1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
7 #include <linux/trace_recursion.h>
8 #include <linux/trace_events.h>
9 #include <linux/ring_buffer.h>
10 #include <linux/trace_clock.h>
11 #include <linux/sched/clock.h>
12 #include <linux/trace_seq.h>
13 #include <linux/spinlock.h>
14 #include <linux/irq_work.h>
15 #include <linux/security.h>
16 #include <linux/uaccess.h>
17 #include <linux/hardirq.h>
18 #include <linux/kthread.h> /* for self test */
19 #include <linux/module.h>
20 #include <linux/percpu.h>
21 #include <linux/mutex.h>
22 #include <linux/delay.h>
23 #include <linux/slab.h>
24 #include <linux/init.h>
25 #include <linux/hash.h>
26 #include <linux/list.h>
27 #include <linux/cpu.h>
28 #include <linux/oom.h>
30 #include <asm/local.h>
33 * The "absolute" timestamp in the buffer is only 59 bits.
34 * If a clock has the 5 MSBs set, it needs to be saved and
37 #define TS_MSB (0xf8ULL << 56)
38 #define ABS_TS_MASK (~TS_MSB)
40 static void update_pages_handler(struct work_struct *work);
43 * The ring buffer header is special. We must manually up keep it.
45 int ring_buffer_print_entry_header(struct trace_seq *s)
47 trace_seq_puts(s, "# compressed entry header\n");
48 trace_seq_puts(s, "\ttype_len : 5 bits\n");
49 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
50 trace_seq_puts(s, "\tarray : 32 bits\n");
51 trace_seq_putc(s, '\n');
52 trace_seq_printf(s, "\tpadding : type == %d\n",
53 RINGBUF_TYPE_PADDING);
54 trace_seq_printf(s, "\ttime_extend : type == %d\n",
55 RINGBUF_TYPE_TIME_EXTEND);
56 trace_seq_printf(s, "\ttime_stamp : type == %d\n",
57 RINGBUF_TYPE_TIME_STAMP);
58 trace_seq_printf(s, "\tdata max type_len == %d\n",
59 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
61 return !trace_seq_has_overflowed(s);
65 * The ring buffer is made up of a list of pages. A separate list of pages is
66 * allocated for each CPU. A writer may only write to a buffer that is
67 * associated with the CPU it is currently executing on. A reader may read
68 * from any per cpu buffer.
70 * The reader is special. For each per cpu buffer, the reader has its own
71 * reader page. When a reader has read the entire reader page, this reader
72 * page is swapped with another page in the ring buffer.
74 * Now, as long as the writer is off the reader page, the reader can do what
75 * ever it wants with that page. The writer will never write to that page
76 * again (as long as it is out of the ring buffer).
78 * Here's some silly ASCII art.
81 * |reader| RING BUFFER
83 * +------+ +---+ +---+ +---+
92 * |reader| RING BUFFER
93 * |page |------------------v
94 * +------+ +---+ +---+ +---+
103 * |reader| RING BUFFER
104 * |page |------------------v
105 * +------+ +---+ +---+ +---+
107 * | +---+ +---+ +---+
110 * +------------------------------+
114 * |buffer| RING BUFFER
115 * |page |------------------v
116 * +------+ +---+ +---+ +---+
118 * | New +---+ +---+ +---+
121 * +------------------------------+
124 * After we make this swap, the reader can hand this page off to the splice
125 * code and be done with it. It can even allocate a new page if it needs to
126 * and swap that into the ring buffer.
128 * We will be using cmpxchg soon to make all this lockless.
132 /* Used for individual buffers (after the counter) */
133 #define RB_BUFFER_OFF (1 << 20)
135 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
137 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
138 #define RB_ALIGNMENT 4U
139 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
140 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
142 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
143 # define RB_FORCE_8BYTE_ALIGNMENT 0
144 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
146 # define RB_FORCE_8BYTE_ALIGNMENT 1
147 # define RB_ARCH_ALIGNMENT 8U
150 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
152 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
153 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
156 RB_LEN_TIME_EXTEND = 8,
157 RB_LEN_TIME_STAMP = 8,
160 #define skip_time_extend(event) \
161 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
163 #define extended_time(event) \
164 (event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
166 static inline bool rb_null_event(struct ring_buffer_event *event)
168 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
171 static void rb_event_set_padding(struct ring_buffer_event *event)
173 /* padding has a NULL time_delta */
174 event->type_len = RINGBUF_TYPE_PADDING;
175 event->time_delta = 0;
179 rb_event_data_length(struct ring_buffer_event *event)
184 length = event->type_len * RB_ALIGNMENT;
186 length = event->array[0];
187 return length + RB_EVNT_HDR_SIZE;
191 * Return the length of the given event. Will return
192 * the length of the time extend if the event is a
195 static inline unsigned
196 rb_event_length(struct ring_buffer_event *event)
198 switch (event->type_len) {
199 case RINGBUF_TYPE_PADDING:
200 if (rb_null_event(event))
203 return event->array[0] + RB_EVNT_HDR_SIZE;
205 case RINGBUF_TYPE_TIME_EXTEND:
206 return RB_LEN_TIME_EXTEND;
208 case RINGBUF_TYPE_TIME_STAMP:
209 return RB_LEN_TIME_STAMP;
211 case RINGBUF_TYPE_DATA:
212 return rb_event_data_length(event);
221 * Return total length of time extend and data,
222 * or just the event length for all other events.
224 static inline unsigned
225 rb_event_ts_length(struct ring_buffer_event *event)
229 if (extended_time(event)) {
230 /* time extends include the data event after it */
231 len = RB_LEN_TIME_EXTEND;
232 event = skip_time_extend(event);
234 return len + rb_event_length(event);
238 * ring_buffer_event_length - return the length of the event
239 * @event: the event to get the length of
241 * Returns the size of the data load of a data event.
242 * If the event is something other than a data event, it
243 * returns the size of the event itself. With the exception
244 * of a TIME EXTEND, where it still returns the size of the
245 * data load of the data event after it.
247 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
251 if (extended_time(event))
252 event = skip_time_extend(event);
254 length = rb_event_length(event);
255 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
257 length -= RB_EVNT_HDR_SIZE;
258 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
259 length -= sizeof(event->array[0]);
262 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
264 /* inline for ring buffer fast paths */
265 static __always_inline void *
266 rb_event_data(struct ring_buffer_event *event)
268 if (extended_time(event))
269 event = skip_time_extend(event);
270 WARN_ON_ONCE(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
271 /* If length is in len field, then array[0] has the data */
273 return (void *)&event->array[0];
274 /* Otherwise length is in array[0] and array[1] has the data */
275 return (void *)&event->array[1];
279 * ring_buffer_event_data - return the data of the event
280 * @event: the event to get the data from
282 void *ring_buffer_event_data(struct ring_buffer_event *event)
284 return rb_event_data(event);
286 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
288 #define for_each_buffer_cpu(buffer, cpu) \
289 for_each_cpu(cpu, buffer->cpumask)
291 #define for_each_online_buffer_cpu(buffer, cpu) \
292 for_each_cpu_and(cpu, buffer->cpumask, cpu_online_mask)
295 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
296 #define TS_DELTA_TEST (~TS_MASK)
298 static u64 rb_event_time_stamp(struct ring_buffer_event *event)
302 ts = event->array[0];
304 ts += event->time_delta;
309 /* Flag when events were overwritten */
310 #define RB_MISSED_EVENTS (1 << 31)
311 /* Missed count stored at end */
312 #define RB_MISSED_STORED (1 << 30)
314 struct buffer_data_page {
315 u64 time_stamp; /* page time stamp */
316 local_t commit; /* write committed index */
317 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
321 * Note, the buffer_page list must be first. The buffer pages
322 * are allocated in cache lines, which means that each buffer
323 * page will be at the beginning of a cache line, and thus
324 * the least significant bits will be zero. We use this to
325 * add flags in the list struct pointers, to make the ring buffer
329 struct list_head list; /* list of buffer pages */
330 local_t write; /* index for next write */
331 unsigned read; /* index for next read */
332 local_t entries; /* entries on this page */
333 unsigned long real_end; /* real end of data */
334 struct buffer_data_page *page; /* Actual data page */
338 * The buffer page counters, write and entries, must be reset
339 * atomically when crossing page boundaries. To synchronize this
340 * update, two counters are inserted into the number. One is
341 * the actual counter for the write position or count on the page.
343 * The other is a counter of updaters. Before an update happens
344 * the update partition of the counter is incremented. This will
345 * allow the updater to update the counter atomically.
347 * The counter is 20 bits, and the state data is 12.
349 #define RB_WRITE_MASK 0xfffff
350 #define RB_WRITE_INTCNT (1 << 20)
352 static void rb_init_page(struct buffer_data_page *bpage)
354 local_set(&bpage->commit, 0);
357 static __always_inline unsigned int rb_page_commit(struct buffer_page *bpage)
359 return local_read(&bpage->page->commit);
362 static void free_buffer_page(struct buffer_page *bpage)
364 free_page((unsigned long)bpage->page);
369 * We need to fit the time_stamp delta into 27 bits.
371 static inline bool test_time_stamp(u64 delta)
373 return !!(delta & TS_DELTA_TEST);
376 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
378 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
379 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
381 int ring_buffer_print_page_header(struct trace_seq *s)
383 struct buffer_data_page field;
385 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
386 "offset:0;\tsize:%u;\tsigned:%u;\n",
387 (unsigned int)sizeof(field.time_stamp),
388 (unsigned int)is_signed_type(u64));
390 trace_seq_printf(s, "\tfield: local_t commit;\t"
391 "offset:%u;\tsize:%u;\tsigned:%u;\n",
392 (unsigned int)offsetof(typeof(field), commit),
393 (unsigned int)sizeof(field.commit),
394 (unsigned int)is_signed_type(long));
396 trace_seq_printf(s, "\tfield: int overwrite;\t"
397 "offset:%u;\tsize:%u;\tsigned:%u;\n",
398 (unsigned int)offsetof(typeof(field), commit),
400 (unsigned int)is_signed_type(long));
402 trace_seq_printf(s, "\tfield: char data;\t"
403 "offset:%u;\tsize:%u;\tsigned:%u;\n",
404 (unsigned int)offsetof(typeof(field), data),
405 (unsigned int)BUF_PAGE_SIZE,
406 (unsigned int)is_signed_type(char));
408 return !trace_seq_has_overflowed(s);
412 struct irq_work work;
413 wait_queue_head_t waiters;
414 wait_queue_head_t full_waiters;
416 bool waiters_pending;
417 bool full_waiters_pending;
422 * Structure to hold event state and handle nested events.
424 struct rb_event_info {
429 unsigned long length;
430 struct buffer_page *tail_page;
435 * Used for the add_timestamp
437 * EXTEND - wants a time extend
438 * ABSOLUTE - the buffer requests all events to have absolute time stamps
439 * FORCE - force a full time stamp.
442 RB_ADD_STAMP_NONE = 0,
443 RB_ADD_STAMP_EXTEND = BIT(1),
444 RB_ADD_STAMP_ABSOLUTE = BIT(2),
445 RB_ADD_STAMP_FORCE = BIT(3)
448 * Used for which event context the event is in.
455 * See trace_recursive_lock() comment below for more details.
466 #if BITS_PER_LONG == 32
470 /* To test on 64 bit machines */
475 struct rb_time_struct {
482 #include <asm/local64.h>
483 struct rb_time_struct {
487 typedef struct rb_time_struct rb_time_t;
492 * head_page == tail_page && head == tail then buffer is empty.
494 struct ring_buffer_per_cpu {
496 atomic_t record_disabled;
497 atomic_t resize_disabled;
498 struct trace_buffer *buffer;
499 raw_spinlock_t reader_lock; /* serialize readers */
500 arch_spinlock_t lock;
501 struct lock_class_key lock_key;
502 struct buffer_data_page *free_page;
503 unsigned long nr_pages;
504 unsigned int current_context;
505 struct list_head *pages;
506 struct buffer_page *head_page; /* read from head */
507 struct buffer_page *tail_page; /* write to tail */
508 struct buffer_page *commit_page; /* committed pages */
509 struct buffer_page *reader_page;
510 unsigned long lost_events;
511 unsigned long last_overrun;
513 local_t entries_bytes;
516 local_t commit_overrun;
517 local_t dropped_events;
520 local_t pages_touched;
523 long last_pages_touch;
524 size_t shortest_full;
526 unsigned long read_bytes;
527 rb_time_t write_stamp;
528 rb_time_t before_stamp;
529 u64 event_stamp[MAX_NEST];
531 /* pages removed since last reset */
532 unsigned long pages_removed;
533 /* ring buffer pages to update, > 0 to add, < 0 to remove */
534 long nr_pages_to_update;
535 struct list_head new_pages; /* new pages to add */
536 struct work_struct update_pages_work;
537 struct completion update_done;
539 struct rb_irq_work irq_work;
542 struct trace_buffer {
545 atomic_t record_disabled;
547 cpumask_var_t cpumask;
549 struct lock_class_key *reader_lock_key;
553 struct ring_buffer_per_cpu **buffers;
555 struct hlist_node node;
558 struct rb_irq_work irq_work;
562 struct ring_buffer_iter {
563 struct ring_buffer_per_cpu *cpu_buffer;
565 unsigned long next_event;
566 struct buffer_page *head_page;
567 struct buffer_page *cache_reader_page;
568 unsigned long cache_read;
569 unsigned long cache_pages_removed;
572 struct ring_buffer_event *event;
579 * On 32 bit machines, local64_t is very expensive. As the ring
580 * buffer doesn't need all the features of a true 64 bit atomic,
581 * on 32 bit, it uses these functions (64 still uses local64_t).
583 * For the ring buffer, 64 bit required operations for the time is
586 * - Reads may fail if it interrupted a modification of the time stamp.
587 * It will succeed if it did not interrupt another write even if
588 * the read itself is interrupted by a write.
589 * It returns whether it was successful or not.
591 * - Writes always succeed and will overwrite other writes and writes
592 * that were done by events interrupting the current write.
594 * - A write followed by a read of the same time stamp will always succeed,
595 * but may not contain the same value.
597 * - A cmpxchg will fail if it interrupted another write or cmpxchg.
598 * Other than that, it acts like a normal cmpxchg.
600 * The 60 bit time stamp is broken up by 30 bits in a top and bottom half
601 * (bottom being the least significant 30 bits of the 60 bit time stamp).
603 * The two most significant bits of each half holds a 2 bit counter (0-3).
604 * Each update will increment this counter by one.
605 * When reading the top and bottom, if the two counter bits match then the
606 * top and bottom together make a valid 60 bit number.
608 #define RB_TIME_SHIFT 30
609 #define RB_TIME_VAL_MASK ((1 << RB_TIME_SHIFT) - 1)
610 #define RB_TIME_MSB_SHIFT 60
612 static inline int rb_time_cnt(unsigned long val)
614 return (val >> RB_TIME_SHIFT) & 3;
617 static inline u64 rb_time_val(unsigned long top, unsigned long bottom)
621 val = top & RB_TIME_VAL_MASK;
622 val <<= RB_TIME_SHIFT;
623 val |= bottom & RB_TIME_VAL_MASK;
628 static inline bool __rb_time_read(rb_time_t *t, u64 *ret, unsigned long *cnt)
630 unsigned long top, bottom, msb;
634 * If the read is interrupted by a write, then the cnt will
635 * be different. Loop until both top and bottom have been read
636 * without interruption.
639 c = local_read(&t->cnt);
640 top = local_read(&t->top);
641 bottom = local_read(&t->bottom);
642 msb = local_read(&t->msb);
643 } while (c != local_read(&t->cnt));
645 *cnt = rb_time_cnt(top);
647 /* If top, msb or bottom counts don't match, this interrupted a write */
648 if (*cnt != rb_time_cnt(msb) || *cnt != rb_time_cnt(bottom))
651 /* The shift to msb will lose its cnt bits */
652 *ret = rb_time_val(top, bottom) | ((u64)msb << RB_TIME_MSB_SHIFT);
656 static bool rb_time_read(rb_time_t *t, u64 *ret)
660 return __rb_time_read(t, ret, &cnt);
663 static inline unsigned long rb_time_val_cnt(unsigned long val, unsigned long cnt)
665 return (val & RB_TIME_VAL_MASK) | ((cnt & 3) << RB_TIME_SHIFT);
668 static inline void rb_time_split(u64 val, unsigned long *top, unsigned long *bottom,
671 *top = (unsigned long)((val >> RB_TIME_SHIFT) & RB_TIME_VAL_MASK);
672 *bottom = (unsigned long)(val & RB_TIME_VAL_MASK);
673 *msb = (unsigned long)(val >> RB_TIME_MSB_SHIFT);
676 static inline void rb_time_val_set(local_t *t, unsigned long val, unsigned long cnt)
678 val = rb_time_val_cnt(val, cnt);
682 static void rb_time_set(rb_time_t *t, u64 val)
684 unsigned long cnt, top, bottom, msb;
686 rb_time_split(val, &top, &bottom, &msb);
688 /* Writes always succeed with a valid number even if it gets interrupted. */
690 cnt = local_inc_return(&t->cnt);
691 rb_time_val_set(&t->top, top, cnt);
692 rb_time_val_set(&t->bottom, bottom, cnt);
693 rb_time_val_set(&t->msb, val >> RB_TIME_MSB_SHIFT, cnt);
694 } while (cnt != local_read(&t->cnt));
698 rb_time_read_cmpxchg(local_t *l, unsigned long expect, unsigned long set)
700 return local_try_cmpxchg(l, &expect, set);
705 /* local64_t always succeeds */
707 static inline bool rb_time_read(rb_time_t *t, u64 *ret)
709 *ret = local64_read(&t->time);
712 static void rb_time_set(rb_time_t *t, u64 val)
714 local64_set(&t->time, val);
719 * Enable this to make sure that the event passed to
720 * ring_buffer_event_time_stamp() is not committed and also
721 * is on the buffer that it passed in.
723 //#define RB_VERIFY_EVENT
724 #ifdef RB_VERIFY_EVENT
725 static struct list_head *rb_list_head(struct list_head *list);
726 static void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
729 struct buffer_page *page = cpu_buffer->commit_page;
730 struct buffer_page *tail_page = READ_ONCE(cpu_buffer->tail_page);
731 struct list_head *next;
733 unsigned long addr = (unsigned long)event;
737 /* Make sure the event exists and is not committed yet */
739 if (page == tail_page || WARN_ON_ONCE(stop++ > 100))
741 commit = local_read(&page->page->commit);
742 write = local_read(&page->write);
743 if (addr >= (unsigned long)&page->page->data[commit] &&
744 addr < (unsigned long)&page->page->data[write])
747 next = rb_list_head(page->list.next);
748 page = list_entry(next, struct buffer_page, list);
753 static inline void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
760 * The absolute time stamp drops the 5 MSBs and some clocks may
761 * require them. The rb_fix_abs_ts() will take a previous full
762 * time stamp, and add the 5 MSB of that time stamp on to the
763 * saved absolute time stamp. Then they are compared in case of
764 * the unlikely event that the latest time stamp incremented
767 static inline u64 rb_fix_abs_ts(u64 abs, u64 save_ts)
769 if (save_ts & TS_MSB) {
770 abs |= save_ts & TS_MSB;
771 /* Check for overflow */
772 if (unlikely(abs < save_ts))
778 static inline u64 rb_time_stamp(struct trace_buffer *buffer);
781 * ring_buffer_event_time_stamp - return the event's current time stamp
782 * @buffer: The buffer that the event is on
783 * @event: the event to get the time stamp of
785 * Note, this must be called after @event is reserved, and before it is
786 * committed to the ring buffer. And must be called from the same
787 * context where the event was reserved (normal, softirq, irq, etc).
789 * Returns the time stamp associated with the current event.
790 * If the event has an extended time stamp, then that is used as
791 * the time stamp to return.
792 * In the highly unlikely case that the event was nested more than
793 * the max nesting, then the write_stamp of the buffer is returned,
794 * otherwise current time is returned, but that really neither of
795 * the last two cases should ever happen.
797 u64 ring_buffer_event_time_stamp(struct trace_buffer *buffer,
798 struct ring_buffer_event *event)
800 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[smp_processor_id()];
804 /* If the event includes an absolute time, then just use that */
805 if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
806 ts = rb_event_time_stamp(event);
807 return rb_fix_abs_ts(ts, cpu_buffer->tail_page->page->time_stamp);
810 nest = local_read(&cpu_buffer->committing);
811 verify_event(cpu_buffer, event);
812 if (WARN_ON_ONCE(!nest))
815 /* Read the current saved nesting level time stamp */
816 if (likely(--nest < MAX_NEST))
817 return cpu_buffer->event_stamp[nest];
819 /* Shouldn't happen, warn if it does */
820 WARN_ONCE(1, "nest (%d) greater than max", nest);
823 /* Can only fail on 32 bit */
824 if (!rb_time_read(&cpu_buffer->write_stamp, &ts))
825 /* Screw it, just read the current time */
826 ts = rb_time_stamp(cpu_buffer->buffer);
832 * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
833 * @buffer: The ring_buffer to get the number of pages from
834 * @cpu: The cpu of the ring_buffer to get the number of pages from
836 * Returns the number of pages used by a per_cpu buffer of the ring buffer.
838 size_t ring_buffer_nr_pages(struct trace_buffer *buffer, int cpu)
840 return buffer->buffers[cpu]->nr_pages;
844 * ring_buffer_nr_dirty_pages - get the number of used pages in the ring buffer
845 * @buffer: The ring_buffer to get the number of pages from
846 * @cpu: The cpu of the ring_buffer to get the number of pages from
848 * Returns the number of pages that have content in the ring buffer.
850 size_t ring_buffer_nr_dirty_pages(struct trace_buffer *buffer, int cpu)
856 read = local_read(&buffer->buffers[cpu]->pages_read);
857 lost = local_read(&buffer->buffers[cpu]->pages_lost);
858 cnt = local_read(&buffer->buffers[cpu]->pages_touched);
860 if (WARN_ON_ONCE(cnt < lost))
865 /* The reader can read an empty page, but not more than that */
867 WARN_ON_ONCE(read > cnt + 1);
874 static __always_inline bool full_hit(struct trace_buffer *buffer, int cpu, int full)
876 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
880 nr_pages = cpu_buffer->nr_pages;
881 if (!nr_pages || !full)
885 * Add one as dirty will never equal nr_pages, as the sub-buffer
886 * that the writer is on is not counted as dirty.
887 * This is needed if "buffer_percent" is set to 100.
889 dirty = ring_buffer_nr_dirty_pages(buffer, cpu) + 1;
891 return (dirty * 100) >= (full * nr_pages);
895 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
897 * Schedules a delayed work to wake up any task that is blocked on the
898 * ring buffer waiters queue.
900 static void rb_wake_up_waiters(struct irq_work *work)
902 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
904 wake_up_all(&rbwork->waiters);
905 if (rbwork->full_waiters_pending || rbwork->wakeup_full) {
906 rbwork->wakeup_full = false;
907 rbwork->full_waiters_pending = false;
908 wake_up_all(&rbwork->full_waiters);
913 * ring_buffer_wake_waiters - wake up any waiters on this ring buffer
914 * @buffer: The ring buffer to wake waiters on
915 * @cpu: The CPU buffer to wake waiters on
917 * In the case of a file that represents a ring buffer is closing,
918 * it is prudent to wake up any waiters that are on this.
920 void ring_buffer_wake_waiters(struct trace_buffer *buffer, int cpu)
922 struct ring_buffer_per_cpu *cpu_buffer;
923 struct rb_irq_work *rbwork;
928 if (cpu == RING_BUFFER_ALL_CPUS) {
930 /* Wake up individual ones too. One level recursion */
931 for_each_buffer_cpu(buffer, cpu)
932 ring_buffer_wake_waiters(buffer, cpu);
934 rbwork = &buffer->irq_work;
936 if (WARN_ON_ONCE(!buffer->buffers))
938 if (WARN_ON_ONCE(cpu >= nr_cpu_ids))
941 cpu_buffer = buffer->buffers[cpu];
942 /* The CPU buffer may not have been initialized yet */
945 rbwork = &cpu_buffer->irq_work;
948 rbwork->wait_index++;
949 /* make sure the waiters see the new index */
952 /* This can be called in any context */
953 irq_work_queue(&rbwork->work);
957 * ring_buffer_wait - wait for input to the ring buffer
958 * @buffer: buffer to wait on
959 * @cpu: the cpu buffer to wait on
960 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
962 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
963 * as data is added to any of the @buffer's cpu buffers. Otherwise
964 * it will wait for data to be added to a specific cpu buffer.
966 int ring_buffer_wait(struct trace_buffer *buffer, int cpu, int full)
968 struct ring_buffer_per_cpu *cpu_buffer;
970 struct rb_irq_work *work;
975 * Depending on what the caller is waiting for, either any
976 * data in any cpu buffer, or a specific buffer, put the
977 * caller on the appropriate wait queue.
979 if (cpu == RING_BUFFER_ALL_CPUS) {
980 work = &buffer->irq_work;
981 /* Full only makes sense on per cpu reads */
984 if (!cpumask_test_cpu(cpu, buffer->cpumask))
986 cpu_buffer = buffer->buffers[cpu];
987 work = &cpu_buffer->irq_work;
990 wait_index = READ_ONCE(work->wait_index);
994 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
996 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
999 * The events can happen in critical sections where
1000 * checking a work queue can cause deadlocks.
1001 * After adding a task to the queue, this flag is set
1002 * only to notify events to try to wake up the queue
1005 * We don't clear it even if the buffer is no longer
1006 * empty. The flag only causes the next event to run
1007 * irq_work to do the work queue wake up. The worse
1008 * that can happen if we race with !trace_empty() is that
1009 * an event will cause an irq_work to try to wake up
1012 * There's no reason to protect this flag either, as
1013 * the work queue and irq_work logic will do the necessary
1014 * synchronization for the wake ups. The only thing
1015 * that is necessary is that the wake up happens after
1016 * a task has been queued. It's OK for spurious wake ups.
1019 work->full_waiters_pending = true;
1021 work->waiters_pending = true;
1023 if (signal_pending(current)) {
1028 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
1031 if (cpu != RING_BUFFER_ALL_CPUS &&
1032 !ring_buffer_empty_cpu(buffer, cpu)) {
1033 unsigned long flags;
1040 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
1041 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
1042 done = !pagebusy && full_hit(buffer, cpu, full);
1044 if (!cpu_buffer->shortest_full ||
1045 cpu_buffer->shortest_full > full)
1046 cpu_buffer->shortest_full = full;
1047 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
1054 /* Make sure to see the new wait index */
1056 if (wait_index != work->wait_index)
1061 finish_wait(&work->full_waiters, &wait);
1063 finish_wait(&work->waiters, &wait);
1069 * ring_buffer_poll_wait - poll on buffer input
1070 * @buffer: buffer to wait on
1071 * @cpu: the cpu buffer to wait on
1072 * @filp: the file descriptor
1073 * @poll_table: The poll descriptor
1074 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
1076 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
1077 * as data is added to any of the @buffer's cpu buffers. Otherwise
1078 * it will wait for data to be added to a specific cpu buffer.
1080 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
1083 __poll_t ring_buffer_poll_wait(struct trace_buffer *buffer, int cpu,
1084 struct file *filp, poll_table *poll_table, int full)
1086 struct ring_buffer_per_cpu *cpu_buffer;
1087 struct rb_irq_work *work;
1089 if (cpu == RING_BUFFER_ALL_CPUS) {
1090 work = &buffer->irq_work;
1093 if (!cpumask_test_cpu(cpu, buffer->cpumask))
1096 cpu_buffer = buffer->buffers[cpu];
1097 work = &cpu_buffer->irq_work;
1101 poll_wait(filp, &work->full_waiters, poll_table);
1102 work->full_waiters_pending = true;
1103 if (!cpu_buffer->shortest_full ||
1104 cpu_buffer->shortest_full > full)
1105 cpu_buffer->shortest_full = full;
1107 poll_wait(filp, &work->waiters, poll_table);
1108 work->waiters_pending = true;
1112 * There's a tight race between setting the waiters_pending and
1113 * checking if the ring buffer is empty. Once the waiters_pending bit
1114 * is set, the next event will wake the task up, but we can get stuck
1115 * if there's only a single event in.
1117 * FIXME: Ideally, we need a memory barrier on the writer side as well,
1118 * but adding a memory barrier to all events will cause too much of a
1119 * performance hit in the fast path. We only need a memory barrier when
1120 * the buffer goes from empty to having content. But as this race is
1121 * extremely small, and it's not a problem if another event comes in, we
1122 * will fix it later.
1127 return full_hit(buffer, cpu, full) ? EPOLLIN | EPOLLRDNORM : 0;
1129 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
1130 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
1131 return EPOLLIN | EPOLLRDNORM;
1135 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
1136 #define RB_WARN_ON(b, cond) \
1138 int _____ret = unlikely(cond); \
1140 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
1141 struct ring_buffer_per_cpu *__b = \
1143 atomic_inc(&__b->buffer->record_disabled); \
1145 atomic_inc(&b->record_disabled); \
1151 /* Up this if you want to test the TIME_EXTENTS and normalization */
1152 #define DEBUG_SHIFT 0
1154 static inline u64 rb_time_stamp(struct trace_buffer *buffer)
1158 /* Skip retpolines :-( */
1159 if (IS_ENABLED(CONFIG_RETPOLINE) && likely(buffer->clock == trace_clock_local))
1160 ts = trace_clock_local();
1162 ts = buffer->clock();
1164 /* shift to debug/test normalization and TIME_EXTENTS */
1165 return ts << DEBUG_SHIFT;
1168 u64 ring_buffer_time_stamp(struct trace_buffer *buffer)
1172 preempt_disable_notrace();
1173 time = rb_time_stamp(buffer);
1174 preempt_enable_notrace();
1178 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
1180 void ring_buffer_normalize_time_stamp(struct trace_buffer *buffer,
1183 /* Just stupid testing the normalize function and deltas */
1184 *ts >>= DEBUG_SHIFT;
1186 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
1189 * Making the ring buffer lockless makes things tricky.
1190 * Although writes only happen on the CPU that they are on,
1191 * and they only need to worry about interrupts. Reads can
1192 * happen on any CPU.
1194 * The reader page is always off the ring buffer, but when the
1195 * reader finishes with a page, it needs to swap its page with
1196 * a new one from the buffer. The reader needs to take from
1197 * the head (writes go to the tail). But if a writer is in overwrite
1198 * mode and wraps, it must push the head page forward.
1200 * Here lies the problem.
1202 * The reader must be careful to replace only the head page, and
1203 * not another one. As described at the top of the file in the
1204 * ASCII art, the reader sets its old page to point to the next
1205 * page after head. It then sets the page after head to point to
1206 * the old reader page. But if the writer moves the head page
1207 * during this operation, the reader could end up with the tail.
1209 * We use cmpxchg to help prevent this race. We also do something
1210 * special with the page before head. We set the LSB to 1.
1212 * When the writer must push the page forward, it will clear the
1213 * bit that points to the head page, move the head, and then set
1214 * the bit that points to the new head page.
1216 * We also don't want an interrupt coming in and moving the head
1217 * page on another writer. Thus we use the second LSB to catch
1220 * head->list->prev->next bit 1 bit 0
1223 * Points to head page 0 1
1226 * Note we can not trust the prev pointer of the head page, because:
1228 * +----+ +-----+ +-----+
1229 * | |------>| T |---X--->| N |
1231 * +----+ +-----+ +-----+
1234 * +----------| R |----------+ |
1238 * Key: ---X--> HEAD flag set in pointer
1243 * (see __rb_reserve_next() to see where this happens)
1245 * What the above shows is that the reader just swapped out
1246 * the reader page with a page in the buffer, but before it
1247 * could make the new header point back to the new page added
1248 * it was preempted by a writer. The writer moved forward onto
1249 * the new page added by the reader and is about to move forward
1252 * You can see, it is legitimate for the previous pointer of
1253 * the head (or any page) not to point back to itself. But only
1257 #define RB_PAGE_NORMAL 0UL
1258 #define RB_PAGE_HEAD 1UL
1259 #define RB_PAGE_UPDATE 2UL
1262 #define RB_FLAG_MASK 3UL
1264 /* PAGE_MOVED is not part of the mask */
1265 #define RB_PAGE_MOVED 4UL
1268 * rb_list_head - remove any bit
1270 static struct list_head *rb_list_head(struct list_head *list)
1272 unsigned long val = (unsigned long)list;
1274 return (struct list_head *)(val & ~RB_FLAG_MASK);
1278 * rb_is_head_page - test if the given page is the head page
1280 * Because the reader may move the head_page pointer, we can
1281 * not trust what the head page is (it may be pointing to
1282 * the reader page). But if the next page is a header page,
1283 * its flags will be non zero.
1286 rb_is_head_page(struct buffer_page *page, struct list_head *list)
1290 val = (unsigned long)list->next;
1292 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
1293 return RB_PAGE_MOVED;
1295 return val & RB_FLAG_MASK;
1301 * The unique thing about the reader page, is that, if the
1302 * writer is ever on it, the previous pointer never points
1303 * back to the reader page.
1305 static bool rb_is_reader_page(struct buffer_page *page)
1307 struct list_head *list = page->list.prev;
1309 return rb_list_head(list->next) != &page->list;
1313 * rb_set_list_to_head - set a list_head to be pointing to head.
1315 static void rb_set_list_to_head(struct list_head *list)
1319 ptr = (unsigned long *)&list->next;
1320 *ptr |= RB_PAGE_HEAD;
1321 *ptr &= ~RB_PAGE_UPDATE;
1325 * rb_head_page_activate - sets up head page
1327 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
1329 struct buffer_page *head;
1331 head = cpu_buffer->head_page;
1336 * Set the previous list pointer to have the HEAD flag.
1338 rb_set_list_to_head(head->list.prev);
1341 static void rb_list_head_clear(struct list_head *list)
1343 unsigned long *ptr = (unsigned long *)&list->next;
1345 *ptr &= ~RB_FLAG_MASK;
1349 * rb_head_page_deactivate - clears head page ptr (for free list)
1352 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
1354 struct list_head *hd;
1356 /* Go through the whole list and clear any pointers found. */
1357 rb_list_head_clear(cpu_buffer->pages);
1359 list_for_each(hd, cpu_buffer->pages)
1360 rb_list_head_clear(hd);
1363 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
1364 struct buffer_page *head,
1365 struct buffer_page *prev,
1366 int old_flag, int new_flag)
1368 struct list_head *list;
1369 unsigned long val = (unsigned long)&head->list;
1374 val &= ~RB_FLAG_MASK;
1376 ret = cmpxchg((unsigned long *)&list->next,
1377 val | old_flag, val | new_flag);
1379 /* check if the reader took the page */
1380 if ((ret & ~RB_FLAG_MASK) != val)
1381 return RB_PAGE_MOVED;
1383 return ret & RB_FLAG_MASK;
1386 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
1387 struct buffer_page *head,
1388 struct buffer_page *prev,
1391 return rb_head_page_set(cpu_buffer, head, prev,
1392 old_flag, RB_PAGE_UPDATE);
1395 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
1396 struct buffer_page *head,
1397 struct buffer_page *prev,
1400 return rb_head_page_set(cpu_buffer, head, prev,
1401 old_flag, RB_PAGE_HEAD);
1404 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
1405 struct buffer_page *head,
1406 struct buffer_page *prev,
1409 return rb_head_page_set(cpu_buffer, head, prev,
1410 old_flag, RB_PAGE_NORMAL);
1413 static inline void rb_inc_page(struct buffer_page **bpage)
1415 struct list_head *p = rb_list_head((*bpage)->list.next);
1417 *bpage = list_entry(p, struct buffer_page, list);
1420 static struct buffer_page *
1421 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1423 struct buffer_page *head;
1424 struct buffer_page *page;
1425 struct list_head *list;
1428 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1432 list = cpu_buffer->pages;
1433 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1436 page = head = cpu_buffer->head_page;
1438 * It is possible that the writer moves the header behind
1439 * where we started, and we miss in one loop.
1440 * A second loop should grab the header, but we'll do
1441 * three loops just because I'm paranoid.
1443 for (i = 0; i < 3; i++) {
1445 if (rb_is_head_page(page, page->list.prev)) {
1446 cpu_buffer->head_page = page;
1450 } while (page != head);
1453 RB_WARN_ON(cpu_buffer, 1);
1458 static bool rb_head_page_replace(struct buffer_page *old,
1459 struct buffer_page *new)
1461 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1464 val = *ptr & ~RB_FLAG_MASK;
1465 val |= RB_PAGE_HEAD;
1467 return try_cmpxchg(ptr, &val, (unsigned long)&new->list);
1471 * rb_tail_page_update - move the tail page forward
1473 static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1474 struct buffer_page *tail_page,
1475 struct buffer_page *next_page)
1477 unsigned long old_entries;
1478 unsigned long old_write;
1481 * The tail page now needs to be moved forward.
1483 * We need to reset the tail page, but without messing
1484 * with possible erasing of data brought in by interrupts
1485 * that have moved the tail page and are currently on it.
1487 * We add a counter to the write field to denote this.
1489 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1490 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1492 local_inc(&cpu_buffer->pages_touched);
1494 * Just make sure we have seen our old_write and synchronize
1495 * with any interrupts that come in.
1500 * If the tail page is still the same as what we think
1501 * it is, then it is up to us to update the tail
1504 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1505 /* Zero the write counter */
1506 unsigned long val = old_write & ~RB_WRITE_MASK;
1507 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1510 * This will only succeed if an interrupt did
1511 * not come in and change it. In which case, we
1512 * do not want to modify it.
1514 * We add (void) to let the compiler know that we do not care
1515 * about the return value of these functions. We use the
1516 * cmpxchg to only update if an interrupt did not already
1517 * do it for us. If the cmpxchg fails, we don't care.
1519 (void)local_cmpxchg(&next_page->write, old_write, val);
1520 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1523 * No need to worry about races with clearing out the commit.
1524 * it only can increment when a commit takes place. But that
1525 * only happens in the outer most nested commit.
1527 local_set(&next_page->page->commit, 0);
1529 /* Again, either we update tail_page or an interrupt does */
1530 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1534 static void rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1535 struct buffer_page *bpage)
1537 unsigned long val = (unsigned long)bpage;
1539 RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK);
1543 * rb_check_pages - integrity check of buffer pages
1544 * @cpu_buffer: CPU buffer with pages to test
1546 * As a safety measure we check to make sure the data pages have not
1549 static void rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1551 struct list_head *head = rb_list_head(cpu_buffer->pages);
1552 struct list_head *tmp;
1554 if (RB_WARN_ON(cpu_buffer,
1555 rb_list_head(rb_list_head(head->next)->prev) != head))
1558 if (RB_WARN_ON(cpu_buffer,
1559 rb_list_head(rb_list_head(head->prev)->next) != head))
1562 for (tmp = rb_list_head(head->next); tmp != head; tmp = rb_list_head(tmp->next)) {
1563 if (RB_WARN_ON(cpu_buffer,
1564 rb_list_head(rb_list_head(tmp->next)->prev) != tmp))
1567 if (RB_WARN_ON(cpu_buffer,
1568 rb_list_head(rb_list_head(tmp->prev)->next) != tmp))
1573 static int __rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1574 long nr_pages, struct list_head *pages)
1576 struct buffer_page *bpage, *tmp;
1577 bool user_thread = current->mm != NULL;
1582 * Check if the available memory is there first.
1583 * Note, si_mem_available() only gives us a rough estimate of available
1584 * memory. It may not be accurate. But we don't care, we just want
1585 * to prevent doing any allocation when it is obvious that it is
1586 * not going to succeed.
1588 i = si_mem_available();
1593 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1594 * gracefully without invoking oom-killer and the system is not
1597 mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1600 * If a user thread allocates too much, and si_mem_available()
1601 * reports there's enough memory, even though there is not.
1602 * Make sure the OOM killer kills this thread. This can happen
1603 * even with RETRY_MAYFAIL because another task may be doing
1604 * an allocation after this task has taken all memory.
1605 * This is the task the OOM killer needs to take out during this
1606 * loop, even if it was triggered by an allocation somewhere else.
1609 set_current_oom_origin();
1610 for (i = 0; i < nr_pages; i++) {
1613 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1614 mflags, cpu_to_node(cpu_buffer->cpu));
1618 rb_check_bpage(cpu_buffer, bpage);
1620 list_add(&bpage->list, pages);
1622 page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu), mflags, 0);
1625 bpage->page = page_address(page);
1626 rb_init_page(bpage->page);
1628 if (user_thread && fatal_signal_pending(current))
1632 clear_current_oom_origin();
1637 list_for_each_entry_safe(bpage, tmp, pages, list) {
1638 list_del_init(&bpage->list);
1639 free_buffer_page(bpage);
1642 clear_current_oom_origin();
1647 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1648 unsigned long nr_pages)
1654 if (__rb_allocate_pages(cpu_buffer, nr_pages, &pages))
1658 * The ring buffer page list is a circular list that does not
1659 * start and end with a list head. All page list items point to
1662 cpu_buffer->pages = pages.next;
1665 cpu_buffer->nr_pages = nr_pages;
1667 rb_check_pages(cpu_buffer);
1672 static struct ring_buffer_per_cpu *
1673 rb_allocate_cpu_buffer(struct trace_buffer *buffer, long nr_pages, int cpu)
1675 struct ring_buffer_per_cpu *cpu_buffer;
1676 struct buffer_page *bpage;
1680 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1681 GFP_KERNEL, cpu_to_node(cpu));
1685 cpu_buffer->cpu = cpu;
1686 cpu_buffer->buffer = buffer;
1687 raw_spin_lock_init(&cpu_buffer->reader_lock);
1688 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1689 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1690 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1691 init_completion(&cpu_buffer->update_done);
1692 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1693 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1694 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1696 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1697 GFP_KERNEL, cpu_to_node(cpu));
1699 goto fail_free_buffer;
1701 rb_check_bpage(cpu_buffer, bpage);
1703 cpu_buffer->reader_page = bpage;
1704 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1706 goto fail_free_reader;
1707 bpage->page = page_address(page);
1708 rb_init_page(bpage->page);
1710 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1711 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1713 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1715 goto fail_free_reader;
1717 cpu_buffer->head_page
1718 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1719 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1721 rb_head_page_activate(cpu_buffer);
1726 free_buffer_page(cpu_buffer->reader_page);
1733 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1735 struct list_head *head = cpu_buffer->pages;
1736 struct buffer_page *bpage, *tmp;
1738 irq_work_sync(&cpu_buffer->irq_work.work);
1740 free_buffer_page(cpu_buffer->reader_page);
1743 rb_head_page_deactivate(cpu_buffer);
1745 list_for_each_entry_safe(bpage, tmp, head, list) {
1746 list_del_init(&bpage->list);
1747 free_buffer_page(bpage);
1749 bpage = list_entry(head, struct buffer_page, list);
1750 free_buffer_page(bpage);
1753 free_page((unsigned long)cpu_buffer->free_page);
1759 * __ring_buffer_alloc - allocate a new ring_buffer
1760 * @size: the size in bytes per cpu that is needed.
1761 * @flags: attributes to set for the ring buffer.
1762 * @key: ring buffer reader_lock_key.
1764 * Currently the only flag that is available is the RB_FL_OVERWRITE
1765 * flag. This flag means that the buffer will overwrite old data
1766 * when the buffer wraps. If this flag is not set, the buffer will
1767 * drop data when the tail hits the head.
1769 struct trace_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1770 struct lock_class_key *key)
1772 struct trace_buffer *buffer;
1778 /* keep it in its own cache line */
1779 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1784 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1785 goto fail_free_buffer;
1787 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1788 buffer->flags = flags;
1789 buffer->clock = trace_clock_local;
1790 buffer->reader_lock_key = key;
1792 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1793 init_waitqueue_head(&buffer->irq_work.waiters);
1795 /* need at least two pages */
1799 buffer->cpus = nr_cpu_ids;
1801 bsize = sizeof(void *) * nr_cpu_ids;
1802 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1804 if (!buffer->buffers)
1805 goto fail_free_cpumask;
1807 cpu = raw_smp_processor_id();
1808 cpumask_set_cpu(cpu, buffer->cpumask);
1809 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1810 if (!buffer->buffers[cpu])
1811 goto fail_free_buffers;
1813 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1815 goto fail_free_buffers;
1817 mutex_init(&buffer->mutex);
1822 for_each_buffer_cpu(buffer, cpu) {
1823 if (buffer->buffers[cpu])
1824 rb_free_cpu_buffer(buffer->buffers[cpu]);
1826 kfree(buffer->buffers);
1829 free_cpumask_var(buffer->cpumask);
1835 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1838 * ring_buffer_free - free a ring buffer.
1839 * @buffer: the buffer to free.
1842 ring_buffer_free(struct trace_buffer *buffer)
1846 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1848 irq_work_sync(&buffer->irq_work.work);
1850 for_each_buffer_cpu(buffer, cpu)
1851 rb_free_cpu_buffer(buffer->buffers[cpu]);
1853 kfree(buffer->buffers);
1854 free_cpumask_var(buffer->cpumask);
1858 EXPORT_SYMBOL_GPL(ring_buffer_free);
1860 void ring_buffer_set_clock(struct trace_buffer *buffer,
1863 buffer->clock = clock;
1866 void ring_buffer_set_time_stamp_abs(struct trace_buffer *buffer, bool abs)
1868 buffer->time_stamp_abs = abs;
1871 bool ring_buffer_time_stamp_abs(struct trace_buffer *buffer)
1873 return buffer->time_stamp_abs;
1876 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1878 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1880 return local_read(&bpage->entries) & RB_WRITE_MASK;
1883 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1885 return local_read(&bpage->write) & RB_WRITE_MASK;
1889 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1891 struct list_head *tail_page, *to_remove, *next_page;
1892 struct buffer_page *to_remove_page, *tmp_iter_page;
1893 struct buffer_page *last_page, *first_page;
1894 unsigned long nr_removed;
1895 unsigned long head_bit;
1900 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1901 atomic_inc(&cpu_buffer->record_disabled);
1903 * We don't race with the readers since we have acquired the reader
1904 * lock. We also don't race with writers after disabling recording.
1905 * This makes it easy to figure out the first and the last page to be
1906 * removed from the list. We unlink all the pages in between including
1907 * the first and last pages. This is done in a busy loop so that we
1908 * lose the least number of traces.
1909 * The pages are freed after we restart recording and unlock readers.
1911 tail_page = &cpu_buffer->tail_page->list;
1914 * tail page might be on reader page, we remove the next page
1915 * from the ring buffer
1917 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1918 tail_page = rb_list_head(tail_page->next);
1919 to_remove = tail_page;
1921 /* start of pages to remove */
1922 first_page = list_entry(rb_list_head(to_remove->next),
1923 struct buffer_page, list);
1925 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1926 to_remove = rb_list_head(to_remove)->next;
1927 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1929 /* Read iterators need to reset themselves when some pages removed */
1930 cpu_buffer->pages_removed += nr_removed;
1932 next_page = rb_list_head(to_remove)->next;
1935 * Now we remove all pages between tail_page and next_page.
1936 * Make sure that we have head_bit value preserved for the
1939 tail_page->next = (struct list_head *)((unsigned long)next_page |
1941 next_page = rb_list_head(next_page);
1942 next_page->prev = tail_page;
1944 /* make sure pages points to a valid page in the ring buffer */
1945 cpu_buffer->pages = next_page;
1947 /* update head page */
1949 cpu_buffer->head_page = list_entry(next_page,
1950 struct buffer_page, list);
1952 /* pages are removed, resume tracing and then free the pages */
1953 atomic_dec(&cpu_buffer->record_disabled);
1954 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1956 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1958 /* last buffer page to remove */
1959 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1961 tmp_iter_page = first_page;
1966 to_remove_page = tmp_iter_page;
1967 rb_inc_page(&tmp_iter_page);
1969 /* update the counters */
1970 page_entries = rb_page_entries(to_remove_page);
1973 * If something was added to this page, it was full
1974 * since it is not the tail page. So we deduct the
1975 * bytes consumed in ring buffer from here.
1976 * Increment overrun to account for the lost events.
1978 local_add(page_entries, &cpu_buffer->overrun);
1979 local_sub(rb_page_commit(to_remove_page), &cpu_buffer->entries_bytes);
1980 local_inc(&cpu_buffer->pages_lost);
1984 * We have already removed references to this list item, just
1985 * free up the buffer_page and its page
1987 free_buffer_page(to_remove_page);
1990 } while (to_remove_page != last_page);
1992 RB_WARN_ON(cpu_buffer, nr_removed);
1994 return nr_removed == 0;
1998 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
2000 struct list_head *pages = &cpu_buffer->new_pages;
2001 unsigned long flags;
2005 /* Can be called at early boot up, where interrupts must not been enabled */
2006 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2008 * We are holding the reader lock, so the reader page won't be swapped
2009 * in the ring buffer. Now we are racing with the writer trying to
2010 * move head page and the tail page.
2011 * We are going to adapt the reader page update process where:
2012 * 1. We first splice the start and end of list of new pages between
2013 * the head page and its previous page.
2014 * 2. We cmpxchg the prev_page->next to point from head page to the
2015 * start of new pages list.
2016 * 3. Finally, we update the head->prev to the end of new list.
2018 * We will try this process 10 times, to make sure that we don't keep
2024 struct list_head *head_page, *prev_page, *r;
2025 struct list_head *last_page, *first_page;
2026 struct list_head *head_page_with_bit;
2027 struct buffer_page *hpage = rb_set_head_page(cpu_buffer);
2031 head_page = &hpage->list;
2032 prev_page = head_page->prev;
2034 first_page = pages->next;
2035 last_page = pages->prev;
2037 head_page_with_bit = (struct list_head *)
2038 ((unsigned long)head_page | RB_PAGE_HEAD);
2040 last_page->next = head_page_with_bit;
2041 first_page->prev = prev_page;
2043 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
2045 if (r == head_page_with_bit) {
2047 * yay, we replaced the page pointer to our new list,
2048 * now, we just have to update to head page's prev
2049 * pointer to point to end of list
2051 head_page->prev = last_page;
2058 INIT_LIST_HEAD(pages);
2060 * If we weren't successful in adding in new pages, warn and stop
2063 RB_WARN_ON(cpu_buffer, !success);
2064 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2066 /* free pages if they weren't inserted */
2068 struct buffer_page *bpage, *tmp;
2069 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2071 list_del_init(&bpage->list);
2072 free_buffer_page(bpage);
2078 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
2082 if (cpu_buffer->nr_pages_to_update > 0)
2083 success = rb_insert_pages(cpu_buffer);
2085 success = rb_remove_pages(cpu_buffer,
2086 -cpu_buffer->nr_pages_to_update);
2089 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
2092 static void update_pages_handler(struct work_struct *work)
2094 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
2095 struct ring_buffer_per_cpu, update_pages_work);
2096 rb_update_pages(cpu_buffer);
2097 complete(&cpu_buffer->update_done);
2101 * ring_buffer_resize - resize the ring buffer
2102 * @buffer: the buffer to resize.
2103 * @size: the new size.
2104 * @cpu_id: the cpu buffer to resize
2106 * Minimum size is 2 * BUF_PAGE_SIZE.
2108 * Returns 0 on success and < 0 on failure.
2110 int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
2113 struct ring_buffer_per_cpu *cpu_buffer;
2114 unsigned long nr_pages;
2118 * Always succeed at resizing a non-existent buffer:
2123 /* Make sure the requested buffer exists */
2124 if (cpu_id != RING_BUFFER_ALL_CPUS &&
2125 !cpumask_test_cpu(cpu_id, buffer->cpumask))
2128 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
2130 /* we need a minimum of two pages */
2134 /* prevent another thread from changing buffer sizes */
2135 mutex_lock(&buffer->mutex);
2136 atomic_inc(&buffer->resizing);
2138 if (cpu_id == RING_BUFFER_ALL_CPUS) {
2140 * Don't succeed if resizing is disabled, as a reader might be
2141 * manipulating the ring buffer and is expecting a sane state while
2144 for_each_buffer_cpu(buffer, cpu) {
2145 cpu_buffer = buffer->buffers[cpu];
2146 if (atomic_read(&cpu_buffer->resize_disabled)) {
2148 goto out_err_unlock;
2152 /* calculate the pages to update */
2153 for_each_buffer_cpu(buffer, cpu) {
2154 cpu_buffer = buffer->buffers[cpu];
2156 cpu_buffer->nr_pages_to_update = nr_pages -
2157 cpu_buffer->nr_pages;
2159 * nothing more to do for removing pages or no update
2161 if (cpu_buffer->nr_pages_to_update <= 0)
2164 * to add pages, make sure all new pages can be
2165 * allocated without receiving ENOMEM
2167 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2168 if (__rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2169 &cpu_buffer->new_pages)) {
2170 /* not enough memory for new pages */
2180 * Fire off all the required work handlers
2181 * We can't schedule on offline CPUs, but it's not necessary
2182 * since we can change their buffer sizes without any race.
2184 for_each_buffer_cpu(buffer, cpu) {
2185 cpu_buffer = buffer->buffers[cpu];
2186 if (!cpu_buffer->nr_pages_to_update)
2189 /* Can't run something on an offline CPU. */
2190 if (!cpu_online(cpu)) {
2191 rb_update_pages(cpu_buffer);
2192 cpu_buffer->nr_pages_to_update = 0;
2194 /* Run directly if possible. */
2196 if (cpu != smp_processor_id()) {
2198 schedule_work_on(cpu,
2199 &cpu_buffer->update_pages_work);
2201 update_pages_handler(&cpu_buffer->update_pages_work);
2207 /* wait for all the updates to complete */
2208 for_each_buffer_cpu(buffer, cpu) {
2209 cpu_buffer = buffer->buffers[cpu];
2210 if (!cpu_buffer->nr_pages_to_update)
2213 if (cpu_online(cpu))
2214 wait_for_completion(&cpu_buffer->update_done);
2215 cpu_buffer->nr_pages_to_update = 0;
2220 cpu_buffer = buffer->buffers[cpu_id];
2222 if (nr_pages == cpu_buffer->nr_pages)
2226 * Don't succeed if resizing is disabled, as a reader might be
2227 * manipulating the ring buffer and is expecting a sane state while
2230 if (atomic_read(&cpu_buffer->resize_disabled)) {
2232 goto out_err_unlock;
2235 cpu_buffer->nr_pages_to_update = nr_pages -
2236 cpu_buffer->nr_pages;
2238 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2239 if (cpu_buffer->nr_pages_to_update > 0 &&
2240 __rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2241 &cpu_buffer->new_pages)) {
2248 /* Can't run something on an offline CPU. */
2249 if (!cpu_online(cpu_id))
2250 rb_update_pages(cpu_buffer);
2252 /* Run directly if possible. */
2254 if (cpu_id == smp_processor_id()) {
2255 rb_update_pages(cpu_buffer);
2259 schedule_work_on(cpu_id,
2260 &cpu_buffer->update_pages_work);
2261 wait_for_completion(&cpu_buffer->update_done);
2265 cpu_buffer->nr_pages_to_update = 0;
2271 * The ring buffer resize can happen with the ring buffer
2272 * enabled, so that the update disturbs the tracing as little
2273 * as possible. But if the buffer is disabled, we do not need
2274 * to worry about that, and we can take the time to verify
2275 * that the buffer is not corrupt.
2277 if (atomic_read(&buffer->record_disabled)) {
2278 atomic_inc(&buffer->record_disabled);
2280 * Even though the buffer was disabled, we must make sure
2281 * that it is truly disabled before calling rb_check_pages.
2282 * There could have been a race between checking
2283 * record_disable and incrementing it.
2286 for_each_buffer_cpu(buffer, cpu) {
2287 cpu_buffer = buffer->buffers[cpu];
2288 rb_check_pages(cpu_buffer);
2290 atomic_dec(&buffer->record_disabled);
2293 atomic_dec(&buffer->resizing);
2294 mutex_unlock(&buffer->mutex);
2298 for_each_buffer_cpu(buffer, cpu) {
2299 struct buffer_page *bpage, *tmp;
2301 cpu_buffer = buffer->buffers[cpu];
2302 cpu_buffer->nr_pages_to_update = 0;
2304 if (list_empty(&cpu_buffer->new_pages))
2307 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2309 list_del_init(&bpage->list);
2310 free_buffer_page(bpage);
2314 atomic_dec(&buffer->resizing);
2315 mutex_unlock(&buffer->mutex);
2318 EXPORT_SYMBOL_GPL(ring_buffer_resize);
2320 void ring_buffer_change_overwrite(struct trace_buffer *buffer, int val)
2322 mutex_lock(&buffer->mutex);
2324 buffer->flags |= RB_FL_OVERWRITE;
2326 buffer->flags &= ~RB_FL_OVERWRITE;
2327 mutex_unlock(&buffer->mutex);
2329 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
2331 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
2333 return bpage->page->data + index;
2336 static __always_inline struct ring_buffer_event *
2337 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
2339 return __rb_page_index(cpu_buffer->reader_page,
2340 cpu_buffer->reader_page->read);
2343 static struct ring_buffer_event *
2344 rb_iter_head_event(struct ring_buffer_iter *iter)
2346 struct ring_buffer_event *event;
2347 struct buffer_page *iter_head_page = iter->head_page;
2348 unsigned long commit;
2351 if (iter->head != iter->next_event)
2355 * When the writer goes across pages, it issues a cmpxchg which
2356 * is a mb(), which will synchronize with the rmb here.
2357 * (see rb_tail_page_update() and __rb_reserve_next())
2359 commit = rb_page_commit(iter_head_page);
2362 /* An event needs to be at least 8 bytes in size */
2363 if (iter->head > commit - 8)
2366 event = __rb_page_index(iter_head_page, iter->head);
2367 length = rb_event_length(event);
2370 * READ_ONCE() doesn't work on functions and we don't want the
2371 * compiler doing any crazy optimizations with length.
2375 if ((iter->head + length) > commit || length > BUF_PAGE_SIZE)
2376 /* Writer corrupted the read? */
2379 memcpy(iter->event, event, length);
2381 * If the page stamp is still the same after this rmb() then the
2382 * event was safely copied without the writer entering the page.
2386 /* Make sure the page didn't change since we read this */
2387 if (iter->page_stamp != iter_head_page->page->time_stamp ||
2388 commit > rb_page_commit(iter_head_page))
2391 iter->next_event = iter->head + length;
2394 /* Reset to the beginning */
2395 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2397 iter->next_event = 0;
2398 iter->missed_events = 1;
2402 /* Size is determined by what has been committed */
2403 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
2405 return rb_page_commit(bpage);
2408 static __always_inline unsigned
2409 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
2411 return rb_page_commit(cpu_buffer->commit_page);
2414 static __always_inline unsigned
2415 rb_event_index(struct ring_buffer_event *event)
2417 unsigned long addr = (unsigned long)event;
2419 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
2422 static void rb_inc_iter(struct ring_buffer_iter *iter)
2424 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2427 * The iterator could be on the reader page (it starts there).
2428 * But the head could have moved, since the reader was
2429 * found. Check for this case and assign the iterator
2430 * to the head page instead of next.
2432 if (iter->head_page == cpu_buffer->reader_page)
2433 iter->head_page = rb_set_head_page(cpu_buffer);
2435 rb_inc_page(&iter->head_page);
2437 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2439 iter->next_event = 0;
2443 * rb_handle_head_page - writer hit the head page
2445 * Returns: +1 to retry page
2450 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2451 struct buffer_page *tail_page,
2452 struct buffer_page *next_page)
2454 struct buffer_page *new_head;
2459 entries = rb_page_entries(next_page);
2462 * The hard part is here. We need to move the head
2463 * forward, and protect against both readers on
2464 * other CPUs and writers coming in via interrupts.
2466 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2470 * type can be one of four:
2471 * NORMAL - an interrupt already moved it for us
2472 * HEAD - we are the first to get here.
2473 * UPDATE - we are the interrupt interrupting
2475 * MOVED - a reader on another CPU moved the next
2476 * pointer to its reader page. Give up
2483 * We changed the head to UPDATE, thus
2484 * it is our responsibility to update
2487 local_add(entries, &cpu_buffer->overrun);
2488 local_sub(rb_page_commit(next_page), &cpu_buffer->entries_bytes);
2489 local_inc(&cpu_buffer->pages_lost);
2492 * The entries will be zeroed out when we move the
2496 /* still more to do */
2499 case RB_PAGE_UPDATE:
2501 * This is an interrupt that interrupt the
2502 * previous update. Still more to do.
2505 case RB_PAGE_NORMAL:
2507 * An interrupt came in before the update
2508 * and processed this for us.
2509 * Nothing left to do.
2514 * The reader is on another CPU and just did
2515 * a swap with our next_page.
2520 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2525 * Now that we are here, the old head pointer is
2526 * set to UPDATE. This will keep the reader from
2527 * swapping the head page with the reader page.
2528 * The reader (on another CPU) will spin till
2531 * We just need to protect against interrupts
2532 * doing the job. We will set the next pointer
2533 * to HEAD. After that, we set the old pointer
2534 * to NORMAL, but only if it was HEAD before.
2535 * otherwise we are an interrupt, and only
2536 * want the outer most commit to reset it.
2538 new_head = next_page;
2539 rb_inc_page(&new_head);
2541 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2545 * Valid returns are:
2546 * HEAD - an interrupt came in and already set it.
2547 * NORMAL - One of two things:
2548 * 1) We really set it.
2549 * 2) A bunch of interrupts came in and moved
2550 * the page forward again.
2554 case RB_PAGE_NORMAL:
2558 RB_WARN_ON(cpu_buffer, 1);
2563 * It is possible that an interrupt came in,
2564 * set the head up, then more interrupts came in
2565 * and moved it again. When we get back here,
2566 * the page would have been set to NORMAL but we
2567 * just set it back to HEAD.
2569 * How do you detect this? Well, if that happened
2570 * the tail page would have moved.
2572 if (ret == RB_PAGE_NORMAL) {
2573 struct buffer_page *buffer_tail_page;
2575 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2577 * If the tail had moved passed next, then we need
2578 * to reset the pointer.
2580 if (buffer_tail_page != tail_page &&
2581 buffer_tail_page != next_page)
2582 rb_head_page_set_normal(cpu_buffer, new_head,
2588 * If this was the outer most commit (the one that
2589 * changed the original pointer from HEAD to UPDATE),
2590 * then it is up to us to reset it to NORMAL.
2592 if (type == RB_PAGE_HEAD) {
2593 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2596 if (RB_WARN_ON(cpu_buffer,
2597 ret != RB_PAGE_UPDATE))
2605 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2606 unsigned long tail, struct rb_event_info *info)
2608 struct buffer_page *tail_page = info->tail_page;
2609 struct ring_buffer_event *event;
2610 unsigned long length = info->length;
2613 * Only the event that crossed the page boundary
2614 * must fill the old tail_page with padding.
2616 if (tail >= BUF_PAGE_SIZE) {
2618 * If the page was filled, then we still need
2619 * to update the real_end. Reset it to zero
2620 * and the reader will ignore it.
2622 if (tail == BUF_PAGE_SIZE)
2623 tail_page->real_end = 0;
2625 local_sub(length, &tail_page->write);
2629 event = __rb_page_index(tail_page, tail);
2632 * Save the original length to the meta data.
2633 * This will be used by the reader to add lost event
2636 tail_page->real_end = tail;
2639 * If this event is bigger than the minimum size, then
2640 * we need to be careful that we don't subtract the
2641 * write counter enough to allow another writer to slip
2643 * We put in a discarded commit instead, to make sure
2644 * that this space is not used again, and this space will
2645 * not be accounted into 'entries_bytes'.
2647 * If we are less than the minimum size, we don't need to
2650 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2651 /* No room for any events */
2653 /* Mark the rest of the page with padding */
2654 rb_event_set_padding(event);
2656 /* Make sure the padding is visible before the write update */
2659 /* Set the write back to the previous setting */
2660 local_sub(length, &tail_page->write);
2664 /* Put in a discarded event */
2665 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2666 event->type_len = RINGBUF_TYPE_PADDING;
2667 /* time delta must be non zero */
2668 event->time_delta = 1;
2670 /* account for padding bytes */
2671 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2673 /* Make sure the padding is visible before the tail_page->write update */
2676 /* Set write to end of buffer */
2677 length = (tail + length) - BUF_PAGE_SIZE;
2678 local_sub(length, &tail_page->write);
2681 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2684 * This is the slow path, force gcc not to inline it.
2686 static noinline struct ring_buffer_event *
2687 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2688 unsigned long tail, struct rb_event_info *info)
2690 struct buffer_page *tail_page = info->tail_page;
2691 struct buffer_page *commit_page = cpu_buffer->commit_page;
2692 struct trace_buffer *buffer = cpu_buffer->buffer;
2693 struct buffer_page *next_page;
2696 next_page = tail_page;
2698 rb_inc_page(&next_page);
2701 * If for some reason, we had an interrupt storm that made
2702 * it all the way around the buffer, bail, and warn
2705 if (unlikely(next_page == commit_page)) {
2706 local_inc(&cpu_buffer->commit_overrun);
2711 * This is where the fun begins!
2713 * We are fighting against races between a reader that
2714 * could be on another CPU trying to swap its reader
2715 * page with the buffer head.
2717 * We are also fighting against interrupts coming in and
2718 * moving the head or tail on us as well.
2720 * If the next page is the head page then we have filled
2721 * the buffer, unless the commit page is still on the
2724 if (rb_is_head_page(next_page, &tail_page->list)) {
2727 * If the commit is not on the reader page, then
2728 * move the header page.
2730 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2732 * If we are not in overwrite mode,
2733 * this is easy, just stop here.
2735 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2736 local_inc(&cpu_buffer->dropped_events);
2740 ret = rb_handle_head_page(cpu_buffer,
2749 * We need to be careful here too. The
2750 * commit page could still be on the reader
2751 * page. We could have a small buffer, and
2752 * have filled up the buffer with events
2753 * from interrupts and such, and wrapped.
2755 * Note, if the tail page is also on the
2756 * reader_page, we let it move out.
2758 if (unlikely((cpu_buffer->commit_page !=
2759 cpu_buffer->tail_page) &&
2760 (cpu_buffer->commit_page ==
2761 cpu_buffer->reader_page))) {
2762 local_inc(&cpu_buffer->commit_overrun);
2768 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2772 rb_reset_tail(cpu_buffer, tail, info);
2774 /* Commit what we have for now. */
2775 rb_end_commit(cpu_buffer);
2776 /* rb_end_commit() decs committing */
2777 local_inc(&cpu_buffer->committing);
2779 /* fail and let the caller try again */
2780 return ERR_PTR(-EAGAIN);
2784 rb_reset_tail(cpu_buffer, tail, info);
2790 static struct ring_buffer_event *
2791 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2794 event->type_len = RINGBUF_TYPE_TIME_STAMP;
2796 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2798 /* Not the first event on the page, or not delta? */
2799 if (abs || rb_event_index(event)) {
2800 event->time_delta = delta & TS_MASK;
2801 event->array[0] = delta >> TS_SHIFT;
2803 /* nope, just zero it */
2804 event->time_delta = 0;
2805 event->array[0] = 0;
2808 return skip_time_extend(event);
2811 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2812 static inline bool sched_clock_stable(void)
2819 rb_check_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2820 struct rb_event_info *info)
2824 WARN_ONCE(1, "Delta way too big! %llu ts=%llu before=%llu after=%llu write stamp=%llu\n%s",
2825 (unsigned long long)info->delta,
2826 (unsigned long long)info->ts,
2827 (unsigned long long)info->before,
2828 (unsigned long long)info->after,
2829 (unsigned long long)(rb_time_read(&cpu_buffer->write_stamp, &write_stamp) ? write_stamp : 0),
2830 sched_clock_stable() ? "" :
2831 "If you just came from a suspend/resume,\n"
2832 "please switch to the trace global clock:\n"
2833 " echo global > /sys/kernel/tracing/trace_clock\n"
2834 "or add trace_clock=global to the kernel command line\n");
2837 static void rb_add_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2838 struct ring_buffer_event **event,
2839 struct rb_event_info *info,
2841 unsigned int *length)
2843 bool abs = info->add_timestamp &
2844 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE);
2846 if (unlikely(info->delta > (1ULL << 59))) {
2848 * Some timers can use more than 59 bits, and when a timestamp
2849 * is added to the buffer, it will lose those bits.
2851 if (abs && (info->ts & TS_MSB)) {
2852 info->delta &= ABS_TS_MASK;
2854 /* did the clock go backwards */
2855 } else if (info->before == info->after && info->before > info->ts) {
2856 /* not interrupted */
2860 * This is possible with a recalibrating of the TSC.
2861 * Do not produce a call stack, but just report it.
2865 pr_warn("Ring buffer clock went backwards: %llu -> %llu\n",
2866 info->before, info->ts);
2869 rb_check_timestamp(cpu_buffer, info);
2873 *event = rb_add_time_stamp(*event, info->delta, abs);
2874 *length -= RB_LEN_TIME_EXTEND;
2879 * rb_update_event - update event type and data
2880 * @cpu_buffer: The per cpu buffer of the @event
2881 * @event: the event to update
2882 * @info: The info to update the @event with (contains length and delta)
2884 * Update the type and data fields of the @event. The length
2885 * is the actual size that is written to the ring buffer,
2886 * and with this, we can determine what to place into the
2890 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2891 struct ring_buffer_event *event,
2892 struct rb_event_info *info)
2894 unsigned length = info->length;
2895 u64 delta = info->delta;
2896 unsigned int nest = local_read(&cpu_buffer->committing) - 1;
2898 if (!WARN_ON_ONCE(nest >= MAX_NEST))
2899 cpu_buffer->event_stamp[nest] = info->ts;
2902 * If we need to add a timestamp, then we
2903 * add it to the start of the reserved space.
2905 if (unlikely(info->add_timestamp))
2906 rb_add_timestamp(cpu_buffer, &event, info, &delta, &length);
2908 event->time_delta = delta;
2909 length -= RB_EVNT_HDR_SIZE;
2910 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2911 event->type_len = 0;
2912 event->array[0] = length;
2914 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2917 static unsigned rb_calculate_event_length(unsigned length)
2919 struct ring_buffer_event event; /* Used only for sizeof array */
2921 /* zero length can cause confusions */
2925 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2926 length += sizeof(event.array[0]);
2928 length += RB_EVNT_HDR_SIZE;
2929 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2932 * In case the time delta is larger than the 27 bits for it
2933 * in the header, we need to add a timestamp. If another
2934 * event comes in when trying to discard this one to increase
2935 * the length, then the timestamp will be added in the allocated
2936 * space of this event. If length is bigger than the size needed
2937 * for the TIME_EXTEND, then padding has to be used. The events
2938 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2939 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2940 * As length is a multiple of 4, we only need to worry if it
2941 * is 12 (RB_LEN_TIME_EXTEND + 4).
2943 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2944 length += RB_ALIGNMENT;
2950 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2951 struct ring_buffer_event *event)
2953 unsigned long new_index, old_index;
2954 struct buffer_page *bpage;
2957 new_index = rb_event_index(event);
2958 old_index = new_index + rb_event_ts_length(event);
2959 addr = (unsigned long)event;
2962 bpage = READ_ONCE(cpu_buffer->tail_page);
2965 * Make sure the tail_page is still the same and
2966 * the next write location is the end of this event
2968 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2969 unsigned long write_mask =
2970 local_read(&bpage->write) & ~RB_WRITE_MASK;
2971 unsigned long event_length = rb_event_length(event);
2974 * For the before_stamp to be different than the write_stamp
2975 * to make sure that the next event adds an absolute
2976 * value and does not rely on the saved write stamp, which
2977 * is now going to be bogus.
2979 * By setting the before_stamp to zero, the next event
2980 * is not going to use the write_stamp and will instead
2981 * create an absolute timestamp. This means there's no
2982 * reason to update the wirte_stamp!
2984 rb_time_set(&cpu_buffer->before_stamp, 0);
2987 * If an event were to come in now, it would see that the
2988 * write_stamp and the before_stamp are different, and assume
2989 * that this event just added itself before updating
2990 * the write stamp. The interrupting event will fix the
2991 * write stamp for us, and use an absolute timestamp.
2995 * This is on the tail page. It is possible that
2996 * a write could come in and move the tail page
2997 * and write to the next page. That is fine
2998 * because we just shorten what is on this page.
3000 old_index += write_mask;
3001 new_index += write_mask;
3003 /* caution: old_index gets updated on cmpxchg failure */
3004 if (local_try_cmpxchg(&bpage->write, &old_index, new_index)) {
3005 /* update counters */
3006 local_sub(event_length, &cpu_buffer->entries_bytes);
3011 /* could not discard */
3015 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
3017 local_inc(&cpu_buffer->committing);
3018 local_inc(&cpu_buffer->commits);
3021 static __always_inline void
3022 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
3024 unsigned long max_count;
3027 * We only race with interrupts and NMIs on this CPU.
3028 * If we own the commit event, then we can commit
3029 * all others that interrupted us, since the interruptions
3030 * are in stack format (they finish before they come
3031 * back to us). This allows us to do a simple loop to
3032 * assign the commit to the tail.
3035 max_count = cpu_buffer->nr_pages * 100;
3037 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
3038 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
3040 if (RB_WARN_ON(cpu_buffer,
3041 rb_is_reader_page(cpu_buffer->tail_page)))
3044 * No need for a memory barrier here, as the update
3045 * of the tail_page did it for this page.
3047 local_set(&cpu_buffer->commit_page->page->commit,
3048 rb_page_write(cpu_buffer->commit_page));
3049 rb_inc_page(&cpu_buffer->commit_page);
3050 /* add barrier to keep gcc from optimizing too much */
3053 while (rb_commit_index(cpu_buffer) !=
3054 rb_page_write(cpu_buffer->commit_page)) {
3056 /* Make sure the readers see the content of what is committed. */
3058 local_set(&cpu_buffer->commit_page->page->commit,
3059 rb_page_write(cpu_buffer->commit_page));
3060 RB_WARN_ON(cpu_buffer,
3061 local_read(&cpu_buffer->commit_page->page->commit) &
3066 /* again, keep gcc from optimizing */
3070 * If an interrupt came in just after the first while loop
3071 * and pushed the tail page forward, we will be left with
3072 * a dangling commit that will never go forward.
3074 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
3078 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
3080 unsigned long commits;
3082 if (RB_WARN_ON(cpu_buffer,
3083 !local_read(&cpu_buffer->committing)))
3087 commits = local_read(&cpu_buffer->commits);
3088 /* synchronize with interrupts */
3090 if (local_read(&cpu_buffer->committing) == 1)
3091 rb_set_commit_to_write(cpu_buffer);
3093 local_dec(&cpu_buffer->committing);
3095 /* synchronize with interrupts */
3099 * Need to account for interrupts coming in between the
3100 * updating of the commit page and the clearing of the
3101 * committing counter.
3103 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
3104 !local_read(&cpu_buffer->committing)) {
3105 local_inc(&cpu_buffer->committing);
3110 static inline void rb_event_discard(struct ring_buffer_event *event)
3112 if (extended_time(event))
3113 event = skip_time_extend(event);
3115 /* array[0] holds the actual length for the discarded event */
3116 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
3117 event->type_len = RINGBUF_TYPE_PADDING;
3118 /* time delta must be non zero */
3119 if (!event->time_delta)
3120 event->time_delta = 1;
3123 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer)
3125 local_inc(&cpu_buffer->entries);
3126 rb_end_commit(cpu_buffer);
3129 static __always_inline void
3130 rb_wakeups(struct trace_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
3132 if (buffer->irq_work.waiters_pending) {
3133 buffer->irq_work.waiters_pending = false;
3134 /* irq_work_queue() supplies it's own memory barriers */
3135 irq_work_queue(&buffer->irq_work.work);
3138 if (cpu_buffer->irq_work.waiters_pending) {
3139 cpu_buffer->irq_work.waiters_pending = false;
3140 /* irq_work_queue() supplies it's own memory barriers */
3141 irq_work_queue(&cpu_buffer->irq_work.work);
3144 if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
3147 if (cpu_buffer->reader_page == cpu_buffer->commit_page)
3150 if (!cpu_buffer->irq_work.full_waiters_pending)
3153 cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
3155 if (!full_hit(buffer, cpu_buffer->cpu, cpu_buffer->shortest_full))
3158 cpu_buffer->irq_work.wakeup_full = true;
3159 cpu_buffer->irq_work.full_waiters_pending = false;
3160 /* irq_work_queue() supplies it's own memory barriers */
3161 irq_work_queue(&cpu_buffer->irq_work.work);
3164 #ifdef CONFIG_RING_BUFFER_RECORD_RECURSION
3165 # define do_ring_buffer_record_recursion() \
3166 do_ftrace_record_recursion(_THIS_IP_, _RET_IP_)
3168 # define do_ring_buffer_record_recursion() do { } while (0)
3172 * The lock and unlock are done within a preempt disable section.
3173 * The current_context per_cpu variable can only be modified
3174 * by the current task between lock and unlock. But it can
3175 * be modified more than once via an interrupt. To pass this
3176 * information from the lock to the unlock without having to
3177 * access the 'in_interrupt()' functions again (which do show
3178 * a bit of overhead in something as critical as function tracing,
3179 * we use a bitmask trick.
3181 * bit 1 = NMI context
3182 * bit 2 = IRQ context
3183 * bit 3 = SoftIRQ context
3184 * bit 4 = normal context.
3186 * This works because this is the order of contexts that can
3187 * preempt other contexts. A SoftIRQ never preempts an IRQ
3190 * When the context is determined, the corresponding bit is
3191 * checked and set (if it was set, then a recursion of that context
3194 * On unlock, we need to clear this bit. To do so, just subtract
3195 * 1 from the current_context and AND it to itself.
3199 * 101 & 100 = 100 (clearing bit zero)
3202 * 1010 & 1001 = 1000 (clearing bit 1)
3204 * The least significant bit can be cleared this way, and it
3205 * just so happens that it is the same bit corresponding to
3206 * the current context.
3208 * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
3209 * is set when a recursion is detected at the current context, and if
3210 * the TRANSITION bit is already set, it will fail the recursion.
3211 * This is needed because there's a lag between the changing of
3212 * interrupt context and updating the preempt count. In this case,
3213 * a false positive will be found. To handle this, one extra recursion
3214 * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
3215 * bit is already set, then it is considered a recursion and the function
3216 * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
3218 * On the trace_recursive_unlock(), the TRANSITION bit will be the first
3219 * to be cleared. Even if it wasn't the context that set it. That is,
3220 * if an interrupt comes in while NORMAL bit is set and the ring buffer
3221 * is called before preempt_count() is updated, since the check will
3222 * be on the NORMAL bit, the TRANSITION bit will then be set. If an
3223 * NMI then comes in, it will set the NMI bit, but when the NMI code
3224 * does the trace_recursive_unlock() it will clear the TRANSITION bit
3225 * and leave the NMI bit set. But this is fine, because the interrupt
3226 * code that set the TRANSITION bit will then clear the NMI bit when it
3227 * calls trace_recursive_unlock(). If another NMI comes in, it will
3228 * set the TRANSITION bit and continue.
3230 * Note: The TRANSITION bit only handles a single transition between context.
3233 static __always_inline bool
3234 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
3236 unsigned int val = cpu_buffer->current_context;
3237 int bit = interrupt_context_level();
3239 bit = RB_CTX_NORMAL - bit;
3241 if (unlikely(val & (1 << (bit + cpu_buffer->nest)))) {
3243 * It is possible that this was called by transitioning
3244 * between interrupt context, and preempt_count() has not
3245 * been updated yet. In this case, use the TRANSITION bit.
3247 bit = RB_CTX_TRANSITION;
3248 if (val & (1 << (bit + cpu_buffer->nest))) {
3249 do_ring_buffer_record_recursion();
3254 val |= (1 << (bit + cpu_buffer->nest));
3255 cpu_buffer->current_context = val;
3260 static __always_inline void
3261 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
3263 cpu_buffer->current_context &=
3264 cpu_buffer->current_context - (1 << cpu_buffer->nest);
3267 /* The recursive locking above uses 5 bits */
3268 #define NESTED_BITS 5
3271 * ring_buffer_nest_start - Allow to trace while nested
3272 * @buffer: The ring buffer to modify
3274 * The ring buffer has a safety mechanism to prevent recursion.
3275 * But there may be a case where a trace needs to be done while
3276 * tracing something else. In this case, calling this function
3277 * will allow this function to nest within a currently active
3278 * ring_buffer_lock_reserve().
3280 * Call this function before calling another ring_buffer_lock_reserve() and
3281 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
3283 void ring_buffer_nest_start(struct trace_buffer *buffer)
3285 struct ring_buffer_per_cpu *cpu_buffer;
3288 /* Enabled by ring_buffer_nest_end() */
3289 preempt_disable_notrace();
3290 cpu = raw_smp_processor_id();
3291 cpu_buffer = buffer->buffers[cpu];
3292 /* This is the shift value for the above recursive locking */
3293 cpu_buffer->nest += NESTED_BITS;
3297 * ring_buffer_nest_end - Allow to trace while nested
3298 * @buffer: The ring buffer to modify
3300 * Must be called after ring_buffer_nest_start() and after the
3301 * ring_buffer_unlock_commit().
3303 void ring_buffer_nest_end(struct trace_buffer *buffer)
3305 struct ring_buffer_per_cpu *cpu_buffer;
3308 /* disabled by ring_buffer_nest_start() */
3309 cpu = raw_smp_processor_id();
3310 cpu_buffer = buffer->buffers[cpu];
3311 /* This is the shift value for the above recursive locking */
3312 cpu_buffer->nest -= NESTED_BITS;
3313 preempt_enable_notrace();
3317 * ring_buffer_unlock_commit - commit a reserved
3318 * @buffer: The buffer to commit to
3320 * This commits the data to the ring buffer, and releases any locks held.
3322 * Must be paired with ring_buffer_lock_reserve.
3324 int ring_buffer_unlock_commit(struct trace_buffer *buffer)
3326 struct ring_buffer_per_cpu *cpu_buffer;
3327 int cpu = raw_smp_processor_id();
3329 cpu_buffer = buffer->buffers[cpu];
3331 rb_commit(cpu_buffer);
3333 rb_wakeups(buffer, cpu_buffer);
3335 trace_recursive_unlock(cpu_buffer);
3337 preempt_enable_notrace();
3341 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
3343 /* Special value to validate all deltas on a page. */
3344 #define CHECK_FULL_PAGE 1L
3346 #ifdef CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS
3347 static void dump_buffer_page(struct buffer_data_page *bpage,
3348 struct rb_event_info *info,
3351 struct ring_buffer_event *event;
3355 ts = bpage->time_stamp;
3356 pr_warn(" [%lld] PAGE TIME STAMP\n", ts);
3358 for (e = 0; e < tail; e += rb_event_length(event)) {
3360 event = (struct ring_buffer_event *)(bpage->data + e);
3362 switch (event->type_len) {
3364 case RINGBUF_TYPE_TIME_EXTEND:
3365 delta = rb_event_time_stamp(event);
3367 pr_warn(" [%lld] delta:%lld TIME EXTEND\n", ts, delta);
3370 case RINGBUF_TYPE_TIME_STAMP:
3371 delta = rb_event_time_stamp(event);
3372 ts = rb_fix_abs_ts(delta, ts);
3373 pr_warn(" [%lld] absolute:%lld TIME STAMP\n", ts, delta);
3376 case RINGBUF_TYPE_PADDING:
3377 ts += event->time_delta;
3378 pr_warn(" [%lld] delta:%d PADDING\n", ts, event->time_delta);
3381 case RINGBUF_TYPE_DATA:
3382 ts += event->time_delta;
3383 pr_warn(" [%lld] delta:%d\n", ts, event->time_delta);
3392 static DEFINE_PER_CPU(atomic_t, checking);
3393 static atomic_t ts_dump;
3396 * Check if the current event time stamp matches the deltas on
3399 static void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3400 struct rb_event_info *info,
3403 struct ring_buffer_event *event;
3404 struct buffer_data_page *bpage;
3409 bpage = info->tail_page->page;
3411 if (tail == CHECK_FULL_PAGE) {
3413 tail = local_read(&bpage->commit);
3414 } else if (info->add_timestamp &
3415 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE)) {
3416 /* Ignore events with absolute time stamps */
3421 * Do not check the first event (skip possible extends too).
3422 * Also do not check if previous events have not been committed.
3424 if (tail <= 8 || tail > local_read(&bpage->commit))
3428 * If this interrupted another event,
3430 if (atomic_inc_return(this_cpu_ptr(&checking)) != 1)
3433 ts = bpage->time_stamp;
3435 for (e = 0; e < tail; e += rb_event_length(event)) {
3437 event = (struct ring_buffer_event *)(bpage->data + e);
3439 switch (event->type_len) {
3441 case RINGBUF_TYPE_TIME_EXTEND:
3442 delta = rb_event_time_stamp(event);
3446 case RINGBUF_TYPE_TIME_STAMP:
3447 delta = rb_event_time_stamp(event);
3448 ts = rb_fix_abs_ts(delta, ts);
3451 case RINGBUF_TYPE_PADDING:
3452 if (event->time_delta == 1)
3455 case RINGBUF_TYPE_DATA:
3456 ts += event->time_delta;
3460 RB_WARN_ON(cpu_buffer, 1);
3463 if ((full && ts > info->ts) ||
3464 (!full && ts + info->delta != info->ts)) {
3465 /* If another report is happening, ignore this one */
3466 if (atomic_inc_return(&ts_dump) != 1) {
3467 atomic_dec(&ts_dump);
3470 atomic_inc(&cpu_buffer->record_disabled);
3471 /* There's some cases in boot up that this can happen */
3472 WARN_ON_ONCE(system_state != SYSTEM_BOOTING);
3473 pr_warn("[CPU: %d]TIME DOES NOT MATCH expected:%lld actual:%lld delta:%lld before:%lld after:%lld%s\n",
3475 ts + info->delta, info->ts, info->delta,
3476 info->before, info->after,
3477 full ? " (full)" : "");
3478 dump_buffer_page(bpage, info, tail);
3479 atomic_dec(&ts_dump);
3480 /* Do not re-enable checking */
3484 atomic_dec(this_cpu_ptr(&checking));
3487 static inline void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3488 struct rb_event_info *info,
3492 #endif /* CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS */
3494 static struct ring_buffer_event *
3495 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
3496 struct rb_event_info *info)
3498 struct ring_buffer_event *event;
3499 struct buffer_page *tail_page;
3500 unsigned long tail, write, w;
3504 /* Don't let the compiler play games with cpu_buffer->tail_page */
3505 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
3507 /*A*/ w = local_read(&tail_page->write) & RB_WRITE_MASK;
3509 b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3510 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3512 info->ts = rb_time_stamp(cpu_buffer->buffer);
3514 if ((info->add_timestamp & RB_ADD_STAMP_ABSOLUTE)) {
3515 info->delta = info->ts;
3518 * If interrupting an event time update, we may need an
3519 * absolute timestamp.
3520 * Don't bother if this is the start of a new page (w == 0).
3523 /* Use the sub-buffer timestamp */
3525 } else if (unlikely(!a_ok || !b_ok || info->before != info->after)) {
3526 info->add_timestamp |= RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND;
3527 info->length += RB_LEN_TIME_EXTEND;
3529 info->delta = info->ts - info->after;
3530 if (unlikely(test_time_stamp(info->delta))) {
3531 info->add_timestamp |= RB_ADD_STAMP_EXTEND;
3532 info->length += RB_LEN_TIME_EXTEND;
3537 /*B*/ rb_time_set(&cpu_buffer->before_stamp, info->ts);
3539 /*C*/ write = local_add_return(info->length, &tail_page->write);
3541 /* set write to only the index of the write */
3542 write &= RB_WRITE_MASK;
3544 tail = write - info->length;
3546 /* See if we shot pass the end of this buffer page */
3547 if (unlikely(write > BUF_PAGE_SIZE)) {
3548 check_buffer(cpu_buffer, info, CHECK_FULL_PAGE);
3549 return rb_move_tail(cpu_buffer, tail, info);
3552 if (likely(tail == w)) {
3553 /* Nothing interrupted us between A and C */
3554 /*D*/ rb_time_set(&cpu_buffer->write_stamp, info->ts);
3556 * If something came in between C and D, the write stamp
3557 * may now not be in sync. But that's fine as the before_stamp
3558 * will be different and then next event will just be forced
3559 * to use an absolute timestamp.
3561 if (likely(!(info->add_timestamp &
3562 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3563 /* This did not interrupt any time update */
3564 info->delta = info->ts - info->after;
3566 /* Just use full timestamp for interrupting event */
3567 info->delta = info->ts;
3568 check_buffer(cpu_buffer, info, tail);
3571 /* SLOW PATH - Interrupted between A and C */
3573 /* Save the old before_stamp */
3574 a_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3575 RB_WARN_ON(cpu_buffer, !a_ok);
3578 * Read a new timestamp and update the before_stamp to make
3579 * the next event after this one force using an absolute
3580 * timestamp. This is in case an interrupt were to come in
3583 ts = rb_time_stamp(cpu_buffer->buffer);
3584 rb_time_set(&cpu_buffer->before_stamp, ts);
3587 /*E*/ a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3588 /* Was interrupted before here, write_stamp must be valid */
3589 RB_WARN_ON(cpu_buffer, !a_ok);
3591 /*F*/ if (write == (local_read(&tail_page->write) & RB_WRITE_MASK) &&
3592 info->after == info->before && info->after < ts) {
3594 * Nothing came after this event between C and F, it is
3595 * safe to use info->after for the delta as it
3596 * matched info->before and is still valid.
3598 info->delta = ts - info->after;
3601 * Interrupted between C and F:
3602 * Lost the previous events time stamp. Just set the
3603 * delta to zero, and this will be the same time as
3604 * the event this event interrupted. And the events that
3605 * came after this will still be correct (as they would
3606 * have built their delta on the previous event.
3611 info->add_timestamp &= ~RB_ADD_STAMP_FORCE;
3615 * If this is the first commit on the page, then it has the same
3616 * timestamp as the page itself.
3618 if (unlikely(!tail && !(info->add_timestamp &
3619 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3622 /* We reserved something on the buffer */
3624 event = __rb_page_index(tail_page, tail);
3625 rb_update_event(cpu_buffer, event, info);
3627 local_inc(&tail_page->entries);
3630 * If this is the first commit on the page, then update
3633 if (unlikely(!tail))
3634 tail_page->page->time_stamp = info->ts;
3636 /* account for these added bytes */
3637 local_add(info->length, &cpu_buffer->entries_bytes);
3642 static __always_inline struct ring_buffer_event *
3643 rb_reserve_next_event(struct trace_buffer *buffer,
3644 struct ring_buffer_per_cpu *cpu_buffer,
3645 unsigned long length)
3647 struct ring_buffer_event *event;
3648 struct rb_event_info info;
3652 /* ring buffer does cmpxchg, make sure it is safe in NMI context */
3653 if (!IS_ENABLED(CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG) &&
3654 (unlikely(in_nmi()))) {
3658 rb_start_commit(cpu_buffer);
3659 /* The commit page can not change after this */
3661 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3663 * Due to the ability to swap a cpu buffer from a buffer
3664 * it is possible it was swapped before we committed.
3665 * (committing stops a swap). We check for it here and
3666 * if it happened, we have to fail the write.
3669 if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
3670 local_dec(&cpu_buffer->committing);
3671 local_dec(&cpu_buffer->commits);
3676 info.length = rb_calculate_event_length(length);
3678 if (ring_buffer_time_stamp_abs(cpu_buffer->buffer)) {
3679 add_ts_default = RB_ADD_STAMP_ABSOLUTE;
3680 info.length += RB_LEN_TIME_EXTEND;
3681 if (info.length > BUF_MAX_DATA_SIZE)
3684 add_ts_default = RB_ADD_STAMP_NONE;
3688 info.add_timestamp = add_ts_default;
3692 * We allow for interrupts to reenter here and do a trace.
3693 * If one does, it will cause this original code to loop
3694 * back here. Even with heavy interrupts happening, this
3695 * should only happen a few times in a row. If this happens
3696 * 1000 times in a row, there must be either an interrupt
3697 * storm or we have something buggy.
3700 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
3703 event = __rb_reserve_next(cpu_buffer, &info);
3705 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
3706 if (info.add_timestamp & (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND))
3707 info.length -= RB_LEN_TIME_EXTEND;
3714 rb_end_commit(cpu_buffer);
3719 * ring_buffer_lock_reserve - reserve a part of the buffer
3720 * @buffer: the ring buffer to reserve from
3721 * @length: the length of the data to reserve (excluding event header)
3723 * Returns a reserved event on the ring buffer to copy directly to.
3724 * The user of this interface will need to get the body to write into
3725 * and can use the ring_buffer_event_data() interface.
3727 * The length is the length of the data needed, not the event length
3728 * which also includes the event header.
3730 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
3731 * If NULL is returned, then nothing has been allocated or locked.
3733 struct ring_buffer_event *
3734 ring_buffer_lock_reserve(struct trace_buffer *buffer, unsigned long length)
3736 struct ring_buffer_per_cpu *cpu_buffer;
3737 struct ring_buffer_event *event;
3740 /* If we are tracing schedule, we don't want to recurse */
3741 preempt_disable_notrace();
3743 if (unlikely(atomic_read(&buffer->record_disabled)))
3746 cpu = raw_smp_processor_id();
3748 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
3751 cpu_buffer = buffer->buffers[cpu];
3753 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
3756 if (unlikely(length > BUF_MAX_DATA_SIZE))
3759 if (unlikely(trace_recursive_lock(cpu_buffer)))
3762 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3769 trace_recursive_unlock(cpu_buffer);
3771 preempt_enable_notrace();
3774 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3777 * Decrement the entries to the page that an event is on.
3778 * The event does not even need to exist, only the pointer
3779 * to the page it is on. This may only be called before the commit
3783 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3784 struct ring_buffer_event *event)
3786 unsigned long addr = (unsigned long)event;
3787 struct buffer_page *bpage = cpu_buffer->commit_page;
3788 struct buffer_page *start;
3792 /* Do the likely case first */
3793 if (likely(bpage->page == (void *)addr)) {
3794 local_dec(&bpage->entries);
3799 * Because the commit page may be on the reader page we
3800 * start with the next page and check the end loop there.
3802 rb_inc_page(&bpage);
3805 if (bpage->page == (void *)addr) {
3806 local_dec(&bpage->entries);
3809 rb_inc_page(&bpage);
3810 } while (bpage != start);
3812 /* commit not part of this buffer?? */
3813 RB_WARN_ON(cpu_buffer, 1);
3817 * ring_buffer_discard_commit - discard an event that has not been committed
3818 * @buffer: the ring buffer
3819 * @event: non committed event to discard
3821 * Sometimes an event that is in the ring buffer needs to be ignored.
3822 * This function lets the user discard an event in the ring buffer
3823 * and then that event will not be read later.
3825 * This function only works if it is called before the item has been
3826 * committed. It will try to free the event from the ring buffer
3827 * if another event has not been added behind it.
3829 * If another event has been added behind it, it will set the event
3830 * up as discarded, and perform the commit.
3832 * If this function is called, do not call ring_buffer_unlock_commit on
3835 void ring_buffer_discard_commit(struct trace_buffer *buffer,
3836 struct ring_buffer_event *event)
3838 struct ring_buffer_per_cpu *cpu_buffer;
3841 /* The event is discarded regardless */
3842 rb_event_discard(event);
3844 cpu = smp_processor_id();
3845 cpu_buffer = buffer->buffers[cpu];
3848 * This must only be called if the event has not been
3849 * committed yet. Thus we can assume that preemption
3850 * is still disabled.
3852 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3854 rb_decrement_entry(cpu_buffer, event);
3855 if (rb_try_to_discard(cpu_buffer, event))
3859 rb_end_commit(cpu_buffer);
3861 trace_recursive_unlock(cpu_buffer);
3863 preempt_enable_notrace();
3866 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3869 * ring_buffer_write - write data to the buffer without reserving
3870 * @buffer: The ring buffer to write to.
3871 * @length: The length of the data being written (excluding the event header)
3872 * @data: The data to write to the buffer.
3874 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3875 * one function. If you already have the data to write to the buffer, it
3876 * may be easier to simply call this function.
3878 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3879 * and not the length of the event which would hold the header.
3881 int ring_buffer_write(struct trace_buffer *buffer,
3882 unsigned long length,
3885 struct ring_buffer_per_cpu *cpu_buffer;
3886 struct ring_buffer_event *event;
3891 preempt_disable_notrace();
3893 if (atomic_read(&buffer->record_disabled))
3896 cpu = raw_smp_processor_id();
3898 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3901 cpu_buffer = buffer->buffers[cpu];
3903 if (atomic_read(&cpu_buffer->record_disabled))
3906 if (length > BUF_MAX_DATA_SIZE)
3909 if (unlikely(trace_recursive_lock(cpu_buffer)))
3912 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3916 body = rb_event_data(event);
3918 memcpy(body, data, length);
3920 rb_commit(cpu_buffer);
3922 rb_wakeups(buffer, cpu_buffer);
3927 trace_recursive_unlock(cpu_buffer);
3930 preempt_enable_notrace();
3934 EXPORT_SYMBOL_GPL(ring_buffer_write);
3936 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3938 struct buffer_page *reader = cpu_buffer->reader_page;
3939 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3940 struct buffer_page *commit = cpu_buffer->commit_page;
3942 /* In case of error, head will be NULL */
3943 if (unlikely(!head))
3946 /* Reader should exhaust content in reader page */
3947 if (reader->read != rb_page_commit(reader))
3951 * If writers are committing on the reader page, knowing all
3952 * committed content has been read, the ring buffer is empty.
3954 if (commit == reader)
3958 * If writers are committing on a page other than reader page
3959 * and head page, there should always be content to read.
3965 * Writers are committing on the head page, we just need
3966 * to care about there're committed data, and the reader will
3967 * swap reader page with head page when it is to read data.
3969 return rb_page_commit(commit) == 0;
3973 * ring_buffer_record_disable - stop all writes into the buffer
3974 * @buffer: The ring buffer to stop writes to.
3976 * This prevents all writes to the buffer. Any attempt to write
3977 * to the buffer after this will fail and return NULL.
3979 * The caller should call synchronize_rcu() after this.
3981 void ring_buffer_record_disable(struct trace_buffer *buffer)
3983 atomic_inc(&buffer->record_disabled);
3985 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3988 * ring_buffer_record_enable - enable writes to the buffer
3989 * @buffer: The ring buffer to enable writes
3991 * Note, multiple disables will need the same number of enables
3992 * to truly enable the writing (much like preempt_disable).
3994 void ring_buffer_record_enable(struct trace_buffer *buffer)
3996 atomic_dec(&buffer->record_disabled);
3998 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
4001 * ring_buffer_record_off - stop all writes into the buffer
4002 * @buffer: The ring buffer to stop writes to.
4004 * This prevents all writes to the buffer. Any attempt to write
4005 * to the buffer after this will fail and return NULL.
4007 * This is different than ring_buffer_record_disable() as
4008 * it works like an on/off switch, where as the disable() version
4009 * must be paired with a enable().
4011 void ring_buffer_record_off(struct trace_buffer *buffer)
4014 unsigned int new_rd;
4016 rd = atomic_read(&buffer->record_disabled);
4018 new_rd = rd | RB_BUFFER_OFF;
4019 } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
4021 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
4024 * ring_buffer_record_on - restart writes into the buffer
4025 * @buffer: The ring buffer to start writes to.
4027 * This enables all writes to the buffer that was disabled by
4028 * ring_buffer_record_off().
4030 * This is different than ring_buffer_record_enable() as
4031 * it works like an on/off switch, where as the enable() version
4032 * must be paired with a disable().
4034 void ring_buffer_record_on(struct trace_buffer *buffer)
4037 unsigned int new_rd;
4039 rd = atomic_read(&buffer->record_disabled);
4041 new_rd = rd & ~RB_BUFFER_OFF;
4042 } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
4044 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
4047 * ring_buffer_record_is_on - return true if the ring buffer can write
4048 * @buffer: The ring buffer to see if write is enabled
4050 * Returns true if the ring buffer is in a state that it accepts writes.
4052 bool ring_buffer_record_is_on(struct trace_buffer *buffer)
4054 return !atomic_read(&buffer->record_disabled);
4058 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
4059 * @buffer: The ring buffer to see if write is set enabled
4061 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
4062 * Note that this does NOT mean it is in a writable state.
4064 * It may return true when the ring buffer has been disabled by
4065 * ring_buffer_record_disable(), as that is a temporary disabling of
4068 bool ring_buffer_record_is_set_on(struct trace_buffer *buffer)
4070 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
4074 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
4075 * @buffer: The ring buffer to stop writes to.
4076 * @cpu: The CPU buffer to stop
4078 * This prevents all writes to the buffer. Any attempt to write
4079 * to the buffer after this will fail and return NULL.
4081 * The caller should call synchronize_rcu() after this.
4083 void ring_buffer_record_disable_cpu(struct trace_buffer *buffer, int cpu)
4085 struct ring_buffer_per_cpu *cpu_buffer;
4087 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4090 cpu_buffer = buffer->buffers[cpu];
4091 atomic_inc(&cpu_buffer->record_disabled);
4093 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
4096 * ring_buffer_record_enable_cpu - enable writes to the buffer
4097 * @buffer: The ring buffer to enable writes
4098 * @cpu: The CPU to enable.
4100 * Note, multiple disables will need the same number of enables
4101 * to truly enable the writing (much like preempt_disable).
4103 void ring_buffer_record_enable_cpu(struct trace_buffer *buffer, int cpu)
4105 struct ring_buffer_per_cpu *cpu_buffer;
4107 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4110 cpu_buffer = buffer->buffers[cpu];
4111 atomic_dec(&cpu_buffer->record_disabled);
4113 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
4116 * The total entries in the ring buffer is the running counter
4117 * of entries entered into the ring buffer, minus the sum of
4118 * the entries read from the ring buffer and the number of
4119 * entries that were overwritten.
4121 static inline unsigned long
4122 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
4124 return local_read(&cpu_buffer->entries) -
4125 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
4129 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
4130 * @buffer: The ring buffer
4131 * @cpu: The per CPU buffer to read from.
4133 u64 ring_buffer_oldest_event_ts(struct trace_buffer *buffer, int cpu)
4135 unsigned long flags;
4136 struct ring_buffer_per_cpu *cpu_buffer;
4137 struct buffer_page *bpage;
4140 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4143 cpu_buffer = buffer->buffers[cpu];
4144 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4146 * if the tail is on reader_page, oldest time stamp is on the reader
4149 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
4150 bpage = cpu_buffer->reader_page;
4152 bpage = rb_set_head_page(cpu_buffer);
4154 ret = bpage->page->time_stamp;
4155 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4159 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
4162 * ring_buffer_bytes_cpu - get the number of bytes unconsumed in a cpu buffer
4163 * @buffer: The ring buffer
4164 * @cpu: The per CPU buffer to read from.
4166 unsigned long ring_buffer_bytes_cpu(struct trace_buffer *buffer, int cpu)
4168 struct ring_buffer_per_cpu *cpu_buffer;
4171 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4174 cpu_buffer = buffer->buffers[cpu];
4175 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
4179 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
4182 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
4183 * @buffer: The ring buffer
4184 * @cpu: The per CPU buffer to get the entries from.
4186 unsigned long ring_buffer_entries_cpu(struct trace_buffer *buffer, int cpu)
4188 struct ring_buffer_per_cpu *cpu_buffer;
4190 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4193 cpu_buffer = buffer->buffers[cpu];
4195 return rb_num_of_entries(cpu_buffer);
4197 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
4200 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
4201 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
4202 * @buffer: The ring buffer
4203 * @cpu: The per CPU buffer to get the number of overruns from
4205 unsigned long ring_buffer_overrun_cpu(struct trace_buffer *buffer, int cpu)
4207 struct ring_buffer_per_cpu *cpu_buffer;
4210 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4213 cpu_buffer = buffer->buffers[cpu];
4214 ret = local_read(&cpu_buffer->overrun);
4218 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
4221 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
4222 * commits failing due to the buffer wrapping around while there are uncommitted
4223 * events, such as during an interrupt storm.
4224 * @buffer: The ring buffer
4225 * @cpu: The per CPU buffer to get the number of overruns from
4228 ring_buffer_commit_overrun_cpu(struct trace_buffer *buffer, int cpu)
4230 struct ring_buffer_per_cpu *cpu_buffer;
4233 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4236 cpu_buffer = buffer->buffers[cpu];
4237 ret = local_read(&cpu_buffer->commit_overrun);
4241 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
4244 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
4245 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
4246 * @buffer: The ring buffer
4247 * @cpu: The per CPU buffer to get the number of overruns from
4250 ring_buffer_dropped_events_cpu(struct trace_buffer *buffer, int cpu)
4252 struct ring_buffer_per_cpu *cpu_buffer;
4255 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4258 cpu_buffer = buffer->buffers[cpu];
4259 ret = local_read(&cpu_buffer->dropped_events);
4263 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
4266 * ring_buffer_read_events_cpu - get the number of events successfully read
4267 * @buffer: The ring buffer
4268 * @cpu: The per CPU buffer to get the number of events read
4271 ring_buffer_read_events_cpu(struct trace_buffer *buffer, int cpu)
4273 struct ring_buffer_per_cpu *cpu_buffer;
4275 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4278 cpu_buffer = buffer->buffers[cpu];
4279 return cpu_buffer->read;
4281 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
4284 * ring_buffer_entries - get the number of entries in a buffer
4285 * @buffer: The ring buffer
4287 * Returns the total number of entries in the ring buffer
4290 unsigned long ring_buffer_entries(struct trace_buffer *buffer)
4292 struct ring_buffer_per_cpu *cpu_buffer;
4293 unsigned long entries = 0;
4296 /* if you care about this being correct, lock the buffer */
4297 for_each_buffer_cpu(buffer, cpu) {
4298 cpu_buffer = buffer->buffers[cpu];
4299 entries += rb_num_of_entries(cpu_buffer);
4304 EXPORT_SYMBOL_GPL(ring_buffer_entries);
4307 * ring_buffer_overruns - get the number of overruns in buffer
4308 * @buffer: The ring buffer
4310 * Returns the total number of overruns in the ring buffer
4313 unsigned long ring_buffer_overruns(struct trace_buffer *buffer)
4315 struct ring_buffer_per_cpu *cpu_buffer;
4316 unsigned long overruns = 0;
4319 /* if you care about this being correct, lock the buffer */
4320 for_each_buffer_cpu(buffer, cpu) {
4321 cpu_buffer = buffer->buffers[cpu];
4322 overruns += local_read(&cpu_buffer->overrun);
4327 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
4329 static void rb_iter_reset(struct ring_buffer_iter *iter)
4331 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4333 /* Iterator usage is expected to have record disabled */
4334 iter->head_page = cpu_buffer->reader_page;
4335 iter->head = cpu_buffer->reader_page->read;
4336 iter->next_event = iter->head;
4338 iter->cache_reader_page = iter->head_page;
4339 iter->cache_read = cpu_buffer->read;
4340 iter->cache_pages_removed = cpu_buffer->pages_removed;
4343 iter->read_stamp = cpu_buffer->read_stamp;
4344 iter->page_stamp = cpu_buffer->reader_page->page->time_stamp;
4346 iter->read_stamp = iter->head_page->page->time_stamp;
4347 iter->page_stamp = iter->read_stamp;
4352 * ring_buffer_iter_reset - reset an iterator
4353 * @iter: The iterator to reset
4355 * Resets the iterator, so that it will start from the beginning
4358 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
4360 struct ring_buffer_per_cpu *cpu_buffer;
4361 unsigned long flags;
4366 cpu_buffer = iter->cpu_buffer;
4368 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4369 rb_iter_reset(iter);
4370 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4372 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
4375 * ring_buffer_iter_empty - check if an iterator has no more to read
4376 * @iter: The iterator to check
4378 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
4380 struct ring_buffer_per_cpu *cpu_buffer;
4381 struct buffer_page *reader;
4382 struct buffer_page *head_page;
4383 struct buffer_page *commit_page;
4384 struct buffer_page *curr_commit_page;
4389 cpu_buffer = iter->cpu_buffer;
4390 reader = cpu_buffer->reader_page;
4391 head_page = cpu_buffer->head_page;
4392 commit_page = cpu_buffer->commit_page;
4393 commit_ts = commit_page->page->time_stamp;
4396 * When the writer goes across pages, it issues a cmpxchg which
4397 * is a mb(), which will synchronize with the rmb here.
4398 * (see rb_tail_page_update())
4401 commit = rb_page_commit(commit_page);
4402 /* We want to make sure that the commit page doesn't change */
4405 /* Make sure commit page didn't change */
4406 curr_commit_page = READ_ONCE(cpu_buffer->commit_page);
4407 curr_commit_ts = READ_ONCE(curr_commit_page->page->time_stamp);
4409 /* If the commit page changed, then there's more data */
4410 if (curr_commit_page != commit_page ||
4411 curr_commit_ts != commit_ts)
4414 /* Still racy, as it may return a false positive, but that's OK */
4415 return ((iter->head_page == commit_page && iter->head >= commit) ||
4416 (iter->head_page == reader && commit_page == head_page &&
4417 head_page->read == commit &&
4418 iter->head == rb_page_commit(cpu_buffer->reader_page)));
4420 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
4423 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
4424 struct ring_buffer_event *event)
4428 switch (event->type_len) {
4429 case RINGBUF_TYPE_PADDING:
4432 case RINGBUF_TYPE_TIME_EXTEND:
4433 delta = rb_event_time_stamp(event);
4434 cpu_buffer->read_stamp += delta;
4437 case RINGBUF_TYPE_TIME_STAMP:
4438 delta = rb_event_time_stamp(event);
4439 delta = rb_fix_abs_ts(delta, cpu_buffer->read_stamp);
4440 cpu_buffer->read_stamp = delta;
4443 case RINGBUF_TYPE_DATA:
4444 cpu_buffer->read_stamp += event->time_delta;
4448 RB_WARN_ON(cpu_buffer, 1);
4453 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
4454 struct ring_buffer_event *event)
4458 switch (event->type_len) {
4459 case RINGBUF_TYPE_PADDING:
4462 case RINGBUF_TYPE_TIME_EXTEND:
4463 delta = rb_event_time_stamp(event);
4464 iter->read_stamp += delta;
4467 case RINGBUF_TYPE_TIME_STAMP:
4468 delta = rb_event_time_stamp(event);
4469 delta = rb_fix_abs_ts(delta, iter->read_stamp);
4470 iter->read_stamp = delta;
4473 case RINGBUF_TYPE_DATA:
4474 iter->read_stamp += event->time_delta;
4478 RB_WARN_ON(iter->cpu_buffer, 1);
4482 static struct buffer_page *
4483 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
4485 struct buffer_page *reader = NULL;
4486 unsigned long overwrite;
4487 unsigned long flags;
4491 local_irq_save(flags);
4492 arch_spin_lock(&cpu_buffer->lock);
4496 * This should normally only loop twice. But because the
4497 * start of the reader inserts an empty page, it causes
4498 * a case where we will loop three times. There should be no
4499 * reason to loop four times (that I know of).
4501 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
4506 reader = cpu_buffer->reader_page;
4508 /* If there's more to read, return this page */
4509 if (cpu_buffer->reader_page->read < rb_page_size(reader))
4512 /* Never should we have an index greater than the size */
4513 if (RB_WARN_ON(cpu_buffer,
4514 cpu_buffer->reader_page->read > rb_page_size(reader)))
4517 /* check if we caught up to the tail */
4519 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
4522 /* Don't bother swapping if the ring buffer is empty */
4523 if (rb_num_of_entries(cpu_buffer) == 0)
4527 * Reset the reader page to size zero.
4529 local_set(&cpu_buffer->reader_page->write, 0);
4530 local_set(&cpu_buffer->reader_page->entries, 0);
4531 local_set(&cpu_buffer->reader_page->page->commit, 0);
4532 cpu_buffer->reader_page->real_end = 0;
4536 * Splice the empty reader page into the list around the head.
4538 reader = rb_set_head_page(cpu_buffer);
4541 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
4542 cpu_buffer->reader_page->list.prev = reader->list.prev;
4545 * cpu_buffer->pages just needs to point to the buffer, it
4546 * has no specific buffer page to point to. Lets move it out
4547 * of our way so we don't accidentally swap it.
4549 cpu_buffer->pages = reader->list.prev;
4551 /* The reader page will be pointing to the new head */
4552 rb_set_list_to_head(&cpu_buffer->reader_page->list);
4555 * We want to make sure we read the overruns after we set up our
4556 * pointers to the next object. The writer side does a
4557 * cmpxchg to cross pages which acts as the mb on the writer
4558 * side. Note, the reader will constantly fail the swap
4559 * while the writer is updating the pointers, so this
4560 * guarantees that the overwrite recorded here is the one we
4561 * want to compare with the last_overrun.
4564 overwrite = local_read(&(cpu_buffer->overrun));
4567 * Here's the tricky part.
4569 * We need to move the pointer past the header page.
4570 * But we can only do that if a writer is not currently
4571 * moving it. The page before the header page has the
4572 * flag bit '1' set if it is pointing to the page we want.
4573 * but if the writer is in the process of moving it
4574 * than it will be '2' or already moved '0'.
4577 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
4580 * If we did not convert it, then we must try again.
4586 * Yay! We succeeded in replacing the page.
4588 * Now make the new head point back to the reader page.
4590 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
4591 rb_inc_page(&cpu_buffer->head_page);
4593 local_inc(&cpu_buffer->pages_read);
4595 /* Finally update the reader page to the new head */
4596 cpu_buffer->reader_page = reader;
4597 cpu_buffer->reader_page->read = 0;
4599 if (overwrite != cpu_buffer->last_overrun) {
4600 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
4601 cpu_buffer->last_overrun = overwrite;
4607 /* Update the read_stamp on the first event */
4608 if (reader && reader->read == 0)
4609 cpu_buffer->read_stamp = reader->page->time_stamp;
4611 arch_spin_unlock(&cpu_buffer->lock);
4612 local_irq_restore(flags);
4615 * The writer has preempt disable, wait for it. But not forever
4616 * Although, 1 second is pretty much "forever"
4618 #define USECS_WAIT 1000000
4619 for (nr_loops = 0; nr_loops < USECS_WAIT; nr_loops++) {
4620 /* If the write is past the end of page, a writer is still updating it */
4621 if (likely(!reader || rb_page_write(reader) <= BUF_PAGE_SIZE))
4626 /* Get the latest version of the reader write value */
4630 /* The writer is not moving forward? Something is wrong */
4631 if (RB_WARN_ON(cpu_buffer, nr_loops == USECS_WAIT))
4635 * Make sure we see any padding after the write update
4636 * (see rb_reset_tail()).
4638 * In addition, a writer may be writing on the reader page
4639 * if the page has not been fully filled, so the read barrier
4640 * is also needed to make sure we see the content of what is
4641 * committed by the writer (see rb_set_commit_to_write()).
4649 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
4651 struct ring_buffer_event *event;
4652 struct buffer_page *reader;
4655 reader = rb_get_reader_page(cpu_buffer);
4657 /* This function should not be called when buffer is empty */
4658 if (RB_WARN_ON(cpu_buffer, !reader))
4661 event = rb_reader_event(cpu_buffer);
4663 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
4666 rb_update_read_stamp(cpu_buffer, event);
4668 length = rb_event_length(event);
4669 cpu_buffer->reader_page->read += length;
4670 cpu_buffer->read_bytes += length;
4673 static void rb_advance_iter(struct ring_buffer_iter *iter)
4675 struct ring_buffer_per_cpu *cpu_buffer;
4677 cpu_buffer = iter->cpu_buffer;
4679 /* If head == next_event then we need to jump to the next event */
4680 if (iter->head == iter->next_event) {
4681 /* If the event gets overwritten again, there's nothing to do */
4682 if (rb_iter_head_event(iter) == NULL)
4686 iter->head = iter->next_event;
4689 * Check if we are at the end of the buffer.
4691 if (iter->next_event >= rb_page_size(iter->head_page)) {
4692 /* discarded commits can make the page empty */
4693 if (iter->head_page == cpu_buffer->commit_page)
4699 rb_update_iter_read_stamp(iter, iter->event);
4702 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
4704 return cpu_buffer->lost_events;
4707 static struct ring_buffer_event *
4708 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
4709 unsigned long *lost_events)
4711 struct ring_buffer_event *event;
4712 struct buffer_page *reader;
4719 * We repeat when a time extend is encountered.
4720 * Since the time extend is always attached to a data event,
4721 * we should never loop more than once.
4722 * (We never hit the following condition more than twice).
4724 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
4727 reader = rb_get_reader_page(cpu_buffer);
4731 event = rb_reader_event(cpu_buffer);
4733 switch (event->type_len) {
4734 case RINGBUF_TYPE_PADDING:
4735 if (rb_null_event(event))
4736 RB_WARN_ON(cpu_buffer, 1);
4738 * Because the writer could be discarding every
4739 * event it creates (which would probably be bad)
4740 * if we were to go back to "again" then we may never
4741 * catch up, and will trigger the warn on, or lock
4742 * the box. Return the padding, and we will release
4743 * the current locks, and try again.
4747 case RINGBUF_TYPE_TIME_EXTEND:
4748 /* Internal data, OK to advance */
4749 rb_advance_reader(cpu_buffer);
4752 case RINGBUF_TYPE_TIME_STAMP:
4754 *ts = rb_event_time_stamp(event);
4755 *ts = rb_fix_abs_ts(*ts, reader->page->time_stamp);
4756 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4757 cpu_buffer->cpu, ts);
4759 /* Internal data, OK to advance */
4760 rb_advance_reader(cpu_buffer);
4763 case RINGBUF_TYPE_DATA:
4765 *ts = cpu_buffer->read_stamp + event->time_delta;
4766 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4767 cpu_buffer->cpu, ts);
4770 *lost_events = rb_lost_events(cpu_buffer);
4774 RB_WARN_ON(cpu_buffer, 1);
4779 EXPORT_SYMBOL_GPL(ring_buffer_peek);
4781 static struct ring_buffer_event *
4782 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4784 struct trace_buffer *buffer;
4785 struct ring_buffer_per_cpu *cpu_buffer;
4786 struct ring_buffer_event *event;
4792 cpu_buffer = iter->cpu_buffer;
4793 buffer = cpu_buffer->buffer;
4796 * Check if someone performed a consuming read to the buffer
4797 * or removed some pages from the buffer. In these cases,
4798 * iterator was invalidated and we need to reset it.
4800 if (unlikely(iter->cache_read != cpu_buffer->read ||
4801 iter->cache_reader_page != cpu_buffer->reader_page ||
4802 iter->cache_pages_removed != cpu_buffer->pages_removed))
4803 rb_iter_reset(iter);
4806 if (ring_buffer_iter_empty(iter))
4810 * As the writer can mess with what the iterator is trying
4811 * to read, just give up if we fail to get an event after
4812 * three tries. The iterator is not as reliable when reading
4813 * the ring buffer with an active write as the consumer is.
4814 * Do not warn if the three failures is reached.
4819 if (rb_per_cpu_empty(cpu_buffer))
4822 if (iter->head >= rb_page_size(iter->head_page)) {
4827 event = rb_iter_head_event(iter);
4831 switch (event->type_len) {
4832 case RINGBUF_TYPE_PADDING:
4833 if (rb_null_event(event)) {
4837 rb_advance_iter(iter);
4840 case RINGBUF_TYPE_TIME_EXTEND:
4841 /* Internal data, OK to advance */
4842 rb_advance_iter(iter);
4845 case RINGBUF_TYPE_TIME_STAMP:
4847 *ts = rb_event_time_stamp(event);
4848 *ts = rb_fix_abs_ts(*ts, iter->head_page->page->time_stamp);
4849 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4850 cpu_buffer->cpu, ts);
4852 /* Internal data, OK to advance */
4853 rb_advance_iter(iter);
4856 case RINGBUF_TYPE_DATA:
4858 *ts = iter->read_stamp + event->time_delta;
4859 ring_buffer_normalize_time_stamp(buffer,
4860 cpu_buffer->cpu, ts);
4865 RB_WARN_ON(cpu_buffer, 1);
4870 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4872 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4874 if (likely(!in_nmi())) {
4875 raw_spin_lock(&cpu_buffer->reader_lock);
4880 * If an NMI die dumps out the content of the ring buffer
4881 * trylock must be used to prevent a deadlock if the NMI
4882 * preempted a task that holds the ring buffer locks. If
4883 * we get the lock then all is fine, if not, then continue
4884 * to do the read, but this can corrupt the ring buffer,
4885 * so it must be permanently disabled from future writes.
4886 * Reading from NMI is a oneshot deal.
4888 if (raw_spin_trylock(&cpu_buffer->reader_lock))
4891 /* Continue without locking, but disable the ring buffer */
4892 atomic_inc(&cpu_buffer->record_disabled);
4897 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4900 raw_spin_unlock(&cpu_buffer->reader_lock);
4904 * ring_buffer_peek - peek at the next event to be read
4905 * @buffer: The ring buffer to read
4906 * @cpu: The cpu to peak at
4907 * @ts: The timestamp counter of this event.
4908 * @lost_events: a variable to store if events were lost (may be NULL)
4910 * This will return the event that will be read next, but does
4911 * not consume the data.
4913 struct ring_buffer_event *
4914 ring_buffer_peek(struct trace_buffer *buffer, int cpu, u64 *ts,
4915 unsigned long *lost_events)
4917 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4918 struct ring_buffer_event *event;
4919 unsigned long flags;
4922 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4926 local_irq_save(flags);
4927 dolock = rb_reader_lock(cpu_buffer);
4928 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4929 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4930 rb_advance_reader(cpu_buffer);
4931 rb_reader_unlock(cpu_buffer, dolock);
4932 local_irq_restore(flags);
4934 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4940 /** ring_buffer_iter_dropped - report if there are dropped events
4941 * @iter: The ring buffer iterator
4943 * Returns true if there was dropped events since the last peek.
4945 bool ring_buffer_iter_dropped(struct ring_buffer_iter *iter)
4947 bool ret = iter->missed_events != 0;
4949 iter->missed_events = 0;
4952 EXPORT_SYMBOL_GPL(ring_buffer_iter_dropped);
4955 * ring_buffer_iter_peek - peek at the next event to be read
4956 * @iter: The ring buffer iterator
4957 * @ts: The timestamp counter of this event.
4959 * This will return the event that will be read next, but does
4960 * not increment the iterator.
4962 struct ring_buffer_event *
4963 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4965 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4966 struct ring_buffer_event *event;
4967 unsigned long flags;
4970 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4971 event = rb_iter_peek(iter, ts);
4972 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4974 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4981 * ring_buffer_consume - return an event and consume it
4982 * @buffer: The ring buffer to get the next event from
4983 * @cpu: the cpu to read the buffer from
4984 * @ts: a variable to store the timestamp (may be NULL)
4985 * @lost_events: a variable to store if events were lost (may be NULL)
4987 * Returns the next event in the ring buffer, and that event is consumed.
4988 * Meaning, that sequential reads will keep returning a different event,
4989 * and eventually empty the ring buffer if the producer is slower.
4991 struct ring_buffer_event *
4992 ring_buffer_consume(struct trace_buffer *buffer, int cpu, u64 *ts,
4993 unsigned long *lost_events)
4995 struct ring_buffer_per_cpu *cpu_buffer;
4996 struct ring_buffer_event *event = NULL;
4997 unsigned long flags;
5001 /* might be called in atomic */
5004 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5007 cpu_buffer = buffer->buffers[cpu];
5008 local_irq_save(flags);
5009 dolock = rb_reader_lock(cpu_buffer);
5011 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
5013 cpu_buffer->lost_events = 0;
5014 rb_advance_reader(cpu_buffer);
5017 rb_reader_unlock(cpu_buffer, dolock);
5018 local_irq_restore(flags);
5023 if (event && event->type_len == RINGBUF_TYPE_PADDING)
5028 EXPORT_SYMBOL_GPL(ring_buffer_consume);
5031 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
5032 * @buffer: The ring buffer to read from
5033 * @cpu: The cpu buffer to iterate over
5034 * @flags: gfp flags to use for memory allocation
5036 * This performs the initial preparations necessary to iterate
5037 * through the buffer. Memory is allocated, buffer recording
5038 * is disabled, and the iterator pointer is returned to the caller.
5040 * Disabling buffer recording prevents the reading from being
5041 * corrupted. This is not a consuming read, so a producer is not
5044 * After a sequence of ring_buffer_read_prepare calls, the user is
5045 * expected to make at least one call to ring_buffer_read_prepare_sync.
5046 * Afterwards, ring_buffer_read_start is invoked to get things going
5049 * This overall must be paired with ring_buffer_read_finish.
5051 struct ring_buffer_iter *
5052 ring_buffer_read_prepare(struct trace_buffer *buffer, int cpu, gfp_t flags)
5054 struct ring_buffer_per_cpu *cpu_buffer;
5055 struct ring_buffer_iter *iter;
5057 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5060 iter = kzalloc(sizeof(*iter), flags);
5064 /* Holds the entire event: data and meta data */
5065 iter->event = kmalloc(BUF_PAGE_SIZE, flags);
5071 cpu_buffer = buffer->buffers[cpu];
5073 iter->cpu_buffer = cpu_buffer;
5075 atomic_inc(&cpu_buffer->resize_disabled);
5079 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
5082 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
5084 * All previously invoked ring_buffer_read_prepare calls to prepare
5085 * iterators will be synchronized. Afterwards, read_buffer_read_start
5086 * calls on those iterators are allowed.
5089 ring_buffer_read_prepare_sync(void)
5093 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
5096 * ring_buffer_read_start - start a non consuming read of the buffer
5097 * @iter: The iterator returned by ring_buffer_read_prepare
5099 * This finalizes the startup of an iteration through the buffer.
5100 * The iterator comes from a call to ring_buffer_read_prepare and
5101 * an intervening ring_buffer_read_prepare_sync must have been
5104 * Must be paired with ring_buffer_read_finish.
5107 ring_buffer_read_start(struct ring_buffer_iter *iter)
5109 struct ring_buffer_per_cpu *cpu_buffer;
5110 unsigned long flags;
5115 cpu_buffer = iter->cpu_buffer;
5117 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5118 arch_spin_lock(&cpu_buffer->lock);
5119 rb_iter_reset(iter);
5120 arch_spin_unlock(&cpu_buffer->lock);
5121 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5123 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
5126 * ring_buffer_read_finish - finish reading the iterator of the buffer
5127 * @iter: The iterator retrieved by ring_buffer_start
5129 * This re-enables the recording to the buffer, and frees the
5133 ring_buffer_read_finish(struct ring_buffer_iter *iter)
5135 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5136 unsigned long flags;
5139 * Ring buffer is disabled from recording, here's a good place
5140 * to check the integrity of the ring buffer.
5141 * Must prevent readers from trying to read, as the check
5142 * clears the HEAD page and readers require it.
5144 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5145 rb_check_pages(cpu_buffer);
5146 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5148 atomic_dec(&cpu_buffer->resize_disabled);
5152 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
5155 * ring_buffer_iter_advance - advance the iterator to the next location
5156 * @iter: The ring buffer iterator
5158 * Move the location of the iterator such that the next read will
5159 * be the next location of the iterator.
5161 void ring_buffer_iter_advance(struct ring_buffer_iter *iter)
5163 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5164 unsigned long flags;
5166 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5168 rb_advance_iter(iter);
5170 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5172 EXPORT_SYMBOL_GPL(ring_buffer_iter_advance);
5175 * ring_buffer_size - return the size of the ring buffer (in bytes)
5176 * @buffer: The ring buffer.
5177 * @cpu: The CPU to get ring buffer size from.
5179 unsigned long ring_buffer_size(struct trace_buffer *buffer, int cpu)
5182 * Earlier, this method returned
5183 * BUF_PAGE_SIZE * buffer->nr_pages
5184 * Since the nr_pages field is now removed, we have converted this to
5185 * return the per cpu buffer value.
5187 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5190 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
5192 EXPORT_SYMBOL_GPL(ring_buffer_size);
5194 static void rb_clear_buffer_page(struct buffer_page *page)
5196 local_set(&page->write, 0);
5197 local_set(&page->entries, 0);
5198 rb_init_page(page->page);
5203 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
5205 struct buffer_page *page;
5207 rb_head_page_deactivate(cpu_buffer);
5209 cpu_buffer->head_page
5210 = list_entry(cpu_buffer->pages, struct buffer_page, list);
5211 rb_clear_buffer_page(cpu_buffer->head_page);
5212 list_for_each_entry(page, cpu_buffer->pages, list) {
5213 rb_clear_buffer_page(page);
5216 cpu_buffer->tail_page = cpu_buffer->head_page;
5217 cpu_buffer->commit_page = cpu_buffer->head_page;
5219 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
5220 INIT_LIST_HEAD(&cpu_buffer->new_pages);
5221 rb_clear_buffer_page(cpu_buffer->reader_page);
5223 local_set(&cpu_buffer->entries_bytes, 0);
5224 local_set(&cpu_buffer->overrun, 0);
5225 local_set(&cpu_buffer->commit_overrun, 0);
5226 local_set(&cpu_buffer->dropped_events, 0);
5227 local_set(&cpu_buffer->entries, 0);
5228 local_set(&cpu_buffer->committing, 0);
5229 local_set(&cpu_buffer->commits, 0);
5230 local_set(&cpu_buffer->pages_touched, 0);
5231 local_set(&cpu_buffer->pages_lost, 0);
5232 local_set(&cpu_buffer->pages_read, 0);
5233 cpu_buffer->last_pages_touch = 0;
5234 cpu_buffer->shortest_full = 0;
5235 cpu_buffer->read = 0;
5236 cpu_buffer->read_bytes = 0;
5238 rb_time_set(&cpu_buffer->write_stamp, 0);
5239 rb_time_set(&cpu_buffer->before_stamp, 0);
5241 memset(cpu_buffer->event_stamp, 0, sizeof(cpu_buffer->event_stamp));
5243 cpu_buffer->lost_events = 0;
5244 cpu_buffer->last_overrun = 0;
5246 rb_head_page_activate(cpu_buffer);
5247 cpu_buffer->pages_removed = 0;
5250 /* Must have disabled the cpu buffer then done a synchronize_rcu */
5251 static void reset_disabled_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
5253 unsigned long flags;
5255 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5257 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
5260 arch_spin_lock(&cpu_buffer->lock);
5262 rb_reset_cpu(cpu_buffer);
5264 arch_spin_unlock(&cpu_buffer->lock);
5267 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5271 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
5272 * @buffer: The ring buffer to reset a per cpu buffer of
5273 * @cpu: The CPU buffer to be reset
5275 void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
5277 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5279 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5282 /* prevent another thread from changing buffer sizes */
5283 mutex_lock(&buffer->mutex);
5285 atomic_inc(&cpu_buffer->resize_disabled);
5286 atomic_inc(&cpu_buffer->record_disabled);
5288 /* Make sure all commits have finished */
5291 reset_disabled_cpu_buffer(cpu_buffer);
5293 atomic_dec(&cpu_buffer->record_disabled);
5294 atomic_dec(&cpu_buffer->resize_disabled);
5296 mutex_unlock(&buffer->mutex);
5298 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
5300 /* Flag to ensure proper resetting of atomic variables */
5301 #define RESET_BIT (1 << 30)
5304 * ring_buffer_reset_online_cpus - reset a ring buffer per CPU buffer
5305 * @buffer: The ring buffer to reset a per cpu buffer of
5307 void ring_buffer_reset_online_cpus(struct trace_buffer *buffer)
5309 struct ring_buffer_per_cpu *cpu_buffer;
5312 /* prevent another thread from changing buffer sizes */
5313 mutex_lock(&buffer->mutex);
5315 for_each_online_buffer_cpu(buffer, cpu) {
5316 cpu_buffer = buffer->buffers[cpu];
5318 atomic_add(RESET_BIT, &cpu_buffer->resize_disabled);
5319 atomic_inc(&cpu_buffer->record_disabled);
5322 /* Make sure all commits have finished */
5325 for_each_buffer_cpu(buffer, cpu) {
5326 cpu_buffer = buffer->buffers[cpu];
5329 * If a CPU came online during the synchronize_rcu(), then
5332 if (!(atomic_read(&cpu_buffer->resize_disabled) & RESET_BIT))
5335 reset_disabled_cpu_buffer(cpu_buffer);
5337 atomic_dec(&cpu_buffer->record_disabled);
5338 atomic_sub(RESET_BIT, &cpu_buffer->resize_disabled);
5341 mutex_unlock(&buffer->mutex);
5345 * ring_buffer_reset - reset a ring buffer
5346 * @buffer: The ring buffer to reset all cpu buffers
5348 void ring_buffer_reset(struct trace_buffer *buffer)
5350 struct ring_buffer_per_cpu *cpu_buffer;
5353 /* prevent another thread from changing buffer sizes */
5354 mutex_lock(&buffer->mutex);
5356 for_each_buffer_cpu(buffer, cpu) {
5357 cpu_buffer = buffer->buffers[cpu];
5359 atomic_inc(&cpu_buffer->resize_disabled);
5360 atomic_inc(&cpu_buffer->record_disabled);
5363 /* Make sure all commits have finished */
5366 for_each_buffer_cpu(buffer, cpu) {
5367 cpu_buffer = buffer->buffers[cpu];
5369 reset_disabled_cpu_buffer(cpu_buffer);
5371 atomic_dec(&cpu_buffer->record_disabled);
5372 atomic_dec(&cpu_buffer->resize_disabled);
5375 mutex_unlock(&buffer->mutex);
5377 EXPORT_SYMBOL_GPL(ring_buffer_reset);
5380 * ring_buffer_empty - is the ring buffer empty?
5381 * @buffer: The ring buffer to test
5383 bool ring_buffer_empty(struct trace_buffer *buffer)
5385 struct ring_buffer_per_cpu *cpu_buffer;
5386 unsigned long flags;
5391 /* yes this is racy, but if you don't like the race, lock the buffer */
5392 for_each_buffer_cpu(buffer, cpu) {
5393 cpu_buffer = buffer->buffers[cpu];
5394 local_irq_save(flags);
5395 dolock = rb_reader_lock(cpu_buffer);
5396 ret = rb_per_cpu_empty(cpu_buffer);
5397 rb_reader_unlock(cpu_buffer, dolock);
5398 local_irq_restore(flags);
5406 EXPORT_SYMBOL_GPL(ring_buffer_empty);
5409 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
5410 * @buffer: The ring buffer
5411 * @cpu: The CPU buffer to test
5413 bool ring_buffer_empty_cpu(struct trace_buffer *buffer, int cpu)
5415 struct ring_buffer_per_cpu *cpu_buffer;
5416 unsigned long flags;
5420 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5423 cpu_buffer = buffer->buffers[cpu];
5424 local_irq_save(flags);
5425 dolock = rb_reader_lock(cpu_buffer);
5426 ret = rb_per_cpu_empty(cpu_buffer);
5427 rb_reader_unlock(cpu_buffer, dolock);
5428 local_irq_restore(flags);
5432 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
5434 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
5436 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
5437 * @buffer_a: One buffer to swap with
5438 * @buffer_b: The other buffer to swap with
5439 * @cpu: the CPU of the buffers to swap
5441 * This function is useful for tracers that want to take a "snapshot"
5442 * of a CPU buffer and has another back up buffer lying around.
5443 * it is expected that the tracer handles the cpu buffer not being
5444 * used at the moment.
5446 int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
5447 struct trace_buffer *buffer_b, int cpu)
5449 struct ring_buffer_per_cpu *cpu_buffer_a;
5450 struct ring_buffer_per_cpu *cpu_buffer_b;
5453 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
5454 !cpumask_test_cpu(cpu, buffer_b->cpumask))
5457 cpu_buffer_a = buffer_a->buffers[cpu];
5458 cpu_buffer_b = buffer_b->buffers[cpu];
5460 /* At least make sure the two buffers are somewhat the same */
5461 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
5466 if (atomic_read(&buffer_a->record_disabled))
5469 if (atomic_read(&buffer_b->record_disabled))
5472 if (atomic_read(&cpu_buffer_a->record_disabled))
5475 if (atomic_read(&cpu_buffer_b->record_disabled))
5479 * We can't do a synchronize_rcu here because this
5480 * function can be called in atomic context.
5481 * Normally this will be called from the same CPU as cpu.
5482 * If not it's up to the caller to protect this.
5484 atomic_inc(&cpu_buffer_a->record_disabled);
5485 atomic_inc(&cpu_buffer_b->record_disabled);
5488 if (local_read(&cpu_buffer_a->committing))
5490 if (local_read(&cpu_buffer_b->committing))
5494 * When resize is in progress, we cannot swap it because
5495 * it will mess the state of the cpu buffer.
5497 if (atomic_read(&buffer_a->resizing))
5499 if (atomic_read(&buffer_b->resizing))
5502 buffer_a->buffers[cpu] = cpu_buffer_b;
5503 buffer_b->buffers[cpu] = cpu_buffer_a;
5505 cpu_buffer_b->buffer = buffer_a;
5506 cpu_buffer_a->buffer = buffer_b;
5511 atomic_dec(&cpu_buffer_a->record_disabled);
5512 atomic_dec(&cpu_buffer_b->record_disabled);
5516 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
5517 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
5520 * ring_buffer_alloc_read_page - allocate a page to read from buffer
5521 * @buffer: the buffer to allocate for.
5522 * @cpu: the cpu buffer to allocate.
5524 * This function is used in conjunction with ring_buffer_read_page.
5525 * When reading a full page from the ring buffer, these functions
5526 * can be used to speed up the process. The calling function should
5527 * allocate a few pages first with this function. Then when it
5528 * needs to get pages from the ring buffer, it passes the result
5529 * of this function into ring_buffer_read_page, which will swap
5530 * the page that was allocated, with the read page of the buffer.
5533 * The page allocated, or ERR_PTR
5535 void *ring_buffer_alloc_read_page(struct trace_buffer *buffer, int cpu)
5537 struct ring_buffer_per_cpu *cpu_buffer;
5538 struct buffer_data_page *bpage = NULL;
5539 unsigned long flags;
5542 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5543 return ERR_PTR(-ENODEV);
5545 cpu_buffer = buffer->buffers[cpu];
5546 local_irq_save(flags);
5547 arch_spin_lock(&cpu_buffer->lock);
5549 if (cpu_buffer->free_page) {
5550 bpage = cpu_buffer->free_page;
5551 cpu_buffer->free_page = NULL;
5554 arch_spin_unlock(&cpu_buffer->lock);
5555 local_irq_restore(flags);
5560 page = alloc_pages_node(cpu_to_node(cpu),
5561 GFP_KERNEL | __GFP_NORETRY, 0);
5563 return ERR_PTR(-ENOMEM);
5565 bpage = page_address(page);
5568 rb_init_page(bpage);
5572 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
5575 * ring_buffer_free_read_page - free an allocated read page
5576 * @buffer: the buffer the page was allocate for
5577 * @cpu: the cpu buffer the page came from
5578 * @data: the page to free
5580 * Free a page allocated from ring_buffer_alloc_read_page.
5582 void ring_buffer_free_read_page(struct trace_buffer *buffer, int cpu, void *data)
5584 struct ring_buffer_per_cpu *cpu_buffer;
5585 struct buffer_data_page *bpage = data;
5586 struct page *page = virt_to_page(bpage);
5587 unsigned long flags;
5589 if (!buffer || !buffer->buffers || !buffer->buffers[cpu])
5592 cpu_buffer = buffer->buffers[cpu];
5594 /* If the page is still in use someplace else, we can't reuse it */
5595 if (page_ref_count(page) > 1)
5598 local_irq_save(flags);
5599 arch_spin_lock(&cpu_buffer->lock);
5601 if (!cpu_buffer->free_page) {
5602 cpu_buffer->free_page = bpage;
5606 arch_spin_unlock(&cpu_buffer->lock);
5607 local_irq_restore(flags);
5610 free_page((unsigned long)bpage);
5612 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
5615 * ring_buffer_read_page - extract a page from the ring buffer
5616 * @buffer: buffer to extract from
5617 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
5618 * @len: amount to extract
5619 * @cpu: the cpu of the buffer to extract
5620 * @full: should the extraction only happen when the page is full.
5622 * This function will pull out a page from the ring buffer and consume it.
5623 * @data_page must be the address of the variable that was returned
5624 * from ring_buffer_alloc_read_page. This is because the page might be used
5625 * to swap with a page in the ring buffer.
5628 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
5629 * if (IS_ERR(rpage))
5630 * return PTR_ERR(rpage);
5631 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
5633 * process_page(rpage, ret);
5635 * When @full is set, the function will not return true unless
5636 * the writer is off the reader page.
5638 * Note: it is up to the calling functions to handle sleeps and wakeups.
5639 * The ring buffer can be used anywhere in the kernel and can not
5640 * blindly call wake_up. The layer that uses the ring buffer must be
5641 * responsible for that.
5644 * >=0 if data has been transferred, returns the offset of consumed data.
5645 * <0 if no data has been transferred.
5647 int ring_buffer_read_page(struct trace_buffer *buffer,
5648 void **data_page, size_t len, int cpu, int full)
5650 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5651 struct ring_buffer_event *event;
5652 struct buffer_data_page *bpage;
5653 struct buffer_page *reader;
5654 unsigned long missed_events;
5655 unsigned long flags;
5656 unsigned int commit;
5661 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5665 * If len is not big enough to hold the page header, then
5666 * we can not copy anything.
5668 if (len <= BUF_PAGE_HDR_SIZE)
5671 len -= BUF_PAGE_HDR_SIZE;
5680 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5682 reader = rb_get_reader_page(cpu_buffer);
5686 event = rb_reader_event(cpu_buffer);
5688 read = reader->read;
5689 commit = rb_page_commit(reader);
5691 /* Check if any events were dropped */
5692 missed_events = cpu_buffer->lost_events;
5695 * If this page has been partially read or
5696 * if len is not big enough to read the rest of the page or
5697 * a writer is still on the page, then
5698 * we must copy the data from the page to the buffer.
5699 * Otherwise, we can simply swap the page with the one passed in.
5701 if (read || (len < (commit - read)) ||
5702 cpu_buffer->reader_page == cpu_buffer->commit_page) {
5703 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
5704 unsigned int rpos = read;
5705 unsigned int pos = 0;
5709 * If a full page is expected, this can still be returned
5710 * if there's been a previous partial read and the
5711 * rest of the page can be read and the commit page is off
5715 (!read || (len < (commit - read)) ||
5716 cpu_buffer->reader_page == cpu_buffer->commit_page))
5719 if (len > (commit - read))
5720 len = (commit - read);
5722 /* Always keep the time extend and data together */
5723 size = rb_event_ts_length(event);
5728 /* save the current timestamp, since the user will need it */
5729 save_timestamp = cpu_buffer->read_stamp;
5731 /* Need to copy one event at a time */
5733 /* We need the size of one event, because
5734 * rb_advance_reader only advances by one event,
5735 * whereas rb_event_ts_length may include the size of
5736 * one or two events.
5737 * We have already ensured there's enough space if this
5738 * is a time extend. */
5739 size = rb_event_length(event);
5740 memcpy(bpage->data + pos, rpage->data + rpos, size);
5744 rb_advance_reader(cpu_buffer);
5745 rpos = reader->read;
5751 event = rb_reader_event(cpu_buffer);
5752 /* Always keep the time extend and data together */
5753 size = rb_event_ts_length(event);
5754 } while (len >= size);
5757 local_set(&bpage->commit, pos);
5758 bpage->time_stamp = save_timestamp;
5760 /* we copied everything to the beginning */
5763 /* update the entry counter */
5764 cpu_buffer->read += rb_page_entries(reader);
5765 cpu_buffer->read_bytes += rb_page_commit(reader);
5767 /* swap the pages */
5768 rb_init_page(bpage);
5769 bpage = reader->page;
5770 reader->page = *data_page;
5771 local_set(&reader->write, 0);
5772 local_set(&reader->entries, 0);
5777 * Use the real_end for the data size,
5778 * This gives us a chance to store the lost events
5781 if (reader->real_end)
5782 local_set(&bpage->commit, reader->real_end);
5786 cpu_buffer->lost_events = 0;
5788 commit = local_read(&bpage->commit);
5790 * Set a flag in the commit field if we lost events
5792 if (missed_events) {
5793 /* If there is room at the end of the page to save the
5794 * missed events, then record it there.
5796 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
5797 memcpy(&bpage->data[commit], &missed_events,
5798 sizeof(missed_events));
5799 local_add(RB_MISSED_STORED, &bpage->commit);
5800 commit += sizeof(missed_events);
5802 local_add(RB_MISSED_EVENTS, &bpage->commit);
5806 * This page may be off to user land. Zero it out here.
5808 if (commit < BUF_PAGE_SIZE)
5809 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
5812 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5817 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
5820 * We only allocate new buffers, never free them if the CPU goes down.
5821 * If we were to free the buffer, then the user would lose any trace that was in
5824 int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
5826 struct trace_buffer *buffer;
5829 unsigned long nr_pages;
5831 buffer = container_of(node, struct trace_buffer, node);
5832 if (cpumask_test_cpu(cpu, buffer->cpumask))
5837 /* check if all cpu sizes are same */
5838 for_each_buffer_cpu(buffer, cpu_i) {
5839 /* fill in the size from first enabled cpu */
5841 nr_pages = buffer->buffers[cpu_i]->nr_pages;
5842 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
5847 /* allocate minimum pages, user can later expand it */
5850 buffer->buffers[cpu] =
5851 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
5852 if (!buffer->buffers[cpu]) {
5853 WARN(1, "failed to allocate ring buffer on CPU %u\n",
5858 cpumask_set_cpu(cpu, buffer->cpumask);
5862 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
5864 * This is a basic integrity check of the ring buffer.
5865 * Late in the boot cycle this test will run when configured in.
5866 * It will kick off a thread per CPU that will go into a loop
5867 * writing to the per cpu ring buffer various sizes of data.
5868 * Some of the data will be large items, some small.
5870 * Another thread is created that goes into a spin, sending out
5871 * IPIs to the other CPUs to also write into the ring buffer.
5872 * this is to test the nesting ability of the buffer.
5874 * Basic stats are recorded and reported. If something in the
5875 * ring buffer should happen that's not expected, a big warning
5876 * is displayed and all ring buffers are disabled.
5878 static struct task_struct *rb_threads[NR_CPUS] __initdata;
5880 struct rb_test_data {
5881 struct trace_buffer *buffer;
5882 unsigned long events;
5883 unsigned long bytes_written;
5884 unsigned long bytes_alloc;
5885 unsigned long bytes_dropped;
5886 unsigned long events_nested;
5887 unsigned long bytes_written_nested;
5888 unsigned long bytes_alloc_nested;
5889 unsigned long bytes_dropped_nested;
5890 int min_size_nested;
5891 int max_size_nested;
5898 static struct rb_test_data rb_data[NR_CPUS] __initdata;
5901 #define RB_TEST_BUFFER_SIZE 1048576
5903 static char rb_string[] __initdata =
5904 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
5905 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
5906 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
5908 static bool rb_test_started __initdata;
5915 static __init int rb_write_something(struct rb_test_data *data, bool nested)
5917 struct ring_buffer_event *event;
5918 struct rb_item *item;
5925 /* Have nested writes different that what is written */
5926 cnt = data->cnt + (nested ? 27 : 0);
5928 /* Multiply cnt by ~e, to make some unique increment */
5929 size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
5931 len = size + sizeof(struct rb_item);
5933 started = rb_test_started;
5934 /* read rb_test_started before checking buffer enabled */
5937 event = ring_buffer_lock_reserve(data->buffer, len);
5939 /* Ignore dropped events before test starts. */
5942 data->bytes_dropped += len;
5944 data->bytes_dropped_nested += len;
5949 event_len = ring_buffer_event_length(event);
5951 if (RB_WARN_ON(data->buffer, event_len < len))
5954 item = ring_buffer_event_data(event);
5956 memcpy(item->str, rb_string, size);
5959 data->bytes_alloc_nested += event_len;
5960 data->bytes_written_nested += len;
5961 data->events_nested++;
5962 if (!data->min_size_nested || len < data->min_size_nested)
5963 data->min_size_nested = len;
5964 if (len > data->max_size_nested)
5965 data->max_size_nested = len;
5967 data->bytes_alloc += event_len;
5968 data->bytes_written += len;
5970 if (!data->min_size || len < data->min_size)
5971 data->max_size = len;
5972 if (len > data->max_size)
5973 data->max_size = len;
5977 ring_buffer_unlock_commit(data->buffer);
5982 static __init int rb_test(void *arg)
5984 struct rb_test_data *data = arg;
5986 while (!kthread_should_stop()) {
5987 rb_write_something(data, false);
5990 set_current_state(TASK_INTERRUPTIBLE);
5991 /* Now sleep between a min of 100-300us and a max of 1ms */
5992 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
5998 static __init void rb_ipi(void *ignore)
6000 struct rb_test_data *data;
6001 int cpu = smp_processor_id();
6003 data = &rb_data[cpu];
6004 rb_write_something(data, true);
6007 static __init int rb_hammer_test(void *arg)
6009 while (!kthread_should_stop()) {
6011 /* Send an IPI to all cpus to write data! */
6012 smp_call_function(rb_ipi, NULL, 1);
6013 /* No sleep, but for non preempt, let others run */
6020 static __init int test_ringbuffer(void)
6022 struct task_struct *rb_hammer;
6023 struct trace_buffer *buffer;
6027 if (security_locked_down(LOCKDOWN_TRACEFS)) {
6028 pr_warn("Lockdown is enabled, skipping ring buffer tests\n");
6032 pr_info("Running ring buffer tests...\n");
6034 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
6035 if (WARN_ON(!buffer))
6038 /* Disable buffer so that threads can't write to it yet */
6039 ring_buffer_record_off(buffer);
6041 for_each_online_cpu(cpu) {
6042 rb_data[cpu].buffer = buffer;
6043 rb_data[cpu].cpu = cpu;
6044 rb_data[cpu].cnt = cpu;
6045 rb_threads[cpu] = kthread_run_on_cpu(rb_test, &rb_data[cpu],
6046 cpu, "rbtester/%u");
6047 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
6048 pr_cont("FAILED\n");
6049 ret = PTR_ERR(rb_threads[cpu]);
6054 /* Now create the rb hammer! */
6055 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
6056 if (WARN_ON(IS_ERR(rb_hammer))) {
6057 pr_cont("FAILED\n");
6058 ret = PTR_ERR(rb_hammer);
6062 ring_buffer_record_on(buffer);
6064 * Show buffer is enabled before setting rb_test_started.
6065 * Yes there's a small race window where events could be
6066 * dropped and the thread wont catch it. But when a ring
6067 * buffer gets enabled, there will always be some kind of
6068 * delay before other CPUs see it. Thus, we don't care about
6069 * those dropped events. We care about events dropped after
6070 * the threads see that the buffer is active.
6073 rb_test_started = true;
6075 set_current_state(TASK_INTERRUPTIBLE);
6076 /* Just run for 10 seconds */;
6077 schedule_timeout(10 * HZ);
6079 kthread_stop(rb_hammer);
6082 for_each_online_cpu(cpu) {
6083 if (!rb_threads[cpu])
6085 kthread_stop(rb_threads[cpu]);
6088 ring_buffer_free(buffer);
6093 pr_info("finished\n");
6094 for_each_online_cpu(cpu) {
6095 struct ring_buffer_event *event;
6096 struct rb_test_data *data = &rb_data[cpu];
6097 struct rb_item *item;
6098 unsigned long total_events;
6099 unsigned long total_dropped;
6100 unsigned long total_written;
6101 unsigned long total_alloc;
6102 unsigned long total_read = 0;
6103 unsigned long total_size = 0;
6104 unsigned long total_len = 0;
6105 unsigned long total_lost = 0;
6108 int small_event_size;
6112 total_events = data->events + data->events_nested;
6113 total_written = data->bytes_written + data->bytes_written_nested;
6114 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
6115 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
6117 big_event_size = data->max_size + data->max_size_nested;
6118 small_event_size = data->min_size + data->min_size_nested;
6120 pr_info("CPU %d:\n", cpu);
6121 pr_info(" events: %ld\n", total_events);
6122 pr_info(" dropped bytes: %ld\n", total_dropped);
6123 pr_info(" alloced bytes: %ld\n", total_alloc);
6124 pr_info(" written bytes: %ld\n", total_written);
6125 pr_info(" biggest event: %d\n", big_event_size);
6126 pr_info(" smallest event: %d\n", small_event_size);
6128 if (RB_WARN_ON(buffer, total_dropped))
6133 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
6135 item = ring_buffer_event_data(event);
6136 total_len += ring_buffer_event_length(event);
6137 total_size += item->size + sizeof(struct rb_item);
6138 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
6139 pr_info("FAILED!\n");
6140 pr_info("buffer had: %.*s\n", item->size, item->str);
6141 pr_info("expected: %.*s\n", item->size, rb_string);
6142 RB_WARN_ON(buffer, 1);
6153 pr_info(" read events: %ld\n", total_read);
6154 pr_info(" lost events: %ld\n", total_lost);
6155 pr_info(" total events: %ld\n", total_lost + total_read);
6156 pr_info(" recorded len bytes: %ld\n", total_len);
6157 pr_info(" recorded size bytes: %ld\n", total_size);
6159 pr_info(" With dropped events, record len and size may not match\n"
6160 " alloced and written from above\n");
6162 if (RB_WARN_ON(buffer, total_len != total_alloc ||
6163 total_size != total_written))
6166 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
6172 pr_info("Ring buffer PASSED!\n");
6174 ring_buffer_free(buffer);
6178 late_initcall(test_ringbuffer);
6179 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */