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 /* Can't run something on an offline CPU. */
1693 if (!cpu_online(cpu)) {
1694 rb_update_pages(cpu_buffer);
1695 cpu_buffer->nr_pages_to_update = 0;
1697 schedule_work_on(cpu,
1698 &cpu_buffer->update_pages_work);
1702 /* wait for all the updates to complete */
1703 for_each_buffer_cpu(buffer, cpu) {
1704 cpu_buffer = buffer->buffers[cpu];
1705 if (!cpu_buffer->nr_pages_to_update)
1708 if (cpu_online(cpu))
1709 wait_for_completion(&cpu_buffer->update_done);
1710 cpu_buffer->nr_pages_to_update = 0;
1715 /* Make sure this CPU has been intitialized */
1716 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1719 cpu_buffer = buffer->buffers[cpu_id];
1721 if (nr_pages == cpu_buffer->nr_pages)
1724 cpu_buffer->nr_pages_to_update = nr_pages -
1725 cpu_buffer->nr_pages;
1727 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1728 if (cpu_buffer->nr_pages_to_update > 0 &&
1729 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1730 &cpu_buffer->new_pages, cpu_id)) {
1737 /* Can't run something on an offline CPU. */
1738 if (!cpu_online(cpu_id))
1739 rb_update_pages(cpu_buffer);
1741 schedule_work_on(cpu_id,
1742 &cpu_buffer->update_pages_work);
1743 wait_for_completion(&cpu_buffer->update_done);
1746 cpu_buffer->nr_pages_to_update = 0;
1752 * The ring buffer resize can happen with the ring buffer
1753 * enabled, so that the update disturbs the tracing as little
1754 * as possible. But if the buffer is disabled, we do not need
1755 * to worry about that, and we can take the time to verify
1756 * that the buffer is not corrupt.
1758 if (atomic_read(&buffer->record_disabled)) {
1759 atomic_inc(&buffer->record_disabled);
1761 * Even though the buffer was disabled, we must make sure
1762 * that it is truly disabled before calling rb_check_pages.
1763 * There could have been a race between checking
1764 * record_disable and incrementing it.
1766 synchronize_sched();
1767 for_each_buffer_cpu(buffer, cpu) {
1768 cpu_buffer = buffer->buffers[cpu];
1769 rb_check_pages(cpu_buffer);
1771 atomic_dec(&buffer->record_disabled);
1774 mutex_unlock(&buffer->mutex);
1778 for_each_buffer_cpu(buffer, cpu) {
1779 struct buffer_page *bpage, *tmp;
1781 cpu_buffer = buffer->buffers[cpu];
1782 cpu_buffer->nr_pages_to_update = 0;
1784 if (list_empty(&cpu_buffer->new_pages))
1787 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1789 list_del_init(&bpage->list);
1790 free_buffer_page(bpage);
1793 mutex_unlock(&buffer->mutex);
1796 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1798 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1800 mutex_lock(&buffer->mutex);
1802 buffer->flags |= RB_FL_OVERWRITE;
1804 buffer->flags &= ~RB_FL_OVERWRITE;
1805 mutex_unlock(&buffer->mutex);
1807 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1809 static inline void *
1810 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1812 return bpage->data + index;
1815 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1817 return bpage->page->data + index;
1820 static inline struct ring_buffer_event *
1821 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1823 return __rb_page_index(cpu_buffer->reader_page,
1824 cpu_buffer->reader_page->read);
1827 static inline struct ring_buffer_event *
1828 rb_iter_head_event(struct ring_buffer_iter *iter)
1830 return __rb_page_index(iter->head_page, iter->head);
1833 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1835 return local_read(&bpage->page->commit);
1838 /* Size is determined by what has been committed */
1839 static inline unsigned rb_page_size(struct buffer_page *bpage)
1841 return rb_page_commit(bpage);
1844 static inline unsigned
1845 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1847 return rb_page_commit(cpu_buffer->commit_page);
1850 static inline unsigned
1851 rb_event_index(struct ring_buffer_event *event)
1853 unsigned long addr = (unsigned long)event;
1855 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1859 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1860 struct ring_buffer_event *event)
1862 unsigned long addr = (unsigned long)event;
1863 unsigned long index;
1865 index = rb_event_index(event);
1868 return cpu_buffer->commit_page->page == (void *)addr &&
1869 rb_commit_index(cpu_buffer) == index;
1873 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1875 unsigned long max_count;
1878 * We only race with interrupts and NMIs on this CPU.
1879 * If we own the commit event, then we can commit
1880 * all others that interrupted us, since the interruptions
1881 * are in stack format (they finish before they come
1882 * back to us). This allows us to do a simple loop to
1883 * assign the commit to the tail.
1886 max_count = cpu_buffer->nr_pages * 100;
1888 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1889 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1891 if (RB_WARN_ON(cpu_buffer,
1892 rb_is_reader_page(cpu_buffer->tail_page)))
1894 local_set(&cpu_buffer->commit_page->page->commit,
1895 rb_page_write(cpu_buffer->commit_page));
1896 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1897 cpu_buffer->write_stamp =
1898 cpu_buffer->commit_page->page->time_stamp;
1899 /* add barrier to keep gcc from optimizing too much */
1902 while (rb_commit_index(cpu_buffer) !=
1903 rb_page_write(cpu_buffer->commit_page)) {
1905 local_set(&cpu_buffer->commit_page->page->commit,
1906 rb_page_write(cpu_buffer->commit_page));
1907 RB_WARN_ON(cpu_buffer,
1908 local_read(&cpu_buffer->commit_page->page->commit) &
1913 /* again, keep gcc from optimizing */
1917 * If an interrupt came in just after the first while loop
1918 * and pushed the tail page forward, we will be left with
1919 * a dangling commit that will never go forward.
1921 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1925 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1927 cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1928 cpu_buffer->reader_page->read = 0;
1931 static void rb_inc_iter(struct ring_buffer_iter *iter)
1933 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1936 * The iterator could be on the reader page (it starts there).
1937 * But the head could have moved, since the reader was
1938 * found. Check for this case and assign the iterator
1939 * to the head page instead of next.
1941 if (iter->head_page == cpu_buffer->reader_page)
1942 iter->head_page = rb_set_head_page(cpu_buffer);
1944 rb_inc_page(cpu_buffer, &iter->head_page);
1946 iter->read_stamp = iter->head_page->page->time_stamp;
1950 /* Slow path, do not inline */
1951 static noinline struct ring_buffer_event *
1952 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
1954 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
1956 /* Not the first event on the page? */
1957 if (rb_event_index(event)) {
1958 event->time_delta = delta & TS_MASK;
1959 event->array[0] = delta >> TS_SHIFT;
1961 /* nope, just zero it */
1962 event->time_delta = 0;
1963 event->array[0] = 0;
1966 return skip_time_extend(event);
1970 * rb_update_event - update event type and data
1971 * @event: the event to update
1972 * @type: the type of event
1973 * @length: the size of the event field in the ring buffer
1975 * Update the type and data fields of the event. The length
1976 * is the actual size that is written to the ring buffer,
1977 * and with this, we can determine what to place into the
1981 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
1982 struct ring_buffer_event *event, unsigned length,
1983 int add_timestamp, u64 delta)
1985 /* Only a commit updates the timestamp */
1986 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
1990 * If we need to add a timestamp, then we
1991 * add it to the start of the resevered space.
1993 if (unlikely(add_timestamp)) {
1994 event = rb_add_time_stamp(event, delta);
1995 length -= RB_LEN_TIME_EXTEND;
1999 event->time_delta = delta;
2000 length -= RB_EVNT_HDR_SIZE;
2001 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2002 event->type_len = 0;
2003 event->array[0] = length;
2005 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2009 * rb_handle_head_page - writer hit the head page
2011 * Returns: +1 to retry page
2016 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2017 struct buffer_page *tail_page,
2018 struct buffer_page *next_page)
2020 struct buffer_page *new_head;
2025 entries = rb_page_entries(next_page);
2028 * The hard part is here. We need to move the head
2029 * forward, and protect against both readers on
2030 * other CPUs and writers coming in via interrupts.
2032 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2036 * type can be one of four:
2037 * NORMAL - an interrupt already moved it for us
2038 * HEAD - we are the first to get here.
2039 * UPDATE - we are the interrupt interrupting
2041 * MOVED - a reader on another CPU moved the next
2042 * pointer to its reader page. Give up
2049 * We changed the head to UPDATE, thus
2050 * it is our responsibility to update
2053 local_add(entries, &cpu_buffer->overrun);
2054 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2057 * The entries will be zeroed out when we move the
2061 /* still more to do */
2064 case RB_PAGE_UPDATE:
2066 * This is an interrupt that interrupt the
2067 * previous update. Still more to do.
2070 case RB_PAGE_NORMAL:
2072 * An interrupt came in before the update
2073 * and processed this for us.
2074 * Nothing left to do.
2079 * The reader is on another CPU and just did
2080 * a swap with our next_page.
2085 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2090 * Now that we are here, the old head pointer is
2091 * set to UPDATE. This will keep the reader from
2092 * swapping the head page with the reader page.
2093 * The reader (on another CPU) will spin till
2096 * We just need to protect against interrupts
2097 * doing the job. We will set the next pointer
2098 * to HEAD. After that, we set the old pointer
2099 * to NORMAL, but only if it was HEAD before.
2100 * otherwise we are an interrupt, and only
2101 * want the outer most commit to reset it.
2103 new_head = next_page;
2104 rb_inc_page(cpu_buffer, &new_head);
2106 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2110 * Valid returns are:
2111 * HEAD - an interrupt came in and already set it.
2112 * NORMAL - One of two things:
2113 * 1) We really set it.
2114 * 2) A bunch of interrupts came in and moved
2115 * the page forward again.
2119 case RB_PAGE_NORMAL:
2123 RB_WARN_ON(cpu_buffer, 1);
2128 * It is possible that an interrupt came in,
2129 * set the head up, then more interrupts came in
2130 * and moved it again. When we get back here,
2131 * the page would have been set to NORMAL but we
2132 * just set it back to HEAD.
2134 * How do you detect this? Well, if that happened
2135 * the tail page would have moved.
2137 if (ret == RB_PAGE_NORMAL) {
2139 * If the tail had moved passed next, then we need
2140 * to reset the pointer.
2142 if (cpu_buffer->tail_page != tail_page &&
2143 cpu_buffer->tail_page != next_page)
2144 rb_head_page_set_normal(cpu_buffer, new_head,
2150 * If this was the outer most commit (the one that
2151 * changed the original pointer from HEAD to UPDATE),
2152 * then it is up to us to reset it to NORMAL.
2154 if (type == RB_PAGE_HEAD) {
2155 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2158 if (RB_WARN_ON(cpu_buffer,
2159 ret != RB_PAGE_UPDATE))
2166 static unsigned rb_calculate_event_length(unsigned length)
2168 struct ring_buffer_event event; /* Used only for sizeof array */
2170 /* zero length can cause confusions */
2174 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2175 length += sizeof(event.array[0]);
2177 length += RB_EVNT_HDR_SIZE;
2178 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2184 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2185 struct buffer_page *tail_page,
2186 unsigned long tail, unsigned long length)
2188 struct ring_buffer_event *event;
2191 * Only the event that crossed the page boundary
2192 * must fill the old tail_page with padding.
2194 if (tail >= BUF_PAGE_SIZE) {
2196 * If the page was filled, then we still need
2197 * to update the real_end. Reset it to zero
2198 * and the reader will ignore it.
2200 if (tail == BUF_PAGE_SIZE)
2201 tail_page->real_end = 0;
2203 local_sub(length, &tail_page->write);
2207 event = __rb_page_index(tail_page, tail);
2208 kmemcheck_annotate_bitfield(event, bitfield);
2210 /* account for padding bytes */
2211 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2214 * Save the original length to the meta data.
2215 * This will be used by the reader to add lost event
2218 tail_page->real_end = tail;
2221 * If this event is bigger than the minimum size, then
2222 * we need to be careful that we don't subtract the
2223 * write counter enough to allow another writer to slip
2225 * We put in a discarded commit instead, to make sure
2226 * that this space is not used again.
2228 * If we are less than the minimum size, we don't need to
2231 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2232 /* No room for any events */
2234 /* Mark the rest of the page with padding */
2235 rb_event_set_padding(event);
2237 /* Set the write back to the previous setting */
2238 local_sub(length, &tail_page->write);
2242 /* Put in a discarded event */
2243 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2244 event->type_len = RINGBUF_TYPE_PADDING;
2245 /* time delta must be non zero */
2246 event->time_delta = 1;
2248 /* Set write to end of buffer */
2249 length = (tail + length) - BUF_PAGE_SIZE;
2250 local_sub(length, &tail_page->write);
2254 * This is the slow path, force gcc not to inline it.
2256 static noinline struct ring_buffer_event *
2257 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2258 unsigned long length, unsigned long tail,
2259 struct buffer_page *tail_page, u64 ts)
2261 struct buffer_page *commit_page = cpu_buffer->commit_page;
2262 struct ring_buffer *buffer = cpu_buffer->buffer;
2263 struct buffer_page *next_page;
2266 next_page = tail_page;
2268 rb_inc_page(cpu_buffer, &next_page);
2271 * If for some reason, we had an interrupt storm that made
2272 * it all the way around the buffer, bail, and warn
2275 if (unlikely(next_page == commit_page)) {
2276 local_inc(&cpu_buffer->commit_overrun);
2281 * This is where the fun begins!
2283 * We are fighting against races between a reader that
2284 * could be on another CPU trying to swap its reader
2285 * page with the buffer head.
2287 * We are also fighting against interrupts coming in and
2288 * moving the head or tail on us as well.
2290 * If the next page is the head page then we have filled
2291 * the buffer, unless the commit page is still on the
2294 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2297 * If the commit is not on the reader page, then
2298 * move the header page.
2300 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2302 * If we are not in overwrite mode,
2303 * this is easy, just stop here.
2305 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2306 local_inc(&cpu_buffer->dropped_events);
2310 ret = rb_handle_head_page(cpu_buffer,
2319 * We need to be careful here too. The
2320 * commit page could still be on the reader
2321 * page. We could have a small buffer, and
2322 * have filled up the buffer with events
2323 * from interrupts and such, and wrapped.
2325 * Note, if the tail page is also the on the
2326 * reader_page, we let it move out.
2328 if (unlikely((cpu_buffer->commit_page !=
2329 cpu_buffer->tail_page) &&
2330 (cpu_buffer->commit_page ==
2331 cpu_buffer->reader_page))) {
2332 local_inc(&cpu_buffer->commit_overrun);
2338 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
2341 * Nested commits always have zero deltas, so
2342 * just reread the time stamp
2344 ts = rb_time_stamp(buffer);
2345 next_page->page->time_stamp = ts;
2350 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2352 /* fail and let the caller try again */
2353 return ERR_PTR(-EAGAIN);
2357 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2362 static struct ring_buffer_event *
2363 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2364 unsigned long length, u64 ts,
2365 u64 delta, int add_timestamp)
2367 struct buffer_page *tail_page;
2368 struct ring_buffer_event *event;
2369 unsigned long tail, write;
2372 * If the time delta since the last event is too big to
2373 * hold in the time field of the event, then we append a
2374 * TIME EXTEND event ahead of the data event.
2376 if (unlikely(add_timestamp))
2377 length += RB_LEN_TIME_EXTEND;
2379 tail_page = cpu_buffer->tail_page;
2380 write = local_add_return(length, &tail_page->write);
2382 /* set write to only the index of the write */
2383 write &= RB_WRITE_MASK;
2384 tail = write - length;
2387 * If this is the first commit on the page, then it has the same
2388 * timestamp as the page itself.
2393 /* See if we shot pass the end of this buffer page */
2394 if (unlikely(write > BUF_PAGE_SIZE))
2395 return rb_move_tail(cpu_buffer, length, tail,
2398 /* We reserved something on the buffer */
2400 event = __rb_page_index(tail_page, tail);
2401 kmemcheck_annotate_bitfield(event, bitfield);
2402 rb_update_event(cpu_buffer, event, length, add_timestamp, delta);
2404 local_inc(&tail_page->entries);
2407 * If this is the first commit on the page, then update
2411 tail_page->page->time_stamp = ts;
2413 /* account for these added bytes */
2414 local_add(length, &cpu_buffer->entries_bytes);
2420 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2421 struct ring_buffer_event *event)
2423 unsigned long new_index, old_index;
2424 struct buffer_page *bpage;
2425 unsigned long index;
2428 new_index = rb_event_index(event);
2429 old_index = new_index + rb_event_ts_length(event);
2430 addr = (unsigned long)event;
2433 bpage = cpu_buffer->tail_page;
2435 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2436 unsigned long write_mask =
2437 local_read(&bpage->write) & ~RB_WRITE_MASK;
2438 unsigned long event_length = rb_event_length(event);
2440 * This is on the tail page. It is possible that
2441 * a write could come in and move the tail page
2442 * and write to the next page. That is fine
2443 * because we just shorten what is on this page.
2445 old_index += write_mask;
2446 new_index += write_mask;
2447 index = local_cmpxchg(&bpage->write, old_index, new_index);
2448 if (index == old_index) {
2449 /* update counters */
2450 local_sub(event_length, &cpu_buffer->entries_bytes);
2455 /* could not discard */
2459 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2461 local_inc(&cpu_buffer->committing);
2462 local_inc(&cpu_buffer->commits);
2465 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2467 unsigned long commits;
2469 if (RB_WARN_ON(cpu_buffer,
2470 !local_read(&cpu_buffer->committing)))
2474 commits = local_read(&cpu_buffer->commits);
2475 /* synchronize with interrupts */
2477 if (local_read(&cpu_buffer->committing) == 1)
2478 rb_set_commit_to_write(cpu_buffer);
2480 local_dec(&cpu_buffer->committing);
2482 /* synchronize with interrupts */
2486 * Need to account for interrupts coming in between the
2487 * updating of the commit page and the clearing of the
2488 * committing counter.
2490 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2491 !local_read(&cpu_buffer->committing)) {
2492 local_inc(&cpu_buffer->committing);
2497 static struct ring_buffer_event *
2498 rb_reserve_next_event(struct ring_buffer *buffer,
2499 struct ring_buffer_per_cpu *cpu_buffer,
2500 unsigned long length)
2502 struct ring_buffer_event *event;
2508 rb_start_commit(cpu_buffer);
2510 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2512 * Due to the ability to swap a cpu buffer from a buffer
2513 * it is possible it was swapped before we committed.
2514 * (committing stops a swap). We check for it here and
2515 * if it happened, we have to fail the write.
2518 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2519 local_dec(&cpu_buffer->committing);
2520 local_dec(&cpu_buffer->commits);
2525 length = rb_calculate_event_length(length);
2531 * We allow for interrupts to reenter here and do a trace.
2532 * If one does, it will cause this original code to loop
2533 * back here. Even with heavy interrupts happening, this
2534 * should only happen a few times in a row. If this happens
2535 * 1000 times in a row, there must be either an interrupt
2536 * storm or we have something buggy.
2539 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2542 ts = rb_time_stamp(cpu_buffer->buffer);
2543 diff = ts - cpu_buffer->write_stamp;
2545 /* make sure this diff is calculated here */
2548 /* Did the write stamp get updated already? */
2549 if (likely(ts >= cpu_buffer->write_stamp)) {
2551 if (unlikely(test_time_stamp(delta))) {
2552 int local_clock_stable = 1;
2553 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2554 local_clock_stable = sched_clock_stable();
2556 WARN_ONCE(delta > (1ULL << 59),
2557 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2558 (unsigned long long)delta,
2559 (unsigned long long)ts,
2560 (unsigned long long)cpu_buffer->write_stamp,
2561 local_clock_stable ? "" :
2562 "If you just came from a suspend/resume,\n"
2563 "please switch to the trace global clock:\n"
2564 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2569 event = __rb_reserve_next(cpu_buffer, length, ts,
2570 delta, add_timestamp);
2571 if (unlikely(PTR_ERR(event) == -EAGAIN))
2580 rb_end_commit(cpu_buffer);
2584 #ifdef CONFIG_TRACING
2587 * The lock and unlock are done within a preempt disable section.
2588 * The current_context per_cpu variable can only be modified
2589 * by the current task between lock and unlock. But it can
2590 * be modified more than once via an interrupt. To pass this
2591 * information from the lock to the unlock without having to
2592 * access the 'in_interrupt()' functions again (which do show
2593 * a bit of overhead in something as critical as function tracing,
2594 * we use a bitmask trick.
2596 * bit 0 = NMI context
2597 * bit 1 = IRQ context
2598 * bit 2 = SoftIRQ context
2599 * bit 3 = normal context.
2601 * This works because this is the order of contexts that can
2602 * preempt other contexts. A SoftIRQ never preempts an IRQ
2605 * When the context is determined, the corresponding bit is
2606 * checked and set (if it was set, then a recursion of that context
2609 * On unlock, we need to clear this bit. To do so, just subtract
2610 * 1 from the current_context and AND it to itself.
2614 * 101 & 100 = 100 (clearing bit zero)
2617 * 1010 & 1001 = 1000 (clearing bit 1)
2619 * The least significant bit can be cleared this way, and it
2620 * just so happens that it is the same bit corresponding to
2621 * the current context.
2623 static DEFINE_PER_CPU(unsigned int, current_context);
2625 static __always_inline int trace_recursive_lock(void)
2627 unsigned int val = this_cpu_read(current_context);
2630 if (in_interrupt()) {
2640 if (unlikely(val & (1 << bit)))
2644 this_cpu_write(current_context, val);
2649 static __always_inline void trace_recursive_unlock(void)
2651 unsigned int val = this_cpu_read(current_context);
2654 val &= this_cpu_read(current_context);
2655 this_cpu_write(current_context, val);
2660 #define trace_recursive_lock() (0)
2661 #define trace_recursive_unlock() do { } while (0)
2666 * ring_buffer_lock_reserve - reserve a part of the buffer
2667 * @buffer: the ring buffer to reserve from
2668 * @length: the length of the data to reserve (excluding event header)
2670 * Returns a reseverd event on the ring buffer to copy directly to.
2671 * The user of this interface will need to get the body to write into
2672 * and can use the ring_buffer_event_data() interface.
2674 * The length is the length of the data needed, not the event length
2675 * which also includes the event header.
2677 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2678 * If NULL is returned, then nothing has been allocated or locked.
2680 struct ring_buffer_event *
2681 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2683 struct ring_buffer_per_cpu *cpu_buffer;
2684 struct ring_buffer_event *event;
2687 if (ring_buffer_flags != RB_BUFFERS_ON)
2690 /* If we are tracing schedule, we don't want to recurse */
2691 preempt_disable_notrace();
2693 if (atomic_read(&buffer->record_disabled))
2696 if (trace_recursive_lock())
2699 cpu = raw_smp_processor_id();
2701 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2704 cpu_buffer = buffer->buffers[cpu];
2706 if (atomic_read(&cpu_buffer->record_disabled))
2709 if (length > BUF_MAX_DATA_SIZE)
2712 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2719 trace_recursive_unlock();
2722 preempt_enable_notrace();
2725 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2728 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2729 struct ring_buffer_event *event)
2734 * The event first in the commit queue updates the
2737 if (rb_event_is_commit(cpu_buffer, event)) {
2739 * A commit event that is first on a page
2740 * updates the write timestamp with the page stamp
2742 if (!rb_event_index(event))
2743 cpu_buffer->write_stamp =
2744 cpu_buffer->commit_page->page->time_stamp;
2745 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2746 delta = event->array[0];
2748 delta += event->time_delta;
2749 cpu_buffer->write_stamp += delta;
2751 cpu_buffer->write_stamp += event->time_delta;
2755 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2756 struct ring_buffer_event *event)
2758 local_inc(&cpu_buffer->entries);
2759 rb_update_write_stamp(cpu_buffer, event);
2760 rb_end_commit(cpu_buffer);
2763 static __always_inline void
2764 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2766 if (buffer->irq_work.waiters_pending) {
2767 buffer->irq_work.waiters_pending = false;
2768 /* irq_work_queue() supplies it's own memory barriers */
2769 irq_work_queue(&buffer->irq_work.work);
2772 if (cpu_buffer->irq_work.waiters_pending) {
2773 cpu_buffer->irq_work.waiters_pending = false;
2774 /* irq_work_queue() supplies it's own memory barriers */
2775 irq_work_queue(&cpu_buffer->irq_work.work);
2780 * ring_buffer_unlock_commit - commit a reserved
2781 * @buffer: The buffer to commit to
2782 * @event: The event pointer to commit.
2784 * This commits the data to the ring buffer, and releases any locks held.
2786 * Must be paired with ring_buffer_lock_reserve.
2788 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2789 struct ring_buffer_event *event)
2791 struct ring_buffer_per_cpu *cpu_buffer;
2792 int cpu = raw_smp_processor_id();
2794 cpu_buffer = buffer->buffers[cpu];
2796 rb_commit(cpu_buffer, event);
2798 rb_wakeups(buffer, cpu_buffer);
2800 trace_recursive_unlock();
2802 preempt_enable_notrace();
2806 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2808 static inline void rb_event_discard(struct ring_buffer_event *event)
2810 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2811 event = skip_time_extend(event);
2813 /* array[0] holds the actual length for the discarded event */
2814 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2815 event->type_len = RINGBUF_TYPE_PADDING;
2816 /* time delta must be non zero */
2817 if (!event->time_delta)
2818 event->time_delta = 1;
2822 * Decrement the entries to the page that an event is on.
2823 * The event does not even need to exist, only the pointer
2824 * to the page it is on. This may only be called before the commit
2828 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2829 struct ring_buffer_event *event)
2831 unsigned long addr = (unsigned long)event;
2832 struct buffer_page *bpage = cpu_buffer->commit_page;
2833 struct buffer_page *start;
2837 /* Do the likely case first */
2838 if (likely(bpage->page == (void *)addr)) {
2839 local_dec(&bpage->entries);
2844 * Because the commit page may be on the reader page we
2845 * start with the next page and check the end loop there.
2847 rb_inc_page(cpu_buffer, &bpage);
2850 if (bpage->page == (void *)addr) {
2851 local_dec(&bpage->entries);
2854 rb_inc_page(cpu_buffer, &bpage);
2855 } while (bpage != start);
2857 /* commit not part of this buffer?? */
2858 RB_WARN_ON(cpu_buffer, 1);
2862 * ring_buffer_commit_discard - discard an event that has not been committed
2863 * @buffer: the ring buffer
2864 * @event: non committed event to discard
2866 * Sometimes an event that is in the ring buffer needs to be ignored.
2867 * This function lets the user discard an event in the ring buffer
2868 * and then that event will not be read later.
2870 * This function only works if it is called before the the item has been
2871 * committed. It will try to free the event from the ring buffer
2872 * if another event has not been added behind it.
2874 * If another event has been added behind it, it will set the event
2875 * up as discarded, and perform the commit.
2877 * If this function is called, do not call ring_buffer_unlock_commit on
2880 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2881 struct ring_buffer_event *event)
2883 struct ring_buffer_per_cpu *cpu_buffer;
2886 /* The event is discarded regardless */
2887 rb_event_discard(event);
2889 cpu = smp_processor_id();
2890 cpu_buffer = buffer->buffers[cpu];
2893 * This must only be called if the event has not been
2894 * committed yet. Thus we can assume that preemption
2895 * is still disabled.
2897 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2899 rb_decrement_entry(cpu_buffer, event);
2900 if (rb_try_to_discard(cpu_buffer, event))
2904 * The commit is still visible by the reader, so we
2905 * must still update the timestamp.
2907 rb_update_write_stamp(cpu_buffer, event);
2909 rb_end_commit(cpu_buffer);
2911 trace_recursive_unlock();
2913 preempt_enable_notrace();
2916 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2919 * ring_buffer_write - write data to the buffer without reserving
2920 * @buffer: The ring buffer to write to.
2921 * @length: The length of the data being written (excluding the event header)
2922 * @data: The data to write to the buffer.
2924 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2925 * one function. If you already have the data to write to the buffer, it
2926 * may be easier to simply call this function.
2928 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2929 * and not the length of the event which would hold the header.
2931 int ring_buffer_write(struct ring_buffer *buffer,
2932 unsigned long length,
2935 struct ring_buffer_per_cpu *cpu_buffer;
2936 struct ring_buffer_event *event;
2941 if (ring_buffer_flags != RB_BUFFERS_ON)
2944 preempt_disable_notrace();
2946 if (atomic_read(&buffer->record_disabled))
2949 cpu = raw_smp_processor_id();
2951 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2954 cpu_buffer = buffer->buffers[cpu];
2956 if (atomic_read(&cpu_buffer->record_disabled))
2959 if (length > BUF_MAX_DATA_SIZE)
2962 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2966 body = rb_event_data(event);
2968 memcpy(body, data, length);
2970 rb_commit(cpu_buffer, event);
2972 rb_wakeups(buffer, cpu_buffer);
2976 preempt_enable_notrace();
2980 EXPORT_SYMBOL_GPL(ring_buffer_write);
2982 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
2984 struct buffer_page *reader = cpu_buffer->reader_page;
2985 struct buffer_page *head = rb_set_head_page(cpu_buffer);
2986 struct buffer_page *commit = cpu_buffer->commit_page;
2988 /* In case of error, head will be NULL */
2989 if (unlikely(!head))
2992 return reader->read == rb_page_commit(reader) &&
2993 (commit == reader ||
2995 head->read == rb_page_commit(commit)));
2999 * ring_buffer_record_disable - stop all writes into the buffer
3000 * @buffer: The ring buffer to stop writes to.
3002 * This prevents all writes to the buffer. Any attempt to write
3003 * to the buffer after this will fail and return NULL.
3005 * The caller should call synchronize_sched() after this.
3007 void ring_buffer_record_disable(struct ring_buffer *buffer)
3009 atomic_inc(&buffer->record_disabled);
3011 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3014 * ring_buffer_record_enable - enable writes to the buffer
3015 * @buffer: The ring buffer to enable writes
3017 * Note, multiple disables will need the same number of enables
3018 * to truly enable the writing (much like preempt_disable).
3020 void ring_buffer_record_enable(struct ring_buffer *buffer)
3022 atomic_dec(&buffer->record_disabled);
3024 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3027 * ring_buffer_record_off - stop all writes into the buffer
3028 * @buffer: The ring buffer to stop writes to.
3030 * This prevents all writes to the buffer. Any attempt to write
3031 * to the buffer after this will fail and return NULL.
3033 * This is different than ring_buffer_record_disable() as
3034 * it works like an on/off switch, where as the disable() version
3035 * must be paired with a enable().
3037 void ring_buffer_record_off(struct ring_buffer *buffer)
3040 unsigned int new_rd;
3043 rd = atomic_read(&buffer->record_disabled);
3044 new_rd = rd | RB_BUFFER_OFF;
3045 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3047 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3050 * ring_buffer_record_on - restart writes into the buffer
3051 * @buffer: The ring buffer to start writes to.
3053 * This enables all writes to the buffer that was disabled by
3054 * ring_buffer_record_off().
3056 * This is different than ring_buffer_record_enable() as
3057 * it works like an on/off switch, where as the enable() version
3058 * must be paired with a disable().
3060 void ring_buffer_record_on(struct ring_buffer *buffer)
3063 unsigned int new_rd;
3066 rd = atomic_read(&buffer->record_disabled);
3067 new_rd = rd & ~RB_BUFFER_OFF;
3068 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3070 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3073 * ring_buffer_record_is_on - return true if the ring buffer can write
3074 * @buffer: The ring buffer to see if write is enabled
3076 * Returns true if the ring buffer is in a state that it accepts writes.
3078 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3080 return !atomic_read(&buffer->record_disabled);
3084 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3085 * @buffer: The ring buffer to stop writes to.
3086 * @cpu: The CPU buffer to stop
3088 * This prevents all writes to the buffer. Any attempt to write
3089 * to the buffer after this will fail and return NULL.
3091 * The caller should call synchronize_sched() after this.
3093 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3095 struct ring_buffer_per_cpu *cpu_buffer;
3097 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3100 cpu_buffer = buffer->buffers[cpu];
3101 atomic_inc(&cpu_buffer->record_disabled);
3103 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3106 * ring_buffer_record_enable_cpu - enable writes to the buffer
3107 * @buffer: The ring buffer to enable writes
3108 * @cpu: The CPU to enable.
3110 * Note, multiple disables will need the same number of enables
3111 * to truly enable the writing (much like preempt_disable).
3113 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3115 struct ring_buffer_per_cpu *cpu_buffer;
3117 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3120 cpu_buffer = buffer->buffers[cpu];
3121 atomic_dec(&cpu_buffer->record_disabled);
3123 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3126 * The total entries in the ring buffer is the running counter
3127 * of entries entered into the ring buffer, minus the sum of
3128 * the entries read from the ring buffer and the number of
3129 * entries that were overwritten.
3131 static inline unsigned long
3132 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3134 return local_read(&cpu_buffer->entries) -
3135 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3139 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3140 * @buffer: The ring buffer
3141 * @cpu: The per CPU buffer to read from.
3143 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3145 unsigned long flags;
3146 struct ring_buffer_per_cpu *cpu_buffer;
3147 struct buffer_page *bpage;
3150 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3153 cpu_buffer = buffer->buffers[cpu];
3154 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3156 * if the tail is on reader_page, oldest time stamp is on the reader
3159 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3160 bpage = cpu_buffer->reader_page;
3162 bpage = rb_set_head_page(cpu_buffer);
3164 ret = bpage->page->time_stamp;
3165 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3169 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3172 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3173 * @buffer: The ring buffer
3174 * @cpu: The per CPU buffer to read from.
3176 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3178 struct ring_buffer_per_cpu *cpu_buffer;
3181 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3184 cpu_buffer = buffer->buffers[cpu];
3185 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3189 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3192 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3193 * @buffer: The ring buffer
3194 * @cpu: The per CPU buffer to get the entries from.
3196 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3198 struct ring_buffer_per_cpu *cpu_buffer;
3200 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3203 cpu_buffer = buffer->buffers[cpu];
3205 return rb_num_of_entries(cpu_buffer);
3207 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3210 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3211 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3212 * @buffer: The ring buffer
3213 * @cpu: The per CPU buffer to get the number of overruns from
3215 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3217 struct ring_buffer_per_cpu *cpu_buffer;
3220 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3223 cpu_buffer = buffer->buffers[cpu];
3224 ret = local_read(&cpu_buffer->overrun);
3228 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3231 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3232 * commits failing due to the buffer wrapping around while there are uncommitted
3233 * events, such as during an interrupt storm.
3234 * @buffer: The ring buffer
3235 * @cpu: The per CPU buffer to get the number of overruns from
3238 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3240 struct ring_buffer_per_cpu *cpu_buffer;
3243 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3246 cpu_buffer = buffer->buffers[cpu];
3247 ret = local_read(&cpu_buffer->commit_overrun);
3251 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3254 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3255 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3256 * @buffer: The ring buffer
3257 * @cpu: The per CPU buffer to get the number of overruns from
3260 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3262 struct ring_buffer_per_cpu *cpu_buffer;
3265 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3268 cpu_buffer = buffer->buffers[cpu];
3269 ret = local_read(&cpu_buffer->dropped_events);
3273 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3276 * ring_buffer_read_events_cpu - get the number of events successfully read
3277 * @buffer: The ring buffer
3278 * @cpu: The per CPU buffer to get the number of events read
3281 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3283 struct ring_buffer_per_cpu *cpu_buffer;
3285 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3288 cpu_buffer = buffer->buffers[cpu];
3289 return cpu_buffer->read;
3291 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3294 * ring_buffer_entries - get the number of entries in a buffer
3295 * @buffer: The ring buffer
3297 * Returns the total number of entries in the ring buffer
3300 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3302 struct ring_buffer_per_cpu *cpu_buffer;
3303 unsigned long entries = 0;
3306 /* if you care about this being correct, lock the buffer */
3307 for_each_buffer_cpu(buffer, cpu) {
3308 cpu_buffer = buffer->buffers[cpu];
3309 entries += rb_num_of_entries(cpu_buffer);
3314 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3317 * ring_buffer_overruns - get the number of overruns in buffer
3318 * @buffer: The ring buffer
3320 * Returns the total number of overruns in the ring buffer
3323 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3325 struct ring_buffer_per_cpu *cpu_buffer;
3326 unsigned long overruns = 0;
3329 /* if you care about this being correct, lock the buffer */
3330 for_each_buffer_cpu(buffer, cpu) {
3331 cpu_buffer = buffer->buffers[cpu];
3332 overruns += local_read(&cpu_buffer->overrun);
3337 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3339 static void rb_iter_reset(struct ring_buffer_iter *iter)
3341 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3343 /* Iterator usage is expected to have record disabled */
3344 iter->head_page = cpu_buffer->reader_page;
3345 iter->head = cpu_buffer->reader_page->read;
3347 iter->cache_reader_page = iter->head_page;
3348 iter->cache_read = iter->head;
3351 iter->read_stamp = cpu_buffer->read_stamp;
3353 iter->read_stamp = iter->head_page->page->time_stamp;
3357 * ring_buffer_iter_reset - reset an iterator
3358 * @iter: The iterator to reset
3360 * Resets the iterator, so that it will start from the beginning
3363 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3365 struct ring_buffer_per_cpu *cpu_buffer;
3366 unsigned long flags;
3371 cpu_buffer = iter->cpu_buffer;
3373 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3374 rb_iter_reset(iter);
3375 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3377 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3380 * ring_buffer_iter_empty - check if an iterator has no more to read
3381 * @iter: The iterator to check
3383 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3385 struct ring_buffer_per_cpu *cpu_buffer;
3387 cpu_buffer = iter->cpu_buffer;
3389 return iter->head_page == cpu_buffer->commit_page &&
3390 iter->head == rb_commit_index(cpu_buffer);
3392 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3395 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3396 struct ring_buffer_event *event)
3400 switch (event->type_len) {
3401 case RINGBUF_TYPE_PADDING:
3404 case RINGBUF_TYPE_TIME_EXTEND:
3405 delta = event->array[0];
3407 delta += event->time_delta;
3408 cpu_buffer->read_stamp += delta;
3411 case RINGBUF_TYPE_TIME_STAMP:
3412 /* FIXME: not implemented */
3415 case RINGBUF_TYPE_DATA:
3416 cpu_buffer->read_stamp += event->time_delta;
3426 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3427 struct ring_buffer_event *event)
3431 switch (event->type_len) {
3432 case RINGBUF_TYPE_PADDING:
3435 case RINGBUF_TYPE_TIME_EXTEND:
3436 delta = event->array[0];
3438 delta += event->time_delta;
3439 iter->read_stamp += delta;
3442 case RINGBUF_TYPE_TIME_STAMP:
3443 /* FIXME: not implemented */
3446 case RINGBUF_TYPE_DATA:
3447 iter->read_stamp += event->time_delta;
3456 static struct buffer_page *
3457 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3459 struct buffer_page *reader = NULL;
3460 unsigned long overwrite;
3461 unsigned long flags;
3465 local_irq_save(flags);
3466 arch_spin_lock(&cpu_buffer->lock);
3470 * This should normally only loop twice. But because the
3471 * start of the reader inserts an empty page, it causes
3472 * a case where we will loop three times. There should be no
3473 * reason to loop four times (that I know of).
3475 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3480 reader = cpu_buffer->reader_page;
3482 /* If there's more to read, return this page */
3483 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3486 /* Never should we have an index greater than the size */
3487 if (RB_WARN_ON(cpu_buffer,
3488 cpu_buffer->reader_page->read > rb_page_size(reader)))
3491 /* check if we caught up to the tail */
3493 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3496 /* Don't bother swapping if the ring buffer is empty */
3497 if (rb_num_of_entries(cpu_buffer) == 0)
3501 * Reset the reader page to size zero.
3503 local_set(&cpu_buffer->reader_page->write, 0);
3504 local_set(&cpu_buffer->reader_page->entries, 0);
3505 local_set(&cpu_buffer->reader_page->page->commit, 0);
3506 cpu_buffer->reader_page->real_end = 0;
3510 * Splice the empty reader page into the list around the head.
3512 reader = rb_set_head_page(cpu_buffer);
3515 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3516 cpu_buffer->reader_page->list.prev = reader->list.prev;
3519 * cpu_buffer->pages just needs to point to the buffer, it
3520 * has no specific buffer page to point to. Lets move it out
3521 * of our way so we don't accidentally swap it.
3523 cpu_buffer->pages = reader->list.prev;
3525 /* The reader page will be pointing to the new head */
3526 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3529 * We want to make sure we read the overruns after we set up our
3530 * pointers to the next object. The writer side does a
3531 * cmpxchg to cross pages which acts as the mb on the writer
3532 * side. Note, the reader will constantly fail the swap
3533 * while the writer is updating the pointers, so this
3534 * guarantees that the overwrite recorded here is the one we
3535 * want to compare with the last_overrun.
3538 overwrite = local_read(&(cpu_buffer->overrun));
3541 * Here's the tricky part.
3543 * We need to move the pointer past the header page.
3544 * But we can only do that if a writer is not currently
3545 * moving it. The page before the header page has the
3546 * flag bit '1' set if it is pointing to the page we want.
3547 * but if the writer is in the process of moving it
3548 * than it will be '2' or already moved '0'.
3551 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3554 * If we did not convert it, then we must try again.
3560 * Yeah! We succeeded in replacing the page.
3562 * Now make the new head point back to the reader page.
3564 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3565 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3567 /* Finally update the reader page to the new head */
3568 cpu_buffer->reader_page = reader;
3569 rb_reset_reader_page(cpu_buffer);
3571 if (overwrite != cpu_buffer->last_overrun) {
3572 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3573 cpu_buffer->last_overrun = overwrite;
3579 arch_spin_unlock(&cpu_buffer->lock);
3580 local_irq_restore(flags);
3585 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3587 struct ring_buffer_event *event;
3588 struct buffer_page *reader;
3591 reader = rb_get_reader_page(cpu_buffer);
3593 /* This function should not be called when buffer is empty */
3594 if (RB_WARN_ON(cpu_buffer, !reader))
3597 event = rb_reader_event(cpu_buffer);
3599 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3602 rb_update_read_stamp(cpu_buffer, event);
3604 length = rb_event_length(event);
3605 cpu_buffer->reader_page->read += length;
3608 static void rb_advance_iter(struct ring_buffer_iter *iter)
3610 struct ring_buffer_per_cpu *cpu_buffer;
3611 struct ring_buffer_event *event;
3614 cpu_buffer = iter->cpu_buffer;
3617 * Check if we are at the end of the buffer.
3619 if (iter->head >= rb_page_size(iter->head_page)) {
3620 /* discarded commits can make the page empty */
3621 if (iter->head_page == cpu_buffer->commit_page)
3627 event = rb_iter_head_event(iter);
3629 length = rb_event_length(event);
3632 * This should not be called to advance the header if we are
3633 * at the tail of the buffer.
3635 if (RB_WARN_ON(cpu_buffer,
3636 (iter->head_page == cpu_buffer->commit_page) &&
3637 (iter->head + length > rb_commit_index(cpu_buffer))))
3640 rb_update_iter_read_stamp(iter, event);
3642 iter->head += length;
3644 /* check for end of page padding */
3645 if ((iter->head >= rb_page_size(iter->head_page)) &&
3646 (iter->head_page != cpu_buffer->commit_page))
3650 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3652 return cpu_buffer->lost_events;
3655 static struct ring_buffer_event *
3656 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3657 unsigned long *lost_events)
3659 struct ring_buffer_event *event;
3660 struct buffer_page *reader;
3665 * We repeat when a time extend is encountered.
3666 * Since the time extend is always attached to a data event,
3667 * we should never loop more than once.
3668 * (We never hit the following condition more than twice).
3670 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3673 reader = rb_get_reader_page(cpu_buffer);
3677 event = rb_reader_event(cpu_buffer);
3679 switch (event->type_len) {
3680 case RINGBUF_TYPE_PADDING:
3681 if (rb_null_event(event))
3682 RB_WARN_ON(cpu_buffer, 1);
3684 * Because the writer could be discarding every
3685 * event it creates (which would probably be bad)
3686 * if we were to go back to "again" then we may never
3687 * catch up, and will trigger the warn on, or lock
3688 * the box. Return the padding, and we will release
3689 * the current locks, and try again.
3693 case RINGBUF_TYPE_TIME_EXTEND:
3694 /* Internal data, OK to advance */
3695 rb_advance_reader(cpu_buffer);
3698 case RINGBUF_TYPE_TIME_STAMP:
3699 /* FIXME: not implemented */
3700 rb_advance_reader(cpu_buffer);
3703 case RINGBUF_TYPE_DATA:
3705 *ts = cpu_buffer->read_stamp + event->time_delta;
3706 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3707 cpu_buffer->cpu, ts);
3710 *lost_events = rb_lost_events(cpu_buffer);
3719 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3721 static struct ring_buffer_event *
3722 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3724 struct ring_buffer *buffer;
3725 struct ring_buffer_per_cpu *cpu_buffer;
3726 struct ring_buffer_event *event;
3729 cpu_buffer = iter->cpu_buffer;
3730 buffer = cpu_buffer->buffer;
3733 * Check if someone performed a consuming read to
3734 * the buffer. A consuming read invalidates the iterator
3735 * and we need to reset the iterator in this case.
3737 if (unlikely(iter->cache_read != cpu_buffer->read ||
3738 iter->cache_reader_page != cpu_buffer->reader_page))
3739 rb_iter_reset(iter);
3742 if (ring_buffer_iter_empty(iter))
3746 * We repeat when a time extend is encountered or we hit
3747 * the end of the page. Since the time extend is always attached
3748 * to a data event, we should never loop more than three times.
3749 * Once for going to next page, once on time extend, and
3750 * finally once to get the event.
3751 * (We never hit the following condition more than thrice).
3753 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3756 if (rb_per_cpu_empty(cpu_buffer))
3759 if (iter->head >= rb_page_size(iter->head_page)) {
3764 event = rb_iter_head_event(iter);
3766 switch (event->type_len) {
3767 case RINGBUF_TYPE_PADDING:
3768 if (rb_null_event(event)) {
3772 rb_advance_iter(iter);
3775 case RINGBUF_TYPE_TIME_EXTEND:
3776 /* Internal data, OK to advance */
3777 rb_advance_iter(iter);
3780 case RINGBUF_TYPE_TIME_STAMP:
3781 /* FIXME: not implemented */
3782 rb_advance_iter(iter);
3785 case RINGBUF_TYPE_DATA:
3787 *ts = iter->read_stamp + event->time_delta;
3788 ring_buffer_normalize_time_stamp(buffer,
3789 cpu_buffer->cpu, ts);
3799 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3801 static inline int rb_ok_to_lock(void)
3804 * If an NMI die dumps out the content of the ring buffer
3805 * do not grab locks. We also permanently disable the ring
3806 * buffer too. A one time deal is all you get from reading
3807 * the ring buffer from an NMI.
3809 if (likely(!in_nmi()))
3812 tracing_off_permanent();
3817 * ring_buffer_peek - peek at the next event to be read
3818 * @buffer: The ring buffer to read
3819 * @cpu: The cpu to peak at
3820 * @ts: The timestamp counter of this event.
3821 * @lost_events: a variable to store if events were lost (may be NULL)
3823 * This will return the event that will be read next, but does
3824 * not consume the data.
3826 struct ring_buffer_event *
3827 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3828 unsigned long *lost_events)
3830 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3831 struct ring_buffer_event *event;
3832 unsigned long flags;
3835 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3838 dolock = rb_ok_to_lock();
3840 local_irq_save(flags);
3842 raw_spin_lock(&cpu_buffer->reader_lock);
3843 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3844 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3845 rb_advance_reader(cpu_buffer);
3847 raw_spin_unlock(&cpu_buffer->reader_lock);
3848 local_irq_restore(flags);
3850 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3857 * ring_buffer_iter_peek - peek at the next event to be read
3858 * @iter: The ring buffer iterator
3859 * @ts: The timestamp counter of this event.
3861 * This will return the event that will be read next, but does
3862 * not increment the iterator.
3864 struct ring_buffer_event *
3865 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3867 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3868 struct ring_buffer_event *event;
3869 unsigned long flags;
3872 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3873 event = rb_iter_peek(iter, ts);
3874 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3876 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3883 * ring_buffer_consume - return an event and consume it
3884 * @buffer: The ring buffer to get the next event from
3885 * @cpu: the cpu to read the buffer from
3886 * @ts: a variable to store the timestamp (may be NULL)
3887 * @lost_events: a variable to store if events were lost (may be NULL)
3889 * Returns the next event in the ring buffer, and that event is consumed.
3890 * Meaning, that sequential reads will keep returning a different event,
3891 * and eventually empty the ring buffer if the producer is slower.
3893 struct ring_buffer_event *
3894 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3895 unsigned long *lost_events)
3897 struct ring_buffer_per_cpu *cpu_buffer;
3898 struct ring_buffer_event *event = NULL;
3899 unsigned long flags;
3902 dolock = rb_ok_to_lock();
3905 /* might be called in atomic */
3908 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3911 cpu_buffer = buffer->buffers[cpu];
3912 local_irq_save(flags);
3914 raw_spin_lock(&cpu_buffer->reader_lock);
3916 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3918 cpu_buffer->lost_events = 0;
3919 rb_advance_reader(cpu_buffer);
3923 raw_spin_unlock(&cpu_buffer->reader_lock);
3924 local_irq_restore(flags);
3929 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3934 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3937 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3938 * @buffer: The ring buffer to read from
3939 * @cpu: The cpu buffer to iterate over
3941 * This performs the initial preparations necessary to iterate
3942 * through the buffer. Memory is allocated, buffer recording
3943 * is disabled, and the iterator pointer is returned to the caller.
3945 * Disabling buffer recordng prevents the reading from being
3946 * corrupted. This is not a consuming read, so a producer is not
3949 * After a sequence of ring_buffer_read_prepare calls, the user is
3950 * expected to make at least one call to ring_buffer_read_prepare_sync.
3951 * Afterwards, ring_buffer_read_start is invoked to get things going
3954 * This overall must be paired with ring_buffer_read_finish.
3956 struct ring_buffer_iter *
3957 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3959 struct ring_buffer_per_cpu *cpu_buffer;
3960 struct ring_buffer_iter *iter;
3962 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3965 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3969 cpu_buffer = buffer->buffers[cpu];
3971 iter->cpu_buffer = cpu_buffer;
3973 atomic_inc(&buffer->resize_disabled);
3974 atomic_inc(&cpu_buffer->record_disabled);
3978 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
3981 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
3983 * All previously invoked ring_buffer_read_prepare calls to prepare
3984 * iterators will be synchronized. Afterwards, read_buffer_read_start
3985 * calls on those iterators are allowed.
3988 ring_buffer_read_prepare_sync(void)
3990 synchronize_sched();
3992 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
3995 * ring_buffer_read_start - start a non consuming read of the buffer
3996 * @iter: The iterator returned by ring_buffer_read_prepare
3998 * This finalizes the startup of an iteration through the buffer.
3999 * The iterator comes from a call to ring_buffer_read_prepare and
4000 * an intervening ring_buffer_read_prepare_sync must have been
4003 * Must be paired with ring_buffer_read_finish.
4006 ring_buffer_read_start(struct ring_buffer_iter *iter)
4008 struct ring_buffer_per_cpu *cpu_buffer;
4009 unsigned long flags;
4014 cpu_buffer = iter->cpu_buffer;
4016 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4017 arch_spin_lock(&cpu_buffer->lock);
4018 rb_iter_reset(iter);
4019 arch_spin_unlock(&cpu_buffer->lock);
4020 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4022 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4025 * ring_buffer_read_finish - finish reading the iterator of the buffer
4026 * @iter: The iterator retrieved by ring_buffer_start
4028 * This re-enables the recording to the buffer, and frees the
4032 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4034 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4035 unsigned long flags;
4038 * Ring buffer is disabled from recording, here's a good place
4039 * to check the integrity of the ring buffer.
4040 * Must prevent readers from trying to read, as the check
4041 * clears the HEAD page and readers require it.
4043 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4044 rb_check_pages(cpu_buffer);
4045 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4047 atomic_dec(&cpu_buffer->record_disabled);
4048 atomic_dec(&cpu_buffer->buffer->resize_disabled);
4051 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4054 * ring_buffer_read - read the next item in the ring buffer by the iterator
4055 * @iter: The ring buffer iterator
4056 * @ts: The time stamp of the event read.
4058 * This reads the next event in the ring buffer and increments the iterator.
4060 struct ring_buffer_event *
4061 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4063 struct ring_buffer_event *event;
4064 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4065 unsigned long flags;
4067 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4069 event = rb_iter_peek(iter, ts);
4073 if (event->type_len == RINGBUF_TYPE_PADDING)
4076 rb_advance_iter(iter);
4078 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4082 EXPORT_SYMBOL_GPL(ring_buffer_read);
4085 * ring_buffer_size - return the size of the ring buffer (in bytes)
4086 * @buffer: The ring buffer.
4088 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4091 * Earlier, this method returned
4092 * BUF_PAGE_SIZE * buffer->nr_pages
4093 * Since the nr_pages field is now removed, we have converted this to
4094 * return the per cpu buffer value.
4096 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4099 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4101 EXPORT_SYMBOL_GPL(ring_buffer_size);
4104 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4106 rb_head_page_deactivate(cpu_buffer);
4108 cpu_buffer->head_page
4109 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4110 local_set(&cpu_buffer->head_page->write, 0);
4111 local_set(&cpu_buffer->head_page->entries, 0);
4112 local_set(&cpu_buffer->head_page->page->commit, 0);
4114 cpu_buffer->head_page->read = 0;
4116 cpu_buffer->tail_page = cpu_buffer->head_page;
4117 cpu_buffer->commit_page = cpu_buffer->head_page;
4119 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4120 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4121 local_set(&cpu_buffer->reader_page->write, 0);
4122 local_set(&cpu_buffer->reader_page->entries, 0);
4123 local_set(&cpu_buffer->reader_page->page->commit, 0);
4124 cpu_buffer->reader_page->read = 0;
4126 local_set(&cpu_buffer->entries_bytes, 0);
4127 local_set(&cpu_buffer->overrun, 0);
4128 local_set(&cpu_buffer->commit_overrun, 0);
4129 local_set(&cpu_buffer->dropped_events, 0);
4130 local_set(&cpu_buffer->entries, 0);
4131 local_set(&cpu_buffer->committing, 0);
4132 local_set(&cpu_buffer->commits, 0);
4133 cpu_buffer->read = 0;
4134 cpu_buffer->read_bytes = 0;
4136 cpu_buffer->write_stamp = 0;
4137 cpu_buffer->read_stamp = 0;
4139 cpu_buffer->lost_events = 0;
4140 cpu_buffer->last_overrun = 0;
4142 rb_head_page_activate(cpu_buffer);
4146 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4147 * @buffer: The ring buffer to reset a per cpu buffer of
4148 * @cpu: The CPU buffer to be reset
4150 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4152 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4153 unsigned long flags;
4155 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4158 atomic_inc(&buffer->resize_disabled);
4159 atomic_inc(&cpu_buffer->record_disabled);
4161 /* Make sure all commits have finished */
4162 synchronize_sched();
4164 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4166 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4169 arch_spin_lock(&cpu_buffer->lock);
4171 rb_reset_cpu(cpu_buffer);
4173 arch_spin_unlock(&cpu_buffer->lock);
4176 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4178 atomic_dec(&cpu_buffer->record_disabled);
4179 atomic_dec(&buffer->resize_disabled);
4181 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4184 * ring_buffer_reset - reset a ring buffer
4185 * @buffer: The ring buffer to reset all cpu buffers
4187 void ring_buffer_reset(struct ring_buffer *buffer)
4191 for_each_buffer_cpu(buffer, cpu)
4192 ring_buffer_reset_cpu(buffer, cpu);
4194 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4197 * rind_buffer_empty - is the ring buffer empty?
4198 * @buffer: The ring buffer to test
4200 int ring_buffer_empty(struct ring_buffer *buffer)
4202 struct ring_buffer_per_cpu *cpu_buffer;
4203 unsigned long flags;
4208 dolock = rb_ok_to_lock();
4210 /* yes this is racy, but if you don't like the race, lock the buffer */
4211 for_each_buffer_cpu(buffer, cpu) {
4212 cpu_buffer = buffer->buffers[cpu];
4213 local_irq_save(flags);
4215 raw_spin_lock(&cpu_buffer->reader_lock);
4216 ret = rb_per_cpu_empty(cpu_buffer);
4218 raw_spin_unlock(&cpu_buffer->reader_lock);
4219 local_irq_restore(flags);
4227 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4230 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4231 * @buffer: The ring buffer
4232 * @cpu: The CPU buffer to test
4234 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4236 struct ring_buffer_per_cpu *cpu_buffer;
4237 unsigned long flags;
4241 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4244 dolock = rb_ok_to_lock();
4246 cpu_buffer = buffer->buffers[cpu];
4247 local_irq_save(flags);
4249 raw_spin_lock(&cpu_buffer->reader_lock);
4250 ret = rb_per_cpu_empty(cpu_buffer);
4252 raw_spin_unlock(&cpu_buffer->reader_lock);
4253 local_irq_restore(flags);
4257 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4259 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4261 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4262 * @buffer_a: One buffer to swap with
4263 * @buffer_b: The other buffer to swap with
4265 * This function is useful for tracers that want to take a "snapshot"
4266 * of a CPU buffer and has another back up buffer lying around.
4267 * it is expected that the tracer handles the cpu buffer not being
4268 * used at the moment.
4270 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4271 struct ring_buffer *buffer_b, int cpu)
4273 struct ring_buffer_per_cpu *cpu_buffer_a;
4274 struct ring_buffer_per_cpu *cpu_buffer_b;
4277 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4278 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4281 cpu_buffer_a = buffer_a->buffers[cpu];
4282 cpu_buffer_b = buffer_b->buffers[cpu];
4284 /* At least make sure the two buffers are somewhat the same */
4285 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4290 if (ring_buffer_flags != RB_BUFFERS_ON)
4293 if (atomic_read(&buffer_a->record_disabled))
4296 if (atomic_read(&buffer_b->record_disabled))
4299 if (atomic_read(&cpu_buffer_a->record_disabled))
4302 if (atomic_read(&cpu_buffer_b->record_disabled))
4306 * We can't do a synchronize_sched here because this
4307 * function can be called in atomic context.
4308 * Normally this will be called from the same CPU as cpu.
4309 * If not it's up to the caller to protect this.
4311 atomic_inc(&cpu_buffer_a->record_disabled);
4312 atomic_inc(&cpu_buffer_b->record_disabled);
4315 if (local_read(&cpu_buffer_a->committing))
4317 if (local_read(&cpu_buffer_b->committing))
4320 buffer_a->buffers[cpu] = cpu_buffer_b;
4321 buffer_b->buffers[cpu] = cpu_buffer_a;
4323 cpu_buffer_b->buffer = buffer_a;
4324 cpu_buffer_a->buffer = buffer_b;
4329 atomic_dec(&cpu_buffer_a->record_disabled);
4330 atomic_dec(&cpu_buffer_b->record_disabled);
4334 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4335 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4338 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4339 * @buffer: the buffer to allocate for.
4340 * @cpu: the cpu buffer to allocate.
4342 * This function is used in conjunction with ring_buffer_read_page.
4343 * When reading a full page from the ring buffer, these functions
4344 * can be used to speed up the process. The calling function should
4345 * allocate a few pages first with this function. Then when it
4346 * needs to get pages from the ring buffer, it passes the result
4347 * of this function into ring_buffer_read_page, which will swap
4348 * the page that was allocated, with the read page of the buffer.
4351 * The page allocated, or NULL on error.
4353 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4355 struct buffer_data_page *bpage;
4358 page = alloc_pages_node(cpu_to_node(cpu),
4359 GFP_KERNEL | __GFP_NORETRY, 0);
4363 bpage = page_address(page);
4365 rb_init_page(bpage);
4369 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4372 * ring_buffer_free_read_page - free an allocated read page
4373 * @buffer: the buffer the page was allocate for
4374 * @data: the page to free
4376 * Free a page allocated from ring_buffer_alloc_read_page.
4378 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4380 free_page((unsigned long)data);
4382 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4385 * ring_buffer_read_page - extract a page from the ring buffer
4386 * @buffer: buffer to extract from
4387 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4388 * @len: amount to extract
4389 * @cpu: the cpu of the buffer to extract
4390 * @full: should the extraction only happen when the page is full.
4392 * This function will pull out a page from the ring buffer and consume it.
4393 * @data_page must be the address of the variable that was returned
4394 * from ring_buffer_alloc_read_page. This is because the page might be used
4395 * to swap with a page in the ring buffer.
4398 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4401 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4403 * process_page(rpage, ret);
4405 * When @full is set, the function will not return true unless
4406 * the writer is off the reader page.
4408 * Note: it is up to the calling functions to handle sleeps and wakeups.
4409 * The ring buffer can be used anywhere in the kernel and can not
4410 * blindly call wake_up. The layer that uses the ring buffer must be
4411 * responsible for that.
4414 * >=0 if data has been transferred, returns the offset of consumed data.
4415 * <0 if no data has been transferred.
4417 int ring_buffer_read_page(struct ring_buffer *buffer,
4418 void **data_page, size_t len, int cpu, int full)
4420 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4421 struct ring_buffer_event *event;
4422 struct buffer_data_page *bpage;
4423 struct buffer_page *reader;
4424 unsigned long missed_events;
4425 unsigned long flags;
4426 unsigned int commit;
4431 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4435 * If len is not big enough to hold the page header, then
4436 * we can not copy anything.
4438 if (len <= BUF_PAGE_HDR_SIZE)
4441 len -= BUF_PAGE_HDR_SIZE;
4450 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4452 reader = rb_get_reader_page(cpu_buffer);
4456 event = rb_reader_event(cpu_buffer);
4458 read = reader->read;
4459 commit = rb_page_commit(reader);
4461 /* Check if any events were dropped */
4462 missed_events = cpu_buffer->lost_events;
4465 * If this page has been partially read or
4466 * if len is not big enough to read the rest of the page or
4467 * a writer is still on the page, then
4468 * we must copy the data from the page to the buffer.
4469 * Otherwise, we can simply swap the page with the one passed in.
4471 if (read || (len < (commit - read)) ||
4472 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4473 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4474 unsigned int rpos = read;
4475 unsigned int pos = 0;
4481 if (len > (commit - read))
4482 len = (commit - read);
4484 /* Always keep the time extend and data together */
4485 size = rb_event_ts_length(event);
4490 /* save the current timestamp, since the user will need it */
4491 save_timestamp = cpu_buffer->read_stamp;
4493 /* Need to copy one event at a time */
4495 /* We need the size of one event, because
4496 * rb_advance_reader only advances by one event,
4497 * whereas rb_event_ts_length may include the size of
4498 * one or two events.
4499 * We have already ensured there's enough space if this
4500 * is a time extend. */
4501 size = rb_event_length(event);
4502 memcpy(bpage->data + pos, rpage->data + rpos, size);
4506 rb_advance_reader(cpu_buffer);
4507 rpos = reader->read;
4513 event = rb_reader_event(cpu_buffer);
4514 /* Always keep the time extend and data together */
4515 size = rb_event_ts_length(event);
4516 } while (len >= size);
4519 local_set(&bpage->commit, pos);
4520 bpage->time_stamp = save_timestamp;
4522 /* we copied everything to the beginning */
4525 /* update the entry counter */
4526 cpu_buffer->read += rb_page_entries(reader);
4527 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4529 /* swap the pages */
4530 rb_init_page(bpage);
4531 bpage = reader->page;
4532 reader->page = *data_page;
4533 local_set(&reader->write, 0);
4534 local_set(&reader->entries, 0);
4539 * Use the real_end for the data size,
4540 * This gives us a chance to store the lost events
4543 if (reader->real_end)
4544 local_set(&bpage->commit, reader->real_end);
4548 cpu_buffer->lost_events = 0;
4550 commit = local_read(&bpage->commit);
4552 * Set a flag in the commit field if we lost events
4554 if (missed_events) {
4555 /* If there is room at the end of the page to save the
4556 * missed events, then record it there.
4558 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4559 memcpy(&bpage->data[commit], &missed_events,
4560 sizeof(missed_events));
4561 local_add(RB_MISSED_STORED, &bpage->commit);
4562 commit += sizeof(missed_events);
4564 local_add(RB_MISSED_EVENTS, &bpage->commit);
4568 * This page may be off to user land. Zero it out here.
4570 if (commit < BUF_PAGE_SIZE)
4571 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4574 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4579 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4581 #ifdef CONFIG_HOTPLUG_CPU
4582 static int rb_cpu_notify(struct notifier_block *self,
4583 unsigned long action, void *hcpu)
4585 struct ring_buffer *buffer =
4586 container_of(self, struct ring_buffer, cpu_notify);
4587 long cpu = (long)hcpu;
4588 int cpu_i, nr_pages_same;
4589 unsigned int nr_pages;
4592 case CPU_UP_PREPARE:
4593 case CPU_UP_PREPARE_FROZEN:
4594 if (cpumask_test_cpu(cpu, buffer->cpumask))
4599 /* check if all cpu sizes are same */
4600 for_each_buffer_cpu(buffer, cpu_i) {
4601 /* fill in the size from first enabled cpu */
4603 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4604 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4609 /* allocate minimum pages, user can later expand it */
4612 buffer->buffers[cpu] =
4613 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4614 if (!buffer->buffers[cpu]) {
4615 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4620 cpumask_set_cpu(cpu, buffer->cpumask);
4622 case CPU_DOWN_PREPARE:
4623 case CPU_DOWN_PREPARE_FROZEN:
4626 * If we were to free the buffer, then the user would
4627 * lose any trace that was in the buffer.
4637 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4639 * This is a basic integrity check of the ring buffer.
4640 * Late in the boot cycle this test will run when configured in.
4641 * It will kick off a thread per CPU that will go into a loop
4642 * writing to the per cpu ring buffer various sizes of data.
4643 * Some of the data will be large items, some small.
4645 * Another thread is created that goes into a spin, sending out
4646 * IPIs to the other CPUs to also write into the ring buffer.
4647 * this is to test the nesting ability of the buffer.
4649 * Basic stats are recorded and reported. If something in the
4650 * ring buffer should happen that's not expected, a big warning
4651 * is displayed and all ring buffers are disabled.
4653 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4655 struct rb_test_data {
4656 struct ring_buffer *buffer;
4657 unsigned long events;
4658 unsigned long bytes_written;
4659 unsigned long bytes_alloc;
4660 unsigned long bytes_dropped;
4661 unsigned long events_nested;
4662 unsigned long bytes_written_nested;
4663 unsigned long bytes_alloc_nested;
4664 unsigned long bytes_dropped_nested;
4665 int min_size_nested;
4666 int max_size_nested;
4673 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4676 #define RB_TEST_BUFFER_SIZE 1048576
4678 static char rb_string[] __initdata =
4679 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4680 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4681 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4683 static bool rb_test_started __initdata;
4690 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4692 struct ring_buffer_event *event;
4693 struct rb_item *item;
4700 /* Have nested writes different that what is written */
4701 cnt = data->cnt + (nested ? 27 : 0);
4703 /* Multiply cnt by ~e, to make some unique increment */
4704 size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4706 len = size + sizeof(struct rb_item);
4708 started = rb_test_started;
4709 /* read rb_test_started before checking buffer enabled */
4712 event = ring_buffer_lock_reserve(data->buffer, len);
4714 /* Ignore dropped events before test starts. */
4717 data->bytes_dropped += len;
4719 data->bytes_dropped_nested += len;
4724 event_len = ring_buffer_event_length(event);
4726 if (RB_WARN_ON(data->buffer, event_len < len))
4729 item = ring_buffer_event_data(event);
4731 memcpy(item->str, rb_string, size);
4734 data->bytes_alloc_nested += event_len;
4735 data->bytes_written_nested += len;
4736 data->events_nested++;
4737 if (!data->min_size_nested || len < data->min_size_nested)
4738 data->min_size_nested = len;
4739 if (len > data->max_size_nested)
4740 data->max_size_nested = len;
4742 data->bytes_alloc += event_len;
4743 data->bytes_written += len;
4745 if (!data->min_size || len < data->min_size)
4746 data->max_size = len;
4747 if (len > data->max_size)
4748 data->max_size = len;
4752 ring_buffer_unlock_commit(data->buffer, event);
4757 static __init int rb_test(void *arg)
4759 struct rb_test_data *data = arg;
4761 while (!kthread_should_stop()) {
4762 rb_write_something(data, false);
4765 set_current_state(TASK_INTERRUPTIBLE);
4766 /* Now sleep between a min of 100-300us and a max of 1ms */
4767 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4773 static __init void rb_ipi(void *ignore)
4775 struct rb_test_data *data;
4776 int cpu = smp_processor_id();
4778 data = &rb_data[cpu];
4779 rb_write_something(data, true);
4782 static __init int rb_hammer_test(void *arg)
4784 while (!kthread_should_stop()) {
4786 /* Send an IPI to all cpus to write data! */
4787 smp_call_function(rb_ipi, NULL, 1);
4788 /* No sleep, but for non preempt, let others run */
4795 static __init int test_ringbuffer(void)
4797 struct task_struct *rb_hammer;
4798 struct ring_buffer *buffer;
4802 pr_info("Running ring buffer tests...\n");
4804 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4805 if (WARN_ON(!buffer))
4808 /* Disable buffer so that threads can't write to it yet */
4809 ring_buffer_record_off(buffer);
4811 for_each_online_cpu(cpu) {
4812 rb_data[cpu].buffer = buffer;
4813 rb_data[cpu].cpu = cpu;
4814 rb_data[cpu].cnt = cpu;
4815 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4816 "rbtester/%d", cpu);
4817 if (WARN_ON(!rb_threads[cpu])) {
4818 pr_cont("FAILED\n");
4823 kthread_bind(rb_threads[cpu], cpu);
4824 wake_up_process(rb_threads[cpu]);
4827 /* Now create the rb hammer! */
4828 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4829 if (WARN_ON(!rb_hammer)) {
4830 pr_cont("FAILED\n");
4835 ring_buffer_record_on(buffer);
4837 * Show buffer is enabled before setting rb_test_started.
4838 * Yes there's a small race window where events could be
4839 * dropped and the thread wont catch it. But when a ring
4840 * buffer gets enabled, there will always be some kind of
4841 * delay before other CPUs see it. Thus, we don't care about
4842 * those dropped events. We care about events dropped after
4843 * the threads see that the buffer is active.
4846 rb_test_started = true;
4848 set_current_state(TASK_INTERRUPTIBLE);
4849 /* Just run for 10 seconds */;
4850 schedule_timeout(10 * HZ);
4852 kthread_stop(rb_hammer);
4855 for_each_online_cpu(cpu) {
4856 if (!rb_threads[cpu])
4858 kthread_stop(rb_threads[cpu]);
4861 ring_buffer_free(buffer);
4866 pr_info("finished\n");
4867 for_each_online_cpu(cpu) {
4868 struct ring_buffer_event *event;
4869 struct rb_test_data *data = &rb_data[cpu];
4870 struct rb_item *item;
4871 unsigned long total_events;
4872 unsigned long total_dropped;
4873 unsigned long total_written;
4874 unsigned long total_alloc;
4875 unsigned long total_read = 0;
4876 unsigned long total_size = 0;
4877 unsigned long total_len = 0;
4878 unsigned long total_lost = 0;
4881 int small_event_size;
4885 total_events = data->events + data->events_nested;
4886 total_written = data->bytes_written + data->bytes_written_nested;
4887 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
4888 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
4890 big_event_size = data->max_size + data->max_size_nested;
4891 small_event_size = data->min_size + data->min_size_nested;
4893 pr_info("CPU %d:\n", cpu);
4894 pr_info(" events: %ld\n", total_events);
4895 pr_info(" dropped bytes: %ld\n", total_dropped);
4896 pr_info(" alloced bytes: %ld\n", total_alloc);
4897 pr_info(" written bytes: %ld\n", total_written);
4898 pr_info(" biggest event: %d\n", big_event_size);
4899 pr_info(" smallest event: %d\n", small_event_size);
4901 if (RB_WARN_ON(buffer, total_dropped))
4906 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
4908 item = ring_buffer_event_data(event);
4909 total_len += ring_buffer_event_length(event);
4910 total_size += item->size + sizeof(struct rb_item);
4911 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
4912 pr_info("FAILED!\n");
4913 pr_info("buffer had: %.*s\n", item->size, item->str);
4914 pr_info("expected: %.*s\n", item->size, rb_string);
4915 RB_WARN_ON(buffer, 1);
4926 pr_info(" read events: %ld\n", total_read);
4927 pr_info(" lost events: %ld\n", total_lost);
4928 pr_info(" total events: %ld\n", total_lost + total_read);
4929 pr_info(" recorded len bytes: %ld\n", total_len);
4930 pr_info(" recorded size bytes: %ld\n", total_size);
4932 pr_info(" With dropped events, record len and size may not match\n"
4933 " alloced and written from above\n");
4935 if (RB_WARN_ON(buffer, total_len != total_alloc ||
4936 total_size != total_written))
4939 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
4945 pr_info("Ring buffer PASSED!\n");
4947 ring_buffer_free(buffer);
4951 late_initcall(test_ringbuffer);
4952 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */