4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 #include <linux/ftrace_event.h>
7 #include <linux/ring_buffer.h>
8 #include <linux/trace_clock.h>
9 #include <linux/trace_seq.h>
10 #include <linux/spinlock.h>
11 #include <linux/irq_work.h>
12 #include <linux/debugfs.h>
13 #include <linux/uaccess.h>
14 #include <linux/hardirq.h>
15 #include <linux/kthread.h> /* for self test */
16 #include <linux/kmemcheck.h>
17 #include <linux/module.h>
18 #include <linux/percpu.h>
19 #include <linux/mutex.h>
20 #include <linux/delay.h>
21 #include <linux/slab.h>
22 #include <linux/init.h>
23 #include <linux/hash.h>
24 #include <linux/list.h>
25 #include <linux/cpu.h>
28 #include <asm/local.h>
30 static void update_pages_handler(struct work_struct *work);
33 * The ring buffer header is special. We must manually up keep it.
35 int ring_buffer_print_entry_header(struct trace_seq *s)
39 ret = trace_seq_printf(s, "# compressed entry header\n");
40 ret = trace_seq_printf(s, "\ttype_len : 5 bits\n");
41 ret = trace_seq_printf(s, "\ttime_delta : 27 bits\n");
42 ret = trace_seq_printf(s, "\tarray : 32 bits\n");
43 ret = trace_seq_printf(s, "\n");
44 ret = trace_seq_printf(s, "\tpadding : type == %d\n",
45 RINGBUF_TYPE_PADDING);
46 ret = trace_seq_printf(s, "\ttime_extend : type == %d\n",
47 RINGBUF_TYPE_TIME_EXTEND);
48 ret = trace_seq_printf(s, "\tdata max type_len == %d\n",
49 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
55 * The ring buffer is made up of a list of pages. A separate list of pages is
56 * allocated for each CPU. A writer may only write to a buffer that is
57 * associated with the CPU it is currently executing on. A reader may read
58 * from any per cpu buffer.
60 * The reader is special. For each per cpu buffer, the reader has its own
61 * reader page. When a reader has read the entire reader page, this reader
62 * page is swapped with another page in the ring buffer.
64 * Now, as long as the writer is off the reader page, the reader can do what
65 * ever it wants with that page. The writer will never write to that page
66 * again (as long as it is out of the ring buffer).
68 * Here's some silly ASCII art.
71 * |reader| RING BUFFER
73 * +------+ +---+ +---+ +---+
82 * |reader| RING BUFFER
83 * |page |------------------v
84 * +------+ +---+ +---+ +---+
93 * |reader| RING BUFFER
94 * |page |------------------v
95 * +------+ +---+ +---+ +---+
100 * +------------------------------+
104 * |buffer| RING BUFFER
105 * |page |------------------v
106 * +------+ +---+ +---+ +---+
108 * | New +---+ +---+ +---+
111 * +------------------------------+
114 * After we make this swap, the reader can hand this page off to the splice
115 * code and be done with it. It can even allocate a new page if it needs to
116 * and swap that into the ring buffer.
118 * We will be using cmpxchg soon to make all this lockless.
123 * A fast way to enable or disable all ring buffers is to
124 * call tracing_on or tracing_off. Turning off the ring buffers
125 * prevents all ring buffers from being recorded to.
126 * Turning this switch on, makes it OK to write to the
127 * ring buffer, if the ring buffer is enabled itself.
129 * There's three layers that must be on in order to write
130 * to the ring buffer.
132 * 1) This global flag must be set.
133 * 2) The ring buffer must be enabled for recording.
134 * 3) The per cpu buffer must be enabled for recording.
136 * In case of an anomaly, this global flag has a bit set that
137 * will permantly disable all ring buffers.
141 * Global flag to disable all recording to ring buffers
142 * This has two bits: ON, DISABLED
146 * 0 0 : ring buffers are off
147 * 1 0 : ring buffers are on
148 * X 1 : ring buffers are permanently disabled
152 RB_BUFFERS_ON_BIT = 0,
153 RB_BUFFERS_DISABLED_BIT = 1,
157 RB_BUFFERS_ON = 1 << RB_BUFFERS_ON_BIT,
158 RB_BUFFERS_DISABLED = 1 << RB_BUFFERS_DISABLED_BIT,
161 static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON;
163 /* Used for individual buffers (after the counter) */
164 #define RB_BUFFER_OFF (1 << 20)
166 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
169 * tracing_off_permanent - permanently disable ring buffers
171 * This function, once called, will disable all ring buffers
174 void tracing_off_permanent(void)
176 set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags);
179 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
180 #define RB_ALIGNMENT 4U
181 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
182 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
184 #if !defined(CONFIG_64BIT) || defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
185 # define RB_FORCE_8BYTE_ALIGNMENT 0
186 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
188 # define RB_FORCE_8BYTE_ALIGNMENT 1
189 # define RB_ARCH_ALIGNMENT 8U
192 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
193 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
196 RB_LEN_TIME_EXTEND = 8,
197 RB_LEN_TIME_STAMP = 16,
200 #define skip_time_extend(event) \
201 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
203 static inline int rb_null_event(struct ring_buffer_event *event)
205 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
208 static void rb_event_set_padding(struct ring_buffer_event *event)
210 /* padding has a NULL time_delta */
211 event->type_len = RINGBUF_TYPE_PADDING;
212 event->time_delta = 0;
216 rb_event_data_length(struct ring_buffer_event *event)
221 length = event->type_len * RB_ALIGNMENT;
223 length = event->array[0];
224 return length + RB_EVNT_HDR_SIZE;
228 * Return the length of the given event. Will return
229 * the length of the time extend if the event is a
232 static inline unsigned
233 rb_event_length(struct ring_buffer_event *event)
235 switch (event->type_len) {
236 case RINGBUF_TYPE_PADDING:
237 if (rb_null_event(event))
240 return event->array[0] + RB_EVNT_HDR_SIZE;
242 case RINGBUF_TYPE_TIME_EXTEND:
243 return RB_LEN_TIME_EXTEND;
245 case RINGBUF_TYPE_TIME_STAMP:
246 return RB_LEN_TIME_STAMP;
248 case RINGBUF_TYPE_DATA:
249 return rb_event_data_length(event);
258 * Return total length of time extend and data,
259 * or just the event length for all other events.
261 static inline unsigned
262 rb_event_ts_length(struct ring_buffer_event *event)
266 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
267 /* time extends include the data event after it */
268 len = RB_LEN_TIME_EXTEND;
269 event = skip_time_extend(event);
271 return len + rb_event_length(event);
275 * ring_buffer_event_length - return the length of the event
276 * @event: the event to get the length of
278 * Returns the size of the data load of a data event.
279 * If the event is something other than a data event, it
280 * returns the size of the event itself. With the exception
281 * of a TIME EXTEND, where it still returns the size of the
282 * data load of the data event after it.
284 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
288 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
289 event = skip_time_extend(event);
291 length = rb_event_length(event);
292 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
294 length -= RB_EVNT_HDR_SIZE;
295 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
296 length -= sizeof(event->array[0]);
299 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
301 /* inline for ring buffer fast paths */
303 rb_event_data(struct ring_buffer_event *event)
305 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
306 event = skip_time_extend(event);
307 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
308 /* If length is in len field, then array[0] has the data */
310 return (void *)&event->array[0];
311 /* Otherwise length is in array[0] and array[1] has the data */
312 return (void *)&event->array[1];
316 * ring_buffer_event_data - return the data of the event
317 * @event: the event to get the data from
319 void *ring_buffer_event_data(struct ring_buffer_event *event)
321 return rb_event_data(event);
323 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
325 #define for_each_buffer_cpu(buffer, cpu) \
326 for_each_cpu(cpu, buffer->cpumask)
329 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
330 #define TS_DELTA_TEST (~TS_MASK)
332 /* Flag when events were overwritten */
333 #define RB_MISSED_EVENTS (1 << 31)
334 /* Missed count stored at end */
335 #define RB_MISSED_STORED (1 << 30)
337 struct buffer_data_page {
338 u64 time_stamp; /* page time stamp */
339 local_t commit; /* write committed index */
340 unsigned char data[]; /* data of buffer page */
344 * Note, the buffer_page list must be first. The buffer pages
345 * are allocated in cache lines, which means that each buffer
346 * page will be at the beginning of a cache line, and thus
347 * the least significant bits will be zero. We use this to
348 * add flags in the list struct pointers, to make the ring buffer
352 struct list_head list; /* list of buffer pages */
353 local_t write; /* index for next write */
354 unsigned read; /* index for next read */
355 local_t entries; /* entries on this page */
356 unsigned long real_end; /* real end of data */
357 struct buffer_data_page *page; /* Actual data page */
361 * The buffer page counters, write and entries, must be reset
362 * atomically when crossing page boundaries. To synchronize this
363 * update, two counters are inserted into the number. One is
364 * the actual counter for the write position or count on the page.
366 * The other is a counter of updaters. Before an update happens
367 * the update partition of the counter is incremented. This will
368 * allow the updater to update the counter atomically.
370 * The counter is 20 bits, and the state data is 12.
372 #define RB_WRITE_MASK 0xfffff
373 #define RB_WRITE_INTCNT (1 << 20)
375 static void rb_init_page(struct buffer_data_page *bpage)
377 local_set(&bpage->commit, 0);
381 * ring_buffer_page_len - the size of data on the page.
382 * @page: The page to read
384 * Returns the amount of data on the page, including buffer page header.
386 size_t ring_buffer_page_len(void *page)
388 return local_read(&((struct buffer_data_page *)page)->commit)
393 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
396 static void free_buffer_page(struct buffer_page *bpage)
398 free_page((unsigned long)bpage->page);
403 * We need to fit the time_stamp delta into 27 bits.
405 static inline int test_time_stamp(u64 delta)
407 if (delta & TS_DELTA_TEST)
412 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
414 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
415 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
417 int ring_buffer_print_page_header(struct trace_seq *s)
419 struct buffer_data_page field;
422 ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t"
423 "offset:0;\tsize:%u;\tsigned:%u;\n",
424 (unsigned int)sizeof(field.time_stamp),
425 (unsigned int)is_signed_type(u64));
427 ret = trace_seq_printf(s, "\tfield: local_t commit;\t"
428 "offset:%u;\tsize:%u;\tsigned:%u;\n",
429 (unsigned int)offsetof(typeof(field), commit),
430 (unsigned int)sizeof(field.commit),
431 (unsigned int)is_signed_type(long));
433 ret = trace_seq_printf(s, "\tfield: int overwrite;\t"
434 "offset:%u;\tsize:%u;\tsigned:%u;\n",
435 (unsigned int)offsetof(typeof(field), commit),
437 (unsigned int)is_signed_type(long));
439 ret = trace_seq_printf(s, "\tfield: char data;\t"
440 "offset:%u;\tsize:%u;\tsigned:%u;\n",
441 (unsigned int)offsetof(typeof(field), data),
442 (unsigned int)BUF_PAGE_SIZE,
443 (unsigned int)is_signed_type(char));
449 struct irq_work work;
450 wait_queue_head_t waiters;
451 bool waiters_pending;
455 * head_page == tail_page && head == tail then buffer is empty.
457 struct ring_buffer_per_cpu {
459 atomic_t record_disabled;
460 struct ring_buffer *buffer;
461 raw_spinlock_t reader_lock; /* serialize readers */
462 arch_spinlock_t lock;
463 struct lock_class_key lock_key;
464 unsigned int nr_pages;
465 struct list_head *pages;
466 struct buffer_page *head_page; /* read from head */
467 struct buffer_page *tail_page; /* write to tail */
468 struct buffer_page *commit_page; /* committed pages */
469 struct buffer_page *reader_page;
470 unsigned long lost_events;
471 unsigned long last_overrun;
472 local_t entries_bytes;
475 local_t commit_overrun;
476 local_t dropped_events;
480 unsigned long read_bytes;
483 /* ring buffer pages to update, > 0 to add, < 0 to remove */
484 int nr_pages_to_update;
485 struct list_head new_pages; /* new pages to add */
486 struct work_struct update_pages_work;
487 struct completion update_done;
489 struct rb_irq_work irq_work;
495 atomic_t record_disabled;
496 atomic_t resize_disabled;
497 cpumask_var_t cpumask;
499 struct lock_class_key *reader_lock_key;
503 struct ring_buffer_per_cpu **buffers;
505 #ifdef CONFIG_HOTPLUG_CPU
506 struct notifier_block cpu_notify;
510 struct rb_irq_work irq_work;
513 struct ring_buffer_iter {
514 struct ring_buffer_per_cpu *cpu_buffer;
516 struct buffer_page *head_page;
517 struct buffer_page *cache_reader_page;
518 unsigned long cache_read;
523 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
525 * Schedules a delayed work to wake up any task that is blocked on the
526 * ring buffer waiters queue.
528 static void rb_wake_up_waiters(struct irq_work *work)
530 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
532 wake_up_all(&rbwork->waiters);
536 * ring_buffer_wait - wait for input to the ring buffer
537 * @buffer: buffer to wait on
538 * @cpu: the cpu buffer to wait on
540 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
541 * as data is added to any of the @buffer's cpu buffers. Otherwise
542 * it will wait for data to be added to a specific cpu buffer.
544 void ring_buffer_wait(struct ring_buffer *buffer, int cpu)
546 struct ring_buffer_per_cpu *cpu_buffer;
548 struct rb_irq_work *work;
551 * Depending on what the caller is waiting for, either any
552 * data in any cpu buffer, or a specific buffer, put the
553 * caller on the appropriate wait queue.
555 if (cpu == RING_BUFFER_ALL_CPUS)
556 work = &buffer->irq_work;
558 cpu_buffer = buffer->buffers[cpu];
559 work = &cpu_buffer->irq_work;
563 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
566 * The events can happen in critical sections where
567 * checking a work queue can cause deadlocks.
568 * After adding a task to the queue, this flag is set
569 * only to notify events to try to wake up the queue
572 * We don't clear it even if the buffer is no longer
573 * empty. The flag only causes the next event to run
574 * irq_work to do the work queue wake up. The worse
575 * that can happen if we race with !trace_empty() is that
576 * an event will cause an irq_work to try to wake up
579 * There's no reason to protect this flag either, as
580 * the work queue and irq_work logic will do the necessary
581 * synchronization for the wake ups. The only thing
582 * that is necessary is that the wake up happens after
583 * a task has been queued. It's OK for spurious wake ups.
585 work->waiters_pending = true;
587 if ((cpu == RING_BUFFER_ALL_CPUS && ring_buffer_empty(buffer)) ||
588 (cpu != RING_BUFFER_ALL_CPUS && ring_buffer_empty_cpu(buffer, cpu)))
591 finish_wait(&work->waiters, &wait);
595 * ring_buffer_poll_wait - poll on buffer input
596 * @buffer: buffer to wait on
597 * @cpu: the cpu buffer to wait on
598 * @filp: the file descriptor
599 * @poll_table: The poll descriptor
601 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
602 * as data is added to any of the @buffer's cpu buffers. Otherwise
603 * it will wait for data to be added to a specific cpu buffer.
605 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
608 int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
609 struct file *filp, poll_table *poll_table)
611 struct ring_buffer_per_cpu *cpu_buffer;
612 struct rb_irq_work *work;
614 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
615 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
616 return POLLIN | POLLRDNORM;
618 if (cpu == RING_BUFFER_ALL_CPUS)
619 work = &buffer->irq_work;
621 cpu_buffer = buffer->buffers[cpu];
622 work = &cpu_buffer->irq_work;
625 work->waiters_pending = true;
626 poll_wait(filp, &work->waiters, poll_table);
628 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
629 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
630 return POLLIN | POLLRDNORM;
634 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
635 #define RB_WARN_ON(b, cond) \
637 int _____ret = unlikely(cond); \
639 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
640 struct ring_buffer_per_cpu *__b = \
642 atomic_inc(&__b->buffer->record_disabled); \
644 atomic_inc(&b->record_disabled); \
650 /* Up this if you want to test the TIME_EXTENTS and normalization */
651 #define DEBUG_SHIFT 0
653 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
655 /* shift to debug/test normalization and TIME_EXTENTS */
656 return buffer->clock() << DEBUG_SHIFT;
659 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
663 preempt_disable_notrace();
664 time = rb_time_stamp(buffer);
665 preempt_enable_no_resched_notrace();
669 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
671 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
674 /* Just stupid testing the normalize function and deltas */
677 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
680 * Making the ring buffer lockless makes things tricky.
681 * Although writes only happen on the CPU that they are on,
682 * and they only need to worry about interrupts. Reads can
685 * The reader page is always off the ring buffer, but when the
686 * reader finishes with a page, it needs to swap its page with
687 * a new one from the buffer. The reader needs to take from
688 * the head (writes go to the tail). But if a writer is in overwrite
689 * mode and wraps, it must push the head page forward.
691 * Here lies the problem.
693 * The reader must be careful to replace only the head page, and
694 * not another one. As described at the top of the file in the
695 * ASCII art, the reader sets its old page to point to the next
696 * page after head. It then sets the page after head to point to
697 * the old reader page. But if the writer moves the head page
698 * during this operation, the reader could end up with the tail.
700 * We use cmpxchg to help prevent this race. We also do something
701 * special with the page before head. We set the LSB to 1.
703 * When the writer must push the page forward, it will clear the
704 * bit that points to the head page, move the head, and then set
705 * the bit that points to the new head page.
707 * We also don't want an interrupt coming in and moving the head
708 * page on another writer. Thus we use the second LSB to catch
711 * head->list->prev->next bit 1 bit 0
714 * Points to head page 0 1
717 * Note we can not trust the prev pointer of the head page, because:
719 * +----+ +-----+ +-----+
720 * | |------>| T |---X--->| N |
722 * +----+ +-----+ +-----+
725 * +----------| R |----------+ |
729 * Key: ---X--> HEAD flag set in pointer
734 * (see __rb_reserve_next() to see where this happens)
736 * What the above shows is that the reader just swapped out
737 * the reader page with a page in the buffer, but before it
738 * could make the new header point back to the new page added
739 * it was preempted by a writer. The writer moved forward onto
740 * the new page added by the reader and is about to move forward
743 * You can see, it is legitimate for the previous pointer of
744 * the head (or any page) not to point back to itself. But only
748 #define RB_PAGE_NORMAL 0UL
749 #define RB_PAGE_HEAD 1UL
750 #define RB_PAGE_UPDATE 2UL
753 #define RB_FLAG_MASK 3UL
755 /* PAGE_MOVED is not part of the mask */
756 #define RB_PAGE_MOVED 4UL
759 * rb_list_head - remove any bit
761 static struct list_head *rb_list_head(struct list_head *list)
763 unsigned long val = (unsigned long)list;
765 return (struct list_head *)(val & ~RB_FLAG_MASK);
769 * rb_is_head_page - test if the given page is the head page
771 * Because the reader may move the head_page pointer, we can
772 * not trust what the head page is (it may be pointing to
773 * the reader page). But if the next page is a header page,
774 * its flags will be non zero.
777 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
778 struct buffer_page *page, struct list_head *list)
782 val = (unsigned long)list->next;
784 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
785 return RB_PAGE_MOVED;
787 return val & RB_FLAG_MASK;
793 * The unique thing about the reader page, is that, if the
794 * writer is ever on it, the previous pointer never points
795 * back to the reader page.
797 static int rb_is_reader_page(struct buffer_page *page)
799 struct list_head *list = page->list.prev;
801 return rb_list_head(list->next) != &page->list;
805 * rb_set_list_to_head - set a list_head to be pointing to head.
807 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
808 struct list_head *list)
812 ptr = (unsigned long *)&list->next;
813 *ptr |= RB_PAGE_HEAD;
814 *ptr &= ~RB_PAGE_UPDATE;
818 * rb_head_page_activate - sets up head page
820 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
822 struct buffer_page *head;
824 head = cpu_buffer->head_page;
829 * Set the previous list pointer to have the HEAD flag.
831 rb_set_list_to_head(cpu_buffer, head->list.prev);
834 static void rb_list_head_clear(struct list_head *list)
836 unsigned long *ptr = (unsigned long *)&list->next;
838 *ptr &= ~RB_FLAG_MASK;
842 * rb_head_page_dactivate - clears head page ptr (for free list)
845 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
847 struct list_head *hd;
849 /* Go through the whole list and clear any pointers found. */
850 rb_list_head_clear(cpu_buffer->pages);
852 list_for_each(hd, cpu_buffer->pages)
853 rb_list_head_clear(hd);
856 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
857 struct buffer_page *head,
858 struct buffer_page *prev,
859 int old_flag, int new_flag)
861 struct list_head *list;
862 unsigned long val = (unsigned long)&head->list;
867 val &= ~RB_FLAG_MASK;
869 ret = cmpxchg((unsigned long *)&list->next,
870 val | old_flag, val | new_flag);
872 /* check if the reader took the page */
873 if ((ret & ~RB_FLAG_MASK) != val)
874 return RB_PAGE_MOVED;
876 return ret & RB_FLAG_MASK;
879 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
880 struct buffer_page *head,
881 struct buffer_page *prev,
884 return rb_head_page_set(cpu_buffer, head, prev,
885 old_flag, RB_PAGE_UPDATE);
888 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
889 struct buffer_page *head,
890 struct buffer_page *prev,
893 return rb_head_page_set(cpu_buffer, head, prev,
894 old_flag, RB_PAGE_HEAD);
897 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
898 struct buffer_page *head,
899 struct buffer_page *prev,
902 return rb_head_page_set(cpu_buffer, head, prev,
903 old_flag, RB_PAGE_NORMAL);
906 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
907 struct buffer_page **bpage)
909 struct list_head *p = rb_list_head((*bpage)->list.next);
911 *bpage = list_entry(p, struct buffer_page, list);
914 static struct buffer_page *
915 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
917 struct buffer_page *head;
918 struct buffer_page *page;
919 struct list_head *list;
922 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
926 list = cpu_buffer->pages;
927 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
930 page = head = cpu_buffer->head_page;
932 * It is possible that the writer moves the header behind
933 * where we started, and we miss in one loop.
934 * A second loop should grab the header, but we'll do
935 * three loops just because I'm paranoid.
937 for (i = 0; i < 3; i++) {
939 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
940 cpu_buffer->head_page = page;
943 rb_inc_page(cpu_buffer, &page);
944 } while (page != head);
947 RB_WARN_ON(cpu_buffer, 1);
952 static int rb_head_page_replace(struct buffer_page *old,
953 struct buffer_page *new)
955 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
959 val = *ptr & ~RB_FLAG_MASK;
962 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
968 * rb_tail_page_update - move the tail page forward
970 * Returns 1 if moved tail page, 0 if someone else did.
972 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
973 struct buffer_page *tail_page,
974 struct buffer_page *next_page)
976 struct buffer_page *old_tail;
977 unsigned long old_entries;
978 unsigned long old_write;
982 * The tail page now needs to be moved forward.
984 * We need to reset the tail page, but without messing
985 * with possible erasing of data brought in by interrupts
986 * that have moved the tail page and are currently on it.
988 * We add a counter to the write field to denote this.
990 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
991 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
994 * Just make sure we have seen our old_write and synchronize
995 * with any interrupts that come in.
1000 * If the tail page is still the same as what we think
1001 * it is, then it is up to us to update the tail
1004 if (tail_page == cpu_buffer->tail_page) {
1005 /* Zero the write counter */
1006 unsigned long val = old_write & ~RB_WRITE_MASK;
1007 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1010 * This will only succeed if an interrupt did
1011 * not come in and change it. In which case, we
1012 * do not want to modify it.
1014 * We add (void) to let the compiler know that we do not care
1015 * about the return value of these functions. We use the
1016 * cmpxchg to only update if an interrupt did not already
1017 * do it for us. If the cmpxchg fails, we don't care.
1019 (void)local_cmpxchg(&next_page->write, old_write, val);
1020 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1023 * No need to worry about races with clearing out the commit.
1024 * it only can increment when a commit takes place. But that
1025 * only happens in the outer most nested commit.
1027 local_set(&next_page->page->commit, 0);
1029 old_tail = cmpxchg(&cpu_buffer->tail_page,
1030 tail_page, next_page);
1032 if (old_tail == tail_page)
1039 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1040 struct buffer_page *bpage)
1042 unsigned long val = (unsigned long)bpage;
1044 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1051 * rb_check_list - make sure a pointer to a list has the last bits zero
1053 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1054 struct list_head *list)
1056 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1058 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1064 * check_pages - integrity check of buffer pages
1065 * @cpu_buffer: CPU buffer with pages to test
1067 * As a safety measure we check to make sure the data pages have not
1070 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1072 struct list_head *head = cpu_buffer->pages;
1073 struct buffer_page *bpage, *tmp;
1075 /* Reset the head page if it exists */
1076 if (cpu_buffer->head_page)
1077 rb_set_head_page(cpu_buffer);
1079 rb_head_page_deactivate(cpu_buffer);
1081 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1083 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1086 if (rb_check_list(cpu_buffer, head))
1089 list_for_each_entry_safe(bpage, tmp, head, list) {
1090 if (RB_WARN_ON(cpu_buffer,
1091 bpage->list.next->prev != &bpage->list))
1093 if (RB_WARN_ON(cpu_buffer,
1094 bpage->list.prev->next != &bpage->list))
1096 if (rb_check_list(cpu_buffer, &bpage->list))
1100 rb_head_page_activate(cpu_buffer);
1105 static int __rb_allocate_pages(int nr_pages, struct list_head *pages, int cpu)
1108 struct buffer_page *bpage, *tmp;
1110 for (i = 0; i < nr_pages; i++) {
1113 * __GFP_NORETRY flag makes sure that the allocation fails
1114 * gracefully without invoking oom-killer and the system is
1117 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1118 GFP_KERNEL | __GFP_NORETRY,
1123 list_add(&bpage->list, pages);
1125 page = alloc_pages_node(cpu_to_node(cpu),
1126 GFP_KERNEL | __GFP_NORETRY, 0);
1129 bpage->page = page_address(page);
1130 rb_init_page(bpage->page);
1136 list_for_each_entry_safe(bpage, tmp, pages, list) {
1137 list_del_init(&bpage->list);
1138 free_buffer_page(bpage);
1144 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1151 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1155 * The ring buffer page list is a circular list that does not
1156 * start and end with a list head. All page list items point to
1159 cpu_buffer->pages = pages.next;
1162 cpu_buffer->nr_pages = nr_pages;
1164 rb_check_pages(cpu_buffer);
1169 static struct ring_buffer_per_cpu *
1170 rb_allocate_cpu_buffer(struct ring_buffer *buffer, int nr_pages, int cpu)
1172 struct ring_buffer_per_cpu *cpu_buffer;
1173 struct buffer_page *bpage;
1177 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1178 GFP_KERNEL, cpu_to_node(cpu));
1182 cpu_buffer->cpu = cpu;
1183 cpu_buffer->buffer = buffer;
1184 raw_spin_lock_init(&cpu_buffer->reader_lock);
1185 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1186 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1187 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1188 init_completion(&cpu_buffer->update_done);
1189 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1190 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1192 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1193 GFP_KERNEL, cpu_to_node(cpu));
1195 goto fail_free_buffer;
1197 rb_check_bpage(cpu_buffer, bpage);
1199 cpu_buffer->reader_page = bpage;
1200 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1202 goto fail_free_reader;
1203 bpage->page = page_address(page);
1204 rb_init_page(bpage->page);
1206 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1207 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1209 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1211 goto fail_free_reader;
1213 cpu_buffer->head_page
1214 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1215 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1217 rb_head_page_activate(cpu_buffer);
1222 free_buffer_page(cpu_buffer->reader_page);
1229 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1231 struct list_head *head = cpu_buffer->pages;
1232 struct buffer_page *bpage, *tmp;
1234 free_buffer_page(cpu_buffer->reader_page);
1236 rb_head_page_deactivate(cpu_buffer);
1239 list_for_each_entry_safe(bpage, tmp, head, list) {
1240 list_del_init(&bpage->list);
1241 free_buffer_page(bpage);
1243 bpage = list_entry(head, struct buffer_page, list);
1244 free_buffer_page(bpage);
1250 #ifdef CONFIG_HOTPLUG_CPU
1251 static int rb_cpu_notify(struct notifier_block *self,
1252 unsigned long action, void *hcpu);
1256 * ring_buffer_alloc - allocate a new ring_buffer
1257 * @size: the size in bytes per cpu that is needed.
1258 * @flags: attributes to set for the ring buffer.
1260 * Currently the only flag that is available is the RB_FL_OVERWRITE
1261 * flag. This flag means that the buffer will overwrite old data
1262 * when the buffer wraps. If this flag is not set, the buffer will
1263 * drop data when the tail hits the head.
1265 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1266 struct lock_class_key *key)
1268 struct ring_buffer *buffer;
1272 /* keep it in its own cache line */
1273 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1278 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1279 goto fail_free_buffer;
1281 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1282 buffer->flags = flags;
1283 buffer->clock = trace_clock_local;
1284 buffer->reader_lock_key = key;
1286 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1287 init_waitqueue_head(&buffer->irq_work.waiters);
1289 /* need at least two pages */
1294 * In case of non-hotplug cpu, if the ring-buffer is allocated
1295 * in early initcall, it will not be notified of secondary cpus.
1296 * In that off case, we need to allocate for all possible cpus.
1298 #ifdef CONFIG_HOTPLUG_CPU
1300 cpumask_copy(buffer->cpumask, cpu_online_mask);
1302 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1304 buffer->cpus = nr_cpu_ids;
1306 bsize = sizeof(void *) * nr_cpu_ids;
1307 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1309 if (!buffer->buffers)
1310 goto fail_free_cpumask;
1312 for_each_buffer_cpu(buffer, cpu) {
1313 buffer->buffers[cpu] =
1314 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1315 if (!buffer->buffers[cpu])
1316 goto fail_free_buffers;
1319 #ifdef CONFIG_HOTPLUG_CPU
1320 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1321 buffer->cpu_notify.priority = 0;
1322 register_cpu_notifier(&buffer->cpu_notify);
1326 mutex_init(&buffer->mutex);
1331 for_each_buffer_cpu(buffer, cpu) {
1332 if (buffer->buffers[cpu])
1333 rb_free_cpu_buffer(buffer->buffers[cpu]);
1335 kfree(buffer->buffers);
1338 free_cpumask_var(buffer->cpumask);
1345 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1348 * ring_buffer_free - free a ring buffer.
1349 * @buffer: the buffer to free.
1352 ring_buffer_free(struct ring_buffer *buffer)
1358 #ifdef CONFIG_HOTPLUG_CPU
1359 unregister_cpu_notifier(&buffer->cpu_notify);
1362 for_each_buffer_cpu(buffer, cpu)
1363 rb_free_cpu_buffer(buffer->buffers[cpu]);
1367 kfree(buffer->buffers);
1368 free_cpumask_var(buffer->cpumask);
1372 EXPORT_SYMBOL_GPL(ring_buffer_free);
1374 void ring_buffer_set_clock(struct ring_buffer *buffer,
1377 buffer->clock = clock;
1380 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1382 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1384 return local_read(&bpage->entries) & RB_WRITE_MASK;
1387 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1389 return local_read(&bpage->write) & RB_WRITE_MASK;
1393 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned int nr_pages)
1395 struct list_head *tail_page, *to_remove, *next_page;
1396 struct buffer_page *to_remove_page, *tmp_iter_page;
1397 struct buffer_page *last_page, *first_page;
1398 unsigned int nr_removed;
1399 unsigned long head_bit;
1404 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1405 atomic_inc(&cpu_buffer->record_disabled);
1407 * We don't race with the readers since we have acquired the reader
1408 * lock. We also don't race with writers after disabling recording.
1409 * This makes it easy to figure out the first and the last page to be
1410 * removed from the list. We unlink all the pages in between including
1411 * the first and last pages. This is done in a busy loop so that we
1412 * lose the least number of traces.
1413 * The pages are freed after we restart recording and unlock readers.
1415 tail_page = &cpu_buffer->tail_page->list;
1418 * tail page might be on reader page, we remove the next page
1419 * from the ring buffer
1421 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1422 tail_page = rb_list_head(tail_page->next);
1423 to_remove = tail_page;
1425 /* start of pages to remove */
1426 first_page = list_entry(rb_list_head(to_remove->next),
1427 struct buffer_page, list);
1429 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1430 to_remove = rb_list_head(to_remove)->next;
1431 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1434 next_page = rb_list_head(to_remove)->next;
1437 * Now we remove all pages between tail_page and next_page.
1438 * Make sure that we have head_bit value preserved for the
1441 tail_page->next = (struct list_head *)((unsigned long)next_page |
1443 next_page = rb_list_head(next_page);
1444 next_page->prev = tail_page;
1446 /* make sure pages points to a valid page in the ring buffer */
1447 cpu_buffer->pages = next_page;
1449 /* update head page */
1451 cpu_buffer->head_page = list_entry(next_page,
1452 struct buffer_page, list);
1455 * change read pointer to make sure any read iterators reset
1458 cpu_buffer->read = 0;
1460 /* pages are removed, resume tracing and then free the pages */
1461 atomic_dec(&cpu_buffer->record_disabled);
1462 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1464 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1466 /* last buffer page to remove */
1467 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1469 tmp_iter_page = first_page;
1472 to_remove_page = tmp_iter_page;
1473 rb_inc_page(cpu_buffer, &tmp_iter_page);
1475 /* update the counters */
1476 page_entries = rb_page_entries(to_remove_page);
1479 * If something was added to this page, it was full
1480 * since it is not the tail page. So we deduct the
1481 * bytes consumed in ring buffer from here.
1482 * Increment overrun to account for the lost events.
1484 local_add(page_entries, &cpu_buffer->overrun);
1485 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1489 * We have already removed references to this list item, just
1490 * free up the buffer_page and its page
1492 free_buffer_page(to_remove_page);
1495 } while (to_remove_page != last_page);
1497 RB_WARN_ON(cpu_buffer, nr_removed);
1499 return nr_removed == 0;
1503 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1505 struct list_head *pages = &cpu_buffer->new_pages;
1506 int retries, success;
1508 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1510 * We are holding the reader lock, so the reader page won't be swapped
1511 * in the ring buffer. Now we are racing with the writer trying to
1512 * move head page and the tail page.
1513 * We are going to adapt the reader page update process where:
1514 * 1. We first splice the start and end of list of new pages between
1515 * the head page and its previous page.
1516 * 2. We cmpxchg the prev_page->next to point from head page to the
1517 * start of new pages list.
1518 * 3. Finally, we update the head->prev to the end of new list.
1520 * We will try this process 10 times, to make sure that we don't keep
1526 struct list_head *head_page, *prev_page, *r;
1527 struct list_head *last_page, *first_page;
1528 struct list_head *head_page_with_bit;
1530 head_page = &rb_set_head_page(cpu_buffer)->list;
1533 prev_page = head_page->prev;
1535 first_page = pages->next;
1536 last_page = pages->prev;
1538 head_page_with_bit = (struct list_head *)
1539 ((unsigned long)head_page | RB_PAGE_HEAD);
1541 last_page->next = head_page_with_bit;
1542 first_page->prev = prev_page;
1544 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1546 if (r == head_page_with_bit) {
1548 * yay, we replaced the page pointer to our new list,
1549 * now, we just have to update to head page's prev
1550 * pointer to point to end of list
1552 head_page->prev = last_page;
1559 INIT_LIST_HEAD(pages);
1561 * If we weren't successful in adding in new pages, warn and stop
1564 RB_WARN_ON(cpu_buffer, !success);
1565 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1567 /* free pages if they weren't inserted */
1569 struct buffer_page *bpage, *tmp;
1570 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1572 list_del_init(&bpage->list);
1573 free_buffer_page(bpage);
1579 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1583 if (cpu_buffer->nr_pages_to_update > 0)
1584 success = rb_insert_pages(cpu_buffer);
1586 success = rb_remove_pages(cpu_buffer,
1587 -cpu_buffer->nr_pages_to_update);
1590 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1593 static void update_pages_handler(struct work_struct *work)
1595 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1596 struct ring_buffer_per_cpu, update_pages_work);
1597 rb_update_pages(cpu_buffer);
1598 complete(&cpu_buffer->update_done);
1602 * ring_buffer_resize - resize the ring buffer
1603 * @buffer: the buffer to resize.
1604 * @size: the new size.
1606 * Minimum size is 2 * BUF_PAGE_SIZE.
1608 * Returns 0 on success and < 0 on failure.
1610 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1613 struct ring_buffer_per_cpu *cpu_buffer;
1618 * Always succeed at resizing a non-existent buffer:
1623 /* Make sure the requested buffer exists */
1624 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1625 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1628 size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1629 size *= BUF_PAGE_SIZE;
1631 /* we need a minimum of two pages */
1632 if (size < BUF_PAGE_SIZE * 2)
1633 size = BUF_PAGE_SIZE * 2;
1635 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1638 * Don't succeed if resizing is disabled, as a reader might be
1639 * manipulating the ring buffer and is expecting a sane state while
1642 if (atomic_read(&buffer->resize_disabled))
1645 /* prevent another thread from changing buffer sizes */
1646 mutex_lock(&buffer->mutex);
1648 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1649 /* calculate the pages to update */
1650 for_each_buffer_cpu(buffer, cpu) {
1651 cpu_buffer = buffer->buffers[cpu];
1653 cpu_buffer->nr_pages_to_update = nr_pages -
1654 cpu_buffer->nr_pages;
1656 * nothing more to do for removing pages or no update
1658 if (cpu_buffer->nr_pages_to_update <= 0)
1661 * to add pages, make sure all new pages can be
1662 * allocated without receiving ENOMEM
1664 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1665 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1666 &cpu_buffer->new_pages, cpu)) {
1667 /* not enough memory for new pages */
1675 * Fire off all the required work handlers
1676 * We can't schedule on offline CPUs, but it's not necessary
1677 * since we can change their buffer sizes without any race.
1679 for_each_buffer_cpu(buffer, cpu) {
1680 cpu_buffer = buffer->buffers[cpu];
1681 if (!cpu_buffer->nr_pages_to_update)
1684 /* The update must run on the CPU that is being updated. */
1686 if (cpu == smp_processor_id() || !cpu_online(cpu)) {
1687 rb_update_pages(cpu_buffer);
1688 cpu_buffer->nr_pages_to_update = 0;
1691 * Can not disable preemption for schedule_work_on()
1695 schedule_work_on(cpu,
1696 &cpu_buffer->update_pages_work);
1702 /* wait for all the updates to complete */
1703 for_each_buffer_cpu(buffer, cpu) {
1704 cpu_buffer = buffer->buffers[cpu];
1705 if (!cpu_buffer->nr_pages_to_update)
1708 if (cpu_online(cpu))
1709 wait_for_completion(&cpu_buffer->update_done);
1710 cpu_buffer->nr_pages_to_update = 0;
1715 /* Make sure this CPU has been intitialized */
1716 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1719 cpu_buffer = buffer->buffers[cpu_id];
1721 if (nr_pages == cpu_buffer->nr_pages)
1724 cpu_buffer->nr_pages_to_update = nr_pages -
1725 cpu_buffer->nr_pages;
1727 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1728 if (cpu_buffer->nr_pages_to_update > 0 &&
1729 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1730 &cpu_buffer->new_pages, cpu_id)) {
1738 /* The update must run on the CPU that is being updated. */
1739 if (cpu_id == smp_processor_id() || !cpu_online(cpu_id))
1740 rb_update_pages(cpu_buffer);
1743 * Can not disable preemption for schedule_work_on()
1747 schedule_work_on(cpu_id,
1748 &cpu_buffer->update_pages_work);
1749 wait_for_completion(&cpu_buffer->update_done);
1754 cpu_buffer->nr_pages_to_update = 0;
1760 * The ring buffer resize can happen with the ring buffer
1761 * enabled, so that the update disturbs the tracing as little
1762 * as possible. But if the buffer is disabled, we do not need
1763 * to worry about that, and we can take the time to verify
1764 * that the buffer is not corrupt.
1766 if (atomic_read(&buffer->record_disabled)) {
1767 atomic_inc(&buffer->record_disabled);
1769 * Even though the buffer was disabled, we must make sure
1770 * that it is truly disabled before calling rb_check_pages.
1771 * There could have been a race between checking
1772 * record_disable and incrementing it.
1774 synchronize_sched();
1775 for_each_buffer_cpu(buffer, cpu) {
1776 cpu_buffer = buffer->buffers[cpu];
1777 rb_check_pages(cpu_buffer);
1779 atomic_dec(&buffer->record_disabled);
1782 mutex_unlock(&buffer->mutex);
1786 for_each_buffer_cpu(buffer, cpu) {
1787 struct buffer_page *bpage, *tmp;
1789 cpu_buffer = buffer->buffers[cpu];
1790 cpu_buffer->nr_pages_to_update = 0;
1792 if (list_empty(&cpu_buffer->new_pages))
1795 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1797 list_del_init(&bpage->list);
1798 free_buffer_page(bpage);
1801 mutex_unlock(&buffer->mutex);
1804 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1806 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1808 mutex_lock(&buffer->mutex);
1810 buffer->flags |= RB_FL_OVERWRITE;
1812 buffer->flags &= ~RB_FL_OVERWRITE;
1813 mutex_unlock(&buffer->mutex);
1815 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1817 static inline void *
1818 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1820 return bpage->data + index;
1823 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1825 return bpage->page->data + index;
1828 static inline struct ring_buffer_event *
1829 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1831 return __rb_page_index(cpu_buffer->reader_page,
1832 cpu_buffer->reader_page->read);
1835 static inline struct ring_buffer_event *
1836 rb_iter_head_event(struct ring_buffer_iter *iter)
1838 return __rb_page_index(iter->head_page, iter->head);
1841 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1843 return local_read(&bpage->page->commit);
1846 /* Size is determined by what has been committed */
1847 static inline unsigned rb_page_size(struct buffer_page *bpage)
1849 return rb_page_commit(bpage);
1852 static inline unsigned
1853 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1855 return rb_page_commit(cpu_buffer->commit_page);
1858 static inline unsigned
1859 rb_event_index(struct ring_buffer_event *event)
1861 unsigned long addr = (unsigned long)event;
1863 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1867 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1868 struct ring_buffer_event *event)
1870 unsigned long addr = (unsigned long)event;
1871 unsigned long index;
1873 index = rb_event_index(event);
1876 return cpu_buffer->commit_page->page == (void *)addr &&
1877 rb_commit_index(cpu_buffer) == index;
1881 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1883 unsigned long max_count;
1886 * We only race with interrupts and NMIs on this CPU.
1887 * If we own the commit event, then we can commit
1888 * all others that interrupted us, since the interruptions
1889 * are in stack format (they finish before they come
1890 * back to us). This allows us to do a simple loop to
1891 * assign the commit to the tail.
1894 max_count = cpu_buffer->nr_pages * 100;
1896 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1897 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1899 if (RB_WARN_ON(cpu_buffer,
1900 rb_is_reader_page(cpu_buffer->tail_page)))
1902 local_set(&cpu_buffer->commit_page->page->commit,
1903 rb_page_write(cpu_buffer->commit_page));
1904 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1905 cpu_buffer->write_stamp =
1906 cpu_buffer->commit_page->page->time_stamp;
1907 /* add barrier to keep gcc from optimizing too much */
1910 while (rb_commit_index(cpu_buffer) !=
1911 rb_page_write(cpu_buffer->commit_page)) {
1913 local_set(&cpu_buffer->commit_page->page->commit,
1914 rb_page_write(cpu_buffer->commit_page));
1915 RB_WARN_ON(cpu_buffer,
1916 local_read(&cpu_buffer->commit_page->page->commit) &
1921 /* again, keep gcc from optimizing */
1925 * If an interrupt came in just after the first while loop
1926 * and pushed the tail page forward, we will be left with
1927 * a dangling commit that will never go forward.
1929 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1933 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1935 cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1936 cpu_buffer->reader_page->read = 0;
1939 static void rb_inc_iter(struct ring_buffer_iter *iter)
1941 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1944 * The iterator could be on the reader page (it starts there).
1945 * But the head could have moved, since the reader was
1946 * found. Check for this case and assign the iterator
1947 * to the head page instead of next.
1949 if (iter->head_page == cpu_buffer->reader_page)
1950 iter->head_page = rb_set_head_page(cpu_buffer);
1952 rb_inc_page(cpu_buffer, &iter->head_page);
1954 iter->read_stamp = iter->head_page->page->time_stamp;
1958 /* Slow path, do not inline */
1959 static noinline struct ring_buffer_event *
1960 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
1962 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
1964 /* Not the first event on the page? */
1965 if (rb_event_index(event)) {
1966 event->time_delta = delta & TS_MASK;
1967 event->array[0] = delta >> TS_SHIFT;
1969 /* nope, just zero it */
1970 event->time_delta = 0;
1971 event->array[0] = 0;
1974 return skip_time_extend(event);
1978 * rb_update_event - update event type and data
1979 * @event: the even to update
1980 * @type: the type of event
1981 * @length: the size of the event field in the ring buffer
1983 * Update the type and data fields of the event. The length
1984 * is the actual size that is written to the ring buffer,
1985 * and with this, we can determine what to place into the
1989 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
1990 struct ring_buffer_event *event, unsigned length,
1991 int add_timestamp, u64 delta)
1993 /* Only a commit updates the timestamp */
1994 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
1998 * If we need to add a timestamp, then we
1999 * add it to the start of the resevered space.
2001 if (unlikely(add_timestamp)) {
2002 event = rb_add_time_stamp(event, delta);
2003 length -= RB_LEN_TIME_EXTEND;
2007 event->time_delta = delta;
2008 length -= RB_EVNT_HDR_SIZE;
2009 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2010 event->type_len = 0;
2011 event->array[0] = length;
2013 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2017 * rb_handle_head_page - writer hit the head page
2019 * Returns: +1 to retry page
2024 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2025 struct buffer_page *tail_page,
2026 struct buffer_page *next_page)
2028 struct buffer_page *new_head;
2033 entries = rb_page_entries(next_page);
2036 * The hard part is here. We need to move the head
2037 * forward, and protect against both readers on
2038 * other CPUs and writers coming in via interrupts.
2040 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2044 * type can be one of four:
2045 * NORMAL - an interrupt already moved it for us
2046 * HEAD - we are the first to get here.
2047 * UPDATE - we are the interrupt interrupting
2049 * MOVED - a reader on another CPU moved the next
2050 * pointer to its reader page. Give up
2057 * We changed the head to UPDATE, thus
2058 * it is our responsibility to update
2061 local_add(entries, &cpu_buffer->overrun);
2062 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2065 * The entries will be zeroed out when we move the
2069 /* still more to do */
2072 case RB_PAGE_UPDATE:
2074 * This is an interrupt that interrupt the
2075 * previous update. Still more to do.
2078 case RB_PAGE_NORMAL:
2080 * An interrupt came in before the update
2081 * and processed this for us.
2082 * Nothing left to do.
2087 * The reader is on another CPU and just did
2088 * a swap with our next_page.
2093 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2098 * Now that we are here, the old head pointer is
2099 * set to UPDATE. This will keep the reader from
2100 * swapping the head page with the reader page.
2101 * The reader (on another CPU) will spin till
2104 * We just need to protect against interrupts
2105 * doing the job. We will set the next pointer
2106 * to HEAD. After that, we set the old pointer
2107 * to NORMAL, but only if it was HEAD before.
2108 * otherwise we are an interrupt, and only
2109 * want the outer most commit to reset it.
2111 new_head = next_page;
2112 rb_inc_page(cpu_buffer, &new_head);
2114 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2118 * Valid returns are:
2119 * HEAD - an interrupt came in and already set it.
2120 * NORMAL - One of two things:
2121 * 1) We really set it.
2122 * 2) A bunch of interrupts came in and moved
2123 * the page forward again.
2127 case RB_PAGE_NORMAL:
2131 RB_WARN_ON(cpu_buffer, 1);
2136 * It is possible that an interrupt came in,
2137 * set the head up, then more interrupts came in
2138 * and moved it again. When we get back here,
2139 * the page would have been set to NORMAL but we
2140 * just set it back to HEAD.
2142 * How do you detect this? Well, if that happened
2143 * the tail page would have moved.
2145 if (ret == RB_PAGE_NORMAL) {
2147 * If the tail had moved passed next, then we need
2148 * to reset the pointer.
2150 if (cpu_buffer->tail_page != tail_page &&
2151 cpu_buffer->tail_page != next_page)
2152 rb_head_page_set_normal(cpu_buffer, new_head,
2158 * If this was the outer most commit (the one that
2159 * changed the original pointer from HEAD to UPDATE),
2160 * then it is up to us to reset it to NORMAL.
2162 if (type == RB_PAGE_HEAD) {
2163 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2166 if (RB_WARN_ON(cpu_buffer,
2167 ret != RB_PAGE_UPDATE))
2174 static unsigned rb_calculate_event_length(unsigned length)
2176 struct ring_buffer_event event; /* Used only for sizeof array */
2178 /* zero length can cause confusions */
2182 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2183 length += sizeof(event.array[0]);
2185 length += RB_EVNT_HDR_SIZE;
2186 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2192 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2193 struct buffer_page *tail_page,
2194 unsigned long tail, unsigned long length)
2196 struct ring_buffer_event *event;
2199 * Only the event that crossed the page boundary
2200 * must fill the old tail_page with padding.
2202 if (tail >= BUF_PAGE_SIZE) {
2204 * If the page was filled, then we still need
2205 * to update the real_end. Reset it to zero
2206 * and the reader will ignore it.
2208 if (tail == BUF_PAGE_SIZE)
2209 tail_page->real_end = 0;
2211 local_sub(length, &tail_page->write);
2215 event = __rb_page_index(tail_page, tail);
2216 kmemcheck_annotate_bitfield(event, bitfield);
2218 /* account for padding bytes */
2219 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2222 * Save the original length to the meta data.
2223 * This will be used by the reader to add lost event
2226 tail_page->real_end = tail;
2229 * If this event is bigger than the minimum size, then
2230 * we need to be careful that we don't subtract the
2231 * write counter enough to allow another writer to slip
2233 * We put in a discarded commit instead, to make sure
2234 * that this space is not used again.
2236 * If we are less than the minimum size, we don't need to
2239 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2240 /* No room for any events */
2242 /* Mark the rest of the page with padding */
2243 rb_event_set_padding(event);
2245 /* Set the write back to the previous setting */
2246 local_sub(length, &tail_page->write);
2250 /* Put in a discarded event */
2251 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2252 event->type_len = RINGBUF_TYPE_PADDING;
2253 /* time delta must be non zero */
2254 event->time_delta = 1;
2256 /* Set write to end of buffer */
2257 length = (tail + length) - BUF_PAGE_SIZE;
2258 local_sub(length, &tail_page->write);
2262 * This is the slow path, force gcc not to inline it.
2264 static noinline struct ring_buffer_event *
2265 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2266 unsigned long length, unsigned long tail,
2267 struct buffer_page *tail_page, u64 ts)
2269 struct buffer_page *commit_page = cpu_buffer->commit_page;
2270 struct ring_buffer *buffer = cpu_buffer->buffer;
2271 struct buffer_page *next_page;
2274 next_page = tail_page;
2276 rb_inc_page(cpu_buffer, &next_page);
2279 * If for some reason, we had an interrupt storm that made
2280 * it all the way around the buffer, bail, and warn
2283 if (unlikely(next_page == commit_page)) {
2284 local_inc(&cpu_buffer->commit_overrun);
2289 * This is where the fun begins!
2291 * We are fighting against races between a reader that
2292 * could be on another CPU trying to swap its reader
2293 * page with the buffer head.
2295 * We are also fighting against interrupts coming in and
2296 * moving the head or tail on us as well.
2298 * If the next page is the head page then we have filled
2299 * the buffer, unless the commit page is still on the
2302 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2305 * If the commit is not on the reader page, then
2306 * move the header page.
2308 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2310 * If we are not in overwrite mode,
2311 * this is easy, just stop here.
2313 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2314 local_inc(&cpu_buffer->dropped_events);
2318 ret = rb_handle_head_page(cpu_buffer,
2327 * We need to be careful here too. The
2328 * commit page could still be on the reader
2329 * page. We could have a small buffer, and
2330 * have filled up the buffer with events
2331 * from interrupts and such, and wrapped.
2333 * Note, if the tail page is also the on the
2334 * reader_page, we let it move out.
2336 if (unlikely((cpu_buffer->commit_page !=
2337 cpu_buffer->tail_page) &&
2338 (cpu_buffer->commit_page ==
2339 cpu_buffer->reader_page))) {
2340 local_inc(&cpu_buffer->commit_overrun);
2346 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
2349 * Nested commits always have zero deltas, so
2350 * just reread the time stamp
2352 ts = rb_time_stamp(buffer);
2353 next_page->page->time_stamp = ts;
2358 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2360 /* fail and let the caller try again */
2361 return ERR_PTR(-EAGAIN);
2365 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2370 static struct ring_buffer_event *
2371 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2372 unsigned long length, u64 ts,
2373 u64 delta, int add_timestamp)
2375 struct buffer_page *tail_page;
2376 struct ring_buffer_event *event;
2377 unsigned long tail, write;
2380 * If the time delta since the last event is too big to
2381 * hold in the time field of the event, then we append a
2382 * TIME EXTEND event ahead of the data event.
2384 if (unlikely(add_timestamp))
2385 length += RB_LEN_TIME_EXTEND;
2387 tail_page = cpu_buffer->tail_page;
2388 write = local_add_return(length, &tail_page->write);
2390 /* set write to only the index of the write */
2391 write &= RB_WRITE_MASK;
2392 tail = write - length;
2394 /* See if we shot pass the end of this buffer page */
2395 if (unlikely(write > BUF_PAGE_SIZE))
2396 return rb_move_tail(cpu_buffer, length, tail,
2399 /* We reserved something on the buffer */
2401 event = __rb_page_index(tail_page, tail);
2402 kmemcheck_annotate_bitfield(event, bitfield);
2403 rb_update_event(cpu_buffer, event, length, add_timestamp, delta);
2405 local_inc(&tail_page->entries);
2408 * If this is the first commit on the page, then update
2412 tail_page->page->time_stamp = ts;
2414 /* account for these added bytes */
2415 local_add(length, &cpu_buffer->entries_bytes);
2421 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2422 struct ring_buffer_event *event)
2424 unsigned long new_index, old_index;
2425 struct buffer_page *bpage;
2426 unsigned long index;
2429 new_index = rb_event_index(event);
2430 old_index = new_index + rb_event_ts_length(event);
2431 addr = (unsigned long)event;
2434 bpage = cpu_buffer->tail_page;
2436 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2437 unsigned long write_mask =
2438 local_read(&bpage->write) & ~RB_WRITE_MASK;
2439 unsigned long event_length = rb_event_length(event);
2441 * This is on the tail page. It is possible that
2442 * a write could come in and move the tail page
2443 * and write to the next page. That is fine
2444 * because we just shorten what is on this page.
2446 old_index += write_mask;
2447 new_index += write_mask;
2448 index = local_cmpxchg(&bpage->write, old_index, new_index);
2449 if (index == old_index) {
2450 /* update counters */
2451 local_sub(event_length, &cpu_buffer->entries_bytes);
2456 /* could not discard */
2460 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2462 local_inc(&cpu_buffer->committing);
2463 local_inc(&cpu_buffer->commits);
2466 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2468 unsigned long commits;
2470 if (RB_WARN_ON(cpu_buffer,
2471 !local_read(&cpu_buffer->committing)))
2475 commits = local_read(&cpu_buffer->commits);
2476 /* synchronize with interrupts */
2478 if (local_read(&cpu_buffer->committing) == 1)
2479 rb_set_commit_to_write(cpu_buffer);
2481 local_dec(&cpu_buffer->committing);
2483 /* synchronize with interrupts */
2487 * Need to account for interrupts coming in between the
2488 * updating of the commit page and the clearing of the
2489 * committing counter.
2491 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2492 !local_read(&cpu_buffer->committing)) {
2493 local_inc(&cpu_buffer->committing);
2498 static struct ring_buffer_event *
2499 rb_reserve_next_event(struct ring_buffer *buffer,
2500 struct ring_buffer_per_cpu *cpu_buffer,
2501 unsigned long length)
2503 struct ring_buffer_event *event;
2509 rb_start_commit(cpu_buffer);
2511 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2513 * Due to the ability to swap a cpu buffer from a buffer
2514 * it is possible it was swapped before we committed.
2515 * (committing stops a swap). We check for it here and
2516 * if it happened, we have to fail the write.
2519 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2520 local_dec(&cpu_buffer->committing);
2521 local_dec(&cpu_buffer->commits);
2526 length = rb_calculate_event_length(length);
2532 * We allow for interrupts to reenter here and do a trace.
2533 * If one does, it will cause this original code to loop
2534 * back here. Even with heavy interrupts happening, this
2535 * should only happen a few times in a row. If this happens
2536 * 1000 times in a row, there must be either an interrupt
2537 * storm or we have something buggy.
2540 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2543 ts = rb_time_stamp(cpu_buffer->buffer);
2544 diff = ts - cpu_buffer->write_stamp;
2546 /* make sure this diff is calculated here */
2549 /* Did the write stamp get updated already? */
2550 if (likely(ts >= cpu_buffer->write_stamp)) {
2552 if (unlikely(test_time_stamp(delta))) {
2553 int local_clock_stable = 1;
2554 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2555 local_clock_stable = sched_clock_stable;
2557 WARN_ONCE(delta > (1ULL << 59),
2558 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2559 (unsigned long long)delta,
2560 (unsigned long long)ts,
2561 (unsigned long long)cpu_buffer->write_stamp,
2562 local_clock_stable ? "" :
2563 "If you just came from a suspend/resume,\n"
2564 "please switch to the trace global clock:\n"
2565 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2570 event = __rb_reserve_next(cpu_buffer, length, ts,
2571 delta, add_timestamp);
2572 if (unlikely(PTR_ERR(event) == -EAGAIN))
2581 rb_end_commit(cpu_buffer);
2585 #ifdef CONFIG_TRACING
2588 * The lock and unlock are done within a preempt disable section.
2589 * The current_context per_cpu variable can only be modified
2590 * by the current task between lock and unlock. But it can
2591 * be modified more than once via an interrupt. To pass this
2592 * information from the lock to the unlock without having to
2593 * access the 'in_interrupt()' functions again (which do show
2594 * a bit of overhead in something as critical as function tracing,
2595 * we use a bitmask trick.
2597 * bit 0 = NMI context
2598 * bit 1 = IRQ context
2599 * bit 2 = SoftIRQ context
2600 * bit 3 = normal context.
2602 * This works because this is the order of contexts that can
2603 * preempt other contexts. A SoftIRQ never preempts an IRQ
2606 * When the context is determined, the corresponding bit is
2607 * checked and set (if it was set, then a recursion of that context
2610 * On unlock, we need to clear this bit. To do so, just subtract
2611 * 1 from the current_context and AND it to itself.
2615 * 101 & 100 = 100 (clearing bit zero)
2618 * 1010 & 1001 = 1000 (clearing bit 1)
2620 * The least significant bit can be cleared this way, and it
2621 * just so happens that it is the same bit corresponding to
2622 * the current context.
2624 static DEFINE_PER_CPU(unsigned int, current_context);
2626 static __always_inline int trace_recursive_lock(void)
2628 unsigned int val = this_cpu_read(current_context);
2631 if (in_interrupt()) {
2641 if (unlikely(val & (1 << bit)))
2645 this_cpu_write(current_context, val);
2650 static __always_inline void trace_recursive_unlock(void)
2652 unsigned int val = this_cpu_read(current_context);
2655 val &= this_cpu_read(current_context);
2656 this_cpu_write(current_context, val);
2661 #define trace_recursive_lock() (0)
2662 #define trace_recursive_unlock() do { } while (0)
2667 * ring_buffer_lock_reserve - reserve a part of the buffer
2668 * @buffer: the ring buffer to reserve from
2669 * @length: the length of the data to reserve (excluding event header)
2671 * Returns a reseverd event on the ring buffer to copy directly to.
2672 * The user of this interface will need to get the body to write into
2673 * and can use the ring_buffer_event_data() interface.
2675 * The length is the length of the data needed, not the event length
2676 * which also includes the event header.
2678 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2679 * If NULL is returned, then nothing has been allocated or locked.
2681 struct ring_buffer_event *
2682 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2684 struct ring_buffer_per_cpu *cpu_buffer;
2685 struct ring_buffer_event *event;
2688 if (ring_buffer_flags != RB_BUFFERS_ON)
2691 /* If we are tracing schedule, we don't want to recurse */
2692 preempt_disable_notrace();
2694 if (atomic_read(&buffer->record_disabled))
2697 if (trace_recursive_lock())
2700 cpu = raw_smp_processor_id();
2702 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2705 cpu_buffer = buffer->buffers[cpu];
2707 if (atomic_read(&cpu_buffer->record_disabled))
2710 if (length > BUF_MAX_DATA_SIZE)
2713 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2720 trace_recursive_unlock();
2723 preempt_enable_notrace();
2726 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2729 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2730 struct ring_buffer_event *event)
2735 * The event first in the commit queue updates the
2738 if (rb_event_is_commit(cpu_buffer, event)) {
2740 * A commit event that is first on a page
2741 * updates the write timestamp with the page stamp
2743 if (!rb_event_index(event))
2744 cpu_buffer->write_stamp =
2745 cpu_buffer->commit_page->page->time_stamp;
2746 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2747 delta = event->array[0];
2749 delta += event->time_delta;
2750 cpu_buffer->write_stamp += delta;
2752 cpu_buffer->write_stamp += event->time_delta;
2756 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2757 struct ring_buffer_event *event)
2759 local_inc(&cpu_buffer->entries);
2760 rb_update_write_stamp(cpu_buffer, event);
2761 rb_end_commit(cpu_buffer);
2764 static __always_inline void
2765 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2767 if (buffer->irq_work.waiters_pending) {
2768 buffer->irq_work.waiters_pending = false;
2769 /* irq_work_queue() supplies it's own memory barriers */
2770 irq_work_queue(&buffer->irq_work.work);
2773 if (cpu_buffer->irq_work.waiters_pending) {
2774 cpu_buffer->irq_work.waiters_pending = false;
2775 /* irq_work_queue() supplies it's own memory barriers */
2776 irq_work_queue(&cpu_buffer->irq_work.work);
2781 * ring_buffer_unlock_commit - commit a reserved
2782 * @buffer: The buffer to commit to
2783 * @event: The event pointer to commit.
2785 * This commits the data to the ring buffer, and releases any locks held.
2787 * Must be paired with ring_buffer_lock_reserve.
2789 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2790 struct ring_buffer_event *event)
2792 struct ring_buffer_per_cpu *cpu_buffer;
2793 int cpu = raw_smp_processor_id();
2795 cpu_buffer = buffer->buffers[cpu];
2797 rb_commit(cpu_buffer, event);
2799 rb_wakeups(buffer, cpu_buffer);
2801 trace_recursive_unlock();
2803 preempt_enable_notrace();
2807 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2809 static inline void rb_event_discard(struct ring_buffer_event *event)
2811 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2812 event = skip_time_extend(event);
2814 /* array[0] holds the actual length for the discarded event */
2815 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2816 event->type_len = RINGBUF_TYPE_PADDING;
2817 /* time delta must be non zero */
2818 if (!event->time_delta)
2819 event->time_delta = 1;
2823 * Decrement the entries to the page that an event is on.
2824 * The event does not even need to exist, only the pointer
2825 * to the page it is on. This may only be called before the commit
2829 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2830 struct ring_buffer_event *event)
2832 unsigned long addr = (unsigned long)event;
2833 struct buffer_page *bpage = cpu_buffer->commit_page;
2834 struct buffer_page *start;
2838 /* Do the likely case first */
2839 if (likely(bpage->page == (void *)addr)) {
2840 local_dec(&bpage->entries);
2845 * Because the commit page may be on the reader page we
2846 * start with the next page and check the end loop there.
2848 rb_inc_page(cpu_buffer, &bpage);
2851 if (bpage->page == (void *)addr) {
2852 local_dec(&bpage->entries);
2855 rb_inc_page(cpu_buffer, &bpage);
2856 } while (bpage != start);
2858 /* commit not part of this buffer?? */
2859 RB_WARN_ON(cpu_buffer, 1);
2863 * ring_buffer_commit_discard - discard an event that has not been committed
2864 * @buffer: the ring buffer
2865 * @event: non committed event to discard
2867 * Sometimes an event that is in the ring buffer needs to be ignored.
2868 * This function lets the user discard an event in the ring buffer
2869 * and then that event will not be read later.
2871 * This function only works if it is called before the the item has been
2872 * committed. It will try to free the event from the ring buffer
2873 * if another event has not been added behind it.
2875 * If another event has been added behind it, it will set the event
2876 * up as discarded, and perform the commit.
2878 * If this function is called, do not call ring_buffer_unlock_commit on
2881 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2882 struct ring_buffer_event *event)
2884 struct ring_buffer_per_cpu *cpu_buffer;
2887 /* The event is discarded regardless */
2888 rb_event_discard(event);
2890 cpu = smp_processor_id();
2891 cpu_buffer = buffer->buffers[cpu];
2894 * This must only be called if the event has not been
2895 * committed yet. Thus we can assume that preemption
2896 * is still disabled.
2898 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2900 rb_decrement_entry(cpu_buffer, event);
2901 if (rb_try_to_discard(cpu_buffer, event))
2905 * The commit is still visible by the reader, so we
2906 * must still update the timestamp.
2908 rb_update_write_stamp(cpu_buffer, event);
2910 rb_end_commit(cpu_buffer);
2912 trace_recursive_unlock();
2914 preempt_enable_notrace();
2917 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2920 * ring_buffer_write - write data to the buffer without reserving
2921 * @buffer: The ring buffer to write to.
2922 * @length: The length of the data being written (excluding the event header)
2923 * @data: The data to write to the buffer.
2925 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2926 * one function. If you already have the data to write to the buffer, it
2927 * may be easier to simply call this function.
2929 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2930 * and not the length of the event which would hold the header.
2932 int ring_buffer_write(struct ring_buffer *buffer,
2933 unsigned long length,
2936 struct ring_buffer_per_cpu *cpu_buffer;
2937 struct ring_buffer_event *event;
2942 if (ring_buffer_flags != RB_BUFFERS_ON)
2945 preempt_disable_notrace();
2947 if (atomic_read(&buffer->record_disabled))
2950 cpu = raw_smp_processor_id();
2952 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2955 cpu_buffer = buffer->buffers[cpu];
2957 if (atomic_read(&cpu_buffer->record_disabled))
2960 if (length > BUF_MAX_DATA_SIZE)
2963 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2967 body = rb_event_data(event);
2969 memcpy(body, data, length);
2971 rb_commit(cpu_buffer, event);
2973 rb_wakeups(buffer, cpu_buffer);
2977 preempt_enable_notrace();
2981 EXPORT_SYMBOL_GPL(ring_buffer_write);
2983 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
2985 struct buffer_page *reader = cpu_buffer->reader_page;
2986 struct buffer_page *head = rb_set_head_page(cpu_buffer);
2987 struct buffer_page *commit = cpu_buffer->commit_page;
2989 /* In case of error, head will be NULL */
2990 if (unlikely(!head))
2993 return reader->read == rb_page_commit(reader) &&
2994 (commit == reader ||
2996 head->read == rb_page_commit(commit)));
3000 * ring_buffer_record_disable - stop all writes into the buffer
3001 * @buffer: The ring buffer to stop writes to.
3003 * This prevents all writes to the buffer. Any attempt to write
3004 * to the buffer after this will fail and return NULL.
3006 * The caller should call synchronize_sched() after this.
3008 void ring_buffer_record_disable(struct ring_buffer *buffer)
3010 atomic_inc(&buffer->record_disabled);
3012 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3015 * ring_buffer_record_enable - enable writes to the buffer
3016 * @buffer: The ring buffer to enable writes
3018 * Note, multiple disables will need the same number of enables
3019 * to truly enable the writing (much like preempt_disable).
3021 void ring_buffer_record_enable(struct ring_buffer *buffer)
3023 atomic_dec(&buffer->record_disabled);
3025 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3028 * ring_buffer_record_off - stop all writes into the buffer
3029 * @buffer: The ring buffer to stop writes to.
3031 * This prevents all writes to the buffer. Any attempt to write
3032 * to the buffer after this will fail and return NULL.
3034 * This is different than ring_buffer_record_disable() as
3035 * it works like an on/off switch, where as the disable() version
3036 * must be paired with a enable().
3038 void ring_buffer_record_off(struct ring_buffer *buffer)
3041 unsigned int new_rd;
3044 rd = atomic_read(&buffer->record_disabled);
3045 new_rd = rd | RB_BUFFER_OFF;
3046 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3048 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3051 * ring_buffer_record_on - restart writes into the buffer
3052 * @buffer: The ring buffer to start writes to.
3054 * This enables all writes to the buffer that was disabled by
3055 * ring_buffer_record_off().
3057 * This is different than ring_buffer_record_enable() as
3058 * it works like an on/off switch, where as the enable() version
3059 * must be paired with a disable().
3061 void ring_buffer_record_on(struct ring_buffer *buffer)
3064 unsigned int new_rd;
3067 rd = atomic_read(&buffer->record_disabled);
3068 new_rd = rd & ~RB_BUFFER_OFF;
3069 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3071 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3074 * ring_buffer_record_is_on - return true if the ring buffer can write
3075 * @buffer: The ring buffer to see if write is enabled
3077 * Returns true if the ring buffer is in a state that it accepts writes.
3079 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3081 return !atomic_read(&buffer->record_disabled);
3085 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3086 * @buffer: The ring buffer to stop writes to.
3087 * @cpu: The CPU buffer to stop
3089 * This prevents all writes to the buffer. Any attempt to write
3090 * to the buffer after this will fail and return NULL.
3092 * The caller should call synchronize_sched() after this.
3094 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3096 struct ring_buffer_per_cpu *cpu_buffer;
3098 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3101 cpu_buffer = buffer->buffers[cpu];
3102 atomic_inc(&cpu_buffer->record_disabled);
3104 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3107 * ring_buffer_record_enable_cpu - enable writes to the buffer
3108 * @buffer: The ring buffer to enable writes
3109 * @cpu: The CPU to enable.
3111 * Note, multiple disables will need the same number of enables
3112 * to truly enable the writing (much like preempt_disable).
3114 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3116 struct ring_buffer_per_cpu *cpu_buffer;
3118 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3121 cpu_buffer = buffer->buffers[cpu];
3122 atomic_dec(&cpu_buffer->record_disabled);
3124 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3127 * The total entries in the ring buffer is the running counter
3128 * of entries entered into the ring buffer, minus the sum of
3129 * the entries read from the ring buffer and the number of
3130 * entries that were overwritten.
3132 static inline unsigned long
3133 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3135 return local_read(&cpu_buffer->entries) -
3136 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3140 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3141 * @buffer: The ring buffer
3142 * @cpu: The per CPU buffer to read from.
3144 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3146 unsigned long flags;
3147 struct ring_buffer_per_cpu *cpu_buffer;
3148 struct buffer_page *bpage;
3151 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3154 cpu_buffer = buffer->buffers[cpu];
3155 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3157 * if the tail is on reader_page, oldest time stamp is on the reader
3160 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3161 bpage = cpu_buffer->reader_page;
3163 bpage = rb_set_head_page(cpu_buffer);
3165 ret = bpage->page->time_stamp;
3166 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3170 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3173 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3174 * @buffer: The ring buffer
3175 * @cpu: The per CPU buffer to read from.
3177 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3179 struct ring_buffer_per_cpu *cpu_buffer;
3182 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3185 cpu_buffer = buffer->buffers[cpu];
3186 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3190 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3193 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3194 * @buffer: The ring buffer
3195 * @cpu: The per CPU buffer to get the entries from.
3197 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3199 struct ring_buffer_per_cpu *cpu_buffer;
3201 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3204 cpu_buffer = buffer->buffers[cpu];
3206 return rb_num_of_entries(cpu_buffer);
3208 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3211 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3212 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3213 * @buffer: The ring buffer
3214 * @cpu: The per CPU buffer to get the number of overruns from
3216 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3218 struct ring_buffer_per_cpu *cpu_buffer;
3221 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3224 cpu_buffer = buffer->buffers[cpu];
3225 ret = local_read(&cpu_buffer->overrun);
3229 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3232 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3233 * commits failing due to the buffer wrapping around while there are uncommitted
3234 * events, such as during an interrupt storm.
3235 * @buffer: The ring buffer
3236 * @cpu: The per CPU buffer to get the number of overruns from
3239 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3241 struct ring_buffer_per_cpu *cpu_buffer;
3244 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3247 cpu_buffer = buffer->buffers[cpu];
3248 ret = local_read(&cpu_buffer->commit_overrun);
3252 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3255 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3256 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3257 * @buffer: The ring buffer
3258 * @cpu: The per CPU buffer to get the number of overruns from
3261 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3263 struct ring_buffer_per_cpu *cpu_buffer;
3266 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3269 cpu_buffer = buffer->buffers[cpu];
3270 ret = local_read(&cpu_buffer->dropped_events);
3274 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3277 * ring_buffer_read_events_cpu - get the number of events successfully read
3278 * @buffer: The ring buffer
3279 * @cpu: The per CPU buffer to get the number of events read
3282 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3284 struct ring_buffer_per_cpu *cpu_buffer;
3286 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3289 cpu_buffer = buffer->buffers[cpu];
3290 return cpu_buffer->read;
3292 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3295 * ring_buffer_entries - get the number of entries in a buffer
3296 * @buffer: The ring buffer
3298 * Returns the total number of entries in the ring buffer
3301 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3303 struct ring_buffer_per_cpu *cpu_buffer;
3304 unsigned long entries = 0;
3307 /* if you care about this being correct, lock the buffer */
3308 for_each_buffer_cpu(buffer, cpu) {
3309 cpu_buffer = buffer->buffers[cpu];
3310 entries += rb_num_of_entries(cpu_buffer);
3315 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3318 * ring_buffer_overruns - get the number of overruns in buffer
3319 * @buffer: The ring buffer
3321 * Returns the total number of overruns in the ring buffer
3324 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3326 struct ring_buffer_per_cpu *cpu_buffer;
3327 unsigned long overruns = 0;
3330 /* if you care about this being correct, lock the buffer */
3331 for_each_buffer_cpu(buffer, cpu) {
3332 cpu_buffer = buffer->buffers[cpu];
3333 overruns += local_read(&cpu_buffer->overrun);
3338 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3340 static void rb_iter_reset(struct ring_buffer_iter *iter)
3342 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3344 /* Iterator usage is expected to have record disabled */
3345 if (list_empty(&cpu_buffer->reader_page->list)) {
3346 iter->head_page = rb_set_head_page(cpu_buffer);
3347 if (unlikely(!iter->head_page))
3349 iter->head = iter->head_page->read;
3351 iter->head_page = cpu_buffer->reader_page;
3352 iter->head = cpu_buffer->reader_page->read;
3355 iter->read_stamp = cpu_buffer->read_stamp;
3357 iter->read_stamp = iter->head_page->page->time_stamp;
3358 iter->cache_reader_page = cpu_buffer->reader_page;
3359 iter->cache_read = cpu_buffer->read;
3363 * ring_buffer_iter_reset - reset an iterator
3364 * @iter: The iterator to reset
3366 * Resets the iterator, so that it will start from the beginning
3369 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3371 struct ring_buffer_per_cpu *cpu_buffer;
3372 unsigned long flags;
3377 cpu_buffer = iter->cpu_buffer;
3379 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3380 rb_iter_reset(iter);
3381 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3383 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3386 * ring_buffer_iter_empty - check if an iterator has no more to read
3387 * @iter: The iterator to check
3389 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3391 struct ring_buffer_per_cpu *cpu_buffer;
3393 cpu_buffer = iter->cpu_buffer;
3395 return iter->head_page == cpu_buffer->commit_page &&
3396 iter->head == rb_commit_index(cpu_buffer);
3398 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3401 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3402 struct ring_buffer_event *event)
3406 switch (event->type_len) {
3407 case RINGBUF_TYPE_PADDING:
3410 case RINGBUF_TYPE_TIME_EXTEND:
3411 delta = event->array[0];
3413 delta += event->time_delta;
3414 cpu_buffer->read_stamp += delta;
3417 case RINGBUF_TYPE_TIME_STAMP:
3418 /* FIXME: not implemented */
3421 case RINGBUF_TYPE_DATA:
3422 cpu_buffer->read_stamp += event->time_delta;
3432 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3433 struct ring_buffer_event *event)
3437 switch (event->type_len) {
3438 case RINGBUF_TYPE_PADDING:
3441 case RINGBUF_TYPE_TIME_EXTEND:
3442 delta = event->array[0];
3444 delta += event->time_delta;
3445 iter->read_stamp += delta;
3448 case RINGBUF_TYPE_TIME_STAMP:
3449 /* FIXME: not implemented */
3452 case RINGBUF_TYPE_DATA:
3453 iter->read_stamp += event->time_delta;
3462 static struct buffer_page *
3463 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3465 struct buffer_page *reader = NULL;
3466 unsigned long overwrite;
3467 unsigned long flags;
3471 local_irq_save(flags);
3472 arch_spin_lock(&cpu_buffer->lock);
3476 * This should normally only loop twice. But because the
3477 * start of the reader inserts an empty page, it causes
3478 * a case where we will loop three times. There should be no
3479 * reason to loop four times (that I know of).
3481 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3486 reader = cpu_buffer->reader_page;
3488 /* If there's more to read, return this page */
3489 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3492 /* Never should we have an index greater than the size */
3493 if (RB_WARN_ON(cpu_buffer,
3494 cpu_buffer->reader_page->read > rb_page_size(reader)))
3497 /* check if we caught up to the tail */
3499 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3502 /* Don't bother swapping if the ring buffer is empty */
3503 if (rb_num_of_entries(cpu_buffer) == 0)
3507 * Reset the reader page to size zero.
3509 local_set(&cpu_buffer->reader_page->write, 0);
3510 local_set(&cpu_buffer->reader_page->entries, 0);
3511 local_set(&cpu_buffer->reader_page->page->commit, 0);
3512 cpu_buffer->reader_page->real_end = 0;
3516 * Splice the empty reader page into the list around the head.
3518 reader = rb_set_head_page(cpu_buffer);
3521 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3522 cpu_buffer->reader_page->list.prev = reader->list.prev;
3525 * cpu_buffer->pages just needs to point to the buffer, it
3526 * has no specific buffer page to point to. Lets move it out
3527 * of our way so we don't accidentally swap it.
3529 cpu_buffer->pages = reader->list.prev;
3531 /* The reader page will be pointing to the new head */
3532 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3535 * We want to make sure we read the overruns after we set up our
3536 * pointers to the next object. The writer side does a
3537 * cmpxchg to cross pages which acts as the mb on the writer
3538 * side. Note, the reader will constantly fail the swap
3539 * while the writer is updating the pointers, so this
3540 * guarantees that the overwrite recorded here is the one we
3541 * want to compare with the last_overrun.
3544 overwrite = local_read(&(cpu_buffer->overrun));
3547 * Here's the tricky part.
3549 * We need to move the pointer past the header page.
3550 * But we can only do that if a writer is not currently
3551 * moving it. The page before the header page has the
3552 * flag bit '1' set if it is pointing to the page we want.
3553 * but if the writer is in the process of moving it
3554 * than it will be '2' or already moved '0'.
3557 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3560 * If we did not convert it, then we must try again.
3566 * Yeah! We succeeded in replacing the page.
3568 * Now make the new head point back to the reader page.
3570 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3571 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3573 /* Finally update the reader page to the new head */
3574 cpu_buffer->reader_page = reader;
3575 rb_reset_reader_page(cpu_buffer);
3577 if (overwrite != cpu_buffer->last_overrun) {
3578 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3579 cpu_buffer->last_overrun = overwrite;
3585 arch_spin_unlock(&cpu_buffer->lock);
3586 local_irq_restore(flags);
3591 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3593 struct ring_buffer_event *event;
3594 struct buffer_page *reader;
3597 reader = rb_get_reader_page(cpu_buffer);
3599 /* This function should not be called when buffer is empty */
3600 if (RB_WARN_ON(cpu_buffer, !reader))
3603 event = rb_reader_event(cpu_buffer);
3605 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3608 rb_update_read_stamp(cpu_buffer, event);
3610 length = rb_event_length(event);
3611 cpu_buffer->reader_page->read += length;
3614 static void rb_advance_iter(struct ring_buffer_iter *iter)
3616 struct ring_buffer_per_cpu *cpu_buffer;
3617 struct ring_buffer_event *event;
3620 cpu_buffer = iter->cpu_buffer;
3623 * Check if we are at the end of the buffer.
3625 if (iter->head >= rb_page_size(iter->head_page)) {
3626 /* discarded commits can make the page empty */
3627 if (iter->head_page == cpu_buffer->commit_page)
3633 event = rb_iter_head_event(iter);
3635 length = rb_event_length(event);
3638 * This should not be called to advance the header if we are
3639 * at the tail of the buffer.
3641 if (RB_WARN_ON(cpu_buffer,
3642 (iter->head_page == cpu_buffer->commit_page) &&
3643 (iter->head + length > rb_commit_index(cpu_buffer))))
3646 rb_update_iter_read_stamp(iter, event);
3648 iter->head += length;
3650 /* check for end of page padding */
3651 if ((iter->head >= rb_page_size(iter->head_page)) &&
3652 (iter->head_page != cpu_buffer->commit_page))
3656 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3658 return cpu_buffer->lost_events;
3661 static struct ring_buffer_event *
3662 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3663 unsigned long *lost_events)
3665 struct ring_buffer_event *event;
3666 struct buffer_page *reader;
3671 * We repeat when a time extend is encountered.
3672 * Since the time extend is always attached to a data event,
3673 * we should never loop more than once.
3674 * (We never hit the following condition more than twice).
3676 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3679 reader = rb_get_reader_page(cpu_buffer);
3683 event = rb_reader_event(cpu_buffer);
3685 switch (event->type_len) {
3686 case RINGBUF_TYPE_PADDING:
3687 if (rb_null_event(event))
3688 RB_WARN_ON(cpu_buffer, 1);
3690 * Because the writer could be discarding every
3691 * event it creates (which would probably be bad)
3692 * if we were to go back to "again" then we may never
3693 * catch up, and will trigger the warn on, or lock
3694 * the box. Return the padding, and we will release
3695 * the current locks, and try again.
3699 case RINGBUF_TYPE_TIME_EXTEND:
3700 /* Internal data, OK to advance */
3701 rb_advance_reader(cpu_buffer);
3704 case RINGBUF_TYPE_TIME_STAMP:
3705 /* FIXME: not implemented */
3706 rb_advance_reader(cpu_buffer);
3709 case RINGBUF_TYPE_DATA:
3711 *ts = cpu_buffer->read_stamp + event->time_delta;
3712 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3713 cpu_buffer->cpu, ts);
3716 *lost_events = rb_lost_events(cpu_buffer);
3725 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3727 static struct ring_buffer_event *
3728 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3730 struct ring_buffer *buffer;
3731 struct ring_buffer_per_cpu *cpu_buffer;
3732 struct ring_buffer_event *event;
3735 cpu_buffer = iter->cpu_buffer;
3736 buffer = cpu_buffer->buffer;
3739 * Check if someone performed a consuming read to
3740 * the buffer. A consuming read invalidates the iterator
3741 * and we need to reset the iterator in this case.
3743 if (unlikely(iter->cache_read != cpu_buffer->read ||
3744 iter->cache_reader_page != cpu_buffer->reader_page))
3745 rb_iter_reset(iter);
3748 if (ring_buffer_iter_empty(iter))
3752 * We repeat when a time extend is encountered.
3753 * Since the time extend is always attached to a data event,
3754 * we should never loop more than once.
3755 * (We never hit the following condition more than twice).
3757 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3760 if (rb_per_cpu_empty(cpu_buffer))
3763 if (iter->head >= local_read(&iter->head_page->page->commit)) {
3768 event = rb_iter_head_event(iter);
3770 switch (event->type_len) {
3771 case RINGBUF_TYPE_PADDING:
3772 if (rb_null_event(event)) {
3776 rb_advance_iter(iter);
3779 case RINGBUF_TYPE_TIME_EXTEND:
3780 /* Internal data, OK to advance */
3781 rb_advance_iter(iter);
3784 case RINGBUF_TYPE_TIME_STAMP:
3785 /* FIXME: not implemented */
3786 rb_advance_iter(iter);
3789 case RINGBUF_TYPE_DATA:
3791 *ts = iter->read_stamp + event->time_delta;
3792 ring_buffer_normalize_time_stamp(buffer,
3793 cpu_buffer->cpu, ts);
3803 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3805 static inline int rb_ok_to_lock(void)
3808 * If an NMI die dumps out the content of the ring buffer
3809 * do not grab locks. We also permanently disable the ring
3810 * buffer too. A one time deal is all you get from reading
3811 * the ring buffer from an NMI.
3813 if (likely(!in_nmi()))
3816 tracing_off_permanent();
3821 * ring_buffer_peek - peek at the next event to be read
3822 * @buffer: The ring buffer to read
3823 * @cpu: The cpu to peak at
3824 * @ts: The timestamp counter of this event.
3825 * @lost_events: a variable to store if events were lost (may be NULL)
3827 * This will return the event that will be read next, but does
3828 * not consume the data.
3830 struct ring_buffer_event *
3831 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3832 unsigned long *lost_events)
3834 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3835 struct ring_buffer_event *event;
3836 unsigned long flags;
3839 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3842 dolock = rb_ok_to_lock();
3844 local_irq_save(flags);
3846 raw_spin_lock(&cpu_buffer->reader_lock);
3847 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3848 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3849 rb_advance_reader(cpu_buffer);
3851 raw_spin_unlock(&cpu_buffer->reader_lock);
3852 local_irq_restore(flags);
3854 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3861 * ring_buffer_iter_peek - peek at the next event to be read
3862 * @iter: The ring buffer iterator
3863 * @ts: The timestamp counter of this event.
3865 * This will return the event that will be read next, but does
3866 * not increment the iterator.
3868 struct ring_buffer_event *
3869 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3871 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3872 struct ring_buffer_event *event;
3873 unsigned long flags;
3876 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3877 event = rb_iter_peek(iter, ts);
3878 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3880 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3887 * ring_buffer_consume - return an event and consume it
3888 * @buffer: The ring buffer to get the next event from
3889 * @cpu: the cpu to read the buffer from
3890 * @ts: a variable to store the timestamp (may be NULL)
3891 * @lost_events: a variable to store if events were lost (may be NULL)
3893 * Returns the next event in the ring buffer, and that event is consumed.
3894 * Meaning, that sequential reads will keep returning a different event,
3895 * and eventually empty the ring buffer if the producer is slower.
3897 struct ring_buffer_event *
3898 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3899 unsigned long *lost_events)
3901 struct ring_buffer_per_cpu *cpu_buffer;
3902 struct ring_buffer_event *event = NULL;
3903 unsigned long flags;
3906 dolock = rb_ok_to_lock();
3909 /* might be called in atomic */
3912 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3915 cpu_buffer = buffer->buffers[cpu];
3916 local_irq_save(flags);
3918 raw_spin_lock(&cpu_buffer->reader_lock);
3920 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3922 cpu_buffer->lost_events = 0;
3923 rb_advance_reader(cpu_buffer);
3927 raw_spin_unlock(&cpu_buffer->reader_lock);
3928 local_irq_restore(flags);
3933 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3938 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3941 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3942 * @buffer: The ring buffer to read from
3943 * @cpu: The cpu buffer to iterate over
3945 * This performs the initial preparations necessary to iterate
3946 * through the buffer. Memory is allocated, buffer recording
3947 * is disabled, and the iterator pointer is returned to the caller.
3949 * Disabling buffer recordng prevents the reading from being
3950 * corrupted. This is not a consuming read, so a producer is not
3953 * After a sequence of ring_buffer_read_prepare calls, the user is
3954 * expected to make at least one call to ring_buffer_prepare_sync.
3955 * Afterwards, ring_buffer_read_start is invoked to get things going
3958 * This overall must be paired with ring_buffer_finish.
3960 struct ring_buffer_iter *
3961 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3963 struct ring_buffer_per_cpu *cpu_buffer;
3964 struct ring_buffer_iter *iter;
3966 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3969 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3973 cpu_buffer = buffer->buffers[cpu];
3975 iter->cpu_buffer = cpu_buffer;
3977 atomic_inc(&buffer->resize_disabled);
3978 atomic_inc(&cpu_buffer->record_disabled);
3982 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
3985 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
3987 * All previously invoked ring_buffer_read_prepare calls to prepare
3988 * iterators will be synchronized. Afterwards, read_buffer_read_start
3989 * calls on those iterators are allowed.
3992 ring_buffer_read_prepare_sync(void)
3994 synchronize_sched();
3996 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
3999 * ring_buffer_read_start - start a non consuming read of the buffer
4000 * @iter: The iterator returned by ring_buffer_read_prepare
4002 * This finalizes the startup of an iteration through the buffer.
4003 * The iterator comes from a call to ring_buffer_read_prepare and
4004 * an intervening ring_buffer_read_prepare_sync must have been
4007 * Must be paired with ring_buffer_finish.
4010 ring_buffer_read_start(struct ring_buffer_iter *iter)
4012 struct ring_buffer_per_cpu *cpu_buffer;
4013 unsigned long flags;
4018 cpu_buffer = iter->cpu_buffer;
4020 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4021 arch_spin_lock(&cpu_buffer->lock);
4022 rb_iter_reset(iter);
4023 arch_spin_unlock(&cpu_buffer->lock);
4024 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4026 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4029 * ring_buffer_finish - finish reading the iterator of the buffer
4030 * @iter: The iterator retrieved by ring_buffer_start
4032 * This re-enables the recording to the buffer, and frees the
4036 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4038 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4039 unsigned long flags;
4042 * Ring buffer is disabled from recording, here's a good place
4043 * to check the integrity of the ring buffer.
4044 * Must prevent readers from trying to read, as the check
4045 * clears the HEAD page and readers require it.
4047 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4048 rb_check_pages(cpu_buffer);
4049 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4051 atomic_dec(&cpu_buffer->record_disabled);
4052 atomic_dec(&cpu_buffer->buffer->resize_disabled);
4055 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4058 * ring_buffer_read - read the next item in the ring buffer by the iterator
4059 * @iter: The ring buffer iterator
4060 * @ts: The time stamp of the event read.
4062 * This reads the next event in the ring buffer and increments the iterator.
4064 struct ring_buffer_event *
4065 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4067 struct ring_buffer_event *event;
4068 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4069 unsigned long flags;
4071 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4073 event = rb_iter_peek(iter, ts);
4077 if (event->type_len == RINGBUF_TYPE_PADDING)
4080 rb_advance_iter(iter);
4082 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4086 EXPORT_SYMBOL_GPL(ring_buffer_read);
4089 * ring_buffer_size - return the size of the ring buffer (in bytes)
4090 * @buffer: The ring buffer.
4092 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4095 * Earlier, this method returned
4096 * BUF_PAGE_SIZE * buffer->nr_pages
4097 * Since the nr_pages field is now removed, we have converted this to
4098 * return the per cpu buffer value.
4100 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4103 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4105 EXPORT_SYMBOL_GPL(ring_buffer_size);
4108 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4110 rb_head_page_deactivate(cpu_buffer);
4112 cpu_buffer->head_page
4113 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4114 local_set(&cpu_buffer->head_page->write, 0);
4115 local_set(&cpu_buffer->head_page->entries, 0);
4116 local_set(&cpu_buffer->head_page->page->commit, 0);
4118 cpu_buffer->head_page->read = 0;
4120 cpu_buffer->tail_page = cpu_buffer->head_page;
4121 cpu_buffer->commit_page = cpu_buffer->head_page;
4123 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4124 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4125 local_set(&cpu_buffer->reader_page->write, 0);
4126 local_set(&cpu_buffer->reader_page->entries, 0);
4127 local_set(&cpu_buffer->reader_page->page->commit, 0);
4128 cpu_buffer->reader_page->read = 0;
4130 local_set(&cpu_buffer->entries_bytes, 0);
4131 local_set(&cpu_buffer->overrun, 0);
4132 local_set(&cpu_buffer->commit_overrun, 0);
4133 local_set(&cpu_buffer->dropped_events, 0);
4134 local_set(&cpu_buffer->entries, 0);
4135 local_set(&cpu_buffer->committing, 0);
4136 local_set(&cpu_buffer->commits, 0);
4137 cpu_buffer->read = 0;
4138 cpu_buffer->read_bytes = 0;
4140 cpu_buffer->write_stamp = 0;
4141 cpu_buffer->read_stamp = 0;
4143 cpu_buffer->lost_events = 0;
4144 cpu_buffer->last_overrun = 0;
4146 rb_head_page_activate(cpu_buffer);
4150 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4151 * @buffer: The ring buffer to reset a per cpu buffer of
4152 * @cpu: The CPU buffer to be reset
4154 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4156 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4157 unsigned long flags;
4159 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4162 atomic_inc(&buffer->resize_disabled);
4163 atomic_inc(&cpu_buffer->record_disabled);
4165 /* Make sure all commits have finished */
4166 synchronize_sched();
4168 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4170 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4173 arch_spin_lock(&cpu_buffer->lock);
4175 rb_reset_cpu(cpu_buffer);
4177 arch_spin_unlock(&cpu_buffer->lock);
4180 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4182 atomic_dec(&cpu_buffer->record_disabled);
4183 atomic_dec(&buffer->resize_disabled);
4185 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4188 * ring_buffer_reset - reset a ring buffer
4189 * @buffer: The ring buffer to reset all cpu buffers
4191 void ring_buffer_reset(struct ring_buffer *buffer)
4195 for_each_buffer_cpu(buffer, cpu)
4196 ring_buffer_reset_cpu(buffer, cpu);
4198 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4201 * rind_buffer_empty - is the ring buffer empty?
4202 * @buffer: The ring buffer to test
4204 int ring_buffer_empty(struct ring_buffer *buffer)
4206 struct ring_buffer_per_cpu *cpu_buffer;
4207 unsigned long flags;
4212 dolock = rb_ok_to_lock();
4214 /* yes this is racy, but if you don't like the race, lock the buffer */
4215 for_each_buffer_cpu(buffer, cpu) {
4216 cpu_buffer = buffer->buffers[cpu];
4217 local_irq_save(flags);
4219 raw_spin_lock(&cpu_buffer->reader_lock);
4220 ret = rb_per_cpu_empty(cpu_buffer);
4222 raw_spin_unlock(&cpu_buffer->reader_lock);
4223 local_irq_restore(flags);
4231 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4234 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4235 * @buffer: The ring buffer
4236 * @cpu: The CPU buffer to test
4238 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4240 struct ring_buffer_per_cpu *cpu_buffer;
4241 unsigned long flags;
4245 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4248 dolock = rb_ok_to_lock();
4250 cpu_buffer = buffer->buffers[cpu];
4251 local_irq_save(flags);
4253 raw_spin_lock(&cpu_buffer->reader_lock);
4254 ret = rb_per_cpu_empty(cpu_buffer);
4256 raw_spin_unlock(&cpu_buffer->reader_lock);
4257 local_irq_restore(flags);
4261 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4263 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4265 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4266 * @buffer_a: One buffer to swap with
4267 * @buffer_b: The other buffer to swap with
4269 * This function is useful for tracers that want to take a "snapshot"
4270 * of a CPU buffer and has another back up buffer lying around.
4271 * it is expected that the tracer handles the cpu buffer not being
4272 * used at the moment.
4274 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4275 struct ring_buffer *buffer_b, int cpu)
4277 struct ring_buffer_per_cpu *cpu_buffer_a;
4278 struct ring_buffer_per_cpu *cpu_buffer_b;
4281 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4282 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4285 cpu_buffer_a = buffer_a->buffers[cpu];
4286 cpu_buffer_b = buffer_b->buffers[cpu];
4288 /* At least make sure the two buffers are somewhat the same */
4289 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4294 if (ring_buffer_flags != RB_BUFFERS_ON)
4297 if (atomic_read(&buffer_a->record_disabled))
4300 if (atomic_read(&buffer_b->record_disabled))
4303 if (atomic_read(&cpu_buffer_a->record_disabled))
4306 if (atomic_read(&cpu_buffer_b->record_disabled))
4310 * We can't do a synchronize_sched here because this
4311 * function can be called in atomic context.
4312 * Normally this will be called from the same CPU as cpu.
4313 * If not it's up to the caller to protect this.
4315 atomic_inc(&cpu_buffer_a->record_disabled);
4316 atomic_inc(&cpu_buffer_b->record_disabled);
4319 if (local_read(&cpu_buffer_a->committing))
4321 if (local_read(&cpu_buffer_b->committing))
4324 buffer_a->buffers[cpu] = cpu_buffer_b;
4325 buffer_b->buffers[cpu] = cpu_buffer_a;
4327 cpu_buffer_b->buffer = buffer_a;
4328 cpu_buffer_a->buffer = buffer_b;
4333 atomic_dec(&cpu_buffer_a->record_disabled);
4334 atomic_dec(&cpu_buffer_b->record_disabled);
4338 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4339 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4342 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4343 * @buffer: the buffer to allocate for.
4345 * This function is used in conjunction with ring_buffer_read_page.
4346 * When reading a full page from the ring buffer, these functions
4347 * can be used to speed up the process. The calling function should
4348 * allocate a few pages first with this function. Then when it
4349 * needs to get pages from the ring buffer, it passes the result
4350 * of this function into ring_buffer_read_page, which will swap
4351 * the page that was allocated, with the read page of the buffer.
4354 * The page allocated, or NULL on error.
4356 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4358 struct buffer_data_page *bpage;
4361 page = alloc_pages_node(cpu_to_node(cpu),
4362 GFP_KERNEL | __GFP_NORETRY, 0);
4366 bpage = page_address(page);
4368 rb_init_page(bpage);
4372 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4375 * ring_buffer_free_read_page - free an allocated read page
4376 * @buffer: the buffer the page was allocate for
4377 * @data: the page to free
4379 * Free a page allocated from ring_buffer_alloc_read_page.
4381 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4383 free_page((unsigned long)data);
4385 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4388 * ring_buffer_read_page - extract a page from the ring buffer
4389 * @buffer: buffer to extract from
4390 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4391 * @len: amount to extract
4392 * @cpu: the cpu of the buffer to extract
4393 * @full: should the extraction only happen when the page is full.
4395 * This function will pull out a page from the ring buffer and consume it.
4396 * @data_page must be the address of the variable that was returned
4397 * from ring_buffer_alloc_read_page. This is because the page might be used
4398 * to swap with a page in the ring buffer.
4401 * rpage = ring_buffer_alloc_read_page(buffer);
4404 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4406 * process_page(rpage, ret);
4408 * When @full is set, the function will not return true unless
4409 * the writer is off the reader page.
4411 * Note: it is up to the calling functions to handle sleeps and wakeups.
4412 * The ring buffer can be used anywhere in the kernel and can not
4413 * blindly call wake_up. The layer that uses the ring buffer must be
4414 * responsible for that.
4417 * >=0 if data has been transferred, returns the offset of consumed data.
4418 * <0 if no data has been transferred.
4420 int ring_buffer_read_page(struct ring_buffer *buffer,
4421 void **data_page, size_t len, int cpu, int full)
4423 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4424 struct ring_buffer_event *event;
4425 struct buffer_data_page *bpage;
4426 struct buffer_page *reader;
4427 unsigned long missed_events;
4428 unsigned long flags;
4429 unsigned int commit;
4434 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4438 * If len is not big enough to hold the page header, then
4439 * we can not copy anything.
4441 if (len <= BUF_PAGE_HDR_SIZE)
4444 len -= BUF_PAGE_HDR_SIZE;
4453 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4455 reader = rb_get_reader_page(cpu_buffer);
4459 event = rb_reader_event(cpu_buffer);
4461 read = reader->read;
4462 commit = rb_page_commit(reader);
4464 /* Check if any events were dropped */
4465 missed_events = cpu_buffer->lost_events;
4468 * If this page has been partially read or
4469 * if len is not big enough to read the rest of the page or
4470 * a writer is still on the page, then
4471 * we must copy the data from the page to the buffer.
4472 * Otherwise, we can simply swap the page with the one passed in.
4474 if (read || (len < (commit - read)) ||
4475 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4476 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4477 unsigned int rpos = read;
4478 unsigned int pos = 0;
4484 if (len > (commit - read))
4485 len = (commit - read);
4487 /* Always keep the time extend and data together */
4488 size = rb_event_ts_length(event);
4493 /* save the current timestamp, since the user will need it */
4494 save_timestamp = cpu_buffer->read_stamp;
4496 /* Need to copy one event at a time */
4498 /* We need the size of one event, because
4499 * rb_advance_reader only advances by one event,
4500 * whereas rb_event_ts_length may include the size of
4501 * one or two events.
4502 * We have already ensured there's enough space if this
4503 * is a time extend. */
4504 size = rb_event_length(event);
4505 memcpy(bpage->data + pos, rpage->data + rpos, size);
4509 rb_advance_reader(cpu_buffer);
4510 rpos = reader->read;
4516 event = rb_reader_event(cpu_buffer);
4517 /* Always keep the time extend and data together */
4518 size = rb_event_ts_length(event);
4519 } while (len >= size);
4522 local_set(&bpage->commit, pos);
4523 bpage->time_stamp = save_timestamp;
4525 /* we copied everything to the beginning */
4528 /* update the entry counter */
4529 cpu_buffer->read += rb_page_entries(reader);
4530 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4532 /* swap the pages */
4533 rb_init_page(bpage);
4534 bpage = reader->page;
4535 reader->page = *data_page;
4536 local_set(&reader->write, 0);
4537 local_set(&reader->entries, 0);
4542 * Use the real_end for the data size,
4543 * This gives us a chance to store the lost events
4546 if (reader->real_end)
4547 local_set(&bpage->commit, reader->real_end);
4551 cpu_buffer->lost_events = 0;
4553 commit = local_read(&bpage->commit);
4555 * Set a flag in the commit field if we lost events
4557 if (missed_events) {
4558 /* If there is room at the end of the page to save the
4559 * missed events, then record it there.
4561 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4562 memcpy(&bpage->data[commit], &missed_events,
4563 sizeof(missed_events));
4564 local_add(RB_MISSED_STORED, &bpage->commit);
4565 commit += sizeof(missed_events);
4567 local_add(RB_MISSED_EVENTS, &bpage->commit);
4571 * This page may be off to user land. Zero it out here.
4573 if (commit < BUF_PAGE_SIZE)
4574 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4577 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4582 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4584 #ifdef CONFIG_HOTPLUG_CPU
4585 static int rb_cpu_notify(struct notifier_block *self,
4586 unsigned long action, void *hcpu)
4588 struct ring_buffer *buffer =
4589 container_of(self, struct ring_buffer, cpu_notify);
4590 long cpu = (long)hcpu;
4591 int cpu_i, nr_pages_same;
4592 unsigned int nr_pages;
4595 case CPU_UP_PREPARE:
4596 case CPU_UP_PREPARE_FROZEN:
4597 if (cpumask_test_cpu(cpu, buffer->cpumask))
4602 /* check if all cpu sizes are same */
4603 for_each_buffer_cpu(buffer, cpu_i) {
4604 /* fill in the size from first enabled cpu */
4606 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4607 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4612 /* allocate minimum pages, user can later expand it */
4615 buffer->buffers[cpu] =
4616 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4617 if (!buffer->buffers[cpu]) {
4618 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4623 cpumask_set_cpu(cpu, buffer->cpumask);
4625 case CPU_DOWN_PREPARE:
4626 case CPU_DOWN_PREPARE_FROZEN:
4629 * If we were to free the buffer, then the user would
4630 * lose any trace that was in the buffer.
4640 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4642 * This is a basic integrity check of the ring buffer.
4643 * Late in the boot cycle this test will run when configured in.
4644 * It will kick off a thread per CPU that will go into a loop
4645 * writing to the per cpu ring buffer various sizes of data.
4646 * Some of the data will be large items, some small.
4648 * Another thread is created that goes into a spin, sending out
4649 * IPIs to the other CPUs to also write into the ring buffer.
4650 * this is to test the nesting ability of the buffer.
4652 * Basic stats are recorded and reported. If something in the
4653 * ring buffer should happen that's not expected, a big warning
4654 * is displayed and all ring buffers are disabled.
4656 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4658 struct rb_test_data {
4659 struct ring_buffer *buffer;
4660 unsigned long events;
4661 unsigned long bytes_written;
4662 unsigned long bytes_alloc;
4663 unsigned long bytes_dropped;
4664 unsigned long events_nested;
4665 unsigned long bytes_written_nested;
4666 unsigned long bytes_alloc_nested;
4667 unsigned long bytes_dropped_nested;
4668 int min_size_nested;
4669 int max_size_nested;
4676 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4679 #define RB_TEST_BUFFER_SIZE 1048576
4681 static char rb_string[] __initdata =
4682 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4683 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4684 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4686 static bool rb_test_started __initdata;
4693 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4695 struct ring_buffer_event *event;
4696 struct rb_item *item;
4703 /* Have nested writes different that what is written */
4704 cnt = data->cnt + (nested ? 27 : 0);
4706 /* Multiply cnt by ~e, to make some unique increment */
4707 size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4709 len = size + sizeof(struct rb_item);
4711 started = rb_test_started;
4712 /* read rb_test_started before checking buffer enabled */
4715 event = ring_buffer_lock_reserve(data->buffer, len);
4717 /* Ignore dropped events before test starts. */
4720 data->bytes_dropped += len;
4722 data->bytes_dropped_nested += len;
4727 event_len = ring_buffer_event_length(event);
4729 if (RB_WARN_ON(data->buffer, event_len < len))
4732 item = ring_buffer_event_data(event);
4734 memcpy(item->str, rb_string, size);
4737 data->bytes_alloc_nested += event_len;
4738 data->bytes_written_nested += len;
4739 data->events_nested++;
4740 if (!data->min_size_nested || len < data->min_size_nested)
4741 data->min_size_nested = len;
4742 if (len > data->max_size_nested)
4743 data->max_size_nested = len;
4745 data->bytes_alloc += event_len;
4746 data->bytes_written += len;
4748 if (!data->min_size || len < data->min_size)
4749 data->max_size = len;
4750 if (len > data->max_size)
4751 data->max_size = len;
4755 ring_buffer_unlock_commit(data->buffer, event);
4760 static __init int rb_test(void *arg)
4762 struct rb_test_data *data = arg;
4764 while (!kthread_should_stop()) {
4765 rb_write_something(data, false);
4768 set_current_state(TASK_INTERRUPTIBLE);
4769 /* Now sleep between a min of 100-300us and a max of 1ms */
4770 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4776 static __init void rb_ipi(void *ignore)
4778 struct rb_test_data *data;
4779 int cpu = smp_processor_id();
4781 data = &rb_data[cpu];
4782 rb_write_something(data, true);
4785 static __init int rb_hammer_test(void *arg)
4787 while (!kthread_should_stop()) {
4789 /* Send an IPI to all cpus to write data! */
4790 smp_call_function(rb_ipi, NULL, 1);
4791 /* No sleep, but for non preempt, let others run */
4798 static __init int test_ringbuffer(void)
4800 struct task_struct *rb_hammer;
4801 struct ring_buffer *buffer;
4805 pr_info("Running ring buffer tests...\n");
4807 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4808 if (WARN_ON(!buffer))
4811 /* Disable buffer so that threads can't write to it yet */
4812 ring_buffer_record_off(buffer);
4814 for_each_online_cpu(cpu) {
4815 rb_data[cpu].buffer = buffer;
4816 rb_data[cpu].cpu = cpu;
4817 rb_data[cpu].cnt = cpu;
4818 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4819 "rbtester/%d", cpu);
4820 if (WARN_ON(!rb_threads[cpu])) {
4821 pr_cont("FAILED\n");
4826 kthread_bind(rb_threads[cpu], cpu);
4827 wake_up_process(rb_threads[cpu]);
4830 /* Now create the rb hammer! */
4831 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4832 if (WARN_ON(!rb_hammer)) {
4833 pr_cont("FAILED\n");
4838 ring_buffer_record_on(buffer);
4840 * Show buffer is enabled before setting rb_test_started.
4841 * Yes there's a small race window where events could be
4842 * dropped and the thread wont catch it. But when a ring
4843 * buffer gets enabled, there will always be some kind of
4844 * delay before other CPUs see it. Thus, we don't care about
4845 * those dropped events. We care about events dropped after
4846 * the threads see that the buffer is active.
4849 rb_test_started = true;
4851 set_current_state(TASK_INTERRUPTIBLE);
4852 /* Just run for 10 seconds */;
4853 schedule_timeout(10 * HZ);
4855 kthread_stop(rb_hammer);
4858 for_each_online_cpu(cpu) {
4859 if (!rb_threads[cpu])
4861 kthread_stop(rb_threads[cpu]);
4864 ring_buffer_free(buffer);
4869 pr_info("finished\n");
4870 for_each_online_cpu(cpu) {
4871 struct ring_buffer_event *event;
4872 struct rb_test_data *data = &rb_data[cpu];
4873 struct rb_item *item;
4874 unsigned long total_events;
4875 unsigned long total_dropped;
4876 unsigned long total_written;
4877 unsigned long total_alloc;
4878 unsigned long total_read = 0;
4879 unsigned long total_size = 0;
4880 unsigned long total_len = 0;
4881 unsigned long total_lost = 0;
4884 int small_event_size;
4888 total_events = data->events + data->events_nested;
4889 total_written = data->bytes_written + data->bytes_written_nested;
4890 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
4891 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
4893 big_event_size = data->max_size + data->max_size_nested;
4894 small_event_size = data->min_size + data->min_size_nested;
4896 pr_info("CPU %d:\n", cpu);
4897 pr_info(" events: %ld\n", total_events);
4898 pr_info(" dropped bytes: %ld\n", total_dropped);
4899 pr_info(" alloced bytes: %ld\n", total_alloc);
4900 pr_info(" written bytes: %ld\n", total_written);
4901 pr_info(" biggest event: %d\n", big_event_size);
4902 pr_info(" smallest event: %d\n", small_event_size);
4904 if (RB_WARN_ON(buffer, total_dropped))
4909 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
4911 item = ring_buffer_event_data(event);
4912 total_len += ring_buffer_event_length(event);
4913 total_size += item->size + sizeof(struct rb_item);
4914 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
4915 pr_info("FAILED!\n");
4916 pr_info("buffer had: %.*s\n", item->size, item->str);
4917 pr_info("expected: %.*s\n", item->size, rb_string);
4918 RB_WARN_ON(buffer, 1);
4929 pr_info(" read events: %ld\n", total_read);
4930 pr_info(" lost events: %ld\n", total_lost);
4931 pr_info(" total events: %ld\n", total_lost + total_read);
4932 pr_info(" recorded len bytes: %ld\n", total_len);
4933 pr_info(" recorded size bytes: %ld\n", total_size);
4935 pr_info(" With dropped events, record len and size may not match\n"
4936 " alloced and written from above\n");
4938 if (RB_WARN_ON(buffer, total_len != total_alloc ||
4939 total_size != total_written))
4942 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
4948 pr_info("Ring buffer PASSED!\n");
4950 ring_buffer_free(buffer);
4954 late_initcall(test_ringbuffer);
4955 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */