pages which acts as the mb on the writer
3536 * side. Note, the reader will constantly fail the swap
3537 * while the writer is updating the pointers, so this
3538 * guarantees that the overwrite recorded here is the one we
3539 * want to compare with the last_overrun.
3542 overwrite = local_read(&(cpu_buffer->overrun));
3545 * Here's the tricky part.
3547 * We need to move the pointer past the header page.
3548 * But we can only do that if a writer is not currently
3549 * moving it. The page before the header page has the
3550 * flag bit '1' set if it is pointing to the page we want.
3551 * but if the writer is in the process of moving it
3552 * than it will be '2' or already moved '0'.
3555 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3558 * If we did not convert it, then we must try again.
3564 * Yeah! We succeeded in replacing the page.
3566 * Now make the new head point back to the reader page.
3568 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3569 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3571 /* Finally update the reader page to the new head */
3572 cpu_buffer->reader_page = reader;
3573 rb_reset_reader_page(cpu_buffer);
3575 if (overwrite != cpu_buffer->last_overrun) {
3576 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3577 cpu_buffer->last_overrun = overwrite;
3583 arch_spin_unlock(&cpu_buffer->lock);
3584 local_irq_restore(flags);
3589 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3591 struct ring_buffer_event *event;
3592 struct buffer_page *reader;
3595 reader = rb_get_reader_page(cpu_buffer);
3597 /* This function should not be called when buffer is empty */
3598 if (RB_WARN_ON(cpu_buffer, !reader))
3601 event = rb_reader_event(cpu_buffer);
3603 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3606 rb_update_read_stamp(cpu_buffer, event);
3608 length = rb_event_length(event);
3609 cpu_buffer->reader_page->read += length;
3612 static void rb_advance_iter(struct ring_buffer_iter *iter)
3614 struct ring_buffer_per_cpu *cpu_buffer;
3615 struct ring_buffer_event *event;
3618 cpu_buffer = iter->cpu_buffer;
3621 * Check if we are at the end of the buffer.
3623 if (iter->head >= rb_page_size(iter->head_page)) {
3624 /* discarded commits can make the page empty */
3625 if (iter->head_page == cpu_buffer->commit_page)
3631 event = rb_iter_head_event(iter);
3633 length = rb_event_length(event);
3636 * This should not be called to advance the header if we are
3637 * at the tail of the buffer.
3639 if (RB_WARN_ON(cpu_buffer,
3640 (iter->head_page == cpu_buffer->commit_page) &&
3641 (iter->head + length > rb_commit_index(cpu_buffer))))
3644 rb_update_iter_read_stamp(iter, event);
3646 iter->head += length;
3648 /* check for end of page padding */
3649 if ((iter->head >= rb_page_size(iter->head_page)) &&
3650 (iter->head_page != cpu_buffer->commit_page))
3654 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3656 return cpu_buffer->lost_events;
3659 static struct ring_buffer_event *
3660 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3661 unsigned long *lost_events)
3663 struct ring_buffer_event *event;
3664 struct buffer_page *reader;
3669 * We repeat when a time extend is encountered.
3670 * Since the time extend is always attached to a data event,
3671 * we should never loop more than once.
3672 * (We never hit the following condition more than twice).
3674 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3677 reader = rb_get_reader_page(cpu_buffer);
3681 event = rb_reader_event(cpu_buffer);
3683 switch (event->type_len) {
3684 case RINGBUF_TYPE_PADDING:
3685 if (rb_null_event(event))
3686 RB_WARN_ON(cpu_buffer, 1);
3688 * Because the writer could be discarding every
3689 * event it creates (which would probably be bad)
3690 * if we were to go back to "again" then we may never
3691 * catch up, and will trigger the warn on, or lock
3692 * the box. Return the padding, and we will release
3693 * the current locks, and try again.
3697 case RINGBUF_TYPE_TIME_EXTEND:
3698 /* Internal data, OK to advance */
3699 rb_advance_reader(cpu_buffer);
3702 case RINGBUF_TYPE_TIME_STAMP:
3703 /* FIXME: not implemented */
3704 rb_advance_reader(cpu_buffer);
3707 case RINGBUF_TYPE_DATA:
3709 *ts = cpu_buffer->read_stamp + event->time_delta;
3710 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3711 cpu_buffer->cpu, ts);
3714 *lost_events = rb_lost_events(cpu_buffer);
3723 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3725 static struct ring_buffer_event *
3726 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3728 struct ring_buffer *buffer;
3729 struct ring_buffer_per_cpu *cpu_buffer;
3730 struct ring_buffer_event *event;
3733 cpu_buffer = iter->cpu_buffer;
3734 buffer = cpu_buffer->buffer;
3737 * Check if someone performed a consuming read to
3738 * the buffer. A consuming read invalidates the iterator
3739 * and we need to reset the iterator in this case.
3741 if (unlikely(iter->cache_read != cpu_buffer->read ||
3742 iter->cache_reader_page != cpu_buffer->reader_page))
3743 rb_iter_reset(iter);
3746 if (ring_buffer_iter_empty(iter))
3750 * We repeat when a time extend is encountered.
3751 * Since the time extend is always attached to a data event,
3752 * we should never loop more than once.
3753 * (We never hit the following condition more than twice).
3755 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3758 if (rb_per_cpu_empty(cpu_buffer))
3761 if (iter->head >= local_read(&iter->head_page->page->commit)) {
3766 event = rb_iter_head_event(iter);
3768 switch (event->type_len) {
3769 case RINGBUF_TYPE_PADDING:
3770 if (rb_null_event(event)) {
3774 rb_advance_iter(iter);
3777 case RINGBUF_TYPE_TIME_EXTEND:
3778 /* Internal data, OK to advance */
3779 rb_advance_iter(iter);
3782 case RINGBUF_TYPE_TIME_STAMP:
3783 /* FIXME: not implemented */
3784 rb_advance_iter(iter);
3787 case RINGBUF_TYPE_DATA:
3789 *ts = iter->read_stamp + event->time_delta;
3790 ring_buffer_normalize_time_stamp(buffer,
3791 cpu_buffer->cpu, ts);
3801 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3803 static inline int rb_ok_to_lock(void)
3806 * If an NMI die dumps out the content of the ring buffer
3807 * do not grab locks. We also permanently disable the ring
3808 * buffer too. A one time deal is all you get from reading
3809 * the ring buffer from an NMI.
3811 if (likely(!in_nmi()))
3814 tracing_off_permanent();
3819 * ring_buffer_peek - peek at the next event to be read
3820 * @buffer: The ring buffer to read
3821 * @cpu: The cpu to peak at
3822 * @ts: The timestamp counter of this event.
3823 * @lost_events: a variable to store if events were lost (may be NULL)
3825 * This will return the event that will be read next, but does
3826 * not consume the data.
3828 struct ring_buffer_event *
3829 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3830 unsigned long *lost_events)
3832 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3833 struct ring_buffer_event *event;
3834 unsigned long flags;
3837 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3840 dolock = rb_ok_to_lock();
3842 local_irq_save(flags);
3844 raw_spin_lock(&cpu_buffer->reader_lock);
3845 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3846 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3847 rb_advance_reader(cpu_buffer);
3849 raw_spin_unlock(&cpu_buffer->reader_lock);
3850 local_irq_restore(flags);
3852 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3859 * ring_buffer_iter_peek - peek at the next event to be read
3860 * @iter: The ring buffer iterator
3861 * @ts: The timestamp counter of this event.
3863 * This will return the event that will be read next, but does
3864 * not increment the iterator.
3866 struct ring_buffer_event *
3867 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3869 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3870 struct ring_buffer_event *event;
3871 unsigned long flags;
3874 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3875 event = rb_iter_peek(iter, ts);
3876 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3878 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3885 * ring_buffer_consume - return an event and consume it
3886 * @buffer: The ring buffer to get the next event from
3887 * @cpu: the cpu to read the buffer from
3888 * @ts: a variable to store the timestamp (may be NULL)
3889 * @lost_events: a variable to store if events were lost (may be NULL)
3891 * Returns the next event in the ring buffer, and that event is consumed.
3892 * Meaning, that sequential reads will keep returning a different event,
3893 * and eventually empty the ring buffer if the producer is slower.
3895 struct ring_buffer_event *
3896 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3897 unsigned long *lost_events)
3899 struct ring_buffer_per_cpu *cpu_buffer;
3900 struct ring_buffer_event *event = NULL;
3901 unsigned long flags;
3904 dolock = rb_ok_to_lock();
3907 /* might be called in atomic */
3910 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3913 cpu_buffer = buffer->buffers[cpu];
3914 local_irq_save(flags);
3916 raw_spin_lock(&cpu_buffer->reader_lock);
3918 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3920 cpu_buffer->lost_events = 0;
3921 rb_advance_reader(cpu_buffer);
3925 raw_spin_unlock(&cpu_buffer->reader_lock);
3926 local_irq_restore(flags);
3931 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3936 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3939 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3940 * @buffer: The ring buffer to read from
3941 * @cpu: The cpu buffer to iterate over
3943 * This performs the initial preparations necessary to iterate
3944 * through the buffer. Memory is allocated, buffer recording
3945 * is disabled, and the iterator pointer is returned to the caller.
3947 * Disabling buffer recordng prevents the reading from being
3948 * corrupted. This is not a consuming read, so a producer is not
3951 * After a sequence of ring_buffer_read_prepare calls, the user is
3952 * expected to make at least one call to ring_buffer_prepare_sync.
3953 * Afterwards, ring_buffer_read_start is invoked to get things going
3956 * This overall must be paired with ring_buffer_finish.
3958 struct ring_buffer_iter *
3959 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3961 struct ring_buffer_per_cpu *cpu_buffer;
3962 struct ring_buffer_iter *iter;
3964 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3967 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3971 cpu_buffer = buffer->buffers[cpu];
3973 iter->cpu_buffer = cpu_buffer;
3975 atomic_inc(&buffer->resize_disabled);
3976 atomic_inc(&cpu_buffer->record_disabled);
3980 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
3983 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
3985 * All previously invoked ring_buffer_read_prepare calls to prepare
3986 * iterators will be synchronized. Afterwards, read_buffer_read_start
3987 * calls on those iterators are allowed.
3990 ring_buffer_read_prepare_sync(void)
3992 synchronize_sched();
3994 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
3997 * ring_buffer_read_start - start a non consuming read of the buffer
3998 * @iter: The iterator returned by ring_buffer_read_prepare
4000 * This finalizes the startup of an iteration through the buffer.
4001 * The iterator comes from a call to ring_buffer_read_prepare and
4002 * an intervening ring_buffer_read_prepare_sync must have been
4005 * Must be paired with ring_buffer_finish.
4008 ring_buffer_read_start(struct ring_buffer_iter *iter)
4010 struct ring_buffer_per_cpu *cpu_buffer;
4011 unsigned long flags;
4016 cpu_buffer = iter->cpu_buffer;
4018 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4019 arch_spin_lock(&cpu_buffer->lock);
4020 rb_iter_reset(iter);
4021 arch_spin_unlock(&cpu_buffer->lock);
4022 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4024 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4027 * ring_buffer_finish - finish reading the iterator of the buffer
4028 * @iter: The iterator retrieved by ring_buffer_start
4030 * This re-enables the recording to the buffer, and frees the
4034 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4036 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4037 unsigned long flags;
4040 * Ring buffer is disabled from recording, here's a good place
4041 * to check the integrity of the ring buffer.
4042 * Must prevent readers from trying to read, as the check
4043 * clears the HEAD page and readers require it.
4045 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4046 rb_check_pages(cpu_buffer);
4047 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4049 atomic_dec(&cpu_buffer->record_disabled);
4050 atomic_dec(&cpu_buffer->buffer->resize_disabled);
4053 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4056 * ring_buffer_read - read the next item in the ring buffer by the iterator
4057 * @iter: The ring buffer iterator
4058 * @ts: The time stamp of the event read.
4060 * This reads the next event in the ring buffer and increments the iterator.
4062 struct ring_buffer_event *
4063 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4065 struct ring_buffer_event *event;
4066 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4067 unsigned long flags;
4069 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4071 event = rb_iter_peek(iter, ts);
4075 if (event->type_len == RINGBUF_TYPE_PADDING)
4078 rb_advance_iter(iter);
4080 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4084 EXPORT_SYMBOL_GPL(ring_buffer_read);
4087 * ring_buffer_size - return the size of the ring buffer (in bytes)
4088 * @buffer: The ring buffer.
4090 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4093 * Earlier, this method returned
4094 * BUF_PAGE_SIZE * buffer->nr_pages
4095 * Since the nr_pages field is now removed, we have converted this to
4096 * return the per cpu buffer value.
4098 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4101 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4103 EXPORT_SYMBOL_GPL(ring_buffer_size);
4106 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4108 rb_head_page_deactivate(cpu_buffer);
4110 cpu_buffer->head_page
4111 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4112 local_set(&cpu_buffer->head_page->write, 0);
4113 local_set(&cpu_buffer->head_page->entries, 0);
4114 local_set(&cpu_buffer->head_page->page->commit, 0);
4116 cpu_buffer->head_page->read = 0;
4118 cpu_buffer->tail_page = cpu_buffer->head_page;
4119 cpu_buffer->commit_page = cpu_buffer->head_page;
4121 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4122 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4123 local_set(&cpu_buffer->reader_page->write, 0);
4124 local_set(&cpu_buffer->reader_page->entries, 0);
4125 local_set(&cpu_buffer->reader_page->page->commit, 0);
4126 cpu_buffer->reader_page->read = 0;
4128 local_set(&cpu_buffer->entries_bytes, 0);
4129 local_set(&cpu_buffer->overrun, 0);
4130 local_set(&cpu_buffer->commit_overrun, 0);
4131 local_set(&cpu_buffer->dropped_events, 0);
4132 local_set(&cpu_buffer->entries, 0);
4133 local_set(&cpu_buffer->committing, 0);
4134 local_set(&cpu_buffer->commits, 0);
4135 cpu_buffer->read = 0;
4136 cpu_buffer->read_bytes = 0;
4138 cpu_buffer->write_stamp = 0;
4139 cpu_buffer->read_stamp = 0;
4141 cpu_buffer->lost_events = 0;
4142 cpu_buffer->last_overrun = 0;
4144 rb_head_page_activate(cpu_buffer);
4148 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4149 * @buffer: The ring buffer to reset a per cpu buffer of
4150 * @cpu: The CPU buffer to be reset
4152 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4154 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4155 unsigned long flags;
4157 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4160 atomic_inc(&buffer->resize_disabled);
4161 atomic_inc(&cpu_buffer->record_disabled);
4163 /* Make sure all commits have finished */
4164 synchronize_sched();
4166 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4168 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4171 arch_spin_lock(&cpu_buffer->lock);
4173 rb_reset_cpu(cpu_buffer);
4175 arch_spin_unlock(&cpu_buffer->lock);
4178 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4180 atomic_dec(&cpu_buffer->record_disabled);
4181 atomic_dec(&buffer->resize_disabled);
4183 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4186 * ring_buffer_reset - reset a ring buffer
4187 * @buffer: The ring buffer to reset all cpu buffers
4189 void ring_buffer_reset(struct ring_buffer *buffer)
4193 for_each_buffer_cpu(buffer, cpu)
4194 ring_buffer_reset_cpu(buffer, cpu);
4196 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4199 * rind_buffer_empty - is the ring buffer empty?
4200 * @buffer: The ring buffer to test
4202 int ring_buffer_empty(struct ring_buffer *buffer)
4204 struct ring_buffer_per_cpu *cpu_buffer;
4205 unsigned long flags;
4210 dolock = rb_ok_to_lock();
4212 /* yes this is racy, but if you don't like the race, lock the buffer */
4213 for_each_buffer_cpu(buffer, cpu) {
4214 cpu_buffer = buffer->buffers[cpu];
4215 local_irq_save(flags);
4217 raw_spin_lock(&cpu_buffer->reader_lock);
4218 ret = rb_per_cpu_empty(cpu_buffer);
4220 raw_spin_unlock(&cpu_buffer->reader_lock);
4221 local_irq_restore(flags);
4229 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4232 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4233 * @buffer: The ring buffer
4234 * @cpu: The CPU buffer to test
4236 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4238 struct ring_buffer_per_cpu *cpu_buffer;
4239 unsigned long flags;
4243 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4246 dolock = rb_ok_to_lock();
4248 cpu_buffer = buffer->buffers[cpu];
4249 local_irq_save(flags);
4251 raw_spin_lock(&cpu_buffer->reader_lock);
4252 ret = rb_per_cpu_empty(cpu_buffer);
4254 raw_spin_unlock(&cpu_buffer->reader_lock);
4255 local_irq_restore(flags);
4259 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4261 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4263 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4264 * @buffer_a: One buffer to swap with
4265 * @buffer_b: The other buffer to swap with
4267 * This function is useful for tracers that want to take a "snapshot"
4268 * of a CPU buffer and has another back up buffer lying around.
4269 * it is expected that the tracer handles the cpu buffer not being
4270 * used at the moment.
4272 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4273 struct ring_buffer *buffer_b, int cpu)
4275 struct ring_buffer_per_cpu *cpu_buffer_a;
4276 struct ring_buffer_per_cpu *cpu_buffer_b;
4279 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4280 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4283 cpu_buffer_a = buffer_a->buffers[cpu];
4284 cpu_buffer_b = buffer_b->buffers[cpu];
4286 /* At least make sure the two buffers are somewhat the same */
4287 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4292 if (ring_buffer_flags != RB_BUFFERS_ON)
4295 if (atomic_read(&buffer_a->record_disabled))
4298 if (atomic_read(&buffer_b->record_disabled))
4301 if (atomic_read(&cpu_buffer_a->record_disabled))
4304 if (atomic_read(&cpu_buffer_b->record_disabled))
4308 * We can't do a synchronize_sched here because this
4309 * function can be called in atomic context.
4310 * Normally this will be called from the same CPU as cpu.
4311 * If not it's up to the caller to protect this.
4313 atomic_inc(&cpu_buffer_a->record_disabled);
4314 atomic_inc(&cpu_buffer_b->record_disabled);
4317 if (local_read(&cpu_buffer_a->committing))
4319 if (local_read(&cpu_buffer_b->committing))
4322 buffer_a->buffers[cpu] = cpu_buffer_b;
4323 buffer_b->buffers[cpu] = cpu_buffer_a;
4325 cpu_buffer_b->buffer = buffer_a;
4326 cpu_buffer_a->buffer = buffer_b;
4331 atomic_dec(&cpu_buffer_a->record_disabled);
4332 atomic_dec(&cpu_buffer_b->record_disabled);
4336 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4337 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4340 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4341 * @buffer: the buffer to allocate for.
4343 * This function is used in conjunction with ring_buffer_read_page.
4344 * When reading a full page from the ring buffer, these functions
4345 * can be used to speed up the process. The calling function should
4346 * allocate a few pages first with this function. Then when it
4347 * needs to get pages from the ring buffer, it passes the result
4348 * of this function into ring_buffer_read_page, which will swap
4349 * the page that was allocated, with the read page of the buffer.
4352 * The page allocated, or NULL on error.
4354 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4356 struct buffer_data_page *bpage;
4359 page = alloc_pages_node(cpu_to_node(cpu),
4360 GFP_KERNEL | __GFP_NORETRY, 0);
4364 bpage = page_address(page);
4366 rb_init_page(bpage);
4370 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4373 * ring_buffer_free_read_page - free an allocated read page
4374 * @buffer: the buffer the page was allocate for
4375 * @data: the page to free
4377 * Free a page allocated from ring_buffer_alloc_read_page.
4379 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4381 free_page((unsigned long)data);
4383 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4386 * ring_buffer_read_page - extract a page from the ring buffer
4387 * @buffer: buffer to extract from
4388 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4389 * @len: amount to extract
4390 * @cpu: the cpu of the buffer to extract
4391 * @full: should the extraction only happen when the page is full.
4393 * This function will pull out a page from the ring buffer and consume it.
4394 * @data_page must be the address of the variable that was returned
4395 * from ring_buffer_alloc_read_page. This is because the page might be used
4396 * to swap with a page in the ring buffer.
4399 * rpage = ring_buffer_alloc_read_page(buffer);
4402 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4404 * process_page(rpage, ret);
4406 * When @full is set, the function will not return true unless
4407 * the writer is off the reader page.
4409 * Note: it is up to the calling functions to handle sleeps and wakeups.
4410 * The ring buffer can be used anywhere in the kernel and can not
4411 * blindly call wake_up. The layer that uses the ring buffer must be
4412 * responsible for that.
4415 * >=0 if data has been transferred, returns the offset of consumed data.
4416 * <0 if no data has been transferred.
4418 int ring_buffer_read_page(struct ring_buffer *buffer,
4419 void **data_page, size_t len, int cpu, int full)
4421 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4422 struct ring_buffer_event *event;
4423 struct buffer_data_page *bpage;
4424 struct buffer_page *reader;
4425 unsigned long missed_events;
4426 unsigned long flags;
4427 unsigned int commit;
4432 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4436 * If len is not big enough to hold the page header, then
4437 * we can not copy anything.
4439 if (len <= BUF_PAGE_HDR_SIZE)
4442 len -= BUF_PAGE_HDR_SIZE;
4451 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4453 reader = rb_get_reader_page(cpu_buffer);
4457 event = rb_reader_event(cpu_buffer);
4459 read = reader->read;
4460 commit = rb_page_commit(reader);
4462 /* Check if any events were dropped */
4463 missed_events = cpu_buffer->lost_events;
4466 * If this page has been partially read or
4467 * if len is not big enough to read the rest of the page or
4468 * a writer is still on the page, then
4469 * we must copy the data from the page to the buffer.
4470 * Otherwise, we can simply swap the page with the one passed in.
4472 if (read || (len < (commit - read)) ||
4473 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4474 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4475 unsigned int rpos = read;
4476 unsigned int pos = 0;
4482 if (len > (commit - read))
4483 len = (commit - read);
4485 /* Always keep the time extend and data together */
4486 size = rb_event_ts_length(event);
4491 /* save the current timestamp, since the user will need it */
4492 save_timestamp = cpu_buffer->read_stamp;
4494 /* Need to copy one event at a time */
4496 /* We need the size of one event, because
4497 * rb_advance_reader only advances by one event,
4498 * whereas rb_event_ts_length may include the size of
4499 * one or two events.
4500 * We have already ensured there's enough space if this
4501 * is a time extend. */
4502 size = rb_event_length(event);
4503 memcpy(bpage->data + pos, rpage->data + rpos, size);
4507 rb_advance_reader(cpu_buffer);
4508 rpos = reader->read;
4514 event = rb_reader_event(cpu_buffer);
4515 /* Always keep the time extend and data together */
4516 size = rb_event_ts_length(event);
4517 } while (len >= size);
4520 local_set(&bpage->commit, pos);
4521 bpage->time_stamp = save_timestamp;
4523 /* we copied everything to the beginning */
4526 /* update the entry counter */
4527 cpu_buffer->read += rb_page_entries(reader);
4528 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4530 /* swap the pages */
4531 rb_init_page(bpage);
4532 bpage = reader->page;
4533 reader->page = *data_page;
4534 local_set(&reader->write, 0);
4535 local_set(&reader->entries, 0);
4540 * Use the real_end for the data size,
4541 * This gives us a chance to store the lost events
4544 if (reader->real_end)
4545 local_set(&bpage->commit, reader->real_end);
4549 cpu_buffer->lost_events = 0;
4551 commit = local_read(&bpage->commit);
4553 * Set a flag in the commit field if we lost events
4555 if (missed_events) {
4556 /* If there is room at the end of the page to save the
4557 * missed events, then record it there.
4559 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4560 memcpy(&bpage->data[commit], &missed_events,
4561 sizeof(missed_events));
4562 local_add(RB_MISSED_STORED, &bpage->commit);
4563 commit += sizeof(missed_events);
4565 local_add(RB_MISSED_EVENTS, &bpage->commit);
4569 * This page may be off to user land. Zero it out here.
4571 if (commit < BUF_PAGE_SIZE)
4572 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4575 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4580 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4582 #ifdef CONFIG_HOTPLUG_CPU
4583 static int rb_cpu_notify(struct notifier_block *self,
4584 unsigned long action, void *hcpu)
4586 struct ring_buffer *buffer =
4587 container_of(self, struct ring_buffer, cpu_notify);
4588 long cpu = (long)hcpu;
4589 int cpu_i, nr_pages_same;
4590 unsigned int nr_pages;
4593 case CPU_UP_PREPARE:
4594 case CPU_UP_PREPARE_FROZEN:
4595 if (cpumask_test_cpu(cpu, buffer->cpumask))
4600 /* check if all cpu sizes are same */
4601 for_each_buffer_cpu(buffer, cpu_i) {
4602 /* fill in the size from first enabled cpu */
4604 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4605 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4610 /* allocate minimum pages, user can later expand it */
4613 buffer->buffers[cpu] =
4614 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4615 if (!buffer->buffers[cpu]) {
4616 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4621 cpumask_set_cpu(cpu, buffer->cpumask);
4623 case CPU_DOWN_PREPARE:
4624 case CPU_DOWN_PREPARE_FROZEN:
4627 * If we were to free the buffer, then the user would
4628 * lose any trace that was in the buffer.