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_puts(s, "# compressed entry header\n");
40 ret = trace_seq_puts(s, "\ttype_len : 5 bits\n");
41 ret = trace_seq_puts(s, "\ttime_delta : 27 bits\n");
42 ret = trace_seq_puts(s, "\tarray : 32 bits\n");
43 ret = trace_seq_putc(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 #ifndef CONFIG_HAVE_64BIT_ALIGNED_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 RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
194 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
195 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
198 RB_LEN_TIME_EXTEND = 8,
199 RB_LEN_TIME_STAMP = 16,
202 #define skip_time_extend(event) \
203 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
205 static inline int rb_null_event(struct ring_buffer_event *event)
207 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
210 static void rb_event_set_padding(struct ring_buffer_event *event)
212 /* padding has a NULL time_delta */
213 event->type_len = RINGBUF_TYPE_PADDING;
214 event->time_delta = 0;
218 rb_event_data_length(struct ring_buffer_event *event)
223 length = event->type_len * RB_ALIGNMENT;
225 length = event->array[0];
226 return length + RB_EVNT_HDR_SIZE;
230 * Return the length of the given event. Will return
231 * the length of the time extend if the event is a
234 static inline unsigned
235 rb_event_length(struct ring_buffer_event *event)
237 switch (event->type_len) {
238 case RINGBUF_TYPE_PADDING:
239 if (rb_null_event(event))
242 return event->array[0] + RB_EVNT_HDR_SIZE;
244 case RINGBUF_TYPE_TIME_EXTEND:
245 return RB_LEN_TIME_EXTEND;
247 case RINGBUF_TYPE_TIME_STAMP:
248 return RB_LEN_TIME_STAMP;
250 case RINGBUF_TYPE_DATA:
251 return rb_event_data_length(event);
260 * Return total length of time extend and data,
261 * or just the event length for all other events.
263 static inline unsigned
264 rb_event_ts_length(struct ring_buffer_event *event)
268 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
269 /* time extends include the data event after it */
270 len = RB_LEN_TIME_EXTEND;
271 event = skip_time_extend(event);
273 return len + rb_event_length(event);
277 * ring_buffer_event_length - return the length of the event
278 * @event: the event to get the length of
280 * Returns the size of the data load of a data event.
281 * If the event is something other than a data event, it
282 * returns the size of the event itself. With the exception
283 * of a TIME EXTEND, where it still returns the size of the
284 * data load of the data event after it.
286 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
290 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
291 event = skip_time_extend(event);
293 length = rb_event_length(event);
294 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
296 length -= RB_EVNT_HDR_SIZE;
297 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
298 length -= sizeof(event->array[0]);
301 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
303 /* inline for ring buffer fast paths */
305 rb_event_data(struct ring_buffer_event *event)
307 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
308 event = skip_time_extend(event);
309 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
310 /* If length is in len field, then array[0] has the data */
312 return (void *)&event->array[0];
313 /* Otherwise length is in array[0] and array[1] has the data */
314 return (void *)&event->array[1];
318 * ring_buffer_event_data - return the data of the event
319 * @event: the event to get the data from
321 void *ring_buffer_event_data(struct ring_buffer_event *event)
323 return rb_event_data(event);
325 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
327 #define for_each_buffer_cpu(buffer, cpu) \
328 for_each_cpu(cpu, buffer->cpumask)
331 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
332 #define TS_DELTA_TEST (~TS_MASK)
334 /* Flag when events were overwritten */
335 #define RB_MISSED_EVENTS (1 << 31)
336 /* Missed count stored at end */
337 #define RB_MISSED_STORED (1 << 30)
339 struct buffer_data_page {
340 u64 time_stamp; /* page time stamp */
341 local_t commit; /* write committed index */
342 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
346 * Note, the buffer_page list must be first. The buffer pages
347 * are allocated in cache lines, which means that each buffer
348 * page will be at the beginning of a cache line, and thus
349 * the least significant bits will be zero. We use this to
350 * add flags in the list struct pointers, to make the ring buffer
354 struct list_head list; /* list of buffer pages */
355 local_t write; /* index for next write */
356 unsigned read; /* index for next read */
357 local_t entries; /* entries on this page */
358 unsigned long real_end; /* real end of data */
359 struct buffer_data_page *page; /* Actual data page */
363 * The buffer page counters, write and entries, must be reset
364 * atomically when crossing page boundaries. To synchronize this
365 * update, two counters are inserted into the number. One is
366 * the actual counter for the write position or count on the page.
368 * The other is a counter of updaters. Before an update happens
369 * the update partition of the counter is incremented. This will
370 * allow the updater to update the counter atomically.
372 * The counter is 20 bits, and the state data is 12.
374 #define RB_WRITE_MASK 0xfffff
375 #define RB_WRITE_INTCNT (1 << 20)
377 static void rb_init_page(struct buffer_data_page *bpage)
379 local_set(&bpage->commit, 0);
383 * ring_buffer_page_len - the size of data on the page.
384 * @page: The page to read
386 * Returns the amount of data on the page, including buffer page header.
388 size_t ring_buffer_page_len(void *page)
390 return local_read(&((struct buffer_data_page *)page)->commit)
395 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
398 static void free_buffer_page(struct buffer_page *bpage)
400 free_page((unsigned long)bpage->page);
405 * We need to fit the time_stamp delta into 27 bits.
407 static inline int test_time_stamp(u64 delta)
409 if (delta & TS_DELTA_TEST)
414 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
416 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
417 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
419 int ring_buffer_print_page_header(struct trace_seq *s)
421 struct buffer_data_page field;
424 ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t"
425 "offset:0;\tsize:%u;\tsigned:%u;\n",
426 (unsigned int)sizeof(field.time_stamp),
427 (unsigned int)is_signed_type(u64));
429 ret = trace_seq_printf(s, "\tfield: local_t commit;\t"
430 "offset:%u;\tsize:%u;\tsigned:%u;\n",
431 (unsigned int)offsetof(typeof(field), commit),
432 (unsigned int)sizeof(field.commit),
433 (unsigned int)is_signed_type(long));
435 ret = trace_seq_printf(s, "\tfield: int overwrite;\t"
436 "offset:%u;\tsize:%u;\tsigned:%u;\n",
437 (unsigned int)offsetof(typeof(field), commit),
439 (unsigned int)is_signed_type(long));
441 ret = trace_seq_printf(s, "\tfield: char data;\t"
442 "offset:%u;\tsize:%u;\tsigned:%u;\n",
443 (unsigned int)offsetof(typeof(field), data),
444 (unsigned int)BUF_PAGE_SIZE,
445 (unsigned int)is_signed_type(char));
451 struct irq_work work;
452 wait_queue_head_t waiters;
453 bool waiters_pending;
457 * head_page == tail_page && head == tail then buffer is empty.
459 struct ring_buffer_per_cpu {
461 atomic_t record_disabled;
462 struct ring_buffer *buffer;
463 raw_spinlock_t reader_lock; /* serialize readers */
464 arch_spinlock_t lock;
465 struct lock_class_key lock_key;
466 unsigned int nr_pages;
467 struct list_head *pages;
468 struct buffer_page *head_page; /* read from head */
469 struct buffer_page *tail_page; /* write to tail */
470 struct buffer_page *commit_page; /* committed pages */
471 struct buffer_page *reader_page;
472 unsigned long lost_events;
473 unsigned long last_overrun;
474 local_t entries_bytes;
477 local_t commit_overrun;
478 local_t dropped_events;
482 unsigned long read_bytes;
485 /* ring buffer pages to update, > 0 to add, < 0 to remove */
486 int nr_pages_to_update;
487 struct list_head new_pages; /* new pages to add */
488 struct work_struct update_pages_work;
489 struct completion update_done;
491 struct rb_irq_work irq_work;
497 atomic_t record_disabled;
498 atomic_t resize_disabled;
499 cpumask_var_t cpumask;
501 struct lock_class_key *reader_lock_key;
505 struct ring_buffer_per_cpu **buffers;
507 #ifdef CONFIG_HOTPLUG_CPU
508 struct notifier_block cpu_notify;
512 struct rb_irq_work irq_work;
515 struct ring_buffer_iter {
516 struct ring_buffer_per_cpu *cpu_buffer;
518 struct buffer_page *head_page;
519 struct buffer_page *cache_reader_page;
520 unsigned long cache_read;
525 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
527 * Schedules a delayed work to wake up any task that is blocked on the
528 * ring buffer waiters queue.
530 static void rb_wake_up_waiters(struct irq_work *work)
532 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
534 wake_up_all(&rbwork->waiters);
538 * ring_buffer_wait - wait for input to the ring buffer
539 * @buffer: buffer to wait on
540 * @cpu: the cpu buffer to wait on
542 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
543 * as data is added to any of the @buffer's cpu buffers. Otherwise
544 * it will wait for data to be added to a specific cpu buffer.
546 int ring_buffer_wait(struct ring_buffer *buffer, int cpu)
548 struct ring_buffer_per_cpu *cpu_buffer;
550 struct rb_irq_work *work;
553 * Depending on what the caller is waiting for, either any
554 * data in any cpu buffer, or a specific buffer, put the
555 * caller on the appropriate wait queue.
557 if (cpu == RING_BUFFER_ALL_CPUS)
558 work = &buffer->irq_work;
560 if (!cpumask_test_cpu(cpu, buffer->cpumask))
562 cpu_buffer = buffer->buffers[cpu];
563 work = &cpu_buffer->irq_work;
567 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
570 * The events can happen in critical sections where
571 * checking a work queue can cause deadlocks.
572 * After adding a task to the queue, this flag is set
573 * only to notify events to try to wake up the queue
576 * We don't clear it even if the buffer is no longer
577 * empty. The flag only causes the next event to run
578 * irq_work to do the work queue wake up. The worse
579 * that can happen if we race with !trace_empty() is that
580 * an event will cause an irq_work to try to wake up
583 * There's no reason to protect this flag either, as
584 * the work queue and irq_work logic will do the necessary
585 * synchronization for the wake ups. The only thing
586 * that is necessary is that the wake up happens after
587 * a task has been queued. It's OK for spurious wake ups.
589 work->waiters_pending = true;
591 if ((cpu == RING_BUFFER_ALL_CPUS && ring_buffer_empty(buffer)) ||
592 (cpu != RING_BUFFER_ALL_CPUS && ring_buffer_empty_cpu(buffer, cpu)))
595 finish_wait(&work->waiters, &wait);
600 * ring_buffer_poll_wait - poll on buffer input
601 * @buffer: buffer to wait on
602 * @cpu: the cpu buffer to wait on
603 * @filp: the file descriptor
604 * @poll_table: The poll descriptor
606 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
607 * as data is added to any of the @buffer's cpu buffers. Otherwise
608 * it will wait for data to be added to a specific cpu buffer.
610 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
613 int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
614 struct file *filp, poll_table *poll_table)
616 struct ring_buffer_per_cpu *cpu_buffer;
617 struct rb_irq_work *work;
619 if (cpu == RING_BUFFER_ALL_CPUS)
620 work = &buffer->irq_work;
622 if (!cpumask_test_cpu(cpu, buffer->cpumask))
625 cpu_buffer = buffer->buffers[cpu];
626 work = &cpu_buffer->irq_work;
629 work->waiters_pending = true;
630 poll_wait(filp, &work->waiters, poll_table);
632 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
633 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
634 return POLLIN | POLLRDNORM;
638 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
639 #define RB_WARN_ON(b, cond) \
641 int _____ret = unlikely(cond); \
643 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
644 struct ring_buffer_per_cpu *__b = \
646 atomic_inc(&__b->buffer->record_disabled); \
648 atomic_inc(&b->record_disabled); \
654 /* Up this if you want to test the TIME_EXTENTS and normalization */
655 #define DEBUG_SHIFT 0
657 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
659 /* shift to debug/test normalization and TIME_EXTENTS */
660 return buffer->clock() << DEBUG_SHIFT;
663 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
667 preempt_disable_notrace();
668 time = rb_time_stamp(buffer);
669 preempt_enable_no_resched_notrace();
673 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
675 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
678 /* Just stupid testing the normalize function and deltas */
681 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
684 * Making the ring buffer lockless makes things tricky.
685 * Although writes only happen on the CPU that they are on,
686 * and they only need to worry about interrupts. Reads can
689 * The reader page is always off the ring buffer, but when the
690 * reader finishes with a page, it needs to swap its page with
691 * a new one from the buffer. The reader needs to take from
692 * the head (writes go to the tail). But if a writer is in overwrite
693 * mode and wraps, it must push the head page forward.
695 * Here lies the problem.
697 * The reader must be careful to replace only the head page, and
698 * not another one. As described at the top of the file in the
699 * ASCII art, the reader sets its old page to point to the next
700 * page after head. It then sets the page after head to point to
701 * the old reader page. But if the writer moves the head page
702 * during this operation, the reader could end up with the tail.
704 * We use cmpxchg to help prevent this race. We also do something
705 * special with the page before head. We set the LSB to 1.
707 * When the writer must push the page forward, it will clear the
708 * bit that points to the head page, move the head, and then set
709 * the bit that points to the new head page.
711 * We also don't want an interrupt coming in and moving the head
712 * page on another writer. Thus we use the second LSB to catch
715 * head->list->prev->next bit 1 bit 0
718 * Points to head page 0 1
721 * Note we can not trust the prev pointer of the head page, because:
723 * +----+ +-----+ +-----+
724 * | |------>| T |---X--->| N |
726 * +----+ +-----+ +-----+
729 * +----------| R |----------+ |
733 * Key: ---X--> HEAD flag set in pointer
738 * (see __rb_reserve_next() to see where this happens)
740 * What the above shows is that the reader just swapped out
741 * the reader page with a page in the buffer, but before it
742 * could make the new header point back to the new page added
743 * it was preempted by a writer. The writer moved forward onto
744 * the new page added by the reader and is about to move forward
747 * You can see, it is legitimate for the previous pointer of
748 * the head (or any page) not to point back to itself. But only
752 #define RB_PAGE_NORMAL 0UL
753 #define RB_PAGE_HEAD 1UL
754 #define RB_PAGE_UPDATE 2UL
757 #define RB_FLAG_MASK 3UL
759 /* PAGE_MOVED is not part of the mask */
760 #define RB_PAGE_MOVED 4UL
763 * rb_list_head - remove any bit
765 static struct list_head *rb_list_head(struct list_head *list)
767 unsigned long val = (unsigned long)list;
769 return (struct list_head *)(val & ~RB_FLAG_MASK);
773 * rb_is_head_page - test if the given page is the head page
775 * Because the reader may move the head_page pointer, we can
776 * not trust what the head page is (it may be pointing to
777 * the reader page). But if the next page is a header page,
778 * its flags will be non zero.
781 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
782 struct buffer_page *page, struct list_head *list)
786 val = (unsigned long)list->next;
788 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
789 return RB_PAGE_MOVED;
791 return val & RB_FLAG_MASK;
797 * The unique thing about the reader page, is that, if the
798 * writer is ever on it, the previous pointer never points
799 * back to the reader page.
801 static int rb_is_reader_page(struct buffer_page *page)
803 struct list_head *list = page->list.prev;
805 return rb_list_head(list->next) != &page->list;
809 * rb_set_list_to_head - set a list_head to be pointing to head.
811 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
812 struct list_head *list)
816 ptr = (unsigned long *)&list->next;
817 *ptr |= RB_PAGE_HEAD;
818 *ptr &= ~RB_PAGE_UPDATE;
822 * rb_head_page_activate - sets up head page
824 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
826 struct buffer_page *head;
828 head = cpu_buffer->head_page;
833 * Set the previous list pointer to have the HEAD flag.
835 rb_set_list_to_head(cpu_buffer, head->list.prev);
838 static void rb_list_head_clear(struct list_head *list)
840 unsigned long *ptr = (unsigned long *)&list->next;
842 *ptr &= ~RB_FLAG_MASK;
846 * rb_head_page_dactivate - clears head page ptr (for free list)
849 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
851 struct list_head *hd;
853 /* Go through the whole list and clear any pointers found. */
854 rb_list_head_clear(cpu_buffer->pages);
856 list_for_each(hd, cpu_buffer->pages)
857 rb_list_head_clear(hd);
860 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
861 struct buffer_page *head,
862 struct buffer_page *prev,
863 int old_flag, int new_flag)
865 struct list_head *list;
866 unsigned long val = (unsigned long)&head->list;
871 val &= ~RB_FLAG_MASK;
873 ret = cmpxchg((unsigned long *)&list->next,
874 val | old_flag, val | new_flag);
876 /* check if the reader took the page */
877 if ((ret & ~RB_FLAG_MASK) != val)
878 return RB_PAGE_MOVED;
880 return ret & RB_FLAG_MASK;
883 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
884 struct buffer_page *head,
885 struct buffer_page *prev,
888 return rb_head_page_set(cpu_buffer, head, prev,
889 old_flag, RB_PAGE_UPDATE);
892 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
893 struct buffer_page *head,
894 struct buffer_page *prev,
897 return rb_head_page_set(cpu_buffer, head, prev,
898 old_flag, RB_PAGE_HEAD);
901 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
902 struct buffer_page *head,
903 struct buffer_page *prev,
906 return rb_head_page_set(cpu_buffer, head, prev,
907 old_flag, RB_PAGE_NORMAL);
910 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
911 struct buffer_page **bpage)
913 struct list_head *p = rb_list_head((*bpage)->list.next);
915 *bpage = list_entry(p, struct buffer_page, list);
918 static struct buffer_page *
919 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
921 struct buffer_page *head;
922 struct buffer_page *page;
923 struct list_head *list;
926 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
930 list = cpu_buffer->pages;
931 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
934 page = head = cpu_buffer->head_page;
936 * It is possible that the writer moves the header behind
937 * where we started, and we miss in one loop.
938 * A second loop should grab the header, but we'll do
939 * three loops just because I'm paranoid.
941 for (i = 0; i < 3; i++) {
943 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
944 cpu_buffer->head_page = page;
947 rb_inc_page(cpu_buffer, &page);
948 } while (page != head);
951 RB_WARN_ON(cpu_buffer, 1);
956 static int rb_head_page_replace(struct buffer_page *old,
957 struct buffer_page *new)
959 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
963 val = *ptr & ~RB_FLAG_MASK;
966 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
972 * rb_tail_page_update - move the tail page forward
974 * Returns 1 if moved tail page, 0 if someone else did.
976 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
977 struct buffer_page *tail_page,
978 struct buffer_page *next_page)
980 struct buffer_page *old_tail;
981 unsigned long old_entries;
982 unsigned long old_write;
986 * The tail page now needs to be moved forward.
988 * We need to reset the tail page, but without messing
989 * with possible erasing of data brought in by interrupts
990 * that have moved the tail page and are currently on it.
992 * We add a counter to the write field to denote this.
994 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
995 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
998 * Just make sure we have seen our old_write and synchronize
999 * with any interrupts that come in.
1004 * If the tail page is still the same as what we think
1005 * it is, then it is up to us to update the tail
1008 if (tail_page == cpu_buffer->tail_page) {
1009 /* Zero the write counter */
1010 unsigned long val = old_write & ~RB_WRITE_MASK;
1011 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1014 * This will only succeed if an interrupt did
1015 * not come in and change it. In which case, we
1016 * do not want to modify it.
1018 * We add (void) to let the compiler know that we do not care
1019 * about the return value of these functions. We use the
1020 * cmpxchg to only update if an interrupt did not already
1021 * do it for us. If the cmpxchg fails, we don't care.
1023 (void)local_cmpxchg(&next_page->write, old_write, val);
1024 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1027 * No need to worry about races with clearing out the commit.
1028 * it only can increment when a commit takes place. But that
1029 * only happens in the outer most nested commit.
1031 local_set(&next_page->page->commit, 0);
1033 old_tail = cmpxchg(&cpu_buffer->tail_page,
1034 tail_page, next_page);
1036 if (old_tail == tail_page)
1043 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1044 struct buffer_page *bpage)
1046 unsigned long val = (unsigned long)bpage;
1048 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1055 * rb_check_list - make sure a pointer to a list has the last bits zero
1057 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1058 struct list_head *list)
1060 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1062 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1068 * rb_check_pages - integrity check of buffer pages
1069 * @cpu_buffer: CPU buffer with pages to test
1071 * As a safety measure we check to make sure the data pages have not
1074 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1076 struct list_head *head = cpu_buffer->pages;
1077 struct buffer_page *bpage, *tmp;
1079 /* Reset the head page if it exists */
1080 if (cpu_buffer->head_page)
1081 rb_set_head_page(cpu_buffer);
1083 rb_head_page_deactivate(cpu_buffer);
1085 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1087 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1090 if (rb_check_list(cpu_buffer, head))
1093 list_for_each_entry_safe(bpage, tmp, head, list) {
1094 if (RB_WARN_ON(cpu_buffer,
1095 bpage->list.next->prev != &bpage->list))
1097 if (RB_WARN_ON(cpu_buffer,
1098 bpage->list.prev->next != &bpage->list))
1100 if (rb_check_list(cpu_buffer, &bpage->list))
1104 rb_head_page_activate(cpu_buffer);
1109 static int __rb_allocate_pages(int nr_pages, struct list_head *pages, int cpu)
1112 struct buffer_page *bpage, *tmp;
1114 for (i = 0; i < nr_pages; i++) {
1117 * __GFP_NORETRY flag makes sure that the allocation fails
1118 * gracefully without invoking oom-killer and the system is
1121 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1122 GFP_KERNEL | __GFP_NORETRY,
1127 list_add(&bpage->list, pages);
1129 page = alloc_pages_node(cpu_to_node(cpu),
1130 GFP_KERNEL | __GFP_NORETRY, 0);
1133 bpage->page = page_address(page);
1134 rb_init_page(bpage->page);
1140 list_for_each_entry_safe(bpage, tmp, pages, list) {
1141 list_del_init(&bpage->list);
1142 free_buffer_page(bpage);
1148 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1155 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1159 * The ring buffer page list is a circular list that does not
1160 * start and end with a list head. All page list items point to
1163 cpu_buffer->pages = pages.next;
1166 cpu_buffer->nr_pages = nr_pages;
1168 rb_check_pages(cpu_buffer);
1173 static struct ring_buffer_per_cpu *
1174 rb_allocate_cpu_buffer(struct ring_buffer *buffer, int nr_pages, int cpu)
1176 struct ring_buffer_per_cpu *cpu_buffer;
1177 struct buffer_page *bpage;
1181 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1182 GFP_KERNEL, cpu_to_node(cpu));
1186 cpu_buffer->cpu = cpu;
1187 cpu_buffer->buffer = buffer;
1188 raw_spin_lock_init(&cpu_buffer->reader_lock);
1189 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1190 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1191 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1192 init_completion(&cpu_buffer->update_done);
1193 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1194 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1196 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1197 GFP_KERNEL, cpu_to_node(cpu));
1199 goto fail_free_buffer;
1201 rb_check_bpage(cpu_buffer, bpage);
1203 cpu_buffer->reader_page = bpage;
1204 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1206 goto fail_free_reader;
1207 bpage->page = page_address(page);
1208 rb_init_page(bpage->page);
1210 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1211 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1213 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1215 goto fail_free_reader;
1217 cpu_buffer->head_page
1218 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1219 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1221 rb_head_page_activate(cpu_buffer);
1226 free_buffer_page(cpu_buffer->reader_page);
1233 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1235 struct list_head *head = cpu_buffer->pages;
1236 struct buffer_page *bpage, *tmp;
1238 free_buffer_page(cpu_buffer->reader_page);
1240 rb_head_page_deactivate(cpu_buffer);
1243 list_for_each_entry_safe(bpage, tmp, head, list) {
1244 list_del_init(&bpage->list);
1245 free_buffer_page(bpage);
1247 bpage = list_entry(head, struct buffer_page, list);
1248 free_buffer_page(bpage);
1254 #ifdef CONFIG_HOTPLUG_CPU
1255 static int rb_cpu_notify(struct notifier_block *self,
1256 unsigned long action, void *hcpu);
1260 * __ring_buffer_alloc - allocate a new ring_buffer
1261 * @size: the size in bytes per cpu that is needed.
1262 * @flags: attributes to set for the ring buffer.
1264 * Currently the only flag that is available is the RB_FL_OVERWRITE
1265 * flag. This flag means that the buffer will overwrite old data
1266 * when the buffer wraps. If this flag is not set, the buffer will
1267 * drop data when the tail hits the head.
1269 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1270 struct lock_class_key *key)
1272 struct ring_buffer *buffer;
1276 /* keep it in its own cache line */
1277 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1282 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1283 goto fail_free_buffer;
1285 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1286 buffer->flags = flags;
1287 buffer->clock = trace_clock_local;
1288 buffer->reader_lock_key = key;
1290 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1291 init_waitqueue_head(&buffer->irq_work.waiters);
1293 /* need at least two pages */
1298 * In case of non-hotplug cpu, if the ring-buffer is allocated
1299 * in early initcall, it will not be notified of secondary cpus.
1300 * In that off case, we need to allocate for all possible cpus.
1302 #ifdef CONFIG_HOTPLUG_CPU
1303 cpu_notifier_register_begin();
1304 cpumask_copy(buffer->cpumask, cpu_online_mask);
1306 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1308 buffer->cpus = nr_cpu_ids;
1310 bsize = sizeof(void *) * nr_cpu_ids;
1311 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1313 if (!buffer->buffers)
1314 goto fail_free_cpumask;
1316 for_each_buffer_cpu(buffer, cpu) {
1317 buffer->buffers[cpu] =
1318 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1319 if (!buffer->buffers[cpu])
1320 goto fail_free_buffers;
1323 #ifdef CONFIG_HOTPLUG_CPU
1324 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1325 buffer->cpu_notify.priority = 0;
1326 __register_cpu_notifier(&buffer->cpu_notify);
1327 cpu_notifier_register_done();
1330 mutex_init(&buffer->mutex);
1335 for_each_buffer_cpu(buffer, cpu) {
1336 if (buffer->buffers[cpu])
1337 rb_free_cpu_buffer(buffer->buffers[cpu]);
1339 kfree(buffer->buffers);
1342 free_cpumask_var(buffer->cpumask);
1343 #ifdef CONFIG_HOTPLUG_CPU
1344 cpu_notifier_register_done();
1351 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1354 * ring_buffer_free - free a ring buffer.
1355 * @buffer: the buffer to free.
1358 ring_buffer_free(struct ring_buffer *buffer)
1362 #ifdef CONFIG_HOTPLUG_CPU
1363 cpu_notifier_register_begin();
1364 __unregister_cpu_notifier(&buffer->cpu_notify);
1367 for_each_buffer_cpu(buffer, cpu)
1368 rb_free_cpu_buffer(buffer->buffers[cpu]);
1370 #ifdef CONFIG_HOTPLUG_CPU
1371 cpu_notifier_register_done();
1374 kfree(buffer->buffers);
1375 free_cpumask_var(buffer->cpumask);
1379 EXPORT_SYMBOL_GPL(ring_buffer_free);
1381 void ring_buffer_set_clock(struct ring_buffer *buffer,
1384 buffer->clock = clock;
1387 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1389 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1391 return local_read(&bpage->entries) & RB_WRITE_MASK;
1394 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1396 return local_read(&bpage->write) & RB_WRITE_MASK;
1400 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned int nr_pages)
1402 struct list_head *tail_page, *to_remove, *next_page;
1403 struct buffer_page *to_remove_page, *tmp_iter_page;
1404 struct buffer_page *last_page, *first_page;
1405 unsigned int nr_removed;
1406 unsigned long head_bit;
1411 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1412 atomic_inc(&cpu_buffer->record_disabled);
1414 * We don't race with the readers since we have acquired the reader
1415 * lock. We also don't race with writers after disabling recording.
1416 * This makes it easy to figure out the first and the last page to be
1417 * removed from the list. We unlink all the pages in between including
1418 * the first and last pages. This is done in a busy loop so that we
1419 * lose the least number of traces.
1420 * The pages are freed after we restart recording and unlock readers.
1422 tail_page = &cpu_buffer->tail_page->list;
1425 * tail page might be on reader page, we remove the next page
1426 * from the ring buffer
1428 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1429 tail_page = rb_list_head(tail_page->next);
1430 to_remove = tail_page;
1432 /* start of pages to remove */
1433 first_page = list_entry(rb_list_head(to_remove->next),
1434 struct buffer_page, list);
1436 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1437 to_remove = rb_list_head(to_remove)->next;
1438 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1441 next_page = rb_list_head(to_remove)->next;
1444 * Now we remove all pages between tail_page and next_page.
1445 * Make sure that we have head_bit value preserved for the
1448 tail_page->next = (struct list_head *)((unsigned long)next_page |
1450 next_page = rb_list_head(next_page);
1451 next_page->prev = tail_page;
1453 /* make sure pages points to a valid page in the ring buffer */
1454 cpu_buffer->pages = next_page;
1456 /* update head page */
1458 cpu_buffer->head_page = list_entry(next_page,
1459 struct buffer_page, list);
1462 * change read pointer to make sure any read iterators reset
1465 cpu_buffer->read = 0;
1467 /* pages are removed, resume tracing and then free the pages */
1468 atomic_dec(&cpu_buffer->record_disabled);
1469 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1471 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1473 /* last buffer page to remove */
1474 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1476 tmp_iter_page = first_page;
1479 to_remove_page = tmp_iter_page;
1480 rb_inc_page(cpu_buffer, &tmp_iter_page);
1482 /* update the counters */
1483 page_entries = rb_page_entries(to_remove_page);
1486 * If something was added to this page, it was full
1487 * since it is not the tail page. So we deduct the
1488 * bytes consumed in ring buffer from here.
1489 * Increment overrun to account for the lost events.
1491 local_add(page_entries, &cpu_buffer->overrun);
1492 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1496 * We have already removed references to this list item, just
1497 * free up the buffer_page and its page
1499 free_buffer_page(to_remove_page);
1502 } while (to_remove_page != last_page);
1504 RB_WARN_ON(cpu_buffer, nr_removed);
1506 return nr_removed == 0;
1510 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1512 struct list_head *pages = &cpu_buffer->new_pages;
1513 int retries, success;
1515 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1517 * We are holding the reader lock, so the reader page won't be swapped
1518 * in the ring buffer. Now we are racing with the writer trying to
1519 * move head page and the tail page.
1520 * We are going to adapt the reader page update process where:
1521 * 1. We first splice the start and end of list of new pages between
1522 * the head page and its previous page.
1523 * 2. We cmpxchg the prev_page->next to point from head page to the
1524 * start of new pages list.
1525 * 3. Finally, we update the head->prev to the end of new list.
1527 * We will try this process 10 times, to make sure that we don't keep
1533 struct list_head *head_page, *prev_page, *r;
1534 struct list_head *last_page, *first_page;
1535 struct list_head *head_page_with_bit;
1537 head_page = &rb_set_head_page(cpu_buffer)->list;
1540 prev_page = head_page->prev;
1542 first_page = pages->next;
1543 last_page = pages->prev;
1545 head_page_with_bit = (struct list_head *)
1546 ((unsigned long)head_page | RB_PAGE_HEAD);
1548 last_page->next = head_page_with_bit;
1549 first_page->prev = prev_page;
1551 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1553 if (r == head_page_with_bit) {
1555 * yay, we replaced the page pointer to our new list,
1556 * now, we just have to update to head page's prev
1557 * pointer to point to end of list
1559 head_page->prev = last_page;
1566 INIT_LIST_HEAD(pages);
1568 * If we weren't successful in adding in new pages, warn and stop
1571 RB_WARN_ON(cpu_buffer, !success);
1572 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1574 /* free pages if they weren't inserted */
1576 struct buffer_page *bpage, *tmp;
1577 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1579 list_del_init(&bpage->list);
1580 free_buffer_page(bpage);
1586 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1590 if (cpu_buffer->nr_pages_to_update > 0)
1591 success = rb_insert_pages(cpu_buffer);
1593 success = rb_remove_pages(cpu_buffer,
1594 -cpu_buffer->nr_pages_to_update);
1597 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1600 static void update_pages_handler(struct work_struct *work)
1602 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1603 struct ring_buffer_per_cpu, update_pages_work);
1604 rb_update_pages(cpu_buffer);
1605 complete(&cpu_buffer->update_done);
1609 * ring_buffer_resize - resize the ring buffer
1610 * @buffer: the buffer to resize.
1611 * @size: the new size.
1612 * @cpu_id: the cpu buffer to resize
1614 * Minimum size is 2 * BUF_PAGE_SIZE.
1616 * Returns 0 on success and < 0 on failure.
1618 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1621 struct ring_buffer_per_cpu *cpu_buffer;
1626 * Always succeed at resizing a non-existent buffer:
1631 /* Make sure the requested buffer exists */
1632 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1633 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1636 size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1637 size *= BUF_PAGE_SIZE;
1639 /* we need a minimum of two pages */
1640 if (size < BUF_PAGE_SIZE * 2)
1641 size = BUF_PAGE_SIZE * 2;
1643 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1646 * Don't succeed if resizing is disabled, as a reader might be
1647 * manipulating the ring buffer and is expecting a sane state while
1650 if (atomic_read(&buffer->resize_disabled))
1653 /* prevent another thread from changing buffer sizes */
1654 mutex_lock(&buffer->mutex);
1656 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1657 /* calculate the pages to update */
1658 for_each_buffer_cpu(buffer, cpu) {
1659 cpu_buffer = buffer->buffers[cpu];
1661 cpu_buffer->nr_pages_to_update = nr_pages -
1662 cpu_buffer->nr_pages;
1664 * nothing more to do for removing pages or no update
1666 if (cpu_buffer->nr_pages_to_update <= 0)
1669 * to add pages, make sure all new pages can be
1670 * allocated without receiving ENOMEM
1672 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1673 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1674 &cpu_buffer->new_pages, cpu)) {
1675 /* not enough memory for new pages */
1683 * Fire off all the required work handlers
1684 * We can't schedule on offline CPUs, but it's not necessary
1685 * since we can change their buffer sizes without any race.
1687 for_each_buffer_cpu(buffer, cpu) {
1688 cpu_buffer = buffer->buffers[cpu];
1689 if (!cpu_buffer->nr_pages_to_update)
1692 /* The update must run on the CPU that is being updated. */
1694 if (cpu == smp_processor_id() || !cpu_online(cpu)) {
1695 rb_update_pages(cpu_buffer);
1696 cpu_buffer->nr_pages_to_update = 0;
1699 * Can not disable preemption for schedule_work_on()
1703 schedule_work_on(cpu,
1704 &cpu_buffer->update_pages_work);
1710 /* wait for all the updates to complete */
1711 for_each_buffer_cpu(buffer, cpu) {
1712 cpu_buffer = buffer->buffers[cpu];
1713 if (!cpu_buffer->nr_pages_to_update)
1716 if (cpu_online(cpu))
1717 wait_for_completion(&cpu_buffer->update_done);
1718 cpu_buffer->nr_pages_to_update = 0;
1723 /* Make sure this CPU has been intitialized */
1724 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1727 cpu_buffer = buffer->buffers[cpu_id];
1729 if (nr_pages == cpu_buffer->nr_pages)
1732 cpu_buffer->nr_pages_to_update = nr_pages -
1733 cpu_buffer->nr_pages;
1735 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1736 if (cpu_buffer->nr_pages_to_update > 0 &&
1737 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1738 &cpu_buffer->new_pages, cpu_id)) {
1746 /* The update must run on the CPU that is being updated. */
1747 if (cpu_id == smp_processor_id() || !cpu_online(cpu_id))
1748 rb_update_pages(cpu_buffer);
1751 * Can not disable preemption for schedule_work_on()
1755 schedule_work_on(cpu_id,
1756 &cpu_buffer->update_pages_work);
1757 wait_for_completion(&cpu_buffer->update_done);
1762 cpu_buffer->nr_pages_to_update = 0;
1768 * The ring buffer resize can happen with the ring buffer
1769 * enabled, so that the update disturbs the tracing as little
1770 * as possible. But if the buffer is disabled, we do not need
1771 * to worry about that, and we can take the time to verify
1772 * that the buffer is not corrupt.
1774 if (atomic_read(&buffer->record_disabled)) {
1775 atomic_inc(&buffer->record_disabled);
1777 * Even though the buffer was disabled, we must make sure
1778 * that it is truly disabled before calling rb_check_pages.
1779 * There could have been a race between checking
1780 * record_disable and incrementing it.
1782 synchronize_sched();
1783 for_each_buffer_cpu(buffer, cpu) {
1784 cpu_buffer = buffer->buffers[cpu];
1785 rb_check_pages(cpu_buffer);
1787 atomic_dec(&buffer->record_disabled);
1790 mutex_unlock(&buffer->mutex);
1794 for_each_buffer_cpu(buffer, cpu) {
1795 struct buffer_page *bpage, *tmp;
1797 cpu_buffer = buffer->buffers[cpu];
1798 cpu_buffer->nr_pages_to_update = 0;
1800 if (list_empty(&cpu_buffer->new_pages))
1803 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1805 list_del_init(&bpage->list);
1806 free_buffer_page(bpage);
1809 mutex_unlock(&buffer->mutex);
1812 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1814 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1816 mutex_lock(&buffer->mutex);
1818 buffer->flags |= RB_FL_OVERWRITE;
1820 buffer->flags &= ~RB_FL_OVERWRITE;
1821 mutex_unlock(&buffer->mutex);
1823 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1825 static inline void *
1826 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1828 return bpage->data + index;
1831 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1833 return bpage->page->data + index;
1836 static inline struct ring_buffer_event *
1837 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1839 return __rb_page_index(cpu_buffer->reader_page,
1840 cpu_buffer->reader_page->read);
1843 static inline struct ring_buffer_event *
1844 rb_iter_head_event(struct ring_buffer_iter *iter)
1846 return __rb_page_index(iter->head_page, iter->head);
1849 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1851 return local_read(&bpage->page->commit);
1854 /* Size is determined by what has been committed */
1855 static inline unsigned rb_page_size(struct buffer_page *bpage)
1857 return rb_page_commit(bpage);
1860 static inline unsigned
1861 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1863 return rb_page_commit(cpu_buffer->commit_page);
1866 static inline unsigned
1867 rb_event_index(struct ring_buffer_event *event)
1869 unsigned long addr = (unsigned long)event;
1871 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1875 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1876 struct ring_buffer_event *event)
1878 unsigned long addr = (unsigned long)event;
1879 unsigned long index;
1881 index = rb_event_index(event);
1884 return cpu_buffer->commit_page->page == (void *)addr &&
1885 rb_commit_index(cpu_buffer) == index;
1889 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1891 unsigned long max_count;
1894 * We only race with interrupts and NMIs on this CPU.
1895 * If we own the commit event, then we can commit
1896 * all others that interrupted us, since the interruptions
1897 * are in stack format (they finish before they come
1898 * back to us). This allows us to do a simple loop to
1899 * assign the commit to the tail.
1902 max_count = cpu_buffer->nr_pages * 100;
1904 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1905 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1907 if (RB_WARN_ON(cpu_buffer,
1908 rb_is_reader_page(cpu_buffer->tail_page)))
1910 local_set(&cpu_buffer->commit_page->page->commit,
1911 rb_page_write(cpu_buffer->commit_page));
1912 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1913 cpu_buffer->write_stamp =
1914 cpu_buffer->commit_page->page->time_stamp;
1915 /* add barrier to keep gcc from optimizing too much */
1918 while (rb_commit_index(cpu_buffer) !=
1919 rb_page_write(cpu_buffer->commit_page)) {
1921 local_set(&cpu_buffer->commit_page->page->commit,
1922 rb_page_write(cpu_buffer->commit_page));
1923 RB_WARN_ON(cpu_buffer,
1924 local_read(&cpu_buffer->commit_page->page->commit) &
1929 /* again, keep gcc from optimizing */
1933 * If an interrupt came in just after the first while loop
1934 * and pushed the tail page forward, we will be left with
1935 * a dangling commit that will never go forward.
1937 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1941 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1943 cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1944 cpu_buffer->reader_page->read = 0;
1947 static void rb_inc_iter(struct ring_buffer_iter *iter)
1949 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1952 * The iterator could be on the reader page (it starts there).
1953 * But the head could have moved, since the reader was
1954 * found. Check for this case and assign the iterator
1955 * to the head page instead of next.
1957 if (iter->head_page == cpu_buffer->reader_page)
1958 iter->head_page = rb_set_head_page(cpu_buffer);
1960 rb_inc_page(cpu_buffer, &iter->head_page);
1962 iter->read_stamp = iter->head_page->page->time_stamp;
1966 /* Slow path, do not inline */
1967 static noinline struct ring_buffer_event *
1968 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
1970 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
1972 /* Not the first event on the page? */
1973 if (rb_event_index(event)) {
1974 event->time_delta = delta & TS_MASK;
1975 event->array[0] = delta >> TS_SHIFT;
1977 /* nope, just zero it */
1978 event->time_delta = 0;
1979 event->array[0] = 0;
1982 return skip_time_extend(event);
1986 * rb_update_event - update event type and data
1987 * @event: the even to update
1988 * @type: the type of event
1989 * @length: the size of the event field in the ring buffer
1991 * Update the type and data fields of the event. The length
1992 * is the actual size that is written to the ring buffer,
1993 * and with this, we can determine what to place into the
1997 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
1998 struct ring_buffer_event *event, unsigned length,
1999 int add_timestamp, u64 delta)
2001 /* Only a commit updates the timestamp */
2002 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2006 * If we need to add a timestamp, then we
2007 * add it to the start of the resevered space.
2009 if (unlikely(add_timestamp)) {
2010 event = rb_add_time_stamp(event, delta);
2011 length -= RB_LEN_TIME_EXTEND;
2015 event->time_delta = delta;
2016 length -= RB_EVNT_HDR_SIZE;
2017 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2018 event->type_len = 0;
2019 event->array[0] = length;
2021 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2025 * rb_handle_head_page - writer hit the head page
2027 * Returns: +1 to retry page
2032 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2033 struct buffer_page *tail_page,
2034 struct buffer_page *next_page)
2036 struct buffer_page *new_head;
2041 entries = rb_page_entries(next_page);
2044 * The hard part is here. We need to move the head
2045 * forward, and protect against both readers on
2046 * other CPUs and writers coming in via interrupts.
2048 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2052 * type can be one of four:
2053 * NORMAL - an interrupt already moved it for us
2054 * HEAD - we are the first to get here.
2055 * UPDATE - we are the interrupt interrupting
2057 * MOVED - a reader on another CPU moved the next
2058 * pointer to its reader page. Give up
2065 * We changed the head to UPDATE, thus
2066 * it is our responsibility to update
2069 local_add(entries, &cpu_buffer->overrun);
2070 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2073 * The entries will be zeroed out when we move the
2077 /* still more to do */
2080 case RB_PAGE_UPDATE:
2082 * This is an interrupt that interrupt the
2083 * previous update. Still more to do.
2086 case RB_PAGE_NORMAL:
2088 * An interrupt came in before the update
2089 * and processed this for us.
2090 * Nothing left to do.
2095 * The reader is on another CPU and just did
2096 * a swap with our next_page.
2101 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2106 * Now that we are here, the old head pointer is
2107 * set to UPDATE. This will keep the reader from
2108 * swapping the head page with the reader page.
2109 * The reader (on another CPU) will spin till
2112 * We just need to protect against interrupts
2113 * doing the job. We will set the next pointer
2114 * to HEAD. After that, we set the old pointer
2115 * to NORMAL, but only if it was HEAD before.
2116 * otherwise we are an interrupt, and only
2117 * want the outer most commit to reset it.
2119 new_head = next_page;
2120 rb_inc_page(cpu_buffer, &new_head);
2122 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2126 * Valid returns are:
2127 * HEAD - an interrupt came in and already set it.
2128 * NORMAL - One of two things:
2129 * 1) We really set it.
2130 * 2) A bunch of interrupts came in and moved
2131 * the page forward again.
2135 case RB_PAGE_NORMAL:
2139 RB_WARN_ON(cpu_buffer, 1);
2144 * It is possible that an interrupt came in,
2145 * set the head up, then more interrupts came in
2146 * and moved it again. When we get back here,
2147 * the page would have been set to NORMAL but we
2148 * just set it back to HEAD.
2150 * How do you detect this? Well, if that happened
2151 * the tail page would have moved.
2153 if (ret == RB_PAGE_NORMAL) {
2155 * If the tail had moved passed next, then we need
2156 * to reset the pointer.
2158 if (cpu_buffer->tail_page != tail_page &&
2159 cpu_buffer->tail_page != next_page)
2160 rb_head_page_set_normal(cpu_buffer, new_head,
2166 * If this was the outer most commit (the one that
2167 * changed the original pointer from HEAD to UPDATE),
2168 * then it is up to us to reset it to NORMAL.
2170 if (type == RB_PAGE_HEAD) {
2171 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2174 if (RB_WARN_ON(cpu_buffer,
2175 ret != RB_PAGE_UPDATE))
2182 static unsigned rb_calculate_event_length(unsigned length)
2184 struct ring_buffer_event event; /* Used only for sizeof array */
2186 /* zero length can cause confusions */
2190 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2191 length += sizeof(event.array[0]);
2193 length += RB_EVNT_HDR_SIZE;
2194 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2200 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2201 struct buffer_page *tail_page,
2202 unsigned long tail, unsigned long length)
2204 struct ring_buffer_event *event;
2207 * Only the event that crossed the page boundary
2208 * must fill the old tail_page with padding.
2210 if (tail >= BUF_PAGE_SIZE) {
2212 * If the page was filled, then we still need
2213 * to update the real_end. Reset it to zero
2214 * and the reader will ignore it.
2216 if (tail == BUF_PAGE_SIZE)
2217 tail_page->real_end = 0;
2219 local_sub(length, &tail_page->write);
2223 event = __rb_page_index(tail_page, tail);
2224 kmemcheck_annotate_bitfield(event, bitfield);
2226 /* account for padding bytes */
2227 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2230 * Save the original length to the meta data.
2231 * This will be used by the reader to add lost event
2234 tail_page->real_end = tail;
2237 * If this event is bigger than the minimum size, then
2238 * we need to be careful that we don't subtract the
2239 * write counter enough to allow another writer to slip
2241 * We put in a discarded commit instead, to make sure
2242 * that this space is not used again.
2244 * If we are less than the minimum size, we don't need to
2247 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2248 /* No room for any events */
2250 /* Mark the rest of the page with padding */
2251 rb_event_set_padding(event);
2253 /* Set the write back to the previous setting */
2254 local_sub(length, &tail_page->write);
2258 /* Put in a discarded event */
2259 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2260 event->type_len = RINGBUF_TYPE_PADDING;
2261 /* time delta must be non zero */
2262 event->time_delta = 1;
2264 /* Set write to end of buffer */
2265 length = (tail + length) - BUF_PAGE_SIZE;
2266 local_sub(length, &tail_page->write);
2270 * This is the slow path, force gcc not to inline it.
2272 static noinline struct ring_buffer_event *
2273 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2274 unsigned long length, unsigned long tail,
2275 struct buffer_page *tail_page, u64 ts)
2277 struct buffer_page *commit_page = cpu_buffer->commit_page;
2278 struct ring_buffer *buffer = cpu_buffer->buffer;
2279 struct buffer_page *next_page;
2282 next_page = tail_page;
2284 rb_inc_page(cpu_buffer, &next_page);
2287 * If for some reason, we had an interrupt storm that made
2288 * it all the way around the buffer, bail, and warn
2291 if (unlikely(next_page == commit_page)) {
2292 local_inc(&cpu_buffer->commit_overrun);
2297 * This is where the fun begins!
2299 * We are fighting against races between a reader that
2300 * could be on another CPU trying to swap its reader
2301 * page with the buffer head.
2303 * We are also fighting against interrupts coming in and
2304 * moving the head or tail on us as well.
2306 * If the next page is the head page then we have filled
2307 * the buffer, unless the commit page is still on the
2310 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2313 * If the commit is not on the reader page, then
2314 * move the header page.
2316 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2318 * If we are not in overwrite mode,
2319 * this is easy, just stop here.
2321 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2322 local_inc(&cpu_buffer->dropped_events);
2326 ret = rb_handle_head_page(cpu_buffer,
2335 * We need to be careful here too. The
2336 * commit page could still be on the reader
2337 * page. We could have a small buffer, and
2338 * have filled up the buffer with events
2339 * from interrupts and such, and wrapped.
2341 * Note, if the tail page is also the on the
2342 * reader_page, we let it move out.
2344 if (unlikely((cpu_buffer->commit_page !=
2345 cpu_buffer->tail_page) &&
2346 (cpu_buffer->commit_page ==
2347 cpu_buffer->reader_page))) {
2348 local_inc(&cpu_buffer->commit_overrun);
2354 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
2357 * Nested commits always have zero deltas, so
2358 * just reread the time stamp
2360 ts = rb_time_stamp(buffer);
2361 next_page->page->time_stamp = ts;
2366 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2368 /* fail and let the caller try again */
2369 return ERR_PTR(-EAGAIN);
2373 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2378 static struct ring_buffer_event *
2379 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2380 unsigned long length, u64 ts,
2381 u64 delta, int add_timestamp)
2383 struct buffer_page *tail_page;
2384 struct ring_buffer_event *event;
2385 unsigned long tail, write;
2388 * If the time delta since the last event is too big to
2389 * hold in the time field of the event, then we append a
2390 * TIME EXTEND event ahead of the data event.
2392 if (unlikely(add_timestamp))
2393 length += RB_LEN_TIME_EXTEND;
2395 tail_page = cpu_buffer->tail_page;
2396 write = local_add_return(length, &tail_page->write);
2398 /* set write to only the index of the write */
2399 write &= RB_WRITE_MASK;
2400 tail = write - length;
2403 * If this is the first commit on the page, then it has the same
2404 * timestamp as the page itself.
2409 /* See if we shot pass the end of this buffer page */
2410 if (unlikely(write > BUF_PAGE_SIZE))
2411 return rb_move_tail(cpu_buffer, length, tail,
2414 /* We reserved something on the buffer */
2416 event = __rb_page_index(tail_page, tail);
2417 kmemcheck_annotate_bitfield(event, bitfield);
2418 rb_update_event(cpu_buffer, event, length, add_timestamp, delta);
2420 local_inc(&tail_page->entries);
2423 * If this is the first commit on the page, then update
2427 tail_page->page->time_stamp = ts;
2429 /* account for these added bytes */
2430 local_add(length, &cpu_buffer->entries_bytes);
2436 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2437 struct ring_buffer_event *event)
2439 unsigned long new_index, old_index;
2440 struct buffer_page *bpage;
2441 unsigned long index;
2444 new_index = rb_event_index(event);
2445 old_index = new_index + rb_event_ts_length(event);
2446 addr = (unsigned long)event;
2449 bpage = cpu_buffer->tail_page;
2451 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2452 unsigned long write_mask =
2453 local_read(&bpage->write) & ~RB_WRITE_MASK;
2454 unsigned long event_length = rb_event_length(event);
2456 * This is on the tail page. It is possible that
2457 * a write could come in and move the tail page
2458 * and write to the next page. That is fine
2459 * because we just shorten what is on this page.
2461 old_index += write_mask;
2462 new_index += write_mask;
2463 index = local_cmpxchg(&bpage->write, old_index, new_index);
2464 if (index == old_index) {
2465 /* update counters */
2466 local_sub(event_length, &cpu_buffer->entries_bytes);
2471 /* could not discard */
2475 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2477 local_inc(&cpu_buffer->committing);
2478 local_inc(&cpu_buffer->commits);
2481 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2483 unsigned long commits;
2485 if (RB_WARN_ON(cpu_buffer,
2486 !local_read(&cpu_buffer->committing)))
2490 commits = local_read(&cpu_buffer->commits);
2491 /* synchronize with interrupts */
2493 if (local_read(&cpu_buffer->committing) == 1)
2494 rb_set_commit_to_write(cpu_buffer);
2496 local_dec(&cpu_buffer->committing);
2498 /* synchronize with interrupts */
2502 * Need to account for interrupts coming in between the
2503 * updating of the commit page and the clearing of the
2504 * committing counter.
2506 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2507 !local_read(&cpu_buffer->committing)) {
2508 local_inc(&cpu_buffer->committing);
2513 static struct ring_buffer_event *
2514 rb_reserve_next_event(struct ring_buffer *buffer,
2515 struct ring_buffer_per_cpu *cpu_buffer,
2516 unsigned long length)
2518 struct ring_buffer_event *event;
2524 rb_start_commit(cpu_buffer);
2526 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2528 * Due to the ability to swap a cpu buffer from a buffer
2529 * it is possible it was swapped before we committed.
2530 * (committing stops a swap). We check for it here and
2531 * if it happened, we have to fail the write.
2534 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2535 local_dec(&cpu_buffer->committing);
2536 local_dec(&cpu_buffer->commits);
2541 length = rb_calculate_event_length(length);
2547 * We allow for interrupts to reenter here and do a trace.
2548 * If one does, it will cause this original code to loop
2549 * back here. Even with heavy interrupts happening, this
2550 * should only happen a few times in a row. If this happens
2551 * 1000 times in a row, there must be either an interrupt
2552 * storm or we have something buggy.
2555 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2558 ts = rb_time_stamp(cpu_buffer->buffer);
2559 diff = ts - cpu_buffer->write_stamp;
2561 /* make sure this diff is calculated here */
2564 /* Did the write stamp get updated already? */
2565 if (likely(ts >= cpu_buffer->write_stamp)) {
2567 if (unlikely(test_time_stamp(delta))) {
2568 int local_clock_stable = 1;
2569 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2570 local_clock_stable = sched_clock_stable();
2572 WARN_ONCE(delta > (1ULL << 59),
2573 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2574 (unsigned long long)delta,
2575 (unsigned long long)ts,
2576 (unsigned long long)cpu_buffer->write_stamp,
2577 local_clock_stable ? "" :
2578 "If you just came from a suspend/resume,\n"
2579 "please switch to the trace global clock:\n"
2580 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2585 event = __rb_reserve_next(cpu_buffer, length, ts,
2586 delta, add_timestamp);
2587 if (unlikely(PTR_ERR(event) == -EAGAIN))
2596 rb_end_commit(cpu_buffer);
2600 #ifdef CONFIG_TRACING
2603 * The lock and unlock are done within a preempt disable section.
2604 * The current_context per_cpu variable can only be modified
2605 * by the current task between lock and unlock. But it can
2606 * be modified more than once via an interrupt. To pass this
2607 * information from the lock to the unlock without having to
2608 * access the 'in_interrupt()' functions again (which do show
2609 * a bit of overhead in something as critical as function tracing,
2610 * we use a bitmask trick.
2612 * bit 0 = NMI context
2613 * bit 1 = IRQ context
2614 * bit 2 = SoftIRQ context
2615 * bit 3 = normal context.
2617 * This works because this is the order of contexts that can
2618 * preempt other contexts. A SoftIRQ never preempts an IRQ
2621 * When the context is determined, the corresponding bit is
2622 * checked and set (if it was set, then a recursion of that context
2625 * On unlock, we need to clear this bit. To do so, just subtract
2626 * 1 from the current_context and AND it to itself.
2630 * 101 & 100 = 100 (clearing bit zero)
2633 * 1010 & 1001 = 1000 (clearing bit 1)
2635 * The least significant bit can be cleared this way, and it
2636 * just so happens that it is the same bit corresponding to
2637 * the current context.
2639 static DEFINE_PER_CPU(unsigned int, current_context);
2641 static __always_inline int trace_recursive_lock(void)
2643 unsigned int val = this_cpu_read(current_context);
2646 if (in_interrupt()) {
2656 if (unlikely(val & (1 << bit)))
2660 this_cpu_write(current_context, val);
2665 static __always_inline void trace_recursive_unlock(void)
2667 unsigned int val = this_cpu_read(current_context);
2670 val &= this_cpu_read(current_context);
2671 this_cpu_write(current_context, val);
2676 #define trace_recursive_lock() (0)
2677 #define trace_recursive_unlock() do { } while (0)
2682 * ring_buffer_lock_reserve - reserve a part of the buffer
2683 * @buffer: the ring buffer to reserve from
2684 * @length: the length of the data to reserve (excluding event header)
2686 * Returns a reseverd event on the ring buffer to copy directly to.
2687 * The user of this interface will need to get the body to write into
2688 * and can use the ring_buffer_event_data() interface.
2690 * The length is the length of the data needed, not the event length
2691 * which also includes the event header.
2693 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2694 * If NULL is returned, then nothing has been allocated or locked.
2696 struct ring_buffer_event *
2697 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2699 struct ring_buffer_per_cpu *cpu_buffer;
2700 struct ring_buffer_event *event;
2703 if (ring_buffer_flags != RB_BUFFERS_ON)
2706 /* If we are tracing schedule, we don't want to recurse */
2707 preempt_disable_notrace();
2709 if (atomic_read(&buffer->record_disabled))
2712 if (trace_recursive_lock())
2715 cpu = raw_smp_processor_id();
2717 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2720 cpu_buffer = buffer->buffers[cpu];
2722 if (atomic_read(&cpu_buffer->record_disabled))
2725 if (length > BUF_MAX_DATA_SIZE)
2728 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2735 trace_recursive_unlock();
2738 preempt_enable_notrace();
2741 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2744 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2745 struct ring_buffer_event *event)
2750 * The event first in the commit queue updates the
2753 if (rb_event_is_commit(cpu_buffer, event)) {
2755 * A commit event that is first on a page
2756 * updates the write timestamp with the page stamp
2758 if (!rb_event_index(event))
2759 cpu_buffer->write_stamp =
2760 cpu_buffer->commit_page->page->time_stamp;
2761 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2762 delta = event->array[0];
2764 delta += event->time_delta;
2765 cpu_buffer->write_stamp += delta;
2767 cpu_buffer->write_stamp += event->time_delta;
2771 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2772 struct ring_buffer_event *event)
2774 local_inc(&cpu_buffer->entries);
2775 rb_update_write_stamp(cpu_buffer, event);
2776 rb_end_commit(cpu_buffer);
2779 static __always_inline void
2780 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2782 if (buffer->irq_work.waiters_pending) {
2783 buffer->irq_work.waiters_pending = false;
2784 /* irq_work_queue() supplies it's own memory barriers */
2785 irq_work_queue(&buffer->irq_work.work);
2788 if (cpu_buffer->irq_work.waiters_pending) {
2789 cpu_buffer->irq_work.waiters_pending = false;
2790 /* irq_work_queue() supplies it's own memory barriers */
2791 irq_work_queue(&cpu_buffer->irq_work.work);
2796 * ring_buffer_unlock_commit - commit a reserved
2797 * @buffer: The buffer to commit to
2798 * @event: The event pointer to commit.
2800 * This commits the data to the ring buffer, and releases any locks held.
2802 * Must be paired with ring_buffer_lock_reserve.
2804 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2805 struct ring_buffer_event *event)
2807 struct ring_buffer_per_cpu *cpu_buffer;
2808 int cpu = raw_smp_processor_id();
2810 cpu_buffer = buffer->buffers[cpu];
2812 rb_commit(cpu_buffer, event);
2814 rb_wakeups(buffer, cpu_buffer);
2816 trace_recursive_unlock();
2818 preempt_enable_notrace();
2822 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2824 static inline void rb_event_discard(struct ring_buffer_event *event)
2826 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2827 event = skip_time_extend(event);
2829 /* array[0] holds the actual length for the discarded event */
2830 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2831 event->type_len = RINGBUF_TYPE_PADDING;
2832 /* time delta must be non zero */
2833 if (!event->time_delta)
2834 event->time_delta = 1;
2838 * Decrement the entries to the page that an event is on.
2839 * The event does not even need to exist, only the pointer
2840 * to the page it is on. This may only be called before the commit
2844 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2845 struct ring_buffer_event *event)
2847 unsigned long addr = (unsigned long)event;
2848 struct buffer_page *bpage = cpu_buffer->commit_page;
2849 struct buffer_page *start;
2853 /* Do the likely case first */
2854 if (likely(bpage->page == (void *)addr)) {
2855 local_dec(&bpage->entries);
2860 * Because the commit page may be on the reader page we
2861 * start with the next page and check the end loop there.
2863 rb_inc_page(cpu_buffer, &bpage);
2866 if (bpage->page == (void *)addr) {
2867 local_dec(&bpage->entries);
2870 rb_inc_page(cpu_buffer, &bpage);
2871 } while (bpage != start);
2873 /* commit not part of this buffer?? */
2874 RB_WARN_ON(cpu_buffer, 1);
2878 * ring_buffer_commit_discard - discard an event that has not been committed
2879 * @buffer: the ring buffer
2880 * @event: non committed event to discard
2882 * Sometimes an event that is in the ring buffer needs to be ignored.
2883 * This function lets the user discard an event in the ring buffer
2884 * and then that event will not be read later.
2886 * This function only works if it is called before the the item has been
2887 * committed. It will try to free the event from the ring buffer
2888 * if another event has not been added behind it.
2890 * If another event has been added behind it, it will set the event
2891 * up as discarded, and perform the commit.
2893 * If this function is called, do not call ring_buffer_unlock_commit on
2896 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2897 struct ring_buffer_event *event)
2899 struct ring_buffer_per_cpu *cpu_buffer;
2902 /* The event is discarded regardless */
2903 rb_event_discard(event);
2905 cpu = smp_processor_id();
2906 cpu_buffer = buffer->buffers[cpu];
2909 * This must only be called if the event has not been
2910 * committed yet. Thus we can assume that preemption
2911 * is still disabled.
2913 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2915 rb_decrement_entry(cpu_buffer, event);
2916 if (rb_try_to_discard(cpu_buffer, event))
2920 * The commit is still visible by the reader, so we
2921 * must still update the timestamp.
2923 rb_update_write_stamp(cpu_buffer, event);
2925 rb_end_commit(cpu_buffer);
2927 trace_recursive_unlock();
2929 preempt_enable_notrace();
2932 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2935 * ring_buffer_write - write data to the buffer without reserving
2936 * @buffer: The ring buffer to write to.
2937 * @length: The length of the data being written (excluding the event header)
2938 * @data: The data to write to the buffer.
2940 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2941 * one function. If you already have the data to write to the buffer, it
2942 * may be easier to simply call this function.
2944 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2945 * and not the length of the event which would hold the header.
2947 int ring_buffer_write(struct ring_buffer *buffer,
2948 unsigned long length,
2951 struct ring_buffer_per_cpu *cpu_buffer;
2952 struct ring_buffer_event *event;
2957 if (ring_buffer_flags != RB_BUFFERS_ON)
2960 preempt_disable_notrace();
2962 if (atomic_read(&buffer->record_disabled))
2965 cpu = raw_smp_processor_id();
2967 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2970 cpu_buffer = buffer->buffers[cpu];
2972 if (atomic_read(&cpu_buffer->record_disabled))
2975 if (length > BUF_MAX_DATA_SIZE)
2978 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2982 body = rb_event_data(event);
2984 memcpy(body, data, length);
2986 rb_commit(cpu_buffer, event);
2988 rb_wakeups(buffer, cpu_buffer);
2992 preempt_enable_notrace();
2996 EXPORT_SYMBOL_GPL(ring_buffer_write);
2998 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3000 struct buffer_page *reader = cpu_buffer->reader_page;
3001 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3002 struct buffer_page *commit = cpu_buffer->commit_page;
3004 /* In case of error, head will be NULL */
3005 if (unlikely(!head))
3008 return reader->read == rb_page_commit(reader) &&
3009 (commit == reader ||
3011 head->read == rb_page_commit(commit)));
3015 * ring_buffer_record_disable - stop all writes into the buffer
3016 * @buffer: The ring buffer to stop writes to.
3018 * This prevents all writes to the buffer. Any attempt to write
3019 * to the buffer after this will fail and return NULL.
3021 * The caller should call synchronize_sched() after this.
3023 void ring_buffer_record_disable(struct ring_buffer *buffer)
3025 atomic_inc(&buffer->record_disabled);
3027 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3030 * ring_buffer_record_enable - enable writes to the buffer
3031 * @buffer: The ring buffer to enable writes
3033 * Note, multiple disables will need the same number of enables
3034 * to truly enable the writing (much like preempt_disable).
3036 void ring_buffer_record_enable(struct ring_buffer *buffer)
3038 atomic_dec(&buffer->record_disabled);
3040 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3043 * ring_buffer_record_off - stop all writes into the buffer
3044 * @buffer: The ring buffer to stop writes to.
3046 * This prevents all writes to the buffer. Any attempt to write
3047 * to the buffer after this will fail and return NULL.
3049 * This is different than ring_buffer_record_disable() as
3050 * it works like an on/off switch, where as the disable() version
3051 * must be paired with a enable().
3053 void ring_buffer_record_off(struct ring_buffer *buffer)
3056 unsigned int new_rd;
3059 rd = atomic_read(&buffer->record_disabled);
3060 new_rd = rd | RB_BUFFER_OFF;
3061 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3063 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3066 * ring_buffer_record_on - restart writes into the buffer
3067 * @buffer: The ring buffer to start writes to.
3069 * This enables all writes to the buffer that was disabled by
3070 * ring_buffer_record_off().
3072 * This is different than ring_buffer_record_enable() as
3073 * it works like an on/off switch, where as the enable() version
3074 * must be paired with a disable().
3076 void ring_buffer_record_on(struct ring_buffer *buffer)
3079 unsigned int new_rd;
3082 rd = atomic_read(&buffer->record_disabled);
3083 new_rd = rd & ~RB_BUFFER_OFF;
3084 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3086 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3089 * ring_buffer_record_is_on - return true if the ring buffer can write
3090 * @buffer: The ring buffer to see if write is enabled
3092 * Returns true if the ring buffer is in a state that it accepts writes.
3094 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3096 return !atomic_read(&buffer->record_disabled);
3100 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3101 * @buffer: The ring buffer to stop writes to.
3102 * @cpu: The CPU buffer to stop
3104 * This prevents all writes to the buffer. Any attempt to write
3105 * to the buffer after this will fail and return NULL.
3107 * The caller should call synchronize_sched() after this.
3109 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3111 struct ring_buffer_per_cpu *cpu_buffer;
3113 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3116 cpu_buffer = buffer->buffers[cpu];
3117 atomic_inc(&cpu_buffer->record_disabled);
3119 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3122 * ring_buffer_record_enable_cpu - enable writes to the buffer
3123 * @buffer: The ring buffer to enable writes
3124 * @cpu: The CPU to enable.
3126 * Note, multiple disables will need the same number of enables
3127 * to truly enable the writing (much like preempt_disable).
3129 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3131 struct ring_buffer_per_cpu *cpu_buffer;
3133 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3136 cpu_buffer = buffer->buffers[cpu];
3137 atomic_dec(&cpu_buffer->record_disabled);
3139 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3142 * The total entries in the ring buffer is the running counter
3143 * of entries entered into the ring buffer, minus the sum of
3144 * the entries read from the ring buffer and the number of
3145 * entries that were overwritten.
3147 static inline unsigned long
3148 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3150 return local_read(&cpu_buffer->entries) -
3151 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3155 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3156 * @buffer: The ring buffer
3157 * @cpu: The per CPU buffer to read from.
3159 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3161 unsigned long flags;
3162 struct ring_buffer_per_cpu *cpu_buffer;
3163 struct buffer_page *bpage;
3166 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3169 cpu_buffer = buffer->buffers[cpu];
3170 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3172 * if the tail is on reader_page, oldest time stamp is on the reader
3175 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3176 bpage = cpu_buffer->reader_page;
3178 bpage = rb_set_head_page(cpu_buffer);
3180 ret = bpage->page->time_stamp;
3181 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3185 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3188 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3189 * @buffer: The ring buffer
3190 * @cpu: The per CPU buffer to read from.
3192 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3194 struct ring_buffer_per_cpu *cpu_buffer;
3197 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3200 cpu_buffer = buffer->buffers[cpu];
3201 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3205 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3208 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3209 * @buffer: The ring buffer
3210 * @cpu: The per CPU buffer to get the entries from.
3212 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3214 struct ring_buffer_per_cpu *cpu_buffer;
3216 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3219 cpu_buffer = buffer->buffers[cpu];
3221 return rb_num_of_entries(cpu_buffer);
3223 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3226 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3227 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3228 * @buffer: The ring buffer
3229 * @cpu: The per CPU buffer to get the number of overruns from
3231 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3233 struct ring_buffer_per_cpu *cpu_buffer;
3236 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3239 cpu_buffer = buffer->buffers[cpu];
3240 ret = local_read(&cpu_buffer->overrun);
3244 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3247 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3248 * commits failing due to the buffer wrapping around while there are uncommitted
3249 * events, such as during an interrupt storm.
3250 * @buffer: The ring buffer
3251 * @cpu: The per CPU buffer to get the number of overruns from
3254 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3256 struct ring_buffer_per_cpu *cpu_buffer;
3259 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3262 cpu_buffer = buffer->buffers[cpu];
3263 ret = local_read(&cpu_buffer->commit_overrun);
3267 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3270 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3271 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3272 * @buffer: The ring buffer
3273 * @cpu: The per CPU buffer to get the number of overruns from
3276 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3278 struct ring_buffer_per_cpu *cpu_buffer;
3281 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3284 cpu_buffer = buffer->buffers[cpu];
3285 ret = local_read(&cpu_buffer->dropped_events);
3289 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3292 * ring_buffer_read_events_cpu - get the number of events successfully read
3293 * @buffer: The ring buffer
3294 * @cpu: The per CPU buffer to get the number of events read
3297 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3299 struct ring_buffer_per_cpu *cpu_buffer;
3301 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3304 cpu_buffer = buffer->buffers[cpu];
3305 return cpu_buffer->read;
3307 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3310 * ring_buffer_entries - get the number of entries in a buffer
3311 * @buffer: The ring buffer
3313 * Returns the total number of entries in the ring buffer
3316 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3318 struct ring_buffer_per_cpu *cpu_buffer;
3319 unsigned long entries = 0;
3322 /* if you care about this being correct, lock the buffer */
3323 for_each_buffer_cpu(buffer, cpu) {
3324 cpu_buffer = buffer->buffers[cpu];
3325 entries += rb_num_of_entries(cpu_buffer);
3330 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3333 * ring_buffer_overruns - get the number of overruns in buffer
3334 * @buffer: The ring buffer
3336 * Returns the total number of overruns in the ring buffer
3339 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3341 struct ring_buffer_per_cpu *cpu_buffer;
3342 unsigned long overruns = 0;
3345 /* if you care about this being correct, lock the buffer */
3346 for_each_buffer_cpu(buffer, cpu) {
3347 cpu_buffer = buffer->buffers[cpu];
3348 overruns += local_read(&cpu_buffer->overrun);
3353 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3355 static void rb_iter_reset(struct ring_buffer_iter *iter)
3357 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3359 /* Iterator usage is expected to have record disabled */
3360 if (list_empty(&cpu_buffer->reader_page->list)) {
3361 iter->head_page = rb_set_head_page(cpu_buffer);
3362 if (unlikely(!iter->head_page))
3364 iter->head = iter->head_page->read;
3366 iter->head_page = cpu_buffer->reader_page;
3367 iter->head = cpu_buffer->reader_page->read;
3370 iter->read_stamp = cpu_buffer->read_stamp;
3372 iter->read_stamp = iter->head_page->page->time_stamp;
3373 iter->cache_reader_page = cpu_buffer->reader_page;
3374 iter->cache_read = cpu_buffer->read;
3378 * ring_buffer_iter_reset - reset an iterator
3379 * @iter: The iterator to reset
3381 * Resets the iterator, so that it will start from the beginning
3384 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3386 struct ring_buffer_per_cpu *cpu_buffer;
3387 unsigned long flags;
3392 cpu_buffer = iter->cpu_buffer;
3394 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3395 rb_iter_reset(iter);
3396 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3398 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3401 * ring_buffer_iter_empty - check if an iterator has no more to read
3402 * @iter: The iterator to check
3404 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3406 struct ring_buffer_per_cpu *cpu_buffer;
3408 cpu_buffer = iter->cpu_buffer;
3410 return iter->head_page == cpu_buffer->commit_page &&
3411 iter->head == rb_commit_index(cpu_buffer);
3413 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3416 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3417 struct ring_buffer_event *event)
3421 switch (event->type_len) {
3422 case RINGBUF_TYPE_PADDING:
3425 case RINGBUF_TYPE_TIME_EXTEND:
3426 delta = event->array[0];
3428 delta += event->time_delta;
3429 cpu_buffer->read_stamp += delta;
3432 case RINGBUF_TYPE_TIME_STAMP:
3433 /* FIXME: not implemented */
3436 case RINGBUF_TYPE_DATA:
3437 cpu_buffer->read_stamp += event->time_delta;
3447 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3448 struct ring_buffer_event *event)
3452 switch (event->type_len) {
3453 case RINGBUF_TYPE_PADDING:
3456 case RINGBUF_TYPE_TIME_EXTEND:
3457 delta = event->array[0];
3459 delta += event->time_delta;
3460 iter->read_stamp += delta;
3463 case RINGBUF_TYPE_TIME_STAMP:
3464 /* FIXME: not implemented */
3467 case RINGBUF_TYPE_DATA:
3468 iter->read_stamp += event->time_delta;
3477 static struct buffer_page *
3478 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3480 struct buffer_page *reader = NULL;
3481 unsigned long overwrite;
3482 unsigned long flags;
3486 local_irq_save(flags);
3487 arch_spin_lock(&cpu_buffer->lock);
3491 * This should normally only loop twice. But because the
3492 * start of the reader inserts an empty page, it causes
3493 * a case where we will loop three times. There should be no
3494 * reason to loop four times (that I know of).
3496 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3501 reader = cpu_buffer->reader_page;
3503 /* If there's more to read, return this page */
3504 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3507 /* Never should we have an index greater than the size */
3508 if (RB_WARN_ON(cpu_buffer,
3509 cpu_buffer->reader_page->read > rb_page_size(reader)))
3512 /* check if we caught up to the tail */
3514 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3517 /* Don't bother swapping if the ring buffer is empty */
3518 if (rb_num_of_entries(cpu_buffer) == 0)
3522 * Reset the reader page to size zero.
3524 local_set(&cpu_buffer->reader_page->write, 0);
3525 local_set(&cpu_buffer->reader_page->entries, 0);
3526 local_set(&cpu_buffer->reader_page->page->commit, 0);
3527 cpu_buffer->reader_page->real_end = 0;
3531 * Splice the empty reader page into the list around the head.
3533 reader = rb_set_head_page(cpu_buffer);
3536 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3537 cpu_buffer->reader_page->list.prev = reader->list.prev;
3540 * cpu_buffer->pages just needs to point to the buffer, it
3541 * has no specific buffer page to point to. Lets move it out
3542 * of our way so we don't accidentally swap it.
3544 cpu_buffer->pages = reader->list.prev;
3546 /* The reader page will be pointing to the new head */
3547 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3550 * We want to make sure we read the overruns after we set up our
3551 * pointers to the next object. The writer side does a
3552 * cmpxchg to cross pages which acts as the mb on the writer
3553 * side. Note, the reader will constantly fail the swap
3554 * while the writer is updating the pointers, so this
3555 * guarantees that the overwrite recorded here is the one we
3556 * want to compare with the last_overrun.
3559 overwrite = local_read(&(cpu_buffer->overrun));
3562 * Here's the tricky part.
3564 * We need to move the pointer past the header page.
3565 * But we can only do that if a writer is not currently
3566 * moving it. The page before the header page has the
3567 * flag bit '1' set if it is pointing to the page we want.
3568 * but if the writer is in the process of moving it
3569 * than it will be '2' or already moved '0'.
3572 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3575 * If we did not convert it, then we must try again.
3581 * Yeah! We succeeded in replacing the page.
3583 * Now make the new head point back to the reader page.
3585 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3586 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3588 /* Finally update the reader page to the new head */
3589 cpu_buffer->reader_page = reader;
3590 rb_reset_reader_page(cpu_buffer);
3592 if (overwrite != cpu_buffer->last_overrun) {
3593 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3594 cpu_buffer->last_overrun = overwrite;
3600 arch_spin_unlock(&cpu_buffer->lock);
3601 local_irq_restore(flags);
3606 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3608 struct ring_buffer_event *event;
3609 struct buffer_page *reader;
3612 reader = rb_get_reader_page(cpu_buffer);
3614 /* This function should not be called when buffer is empty */
3615 if (RB_WARN_ON(cpu_buffer, !reader))
3618 event = rb_reader_event(cpu_buffer);
3620 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3623 rb_update_read_stamp(cpu_buffer, event);
3625 length = rb_event_length(event);
3626 cpu_buffer->reader_page->read += length;
3629 static void rb_advance_iter(struct ring_buffer_iter *iter)
3631 struct ring_buffer_per_cpu *cpu_buffer;
3632 struct ring_buffer_event *event;
3635 cpu_buffer = iter->cpu_buffer;
3638 * Check if we are at the end of the buffer.
3640 if (iter->head >= rb_page_size(iter->head_page)) {
3641 /* discarded commits can make the page empty */
3642 if (iter->head_page == cpu_buffer->commit_page)
3648 event = rb_iter_head_event(iter);
3650 length = rb_event_length(event);
3653 * This should not be called to advance the header if we are
3654 * at the tail of the buffer.
3656 if (RB_WARN_ON(cpu_buffer,
3657 (iter->head_page == cpu_buffer->commit_page) &&
3658 (iter->head + length > rb_commit_index(cpu_buffer))))
3661 rb_update_iter_read_stamp(iter, event);
3663 iter->head += length;
3665 /* check for end of page padding */
3666 if ((iter->head >= rb_page_size(iter->head_page)) &&
3667 (iter->head_page != cpu_buffer->commit_page))
3671 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3673 return cpu_buffer->lost_events;
3676 static struct ring_buffer_event *
3677 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3678 unsigned long *lost_events)
3680 struct ring_buffer_event *event;
3681 struct buffer_page *reader;
3686 * We repeat when a time extend is encountered.
3687 * Since the time extend is always attached to a data event,
3688 * we should never loop more than once.
3689 * (We never hit the following condition more than twice).
3691 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3694 reader = rb_get_reader_page(cpu_buffer);
3698 event = rb_reader_event(cpu_buffer);
3700 switch (event->type_len) {
3701 case RINGBUF_TYPE_PADDING:
3702 if (rb_null_event(event))
3703 RB_WARN_ON(cpu_buffer, 1);
3705 * Because the writer could be discarding every
3706 * event it creates (which would probably be bad)
3707 * if we were to go back to "again" then we may never
3708 * catch up, and will trigger the warn on, or lock
3709 * the box. Return the padding, and we will release
3710 * the current locks, and try again.
3714 case RINGBUF_TYPE_TIME_EXTEND:
3715 /* Internal data, OK to advance */
3716 rb_advance_reader(cpu_buffer);
3719 case RINGBUF_TYPE_TIME_STAMP:
3720 /* FIXME: not implemented */
3721 rb_advance_reader(cpu_buffer);
3724 case RINGBUF_TYPE_DATA:
3726 *ts = cpu_buffer->read_stamp + event->time_delta;
3727 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3728 cpu_buffer->cpu, ts);
3731 *lost_events = rb_lost_events(cpu_buffer);
3740 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3742 static struct ring_buffer_event *
3743 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3745 struct ring_buffer *buffer;
3746 struct ring_buffer_per_cpu *cpu_buffer;
3747 struct ring_buffer_event *event;
3750 cpu_buffer = iter->cpu_buffer;
3751 buffer = cpu_buffer->buffer;
3754 * Check if someone performed a consuming read to
3755 * the buffer. A consuming read invalidates the iterator
3756 * and we need to reset the iterator in this case.
3758 if (unlikely(iter->cache_read != cpu_buffer->read ||
3759 iter->cache_reader_page != cpu_buffer->reader_page))
3760 rb_iter_reset(iter);
3763 if (ring_buffer_iter_empty(iter))
3767 * We repeat when a time extend is encountered.
3768 * Since the time extend is always attached to a data event,
3769 * we should never loop more than once.
3770 * (We never hit the following condition more than twice).
3772 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3775 if (rb_per_cpu_empty(cpu_buffer))
3778 if (iter->head >= local_read(&iter->head_page->page->commit)) {
3783 event = rb_iter_head_event(iter);
3785 switch (event->type_len) {
3786 case RINGBUF_TYPE_PADDING:
3787 if (rb_null_event(event)) {
3791 rb_advance_iter(iter);
3794 case RINGBUF_TYPE_TIME_EXTEND:
3795 /* Internal data, OK to advance */
3796 rb_advance_iter(iter);
3799 case RINGBUF_TYPE_TIME_STAMP:
3800 /* FIXME: not implemented */
3801 rb_advance_iter(iter);
3804 case RINGBUF_TYPE_DATA:
3806 *ts = iter->read_stamp + event->time_delta;
3807 ring_buffer_normalize_time_stamp(buffer,
3808 cpu_buffer->cpu, ts);
3818 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3820 static inline int rb_ok_to_lock(void)
3823 * If an NMI die dumps out the content of the ring buffer
3824 * do not grab locks. We also permanently disable the ring
3825 * buffer too. A one time deal is all you get from reading
3826 * the ring buffer from an NMI.
3828 if (likely(!in_nmi()))
3831 tracing_off_permanent();
3836 * ring_buffer_peek - peek at the next event to be read
3837 * @buffer: The ring buffer to read
3838 * @cpu: The cpu to peak at
3839 * @ts: The timestamp counter of this event.
3840 * @lost_events: a variable to store if events were lost (may be NULL)
3842 * This will return the event that will be read next, but does
3843 * not consume the data.
3845 struct ring_buffer_event *
3846 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3847 unsigned long *lost_events)
3849 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3850 struct ring_buffer_event *event;
3851 unsigned long flags;
3854 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3857 dolock = rb_ok_to_lock();
3859 local_irq_save(flags);
3861 raw_spin_lock(&cpu_buffer->reader_lock);
3862 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3863 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3864 rb_advance_reader(cpu_buffer);
3866 raw_spin_unlock(&cpu_buffer->reader_lock);
3867 local_irq_restore(flags);
3869 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3876 * ring_buffer_iter_peek - peek at the next event to be read
3877 * @iter: The ring buffer iterator
3878 * @ts: The timestamp counter of this event.
3880 * This will return the event that will be read next, but does
3881 * not increment the iterator.
3883 struct ring_buffer_event *
3884 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3886 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3887 struct ring_buffer_event *event;
3888 unsigned long flags;
3891 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3892 event = rb_iter_peek(iter, ts);
3893 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3895 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3902 * ring_buffer_consume - return an event and consume it
3903 * @buffer: The ring buffer to get the next event from
3904 * @cpu: the cpu to read the buffer from
3905 * @ts: a variable to store the timestamp (may be NULL)
3906 * @lost_events: a variable to store if events were lost (may be NULL)
3908 * Returns the next event in the ring buffer, and that event is consumed.
3909 * Meaning, that sequential reads will keep returning a different event,
3910 * and eventually empty the ring buffer if the producer is slower.
3912 struct ring_buffer_event *
3913 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3914 unsigned long *lost_events)
3916 struct ring_buffer_per_cpu *cpu_buffer;
3917 struct ring_buffer_event *event = NULL;
3918 unsigned long flags;
3921 dolock = rb_ok_to_lock();
3924 /* might be called in atomic */
3927 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3930 cpu_buffer = buffer->buffers[cpu];
3931 local_irq_save(flags);
3933 raw_spin_lock(&cpu_buffer->reader_lock);
3935 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3937 cpu_buffer->lost_events = 0;
3938 rb_advance_reader(cpu_buffer);
3942 raw_spin_unlock(&cpu_buffer->reader_lock);
3943 local_irq_restore(flags);
3948 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3953 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3956 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3957 * @buffer: The ring buffer to read from
3958 * @cpu: The cpu buffer to iterate over
3960 * This performs the initial preparations necessary to iterate
3961 * through the buffer. Memory is allocated, buffer recording
3962 * is disabled, and the iterator pointer is returned to the caller.
3964 * Disabling buffer recordng prevents the reading from being
3965 * corrupted. This is not a consuming read, so a producer is not
3968 * After a sequence of ring_buffer_read_prepare calls, the user is
3969 * expected to make at least one call to ring_buffer_read_prepare_sync.
3970 * Afterwards, ring_buffer_read_start is invoked to get things going
3973 * This overall must be paired with ring_buffer_read_finish.
3975 struct ring_buffer_iter *
3976 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3978 struct ring_buffer_per_cpu *cpu_buffer;
3979 struct ring_buffer_iter *iter;
3981 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3984 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3988 cpu_buffer = buffer->buffers[cpu];
3990 iter->cpu_buffer = cpu_buffer;
3992 atomic_inc(&buffer->resize_disabled);
3993 atomic_inc(&cpu_buffer->record_disabled);
3997 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4000 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4002 * All previously invoked ring_buffer_read_prepare calls to prepare
4003 * iterators will be synchronized. Afterwards, read_buffer_read_start
4004 * calls on those iterators are allowed.
4007 ring_buffer_read_prepare_sync(void)
4009 synchronize_sched();
4011 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4014 * ring_buffer_read_start - start a non consuming read of the buffer
4015 * @iter: The iterator returned by ring_buffer_read_prepare
4017 * This finalizes the startup of an iteration through the buffer.
4018 * The iterator comes from a call to ring_buffer_read_prepare and
4019 * an intervening ring_buffer_read_prepare_sync must have been
4022 * Must be paired with ring_buffer_read_finish.
4025 ring_buffer_read_start(struct ring_buffer_iter *iter)
4027 struct ring_buffer_per_cpu *cpu_buffer;
4028 unsigned long flags;
4033 cpu_buffer = iter->cpu_buffer;
4035 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4036 arch_spin_lock(&cpu_buffer->lock);
4037 rb_iter_reset(iter);
4038 arch_spin_unlock(&cpu_buffer->lock);
4039 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4041 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4044 * ring_buffer_read_finish - finish reading the iterator of the buffer
4045 * @iter: The iterator retrieved by ring_buffer_start
4047 * This re-enables the recording to the buffer, and frees the
4051 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4053 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4054 unsigned long flags;
4057 * Ring buffer is disabled from recording, here's a good place
4058 * to check the integrity of the ring buffer.
4059 * Must prevent readers from trying to read, as the check
4060 * clears the HEAD page and readers require it.
4062 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4063 rb_check_pages(cpu_buffer);
4064 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4066 atomic_dec(&cpu_buffer->record_disabled);
4067 atomic_dec(&cpu_buffer->buffer->resize_disabled);
4070 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4073 * ring_buffer_read - read the next item in the ring buffer by the iterator
4074 * @iter: The ring buffer iterator
4075 * @ts: The time stamp of the event read.
4077 * This reads the next event in the ring buffer and increments the iterator.
4079 struct ring_buffer_event *
4080 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4082 struct ring_buffer_event *event;
4083 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4084 unsigned long flags;
4086 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4088 event = rb_iter_peek(iter, ts);
4092 if (event->type_len == RINGBUF_TYPE_PADDING)
4095 rb_advance_iter(iter);
4097 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4101 EXPORT_SYMBOL_GPL(ring_buffer_read);
4104 * ring_buffer_size - return the size of the ring buffer (in bytes)
4105 * @buffer: The ring buffer.
4107 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4110 * Earlier, this method returned
4111 * BUF_PAGE_SIZE * buffer->nr_pages
4112 * Since the nr_pages field is now removed, we have converted this to
4113 * return the per cpu buffer value.
4115 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4118 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4120 EXPORT_SYMBOL_GPL(ring_buffer_size);
4123 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4125 rb_head_page_deactivate(cpu_buffer);
4127 cpu_buffer->head_page
4128 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4129 local_set(&cpu_buffer->head_page->write, 0);
4130 local_set(&cpu_buffer->head_page->entries, 0);
4131 local_set(&cpu_buffer->head_page->page->commit, 0);
4133 cpu_buffer->head_page->read = 0;
4135 cpu_buffer->tail_page = cpu_buffer->head_page;
4136 cpu_buffer->commit_page = cpu_buffer->head_page;
4138 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4139 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4140 local_set(&cpu_buffer->reader_page->write, 0);
4141 local_set(&cpu_buffer->reader_page->entries, 0);
4142 local_set(&cpu_buffer->reader_page->page->commit, 0);
4143 cpu_buffer->reader_page->read = 0;
4145 local_set(&cpu_buffer->entries_bytes, 0);
4146 local_set(&cpu_buffer->overrun, 0);
4147 local_set(&cpu_buffer->commit_overrun, 0);
4148 local_set(&cpu_buffer->dropped_events, 0);
4149 local_set(&cpu_buffer->entries, 0);
4150 local_set(&cpu_buffer->committing, 0);
4151 local_set(&cpu_buffer->commits, 0);
4152 cpu_buffer->read = 0;
4153 cpu_buffer->read_bytes = 0;
4155 cpu_buffer->write_stamp = 0;
4156 cpu_buffer->read_stamp = 0;
4158 cpu_buffer->lost_events = 0;
4159 cpu_buffer->last_overrun = 0;
4161 rb_head_page_activate(cpu_buffer);
4165 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4166 * @buffer: The ring buffer to reset a per cpu buffer of
4167 * @cpu: The CPU buffer to be reset
4169 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4171 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4172 unsigned long flags;
4174 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4177 atomic_inc(&buffer->resize_disabled);
4178 atomic_inc(&cpu_buffer->record_disabled);
4180 /* Make sure all commits have finished */
4181 synchronize_sched();
4183 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4185 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4188 arch_spin_lock(&cpu_buffer->lock);
4190 rb_reset_cpu(cpu_buffer);
4192 arch_spin_unlock(&cpu_buffer->lock);
4195 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4197 atomic_dec(&cpu_buffer->record_disabled);
4198 atomic_dec(&buffer->resize_disabled);
4200 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4203 * ring_buffer_reset - reset a ring buffer
4204 * @buffer: The ring buffer to reset all cpu buffers
4206 void ring_buffer_reset(struct ring_buffer *buffer)
4210 for_each_buffer_cpu(buffer, cpu)
4211 ring_buffer_reset_cpu(buffer, cpu);
4213 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4216 * rind_buffer_empty - is the ring buffer empty?
4217 * @buffer: The ring buffer to test
4219 int ring_buffer_empty(struct ring_buffer *buffer)
4221 struct ring_buffer_per_cpu *cpu_buffer;
4222 unsigned long flags;
4227 dolock = rb_ok_to_lock();
4229 /* yes this is racy, but if you don't like the race, lock the buffer */
4230 for_each_buffer_cpu(buffer, cpu) {
4231 cpu_buffer = buffer->buffers[cpu];
4232 local_irq_save(flags);
4234 raw_spin_lock(&cpu_buffer->reader_lock);
4235 ret = rb_per_cpu_empty(cpu_buffer);
4237 raw_spin_unlock(&cpu_buffer->reader_lock);
4238 local_irq_restore(flags);
4246 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4249 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4250 * @buffer: The ring buffer
4251 * @cpu: The CPU buffer to test
4253 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4255 struct ring_buffer_per_cpu *cpu_buffer;
4256 unsigned long flags;
4260 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4263 dolock = rb_ok_to_lock();
4265 cpu_buffer = buffer->buffers[cpu];
4266 local_irq_save(flags);
4268 raw_spin_lock(&cpu_buffer->reader_lock);
4269 ret = rb_per_cpu_empty(cpu_buffer);
4271 raw_spin_unlock(&cpu_buffer->reader_lock);
4272 local_irq_restore(flags);
4276 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4278 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4280 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4281 * @buffer_a: One buffer to swap with
4282 * @buffer_b: The other buffer to swap with
4284 * This function is useful for tracers that want to take a "snapshot"
4285 * of a CPU buffer and has another back up buffer lying around.
4286 * it is expected that the tracer handles the cpu buffer not being
4287 * used at the moment.
4289 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4290 struct ring_buffer *buffer_b, int cpu)
4292 struct ring_buffer_per_cpu *cpu_buffer_a;
4293 struct ring_buffer_per_cpu *cpu_buffer_b;
4296 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4297 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4300 cpu_buffer_a = buffer_a->buffers[cpu];
4301 cpu_buffer_b = buffer_b->buffers[cpu];
4303 /* At least make sure the two buffers are somewhat the same */
4304 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4309 if (ring_buffer_flags != RB_BUFFERS_ON)
4312 if (atomic_read(&buffer_a->record_disabled))
4315 if (atomic_read(&buffer_b->record_disabled))
4318 if (atomic_read(&cpu_buffer_a->record_disabled))
4321 if (atomic_read(&cpu_buffer_b->record_disabled))
4325 * We can't do a synchronize_sched here because this
4326 * function can be called in atomic context.
4327 * Normally this will be called from the same CPU as cpu.
4328 * If not it's up to the caller to protect this.
4330 atomic_inc(&cpu_buffer_a->record_disabled);
4331 atomic_inc(&cpu_buffer_b->record_disabled);
4334 if (local_read(&cpu_buffer_a->committing))
4336 if (local_read(&cpu_buffer_b->committing))
4339 buffer_a->buffers[cpu] = cpu_buffer_b;
4340 buffer_b->buffers[cpu] = cpu_buffer_a;
4342 cpu_buffer_b->buffer = buffer_a;
4343 cpu_buffer_a->buffer = buffer_b;
4348 atomic_dec(&cpu_buffer_a->record_disabled);
4349 atomic_dec(&cpu_buffer_b->record_disabled);
4353 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4354 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4357 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4358 * @buffer: the buffer to allocate for.
4359 * @cpu: the cpu buffer to allocate.
4361 * This function is used in conjunction with ring_buffer_read_page.
4362 * When reading a full page from the ring buffer, these functions
4363 * can be used to speed up the process. The calling function should
4364 * allocate a few pages first with this function. Then when it
4365 * needs to get pages from the ring buffer, it passes the result
4366 * of this function into ring_buffer_read_page, which will swap
4367 * the page that was allocated, with the read page of the buffer.
4370 * The page allocated, or NULL on error.
4372 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4374 struct buffer_data_page *bpage;
4377 page = alloc_pages_node(cpu_to_node(cpu),
4378 GFP_KERNEL | __GFP_NORETRY, 0);
4382 bpage = page_address(page);
4384 rb_init_page(bpage);
4388 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4391 * ring_buffer_free_read_page - free an allocated read page
4392 * @buffer: the buffer the page was allocate for
4393 * @data: the page to free
4395 * Free a page allocated from ring_buffer_alloc_read_page.
4397 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4399 free_page((unsigned long)data);
4401 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4404 * ring_buffer_read_page - extract a page from the ring buffer
4405 * @buffer: buffer to extract from
4406 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4407 * @len: amount to extract
4408 * @cpu: the cpu of the buffer to extract
4409 * @full: should the extraction only happen when the page is full.
4411 * This function will pull out a page from the ring buffer and consume it.
4412 * @data_page must be the address of the variable that was returned
4413 * from ring_buffer_alloc_read_page. This is because the page might be used
4414 * to swap with a page in the ring buffer.
4417 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4420 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4422 * process_page(rpage, ret);
4424 * When @full is set, the function will not return true unless
4425 * the writer is off the reader page.
4427 * Note: it is up to the calling functions to handle sleeps and wakeups.
4428 * The ring buffer can be used anywhere in the kernel and can not
4429 * blindly call wake_up. The layer that uses the ring buffer must be
4430 * responsible for that.
4433 * >=0 if data has been transferred, returns the offset of consumed data.
4434 * <0 if no data has been transferred.
4436 int ring_buffer_read_page(struct ring_buffer *buffer,
4437 void **data_page, size_t len, int cpu, int full)
4439 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4440 struct ring_buffer_event *event;
4441 struct buffer_data_page *bpage;
4442 struct buffer_page *reader;
4443 unsigned long missed_events;
4444 unsigned long flags;
4445 unsigned int commit;
4450 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4454 * If len is not big enough to hold the page header, then
4455 * we can not copy anything.
4457 if (len <= BUF_PAGE_HDR_SIZE)
4460 len -= BUF_PAGE_HDR_SIZE;
4469 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4471 reader = rb_get_reader_page(cpu_buffer);
4475 event = rb_reader_event(cpu_buffer);
4477 read = reader->read;
4478 commit = rb_page_commit(reader);
4480 /* Check if any events were dropped */
4481 missed_events = cpu_buffer->lost_events;
4484 * If this page has been partially read or
4485 * if len is not big enough to read the rest of the page or
4486 * a writer is still on the page, then
4487 * we must copy the data from the page to the buffer.
4488 * Otherwise, we can simply swap the page with the one passed in.
4490 if (read || (len < (commit - read)) ||
4491 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4492 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4493 unsigned int rpos = read;
4494 unsigned int pos = 0;
4500 if (len > (commit - read))
4501 len = (commit - read);
4503 /* Always keep the time extend and data together */
4504 size = rb_event_ts_length(event);
4509 /* save the current timestamp, since the user will need it */
4510 save_timestamp = cpu_buffer->read_stamp;
4512 /* Need to copy one event at a time */
4514 /* We need the size of one event, because
4515 * rb_advance_reader only advances by one event,
4516 * whereas rb_event_ts_length may include the size of
4517 * one or two events.
4518 * We have already ensured there's enough space if this
4519 * is a time extend. */
4520 size = rb_event_length(event);
4521 memcpy(bpage->data + pos, rpage->data + rpos, size);
4525 rb_advance_reader(cpu_buffer);
4526 rpos = reader->read;
4532 event = rb_reader_event(cpu_buffer);
4533 /* Always keep the time extend and data together */
4534 size = rb_event_ts_length(event);
4535 } while (len >= size);
4538 local_set(&bpage->commit, pos);
4539 bpage->time_stamp = save_timestamp;
4541 /* we copied everything to the beginning */
4544 /* update the entry counter */
4545 cpu_buffer->read += rb_page_entries(reader);
4546 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4548 /* swap the pages */
4549 rb_init_page(bpage);
4550 bpage = reader->page;
4551 reader->page = *data_page;
4552 local_set(&reader->write, 0);
4553 local_set(&reader->entries, 0);
4558 * Use the real_end for the data size,
4559 * This gives us a chance to store the lost events
4562 if (reader->real_end)
4563 local_set(&bpage->commit, reader->real_end);
4567 cpu_buffer->lost_events = 0;
4569 commit = local_read(&bpage->commit);
4571 * Set a flag in the commit field if we lost events
4573 if (missed_events) {
4574 /* If there is room at the end of the page to save the
4575 * missed events, then record it there.
4577 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4578 memcpy(&bpage->data[commit], &missed_events,
4579 sizeof(missed_events));
4580 local_add(RB_MISSED_STORED, &bpage->commit);
4581 commit += sizeof(missed_events);
4583 local_add(RB_MISSED_EVENTS, &bpage->commit);
4587 * This page may be off to user land. Zero it out here.
4589 if (commit < BUF_PAGE_SIZE)
4590 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4593 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4598 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4600 #ifdef CONFIG_HOTPLUG_CPU
4601 static int rb_cpu_notify(struct notifier_block *self,
4602 unsigned long action, void *hcpu)
4604 struct ring_buffer *buffer =
4605 container_of(self, struct ring_buffer, cpu_notify);
4606 long cpu = (long)hcpu;
4607 int cpu_i, nr_pages_same;
4608 unsigned int nr_pages;
4611 case CPU_UP_PREPARE:
4612 case CPU_UP_PREPARE_FROZEN:
4613 if (cpumask_test_cpu(cpu, buffer->cpumask))
4618 /* check if all cpu sizes are same */
4619 for_each_buffer_cpu(buffer, cpu_i) {
4620 /* fill in the size from first enabled cpu */
4622 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4623 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4628 /* allocate minimum pages, user can later expand it */
4631 buffer->buffers[cpu] =
4632 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4633 if (!buffer->buffers[cpu]) {
4634 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4639 cpumask_set_cpu(cpu, buffer->cpumask);
4641 case CPU_DOWN_PREPARE:
4642 case CPU_DOWN_PREPARE_FROZEN:
4645 * If we were to free the buffer, then the user would
4646 * lose any trace that was in the buffer.
4656 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4658 * This is a basic integrity check of the ring buffer.
4659 * Late in the boot cycle this test will run when configured in.
4660 * It will kick off a thread per CPU that will go into a loop
4661 * writing to the per cpu ring buffer various sizes of data.
4662 * Some of the data will be large items, some small.
4664 * Another thread is created that goes into a spin, sending out
4665 * IPIs to the other CPUs to also write into the ring buffer.
4666 * this is to test the nesting ability of the buffer.
4668 * Basic stats are recorded and reported. If something in the
4669 * ring buffer should happen that's not expected, a big warning
4670 * is displayed and all ring buffers are disabled.
4672 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4674 struct rb_test_data {
4675 struct ring_buffer *buffer;
4676 unsigned long events;
4677 unsigned long bytes_written;
4678 unsigned long bytes_alloc;
4679 unsigned long bytes_dropped;
4680 unsigned long events_nested;
4681 unsigned long bytes_written_nested;
4682 unsigned long bytes_alloc_nested;
4683 unsigned long bytes_dropped_nested;
4684 int min_size_nested;
4685 int max_size_nested;
4692 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4695 #define RB_TEST_BUFFER_SIZE 1048576
4697 static char rb_string[] __initdata =
4698 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4699 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4700 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4702 static bool rb_test_started __initdata;
4709 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4711 struct ring_buffer_event *event;
4712 struct rb_item *item;
4719 /* Have nested writes different that what is written */
4720 cnt = data->cnt + (nested ? 27 : 0);
4722 /* Multiply cnt by ~e, to make some unique increment */
4723 size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4725 len = size + sizeof(struct rb_item);
4727 started = rb_test_started;
4728 /* read rb_test_started before checking buffer enabled */
4731 event = ring_buffer_lock_reserve(data->buffer, len);
4733 /* Ignore dropped events before test starts. */
4736 data->bytes_dropped += len;
4738 data->bytes_dropped_nested += len;
4743 event_len = ring_buffer_event_length(event);
4745 if (RB_WARN_ON(data->buffer, event_len < len))
4748 item = ring_buffer_event_data(event);
4750 memcpy(item->str, rb_string, size);
4753 data->bytes_alloc_nested += event_len;
4754 data->bytes_written_nested += len;
4755 data->events_nested++;
4756 if (!data->min_size_nested || len < data->min_size_nested)
4757 data->min_size_nested = len;
4758 if (len > data->max_size_nested)
4759 data->max_size_nested = len;
4761 data->bytes_alloc += event_len;
4762 data->bytes_written += len;
4764 if (!data->min_size || len < data->min_size)
4765 data->max_size = len;
4766 if (len > data->max_size)
4767 data->max_size = len;
4771 ring_buffer_unlock_commit(data->buffer, event);
4776 static __init int rb_test(void *arg)
4778 struct rb_test_data *data = arg;
4780 while (!kthread_should_stop()) {
4781 rb_write_something(data, false);
4784 set_current_state(TASK_INTERRUPTIBLE);
4785 /* Now sleep between a min of 100-300us and a max of 1ms */
4786 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4792 static __init void rb_ipi(void *ignore)
4794 struct rb_test_data *data;
4795 int cpu = smp_processor_id();
4797 data = &rb_data[cpu];
4798 rb_write_something(data, true);
4801 static __init int rb_hammer_test(void *arg)
4803 while (!kthread_should_stop()) {
4805 /* Send an IPI to all cpus to write data! */
4806 smp_call_function(rb_ipi, NULL, 1);
4807 /* No sleep, but for non preempt, let others run */
4814 static __init int test_ringbuffer(void)
4816 struct task_struct *rb_hammer;
4817 struct ring_buffer *buffer;
4821 pr_info("Running ring buffer tests...\n");
4823 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4824 if (WARN_ON(!buffer))
4827 /* Disable buffer so that threads can't write to it yet */
4828 ring_buffer_record_off(buffer);
4830 for_each_online_cpu(cpu) {
4831 rb_data[cpu].buffer = buffer;
4832 rb_data[cpu].cpu = cpu;
4833 rb_data[cpu].cnt = cpu;
4834 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4835 "rbtester/%d", cpu);
4836 if (WARN_ON(!rb_threads[cpu])) {
4837 pr_cont("FAILED\n");
4842 kthread_bind(rb_threads[cpu], cpu);
4843 wake_up_process(rb_threads[cpu]);
4846 /* Now create the rb hammer! */
4847 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4848 if (WARN_ON(!rb_hammer)) {
4849 pr_cont("FAILED\n");
4854 ring_buffer_record_on(buffer);
4856 * Show buffer is enabled before setting rb_test_started.
4857 * Yes there's a small race window where events could be
4858 * dropped and the thread wont catch it. But when a ring
4859 * buffer gets enabled, there will always be some kind of
4860 * delay before other CPUs see it. Thus, we don't care about
4861 * those dropped events. We care about events dropped after
4862 * the threads see that the buffer is active.
4865 rb_test_started = true;
4867 set_current_state(TASK_INTERRUPTIBLE);
4868 /* Just run for 10 seconds */;
4869 schedule_timeout(10 * HZ);
4871 kthread_stop(rb_hammer);
4874 for_each_online_cpu(cpu) {
4875 if (!rb_threads[cpu])
4877 kthread_stop(rb_threads[cpu]);
4880 ring_buffer_free(buffer);
4885 pr_info("finished\n");
4886 for_each_online_cpu(cpu) {
4887 struct ring_buffer_event *event;
4888 struct rb_test_data *data = &rb_data[cpu];
4889 struct rb_item *item;
4890 unsigned long total_events;
4891 unsigned long total_dropped;
4892 unsigned long total_written;
4893 unsigned long total_alloc;
4894 unsigned long total_read = 0;
4895 unsigned long total_size = 0;
4896 unsigned long total_len = 0;
4897 unsigned long total_lost = 0;
4900 int small_event_size;
4904 total_events = data->events + data->events_nested;
4905 total_written = data->bytes_written + data->bytes_written_nested;
4906 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
4907 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
4909 big_event_size = data->max_size + data->max_size_nested;
4910 small_event_size = data->min_size + data->min_size_nested;
4912 pr_info("CPU %d:\n", cpu);
4913 pr_info(" events: %ld\n", total_events);
4914 pr_info(" dropped bytes: %ld\n", total_dropped);
4915 pr_info(" alloced bytes: %ld\n", total_alloc);
4916 pr_info(" written bytes: %ld\n", total_written);
4917 pr_info(" biggest event: %d\n", big_event_size);
4918 pr_info(" smallest event: %d\n", small_event_size);
4920 if (RB_WARN_ON(buffer, total_dropped))
4925 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
4927 item = ring_buffer_event_data(event);
4928 total_len += ring_buffer_event_length(event);
4929 total_size += item->size + sizeof(struct rb_item);
4930 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
4931 pr_info("FAILED!\n");
4932 pr_info("buffer had: %.*s\n", item->size, item->str);
4933 pr_info("expected: %.*s\n", item->size, rb_string);
4934 RB_WARN_ON(buffer, 1);
4945 pr_info(" read events: %ld\n", total_read);
4946 pr_info(" lost events: %ld\n", total_lost);
4947 pr_info(" total events: %ld\n", total_lost + total_read);
4948 pr_info(" recorded len bytes: %ld\n", total_len);
4949 pr_info(" recorded size bytes: %ld\n", total_size);
4951 pr_info(" With dropped events, record len and size may not match\n"
4952 " alloced and written from above\n");
4954 if (RB_WARN_ON(buffer, total_len != total_alloc ||
4955 total_size != total_written))
4958 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
4964 pr_info("Ring buffer PASSED!\n");
4966 ring_buffer_free(buffer);
4970 late_initcall(test_ringbuffer);
4971 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */