4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 #include <linux/trace_events.h>
7 #include <linux/ring_buffer.h>
8 #include <linux/trace_clock.h>
9 #include <linux/sched/clock.h>
10 #include <linux/trace_seq.h>
11 #include <linux/spinlock.h>
12 #include <linux/irq_work.h>
13 #include <linux/uaccess.h>
14 #include <linux/hardirq.h>
15 #include <linux/kthread.h> /* for self test */
16 #include <linux/module.h>
17 #include <linux/percpu.h>
18 #include <linux/mutex.h>
19 #include <linux/delay.h>
20 #include <linux/slab.h>
21 #include <linux/init.h>
22 #include <linux/hash.h>
23 #include <linux/list.h>
24 #include <linux/cpu.h>
26 #include <asm/local.h>
28 static void update_pages_handler(struct work_struct *work);
31 * The ring buffer header is special. We must manually up keep it.
33 int ring_buffer_print_entry_header(struct trace_seq *s)
35 trace_seq_puts(s, "# compressed entry header\n");
36 trace_seq_puts(s, "\ttype_len : 5 bits\n");
37 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
38 trace_seq_puts(s, "\tarray : 32 bits\n");
39 trace_seq_putc(s, '\n');
40 trace_seq_printf(s, "\tpadding : type == %d\n",
41 RINGBUF_TYPE_PADDING);
42 trace_seq_printf(s, "\ttime_extend : type == %d\n",
43 RINGBUF_TYPE_TIME_EXTEND);
44 trace_seq_printf(s, "\tdata max type_len == %d\n",
45 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
47 return !trace_seq_has_overflowed(s);
51 * The ring buffer is made up of a list of pages. A separate list of pages is
52 * allocated for each CPU. A writer may only write to a buffer that is
53 * associated with the CPU it is currently executing on. A reader may read
54 * from any per cpu buffer.
56 * The reader is special. For each per cpu buffer, the reader has its own
57 * reader page. When a reader has read the entire reader page, this reader
58 * page is swapped with another page in the ring buffer.
60 * Now, as long as the writer is off the reader page, the reader can do what
61 * ever it wants with that page. The writer will never write to that page
62 * again (as long as it is out of the ring buffer).
64 * Here's some silly ASCII art.
67 * |reader| RING BUFFER
69 * +------+ +---+ +---+ +---+
78 * |reader| RING BUFFER
79 * |page |------------------v
80 * +------+ +---+ +---+ +---+
89 * |reader| RING BUFFER
90 * |page |------------------v
91 * +------+ +---+ +---+ +---+
96 * +------------------------------+
100 * |buffer| RING BUFFER
101 * |page |------------------v
102 * +------+ +---+ +---+ +---+
104 * | New +---+ +---+ +---+
107 * +------------------------------+
110 * After we make this swap, the reader can hand this page off to the splice
111 * code and be done with it. It can even allocate a new page if it needs to
112 * and swap that into the ring buffer.
114 * We will be using cmpxchg soon to make all this lockless.
118 /* Used for individual buffers (after the counter) */
119 #define RB_BUFFER_OFF (1 << 20)
121 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
123 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
124 #define RB_ALIGNMENT 4U
125 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
126 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
128 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
129 # define RB_FORCE_8BYTE_ALIGNMENT 0
130 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
132 # define RB_FORCE_8BYTE_ALIGNMENT 1
133 # define RB_ARCH_ALIGNMENT 8U
136 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
138 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
139 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
142 RB_LEN_TIME_EXTEND = 8,
143 RB_LEN_TIME_STAMP = 16,
146 #define skip_time_extend(event) \
147 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
149 static inline int rb_null_event(struct ring_buffer_event *event)
151 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
154 static void rb_event_set_padding(struct ring_buffer_event *event)
156 /* padding has a NULL time_delta */
157 event->type_len = RINGBUF_TYPE_PADDING;
158 event->time_delta = 0;
162 rb_event_data_length(struct ring_buffer_event *event)
167 length = event->type_len * RB_ALIGNMENT;
169 length = event->array[0];
170 return length + RB_EVNT_HDR_SIZE;
174 * Return the length of the given event. Will return
175 * the length of the time extend if the event is a
178 static inline unsigned
179 rb_event_length(struct ring_buffer_event *event)
181 switch (event->type_len) {
182 case RINGBUF_TYPE_PADDING:
183 if (rb_null_event(event))
186 return event->array[0] + RB_EVNT_HDR_SIZE;
188 case RINGBUF_TYPE_TIME_EXTEND:
189 return RB_LEN_TIME_EXTEND;
191 case RINGBUF_TYPE_TIME_STAMP:
192 return RB_LEN_TIME_STAMP;
194 case RINGBUF_TYPE_DATA:
195 return rb_event_data_length(event);
204 * Return total length of time extend and data,
205 * or just the event length for all other events.
207 static inline unsigned
208 rb_event_ts_length(struct ring_buffer_event *event)
212 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
213 /* time extends include the data event after it */
214 len = RB_LEN_TIME_EXTEND;
215 event = skip_time_extend(event);
217 return len + rb_event_length(event);
221 * ring_buffer_event_length - return the length of the event
222 * @event: the event to get the length of
224 * Returns the size of the data load of a data event.
225 * If the event is something other than a data event, it
226 * returns the size of the event itself. With the exception
227 * of a TIME EXTEND, where it still returns the size of the
228 * data load of the data event after it.
230 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
234 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
235 event = skip_time_extend(event);
237 length = rb_event_length(event);
238 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
240 length -= RB_EVNT_HDR_SIZE;
241 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
242 length -= sizeof(event->array[0]);
245 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
247 /* inline for ring buffer fast paths */
248 static __always_inline void *
249 rb_event_data(struct ring_buffer_event *event)
251 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
252 event = skip_time_extend(event);
253 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
254 /* If length is in len field, then array[0] has the data */
256 return (void *)&event->array[0];
257 /* Otherwise length is in array[0] and array[1] has the data */
258 return (void *)&event->array[1];
262 * ring_buffer_event_data - return the data of the event
263 * @event: the event to get the data from
265 void *ring_buffer_event_data(struct ring_buffer_event *event)
267 return rb_event_data(event);
269 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
271 #define for_each_buffer_cpu(buffer, cpu) \
272 for_each_cpu(cpu, buffer->cpumask)
275 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
276 #define TS_DELTA_TEST (~TS_MASK)
278 /* Flag when events were overwritten */
279 #define RB_MISSED_EVENTS (1 << 31)
280 /* Missed count stored at end */
281 #define RB_MISSED_STORED (1 << 30)
283 #define RB_MISSED_FLAGS (RB_MISSED_EVENTS|RB_MISSED_STORED)
285 struct buffer_data_page {
286 u64 time_stamp; /* page time stamp */
287 local_t commit; /* write committed index */
288 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
292 * Note, the buffer_page list must be first. The buffer pages
293 * are allocated in cache lines, which means that each buffer
294 * page will be at the beginning of a cache line, and thus
295 * the least significant bits will be zero. We use this to
296 * add flags in the list struct pointers, to make the ring buffer
300 struct list_head list; /* list of buffer pages */
301 local_t write; /* index for next write */
302 unsigned read; /* index for next read */
303 local_t entries; /* entries on this page */
304 unsigned long real_end; /* real end of data */
305 struct buffer_data_page *page; /* Actual data page */
309 * The buffer page counters, write and entries, must be reset
310 * atomically when crossing page boundaries. To synchronize this
311 * update, two counters are inserted into the number. One is
312 * the actual counter for the write position or count on the page.
314 * The other is a counter of updaters. Before an update happens
315 * the update partition of the counter is incremented. This will
316 * allow the updater to update the counter atomically.
318 * The counter is 20 bits, and the state data is 12.
320 #define RB_WRITE_MASK 0xfffff
321 #define RB_WRITE_INTCNT (1 << 20)
323 static void rb_init_page(struct buffer_data_page *bpage)
325 local_set(&bpage->commit, 0);
329 * ring_buffer_page_len - the size of data on the page.
330 * @page: The page to read
332 * Returns the amount of data on the page, including buffer page header.
334 size_t ring_buffer_page_len(void *page)
336 struct buffer_data_page *bpage = page;
338 return (local_read(&bpage->commit) & ~RB_MISSED_FLAGS)
343 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
346 static void free_buffer_page(struct buffer_page *bpage)
348 free_page((unsigned long)bpage->page);
353 * We need to fit the time_stamp delta into 27 bits.
355 static inline int test_time_stamp(u64 delta)
357 if (delta & TS_DELTA_TEST)
362 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
364 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
365 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
367 int ring_buffer_print_page_header(struct trace_seq *s)
369 struct buffer_data_page field;
371 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
372 "offset:0;\tsize:%u;\tsigned:%u;\n",
373 (unsigned int)sizeof(field.time_stamp),
374 (unsigned int)is_signed_type(u64));
376 trace_seq_printf(s, "\tfield: local_t commit;\t"
377 "offset:%u;\tsize:%u;\tsigned:%u;\n",
378 (unsigned int)offsetof(typeof(field), commit),
379 (unsigned int)sizeof(field.commit),
380 (unsigned int)is_signed_type(long));
382 trace_seq_printf(s, "\tfield: int overwrite;\t"
383 "offset:%u;\tsize:%u;\tsigned:%u;\n",
384 (unsigned int)offsetof(typeof(field), commit),
386 (unsigned int)is_signed_type(long));
388 trace_seq_printf(s, "\tfield: char data;\t"
389 "offset:%u;\tsize:%u;\tsigned:%u;\n",
390 (unsigned int)offsetof(typeof(field), data),
391 (unsigned int)BUF_PAGE_SIZE,
392 (unsigned int)is_signed_type(char));
394 return !trace_seq_has_overflowed(s);
398 struct irq_work work;
399 wait_queue_head_t waiters;
400 wait_queue_head_t full_waiters;
401 bool waiters_pending;
402 bool full_waiters_pending;
407 * Structure to hold event state and handle nested events.
409 struct rb_event_info {
412 unsigned long length;
413 struct buffer_page *tail_page;
418 * Used for which event context the event is in.
425 * See trace_recursive_lock() comment below for more details.
437 * head_page == tail_page && head == tail then buffer is empty.
439 struct ring_buffer_per_cpu {
441 atomic_t record_disabled;
442 struct ring_buffer *buffer;
443 raw_spinlock_t reader_lock; /* serialize readers */
444 arch_spinlock_t lock;
445 struct lock_class_key lock_key;
446 struct buffer_data_page *free_page;
447 unsigned long nr_pages;
448 unsigned int current_context;
449 struct list_head *pages;
450 struct buffer_page *head_page; /* read from head */
451 struct buffer_page *tail_page; /* write to tail */
452 struct buffer_page *commit_page; /* committed pages */
453 struct buffer_page *reader_page;
454 unsigned long lost_events;
455 unsigned long last_overrun;
456 local_t entries_bytes;
459 local_t commit_overrun;
460 local_t dropped_events;
464 unsigned long read_bytes;
467 /* ring buffer pages to update, > 0 to add, < 0 to remove */
468 long nr_pages_to_update;
469 struct list_head new_pages; /* new pages to add */
470 struct work_struct update_pages_work;
471 struct completion update_done;
473 struct rb_irq_work irq_work;
479 atomic_t record_disabled;
480 atomic_t resize_disabled;
481 cpumask_var_t cpumask;
483 struct lock_class_key *reader_lock_key;
487 struct ring_buffer_per_cpu **buffers;
489 struct hlist_node node;
492 struct rb_irq_work irq_work;
495 struct ring_buffer_iter {
496 struct ring_buffer_per_cpu *cpu_buffer;
498 struct buffer_page *head_page;
499 struct buffer_page *cache_reader_page;
500 unsigned long cache_read;
505 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
507 * Schedules a delayed work to wake up any task that is blocked on the
508 * ring buffer waiters queue.
510 static void rb_wake_up_waiters(struct irq_work *work)
512 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
514 wake_up_all(&rbwork->waiters);
515 if (rbwork->wakeup_full) {
516 rbwork->wakeup_full = false;
517 wake_up_all(&rbwork->full_waiters);
522 * ring_buffer_wait - wait for input to the ring buffer
523 * @buffer: buffer to wait on
524 * @cpu: the cpu buffer to wait on
525 * @full: wait until a full page is available, if @cpu != RING_BUFFER_ALL_CPUS
527 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
528 * as data is added to any of the @buffer's cpu buffers. Otherwise
529 * it will wait for data to be added to a specific cpu buffer.
531 int ring_buffer_wait(struct ring_buffer *buffer, int cpu, bool full)
533 struct ring_buffer_per_cpu *uninitialized_var(cpu_buffer);
535 struct rb_irq_work *work;
539 * Depending on what the caller is waiting for, either any
540 * data in any cpu buffer, or a specific buffer, put the
541 * caller on the appropriate wait queue.
543 if (cpu == RING_BUFFER_ALL_CPUS) {
544 work = &buffer->irq_work;
545 /* Full only makes sense on per cpu reads */
548 if (!cpumask_test_cpu(cpu, buffer->cpumask))
550 cpu_buffer = buffer->buffers[cpu];
551 work = &cpu_buffer->irq_work;
557 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
559 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
562 * The events can happen in critical sections where
563 * checking a work queue can cause deadlocks.
564 * After adding a task to the queue, this flag is set
565 * only to notify events to try to wake up the queue
568 * We don't clear it even if the buffer is no longer
569 * empty. The flag only causes the next event to run
570 * irq_work to do the work queue wake up. The worse
571 * that can happen if we race with !trace_empty() is that
572 * an event will cause an irq_work to try to wake up
575 * There's no reason to protect this flag either, as
576 * the work queue and irq_work logic will do the necessary
577 * synchronization for the wake ups. The only thing
578 * that is necessary is that the wake up happens after
579 * a task has been queued. It's OK for spurious wake ups.
582 work->full_waiters_pending = true;
584 work->waiters_pending = true;
586 if (signal_pending(current)) {
591 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
594 if (cpu != RING_BUFFER_ALL_CPUS &&
595 !ring_buffer_empty_cpu(buffer, cpu)) {
602 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
603 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
604 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
614 finish_wait(&work->full_waiters, &wait);
616 finish_wait(&work->waiters, &wait);
622 * ring_buffer_poll_wait - poll on buffer input
623 * @buffer: buffer to wait on
624 * @cpu: the cpu buffer to wait on
625 * @filp: the file descriptor
626 * @poll_table: The poll descriptor
628 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
629 * as data is added to any of the @buffer's cpu buffers. Otherwise
630 * it will wait for data to be added to a specific cpu buffer.
632 * Returns POLLIN | POLLRDNORM if data exists in the buffers,
635 int ring_buffer_poll_wait(struct ring_buffer *buffer, int cpu,
636 struct file *filp, poll_table *poll_table)
638 struct ring_buffer_per_cpu *cpu_buffer;
639 struct rb_irq_work *work;
641 if (cpu == RING_BUFFER_ALL_CPUS)
642 work = &buffer->irq_work;
644 if (!cpumask_test_cpu(cpu, buffer->cpumask))
647 cpu_buffer = buffer->buffers[cpu];
648 work = &cpu_buffer->irq_work;
651 poll_wait(filp, &work->waiters, poll_table);
652 work->waiters_pending = true;
654 * There's a tight race between setting the waiters_pending and
655 * checking if the ring buffer is empty. Once the waiters_pending bit
656 * is set, the next event will wake the task up, but we can get stuck
657 * if there's only a single event in.
659 * FIXME: Ideally, we need a memory barrier on the writer side as well,
660 * but adding a memory barrier to all events will cause too much of a
661 * performance hit in the fast path. We only need a memory barrier when
662 * the buffer goes from empty to having content. But as this race is
663 * extremely small, and it's not a problem if another event comes in, we
668 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
669 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
670 return POLLIN | POLLRDNORM;
674 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
675 #define RB_WARN_ON(b, cond) \
677 int _____ret = unlikely(cond); \
679 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
680 struct ring_buffer_per_cpu *__b = \
682 atomic_inc(&__b->buffer->record_disabled); \
684 atomic_inc(&b->record_disabled); \
690 /* Up this if you want to test the TIME_EXTENTS and normalization */
691 #define DEBUG_SHIFT 0
693 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
695 /* shift to debug/test normalization and TIME_EXTENTS */
696 return buffer->clock() << DEBUG_SHIFT;
699 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
703 preempt_disable_notrace();
704 time = rb_time_stamp(buffer);
705 preempt_enable_notrace();
709 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
711 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
714 /* Just stupid testing the normalize function and deltas */
717 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
720 * Making the ring buffer lockless makes things tricky.
721 * Although writes only happen on the CPU that they are on,
722 * and they only need to worry about interrupts. Reads can
725 * The reader page is always off the ring buffer, but when the
726 * reader finishes with a page, it needs to swap its page with
727 * a new one from the buffer. The reader needs to take from
728 * the head (writes go to the tail). But if a writer is in overwrite
729 * mode and wraps, it must push the head page forward.
731 * Here lies the problem.
733 * The reader must be careful to replace only the head page, and
734 * not another one. As described at the top of the file in the
735 * ASCII art, the reader sets its old page to point to the next
736 * page after head. It then sets the page after head to point to
737 * the old reader page. But if the writer moves the head page
738 * during this operation, the reader could end up with the tail.
740 * We use cmpxchg to help prevent this race. We also do something
741 * special with the page before head. We set the LSB to 1.
743 * When the writer must push the page forward, it will clear the
744 * bit that points to the head page, move the head, and then set
745 * the bit that points to the new head page.
747 * We also don't want an interrupt coming in and moving the head
748 * page on another writer. Thus we use the second LSB to catch
751 * head->list->prev->next bit 1 bit 0
754 * Points to head page 0 1
757 * Note we can not trust the prev pointer of the head page, because:
759 * +----+ +-----+ +-----+
760 * | |------>| T |---X--->| N |
762 * +----+ +-----+ +-----+
765 * +----------| R |----------+ |
769 * Key: ---X--> HEAD flag set in pointer
774 * (see __rb_reserve_next() to see where this happens)
776 * What the above shows is that the reader just swapped out
777 * the reader page with a page in the buffer, but before it
778 * could make the new header point back to the new page added
779 * it was preempted by a writer. The writer moved forward onto
780 * the new page added by the reader and is about to move forward
783 * You can see, it is legitimate for the previous pointer of
784 * the head (or any page) not to point back to itself. But only
788 #define RB_PAGE_NORMAL 0UL
789 #define RB_PAGE_HEAD 1UL
790 #define RB_PAGE_UPDATE 2UL
793 #define RB_FLAG_MASK 3UL
795 /* PAGE_MOVED is not part of the mask */
796 #define RB_PAGE_MOVED 4UL
799 * rb_list_head - remove any bit
801 static struct list_head *rb_list_head(struct list_head *list)
803 unsigned long val = (unsigned long)list;
805 return (struct list_head *)(val & ~RB_FLAG_MASK);
809 * rb_is_head_page - test if the given page is the head page
811 * Because the reader may move the head_page pointer, we can
812 * not trust what the head page is (it may be pointing to
813 * the reader page). But if the next page is a header page,
814 * its flags will be non zero.
817 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
818 struct buffer_page *page, struct list_head *list)
822 val = (unsigned long)list->next;
824 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
825 return RB_PAGE_MOVED;
827 return val & RB_FLAG_MASK;
833 * The unique thing about the reader page, is that, if the
834 * writer is ever on it, the previous pointer never points
835 * back to the reader page.
837 static bool rb_is_reader_page(struct buffer_page *page)
839 struct list_head *list = page->list.prev;
841 return rb_list_head(list->next) != &page->list;
845 * rb_set_list_to_head - set a list_head to be pointing to head.
847 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
848 struct list_head *list)
852 ptr = (unsigned long *)&list->next;
853 *ptr |= RB_PAGE_HEAD;
854 *ptr &= ~RB_PAGE_UPDATE;
858 * rb_head_page_activate - sets up head page
860 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
862 struct buffer_page *head;
864 head = cpu_buffer->head_page;
869 * Set the previous list pointer to have the HEAD flag.
871 rb_set_list_to_head(cpu_buffer, head->list.prev);
874 static void rb_list_head_clear(struct list_head *list)
876 unsigned long *ptr = (unsigned long *)&list->next;
878 *ptr &= ~RB_FLAG_MASK;
882 * rb_head_page_dactivate - clears head page ptr (for free list)
885 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
887 struct list_head *hd;
889 /* Go through the whole list and clear any pointers found. */
890 rb_list_head_clear(cpu_buffer->pages);
892 list_for_each(hd, cpu_buffer->pages)
893 rb_list_head_clear(hd);
896 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
897 struct buffer_page *head,
898 struct buffer_page *prev,
899 int old_flag, int new_flag)
901 struct list_head *list;
902 unsigned long val = (unsigned long)&head->list;
907 val &= ~RB_FLAG_MASK;
909 ret = cmpxchg((unsigned long *)&list->next,
910 val | old_flag, val | new_flag);
912 /* check if the reader took the page */
913 if ((ret & ~RB_FLAG_MASK) != val)
914 return RB_PAGE_MOVED;
916 return ret & RB_FLAG_MASK;
919 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
920 struct buffer_page *head,
921 struct buffer_page *prev,
924 return rb_head_page_set(cpu_buffer, head, prev,
925 old_flag, RB_PAGE_UPDATE);
928 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
929 struct buffer_page *head,
930 struct buffer_page *prev,
933 return rb_head_page_set(cpu_buffer, head, prev,
934 old_flag, RB_PAGE_HEAD);
937 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
938 struct buffer_page *head,
939 struct buffer_page *prev,
942 return rb_head_page_set(cpu_buffer, head, prev,
943 old_flag, RB_PAGE_NORMAL);
946 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
947 struct buffer_page **bpage)
949 struct list_head *p = rb_list_head((*bpage)->list.next);
951 *bpage = list_entry(p, struct buffer_page, list);
954 static struct buffer_page *
955 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
957 struct buffer_page *head;
958 struct buffer_page *page;
959 struct list_head *list;
962 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
966 list = cpu_buffer->pages;
967 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
970 page = head = cpu_buffer->head_page;
972 * It is possible that the writer moves the header behind
973 * where we started, and we miss in one loop.
974 * A second loop should grab the header, but we'll do
975 * three loops just because I'm paranoid.
977 for (i = 0; i < 3; i++) {
979 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
980 cpu_buffer->head_page = page;
983 rb_inc_page(cpu_buffer, &page);
984 } while (page != head);
987 RB_WARN_ON(cpu_buffer, 1);
992 static int rb_head_page_replace(struct buffer_page *old,
993 struct buffer_page *new)
995 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
999 val = *ptr & ~RB_FLAG_MASK;
1000 val |= RB_PAGE_HEAD;
1002 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1008 * rb_tail_page_update - move the tail page forward
1010 static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1011 struct buffer_page *tail_page,
1012 struct buffer_page *next_page)
1014 unsigned long old_entries;
1015 unsigned long old_write;
1018 * The tail page now needs to be moved forward.
1020 * We need to reset the tail page, but without messing
1021 * with possible erasing of data brought in by interrupts
1022 * that have moved the tail page and are currently on it.
1024 * We add a counter to the write field to denote this.
1026 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1027 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1030 * Just make sure we have seen our old_write and synchronize
1031 * with any interrupts that come in.
1036 * If the tail page is still the same as what we think
1037 * it is, then it is up to us to update the tail
1040 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1041 /* Zero the write counter */
1042 unsigned long val = old_write & ~RB_WRITE_MASK;
1043 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1046 * This will only succeed if an interrupt did
1047 * not come in and change it. In which case, we
1048 * do not want to modify it.
1050 * We add (void) to let the compiler know that we do not care
1051 * about the return value of these functions. We use the
1052 * cmpxchg to only update if an interrupt did not already
1053 * do it for us. If the cmpxchg fails, we don't care.
1055 (void)local_cmpxchg(&next_page->write, old_write, val);
1056 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1059 * No need to worry about races with clearing out the commit.
1060 * it only can increment when a commit takes place. But that
1061 * only happens in the outer most nested commit.
1063 local_set(&next_page->page->commit, 0);
1065 /* Again, either we update tail_page or an interrupt does */
1066 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1070 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1071 struct buffer_page *bpage)
1073 unsigned long val = (unsigned long)bpage;
1075 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1082 * rb_check_list - make sure a pointer to a list has the last bits zero
1084 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
1085 struct list_head *list)
1087 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
1089 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
1095 * rb_check_pages - integrity check of buffer pages
1096 * @cpu_buffer: CPU buffer with pages to test
1098 * As a safety measure we check to make sure the data pages have not
1101 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1103 struct list_head *head = cpu_buffer->pages;
1104 struct buffer_page *bpage, *tmp;
1106 /* Reset the head page if it exists */
1107 if (cpu_buffer->head_page)
1108 rb_set_head_page(cpu_buffer);
1110 rb_head_page_deactivate(cpu_buffer);
1112 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
1114 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
1117 if (rb_check_list(cpu_buffer, head))
1120 list_for_each_entry_safe(bpage, tmp, head, list) {
1121 if (RB_WARN_ON(cpu_buffer,
1122 bpage->list.next->prev != &bpage->list))
1124 if (RB_WARN_ON(cpu_buffer,
1125 bpage->list.prev->next != &bpage->list))
1127 if (rb_check_list(cpu_buffer, &bpage->list))
1131 rb_head_page_activate(cpu_buffer);
1136 static int __rb_allocate_pages(long nr_pages, struct list_head *pages, int cpu)
1138 struct buffer_page *bpage, *tmp;
1141 /* Check if the available memory is there first */
1142 i = si_mem_available();
1146 for (i = 0; i < nr_pages; i++) {
1149 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1150 * gracefully without invoking oom-killer and the system is not
1153 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1154 GFP_KERNEL | __GFP_RETRY_MAYFAIL,
1159 list_add(&bpage->list, pages);
1161 page = alloc_pages_node(cpu_to_node(cpu),
1162 GFP_KERNEL | __GFP_RETRY_MAYFAIL, 0);
1165 bpage->page = page_address(page);
1166 rb_init_page(bpage->page);
1172 list_for_each_entry_safe(bpage, tmp, pages, list) {
1173 list_del_init(&bpage->list);
1174 free_buffer_page(bpage);
1180 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1181 unsigned long nr_pages)
1187 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1191 * The ring buffer page list is a circular list that does not
1192 * start and end with a list head. All page list items point to
1195 cpu_buffer->pages = pages.next;
1198 cpu_buffer->nr_pages = nr_pages;
1200 rb_check_pages(cpu_buffer);
1205 static struct ring_buffer_per_cpu *
1206 rb_allocate_cpu_buffer(struct ring_buffer *buffer, long nr_pages, int cpu)
1208 struct ring_buffer_per_cpu *cpu_buffer;
1209 struct buffer_page *bpage;
1213 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1214 GFP_KERNEL, cpu_to_node(cpu));
1218 cpu_buffer->cpu = cpu;
1219 cpu_buffer->buffer = buffer;
1220 raw_spin_lock_init(&cpu_buffer->reader_lock);
1221 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1222 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1223 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1224 init_completion(&cpu_buffer->update_done);
1225 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1226 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1227 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1229 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1230 GFP_KERNEL, cpu_to_node(cpu));
1232 goto fail_free_buffer;
1234 rb_check_bpage(cpu_buffer, bpage);
1236 cpu_buffer->reader_page = bpage;
1237 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1239 goto fail_free_reader;
1240 bpage->page = page_address(page);
1241 rb_init_page(bpage->page);
1243 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1244 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1246 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1248 goto fail_free_reader;
1250 cpu_buffer->head_page
1251 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1252 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1254 rb_head_page_activate(cpu_buffer);
1259 free_buffer_page(cpu_buffer->reader_page);
1266 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1268 struct list_head *head = cpu_buffer->pages;
1269 struct buffer_page *bpage, *tmp;
1271 free_buffer_page(cpu_buffer->reader_page);
1273 rb_head_page_deactivate(cpu_buffer);
1276 list_for_each_entry_safe(bpage, tmp, head, list) {
1277 list_del_init(&bpage->list);
1278 free_buffer_page(bpage);
1280 bpage = list_entry(head, struct buffer_page, list);
1281 free_buffer_page(bpage);
1288 * __ring_buffer_alloc - allocate a new ring_buffer
1289 * @size: the size in bytes per cpu that is needed.
1290 * @flags: attributes to set for the ring buffer.
1292 * Currently the only flag that is available is the RB_FL_OVERWRITE
1293 * flag. This flag means that the buffer will overwrite old data
1294 * when the buffer wraps. If this flag is not set, the buffer will
1295 * drop data when the tail hits the head.
1297 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1298 struct lock_class_key *key)
1300 struct ring_buffer *buffer;
1306 /* keep it in its own cache line */
1307 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1312 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1313 goto fail_free_buffer;
1315 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1316 buffer->flags = flags;
1317 buffer->clock = trace_clock_local;
1318 buffer->reader_lock_key = key;
1320 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1321 init_waitqueue_head(&buffer->irq_work.waiters);
1323 /* need at least two pages */
1327 buffer->cpus = nr_cpu_ids;
1329 bsize = sizeof(void *) * nr_cpu_ids;
1330 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1332 if (!buffer->buffers)
1333 goto fail_free_cpumask;
1335 cpu = raw_smp_processor_id();
1336 cpumask_set_cpu(cpu, buffer->cpumask);
1337 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1338 if (!buffer->buffers[cpu])
1339 goto fail_free_buffers;
1341 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1343 goto fail_free_buffers;
1345 mutex_init(&buffer->mutex);
1350 for_each_buffer_cpu(buffer, cpu) {
1351 if (buffer->buffers[cpu])
1352 rb_free_cpu_buffer(buffer->buffers[cpu]);
1354 kfree(buffer->buffers);
1357 free_cpumask_var(buffer->cpumask);
1363 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1366 * ring_buffer_free - free a ring buffer.
1367 * @buffer: the buffer to free.
1370 ring_buffer_free(struct ring_buffer *buffer)
1374 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1376 for_each_buffer_cpu(buffer, cpu)
1377 rb_free_cpu_buffer(buffer->buffers[cpu]);
1379 kfree(buffer->buffers);
1380 free_cpumask_var(buffer->cpumask);
1384 EXPORT_SYMBOL_GPL(ring_buffer_free);
1386 void ring_buffer_set_clock(struct ring_buffer *buffer,
1389 buffer->clock = clock;
1392 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1394 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1396 return local_read(&bpage->entries) & RB_WRITE_MASK;
1399 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1401 return local_read(&bpage->write) & RB_WRITE_MASK;
1405 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1407 struct list_head *tail_page, *to_remove, *next_page;
1408 struct buffer_page *to_remove_page, *tmp_iter_page;
1409 struct buffer_page *last_page, *first_page;
1410 unsigned long nr_removed;
1411 unsigned long head_bit;
1416 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1417 atomic_inc(&cpu_buffer->record_disabled);
1419 * We don't race with the readers since we have acquired the reader
1420 * lock. We also don't race with writers after disabling recording.
1421 * This makes it easy to figure out the first and the last page to be
1422 * removed from the list. We unlink all the pages in between including
1423 * the first and last pages. This is done in a busy loop so that we
1424 * lose the least number of traces.
1425 * The pages are freed after we restart recording and unlock readers.
1427 tail_page = &cpu_buffer->tail_page->list;
1430 * tail page might be on reader page, we remove the next page
1431 * from the ring buffer
1433 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1434 tail_page = rb_list_head(tail_page->next);
1435 to_remove = tail_page;
1437 /* start of pages to remove */
1438 first_page = list_entry(rb_list_head(to_remove->next),
1439 struct buffer_page, list);
1441 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1442 to_remove = rb_list_head(to_remove)->next;
1443 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1446 next_page = rb_list_head(to_remove)->next;
1449 * Now we remove all pages between tail_page and next_page.
1450 * Make sure that we have head_bit value preserved for the
1453 tail_page->next = (struct list_head *)((unsigned long)next_page |
1455 next_page = rb_list_head(next_page);
1456 next_page->prev = tail_page;
1458 /* make sure pages points to a valid page in the ring buffer */
1459 cpu_buffer->pages = next_page;
1461 /* update head page */
1463 cpu_buffer->head_page = list_entry(next_page,
1464 struct buffer_page, list);
1467 * change read pointer to make sure any read iterators reset
1470 cpu_buffer->read = 0;
1472 /* pages are removed, resume tracing and then free the pages */
1473 atomic_dec(&cpu_buffer->record_disabled);
1474 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1476 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1478 /* last buffer page to remove */
1479 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1481 tmp_iter_page = first_page;
1486 to_remove_page = tmp_iter_page;
1487 rb_inc_page(cpu_buffer, &tmp_iter_page);
1489 /* update the counters */
1490 page_entries = rb_page_entries(to_remove_page);
1493 * If something was added to this page, it was full
1494 * since it is not the tail page. So we deduct the
1495 * bytes consumed in ring buffer from here.
1496 * Increment overrun to account for the lost events.
1498 local_add(page_entries, &cpu_buffer->overrun);
1499 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1503 * We have already removed references to this list item, just
1504 * free up the buffer_page and its page
1506 free_buffer_page(to_remove_page);
1509 } while (to_remove_page != last_page);
1511 RB_WARN_ON(cpu_buffer, nr_removed);
1513 return nr_removed == 0;
1517 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1519 struct list_head *pages = &cpu_buffer->new_pages;
1520 int retries, success;
1522 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1524 * We are holding the reader lock, so the reader page won't be swapped
1525 * in the ring buffer. Now we are racing with the writer trying to
1526 * move head page and the tail page.
1527 * We are going to adapt the reader page update process where:
1528 * 1. We first splice the start and end of list of new pages between
1529 * the head page and its previous page.
1530 * 2. We cmpxchg the prev_page->next to point from head page to the
1531 * start of new pages list.
1532 * 3. Finally, we update the head->prev to the end of new list.
1534 * We will try this process 10 times, to make sure that we don't keep
1540 struct list_head *head_page, *prev_page, *r;
1541 struct list_head *last_page, *first_page;
1542 struct list_head *head_page_with_bit;
1544 head_page = &rb_set_head_page(cpu_buffer)->list;
1547 prev_page = head_page->prev;
1549 first_page = pages->next;
1550 last_page = pages->prev;
1552 head_page_with_bit = (struct list_head *)
1553 ((unsigned long)head_page | RB_PAGE_HEAD);
1555 last_page->next = head_page_with_bit;
1556 first_page->prev = prev_page;
1558 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1560 if (r == head_page_with_bit) {
1562 * yay, we replaced the page pointer to our new list,
1563 * now, we just have to update to head page's prev
1564 * pointer to point to end of list
1566 head_page->prev = last_page;
1573 INIT_LIST_HEAD(pages);
1575 * If we weren't successful in adding in new pages, warn and stop
1578 RB_WARN_ON(cpu_buffer, !success);
1579 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1581 /* free pages if they weren't inserted */
1583 struct buffer_page *bpage, *tmp;
1584 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1586 list_del_init(&bpage->list);
1587 free_buffer_page(bpage);
1593 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1597 if (cpu_buffer->nr_pages_to_update > 0)
1598 success = rb_insert_pages(cpu_buffer);
1600 success = rb_remove_pages(cpu_buffer,
1601 -cpu_buffer->nr_pages_to_update);
1604 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1607 static void update_pages_handler(struct work_struct *work)
1609 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1610 struct ring_buffer_per_cpu, update_pages_work);
1611 rb_update_pages(cpu_buffer);
1612 complete(&cpu_buffer->update_done);
1616 * ring_buffer_resize - resize the ring buffer
1617 * @buffer: the buffer to resize.
1618 * @size: the new size.
1619 * @cpu_id: the cpu buffer to resize
1621 * Minimum size is 2 * BUF_PAGE_SIZE.
1623 * Returns 0 on success and < 0 on failure.
1625 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1628 struct ring_buffer_per_cpu *cpu_buffer;
1629 unsigned long nr_pages;
1633 * Always succeed at resizing a non-existent buffer:
1638 /* Make sure the requested buffer exists */
1639 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1640 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1643 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1645 /* we need a minimum of two pages */
1649 size = nr_pages * BUF_PAGE_SIZE;
1652 * Don't succeed if resizing is disabled, as a reader might be
1653 * manipulating the ring buffer and is expecting a sane state while
1656 if (atomic_read(&buffer->resize_disabled))
1659 /* prevent another thread from changing buffer sizes */
1660 mutex_lock(&buffer->mutex);
1662 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1663 /* calculate the pages to update */
1664 for_each_buffer_cpu(buffer, cpu) {
1665 cpu_buffer = buffer->buffers[cpu];
1667 cpu_buffer->nr_pages_to_update = nr_pages -
1668 cpu_buffer->nr_pages;
1670 * nothing more to do for removing pages or no update
1672 if (cpu_buffer->nr_pages_to_update <= 0)
1675 * to add pages, make sure all new pages can be
1676 * allocated without receiving ENOMEM
1678 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1679 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1680 &cpu_buffer->new_pages, cpu)) {
1681 /* not enough memory for new pages */
1689 * Fire off all the required work handlers
1690 * We can't schedule on offline CPUs, but it's not necessary
1691 * since we can change their buffer sizes without any race.
1693 for_each_buffer_cpu(buffer, cpu) {
1694 cpu_buffer = buffer->buffers[cpu];
1695 if (!cpu_buffer->nr_pages_to_update)
1698 /* Can't run something on an offline CPU. */
1699 if (!cpu_online(cpu)) {
1700 rb_update_pages(cpu_buffer);
1701 cpu_buffer->nr_pages_to_update = 0;
1703 schedule_work_on(cpu,
1704 &cpu_buffer->update_pages_work);
1708 /* wait for all the updates to complete */
1709 for_each_buffer_cpu(buffer, cpu) {
1710 cpu_buffer = buffer->buffers[cpu];
1711 if (!cpu_buffer->nr_pages_to_update)
1714 if (cpu_online(cpu))
1715 wait_for_completion(&cpu_buffer->update_done);
1716 cpu_buffer->nr_pages_to_update = 0;
1721 /* Make sure this CPU has been intitialized */
1722 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1725 cpu_buffer = buffer->buffers[cpu_id];
1727 if (nr_pages == cpu_buffer->nr_pages)
1730 cpu_buffer->nr_pages_to_update = nr_pages -
1731 cpu_buffer->nr_pages;
1733 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1734 if (cpu_buffer->nr_pages_to_update > 0 &&
1735 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1736 &cpu_buffer->new_pages, cpu_id)) {
1743 /* Can't run something on an offline CPU. */
1744 if (!cpu_online(cpu_id))
1745 rb_update_pages(cpu_buffer);
1747 schedule_work_on(cpu_id,
1748 &cpu_buffer->update_pages_work);
1749 wait_for_completion(&cpu_buffer->update_done);
1752 cpu_buffer->nr_pages_to_update = 0;
1758 * The ring buffer resize can happen with the ring buffer
1759 * enabled, so that the update disturbs the tracing as little
1760 * as possible. But if the buffer is disabled, we do not need
1761 * to worry about that, and we can take the time to verify
1762 * that the buffer is not corrupt.
1764 if (atomic_read(&buffer->record_disabled)) {
1765 atomic_inc(&buffer->record_disabled);
1767 * Even though the buffer was disabled, we must make sure
1768 * that it is truly disabled before calling rb_check_pages.
1769 * There could have been a race between checking
1770 * record_disable and incrementing it.
1772 synchronize_sched();
1773 for_each_buffer_cpu(buffer, cpu) {
1774 cpu_buffer = buffer->buffers[cpu];
1775 rb_check_pages(cpu_buffer);
1777 atomic_dec(&buffer->record_disabled);
1780 mutex_unlock(&buffer->mutex);
1784 for_each_buffer_cpu(buffer, cpu) {
1785 struct buffer_page *bpage, *tmp;
1787 cpu_buffer = buffer->buffers[cpu];
1788 cpu_buffer->nr_pages_to_update = 0;
1790 if (list_empty(&cpu_buffer->new_pages))
1793 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1795 list_del_init(&bpage->list);
1796 free_buffer_page(bpage);
1799 mutex_unlock(&buffer->mutex);
1802 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1804 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1806 mutex_lock(&buffer->mutex);
1808 buffer->flags |= RB_FL_OVERWRITE;
1810 buffer->flags &= ~RB_FL_OVERWRITE;
1811 mutex_unlock(&buffer->mutex);
1813 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1815 static __always_inline void *
1816 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1818 return bpage->data + index;
1821 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1823 return bpage->page->data + index;
1826 static __always_inline struct ring_buffer_event *
1827 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1829 return __rb_page_index(cpu_buffer->reader_page,
1830 cpu_buffer->reader_page->read);
1833 static __always_inline struct ring_buffer_event *
1834 rb_iter_head_event(struct ring_buffer_iter *iter)
1836 return __rb_page_index(iter->head_page, iter->head);
1839 static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
1841 return local_read(&bpage->page->commit);
1844 /* Size is determined by what has been committed */
1845 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
1847 return rb_page_commit(bpage);
1850 static __always_inline unsigned
1851 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1853 return rb_page_commit(cpu_buffer->commit_page);
1856 static __always_inline unsigned
1857 rb_event_index(struct ring_buffer_event *event)
1859 unsigned long addr = (unsigned long)event;
1861 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1864 static void rb_inc_iter(struct ring_buffer_iter *iter)
1866 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1869 * The iterator could be on the reader page (it starts there).
1870 * But the head could have moved, since the reader was
1871 * found. Check for this case and assign the iterator
1872 * to the head page instead of next.
1874 if (iter->head_page == cpu_buffer->reader_page)
1875 iter->head_page = rb_set_head_page(cpu_buffer);
1877 rb_inc_page(cpu_buffer, &iter->head_page);
1879 iter->read_stamp = iter->head_page->page->time_stamp;
1884 * rb_handle_head_page - writer hit the head page
1886 * Returns: +1 to retry page
1891 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1892 struct buffer_page *tail_page,
1893 struct buffer_page *next_page)
1895 struct buffer_page *new_head;
1900 entries = rb_page_entries(next_page);
1903 * The hard part is here. We need to move the head
1904 * forward, and protect against both readers on
1905 * other CPUs and writers coming in via interrupts.
1907 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1911 * type can be one of four:
1912 * NORMAL - an interrupt already moved it for us
1913 * HEAD - we are the first to get here.
1914 * UPDATE - we are the interrupt interrupting
1916 * MOVED - a reader on another CPU moved the next
1917 * pointer to its reader page. Give up
1924 * We changed the head to UPDATE, thus
1925 * it is our responsibility to update
1928 local_add(entries, &cpu_buffer->overrun);
1929 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1932 * The entries will be zeroed out when we move the
1936 /* still more to do */
1939 case RB_PAGE_UPDATE:
1941 * This is an interrupt that interrupt the
1942 * previous update. Still more to do.
1945 case RB_PAGE_NORMAL:
1947 * An interrupt came in before the update
1948 * and processed this for us.
1949 * Nothing left to do.
1954 * The reader is on another CPU and just did
1955 * a swap with our next_page.
1960 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1965 * Now that we are here, the old head pointer is
1966 * set to UPDATE. This will keep the reader from
1967 * swapping the head page with the reader page.
1968 * The reader (on another CPU) will spin till
1971 * We just need to protect against interrupts
1972 * doing the job. We will set the next pointer
1973 * to HEAD. After that, we set the old pointer
1974 * to NORMAL, but only if it was HEAD before.
1975 * otherwise we are an interrupt, and only
1976 * want the outer most commit to reset it.
1978 new_head = next_page;
1979 rb_inc_page(cpu_buffer, &new_head);
1981 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1985 * Valid returns are:
1986 * HEAD - an interrupt came in and already set it.
1987 * NORMAL - One of two things:
1988 * 1) We really set it.
1989 * 2) A bunch of interrupts came in and moved
1990 * the page forward again.
1994 case RB_PAGE_NORMAL:
1998 RB_WARN_ON(cpu_buffer, 1);
2003 * It is possible that an interrupt came in,
2004 * set the head up, then more interrupts came in
2005 * and moved it again. When we get back here,
2006 * the page would have been set to NORMAL but we
2007 * just set it back to HEAD.
2009 * How do you detect this? Well, if that happened
2010 * the tail page would have moved.
2012 if (ret == RB_PAGE_NORMAL) {
2013 struct buffer_page *buffer_tail_page;
2015 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2017 * If the tail had moved passed next, then we need
2018 * to reset the pointer.
2020 if (buffer_tail_page != tail_page &&
2021 buffer_tail_page != next_page)
2022 rb_head_page_set_normal(cpu_buffer, new_head,
2028 * If this was the outer most commit (the one that
2029 * changed the original pointer from HEAD to UPDATE),
2030 * then it is up to us to reset it to NORMAL.
2032 if (type == RB_PAGE_HEAD) {
2033 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2036 if (RB_WARN_ON(cpu_buffer,
2037 ret != RB_PAGE_UPDATE))
2045 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2046 unsigned long tail, struct rb_event_info *info)
2048 struct buffer_page *tail_page = info->tail_page;
2049 struct ring_buffer_event *event;
2050 unsigned long length = info->length;
2053 * Only the event that crossed the page boundary
2054 * must fill the old tail_page with padding.
2056 if (tail >= BUF_PAGE_SIZE) {
2058 * If the page was filled, then we still need
2059 * to update the real_end. Reset it to zero
2060 * and the reader will ignore it.
2062 if (tail == BUF_PAGE_SIZE)
2063 tail_page->real_end = 0;
2065 local_sub(length, &tail_page->write);
2069 event = __rb_page_index(tail_page, tail);
2071 /* account for padding bytes */
2072 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2075 * Save the original length to the meta data.
2076 * This will be used by the reader to add lost event
2079 tail_page->real_end = tail;
2082 * If this event is bigger than the minimum size, then
2083 * we need to be careful that we don't subtract the
2084 * write counter enough to allow another writer to slip
2086 * We put in a discarded commit instead, to make sure
2087 * that this space is not used again.
2089 * If we are less than the minimum size, we don't need to
2092 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2093 /* No room for any events */
2095 /* Mark the rest of the page with padding */
2096 rb_event_set_padding(event);
2098 /* Set the write back to the previous setting */
2099 local_sub(length, &tail_page->write);
2103 /* Put in a discarded event */
2104 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2105 event->type_len = RINGBUF_TYPE_PADDING;
2106 /* time delta must be non zero */
2107 event->time_delta = 1;
2109 /* Set write to end of buffer */
2110 length = (tail + length) - BUF_PAGE_SIZE;
2111 local_sub(length, &tail_page->write);
2114 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2117 * This is the slow path, force gcc not to inline it.
2119 static noinline struct ring_buffer_event *
2120 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2121 unsigned long tail, struct rb_event_info *info)
2123 struct buffer_page *tail_page = info->tail_page;
2124 struct buffer_page *commit_page = cpu_buffer->commit_page;
2125 struct ring_buffer *buffer = cpu_buffer->buffer;
2126 struct buffer_page *next_page;
2129 next_page = tail_page;
2131 rb_inc_page(cpu_buffer, &next_page);
2134 * If for some reason, we had an interrupt storm that made
2135 * it all the way around the buffer, bail, and warn
2138 if (unlikely(next_page == commit_page)) {
2139 local_inc(&cpu_buffer->commit_overrun);
2144 * This is where the fun begins!
2146 * We are fighting against races between a reader that
2147 * could be on another CPU trying to swap its reader
2148 * page with the buffer head.
2150 * We are also fighting against interrupts coming in and
2151 * moving the head or tail on us as well.
2153 * If the next page is the head page then we have filled
2154 * the buffer, unless the commit page is still on the
2157 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2160 * If the commit is not on the reader page, then
2161 * move the header page.
2163 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2165 * If we are not in overwrite mode,
2166 * this is easy, just stop here.
2168 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2169 local_inc(&cpu_buffer->dropped_events);
2173 ret = rb_handle_head_page(cpu_buffer,
2182 * We need to be careful here too. The
2183 * commit page could still be on the reader
2184 * page. We could have a small buffer, and
2185 * have filled up the buffer with events
2186 * from interrupts and such, and wrapped.
2188 * Note, if the tail page is also the on the
2189 * reader_page, we let it move out.
2191 if (unlikely((cpu_buffer->commit_page !=
2192 cpu_buffer->tail_page) &&
2193 (cpu_buffer->commit_page ==
2194 cpu_buffer->reader_page))) {
2195 local_inc(&cpu_buffer->commit_overrun);
2201 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2205 rb_reset_tail(cpu_buffer, tail, info);
2207 /* Commit what we have for now. */
2208 rb_end_commit(cpu_buffer);
2209 /* rb_end_commit() decs committing */
2210 local_inc(&cpu_buffer->committing);
2212 /* fail and let the caller try again */
2213 return ERR_PTR(-EAGAIN);
2217 rb_reset_tail(cpu_buffer, tail, info);
2222 /* Slow path, do not inline */
2223 static noinline struct ring_buffer_event *
2224 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
2226 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2228 /* Not the first event on the page? */
2229 if (rb_event_index(event)) {
2230 event->time_delta = delta & TS_MASK;
2231 event->array[0] = delta >> TS_SHIFT;
2233 /* nope, just zero it */
2234 event->time_delta = 0;
2235 event->array[0] = 0;
2238 return skip_time_extend(event);
2241 static inline bool rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2242 struct ring_buffer_event *event);
2245 * rb_update_event - update event type and data
2246 * @event: the event to update
2247 * @type: the type of event
2248 * @length: the size of the event field in the ring buffer
2250 * Update the type and data fields of the event. The length
2251 * is the actual size that is written to the ring buffer,
2252 * and with this, we can determine what to place into the
2256 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2257 struct ring_buffer_event *event,
2258 struct rb_event_info *info)
2260 unsigned length = info->length;
2261 u64 delta = info->delta;
2263 /* Only a commit updates the timestamp */
2264 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
2268 * If we need to add a timestamp, then we
2269 * add it to the start of the resevered space.
2271 if (unlikely(info->add_timestamp)) {
2272 event = rb_add_time_stamp(event, delta);
2273 length -= RB_LEN_TIME_EXTEND;
2277 event->time_delta = delta;
2278 length -= RB_EVNT_HDR_SIZE;
2279 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2280 event->type_len = 0;
2281 event->array[0] = length;
2283 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2286 static unsigned rb_calculate_event_length(unsigned length)
2288 struct ring_buffer_event event; /* Used only for sizeof array */
2290 /* zero length can cause confusions */
2294 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2295 length += sizeof(event.array[0]);
2297 length += RB_EVNT_HDR_SIZE;
2298 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2301 * In case the time delta is larger than the 27 bits for it
2302 * in the header, we need to add a timestamp. If another
2303 * event comes in when trying to discard this one to increase
2304 * the length, then the timestamp will be added in the allocated
2305 * space of this event. If length is bigger than the size needed
2306 * for the TIME_EXTEND, then padding has to be used. The events
2307 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2308 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2309 * As length is a multiple of 4, we only need to worry if it
2310 * is 12 (RB_LEN_TIME_EXTEND + 4).
2312 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2313 length += RB_ALIGNMENT;
2318 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2319 static inline bool sched_clock_stable(void)
2326 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2327 struct ring_buffer_event *event)
2329 unsigned long new_index, old_index;
2330 struct buffer_page *bpage;
2331 unsigned long index;
2334 new_index = rb_event_index(event);
2335 old_index = new_index + rb_event_ts_length(event);
2336 addr = (unsigned long)event;
2339 bpage = READ_ONCE(cpu_buffer->tail_page);
2341 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2342 unsigned long write_mask =
2343 local_read(&bpage->write) & ~RB_WRITE_MASK;
2344 unsigned long event_length = rb_event_length(event);
2346 * This is on the tail page. It is possible that
2347 * a write could come in and move the tail page
2348 * and write to the next page. That is fine
2349 * because we just shorten what is on this page.
2351 old_index += write_mask;
2352 new_index += write_mask;
2353 index = local_cmpxchg(&bpage->write, old_index, new_index);
2354 if (index == old_index) {
2355 /* update counters */
2356 local_sub(event_length, &cpu_buffer->entries_bytes);
2361 /* could not discard */
2365 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2367 local_inc(&cpu_buffer->committing);
2368 local_inc(&cpu_buffer->commits);
2371 static __always_inline void
2372 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
2374 unsigned long max_count;
2377 * We only race with interrupts and NMIs on this CPU.
2378 * If we own the commit event, then we can commit
2379 * all others that interrupted us, since the interruptions
2380 * are in stack format (they finish before they come
2381 * back to us). This allows us to do a simple loop to
2382 * assign the commit to the tail.
2385 max_count = cpu_buffer->nr_pages * 100;
2387 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
2388 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
2390 if (RB_WARN_ON(cpu_buffer,
2391 rb_is_reader_page(cpu_buffer->tail_page)))
2393 local_set(&cpu_buffer->commit_page->page->commit,
2394 rb_page_write(cpu_buffer->commit_page));
2395 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
2396 /* Only update the write stamp if the page has an event */
2397 if (rb_page_write(cpu_buffer->commit_page))
2398 cpu_buffer->write_stamp =
2399 cpu_buffer->commit_page->page->time_stamp;
2400 /* add barrier to keep gcc from optimizing too much */
2403 while (rb_commit_index(cpu_buffer) !=
2404 rb_page_write(cpu_buffer->commit_page)) {
2406 local_set(&cpu_buffer->commit_page->page->commit,
2407 rb_page_write(cpu_buffer->commit_page));
2408 RB_WARN_ON(cpu_buffer,
2409 local_read(&cpu_buffer->commit_page->page->commit) &
2414 /* again, keep gcc from optimizing */
2418 * If an interrupt came in just after the first while loop
2419 * and pushed the tail page forward, we will be left with
2420 * a dangling commit that will never go forward.
2422 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
2426 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2428 unsigned long commits;
2430 if (RB_WARN_ON(cpu_buffer,
2431 !local_read(&cpu_buffer->committing)))
2435 commits = local_read(&cpu_buffer->commits);
2436 /* synchronize with interrupts */
2438 if (local_read(&cpu_buffer->committing) == 1)
2439 rb_set_commit_to_write(cpu_buffer);
2441 local_dec(&cpu_buffer->committing);
2443 /* synchronize with interrupts */
2447 * Need to account for interrupts coming in between the
2448 * updating of the commit page and the clearing of the
2449 * committing counter.
2451 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2452 !local_read(&cpu_buffer->committing)) {
2453 local_inc(&cpu_buffer->committing);
2458 static inline void rb_event_discard(struct ring_buffer_event *event)
2460 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2461 event = skip_time_extend(event);
2463 /* array[0] holds the actual length for the discarded event */
2464 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2465 event->type_len = RINGBUF_TYPE_PADDING;
2466 /* time delta must be non zero */
2467 if (!event->time_delta)
2468 event->time_delta = 1;
2471 static __always_inline bool
2472 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
2473 struct ring_buffer_event *event)
2475 unsigned long addr = (unsigned long)event;
2476 unsigned long index;
2478 index = rb_event_index(event);
2481 return cpu_buffer->commit_page->page == (void *)addr &&
2482 rb_commit_index(cpu_buffer) == index;
2485 static __always_inline void
2486 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2487 struct ring_buffer_event *event)
2492 * The event first in the commit queue updates the
2495 if (rb_event_is_commit(cpu_buffer, event)) {
2497 * A commit event that is first on a page
2498 * updates the write timestamp with the page stamp
2500 if (!rb_event_index(event))
2501 cpu_buffer->write_stamp =
2502 cpu_buffer->commit_page->page->time_stamp;
2503 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2504 delta = event->array[0];
2506 delta += event->time_delta;
2507 cpu_buffer->write_stamp += delta;
2509 cpu_buffer->write_stamp += event->time_delta;
2513 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2514 struct ring_buffer_event *event)
2516 local_inc(&cpu_buffer->entries);
2517 rb_update_write_stamp(cpu_buffer, event);
2518 rb_end_commit(cpu_buffer);
2521 static __always_inline void
2522 rb_wakeups(struct ring_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
2526 if (buffer->irq_work.waiters_pending) {
2527 buffer->irq_work.waiters_pending = false;
2528 /* irq_work_queue() supplies it's own memory barriers */
2529 irq_work_queue(&buffer->irq_work.work);
2532 if (cpu_buffer->irq_work.waiters_pending) {
2533 cpu_buffer->irq_work.waiters_pending = false;
2534 /* irq_work_queue() supplies it's own memory barriers */
2535 irq_work_queue(&cpu_buffer->irq_work.work);
2538 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
2540 if (!pagebusy && cpu_buffer->irq_work.full_waiters_pending) {
2541 cpu_buffer->irq_work.wakeup_full = true;
2542 cpu_buffer->irq_work.full_waiters_pending = false;
2543 /* irq_work_queue() supplies it's own memory barriers */
2544 irq_work_queue(&cpu_buffer->irq_work.work);
2549 * The lock and unlock are done within a preempt disable section.
2550 * The current_context per_cpu variable can only be modified
2551 * by the current task between lock and unlock. But it can
2552 * be modified more than once via an interrupt. To pass this
2553 * information from the lock to the unlock without having to
2554 * access the 'in_interrupt()' functions again (which do show
2555 * a bit of overhead in something as critical as function tracing,
2556 * we use a bitmask trick.
2558 * bit 1 = NMI context
2559 * bit 2 = IRQ context
2560 * bit 3 = SoftIRQ context
2561 * bit 4 = normal context.
2563 * This works because this is the order of contexts that can
2564 * preempt other contexts. A SoftIRQ never preempts an IRQ
2567 * When the context is determined, the corresponding bit is
2568 * checked and set (if it was set, then a recursion of that context
2571 * On unlock, we need to clear this bit. To do so, just subtract
2572 * 1 from the current_context and AND it to itself.
2576 * 101 & 100 = 100 (clearing bit zero)
2579 * 1010 & 1001 = 1000 (clearing bit 1)
2581 * The least significant bit can be cleared this way, and it
2582 * just so happens that it is the same bit corresponding to
2583 * the current context.
2585 * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
2586 * is set when a recursion is detected at the current context, and if
2587 * the TRANSITION bit is already set, it will fail the recursion.
2588 * This is needed because there's a lag between the changing of
2589 * interrupt context and updating the preempt count. In this case,
2590 * a false positive will be found. To handle this, one extra recursion
2591 * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
2592 * bit is already set, then it is considered a recursion and the function
2593 * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
2595 * On the trace_recursive_unlock(), the TRANSITION bit will be the first
2596 * to be cleared. Even if it wasn't the context that set it. That is,
2597 * if an interrupt comes in while NORMAL bit is set and the ring buffer
2598 * is called before preempt_count() is updated, since the check will
2599 * be on the NORMAL bit, the TRANSITION bit will then be set. If an
2600 * NMI then comes in, it will set the NMI bit, but when the NMI code
2601 * does the trace_recursive_unlock() it will clear the TRANSTION bit
2602 * and leave the NMI bit set. But this is fine, because the interrupt
2603 * code that set the TRANSITION bit will then clear the NMI bit when it
2604 * calls trace_recursive_unlock(). If another NMI comes in, it will
2605 * set the TRANSITION bit and continue.
2607 * Note: The TRANSITION bit only handles a single transition between context.
2610 static __always_inline int
2611 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
2613 unsigned int val = cpu_buffer->current_context;
2616 if (in_interrupt()) {
2622 bit = RB_CTX_SOFTIRQ;
2624 bit = RB_CTX_NORMAL;
2626 if (unlikely(val & (1 << bit))) {
2628 * It is possible that this was called by transitioning
2629 * between interrupt context, and preempt_count() has not
2630 * been updated yet. In this case, use the TRANSITION bit.
2632 bit = RB_CTX_TRANSITION;
2633 if (val & (1 << bit))
2638 cpu_buffer->current_context = val;
2643 static __always_inline void
2644 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
2646 cpu_buffer->current_context &= cpu_buffer->current_context - 1;
2650 * ring_buffer_unlock_commit - commit a reserved
2651 * @buffer: The buffer to commit to
2652 * @event: The event pointer to commit.
2654 * This commits the data to the ring buffer, and releases any locks held.
2656 * Must be paired with ring_buffer_lock_reserve.
2658 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2659 struct ring_buffer_event *event)
2661 struct ring_buffer_per_cpu *cpu_buffer;
2662 int cpu = raw_smp_processor_id();
2664 cpu_buffer = buffer->buffers[cpu];
2666 rb_commit(cpu_buffer, event);
2668 rb_wakeups(buffer, cpu_buffer);
2670 trace_recursive_unlock(cpu_buffer);
2672 preempt_enable_notrace();
2676 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2678 static noinline void
2679 rb_handle_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2680 struct rb_event_info *info)
2682 WARN_ONCE(info->delta > (1ULL << 59),
2683 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2684 (unsigned long long)info->delta,
2685 (unsigned long long)info->ts,
2686 (unsigned long long)cpu_buffer->write_stamp,
2687 sched_clock_stable() ? "" :
2688 "If you just came from a suspend/resume,\n"
2689 "please switch to the trace global clock:\n"
2690 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2691 info->add_timestamp = 1;
2694 static struct ring_buffer_event *
2695 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2696 struct rb_event_info *info)
2698 struct ring_buffer_event *event;
2699 struct buffer_page *tail_page;
2700 unsigned long tail, write;
2703 * If the time delta since the last event is too big to
2704 * hold in the time field of the event, then we append a
2705 * TIME EXTEND event ahead of the data event.
2707 if (unlikely(info->add_timestamp))
2708 info->length += RB_LEN_TIME_EXTEND;
2710 /* Don't let the compiler play games with cpu_buffer->tail_page */
2711 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
2712 write = local_add_return(info->length, &tail_page->write);
2714 /* set write to only the index of the write */
2715 write &= RB_WRITE_MASK;
2716 tail = write - info->length;
2719 * If this is the first commit on the page, then it has the same
2720 * timestamp as the page itself.
2725 /* See if we shot pass the end of this buffer page */
2726 if (unlikely(write > BUF_PAGE_SIZE))
2727 return rb_move_tail(cpu_buffer, tail, info);
2729 /* We reserved something on the buffer */
2731 event = __rb_page_index(tail_page, tail);
2732 rb_update_event(cpu_buffer, event, info);
2734 local_inc(&tail_page->entries);
2737 * If this is the first commit on the page, then update
2741 tail_page->page->time_stamp = info->ts;
2743 /* account for these added bytes */
2744 local_add(info->length, &cpu_buffer->entries_bytes);
2749 static __always_inline struct ring_buffer_event *
2750 rb_reserve_next_event(struct ring_buffer *buffer,
2751 struct ring_buffer_per_cpu *cpu_buffer,
2752 unsigned long length)
2754 struct ring_buffer_event *event;
2755 struct rb_event_info info;
2759 rb_start_commit(cpu_buffer);
2761 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2763 * Due to the ability to swap a cpu buffer from a buffer
2764 * it is possible it was swapped before we committed.
2765 * (committing stops a swap). We check for it here and
2766 * if it happened, we have to fail the write.
2769 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2770 local_dec(&cpu_buffer->committing);
2771 local_dec(&cpu_buffer->commits);
2776 info.length = rb_calculate_event_length(length);
2778 info.add_timestamp = 0;
2782 * We allow for interrupts to reenter here and do a trace.
2783 * If one does, it will cause this original code to loop
2784 * back here. Even with heavy interrupts happening, this
2785 * should only happen a few times in a row. If this happens
2786 * 1000 times in a row, there must be either an interrupt
2787 * storm or we have something buggy.
2790 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2793 info.ts = rb_time_stamp(cpu_buffer->buffer);
2794 diff = info.ts - cpu_buffer->write_stamp;
2796 /* make sure this diff is calculated here */
2799 /* Did the write stamp get updated already? */
2800 if (likely(info.ts >= cpu_buffer->write_stamp)) {
2802 if (unlikely(test_time_stamp(info.delta)))
2803 rb_handle_timestamp(cpu_buffer, &info);
2806 event = __rb_reserve_next(cpu_buffer, &info);
2808 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
2809 if (info.add_timestamp)
2810 info.length -= RB_LEN_TIME_EXTEND;
2820 rb_end_commit(cpu_buffer);
2825 * ring_buffer_lock_reserve - reserve a part of the buffer
2826 * @buffer: the ring buffer to reserve from
2827 * @length: the length of the data to reserve (excluding event header)
2829 * Returns a reseverd event on the ring buffer to copy directly to.
2830 * The user of this interface will need to get the body to write into
2831 * and can use the ring_buffer_event_data() interface.
2833 * The length is the length of the data needed, not the event length
2834 * which also includes the event header.
2836 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2837 * If NULL is returned, then nothing has been allocated or locked.
2839 struct ring_buffer_event *
2840 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2842 struct ring_buffer_per_cpu *cpu_buffer;
2843 struct ring_buffer_event *event;
2846 /* If we are tracing schedule, we don't want to recurse */
2847 preempt_disable_notrace();
2849 if (unlikely(atomic_read(&buffer->record_disabled)))
2852 cpu = raw_smp_processor_id();
2854 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
2857 cpu_buffer = buffer->buffers[cpu];
2859 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
2862 if (unlikely(length > BUF_MAX_DATA_SIZE))
2865 if (unlikely(trace_recursive_lock(cpu_buffer)))
2868 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2875 trace_recursive_unlock(cpu_buffer);
2877 preempt_enable_notrace();
2880 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2883 * Decrement the entries to the page that an event is on.
2884 * The event does not even need to exist, only the pointer
2885 * to the page it is on. This may only be called before the commit
2889 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2890 struct ring_buffer_event *event)
2892 unsigned long addr = (unsigned long)event;
2893 struct buffer_page *bpage = cpu_buffer->commit_page;
2894 struct buffer_page *start;
2898 /* Do the likely case first */
2899 if (likely(bpage->page == (void *)addr)) {
2900 local_dec(&bpage->entries);
2905 * Because the commit page may be on the reader page we
2906 * start with the next page and check the end loop there.
2908 rb_inc_page(cpu_buffer, &bpage);
2911 if (bpage->page == (void *)addr) {
2912 local_dec(&bpage->entries);
2915 rb_inc_page(cpu_buffer, &bpage);
2916 } while (bpage != start);
2918 /* commit not part of this buffer?? */
2919 RB_WARN_ON(cpu_buffer, 1);
2923 * ring_buffer_commit_discard - discard an event that has not been committed
2924 * @buffer: the ring buffer
2925 * @event: non committed event to discard
2927 * Sometimes an event that is in the ring buffer needs to be ignored.
2928 * This function lets the user discard an event in the ring buffer
2929 * and then that event will not be read later.
2931 * This function only works if it is called before the the item has been
2932 * committed. It will try to free the event from the ring buffer
2933 * if another event has not been added behind it.
2935 * If another event has been added behind it, it will set the event
2936 * up as discarded, and perform the commit.
2938 * If this function is called, do not call ring_buffer_unlock_commit on
2941 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2942 struct ring_buffer_event *event)
2944 struct ring_buffer_per_cpu *cpu_buffer;
2947 /* The event is discarded regardless */
2948 rb_event_discard(event);
2950 cpu = smp_processor_id();
2951 cpu_buffer = buffer->buffers[cpu];
2954 * This must only be called if the event has not been
2955 * committed yet. Thus we can assume that preemption
2956 * is still disabled.
2958 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2960 rb_decrement_entry(cpu_buffer, event);
2961 if (rb_try_to_discard(cpu_buffer, event))
2965 * The commit is still visible by the reader, so we
2966 * must still update the timestamp.
2968 rb_update_write_stamp(cpu_buffer, event);
2970 rb_end_commit(cpu_buffer);
2972 trace_recursive_unlock(cpu_buffer);
2974 preempt_enable_notrace();
2977 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2980 * ring_buffer_write - write data to the buffer without reserving
2981 * @buffer: The ring buffer to write to.
2982 * @length: The length of the data being written (excluding the event header)
2983 * @data: The data to write to the buffer.
2985 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2986 * one function. If you already have the data to write to the buffer, it
2987 * may be easier to simply call this function.
2989 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2990 * and not the length of the event which would hold the header.
2992 int ring_buffer_write(struct ring_buffer *buffer,
2993 unsigned long length,
2996 struct ring_buffer_per_cpu *cpu_buffer;
2997 struct ring_buffer_event *event;
3002 preempt_disable_notrace();
3004 if (atomic_read(&buffer->record_disabled))
3007 cpu = raw_smp_processor_id();
3009 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3012 cpu_buffer = buffer->buffers[cpu];
3014 if (atomic_read(&cpu_buffer->record_disabled))
3017 if (length > BUF_MAX_DATA_SIZE)
3020 if (unlikely(trace_recursive_lock(cpu_buffer)))
3023 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3027 body = rb_event_data(event);
3029 memcpy(body, data, length);
3031 rb_commit(cpu_buffer, event);
3033 rb_wakeups(buffer, cpu_buffer);
3038 trace_recursive_unlock(cpu_buffer);
3041 preempt_enable_notrace();
3045 EXPORT_SYMBOL_GPL(ring_buffer_write);
3047 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3049 struct buffer_page *reader = cpu_buffer->reader_page;
3050 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3051 struct buffer_page *commit = cpu_buffer->commit_page;
3053 /* In case of error, head will be NULL */
3054 if (unlikely(!head))
3057 /* Reader should exhaust content in reader page */
3058 if (reader->read != rb_page_commit(reader))
3062 * If writers are committing on the reader page, knowing all
3063 * committed content has been read, the ring buffer is empty.
3065 if (commit == reader)
3069 * If writers are committing on a page other than reader page
3070 * and head page, there should always be content to read.
3076 * Writers are committing on the head page, we just need
3077 * to care about there're committed data, and the reader will
3078 * swap reader page with head page when it is to read data.
3080 return rb_page_commit(commit) == 0;
3084 * ring_buffer_record_disable - stop all writes into the buffer
3085 * @buffer: The ring buffer to stop writes to.
3087 * This prevents all writes to the buffer. Any attempt to write
3088 * to the buffer after this will fail and return NULL.
3090 * The caller should call synchronize_sched() after this.
3092 void ring_buffer_record_disable(struct ring_buffer *buffer)
3094 atomic_inc(&buffer->record_disabled);
3096 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3099 * ring_buffer_record_enable - enable writes to the buffer
3100 * @buffer: The ring buffer to enable writes
3102 * Note, multiple disables will need the same number of enables
3103 * to truly enable the writing (much like preempt_disable).
3105 void ring_buffer_record_enable(struct ring_buffer *buffer)
3107 atomic_dec(&buffer->record_disabled);
3109 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3112 * ring_buffer_record_off - stop all writes into the buffer
3113 * @buffer: The ring buffer to stop writes to.
3115 * This prevents all writes to the buffer. Any attempt to write
3116 * to the buffer after this will fail and return NULL.
3118 * This is different than ring_buffer_record_disable() as
3119 * it works like an on/off switch, where as the disable() version
3120 * must be paired with a enable().
3122 void ring_buffer_record_off(struct ring_buffer *buffer)
3125 unsigned int new_rd;
3128 rd = atomic_read(&buffer->record_disabled);
3129 new_rd = rd | RB_BUFFER_OFF;
3130 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3132 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
3135 * ring_buffer_record_on - restart writes into the buffer
3136 * @buffer: The ring buffer to start writes to.
3138 * This enables all writes to the buffer that was disabled by
3139 * ring_buffer_record_off().
3141 * This is different than ring_buffer_record_enable() as
3142 * it works like an on/off switch, where as the enable() version
3143 * must be paired with a disable().
3145 void ring_buffer_record_on(struct ring_buffer *buffer)
3148 unsigned int new_rd;
3151 rd = atomic_read(&buffer->record_disabled);
3152 new_rd = rd & ~RB_BUFFER_OFF;
3153 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
3155 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
3158 * ring_buffer_record_is_on - return true if the ring buffer can write
3159 * @buffer: The ring buffer to see if write is enabled
3161 * Returns true if the ring buffer is in a state that it accepts writes.
3163 int ring_buffer_record_is_on(struct ring_buffer *buffer)
3165 return !atomic_read(&buffer->record_disabled);
3169 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
3170 * @buffer: The ring buffer to see if write is set enabled
3172 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
3173 * Note that this does NOT mean it is in a writable state.
3175 * It may return true when the ring buffer has been disabled by
3176 * ring_buffer_record_disable(), as that is a temporary disabling of
3179 int ring_buffer_record_is_set_on(struct ring_buffer *buffer)
3181 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
3185 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
3186 * @buffer: The ring buffer to stop writes to.
3187 * @cpu: The CPU buffer to stop
3189 * This prevents all writes to the buffer. Any attempt to write
3190 * to the buffer after this will fail and return NULL.
3192 * The caller should call synchronize_sched() after this.
3194 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
3196 struct ring_buffer_per_cpu *cpu_buffer;
3198 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3201 cpu_buffer = buffer->buffers[cpu];
3202 atomic_inc(&cpu_buffer->record_disabled);
3204 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
3207 * ring_buffer_record_enable_cpu - enable writes to the buffer
3208 * @buffer: The ring buffer to enable writes
3209 * @cpu: The CPU to enable.
3211 * Note, multiple disables will need the same number of enables
3212 * to truly enable the writing (much like preempt_disable).
3214 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
3216 struct ring_buffer_per_cpu *cpu_buffer;
3218 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3221 cpu_buffer = buffer->buffers[cpu];
3222 atomic_dec(&cpu_buffer->record_disabled);
3224 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
3227 * The total entries in the ring buffer is the running counter
3228 * of entries entered into the ring buffer, minus the sum of
3229 * the entries read from the ring buffer and the number of
3230 * entries that were overwritten.
3232 static inline unsigned long
3233 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
3235 return local_read(&cpu_buffer->entries) -
3236 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
3240 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
3241 * @buffer: The ring buffer
3242 * @cpu: The per CPU buffer to read from.
3244 u64 ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
3246 unsigned long flags;
3247 struct ring_buffer_per_cpu *cpu_buffer;
3248 struct buffer_page *bpage;
3251 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3254 cpu_buffer = buffer->buffers[cpu];
3255 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3257 * if the tail is on reader_page, oldest time stamp is on the reader
3260 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
3261 bpage = cpu_buffer->reader_page;
3263 bpage = rb_set_head_page(cpu_buffer);
3265 ret = bpage->page->time_stamp;
3266 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3270 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
3273 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
3274 * @buffer: The ring buffer
3275 * @cpu: The per CPU buffer to read from.
3277 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
3279 struct ring_buffer_per_cpu *cpu_buffer;
3282 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3285 cpu_buffer = buffer->buffers[cpu];
3286 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
3290 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
3293 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
3294 * @buffer: The ring buffer
3295 * @cpu: The per CPU buffer to get the entries from.
3297 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
3299 struct ring_buffer_per_cpu *cpu_buffer;
3301 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3304 cpu_buffer = buffer->buffers[cpu];
3306 return rb_num_of_entries(cpu_buffer);
3308 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3311 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
3312 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
3313 * @buffer: The ring buffer
3314 * @cpu: The per CPU buffer to get the number of overruns from
3316 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3318 struct ring_buffer_per_cpu *cpu_buffer;
3321 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3324 cpu_buffer = buffer->buffers[cpu];
3325 ret = local_read(&cpu_buffer->overrun);
3329 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3332 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
3333 * commits failing due to the buffer wrapping around while there are uncommitted
3334 * events, such as during an interrupt storm.
3335 * @buffer: The ring buffer
3336 * @cpu: The per CPU buffer to get the number of overruns from
3339 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3341 struct ring_buffer_per_cpu *cpu_buffer;
3344 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3347 cpu_buffer = buffer->buffers[cpu];
3348 ret = local_read(&cpu_buffer->commit_overrun);
3352 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3355 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
3356 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
3357 * @buffer: The ring buffer
3358 * @cpu: The per CPU buffer to get the number of overruns from
3361 ring_buffer_dropped_events_cpu(struct ring_buffer *buffer, int cpu)
3363 struct ring_buffer_per_cpu *cpu_buffer;
3366 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3369 cpu_buffer = buffer->buffers[cpu];
3370 ret = local_read(&cpu_buffer->dropped_events);
3374 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
3377 * ring_buffer_read_events_cpu - get the number of events successfully read
3378 * @buffer: The ring buffer
3379 * @cpu: The per CPU buffer to get the number of events read
3382 ring_buffer_read_events_cpu(struct ring_buffer *buffer, int cpu)
3384 struct ring_buffer_per_cpu *cpu_buffer;
3386 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3389 cpu_buffer = buffer->buffers[cpu];
3390 return cpu_buffer->read;
3392 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
3395 * ring_buffer_entries - get the number of entries in a buffer
3396 * @buffer: The ring buffer
3398 * Returns the total number of entries in the ring buffer
3401 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3403 struct ring_buffer_per_cpu *cpu_buffer;
3404 unsigned long entries = 0;
3407 /* if you care about this being correct, lock the buffer */
3408 for_each_buffer_cpu(buffer, cpu) {
3409 cpu_buffer = buffer->buffers[cpu];
3410 entries += rb_num_of_entries(cpu_buffer);
3415 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3418 * ring_buffer_overruns - get the number of overruns in buffer
3419 * @buffer: The ring buffer
3421 * Returns the total number of overruns in the ring buffer
3424 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3426 struct ring_buffer_per_cpu *cpu_buffer;
3427 unsigned long overruns = 0;
3430 /* if you care about this being correct, lock the buffer */
3431 for_each_buffer_cpu(buffer, cpu) {
3432 cpu_buffer = buffer->buffers[cpu];
3433 overruns += local_read(&cpu_buffer->overrun);
3438 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3440 static void rb_iter_reset(struct ring_buffer_iter *iter)
3442 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3444 /* Iterator usage is expected to have record disabled */
3445 iter->head_page = cpu_buffer->reader_page;
3446 iter->head = cpu_buffer->reader_page->read;
3448 iter->cache_reader_page = iter->head_page;
3449 iter->cache_read = cpu_buffer->read;
3452 iter->read_stamp = cpu_buffer->read_stamp;
3454 iter->read_stamp = iter->head_page->page->time_stamp;
3458 * ring_buffer_iter_reset - reset an iterator
3459 * @iter: The iterator to reset
3461 * Resets the iterator, so that it will start from the beginning
3464 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3466 struct ring_buffer_per_cpu *cpu_buffer;
3467 unsigned long flags;
3472 cpu_buffer = iter->cpu_buffer;
3474 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3475 rb_iter_reset(iter);
3476 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3478 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3481 * ring_buffer_iter_empty - check if an iterator has no more to read
3482 * @iter: The iterator to check
3484 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3486 struct ring_buffer_per_cpu *cpu_buffer;
3487 struct buffer_page *reader;
3488 struct buffer_page *head_page;
3489 struct buffer_page *commit_page;
3492 cpu_buffer = iter->cpu_buffer;
3494 /* Remember, trace recording is off when iterator is in use */
3495 reader = cpu_buffer->reader_page;
3496 head_page = cpu_buffer->head_page;
3497 commit_page = cpu_buffer->commit_page;
3498 commit = rb_page_commit(commit_page);
3500 return ((iter->head_page == commit_page && iter->head == commit) ||
3501 (iter->head_page == reader && commit_page == head_page &&
3502 head_page->read == commit &&
3503 iter->head == rb_page_commit(cpu_buffer->reader_page)));
3505 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3508 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3509 struct ring_buffer_event *event)
3513 switch (event->type_len) {
3514 case RINGBUF_TYPE_PADDING:
3517 case RINGBUF_TYPE_TIME_EXTEND:
3518 delta = event->array[0];
3520 delta += event->time_delta;
3521 cpu_buffer->read_stamp += delta;
3524 case RINGBUF_TYPE_TIME_STAMP:
3525 /* FIXME: not implemented */
3528 case RINGBUF_TYPE_DATA:
3529 cpu_buffer->read_stamp += event->time_delta;
3539 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3540 struct ring_buffer_event *event)
3544 switch (event->type_len) {
3545 case RINGBUF_TYPE_PADDING:
3548 case RINGBUF_TYPE_TIME_EXTEND:
3549 delta = event->array[0];
3551 delta += event->time_delta;
3552 iter->read_stamp += delta;
3555 case RINGBUF_TYPE_TIME_STAMP:
3556 /* FIXME: not implemented */
3559 case RINGBUF_TYPE_DATA:
3560 iter->read_stamp += event->time_delta;
3569 static struct buffer_page *
3570 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3572 struct buffer_page *reader = NULL;
3573 unsigned long overwrite;
3574 unsigned long flags;
3578 local_irq_save(flags);
3579 arch_spin_lock(&cpu_buffer->lock);
3583 * This should normally only loop twice. But because the
3584 * start of the reader inserts an empty page, it causes
3585 * a case where we will loop three times. There should be no
3586 * reason to loop four times (that I know of).
3588 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3593 reader = cpu_buffer->reader_page;
3595 /* If there's more to read, return this page */
3596 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3599 /* Never should we have an index greater than the size */
3600 if (RB_WARN_ON(cpu_buffer,
3601 cpu_buffer->reader_page->read > rb_page_size(reader)))
3604 /* check if we caught up to the tail */
3606 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3609 /* Don't bother swapping if the ring buffer is empty */
3610 if (rb_num_of_entries(cpu_buffer) == 0)
3614 * Reset the reader page to size zero.
3616 local_set(&cpu_buffer->reader_page->write, 0);
3617 local_set(&cpu_buffer->reader_page->entries, 0);
3618 local_set(&cpu_buffer->reader_page->page->commit, 0);
3619 cpu_buffer->reader_page->real_end = 0;
3623 * Splice the empty reader page into the list around the head.
3625 reader = rb_set_head_page(cpu_buffer);
3628 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3629 cpu_buffer->reader_page->list.prev = reader->list.prev;
3632 * cpu_buffer->pages just needs to point to the buffer, it
3633 * has no specific buffer page to point to. Lets move it out
3634 * of our way so we don't accidentally swap it.
3636 cpu_buffer->pages = reader->list.prev;
3638 /* The reader page will be pointing to the new head */
3639 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3642 * We want to make sure we read the overruns after we set up our
3643 * pointers to the next object. The writer side does a
3644 * cmpxchg to cross pages which acts as the mb on the writer
3645 * side. Note, the reader will constantly fail the swap
3646 * while the writer is updating the pointers, so this
3647 * guarantees that the overwrite recorded here is the one we
3648 * want to compare with the last_overrun.
3651 overwrite = local_read(&(cpu_buffer->overrun));
3654 * Here's the tricky part.
3656 * We need to move the pointer past the header page.
3657 * But we can only do that if a writer is not currently
3658 * moving it. The page before the header page has the
3659 * flag bit '1' set if it is pointing to the page we want.
3660 * but if the writer is in the process of moving it
3661 * than it will be '2' or already moved '0'.
3664 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3667 * If we did not convert it, then we must try again.
3673 * Yeah! We succeeded in replacing the page.
3675 * Now make the new head point back to the reader page.
3677 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3678 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3680 /* Finally update the reader page to the new head */
3681 cpu_buffer->reader_page = reader;
3682 cpu_buffer->reader_page->read = 0;
3684 if (overwrite != cpu_buffer->last_overrun) {
3685 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3686 cpu_buffer->last_overrun = overwrite;
3692 /* Update the read_stamp on the first event */
3693 if (reader && reader->read == 0)
3694 cpu_buffer->read_stamp = reader->page->time_stamp;
3696 arch_spin_unlock(&cpu_buffer->lock);
3697 local_irq_restore(flags);
3702 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3704 struct ring_buffer_event *event;
3705 struct buffer_page *reader;
3708 reader = rb_get_reader_page(cpu_buffer);
3710 /* This function should not be called when buffer is empty */
3711 if (RB_WARN_ON(cpu_buffer, !reader))
3714 event = rb_reader_event(cpu_buffer);
3716 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3719 rb_update_read_stamp(cpu_buffer, event);
3721 length = rb_event_length(event);
3722 cpu_buffer->reader_page->read += length;
3725 static void rb_advance_iter(struct ring_buffer_iter *iter)
3727 struct ring_buffer_per_cpu *cpu_buffer;
3728 struct ring_buffer_event *event;
3731 cpu_buffer = iter->cpu_buffer;
3734 * Check if we are at the end of the buffer.
3736 if (iter->head >= rb_page_size(iter->head_page)) {
3737 /* discarded commits can make the page empty */
3738 if (iter->head_page == cpu_buffer->commit_page)
3744 event = rb_iter_head_event(iter);
3746 length = rb_event_length(event);
3749 * This should not be called to advance the header if we are
3750 * at the tail of the buffer.
3752 if (RB_WARN_ON(cpu_buffer,
3753 (iter->head_page == cpu_buffer->commit_page) &&
3754 (iter->head + length > rb_commit_index(cpu_buffer))))
3757 rb_update_iter_read_stamp(iter, event);
3759 iter->head += length;
3761 /* check for end of page padding */
3762 if ((iter->head >= rb_page_size(iter->head_page)) &&
3763 (iter->head_page != cpu_buffer->commit_page))
3767 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3769 return cpu_buffer->lost_events;
3772 static struct ring_buffer_event *
3773 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3774 unsigned long *lost_events)
3776 struct ring_buffer_event *event;
3777 struct buffer_page *reader;
3782 * We repeat when a time extend is encountered.
3783 * Since the time extend is always attached to a data event,
3784 * we should never loop more than once.
3785 * (We never hit the following condition more than twice).
3787 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3790 reader = rb_get_reader_page(cpu_buffer);
3794 event = rb_reader_event(cpu_buffer);
3796 switch (event->type_len) {
3797 case RINGBUF_TYPE_PADDING:
3798 if (rb_null_event(event))
3799 RB_WARN_ON(cpu_buffer, 1);
3801 * Because the writer could be discarding every
3802 * event it creates (which would probably be bad)
3803 * if we were to go back to "again" then we may never
3804 * catch up, and will trigger the warn on, or lock
3805 * the box. Return the padding, and we will release
3806 * the current locks, and try again.
3810 case RINGBUF_TYPE_TIME_EXTEND:
3811 /* Internal data, OK to advance */
3812 rb_advance_reader(cpu_buffer);
3815 case RINGBUF_TYPE_TIME_STAMP:
3816 /* FIXME: not implemented */
3817 rb_advance_reader(cpu_buffer);
3820 case RINGBUF_TYPE_DATA:
3822 *ts = cpu_buffer->read_stamp + event->time_delta;
3823 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3824 cpu_buffer->cpu, ts);
3827 *lost_events = rb_lost_events(cpu_buffer);
3836 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3838 static struct ring_buffer_event *
3839 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3841 struct ring_buffer *buffer;
3842 struct ring_buffer_per_cpu *cpu_buffer;
3843 struct ring_buffer_event *event;
3846 cpu_buffer = iter->cpu_buffer;
3847 buffer = cpu_buffer->buffer;
3850 * Check if someone performed a consuming read to
3851 * the buffer. A consuming read invalidates the iterator
3852 * and we need to reset the iterator in this case.
3854 if (unlikely(iter->cache_read != cpu_buffer->read ||
3855 iter->cache_reader_page != cpu_buffer->reader_page))
3856 rb_iter_reset(iter);
3859 if (ring_buffer_iter_empty(iter))
3863 * We repeat when a time extend is encountered or we hit
3864 * the end of the page. Since the time extend is always attached
3865 * to a data event, we should never loop more than three times.
3866 * Once for going to next page, once on time extend, and
3867 * finally once to get the event.
3868 * (We never hit the following condition more than thrice).
3870 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3))
3873 if (rb_per_cpu_empty(cpu_buffer))
3876 if (iter->head >= rb_page_size(iter->head_page)) {
3881 event = rb_iter_head_event(iter);
3883 switch (event->type_len) {
3884 case RINGBUF_TYPE_PADDING:
3885 if (rb_null_event(event)) {
3889 rb_advance_iter(iter);
3892 case RINGBUF_TYPE_TIME_EXTEND:
3893 /* Internal data, OK to advance */
3894 rb_advance_iter(iter);
3897 case RINGBUF_TYPE_TIME_STAMP:
3898 /* FIXME: not implemented */
3899 rb_advance_iter(iter);
3902 case RINGBUF_TYPE_DATA:
3904 *ts = iter->read_stamp + event->time_delta;
3905 ring_buffer_normalize_time_stamp(buffer,
3906 cpu_buffer->cpu, ts);
3916 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3918 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
3920 if (likely(!in_nmi())) {
3921 raw_spin_lock(&cpu_buffer->reader_lock);
3926 * If an NMI die dumps out the content of the ring buffer
3927 * trylock must be used to prevent a deadlock if the NMI
3928 * preempted a task that holds the ring buffer locks. If
3929 * we get the lock then all is fine, if not, then continue
3930 * to do the read, but this can corrupt the ring buffer,
3931 * so it must be permanently disabled from future writes.
3932 * Reading from NMI is a oneshot deal.
3934 if (raw_spin_trylock(&cpu_buffer->reader_lock))
3937 /* Continue without locking, but disable the ring buffer */
3938 atomic_inc(&cpu_buffer->record_disabled);
3943 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
3946 raw_spin_unlock(&cpu_buffer->reader_lock);
3951 * ring_buffer_peek - peek at the next event to be read
3952 * @buffer: The ring buffer to read
3953 * @cpu: The cpu to peak at
3954 * @ts: The timestamp counter of this event.
3955 * @lost_events: a variable to store if events were lost (may be NULL)
3957 * This will return the event that will be read next, but does
3958 * not consume the data.
3960 struct ring_buffer_event *
3961 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3962 unsigned long *lost_events)
3964 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3965 struct ring_buffer_event *event;
3966 unsigned long flags;
3969 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3973 local_irq_save(flags);
3974 dolock = rb_reader_lock(cpu_buffer);
3975 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3976 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3977 rb_advance_reader(cpu_buffer);
3978 rb_reader_unlock(cpu_buffer, dolock);
3979 local_irq_restore(flags);
3981 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3988 * ring_buffer_iter_peek - peek at the next event to be read
3989 * @iter: The ring buffer iterator
3990 * @ts: The timestamp counter of this event.
3992 * This will return the event that will be read next, but does
3993 * not increment the iterator.
3995 struct ring_buffer_event *
3996 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3998 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3999 struct ring_buffer_event *event;
4000 unsigned long flags;
4003 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4004 event = rb_iter_peek(iter, ts);
4005 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4007 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4014 * ring_buffer_consume - return an event and consume it
4015 * @buffer: The ring buffer to get the next event from
4016 * @cpu: the cpu to read the buffer from
4017 * @ts: a variable to store the timestamp (may be NULL)
4018 * @lost_events: a variable to store if events were lost (may be NULL)
4020 * Returns the next event in the ring buffer, and that event is consumed.
4021 * Meaning, that sequential reads will keep returning a different event,
4022 * and eventually empty the ring buffer if the producer is slower.
4024 struct ring_buffer_event *
4025 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
4026 unsigned long *lost_events)
4028 struct ring_buffer_per_cpu *cpu_buffer;
4029 struct ring_buffer_event *event = NULL;
4030 unsigned long flags;
4034 /* might be called in atomic */
4037 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4040 cpu_buffer = buffer->buffers[cpu];
4041 local_irq_save(flags);
4042 dolock = rb_reader_lock(cpu_buffer);
4044 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4046 cpu_buffer->lost_events = 0;
4047 rb_advance_reader(cpu_buffer);
4050 rb_reader_unlock(cpu_buffer, dolock);
4051 local_irq_restore(flags);
4056 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4061 EXPORT_SYMBOL_GPL(ring_buffer_consume);
4064 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
4065 * @buffer: The ring buffer to read from
4066 * @cpu: The cpu buffer to iterate over
4067 * @flags: gfp flags to use for memory allocation
4069 * This performs the initial preparations necessary to iterate
4070 * through the buffer. Memory is allocated, buffer recording
4071 * is disabled, and the iterator pointer is returned to the caller.
4073 * Disabling buffer recordng prevents the reading from being
4074 * corrupted. This is not a consuming read, so a producer is not
4077 * After a sequence of ring_buffer_read_prepare calls, the user is
4078 * expected to make at least one call to ring_buffer_read_prepare_sync.
4079 * Afterwards, ring_buffer_read_start is invoked to get things going
4082 * This overall must be paired with ring_buffer_read_finish.
4084 struct ring_buffer_iter *
4085 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu, gfp_t flags)
4087 struct ring_buffer_per_cpu *cpu_buffer;
4088 struct ring_buffer_iter *iter;
4090 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4093 iter = kmalloc(sizeof(*iter), flags);
4097 cpu_buffer = buffer->buffers[cpu];
4099 iter->cpu_buffer = cpu_buffer;
4101 atomic_inc(&buffer->resize_disabled);
4102 atomic_inc(&cpu_buffer->record_disabled);
4106 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
4109 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
4111 * All previously invoked ring_buffer_read_prepare calls to prepare
4112 * iterators will be synchronized. Afterwards, read_buffer_read_start
4113 * calls on those iterators are allowed.
4116 ring_buffer_read_prepare_sync(void)
4118 synchronize_sched();
4120 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
4123 * ring_buffer_read_start - start a non consuming read of the buffer
4124 * @iter: The iterator returned by ring_buffer_read_prepare
4126 * This finalizes the startup of an iteration through the buffer.
4127 * The iterator comes from a call to ring_buffer_read_prepare and
4128 * an intervening ring_buffer_read_prepare_sync must have been
4131 * Must be paired with ring_buffer_read_finish.
4134 ring_buffer_read_start(struct ring_buffer_iter *iter)
4136 struct ring_buffer_per_cpu *cpu_buffer;
4137 unsigned long flags;
4142 cpu_buffer = iter->cpu_buffer;
4144 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4145 arch_spin_lock(&cpu_buffer->lock);
4146 rb_iter_reset(iter);
4147 arch_spin_unlock(&cpu_buffer->lock);
4148 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4150 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
4153 * ring_buffer_read_finish - finish reading the iterator of the buffer
4154 * @iter: The iterator retrieved by ring_buffer_start
4156 * This re-enables the recording to the buffer, and frees the
4160 ring_buffer_read_finish(struct ring_buffer_iter *iter)
4162 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4163 unsigned long flags;
4166 * Ring buffer is disabled from recording, here's a good place
4167 * to check the integrity of the ring buffer.
4168 * Must prevent readers from trying to read, as the check
4169 * clears the HEAD page and readers require it.
4171 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4172 rb_check_pages(cpu_buffer);
4173 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4175 atomic_dec(&cpu_buffer->record_disabled);
4176 atomic_dec(&cpu_buffer->buffer->resize_disabled);
4179 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
4182 * ring_buffer_read - read the next item in the ring buffer by the iterator
4183 * @iter: The ring buffer iterator
4184 * @ts: The time stamp of the event read.
4186 * This reads the next event in the ring buffer and increments the iterator.
4188 struct ring_buffer_event *
4189 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
4191 struct ring_buffer_event *event;
4192 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4193 unsigned long flags;
4195 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4197 event = rb_iter_peek(iter, ts);
4201 if (event->type_len == RINGBUF_TYPE_PADDING)
4204 rb_advance_iter(iter);
4206 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4210 EXPORT_SYMBOL_GPL(ring_buffer_read);
4213 * ring_buffer_size - return the size of the ring buffer (in bytes)
4214 * @buffer: The ring buffer.
4216 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
4219 * Earlier, this method returned
4220 * BUF_PAGE_SIZE * buffer->nr_pages
4221 * Since the nr_pages field is now removed, we have converted this to
4222 * return the per cpu buffer value.
4224 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4227 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
4229 EXPORT_SYMBOL_GPL(ring_buffer_size);
4232 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
4234 rb_head_page_deactivate(cpu_buffer);
4236 cpu_buffer->head_page
4237 = list_entry(cpu_buffer->pages, struct buffer_page, list);
4238 local_set(&cpu_buffer->head_page->write, 0);
4239 local_set(&cpu_buffer->head_page->entries, 0);
4240 local_set(&cpu_buffer->head_page->page->commit, 0);
4242 cpu_buffer->head_page->read = 0;
4244 cpu_buffer->tail_page = cpu_buffer->head_page;
4245 cpu_buffer->commit_page = cpu_buffer->head_page;
4247 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
4248 INIT_LIST_HEAD(&cpu_buffer->new_pages);
4249 local_set(&cpu_buffer->reader_page->write, 0);
4250 local_set(&cpu_buffer->reader_page->entries, 0);
4251 local_set(&cpu_buffer->reader_page->page->commit, 0);
4252 cpu_buffer->reader_page->read = 0;
4254 local_set(&cpu_buffer->entries_bytes, 0);
4255 local_set(&cpu_buffer->overrun, 0);
4256 local_set(&cpu_buffer->commit_overrun, 0);
4257 local_set(&cpu_buffer->dropped_events, 0);
4258 local_set(&cpu_buffer->entries, 0);
4259 local_set(&cpu_buffer->committing, 0);
4260 local_set(&cpu_buffer->commits, 0);
4261 cpu_buffer->read = 0;
4262 cpu_buffer->read_bytes = 0;
4264 cpu_buffer->write_stamp = 0;
4265 cpu_buffer->read_stamp = 0;
4267 cpu_buffer->lost_events = 0;
4268 cpu_buffer->last_overrun = 0;
4270 rb_head_page_activate(cpu_buffer);
4274 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
4275 * @buffer: The ring buffer to reset a per cpu buffer of
4276 * @cpu: The CPU buffer to be reset
4278 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
4280 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4281 unsigned long flags;
4283 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4285 /* prevent another thread from changing buffer sizes */
4286 mutex_lock(&buffer->mutex);
4288 atomic_inc(&buffer->resize_disabled);
4289 atomic_inc(&cpu_buffer->record_disabled);
4291 /* Make sure all commits have finished */
4292 synchronize_sched();
4294 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4296 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
4299 arch_spin_lock(&cpu_buffer->lock);
4301 rb_reset_cpu(cpu_buffer);
4303 arch_spin_unlock(&cpu_buffer->lock);
4306 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4308 atomic_dec(&cpu_buffer->record_disabled);
4309 atomic_dec(&buffer->resize_disabled);
4311 mutex_unlock(&buffer->mutex);
4313 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
4316 * ring_buffer_reset - reset a ring buffer
4317 * @buffer: The ring buffer to reset all cpu buffers
4319 void ring_buffer_reset(struct ring_buffer *buffer)
4323 for_each_buffer_cpu(buffer, cpu)
4324 ring_buffer_reset_cpu(buffer, cpu);
4326 EXPORT_SYMBOL_GPL(ring_buffer_reset);
4329 * rind_buffer_empty - is the ring buffer empty?
4330 * @buffer: The ring buffer to test
4332 bool ring_buffer_empty(struct ring_buffer *buffer)
4334 struct ring_buffer_per_cpu *cpu_buffer;
4335 unsigned long flags;
4340 /* yes this is racy, but if you don't like the race, lock the buffer */
4341 for_each_buffer_cpu(buffer, cpu) {
4342 cpu_buffer = buffer->buffers[cpu];
4343 local_irq_save(flags);
4344 dolock = rb_reader_lock(cpu_buffer);
4345 ret = rb_per_cpu_empty(cpu_buffer);
4346 rb_reader_unlock(cpu_buffer, dolock);
4347 local_irq_restore(flags);
4355 EXPORT_SYMBOL_GPL(ring_buffer_empty);
4358 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
4359 * @buffer: The ring buffer
4360 * @cpu: The CPU buffer to test
4362 bool ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
4364 struct ring_buffer_per_cpu *cpu_buffer;
4365 unsigned long flags;
4369 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4372 cpu_buffer = buffer->buffers[cpu];
4373 local_irq_save(flags);
4374 dolock = rb_reader_lock(cpu_buffer);
4375 ret = rb_per_cpu_empty(cpu_buffer);
4376 rb_reader_unlock(cpu_buffer, dolock);
4377 local_irq_restore(flags);
4381 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4383 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4385 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4386 * @buffer_a: One buffer to swap with
4387 * @buffer_b: The other buffer to swap with
4389 * This function is useful for tracers that want to take a "snapshot"
4390 * of a CPU buffer and has another back up buffer lying around.
4391 * it is expected that the tracer handles the cpu buffer not being
4392 * used at the moment.
4394 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4395 struct ring_buffer *buffer_b, int cpu)
4397 struct ring_buffer_per_cpu *cpu_buffer_a;
4398 struct ring_buffer_per_cpu *cpu_buffer_b;
4401 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4402 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4405 cpu_buffer_a = buffer_a->buffers[cpu];
4406 cpu_buffer_b = buffer_b->buffers[cpu];
4408 /* At least make sure the two buffers are somewhat the same */
4409 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4414 if (atomic_read(&buffer_a->record_disabled))
4417 if (atomic_read(&buffer_b->record_disabled))
4420 if (atomic_read(&cpu_buffer_a->record_disabled))
4423 if (atomic_read(&cpu_buffer_b->record_disabled))
4427 * We can't do a synchronize_sched here because this
4428 * function can be called in atomic context.
4429 * Normally this will be called from the same CPU as cpu.
4430 * If not it's up to the caller to protect this.
4432 atomic_inc(&cpu_buffer_a->record_disabled);
4433 atomic_inc(&cpu_buffer_b->record_disabled);
4436 if (local_read(&cpu_buffer_a->committing))
4438 if (local_read(&cpu_buffer_b->committing))
4441 buffer_a->buffers[cpu] = cpu_buffer_b;
4442 buffer_b->buffers[cpu] = cpu_buffer_a;
4444 cpu_buffer_b->buffer = buffer_a;
4445 cpu_buffer_a->buffer = buffer_b;
4450 atomic_dec(&cpu_buffer_a->record_disabled);
4451 atomic_dec(&cpu_buffer_b->record_disabled);
4455 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4456 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4459 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4460 * @buffer: the buffer to allocate for.
4461 * @cpu: the cpu buffer to allocate.
4463 * This function is used in conjunction with ring_buffer_read_page.
4464 * When reading a full page from the ring buffer, these functions
4465 * can be used to speed up the process. The calling function should
4466 * allocate a few pages first with this function. Then when it
4467 * needs to get pages from the ring buffer, it passes the result
4468 * of this function into ring_buffer_read_page, which will swap
4469 * the page that was allocated, with the read page of the buffer.
4472 * The page allocated, or ERR_PTR
4474 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4476 struct ring_buffer_per_cpu *cpu_buffer;
4477 struct buffer_data_page *bpage = NULL;
4478 unsigned long flags;
4481 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4482 return ERR_PTR(-ENODEV);
4484 cpu_buffer = buffer->buffers[cpu];
4485 local_irq_save(flags);
4486 arch_spin_lock(&cpu_buffer->lock);
4488 if (cpu_buffer->free_page) {
4489 bpage = cpu_buffer->free_page;
4490 cpu_buffer->free_page = NULL;
4493 arch_spin_unlock(&cpu_buffer->lock);
4494 local_irq_restore(flags);
4499 page = alloc_pages_node(cpu_to_node(cpu),
4500 GFP_KERNEL | __GFP_NORETRY, 0);
4502 return ERR_PTR(-ENOMEM);
4504 bpage = page_address(page);
4507 rb_init_page(bpage);
4511 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4514 * ring_buffer_free_read_page - free an allocated read page
4515 * @buffer: the buffer the page was allocate for
4516 * @cpu: the cpu buffer the page came from
4517 * @data: the page to free
4519 * Free a page allocated from ring_buffer_alloc_read_page.
4521 void ring_buffer_free_read_page(struct ring_buffer *buffer, int cpu, void *data)
4523 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4524 struct buffer_data_page *bpage = data;
4525 struct page *page = virt_to_page(bpage);
4526 unsigned long flags;
4528 /* If the page is still in use someplace else, we can't reuse it */
4529 if (page_ref_count(page) > 1)
4532 local_irq_save(flags);
4533 arch_spin_lock(&cpu_buffer->lock);
4535 if (!cpu_buffer->free_page) {
4536 cpu_buffer->free_page = bpage;
4540 arch_spin_unlock(&cpu_buffer->lock);
4541 local_irq_restore(flags);
4544 free_page((unsigned long)bpage);
4546 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4549 * ring_buffer_read_page - extract a page from the ring buffer
4550 * @buffer: buffer to extract from
4551 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4552 * @len: amount to extract
4553 * @cpu: the cpu of the buffer to extract
4554 * @full: should the extraction only happen when the page is full.
4556 * This function will pull out a page from the ring buffer and consume it.
4557 * @data_page must be the address of the variable that was returned
4558 * from ring_buffer_alloc_read_page. This is because the page might be used
4559 * to swap with a page in the ring buffer.
4562 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
4563 * if (IS_ERR(rpage))
4564 * return PTR_ERR(rpage);
4565 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4567 * process_page(rpage, ret);
4569 * When @full is set, the function will not return true unless
4570 * the writer is off the reader page.
4572 * Note: it is up to the calling functions to handle sleeps and wakeups.
4573 * The ring buffer can be used anywhere in the kernel and can not
4574 * blindly call wake_up. The layer that uses the ring buffer must be
4575 * responsible for that.
4578 * >=0 if data has been transferred, returns the offset of consumed data.
4579 * <0 if no data has been transferred.
4581 int ring_buffer_read_page(struct ring_buffer *buffer,
4582 void **data_page, size_t len, int cpu, int full)
4584 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4585 struct ring_buffer_event *event;
4586 struct buffer_data_page *bpage;
4587 struct buffer_page *reader;
4588 unsigned long missed_events;
4589 unsigned long flags;
4590 unsigned int commit;
4595 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4599 * If len is not big enough to hold the page header, then
4600 * we can not copy anything.
4602 if (len <= BUF_PAGE_HDR_SIZE)
4605 len -= BUF_PAGE_HDR_SIZE;
4614 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4616 reader = rb_get_reader_page(cpu_buffer);
4620 event = rb_reader_event(cpu_buffer);
4622 read = reader->read;
4623 commit = rb_page_commit(reader);
4625 /* Check if any events were dropped */
4626 missed_events = cpu_buffer->lost_events;
4629 * If this page has been partially read or
4630 * if len is not big enough to read the rest of the page or
4631 * a writer is still on the page, then
4632 * we must copy the data from the page to the buffer.
4633 * Otherwise, we can simply swap the page with the one passed in.
4635 if (read || (len < (commit - read)) ||
4636 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4637 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4638 unsigned int rpos = read;
4639 unsigned int pos = 0;
4645 if (len > (commit - read))
4646 len = (commit - read);
4648 /* Always keep the time extend and data together */
4649 size = rb_event_ts_length(event);
4654 /* save the current timestamp, since the user will need it */
4655 save_timestamp = cpu_buffer->read_stamp;
4657 /* Need to copy one event at a time */
4659 /* We need the size of one event, because
4660 * rb_advance_reader only advances by one event,
4661 * whereas rb_event_ts_length may include the size of
4662 * one or two events.
4663 * We have already ensured there's enough space if this
4664 * is a time extend. */
4665 size = rb_event_length(event);
4666 memcpy(bpage->data + pos, rpage->data + rpos, size);
4670 rb_advance_reader(cpu_buffer);
4671 rpos = reader->read;
4677 event = rb_reader_event(cpu_buffer);
4678 /* Always keep the time extend and data together */
4679 size = rb_event_ts_length(event);
4680 } while (len >= size);
4683 local_set(&bpage->commit, pos);
4684 bpage->time_stamp = save_timestamp;
4686 /* we copied everything to the beginning */
4689 /* update the entry counter */
4690 cpu_buffer->read += rb_page_entries(reader);
4691 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4693 /* swap the pages */
4694 rb_init_page(bpage);
4695 bpage = reader->page;
4696 reader->page = *data_page;
4697 local_set(&reader->write, 0);
4698 local_set(&reader->entries, 0);
4703 * Use the real_end for the data size,
4704 * This gives us a chance to store the lost events
4707 if (reader->real_end)
4708 local_set(&bpage->commit, reader->real_end);
4712 cpu_buffer->lost_events = 0;
4714 commit = local_read(&bpage->commit);
4716 * Set a flag in the commit field if we lost events
4718 if (missed_events) {
4719 /* If there is room at the end of the page to save the
4720 * missed events, then record it there.
4722 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4723 memcpy(&bpage->data[commit], &missed_events,
4724 sizeof(missed_events));
4725 local_add(RB_MISSED_STORED, &bpage->commit);
4726 commit += sizeof(missed_events);
4728 local_add(RB_MISSED_EVENTS, &bpage->commit);
4732 * This page may be off to user land. Zero it out here.
4734 if (commit < BUF_PAGE_SIZE)
4735 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4738 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4743 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4746 * We only allocate new buffers, never free them if the CPU goes down.
4747 * If we were to free the buffer, then the user would lose any trace that was in
4750 int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
4752 struct ring_buffer *buffer;
4755 unsigned long nr_pages;
4757 buffer = container_of(node, struct ring_buffer, node);
4758 if (cpumask_test_cpu(cpu, buffer->cpumask))
4763 /* check if all cpu sizes are same */
4764 for_each_buffer_cpu(buffer, cpu_i) {
4765 /* fill in the size from first enabled cpu */
4767 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4768 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4773 /* allocate minimum pages, user can later expand it */
4776 buffer->buffers[cpu] =
4777 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4778 if (!buffer->buffers[cpu]) {
4779 WARN(1, "failed to allocate ring buffer on CPU %u\n",
4784 cpumask_set_cpu(cpu, buffer->cpumask);
4788 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
4790 * This is a basic integrity check of the ring buffer.
4791 * Late in the boot cycle this test will run when configured in.
4792 * It will kick off a thread per CPU that will go into a loop
4793 * writing to the per cpu ring buffer various sizes of data.
4794 * Some of the data will be large items, some small.
4796 * Another thread is created that goes into a spin, sending out
4797 * IPIs to the other CPUs to also write into the ring buffer.
4798 * this is to test the nesting ability of the buffer.
4800 * Basic stats are recorded and reported. If something in the
4801 * ring buffer should happen that's not expected, a big warning
4802 * is displayed and all ring buffers are disabled.
4804 static struct task_struct *rb_threads[NR_CPUS] __initdata;
4806 struct rb_test_data {
4807 struct ring_buffer *buffer;
4808 unsigned long events;
4809 unsigned long bytes_written;
4810 unsigned long bytes_alloc;
4811 unsigned long bytes_dropped;
4812 unsigned long events_nested;
4813 unsigned long bytes_written_nested;
4814 unsigned long bytes_alloc_nested;
4815 unsigned long bytes_dropped_nested;
4816 int min_size_nested;
4817 int max_size_nested;
4824 static struct rb_test_data rb_data[NR_CPUS] __initdata;
4827 #define RB_TEST_BUFFER_SIZE 1048576
4829 static char rb_string[] __initdata =
4830 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
4831 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
4832 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
4834 static bool rb_test_started __initdata;
4841 static __init int rb_write_something(struct rb_test_data *data, bool nested)
4843 struct ring_buffer_event *event;
4844 struct rb_item *item;
4851 /* Have nested writes different that what is written */
4852 cnt = data->cnt + (nested ? 27 : 0);
4854 /* Multiply cnt by ~e, to make some unique increment */
4855 size = (data->cnt * 68 / 25) % (sizeof(rb_string) - 1);
4857 len = size + sizeof(struct rb_item);
4859 started = rb_test_started;
4860 /* read rb_test_started before checking buffer enabled */
4863 event = ring_buffer_lock_reserve(data->buffer, len);
4865 /* Ignore dropped events before test starts. */
4868 data->bytes_dropped += len;
4870 data->bytes_dropped_nested += len;
4875 event_len = ring_buffer_event_length(event);
4877 if (RB_WARN_ON(data->buffer, event_len < len))
4880 item = ring_buffer_event_data(event);
4882 memcpy(item->str, rb_string, size);
4885 data->bytes_alloc_nested += event_len;
4886 data->bytes_written_nested += len;
4887 data->events_nested++;
4888 if (!data->min_size_nested || len < data->min_size_nested)
4889 data->min_size_nested = len;
4890 if (len > data->max_size_nested)
4891 data->max_size_nested = len;
4893 data->bytes_alloc += event_len;
4894 data->bytes_written += len;
4896 if (!data->min_size || len < data->min_size)
4897 data->max_size = len;
4898 if (len > data->max_size)
4899 data->max_size = len;
4903 ring_buffer_unlock_commit(data->buffer, event);
4908 static __init int rb_test(void *arg)
4910 struct rb_test_data *data = arg;
4912 while (!kthread_should_stop()) {
4913 rb_write_something(data, false);
4916 set_current_state(TASK_INTERRUPTIBLE);
4917 /* Now sleep between a min of 100-300us and a max of 1ms */
4918 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
4924 static __init void rb_ipi(void *ignore)
4926 struct rb_test_data *data;
4927 int cpu = smp_processor_id();
4929 data = &rb_data[cpu];
4930 rb_write_something(data, true);
4933 static __init int rb_hammer_test(void *arg)
4935 while (!kthread_should_stop()) {
4937 /* Send an IPI to all cpus to write data! */
4938 smp_call_function(rb_ipi, NULL, 1);
4939 /* No sleep, but for non preempt, let others run */
4946 static __init int test_ringbuffer(void)
4948 struct task_struct *rb_hammer;
4949 struct ring_buffer *buffer;
4953 pr_info("Running ring buffer tests...\n");
4955 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
4956 if (WARN_ON(!buffer))
4959 /* Disable buffer so that threads can't write to it yet */
4960 ring_buffer_record_off(buffer);
4962 for_each_online_cpu(cpu) {
4963 rb_data[cpu].buffer = buffer;
4964 rb_data[cpu].cpu = cpu;
4965 rb_data[cpu].cnt = cpu;
4966 rb_threads[cpu] = kthread_create(rb_test, &rb_data[cpu],
4967 "rbtester/%d", cpu);
4968 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
4969 pr_cont("FAILED\n");
4970 ret = PTR_ERR(rb_threads[cpu]);
4974 kthread_bind(rb_threads[cpu], cpu);
4975 wake_up_process(rb_threads[cpu]);
4978 /* Now create the rb hammer! */
4979 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
4980 if (WARN_ON(IS_ERR(rb_hammer))) {
4981 pr_cont("FAILED\n");
4982 ret = PTR_ERR(rb_hammer);
4986 ring_buffer_record_on(buffer);
4988 * Show buffer is enabled before setting rb_test_started.
4989 * Yes there's a small race window where events could be
4990 * dropped and the thread wont catch it. But when a ring
4991 * buffer gets enabled, there will always be some kind of
4992 * delay before other CPUs see it. Thus, we don't care about
4993 * those dropped events. We care about events dropped after
4994 * the threads see that the buffer is active.
4997 rb_test_started = true;
4999 set_current_state(TASK_INTERRUPTIBLE);
5000 /* Just run for 10 seconds */;
5001 schedule_timeout(10 * HZ);
5003 kthread_stop(rb_hammer);
5006 for_each_online_cpu(cpu) {
5007 if (!rb_threads[cpu])
5009 kthread_stop(rb_threads[cpu]);
5012 ring_buffer_free(buffer);
5017 pr_info("finished\n");
5018 for_each_online_cpu(cpu) {
5019 struct ring_buffer_event *event;
5020 struct rb_test_data *data = &rb_data[cpu];
5021 struct rb_item *item;
5022 unsigned long total_events;
5023 unsigned long total_dropped;
5024 unsigned long total_written;
5025 unsigned long total_alloc;
5026 unsigned long total_read = 0;
5027 unsigned long total_size = 0;
5028 unsigned long total_len = 0;
5029 unsigned long total_lost = 0;
5032 int small_event_size;
5036 total_events = data->events + data->events_nested;
5037 total_written = data->bytes_written + data->bytes_written_nested;
5038 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
5039 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
5041 big_event_size = data->max_size + data->max_size_nested;
5042 small_event_size = data->min_size + data->min_size_nested;
5044 pr_info("CPU %d:\n", cpu);
5045 pr_info(" events: %ld\n", total_events);
5046 pr_info(" dropped bytes: %ld\n", total_dropped);
5047 pr_info(" alloced bytes: %ld\n", total_alloc);
5048 pr_info(" written bytes: %ld\n", total_written);
5049 pr_info(" biggest event: %d\n", big_event_size);
5050 pr_info(" smallest event: %d\n", small_event_size);
5052 if (RB_WARN_ON(buffer, total_dropped))
5057 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
5059 item = ring_buffer_event_data(event);
5060 total_len += ring_buffer_event_length(event);
5061 total_size += item->size + sizeof(struct rb_item);
5062 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
5063 pr_info("FAILED!\n");
5064 pr_info("buffer had: %.*s\n", item->size, item->str);
5065 pr_info("expected: %.*s\n", item->size, rb_string);
5066 RB_WARN_ON(buffer, 1);
5077 pr_info(" read events: %ld\n", total_read);
5078 pr_info(" lost events: %ld\n", total_lost);
5079 pr_info(" total events: %ld\n", total_lost + total_read);
5080 pr_info(" recorded len bytes: %ld\n", total_len);
5081 pr_info(" recorded size bytes: %ld\n", total_size);
5083 pr_info(" With dropped events, record len and size may not match\n"
5084 " alloced and written from above\n");
5086 if (RB_WARN_ON(buffer, total_len != total_alloc ||
5087 total_size != total_written))
5090 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
5096 pr_info("Ring buffer PASSED!\n");
5098 ring_buffer_free(buffer);
5102 late_initcall(test_ringbuffer);
5103 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */