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 int 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 void free_buffer_page(struct buffer_page *bpage)
359 free_page((unsigned long)bpage->page);
364 * We need to fit the time_stamp delta into 27 bits.
366 static inline int test_time_stamp(u64 delta)
368 if (delta & TS_DELTA_TEST)
373 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
375 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
376 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
378 int ring_buffer_print_page_header(struct trace_seq *s)
380 struct buffer_data_page field;
382 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
383 "offset:0;\tsize:%u;\tsigned:%u;\n",
384 (unsigned int)sizeof(field.time_stamp),
385 (unsigned int)is_signed_type(u64));
387 trace_seq_printf(s, "\tfield: local_t commit;\t"
388 "offset:%u;\tsize:%u;\tsigned:%u;\n",
389 (unsigned int)offsetof(typeof(field), commit),
390 (unsigned int)sizeof(field.commit),
391 (unsigned int)is_signed_type(long));
393 trace_seq_printf(s, "\tfield: int overwrite;\t"
394 "offset:%u;\tsize:%u;\tsigned:%u;\n",
395 (unsigned int)offsetof(typeof(field), commit),
397 (unsigned int)is_signed_type(long));
399 trace_seq_printf(s, "\tfield: char data;\t"
400 "offset:%u;\tsize:%u;\tsigned:%u;\n",
401 (unsigned int)offsetof(typeof(field), data),
402 (unsigned int)BUF_PAGE_SIZE,
403 (unsigned int)is_signed_type(char));
405 return !trace_seq_has_overflowed(s);
409 struct irq_work work;
410 wait_queue_head_t waiters;
411 wait_queue_head_t full_waiters;
413 bool waiters_pending;
414 bool full_waiters_pending;
419 * Structure to hold event state and handle nested events.
421 struct rb_event_info {
426 unsigned long length;
427 struct buffer_page *tail_page;
432 * Used for the add_timestamp
434 * EXTEND - wants a time extend
435 * ABSOLUTE - the buffer requests all events to have absolute time stamps
436 * FORCE - force a full time stamp.
439 RB_ADD_STAMP_NONE = 0,
440 RB_ADD_STAMP_EXTEND = BIT(1),
441 RB_ADD_STAMP_ABSOLUTE = BIT(2),
442 RB_ADD_STAMP_FORCE = BIT(3)
445 * Used for which event context the event is in.
452 * See trace_recursive_lock() comment below for more details.
463 #if BITS_PER_LONG == 32
467 /* To test on 64 bit machines */
472 struct rb_time_struct {
479 #include <asm/local64.h>
480 struct rb_time_struct {
484 typedef struct rb_time_struct rb_time_t;
489 * head_page == tail_page && head == tail then buffer is empty.
491 struct ring_buffer_per_cpu {
493 atomic_t record_disabled;
494 atomic_t resize_disabled;
495 struct trace_buffer *buffer;
496 raw_spinlock_t reader_lock; /* serialize readers */
497 arch_spinlock_t lock;
498 struct lock_class_key lock_key;
499 struct buffer_data_page *free_page;
500 unsigned long nr_pages;
501 unsigned int current_context;
502 struct list_head *pages;
503 struct buffer_page *head_page; /* read from head */
504 struct buffer_page *tail_page; /* write to tail */
505 struct buffer_page *commit_page; /* committed pages */
506 struct buffer_page *reader_page;
507 unsigned long lost_events;
508 unsigned long last_overrun;
510 local_t entries_bytes;
513 local_t commit_overrun;
514 local_t dropped_events;
517 local_t pages_touched;
520 long last_pages_touch;
521 size_t shortest_full;
523 unsigned long read_bytes;
524 rb_time_t write_stamp;
525 rb_time_t before_stamp;
526 u64 event_stamp[MAX_NEST];
528 /* ring buffer pages to update, > 0 to add, < 0 to remove */
529 long nr_pages_to_update;
530 struct list_head new_pages; /* new pages to add */
531 struct work_struct update_pages_work;
532 struct completion update_done;
534 struct rb_irq_work irq_work;
537 struct trace_buffer {
540 atomic_t record_disabled;
541 cpumask_var_t cpumask;
543 struct lock_class_key *reader_lock_key;
547 struct ring_buffer_per_cpu **buffers;
549 struct hlist_node node;
552 struct rb_irq_work irq_work;
556 struct ring_buffer_iter {
557 struct ring_buffer_per_cpu *cpu_buffer;
559 unsigned long next_event;
560 struct buffer_page *head_page;
561 struct buffer_page *cache_reader_page;
562 unsigned long cache_read;
565 struct ring_buffer_event *event;
572 * On 32 bit machines, local64_t is very expensive. As the ring
573 * buffer doesn't need all the features of a true 64 bit atomic,
574 * on 32 bit, it uses these functions (64 still uses local64_t).
576 * For the ring buffer, 64 bit required operations for the time is
579 * - Reads may fail if it interrupted a modification of the time stamp.
580 * It will succeed if it did not interrupt another write even if
581 * the read itself is interrupted by a write.
582 * It returns whether it was successful or not.
584 * - Writes always succeed and will overwrite other writes and writes
585 * that were done by events interrupting the current write.
587 * - A write followed by a read of the same time stamp will always succeed,
588 * but may not contain the same value.
590 * - A cmpxchg will fail if it interrupted another write or cmpxchg.
591 * Other than that, it acts like a normal cmpxchg.
593 * The 60 bit time stamp is broken up by 30 bits in a top and bottom half
594 * (bottom being the least significant 30 bits of the 60 bit time stamp).
596 * The two most significant bits of each half holds a 2 bit counter (0-3).
597 * Each update will increment this counter by one.
598 * When reading the top and bottom, if the two counter bits match then the
599 * top and bottom together make a valid 60 bit number.
601 #define RB_TIME_SHIFT 30
602 #define RB_TIME_VAL_MASK ((1 << RB_TIME_SHIFT) - 1)
603 #define RB_TIME_MSB_SHIFT 60
605 static inline int rb_time_cnt(unsigned long val)
607 return (val >> RB_TIME_SHIFT) & 3;
610 static inline u64 rb_time_val(unsigned long top, unsigned long bottom)
614 val = top & RB_TIME_VAL_MASK;
615 val <<= RB_TIME_SHIFT;
616 val |= bottom & RB_TIME_VAL_MASK;
621 static inline bool __rb_time_read(rb_time_t *t, u64 *ret, unsigned long *cnt)
623 unsigned long top, bottom, msb;
627 * If the read is interrupted by a write, then the cnt will
628 * be different. Loop until both top and bottom have been read
629 * without interruption.
632 c = local_read(&t->cnt);
633 top = local_read(&t->top);
634 bottom = local_read(&t->bottom);
635 msb = local_read(&t->msb);
636 } while (c != local_read(&t->cnt));
638 *cnt = rb_time_cnt(top);
640 /* If top and bottom counts don't match, this interrupted a write */
641 if (*cnt != rb_time_cnt(bottom))
644 /* The shift to msb will lose its cnt bits */
645 *ret = rb_time_val(top, bottom) | ((u64)msb << RB_TIME_MSB_SHIFT);
649 static bool rb_time_read(rb_time_t *t, u64 *ret)
653 return __rb_time_read(t, ret, &cnt);
656 static inline unsigned long rb_time_val_cnt(unsigned long val, unsigned long cnt)
658 return (val & RB_TIME_VAL_MASK) | ((cnt & 3) << RB_TIME_SHIFT);
661 static inline void rb_time_split(u64 val, unsigned long *top, unsigned long *bottom,
664 *top = (unsigned long)((val >> RB_TIME_SHIFT) & RB_TIME_VAL_MASK);
665 *bottom = (unsigned long)(val & RB_TIME_VAL_MASK);
666 *msb = (unsigned long)(val >> RB_TIME_MSB_SHIFT);
669 static inline void rb_time_val_set(local_t *t, unsigned long val, unsigned long cnt)
671 val = rb_time_val_cnt(val, cnt);
675 static void rb_time_set(rb_time_t *t, u64 val)
677 unsigned long cnt, top, bottom, msb;
679 rb_time_split(val, &top, &bottom, &msb);
681 /* Writes always succeed with a valid number even if it gets interrupted. */
683 cnt = local_inc_return(&t->cnt);
684 rb_time_val_set(&t->top, top, cnt);
685 rb_time_val_set(&t->bottom, bottom, cnt);
686 rb_time_val_set(&t->msb, val >> RB_TIME_MSB_SHIFT, cnt);
687 } while (cnt != local_read(&t->cnt));
691 rb_time_read_cmpxchg(local_t *l, unsigned long expect, unsigned long set)
695 ret = local_cmpxchg(l, expect, set);
696 return ret == expect;
699 static int rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
701 unsigned long cnt, top, bottom, msb;
702 unsigned long cnt2, top2, bottom2, msb2;
705 /* The cmpxchg always fails if it interrupted an update */
706 if (!__rb_time_read(t, &val, &cnt2))
712 cnt = local_read(&t->cnt);
713 if ((cnt & 3) != cnt2)
718 rb_time_split(val, &top, &bottom, &msb);
719 top = rb_time_val_cnt(top, cnt);
720 bottom = rb_time_val_cnt(bottom, cnt);
722 rb_time_split(set, &top2, &bottom2, &msb2);
723 top2 = rb_time_val_cnt(top2, cnt2);
724 bottom2 = rb_time_val_cnt(bottom2, cnt2);
726 if (!rb_time_read_cmpxchg(&t->cnt, cnt, cnt2))
728 if (!rb_time_read_cmpxchg(&t->msb, msb, msb2))
730 if (!rb_time_read_cmpxchg(&t->top, top, top2))
732 if (!rb_time_read_cmpxchg(&t->bottom, bottom, bottom2))
739 /* local64_t always succeeds */
741 static inline bool rb_time_read(rb_time_t *t, u64 *ret)
743 *ret = local64_read(&t->time);
746 static void rb_time_set(rb_time_t *t, u64 val)
748 local64_set(&t->time, val);
751 static bool rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
754 val = local64_cmpxchg(&t->time, expect, set);
755 return val == expect;
760 * Enable this to make sure that the event passed to
761 * ring_buffer_event_time_stamp() is not committed and also
762 * is on the buffer that it passed in.
764 //#define RB_VERIFY_EVENT
765 #ifdef RB_VERIFY_EVENT
766 static struct list_head *rb_list_head(struct list_head *list);
767 static void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
770 struct buffer_page *page = cpu_buffer->commit_page;
771 struct buffer_page *tail_page = READ_ONCE(cpu_buffer->tail_page);
772 struct list_head *next;
774 unsigned long addr = (unsigned long)event;
778 /* Make sure the event exists and is not committed yet */
780 if (page == tail_page || WARN_ON_ONCE(stop++ > 100))
782 commit = local_read(&page->page->commit);
783 write = local_read(&page->write);
784 if (addr >= (unsigned long)&page->page->data[commit] &&
785 addr < (unsigned long)&page->page->data[write])
788 next = rb_list_head(page->list.next);
789 page = list_entry(next, struct buffer_page, list);
794 static inline void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
801 * The absolute time stamp drops the 5 MSBs and some clocks may
802 * require them. The rb_fix_abs_ts() will take a previous full
803 * time stamp, and add the 5 MSB of that time stamp on to the
804 * saved absolute time stamp. Then they are compared in case of
805 * the unlikely event that the latest time stamp incremented
808 static inline u64 rb_fix_abs_ts(u64 abs, u64 save_ts)
810 if (save_ts & TS_MSB) {
811 abs |= save_ts & TS_MSB;
812 /* Check for overflow */
813 if (unlikely(abs < save_ts))
819 static inline u64 rb_time_stamp(struct trace_buffer *buffer);
822 * ring_buffer_event_time_stamp - return the event's current time stamp
823 * @buffer: The buffer that the event is on
824 * @event: the event to get the time stamp of
826 * Note, this must be called after @event is reserved, and before it is
827 * committed to the ring buffer. And must be called from the same
828 * context where the event was reserved (normal, softirq, irq, etc).
830 * Returns the time stamp associated with the current event.
831 * If the event has an extended time stamp, then that is used as
832 * the time stamp to return.
833 * In the highly unlikely case that the event was nested more than
834 * the max nesting, then the write_stamp of the buffer is returned,
835 * otherwise current time is returned, but that really neither of
836 * the last two cases should ever happen.
838 u64 ring_buffer_event_time_stamp(struct trace_buffer *buffer,
839 struct ring_buffer_event *event)
841 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[smp_processor_id()];
845 /* If the event includes an absolute time, then just use that */
846 if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
847 ts = rb_event_time_stamp(event);
848 return rb_fix_abs_ts(ts, cpu_buffer->tail_page->page->time_stamp);
851 nest = local_read(&cpu_buffer->committing);
852 verify_event(cpu_buffer, event);
853 if (WARN_ON_ONCE(!nest))
856 /* Read the current saved nesting level time stamp */
857 if (likely(--nest < MAX_NEST))
858 return cpu_buffer->event_stamp[nest];
860 /* Shouldn't happen, warn if it does */
861 WARN_ONCE(1, "nest (%d) greater than max", nest);
864 /* Can only fail on 32 bit */
865 if (!rb_time_read(&cpu_buffer->write_stamp, &ts))
866 /* Screw it, just read the current time */
867 ts = rb_time_stamp(cpu_buffer->buffer);
873 * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
874 * @buffer: The ring_buffer to get the number of pages from
875 * @cpu: The cpu of the ring_buffer to get the number of pages from
877 * Returns the number of pages used by a per_cpu buffer of the ring buffer.
879 size_t ring_buffer_nr_pages(struct trace_buffer *buffer, int cpu)
881 return buffer->buffers[cpu]->nr_pages;
885 * ring_buffer_nr_dirty_pages - get the number of used pages in the ring buffer
886 * @buffer: The ring_buffer to get the number of pages from
887 * @cpu: The cpu of the ring_buffer to get the number of pages from
889 * Returns the number of pages that have content in the ring buffer.
891 size_t ring_buffer_nr_dirty_pages(struct trace_buffer *buffer, int cpu)
897 read = local_read(&buffer->buffers[cpu]->pages_read);
898 lost = local_read(&buffer->buffers[cpu]->pages_lost);
899 cnt = local_read(&buffer->buffers[cpu]->pages_touched);
901 if (WARN_ON_ONCE(cnt < lost))
906 /* The reader can read an empty page, but not more than that */
908 WARN_ON_ONCE(read > cnt + 1);
915 static __always_inline bool full_hit(struct trace_buffer *buffer, int cpu, int full)
917 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
921 nr_pages = cpu_buffer->nr_pages;
922 if (!nr_pages || !full)
925 dirty = ring_buffer_nr_dirty_pages(buffer, cpu);
927 return (dirty * 100) > (full * nr_pages);
931 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
933 * Schedules a delayed work to wake up any task that is blocked on the
934 * ring buffer waiters queue.
936 static void rb_wake_up_waiters(struct irq_work *work)
938 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
940 wake_up_all(&rbwork->waiters);
941 if (rbwork->full_waiters_pending || rbwork->wakeup_full) {
942 rbwork->wakeup_full = false;
943 rbwork->full_waiters_pending = false;
944 wake_up_all(&rbwork->full_waiters);
949 * ring_buffer_wake_waiters - wake up any waiters on this ring buffer
950 * @buffer: The ring buffer to wake waiters on
952 * In the case of a file that represents a ring buffer is closing,
953 * it is prudent to wake up any waiters that are on this.
955 void ring_buffer_wake_waiters(struct trace_buffer *buffer, int cpu)
957 struct ring_buffer_per_cpu *cpu_buffer;
958 struct rb_irq_work *rbwork;
963 if (cpu == RING_BUFFER_ALL_CPUS) {
965 /* Wake up individual ones too. One level recursion */
966 for_each_buffer_cpu(buffer, cpu)
967 ring_buffer_wake_waiters(buffer, cpu);
969 rbwork = &buffer->irq_work;
971 if (WARN_ON_ONCE(!buffer->buffers))
973 if (WARN_ON_ONCE(cpu >= nr_cpu_ids))
976 cpu_buffer = buffer->buffers[cpu];
977 /* The CPU buffer may not have been initialized yet */
980 rbwork = &cpu_buffer->irq_work;
983 rbwork->wait_index++;
984 /* make sure the waiters see the new index */
987 rb_wake_up_waiters(&rbwork->work);
991 * ring_buffer_wait - wait for input to the ring buffer
992 * @buffer: buffer to wait on
993 * @cpu: the cpu buffer to wait on
994 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
996 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
997 * as data is added to any of the @buffer's cpu buffers. Otherwise
998 * it will wait for data to be added to a specific cpu buffer.
1000 int ring_buffer_wait(struct trace_buffer *buffer, int cpu, int full)
1002 struct ring_buffer_per_cpu *cpu_buffer;
1004 struct rb_irq_work *work;
1009 * Depending on what the caller is waiting for, either any
1010 * data in any cpu buffer, or a specific buffer, put the
1011 * caller on the appropriate wait queue.
1013 if (cpu == RING_BUFFER_ALL_CPUS) {
1014 work = &buffer->irq_work;
1015 /* Full only makes sense on per cpu reads */
1018 if (!cpumask_test_cpu(cpu, buffer->cpumask))
1020 cpu_buffer = buffer->buffers[cpu];
1021 work = &cpu_buffer->irq_work;
1024 wait_index = READ_ONCE(work->wait_index);
1028 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
1030 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
1033 * The events can happen in critical sections where
1034 * checking a work queue can cause deadlocks.
1035 * After adding a task to the queue, this flag is set
1036 * only to notify events to try to wake up the queue
1039 * We don't clear it even if the buffer is no longer
1040 * empty. The flag only causes the next event to run
1041 * irq_work to do the work queue wake up. The worse
1042 * that can happen if we race with !trace_empty() is that
1043 * an event will cause an irq_work to try to wake up
1046 * There's no reason to protect this flag either, as
1047 * the work queue and irq_work logic will do the necessary
1048 * synchronization for the wake ups. The only thing
1049 * that is necessary is that the wake up happens after
1050 * a task has been queued. It's OK for spurious wake ups.
1053 work->full_waiters_pending = true;
1055 work->waiters_pending = true;
1057 if (signal_pending(current)) {
1062 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
1065 if (cpu != RING_BUFFER_ALL_CPUS &&
1066 !ring_buffer_empty_cpu(buffer, cpu)) {
1067 unsigned long flags;
1074 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
1075 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
1076 done = !pagebusy && full_hit(buffer, cpu, full);
1078 if (!cpu_buffer->shortest_full ||
1079 cpu_buffer->shortest_full > full)
1080 cpu_buffer->shortest_full = full;
1081 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
1088 /* Make sure to see the new wait index */
1090 if (wait_index != work->wait_index)
1095 finish_wait(&work->full_waiters, &wait);
1097 finish_wait(&work->waiters, &wait);
1103 * ring_buffer_poll_wait - poll on buffer input
1104 * @buffer: buffer to wait on
1105 * @cpu: the cpu buffer to wait on
1106 * @filp: the file descriptor
1107 * @poll_table: The poll descriptor
1108 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
1110 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
1111 * as data is added to any of the @buffer's cpu buffers. Otherwise
1112 * it will wait for data to be added to a specific cpu buffer.
1114 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
1117 __poll_t ring_buffer_poll_wait(struct trace_buffer *buffer, int cpu,
1118 struct file *filp, poll_table *poll_table, int full)
1120 struct ring_buffer_per_cpu *cpu_buffer;
1121 struct rb_irq_work *work;
1123 if (cpu == RING_BUFFER_ALL_CPUS) {
1124 work = &buffer->irq_work;
1127 if (!cpumask_test_cpu(cpu, buffer->cpumask))
1130 cpu_buffer = buffer->buffers[cpu];
1131 work = &cpu_buffer->irq_work;
1135 poll_wait(filp, &work->full_waiters, poll_table);
1136 work->full_waiters_pending = true;
1138 poll_wait(filp, &work->waiters, poll_table);
1139 work->waiters_pending = true;
1143 * There's a tight race between setting the waiters_pending and
1144 * checking if the ring buffer is empty. Once the waiters_pending bit
1145 * is set, the next event will wake the task up, but we can get stuck
1146 * if there's only a single event in.
1148 * FIXME: Ideally, we need a memory barrier on the writer side as well,
1149 * but adding a memory barrier to all events will cause too much of a
1150 * performance hit in the fast path. We only need a memory barrier when
1151 * the buffer goes from empty to having content. But as this race is
1152 * extremely small, and it's not a problem if another event comes in, we
1153 * will fix it later.
1158 return full_hit(buffer, cpu, full) ? EPOLLIN | EPOLLRDNORM : 0;
1160 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
1161 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
1162 return EPOLLIN | EPOLLRDNORM;
1166 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
1167 #define RB_WARN_ON(b, cond) \
1169 int _____ret = unlikely(cond); \
1171 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
1172 struct ring_buffer_per_cpu *__b = \
1174 atomic_inc(&__b->buffer->record_disabled); \
1176 atomic_inc(&b->record_disabled); \
1182 /* Up this if you want to test the TIME_EXTENTS and normalization */
1183 #define DEBUG_SHIFT 0
1185 static inline u64 rb_time_stamp(struct trace_buffer *buffer)
1189 /* Skip retpolines :-( */
1190 if (IS_ENABLED(CONFIG_RETPOLINE) && likely(buffer->clock == trace_clock_local))
1191 ts = trace_clock_local();
1193 ts = buffer->clock();
1195 /* shift to debug/test normalization and TIME_EXTENTS */
1196 return ts << DEBUG_SHIFT;
1199 u64 ring_buffer_time_stamp(struct trace_buffer *buffer)
1203 preempt_disable_notrace();
1204 time = rb_time_stamp(buffer);
1205 preempt_enable_notrace();
1209 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
1211 void ring_buffer_normalize_time_stamp(struct trace_buffer *buffer,
1214 /* Just stupid testing the normalize function and deltas */
1215 *ts >>= DEBUG_SHIFT;
1217 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
1220 * Making the ring buffer lockless makes things tricky.
1221 * Although writes only happen on the CPU that they are on,
1222 * and they only need to worry about interrupts. Reads can
1223 * happen on any CPU.
1225 * The reader page is always off the ring buffer, but when the
1226 * reader finishes with a page, it needs to swap its page with
1227 * a new one from the buffer. The reader needs to take from
1228 * the head (writes go to the tail). But if a writer is in overwrite
1229 * mode and wraps, it must push the head page forward.
1231 * Here lies the problem.
1233 * The reader must be careful to replace only the head page, and
1234 * not another one. As described at the top of the file in the
1235 * ASCII art, the reader sets its old page to point to the next
1236 * page after head. It then sets the page after head to point to
1237 * the old reader page. But if the writer moves the head page
1238 * during this operation, the reader could end up with the tail.
1240 * We use cmpxchg to help prevent this race. We also do something
1241 * special with the page before head. We set the LSB to 1.
1243 * When the writer must push the page forward, it will clear the
1244 * bit that points to the head page, move the head, and then set
1245 * the bit that points to the new head page.
1247 * We also don't want an interrupt coming in and moving the head
1248 * page on another writer. Thus we use the second LSB to catch
1251 * head->list->prev->next bit 1 bit 0
1254 * Points to head page 0 1
1257 * Note we can not trust the prev pointer of the head page, because:
1259 * +----+ +-----+ +-----+
1260 * | |------>| T |---X--->| N |
1262 * +----+ +-----+ +-----+
1265 * +----------| R |----------+ |
1269 * Key: ---X--> HEAD flag set in pointer
1274 * (see __rb_reserve_next() to see where this happens)
1276 * What the above shows is that the reader just swapped out
1277 * the reader page with a page in the buffer, but before it
1278 * could make the new header point back to the new page added
1279 * it was preempted by a writer. The writer moved forward onto
1280 * the new page added by the reader and is about to move forward
1283 * You can see, it is legitimate for the previous pointer of
1284 * the head (or any page) not to point back to itself. But only
1288 #define RB_PAGE_NORMAL 0UL
1289 #define RB_PAGE_HEAD 1UL
1290 #define RB_PAGE_UPDATE 2UL
1293 #define RB_FLAG_MASK 3UL
1295 /* PAGE_MOVED is not part of the mask */
1296 #define RB_PAGE_MOVED 4UL
1299 * rb_list_head - remove any bit
1301 static struct list_head *rb_list_head(struct list_head *list)
1303 unsigned long val = (unsigned long)list;
1305 return (struct list_head *)(val & ~RB_FLAG_MASK);
1309 * rb_is_head_page - test if the given page is the head page
1311 * Because the reader may move the head_page pointer, we can
1312 * not trust what the head page is (it may be pointing to
1313 * the reader page). But if the next page is a header page,
1314 * its flags will be non zero.
1317 rb_is_head_page(struct buffer_page *page, struct list_head *list)
1321 val = (unsigned long)list->next;
1323 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
1324 return RB_PAGE_MOVED;
1326 return val & RB_FLAG_MASK;
1332 * The unique thing about the reader page, is that, if the
1333 * writer is ever on it, the previous pointer never points
1334 * back to the reader page.
1336 static bool rb_is_reader_page(struct buffer_page *page)
1338 struct list_head *list = page->list.prev;
1340 return rb_list_head(list->next) != &page->list;
1344 * rb_set_list_to_head - set a list_head to be pointing to head.
1346 static void rb_set_list_to_head(struct list_head *list)
1350 ptr = (unsigned long *)&list->next;
1351 *ptr |= RB_PAGE_HEAD;
1352 *ptr &= ~RB_PAGE_UPDATE;
1356 * rb_head_page_activate - sets up head page
1358 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
1360 struct buffer_page *head;
1362 head = cpu_buffer->head_page;
1367 * Set the previous list pointer to have the HEAD flag.
1369 rb_set_list_to_head(head->list.prev);
1372 static void rb_list_head_clear(struct list_head *list)
1374 unsigned long *ptr = (unsigned long *)&list->next;
1376 *ptr &= ~RB_FLAG_MASK;
1380 * rb_head_page_deactivate - clears head page ptr (for free list)
1383 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
1385 struct list_head *hd;
1387 /* Go through the whole list and clear any pointers found. */
1388 rb_list_head_clear(cpu_buffer->pages);
1390 list_for_each(hd, cpu_buffer->pages)
1391 rb_list_head_clear(hd);
1394 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
1395 struct buffer_page *head,
1396 struct buffer_page *prev,
1397 int old_flag, int new_flag)
1399 struct list_head *list;
1400 unsigned long val = (unsigned long)&head->list;
1405 val &= ~RB_FLAG_MASK;
1407 ret = cmpxchg((unsigned long *)&list->next,
1408 val | old_flag, val | new_flag);
1410 /* check if the reader took the page */
1411 if ((ret & ~RB_FLAG_MASK) != val)
1412 return RB_PAGE_MOVED;
1414 return ret & RB_FLAG_MASK;
1417 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
1418 struct buffer_page *head,
1419 struct buffer_page *prev,
1422 return rb_head_page_set(cpu_buffer, head, prev,
1423 old_flag, RB_PAGE_UPDATE);
1426 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
1427 struct buffer_page *head,
1428 struct buffer_page *prev,
1431 return rb_head_page_set(cpu_buffer, head, prev,
1432 old_flag, RB_PAGE_HEAD);
1435 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
1436 struct buffer_page *head,
1437 struct buffer_page *prev,
1440 return rb_head_page_set(cpu_buffer, head, prev,
1441 old_flag, RB_PAGE_NORMAL);
1444 static inline void rb_inc_page(struct buffer_page **bpage)
1446 struct list_head *p = rb_list_head((*bpage)->list.next);
1448 *bpage = list_entry(p, struct buffer_page, list);
1451 static struct buffer_page *
1452 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1454 struct buffer_page *head;
1455 struct buffer_page *page;
1456 struct list_head *list;
1459 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1463 list = cpu_buffer->pages;
1464 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1467 page = head = cpu_buffer->head_page;
1469 * It is possible that the writer moves the header behind
1470 * where we started, and we miss in one loop.
1471 * A second loop should grab the header, but we'll do
1472 * three loops just because I'm paranoid.
1474 for (i = 0; i < 3; i++) {
1476 if (rb_is_head_page(page, page->list.prev)) {
1477 cpu_buffer->head_page = page;
1481 } while (page != head);
1484 RB_WARN_ON(cpu_buffer, 1);
1489 static int rb_head_page_replace(struct buffer_page *old,
1490 struct buffer_page *new)
1492 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1496 val = *ptr & ~RB_FLAG_MASK;
1497 val |= RB_PAGE_HEAD;
1499 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
1505 * rb_tail_page_update - move the tail page forward
1507 static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1508 struct buffer_page *tail_page,
1509 struct buffer_page *next_page)
1511 unsigned long old_entries;
1512 unsigned long old_write;
1515 * The tail page now needs to be moved forward.
1517 * We need to reset the tail page, but without messing
1518 * with possible erasing of data brought in by interrupts
1519 * that have moved the tail page and are currently on it.
1521 * We add a counter to the write field to denote this.
1523 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1524 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1526 local_inc(&cpu_buffer->pages_touched);
1528 * Just make sure we have seen our old_write and synchronize
1529 * with any interrupts that come in.
1534 * If the tail page is still the same as what we think
1535 * it is, then it is up to us to update the tail
1538 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1539 /* Zero the write counter */
1540 unsigned long val = old_write & ~RB_WRITE_MASK;
1541 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1544 * This will only succeed if an interrupt did
1545 * not come in and change it. In which case, we
1546 * do not want to modify it.
1548 * We add (void) to let the compiler know that we do not care
1549 * about the return value of these functions. We use the
1550 * cmpxchg to only update if an interrupt did not already
1551 * do it for us. If the cmpxchg fails, we don't care.
1553 (void)local_cmpxchg(&next_page->write, old_write, val);
1554 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1557 * No need to worry about races with clearing out the commit.
1558 * it only can increment when a commit takes place. But that
1559 * only happens in the outer most nested commit.
1561 local_set(&next_page->page->commit, 0);
1563 /* Again, either we update tail_page or an interrupt does */
1564 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1568 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1569 struct buffer_page *bpage)
1571 unsigned long val = (unsigned long)bpage;
1573 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
1580 * rb_check_pages - integrity check of buffer pages
1581 * @cpu_buffer: CPU buffer with pages to test
1583 * As a safety measure we check to make sure the data pages have not
1586 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1588 struct list_head *head = rb_list_head(cpu_buffer->pages);
1589 struct list_head *tmp;
1591 if (RB_WARN_ON(cpu_buffer,
1592 rb_list_head(rb_list_head(head->next)->prev) != head))
1595 if (RB_WARN_ON(cpu_buffer,
1596 rb_list_head(rb_list_head(head->prev)->next) != head))
1599 for (tmp = rb_list_head(head->next); tmp != head; tmp = rb_list_head(tmp->next)) {
1600 if (RB_WARN_ON(cpu_buffer,
1601 rb_list_head(rb_list_head(tmp->next)->prev) != tmp))
1604 if (RB_WARN_ON(cpu_buffer,
1605 rb_list_head(rb_list_head(tmp->prev)->next) != tmp))
1612 static int __rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1613 long nr_pages, struct list_head *pages)
1615 struct buffer_page *bpage, *tmp;
1616 bool user_thread = current->mm != NULL;
1621 * Check if the available memory is there first.
1622 * Note, si_mem_available() only gives us a rough estimate of available
1623 * memory. It may not be accurate. But we don't care, we just want
1624 * to prevent doing any allocation when it is obvious that it is
1625 * not going to succeed.
1627 i = si_mem_available();
1632 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1633 * gracefully without invoking oom-killer and the system is not
1636 mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1639 * If a user thread allocates too much, and si_mem_available()
1640 * reports there's enough memory, even though there is not.
1641 * Make sure the OOM killer kills this thread. This can happen
1642 * even with RETRY_MAYFAIL because another task may be doing
1643 * an allocation after this task has taken all memory.
1644 * This is the task the OOM killer needs to take out during this
1645 * loop, even if it was triggered by an allocation somewhere else.
1648 set_current_oom_origin();
1649 for (i = 0; i < nr_pages; i++) {
1652 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1653 mflags, cpu_to_node(cpu_buffer->cpu));
1657 rb_check_bpage(cpu_buffer, bpage);
1659 list_add(&bpage->list, pages);
1661 page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu), mflags, 0);
1664 bpage->page = page_address(page);
1665 rb_init_page(bpage->page);
1667 if (user_thread && fatal_signal_pending(current))
1671 clear_current_oom_origin();
1676 list_for_each_entry_safe(bpage, tmp, pages, list) {
1677 list_del_init(&bpage->list);
1678 free_buffer_page(bpage);
1681 clear_current_oom_origin();
1686 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1687 unsigned long nr_pages)
1693 if (__rb_allocate_pages(cpu_buffer, nr_pages, &pages))
1697 * The ring buffer page list is a circular list that does not
1698 * start and end with a list head. All page list items point to
1701 cpu_buffer->pages = pages.next;
1704 cpu_buffer->nr_pages = nr_pages;
1706 rb_check_pages(cpu_buffer);
1711 static struct ring_buffer_per_cpu *
1712 rb_allocate_cpu_buffer(struct trace_buffer *buffer, long nr_pages, int cpu)
1714 struct ring_buffer_per_cpu *cpu_buffer;
1715 struct buffer_page *bpage;
1719 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1720 GFP_KERNEL, cpu_to_node(cpu));
1724 cpu_buffer->cpu = cpu;
1725 cpu_buffer->buffer = buffer;
1726 raw_spin_lock_init(&cpu_buffer->reader_lock);
1727 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1728 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1729 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1730 init_completion(&cpu_buffer->update_done);
1731 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1732 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1733 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1735 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1736 GFP_KERNEL, cpu_to_node(cpu));
1738 goto fail_free_buffer;
1740 rb_check_bpage(cpu_buffer, bpage);
1742 cpu_buffer->reader_page = bpage;
1743 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1745 goto fail_free_reader;
1746 bpage->page = page_address(page);
1747 rb_init_page(bpage->page);
1749 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1750 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1752 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1754 goto fail_free_reader;
1756 cpu_buffer->head_page
1757 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1758 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1760 rb_head_page_activate(cpu_buffer);
1765 free_buffer_page(cpu_buffer->reader_page);
1772 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1774 struct list_head *head = cpu_buffer->pages;
1775 struct buffer_page *bpage, *tmp;
1777 free_buffer_page(cpu_buffer->reader_page);
1780 rb_head_page_deactivate(cpu_buffer);
1782 list_for_each_entry_safe(bpage, tmp, head, list) {
1783 list_del_init(&bpage->list);
1784 free_buffer_page(bpage);
1786 bpage = list_entry(head, struct buffer_page, list);
1787 free_buffer_page(bpage);
1794 * __ring_buffer_alloc - allocate a new ring_buffer
1795 * @size: the size in bytes per cpu that is needed.
1796 * @flags: attributes to set for the ring buffer.
1797 * @key: ring buffer reader_lock_key.
1799 * Currently the only flag that is available is the RB_FL_OVERWRITE
1800 * flag. This flag means that the buffer will overwrite old data
1801 * when the buffer wraps. If this flag is not set, the buffer will
1802 * drop data when the tail hits the head.
1804 struct trace_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1805 struct lock_class_key *key)
1807 struct trace_buffer *buffer;
1813 /* keep it in its own cache line */
1814 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1819 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1820 goto fail_free_buffer;
1822 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1823 buffer->flags = flags;
1824 buffer->clock = trace_clock_local;
1825 buffer->reader_lock_key = key;
1827 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1828 init_waitqueue_head(&buffer->irq_work.waiters);
1830 /* need at least two pages */
1834 buffer->cpus = nr_cpu_ids;
1836 bsize = sizeof(void *) * nr_cpu_ids;
1837 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1839 if (!buffer->buffers)
1840 goto fail_free_cpumask;
1842 cpu = raw_smp_processor_id();
1843 cpumask_set_cpu(cpu, buffer->cpumask);
1844 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1845 if (!buffer->buffers[cpu])
1846 goto fail_free_buffers;
1848 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1850 goto fail_free_buffers;
1852 mutex_init(&buffer->mutex);
1857 for_each_buffer_cpu(buffer, cpu) {
1858 if (buffer->buffers[cpu])
1859 rb_free_cpu_buffer(buffer->buffers[cpu]);
1861 kfree(buffer->buffers);
1864 free_cpumask_var(buffer->cpumask);
1870 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1873 * ring_buffer_free - free a ring buffer.
1874 * @buffer: the buffer to free.
1877 ring_buffer_free(struct trace_buffer *buffer)
1881 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1883 for_each_buffer_cpu(buffer, cpu)
1884 rb_free_cpu_buffer(buffer->buffers[cpu]);
1886 kfree(buffer->buffers);
1887 free_cpumask_var(buffer->cpumask);
1891 EXPORT_SYMBOL_GPL(ring_buffer_free);
1893 void ring_buffer_set_clock(struct trace_buffer *buffer,
1896 buffer->clock = clock;
1899 void ring_buffer_set_time_stamp_abs(struct trace_buffer *buffer, bool abs)
1901 buffer->time_stamp_abs = abs;
1904 bool ring_buffer_time_stamp_abs(struct trace_buffer *buffer)
1906 return buffer->time_stamp_abs;
1909 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1911 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1913 return local_read(&bpage->entries) & RB_WRITE_MASK;
1916 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1918 return local_read(&bpage->write) & RB_WRITE_MASK;
1922 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1924 struct list_head *tail_page, *to_remove, *next_page;
1925 struct buffer_page *to_remove_page, *tmp_iter_page;
1926 struct buffer_page *last_page, *first_page;
1927 unsigned long nr_removed;
1928 unsigned long head_bit;
1933 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1934 atomic_inc(&cpu_buffer->record_disabled);
1936 * We don't race with the readers since we have acquired the reader
1937 * lock. We also don't race with writers after disabling recording.
1938 * This makes it easy to figure out the first and the last page to be
1939 * removed from the list. We unlink all the pages in between including
1940 * the first and last pages. This is done in a busy loop so that we
1941 * lose the least number of traces.
1942 * The pages are freed after we restart recording and unlock readers.
1944 tail_page = &cpu_buffer->tail_page->list;
1947 * tail page might be on reader page, we remove the next page
1948 * from the ring buffer
1950 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1951 tail_page = rb_list_head(tail_page->next);
1952 to_remove = tail_page;
1954 /* start of pages to remove */
1955 first_page = list_entry(rb_list_head(to_remove->next),
1956 struct buffer_page, list);
1958 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1959 to_remove = rb_list_head(to_remove)->next;
1960 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1963 next_page = rb_list_head(to_remove)->next;
1966 * Now we remove all pages between tail_page and next_page.
1967 * Make sure that we have head_bit value preserved for the
1970 tail_page->next = (struct list_head *)((unsigned long)next_page |
1972 next_page = rb_list_head(next_page);
1973 next_page->prev = tail_page;
1975 /* make sure pages points to a valid page in the ring buffer */
1976 cpu_buffer->pages = next_page;
1978 /* update head page */
1980 cpu_buffer->head_page = list_entry(next_page,
1981 struct buffer_page, list);
1984 * change read pointer to make sure any read iterators reset
1987 cpu_buffer->read = 0;
1989 /* pages are removed, resume tracing and then free the pages */
1990 atomic_dec(&cpu_buffer->record_disabled);
1991 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1993 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1995 /* last buffer page to remove */
1996 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1998 tmp_iter_page = first_page;
2003 to_remove_page = tmp_iter_page;
2004 rb_inc_page(&tmp_iter_page);
2006 /* update the counters */
2007 page_entries = rb_page_entries(to_remove_page);
2010 * If something was added to this page, it was full
2011 * since it is not the tail page. So we deduct the
2012 * bytes consumed in ring buffer from here.
2013 * Increment overrun to account for the lost events.
2015 local_add(page_entries, &cpu_buffer->overrun);
2016 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2017 local_inc(&cpu_buffer->pages_lost);
2021 * We have already removed references to this list item, just
2022 * free up the buffer_page and its page
2024 free_buffer_page(to_remove_page);
2027 } while (to_remove_page != last_page);
2029 RB_WARN_ON(cpu_buffer, nr_removed);
2031 return nr_removed == 0;
2035 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
2037 struct list_head *pages = &cpu_buffer->new_pages;
2038 int retries, success;
2039 unsigned long flags;
2041 /* Can be called at early boot up, where interrupts must not been enabled */
2042 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2044 * We are holding the reader lock, so the reader page won't be swapped
2045 * in the ring buffer. Now we are racing with the writer trying to
2046 * move head page and the tail page.
2047 * We are going to adapt the reader page update process where:
2048 * 1. We first splice the start and end of list of new pages between
2049 * the head page and its previous page.
2050 * 2. We cmpxchg the prev_page->next to point from head page to the
2051 * start of new pages list.
2052 * 3. Finally, we update the head->prev to the end of new list.
2054 * We will try this process 10 times, to make sure that we don't keep
2060 struct list_head *head_page, *prev_page, *r;
2061 struct list_head *last_page, *first_page;
2062 struct list_head *head_page_with_bit;
2064 head_page = &rb_set_head_page(cpu_buffer)->list;
2067 prev_page = head_page->prev;
2069 first_page = pages->next;
2070 last_page = pages->prev;
2072 head_page_with_bit = (struct list_head *)
2073 ((unsigned long)head_page | RB_PAGE_HEAD);
2075 last_page->next = head_page_with_bit;
2076 first_page->prev = prev_page;
2078 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
2080 if (r == head_page_with_bit) {
2082 * yay, we replaced the page pointer to our new list,
2083 * now, we just have to update to head page's prev
2084 * pointer to point to end of list
2086 head_page->prev = last_page;
2093 INIT_LIST_HEAD(pages);
2095 * If we weren't successful in adding in new pages, warn and stop
2098 RB_WARN_ON(cpu_buffer, !success);
2099 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2101 /* free pages if they weren't inserted */
2103 struct buffer_page *bpage, *tmp;
2104 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2106 list_del_init(&bpage->list);
2107 free_buffer_page(bpage);
2113 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
2117 if (cpu_buffer->nr_pages_to_update > 0)
2118 success = rb_insert_pages(cpu_buffer);
2120 success = rb_remove_pages(cpu_buffer,
2121 -cpu_buffer->nr_pages_to_update);
2124 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
2127 static void update_pages_handler(struct work_struct *work)
2129 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
2130 struct ring_buffer_per_cpu, update_pages_work);
2131 rb_update_pages(cpu_buffer);
2132 complete(&cpu_buffer->update_done);
2136 * ring_buffer_resize - resize the ring buffer
2137 * @buffer: the buffer to resize.
2138 * @size: the new size.
2139 * @cpu_id: the cpu buffer to resize
2141 * Minimum size is 2 * BUF_PAGE_SIZE.
2143 * Returns 0 on success and < 0 on failure.
2145 int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
2148 struct ring_buffer_per_cpu *cpu_buffer;
2149 unsigned long nr_pages;
2153 * Always succeed at resizing a non-existent buffer:
2158 /* Make sure the requested buffer exists */
2159 if (cpu_id != RING_BUFFER_ALL_CPUS &&
2160 !cpumask_test_cpu(cpu_id, buffer->cpumask))
2163 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
2165 /* we need a minimum of two pages */
2169 /* prevent another thread from changing buffer sizes */
2170 mutex_lock(&buffer->mutex);
2173 if (cpu_id == RING_BUFFER_ALL_CPUS) {
2175 * Don't succeed if resizing is disabled, as a reader might be
2176 * manipulating the ring buffer and is expecting a sane state while
2179 for_each_buffer_cpu(buffer, cpu) {
2180 cpu_buffer = buffer->buffers[cpu];
2181 if (atomic_read(&cpu_buffer->resize_disabled)) {
2183 goto out_err_unlock;
2187 /* calculate the pages to update */
2188 for_each_buffer_cpu(buffer, cpu) {
2189 cpu_buffer = buffer->buffers[cpu];
2191 cpu_buffer->nr_pages_to_update = nr_pages -
2192 cpu_buffer->nr_pages;
2194 * nothing more to do for removing pages or no update
2196 if (cpu_buffer->nr_pages_to_update <= 0)
2199 * to add pages, make sure all new pages can be
2200 * allocated without receiving ENOMEM
2202 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2203 if (__rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2204 &cpu_buffer->new_pages)) {
2205 /* not enough memory for new pages */
2213 * Fire off all the required work handlers
2214 * We can't schedule on offline CPUs, but it's not necessary
2215 * since we can change their buffer sizes without any race.
2217 for_each_buffer_cpu(buffer, cpu) {
2218 cpu_buffer = buffer->buffers[cpu];
2219 if (!cpu_buffer->nr_pages_to_update)
2222 /* Can't run something on an offline CPU. */
2223 if (!cpu_online(cpu)) {
2224 rb_update_pages(cpu_buffer);
2225 cpu_buffer->nr_pages_to_update = 0;
2227 /* Run directly if possible. */
2229 if (cpu != smp_processor_id()) {
2231 schedule_work_on(cpu,
2232 &cpu_buffer->update_pages_work);
2234 update_pages_handler(&cpu_buffer->update_pages_work);
2240 /* wait for all the updates to complete */
2241 for_each_buffer_cpu(buffer, cpu) {
2242 cpu_buffer = buffer->buffers[cpu];
2243 if (!cpu_buffer->nr_pages_to_update)
2246 if (cpu_online(cpu))
2247 wait_for_completion(&cpu_buffer->update_done);
2248 cpu_buffer->nr_pages_to_update = 0;
2253 cpu_buffer = buffer->buffers[cpu_id];
2255 if (nr_pages == cpu_buffer->nr_pages)
2259 * Don't succeed if resizing is disabled, as a reader might be
2260 * manipulating the ring buffer and is expecting a sane state while
2263 if (atomic_read(&cpu_buffer->resize_disabled)) {
2265 goto out_err_unlock;
2268 cpu_buffer->nr_pages_to_update = nr_pages -
2269 cpu_buffer->nr_pages;
2271 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2272 if (cpu_buffer->nr_pages_to_update > 0 &&
2273 __rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2274 &cpu_buffer->new_pages)) {
2281 /* Can't run something on an offline CPU. */
2282 if (!cpu_online(cpu_id))
2283 rb_update_pages(cpu_buffer);
2285 /* Run directly if possible. */
2287 if (cpu_id == smp_processor_id()) {
2288 rb_update_pages(cpu_buffer);
2292 schedule_work_on(cpu_id,
2293 &cpu_buffer->update_pages_work);
2294 wait_for_completion(&cpu_buffer->update_done);
2298 cpu_buffer->nr_pages_to_update = 0;
2304 * The ring buffer resize can happen with the ring buffer
2305 * enabled, so that the update disturbs the tracing as little
2306 * as possible. But if the buffer is disabled, we do not need
2307 * to worry about that, and we can take the time to verify
2308 * that the buffer is not corrupt.
2310 if (atomic_read(&buffer->record_disabled)) {
2311 atomic_inc(&buffer->record_disabled);
2313 * Even though the buffer was disabled, we must make sure
2314 * that it is truly disabled before calling rb_check_pages.
2315 * There could have been a race between checking
2316 * record_disable and incrementing it.
2319 for_each_buffer_cpu(buffer, cpu) {
2320 cpu_buffer = buffer->buffers[cpu];
2321 rb_check_pages(cpu_buffer);
2323 atomic_dec(&buffer->record_disabled);
2326 mutex_unlock(&buffer->mutex);
2330 for_each_buffer_cpu(buffer, cpu) {
2331 struct buffer_page *bpage, *tmp;
2333 cpu_buffer = buffer->buffers[cpu];
2334 cpu_buffer->nr_pages_to_update = 0;
2336 if (list_empty(&cpu_buffer->new_pages))
2339 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2341 list_del_init(&bpage->list);
2342 free_buffer_page(bpage);
2346 mutex_unlock(&buffer->mutex);
2349 EXPORT_SYMBOL_GPL(ring_buffer_resize);
2351 void ring_buffer_change_overwrite(struct trace_buffer *buffer, int val)
2353 mutex_lock(&buffer->mutex);
2355 buffer->flags |= RB_FL_OVERWRITE;
2357 buffer->flags &= ~RB_FL_OVERWRITE;
2358 mutex_unlock(&buffer->mutex);
2360 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
2362 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
2364 return bpage->page->data + index;
2367 static __always_inline struct ring_buffer_event *
2368 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
2370 return __rb_page_index(cpu_buffer->reader_page,
2371 cpu_buffer->reader_page->read);
2374 static __always_inline unsigned rb_page_commit(struct buffer_page *bpage)
2376 return local_read(&bpage->page->commit);
2379 static struct ring_buffer_event *
2380 rb_iter_head_event(struct ring_buffer_iter *iter)
2382 struct ring_buffer_event *event;
2383 struct buffer_page *iter_head_page = iter->head_page;
2384 unsigned long commit;
2387 if (iter->head != iter->next_event)
2391 * When the writer goes across pages, it issues a cmpxchg which
2392 * is a mb(), which will synchronize with the rmb here.
2393 * (see rb_tail_page_update() and __rb_reserve_next())
2395 commit = rb_page_commit(iter_head_page);
2397 event = __rb_page_index(iter_head_page, iter->head);
2398 length = rb_event_length(event);
2401 * READ_ONCE() doesn't work on functions and we don't want the
2402 * compiler doing any crazy optimizations with length.
2406 if ((iter->head + length) > commit || length > BUF_MAX_DATA_SIZE)
2407 /* Writer corrupted the read? */
2410 memcpy(iter->event, event, length);
2412 * If the page stamp is still the same after this rmb() then the
2413 * event was safely copied without the writer entering the page.
2417 /* Make sure the page didn't change since we read this */
2418 if (iter->page_stamp != iter_head_page->page->time_stamp ||
2419 commit > rb_page_commit(iter_head_page))
2422 iter->next_event = iter->head + length;
2425 /* Reset to the beginning */
2426 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2428 iter->next_event = 0;
2429 iter->missed_events = 1;
2433 /* Size is determined by what has been committed */
2434 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
2436 return rb_page_commit(bpage);
2439 static __always_inline unsigned
2440 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
2442 return rb_page_commit(cpu_buffer->commit_page);
2445 static __always_inline unsigned
2446 rb_event_index(struct ring_buffer_event *event)
2448 unsigned long addr = (unsigned long)event;
2450 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
2453 static void rb_inc_iter(struct ring_buffer_iter *iter)
2455 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2458 * The iterator could be on the reader page (it starts there).
2459 * But the head could have moved, since the reader was
2460 * found. Check for this case and assign the iterator
2461 * to the head page instead of next.
2463 if (iter->head_page == cpu_buffer->reader_page)
2464 iter->head_page = rb_set_head_page(cpu_buffer);
2466 rb_inc_page(&iter->head_page);
2468 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2470 iter->next_event = 0;
2474 * rb_handle_head_page - writer hit the head page
2476 * Returns: +1 to retry page
2481 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2482 struct buffer_page *tail_page,
2483 struct buffer_page *next_page)
2485 struct buffer_page *new_head;
2490 entries = rb_page_entries(next_page);
2493 * The hard part is here. We need to move the head
2494 * forward, and protect against both readers on
2495 * other CPUs and writers coming in via interrupts.
2497 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2501 * type can be one of four:
2502 * NORMAL - an interrupt already moved it for us
2503 * HEAD - we are the first to get here.
2504 * UPDATE - we are the interrupt interrupting
2506 * MOVED - a reader on another CPU moved the next
2507 * pointer to its reader page. Give up
2514 * We changed the head to UPDATE, thus
2515 * it is our responsibility to update
2518 local_add(entries, &cpu_buffer->overrun);
2519 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
2520 local_inc(&cpu_buffer->pages_lost);
2523 * The entries will be zeroed out when we move the
2527 /* still more to do */
2530 case RB_PAGE_UPDATE:
2532 * This is an interrupt that interrupt the
2533 * previous update. Still more to do.
2536 case RB_PAGE_NORMAL:
2538 * An interrupt came in before the update
2539 * and processed this for us.
2540 * Nothing left to do.
2545 * The reader is on another CPU and just did
2546 * a swap with our next_page.
2551 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2556 * Now that we are here, the old head pointer is
2557 * set to UPDATE. This will keep the reader from
2558 * swapping the head page with the reader page.
2559 * The reader (on another CPU) will spin till
2562 * We just need to protect against interrupts
2563 * doing the job. We will set the next pointer
2564 * to HEAD. After that, we set the old pointer
2565 * to NORMAL, but only if it was HEAD before.
2566 * otherwise we are an interrupt, and only
2567 * want the outer most commit to reset it.
2569 new_head = next_page;
2570 rb_inc_page(&new_head);
2572 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2576 * Valid returns are:
2577 * HEAD - an interrupt came in and already set it.
2578 * NORMAL - One of two things:
2579 * 1) We really set it.
2580 * 2) A bunch of interrupts came in and moved
2581 * the page forward again.
2585 case RB_PAGE_NORMAL:
2589 RB_WARN_ON(cpu_buffer, 1);
2594 * It is possible that an interrupt came in,
2595 * set the head up, then more interrupts came in
2596 * and moved it again. When we get back here,
2597 * the page would have been set to NORMAL but we
2598 * just set it back to HEAD.
2600 * How do you detect this? Well, if that happened
2601 * the tail page would have moved.
2603 if (ret == RB_PAGE_NORMAL) {
2604 struct buffer_page *buffer_tail_page;
2606 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2608 * If the tail had moved passed next, then we need
2609 * to reset the pointer.
2611 if (buffer_tail_page != tail_page &&
2612 buffer_tail_page != next_page)
2613 rb_head_page_set_normal(cpu_buffer, new_head,
2619 * If this was the outer most commit (the one that
2620 * changed the original pointer from HEAD to UPDATE),
2621 * then it is up to us to reset it to NORMAL.
2623 if (type == RB_PAGE_HEAD) {
2624 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2627 if (RB_WARN_ON(cpu_buffer,
2628 ret != RB_PAGE_UPDATE))
2636 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2637 unsigned long tail, struct rb_event_info *info)
2639 struct buffer_page *tail_page = info->tail_page;
2640 struct ring_buffer_event *event;
2641 unsigned long length = info->length;
2644 * Only the event that crossed the page boundary
2645 * must fill the old tail_page with padding.
2647 if (tail >= BUF_PAGE_SIZE) {
2649 * If the page was filled, then we still need
2650 * to update the real_end. Reset it to zero
2651 * and the reader will ignore it.
2653 if (tail == BUF_PAGE_SIZE)
2654 tail_page->real_end = 0;
2656 local_sub(length, &tail_page->write);
2660 event = __rb_page_index(tail_page, tail);
2662 /* account for padding bytes */
2663 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2666 * Save the original length to the meta data.
2667 * This will be used by the reader to add lost event
2670 tail_page->real_end = tail;
2673 * If this event is bigger than the minimum size, then
2674 * we need to be careful that we don't subtract the
2675 * write counter enough to allow another writer to slip
2677 * We put in a discarded commit instead, to make sure
2678 * that this space is not used again.
2680 * If we are less than the minimum size, we don't need to
2683 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2684 /* No room for any events */
2686 /* Mark the rest of the page with padding */
2687 rb_event_set_padding(event);
2689 /* Make sure the padding is visible before the write update */
2692 /* Set the write back to the previous setting */
2693 local_sub(length, &tail_page->write);
2697 /* Put in a discarded event */
2698 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2699 event->type_len = RINGBUF_TYPE_PADDING;
2700 /* time delta must be non zero */
2701 event->time_delta = 1;
2703 /* Make sure the padding is visible before the tail_page->write update */
2706 /* Set write to end of buffer */
2707 length = (tail + length) - BUF_PAGE_SIZE;
2708 local_sub(length, &tail_page->write);
2711 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2714 * This is the slow path, force gcc not to inline it.
2716 static noinline struct ring_buffer_event *
2717 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2718 unsigned long tail, struct rb_event_info *info)
2720 struct buffer_page *tail_page = info->tail_page;
2721 struct buffer_page *commit_page = cpu_buffer->commit_page;
2722 struct trace_buffer *buffer = cpu_buffer->buffer;
2723 struct buffer_page *next_page;
2726 next_page = tail_page;
2728 rb_inc_page(&next_page);
2731 * If for some reason, we had an interrupt storm that made
2732 * it all the way around the buffer, bail, and warn
2735 if (unlikely(next_page == commit_page)) {
2736 local_inc(&cpu_buffer->commit_overrun);
2741 * This is where the fun begins!
2743 * We are fighting against races between a reader that
2744 * could be on another CPU trying to swap its reader
2745 * page with the buffer head.
2747 * We are also fighting against interrupts coming in and
2748 * moving the head or tail on us as well.
2750 * If the next page is the head page then we have filled
2751 * the buffer, unless the commit page is still on the
2754 if (rb_is_head_page(next_page, &tail_page->list)) {
2757 * If the commit is not on the reader page, then
2758 * move the header page.
2760 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2762 * If we are not in overwrite mode,
2763 * this is easy, just stop here.
2765 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2766 local_inc(&cpu_buffer->dropped_events);
2770 ret = rb_handle_head_page(cpu_buffer,
2779 * We need to be careful here too. The
2780 * commit page could still be on the reader
2781 * page. We could have a small buffer, and
2782 * have filled up the buffer with events
2783 * from interrupts and such, and wrapped.
2785 * Note, if the tail page is also on the
2786 * reader_page, we let it move out.
2788 if (unlikely((cpu_buffer->commit_page !=
2789 cpu_buffer->tail_page) &&
2790 (cpu_buffer->commit_page ==
2791 cpu_buffer->reader_page))) {
2792 local_inc(&cpu_buffer->commit_overrun);
2798 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2802 rb_reset_tail(cpu_buffer, tail, info);
2804 /* Commit what we have for now. */
2805 rb_end_commit(cpu_buffer);
2806 /* rb_end_commit() decs committing */
2807 local_inc(&cpu_buffer->committing);
2809 /* fail and let the caller try again */
2810 return ERR_PTR(-EAGAIN);
2814 rb_reset_tail(cpu_buffer, tail, info);
2820 static struct ring_buffer_event *
2821 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2824 event->type_len = RINGBUF_TYPE_TIME_STAMP;
2826 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2828 /* Not the first event on the page, or not delta? */
2829 if (abs || rb_event_index(event)) {
2830 event->time_delta = delta & TS_MASK;
2831 event->array[0] = delta >> TS_SHIFT;
2833 /* nope, just zero it */
2834 event->time_delta = 0;
2835 event->array[0] = 0;
2838 return skip_time_extend(event);
2841 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2842 static inline bool sched_clock_stable(void)
2849 rb_check_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2850 struct rb_event_info *info)
2854 WARN_ONCE(1, "Delta way too big! %llu ts=%llu before=%llu after=%llu write stamp=%llu\n%s",
2855 (unsigned long long)info->delta,
2856 (unsigned long long)info->ts,
2857 (unsigned long long)info->before,
2858 (unsigned long long)info->after,
2859 (unsigned long long)(rb_time_read(&cpu_buffer->write_stamp, &write_stamp) ? write_stamp : 0),
2860 sched_clock_stable() ? "" :
2861 "If you just came from a suspend/resume,\n"
2862 "please switch to the trace global clock:\n"
2863 " echo global > /sys/kernel/tracing/trace_clock\n"
2864 "or add trace_clock=global to the kernel command line\n");
2867 static void rb_add_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2868 struct ring_buffer_event **event,
2869 struct rb_event_info *info,
2871 unsigned int *length)
2873 bool abs = info->add_timestamp &
2874 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE);
2876 if (unlikely(info->delta > (1ULL << 59))) {
2878 * Some timers can use more than 59 bits, and when a timestamp
2879 * is added to the buffer, it will lose those bits.
2881 if (abs && (info->ts & TS_MSB)) {
2882 info->delta &= ABS_TS_MASK;
2884 /* did the clock go backwards */
2885 } else if (info->before == info->after && info->before > info->ts) {
2886 /* not interrupted */
2890 * This is possible with a recalibrating of the TSC.
2891 * Do not produce a call stack, but just report it.
2895 pr_warn("Ring buffer clock went backwards: %llu -> %llu\n",
2896 info->before, info->ts);
2899 rb_check_timestamp(cpu_buffer, info);
2903 *event = rb_add_time_stamp(*event, info->delta, abs);
2904 *length -= RB_LEN_TIME_EXTEND;
2909 * rb_update_event - update event type and data
2910 * @cpu_buffer: The per cpu buffer of the @event
2911 * @event: the event to update
2912 * @info: The info to update the @event with (contains length and delta)
2914 * Update the type and data fields of the @event. The length
2915 * is the actual size that is written to the ring buffer,
2916 * and with this, we can determine what to place into the
2920 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2921 struct ring_buffer_event *event,
2922 struct rb_event_info *info)
2924 unsigned length = info->length;
2925 u64 delta = info->delta;
2926 unsigned int nest = local_read(&cpu_buffer->committing) - 1;
2928 if (!WARN_ON_ONCE(nest >= MAX_NEST))
2929 cpu_buffer->event_stamp[nest] = info->ts;
2932 * If we need to add a timestamp, then we
2933 * add it to the start of the reserved space.
2935 if (unlikely(info->add_timestamp))
2936 rb_add_timestamp(cpu_buffer, &event, info, &delta, &length);
2938 event->time_delta = delta;
2939 length -= RB_EVNT_HDR_SIZE;
2940 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2941 event->type_len = 0;
2942 event->array[0] = length;
2944 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2947 static unsigned rb_calculate_event_length(unsigned length)
2949 struct ring_buffer_event event; /* Used only for sizeof array */
2951 /* zero length can cause confusions */
2955 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2956 length += sizeof(event.array[0]);
2958 length += RB_EVNT_HDR_SIZE;
2959 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2962 * In case the time delta is larger than the 27 bits for it
2963 * in the header, we need to add a timestamp. If another
2964 * event comes in when trying to discard this one to increase
2965 * the length, then the timestamp will be added in the allocated
2966 * space of this event. If length is bigger than the size needed
2967 * for the TIME_EXTEND, then padding has to be used. The events
2968 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2969 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2970 * As length is a multiple of 4, we only need to worry if it
2971 * is 12 (RB_LEN_TIME_EXTEND + 4).
2973 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2974 length += RB_ALIGNMENT;
2979 static u64 rb_time_delta(struct ring_buffer_event *event)
2981 switch (event->type_len) {
2982 case RINGBUF_TYPE_PADDING:
2985 case RINGBUF_TYPE_TIME_EXTEND:
2986 return rb_event_time_stamp(event);
2988 case RINGBUF_TYPE_TIME_STAMP:
2991 case RINGBUF_TYPE_DATA:
2992 return event->time_delta;
2999 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
3000 struct ring_buffer_event *event)
3002 unsigned long new_index, old_index;
3003 struct buffer_page *bpage;
3004 unsigned long index;
3009 new_index = rb_event_index(event);
3010 old_index = new_index + rb_event_ts_length(event);
3011 addr = (unsigned long)event;
3014 bpage = READ_ONCE(cpu_buffer->tail_page);
3016 delta = rb_time_delta(event);
3018 if (!rb_time_read(&cpu_buffer->write_stamp, &write_stamp))
3021 /* Make sure the write stamp is read before testing the location */
3024 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
3025 unsigned long write_mask =
3026 local_read(&bpage->write) & ~RB_WRITE_MASK;
3027 unsigned long event_length = rb_event_length(event);
3029 /* Something came in, can't discard */
3030 if (!rb_time_cmpxchg(&cpu_buffer->write_stamp,
3031 write_stamp, write_stamp - delta))
3035 * It's possible that the event time delta is zero
3036 * (has the same time stamp as the previous event)
3037 * in which case write_stamp and before_stamp could
3038 * be the same. In such a case, force before_stamp
3039 * to be different than write_stamp. It doesn't
3040 * matter what it is, as long as its different.
3043 rb_time_set(&cpu_buffer->before_stamp, 0);
3046 * If an event were to come in now, it would see that the
3047 * write_stamp and the before_stamp are different, and assume
3048 * that this event just added itself before updating
3049 * the write stamp. The interrupting event will fix the
3050 * write stamp for us, and use the before stamp as its delta.
3054 * This is on the tail page. It is possible that
3055 * a write could come in and move the tail page
3056 * and write to the next page. That is fine
3057 * because we just shorten what is on this page.
3059 old_index += write_mask;
3060 new_index += write_mask;
3061 index = local_cmpxchg(&bpage->write, old_index, new_index);
3062 if (index == old_index) {
3063 /* update counters */
3064 local_sub(event_length, &cpu_buffer->entries_bytes);
3069 /* could not discard */
3073 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
3075 local_inc(&cpu_buffer->committing);
3076 local_inc(&cpu_buffer->commits);
3079 static __always_inline void
3080 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
3082 unsigned long max_count;
3085 * We only race with interrupts and NMIs on this CPU.
3086 * If we own the commit event, then we can commit
3087 * all others that interrupted us, since the interruptions
3088 * are in stack format (they finish before they come
3089 * back to us). This allows us to do a simple loop to
3090 * assign the commit to the tail.
3093 max_count = cpu_buffer->nr_pages * 100;
3095 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
3096 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
3098 if (RB_WARN_ON(cpu_buffer,
3099 rb_is_reader_page(cpu_buffer->tail_page)))
3101 local_set(&cpu_buffer->commit_page->page->commit,
3102 rb_page_write(cpu_buffer->commit_page));
3103 rb_inc_page(&cpu_buffer->commit_page);
3104 /* add barrier to keep gcc from optimizing too much */
3107 while (rb_commit_index(cpu_buffer) !=
3108 rb_page_write(cpu_buffer->commit_page)) {
3110 local_set(&cpu_buffer->commit_page->page->commit,
3111 rb_page_write(cpu_buffer->commit_page));
3112 RB_WARN_ON(cpu_buffer,
3113 local_read(&cpu_buffer->commit_page->page->commit) &
3118 /* again, keep gcc from optimizing */
3122 * If an interrupt came in just after the first while loop
3123 * and pushed the tail page forward, we will be left with
3124 * a dangling commit that will never go forward.
3126 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
3130 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
3132 unsigned long commits;
3134 if (RB_WARN_ON(cpu_buffer,
3135 !local_read(&cpu_buffer->committing)))
3139 commits = local_read(&cpu_buffer->commits);
3140 /* synchronize with interrupts */
3142 if (local_read(&cpu_buffer->committing) == 1)
3143 rb_set_commit_to_write(cpu_buffer);
3145 local_dec(&cpu_buffer->committing);
3147 /* synchronize with interrupts */
3151 * Need to account for interrupts coming in between the
3152 * updating of the commit page and the clearing of the
3153 * committing counter.
3155 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
3156 !local_read(&cpu_buffer->committing)) {
3157 local_inc(&cpu_buffer->committing);
3162 static inline void rb_event_discard(struct ring_buffer_event *event)
3164 if (extended_time(event))
3165 event = skip_time_extend(event);
3167 /* array[0] holds the actual length for the discarded event */
3168 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
3169 event->type_len = RINGBUF_TYPE_PADDING;
3170 /* time delta must be non zero */
3171 if (!event->time_delta)
3172 event->time_delta = 1;
3175 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer)
3177 local_inc(&cpu_buffer->entries);
3178 rb_end_commit(cpu_buffer);
3181 static __always_inline void
3182 rb_wakeups(struct trace_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
3184 if (buffer->irq_work.waiters_pending) {
3185 buffer->irq_work.waiters_pending = false;
3186 /* irq_work_queue() supplies it's own memory barriers */
3187 irq_work_queue(&buffer->irq_work.work);
3190 if (cpu_buffer->irq_work.waiters_pending) {
3191 cpu_buffer->irq_work.waiters_pending = false;
3192 /* irq_work_queue() supplies it's own memory barriers */
3193 irq_work_queue(&cpu_buffer->irq_work.work);
3196 if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
3199 if (cpu_buffer->reader_page == cpu_buffer->commit_page)
3202 if (!cpu_buffer->irq_work.full_waiters_pending)
3205 cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
3207 if (!full_hit(buffer, cpu_buffer->cpu, cpu_buffer->shortest_full))
3210 cpu_buffer->irq_work.wakeup_full = true;
3211 cpu_buffer->irq_work.full_waiters_pending = false;
3212 /* irq_work_queue() supplies it's own memory barriers */
3213 irq_work_queue(&cpu_buffer->irq_work.work);
3216 #ifdef CONFIG_RING_BUFFER_RECORD_RECURSION
3217 # define do_ring_buffer_record_recursion() \
3218 do_ftrace_record_recursion(_THIS_IP_, _RET_IP_)
3220 # define do_ring_buffer_record_recursion() do { } while (0)
3224 * The lock and unlock are done within a preempt disable section.
3225 * The current_context per_cpu variable can only be modified
3226 * by the current task between lock and unlock. But it can
3227 * be modified more than once via an interrupt. To pass this
3228 * information from the lock to the unlock without having to
3229 * access the 'in_interrupt()' functions again (which do show
3230 * a bit of overhead in something as critical as function tracing,
3231 * we use a bitmask trick.
3233 * bit 1 = NMI context
3234 * bit 2 = IRQ context
3235 * bit 3 = SoftIRQ context
3236 * bit 4 = normal context.
3238 * This works because this is the order of contexts that can
3239 * preempt other contexts. A SoftIRQ never preempts an IRQ
3242 * When the context is determined, the corresponding bit is
3243 * checked and set (if it was set, then a recursion of that context
3246 * On unlock, we need to clear this bit. To do so, just subtract
3247 * 1 from the current_context and AND it to itself.
3251 * 101 & 100 = 100 (clearing bit zero)
3254 * 1010 & 1001 = 1000 (clearing bit 1)
3256 * The least significant bit can be cleared this way, and it
3257 * just so happens that it is the same bit corresponding to
3258 * the current context.
3260 * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
3261 * is set when a recursion is detected at the current context, and if
3262 * the TRANSITION bit is already set, it will fail the recursion.
3263 * This is needed because there's a lag between the changing of
3264 * interrupt context and updating the preempt count. In this case,
3265 * a false positive will be found. To handle this, one extra recursion
3266 * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
3267 * bit is already set, then it is considered a recursion and the function
3268 * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
3270 * On the trace_recursive_unlock(), the TRANSITION bit will be the first
3271 * to be cleared. Even if it wasn't the context that set it. That is,
3272 * if an interrupt comes in while NORMAL bit is set and the ring buffer
3273 * is called before preempt_count() is updated, since the check will
3274 * be on the NORMAL bit, the TRANSITION bit will then be set. If an
3275 * NMI then comes in, it will set the NMI bit, but when the NMI code
3276 * does the trace_recursive_unlock() it will clear the TRANSITION bit
3277 * and leave the NMI bit set. But this is fine, because the interrupt
3278 * code that set the TRANSITION bit will then clear the NMI bit when it
3279 * calls trace_recursive_unlock(). If another NMI comes in, it will
3280 * set the TRANSITION bit and continue.
3282 * Note: The TRANSITION bit only handles a single transition between context.
3285 static __always_inline int
3286 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
3288 unsigned int val = cpu_buffer->current_context;
3289 int bit = interrupt_context_level();
3291 bit = RB_CTX_NORMAL - bit;
3293 if (unlikely(val & (1 << (bit + cpu_buffer->nest)))) {
3295 * It is possible that this was called by transitioning
3296 * between interrupt context, and preempt_count() has not
3297 * been updated yet. In this case, use the TRANSITION bit.
3299 bit = RB_CTX_TRANSITION;
3300 if (val & (1 << (bit + cpu_buffer->nest))) {
3301 do_ring_buffer_record_recursion();
3306 val |= (1 << (bit + cpu_buffer->nest));
3307 cpu_buffer->current_context = val;
3312 static __always_inline void
3313 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
3315 cpu_buffer->current_context &=
3316 cpu_buffer->current_context - (1 << cpu_buffer->nest);
3319 /* The recursive locking above uses 5 bits */
3320 #define NESTED_BITS 5
3323 * ring_buffer_nest_start - Allow to trace while nested
3324 * @buffer: The ring buffer to modify
3326 * The ring buffer has a safety mechanism to prevent recursion.
3327 * But there may be a case where a trace needs to be done while
3328 * tracing something else. In this case, calling this function
3329 * will allow this function to nest within a currently active
3330 * ring_buffer_lock_reserve().
3332 * Call this function before calling another ring_buffer_lock_reserve() and
3333 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
3335 void ring_buffer_nest_start(struct trace_buffer *buffer)
3337 struct ring_buffer_per_cpu *cpu_buffer;
3340 /* Enabled by ring_buffer_nest_end() */
3341 preempt_disable_notrace();
3342 cpu = raw_smp_processor_id();
3343 cpu_buffer = buffer->buffers[cpu];
3344 /* This is the shift value for the above recursive locking */
3345 cpu_buffer->nest += NESTED_BITS;
3349 * ring_buffer_nest_end - Allow to trace while nested
3350 * @buffer: The ring buffer to modify
3352 * Must be called after ring_buffer_nest_start() and after the
3353 * ring_buffer_unlock_commit().
3355 void ring_buffer_nest_end(struct trace_buffer *buffer)
3357 struct ring_buffer_per_cpu *cpu_buffer;
3360 /* disabled by ring_buffer_nest_start() */
3361 cpu = raw_smp_processor_id();
3362 cpu_buffer = buffer->buffers[cpu];
3363 /* This is the shift value for the above recursive locking */
3364 cpu_buffer->nest -= NESTED_BITS;
3365 preempt_enable_notrace();
3369 * ring_buffer_unlock_commit - commit a reserved
3370 * @buffer: The buffer to commit to
3371 * @event: The event pointer to commit.
3373 * This commits the data to the ring buffer, and releases any locks held.
3375 * Must be paired with ring_buffer_lock_reserve.
3377 int ring_buffer_unlock_commit(struct trace_buffer *buffer)
3379 struct ring_buffer_per_cpu *cpu_buffer;
3380 int cpu = raw_smp_processor_id();
3382 cpu_buffer = buffer->buffers[cpu];
3384 rb_commit(cpu_buffer);
3386 rb_wakeups(buffer, cpu_buffer);
3388 trace_recursive_unlock(cpu_buffer);
3390 preempt_enable_notrace();
3394 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
3396 /* Special value to validate all deltas on a page. */
3397 #define CHECK_FULL_PAGE 1L
3399 #ifdef CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS
3400 static void dump_buffer_page(struct buffer_data_page *bpage,
3401 struct rb_event_info *info,
3404 struct ring_buffer_event *event;
3408 ts = bpage->time_stamp;
3409 pr_warn(" [%lld] PAGE TIME STAMP\n", ts);
3411 for (e = 0; e < tail; e += rb_event_length(event)) {
3413 event = (struct ring_buffer_event *)(bpage->data + e);
3415 switch (event->type_len) {
3417 case RINGBUF_TYPE_TIME_EXTEND:
3418 delta = rb_event_time_stamp(event);
3420 pr_warn(" [%lld] delta:%lld TIME EXTEND\n", ts, delta);
3423 case RINGBUF_TYPE_TIME_STAMP:
3424 delta = rb_event_time_stamp(event);
3425 ts = rb_fix_abs_ts(delta, ts);
3426 pr_warn(" [%lld] absolute:%lld TIME STAMP\n", ts, delta);
3429 case RINGBUF_TYPE_PADDING:
3430 ts += event->time_delta;
3431 pr_warn(" [%lld] delta:%d PADDING\n", ts, event->time_delta);
3434 case RINGBUF_TYPE_DATA:
3435 ts += event->time_delta;
3436 pr_warn(" [%lld] delta:%d\n", ts, event->time_delta);
3445 static DEFINE_PER_CPU(atomic_t, checking);
3446 static atomic_t ts_dump;
3449 * Check if the current event time stamp matches the deltas on
3452 static void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3453 struct rb_event_info *info,
3456 struct ring_buffer_event *event;
3457 struct buffer_data_page *bpage;
3462 bpage = info->tail_page->page;
3464 if (tail == CHECK_FULL_PAGE) {
3466 tail = local_read(&bpage->commit);
3467 } else if (info->add_timestamp &
3468 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE)) {
3469 /* Ignore events with absolute time stamps */
3474 * Do not check the first event (skip possible extends too).
3475 * Also do not check if previous events have not been committed.
3477 if (tail <= 8 || tail > local_read(&bpage->commit))
3481 * If this interrupted another event,
3483 if (atomic_inc_return(this_cpu_ptr(&checking)) != 1)
3486 ts = bpage->time_stamp;
3488 for (e = 0; e < tail; e += rb_event_length(event)) {
3490 event = (struct ring_buffer_event *)(bpage->data + e);
3492 switch (event->type_len) {
3494 case RINGBUF_TYPE_TIME_EXTEND:
3495 delta = rb_event_time_stamp(event);
3499 case RINGBUF_TYPE_TIME_STAMP:
3500 delta = rb_event_time_stamp(event);
3501 ts = rb_fix_abs_ts(delta, ts);
3504 case RINGBUF_TYPE_PADDING:
3505 if (event->time_delta == 1)
3508 case RINGBUF_TYPE_DATA:
3509 ts += event->time_delta;
3513 RB_WARN_ON(cpu_buffer, 1);
3516 if ((full && ts > info->ts) ||
3517 (!full && ts + info->delta != info->ts)) {
3518 /* If another report is happening, ignore this one */
3519 if (atomic_inc_return(&ts_dump) != 1) {
3520 atomic_dec(&ts_dump);
3523 atomic_inc(&cpu_buffer->record_disabled);
3524 /* There's some cases in boot up that this can happen */
3525 WARN_ON_ONCE(system_state != SYSTEM_BOOTING);
3526 pr_warn("[CPU: %d]TIME DOES NOT MATCH expected:%lld actual:%lld delta:%lld before:%lld after:%lld%s\n",
3528 ts + info->delta, info->ts, info->delta,
3529 info->before, info->after,
3530 full ? " (full)" : "");
3531 dump_buffer_page(bpage, info, tail);
3532 atomic_dec(&ts_dump);
3533 /* Do not re-enable checking */
3537 atomic_dec(this_cpu_ptr(&checking));
3540 static inline void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3541 struct rb_event_info *info,
3545 #endif /* CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS */
3547 static struct ring_buffer_event *
3548 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
3549 struct rb_event_info *info)
3551 struct ring_buffer_event *event;
3552 struct buffer_page *tail_page;
3553 unsigned long tail, write, w;
3557 /* Don't let the compiler play games with cpu_buffer->tail_page */
3558 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
3560 /*A*/ w = local_read(&tail_page->write) & RB_WRITE_MASK;
3562 b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3563 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3565 info->ts = rb_time_stamp(cpu_buffer->buffer);
3567 if ((info->add_timestamp & RB_ADD_STAMP_ABSOLUTE)) {
3568 info->delta = info->ts;
3571 * If interrupting an event time update, we may need an
3572 * absolute timestamp.
3573 * Don't bother if this is the start of a new page (w == 0).
3575 if (unlikely(!a_ok || !b_ok || (info->before != info->after && w))) {
3576 info->add_timestamp |= RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND;
3577 info->length += RB_LEN_TIME_EXTEND;
3579 info->delta = info->ts - info->after;
3580 if (unlikely(test_time_stamp(info->delta))) {
3581 info->add_timestamp |= RB_ADD_STAMP_EXTEND;
3582 info->length += RB_LEN_TIME_EXTEND;
3587 /*B*/ rb_time_set(&cpu_buffer->before_stamp, info->ts);
3589 /*C*/ write = local_add_return(info->length, &tail_page->write);
3591 /* set write to only the index of the write */
3592 write &= RB_WRITE_MASK;
3594 tail = write - info->length;
3596 /* See if we shot pass the end of this buffer page */
3597 if (unlikely(write > BUF_PAGE_SIZE)) {
3598 /* before and after may now different, fix it up*/
3599 b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3600 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3601 if (a_ok && b_ok && info->before != info->after)
3602 (void)rb_time_cmpxchg(&cpu_buffer->before_stamp,
3603 info->before, info->after);
3605 check_buffer(cpu_buffer, info, CHECK_FULL_PAGE);
3606 return rb_move_tail(cpu_buffer, tail, info);
3609 if (likely(tail == w)) {
3613 /* Nothing interrupted us between A and C */
3614 /*D*/ rb_time_set(&cpu_buffer->write_stamp, info->ts);
3616 /*E*/ s_ok = rb_time_read(&cpu_buffer->before_stamp, &save_before);
3617 RB_WARN_ON(cpu_buffer, !s_ok);
3618 if (likely(!(info->add_timestamp &
3619 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3620 /* This did not interrupt any time update */
3621 info->delta = info->ts - info->after;
3623 /* Just use full timestamp for interrupting event */
3624 info->delta = info->ts;
3626 check_buffer(cpu_buffer, info, tail);
3627 if (unlikely(info->ts != save_before)) {
3628 /* SLOW PATH - Interrupted between C and E */
3630 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3631 RB_WARN_ON(cpu_buffer, !a_ok);
3633 /* Write stamp must only go forward */
3634 if (save_before > info->after) {
3636 * We do not care about the result, only that
3637 * it gets updated atomically.
3639 (void)rb_time_cmpxchg(&cpu_buffer->write_stamp,
3640 info->after, save_before);
3645 /* SLOW PATH - Interrupted between A and C */
3646 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3647 /* Was interrupted before here, write_stamp must be valid */
3648 RB_WARN_ON(cpu_buffer, !a_ok);
3649 ts = rb_time_stamp(cpu_buffer->buffer);
3651 /*E*/ if (write == (local_read(&tail_page->write) & RB_WRITE_MASK) &&
3653 rb_time_cmpxchg(&cpu_buffer->write_stamp,
3655 /* Nothing came after this event between C and E */
3656 info->delta = ts - info->after;
3659 * Interrupted between C and E:
3660 * Lost the previous events time stamp. Just set the
3661 * delta to zero, and this will be the same time as
3662 * the event this event interrupted. And the events that
3663 * came after this will still be correct (as they would
3664 * have built their delta on the previous event.
3669 info->add_timestamp &= ~RB_ADD_STAMP_FORCE;
3673 * If this is the first commit on the page, then it has the same
3674 * timestamp as the page itself.
3676 if (unlikely(!tail && !(info->add_timestamp &
3677 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3680 /* We reserved something on the buffer */
3682 event = __rb_page_index(tail_page, tail);
3683 rb_update_event(cpu_buffer, event, info);
3685 local_inc(&tail_page->entries);
3688 * If this is the first commit on the page, then update
3691 if (unlikely(!tail))
3692 tail_page->page->time_stamp = info->ts;
3694 /* account for these added bytes */
3695 local_add(info->length, &cpu_buffer->entries_bytes);
3700 static __always_inline struct ring_buffer_event *
3701 rb_reserve_next_event(struct trace_buffer *buffer,
3702 struct ring_buffer_per_cpu *cpu_buffer,
3703 unsigned long length)
3705 struct ring_buffer_event *event;
3706 struct rb_event_info info;
3710 rb_start_commit(cpu_buffer);
3711 /* The commit page can not change after this */
3713 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3715 * Due to the ability to swap a cpu buffer from a buffer
3716 * it is possible it was swapped before we committed.
3717 * (committing stops a swap). We check for it here and
3718 * if it happened, we have to fail the write.
3721 if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
3722 local_dec(&cpu_buffer->committing);
3723 local_dec(&cpu_buffer->commits);
3728 info.length = rb_calculate_event_length(length);
3730 if (ring_buffer_time_stamp_abs(cpu_buffer->buffer)) {
3731 add_ts_default = RB_ADD_STAMP_ABSOLUTE;
3732 info.length += RB_LEN_TIME_EXTEND;
3734 add_ts_default = RB_ADD_STAMP_NONE;
3738 info.add_timestamp = add_ts_default;
3742 * We allow for interrupts to reenter here and do a trace.
3743 * If one does, it will cause this original code to loop
3744 * back here. Even with heavy interrupts happening, this
3745 * should only happen a few times in a row. If this happens
3746 * 1000 times in a row, there must be either an interrupt
3747 * storm or we have something buggy.
3750 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
3753 event = __rb_reserve_next(cpu_buffer, &info);
3755 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
3756 if (info.add_timestamp & (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND))
3757 info.length -= RB_LEN_TIME_EXTEND;
3764 rb_end_commit(cpu_buffer);
3769 * ring_buffer_lock_reserve - reserve a part of the buffer
3770 * @buffer: the ring buffer to reserve from
3771 * @length: the length of the data to reserve (excluding event header)
3773 * Returns a reserved event on the ring buffer to copy directly to.
3774 * The user of this interface will need to get the body to write into
3775 * and can use the ring_buffer_event_data() interface.
3777 * The length is the length of the data needed, not the event length
3778 * which also includes the event header.
3780 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
3781 * If NULL is returned, then nothing has been allocated or locked.
3783 struct ring_buffer_event *
3784 ring_buffer_lock_reserve(struct trace_buffer *buffer, unsigned long length)
3786 struct ring_buffer_per_cpu *cpu_buffer;
3787 struct ring_buffer_event *event;
3790 /* If we are tracing schedule, we don't want to recurse */
3791 preempt_disable_notrace();
3793 if (unlikely(atomic_read(&buffer->record_disabled)))
3796 cpu = raw_smp_processor_id();
3798 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
3801 cpu_buffer = buffer->buffers[cpu];
3803 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
3806 if (unlikely(length > BUF_MAX_DATA_SIZE))
3809 if (unlikely(trace_recursive_lock(cpu_buffer)))
3812 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3819 trace_recursive_unlock(cpu_buffer);
3821 preempt_enable_notrace();
3824 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3827 * Decrement the entries to the page that an event is on.
3828 * The event does not even need to exist, only the pointer
3829 * to the page it is on. This may only be called before the commit
3833 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3834 struct ring_buffer_event *event)
3836 unsigned long addr = (unsigned long)event;
3837 struct buffer_page *bpage = cpu_buffer->commit_page;
3838 struct buffer_page *start;
3842 /* Do the likely case first */
3843 if (likely(bpage->page == (void *)addr)) {
3844 local_dec(&bpage->entries);
3849 * Because the commit page may be on the reader page we
3850 * start with the next page and check the end loop there.
3852 rb_inc_page(&bpage);
3855 if (bpage->page == (void *)addr) {
3856 local_dec(&bpage->entries);
3859 rb_inc_page(&bpage);
3860 } while (bpage != start);
3862 /* commit not part of this buffer?? */
3863 RB_WARN_ON(cpu_buffer, 1);
3867 * ring_buffer_discard_commit - discard an event that has not been committed
3868 * @buffer: the ring buffer
3869 * @event: non committed event to discard
3871 * Sometimes an event that is in the ring buffer needs to be ignored.
3872 * This function lets the user discard an event in the ring buffer
3873 * and then that event will not be read later.
3875 * This function only works if it is called before the item has been
3876 * committed. It will try to free the event from the ring buffer
3877 * if another event has not been added behind it.
3879 * If another event has been added behind it, it will set the event
3880 * up as discarded, and perform the commit.
3882 * If this function is called, do not call ring_buffer_unlock_commit on
3885 void ring_buffer_discard_commit(struct trace_buffer *buffer,
3886 struct ring_buffer_event *event)
3888 struct ring_buffer_per_cpu *cpu_buffer;
3891 /* The event is discarded regardless */
3892 rb_event_discard(event);
3894 cpu = smp_processor_id();
3895 cpu_buffer = buffer->buffers[cpu];
3898 * This must only be called if the event has not been
3899 * committed yet. Thus we can assume that preemption
3900 * is still disabled.
3902 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3904 rb_decrement_entry(cpu_buffer, event);
3905 if (rb_try_to_discard(cpu_buffer, event))
3909 rb_end_commit(cpu_buffer);
3911 trace_recursive_unlock(cpu_buffer);
3913 preempt_enable_notrace();
3916 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3919 * ring_buffer_write - write data to the buffer without reserving
3920 * @buffer: The ring buffer to write to.
3921 * @length: The length of the data being written (excluding the event header)
3922 * @data: The data to write to the buffer.
3924 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3925 * one function. If you already have the data to write to the buffer, it
3926 * may be easier to simply call this function.
3928 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3929 * and not the length of the event which would hold the header.
3931 int ring_buffer_write(struct trace_buffer *buffer,
3932 unsigned long length,
3935 struct ring_buffer_per_cpu *cpu_buffer;
3936 struct ring_buffer_event *event;
3941 preempt_disable_notrace();
3943 if (atomic_read(&buffer->record_disabled))
3946 cpu = raw_smp_processor_id();
3948 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3951 cpu_buffer = buffer->buffers[cpu];
3953 if (atomic_read(&cpu_buffer->record_disabled))
3956 if (length > BUF_MAX_DATA_SIZE)
3959 if (unlikely(trace_recursive_lock(cpu_buffer)))
3962 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3966 body = rb_event_data(event);
3968 memcpy(body, data, length);
3970 rb_commit(cpu_buffer);
3972 rb_wakeups(buffer, cpu_buffer);
3977 trace_recursive_unlock(cpu_buffer);
3980 preempt_enable_notrace();
3984 EXPORT_SYMBOL_GPL(ring_buffer_write);
3986 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3988 struct buffer_page *reader = cpu_buffer->reader_page;
3989 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3990 struct buffer_page *commit = cpu_buffer->commit_page;
3992 /* In case of error, head will be NULL */
3993 if (unlikely(!head))
3996 /* Reader should exhaust content in reader page */
3997 if (reader->read != rb_page_commit(reader))
4001 * If writers are committing on the reader page, knowing all
4002 * committed content has been read, the ring buffer is empty.
4004 if (commit == reader)
4008 * If writers are committing on a page other than reader page
4009 * and head page, there should always be content to read.
4015 * Writers are committing on the head page, we just need
4016 * to care about there're committed data, and the reader will
4017 * swap reader page with head page when it is to read data.
4019 return rb_page_commit(commit) == 0;
4023 * ring_buffer_record_disable - stop all writes into the buffer
4024 * @buffer: The ring buffer to stop writes to.
4026 * This prevents all writes to the buffer. Any attempt to write
4027 * to the buffer after this will fail and return NULL.
4029 * The caller should call synchronize_rcu() after this.
4031 void ring_buffer_record_disable(struct trace_buffer *buffer)
4033 atomic_inc(&buffer->record_disabled);
4035 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
4038 * ring_buffer_record_enable - enable writes to the buffer
4039 * @buffer: The ring buffer to enable writes
4041 * Note, multiple disables will need the same number of enables
4042 * to truly enable the writing (much like preempt_disable).
4044 void ring_buffer_record_enable(struct trace_buffer *buffer)
4046 atomic_dec(&buffer->record_disabled);
4048 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
4051 * ring_buffer_record_off - stop all writes into the buffer
4052 * @buffer: The ring buffer to stop writes to.
4054 * This prevents all writes to the buffer. Any attempt to write
4055 * to the buffer after this will fail and return NULL.
4057 * This is different than ring_buffer_record_disable() as
4058 * it works like an on/off switch, where as the disable() version
4059 * must be paired with a enable().
4061 void ring_buffer_record_off(struct trace_buffer *buffer)
4064 unsigned int new_rd;
4067 rd = atomic_read(&buffer->record_disabled);
4068 new_rd = rd | RB_BUFFER_OFF;
4069 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
4071 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
4074 * ring_buffer_record_on - restart writes into the buffer
4075 * @buffer: The ring buffer to start writes to.
4077 * This enables all writes to the buffer that was disabled by
4078 * ring_buffer_record_off().
4080 * This is different than ring_buffer_record_enable() as
4081 * it works like an on/off switch, where as the enable() version
4082 * must be paired with a disable().
4084 void ring_buffer_record_on(struct trace_buffer *buffer)
4087 unsigned int new_rd;
4090 rd = atomic_read(&buffer->record_disabled);
4091 new_rd = rd & ~RB_BUFFER_OFF;
4092 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
4094 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
4097 * ring_buffer_record_is_on - return true if the ring buffer can write
4098 * @buffer: The ring buffer to see if write is enabled
4100 * Returns true if the ring buffer is in a state that it accepts writes.
4102 bool ring_buffer_record_is_on(struct trace_buffer *buffer)
4104 return !atomic_read(&buffer->record_disabled);
4108 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
4109 * @buffer: The ring buffer to see if write is set enabled
4111 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
4112 * Note that this does NOT mean it is in a writable state.
4114 * It may return true when the ring buffer has been disabled by
4115 * ring_buffer_record_disable(), as that is a temporary disabling of
4118 bool ring_buffer_record_is_set_on(struct trace_buffer *buffer)
4120 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
4124 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
4125 * @buffer: The ring buffer to stop writes to.
4126 * @cpu: The CPU buffer to stop
4128 * This prevents all writes to the buffer. Any attempt to write
4129 * to the buffer after this will fail and return NULL.
4131 * The caller should call synchronize_rcu() after this.
4133 void ring_buffer_record_disable_cpu(struct trace_buffer *buffer, int cpu)
4135 struct ring_buffer_per_cpu *cpu_buffer;
4137 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4140 cpu_buffer = buffer->buffers[cpu];
4141 atomic_inc(&cpu_buffer->record_disabled);
4143 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
4146 * ring_buffer_record_enable_cpu - enable writes to the buffer
4147 * @buffer: The ring buffer to enable writes
4148 * @cpu: The CPU to enable.
4150 * Note, multiple disables will need the same number of enables
4151 * to truly enable the writing (much like preempt_disable).
4153 void ring_buffer_record_enable_cpu(struct trace_buffer *buffer, int cpu)
4155 struct ring_buffer_per_cpu *cpu_buffer;
4157 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4160 cpu_buffer = buffer->buffers[cpu];
4161 atomic_dec(&cpu_buffer->record_disabled);
4163 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
4166 * The total entries in the ring buffer is the running counter
4167 * of entries entered into the ring buffer, minus the sum of
4168 * the entries read from the ring buffer and the number of
4169 * entries that were overwritten.
4171 static inline unsigned long
4172 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
4174 return local_read(&cpu_buffer->entries) -
4175 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
4179 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
4180 * @buffer: The ring buffer
4181 * @cpu: The per CPU buffer to read from.
4183 u64 ring_buffer_oldest_event_ts(struct trace_buffer *buffer, int cpu)
4185 unsigned long flags;
4186 struct ring_buffer_per_cpu *cpu_buffer;
4187 struct buffer_page *bpage;
4190 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4193 cpu_buffer = buffer->buffers[cpu];
4194 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4196 * if the tail is on reader_page, oldest time stamp is on the reader
4199 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
4200 bpage = cpu_buffer->reader_page;
4202 bpage = rb_set_head_page(cpu_buffer);
4204 ret = bpage->page->time_stamp;
4205 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4209 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
4212 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
4213 * @buffer: The ring buffer
4214 * @cpu: The per CPU buffer to read from.
4216 unsigned long ring_buffer_bytes_cpu(struct trace_buffer *buffer, int cpu)
4218 struct ring_buffer_per_cpu *cpu_buffer;
4221 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4224 cpu_buffer = buffer->buffers[cpu];
4225 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
4229 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
4232 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
4233 * @buffer: The ring buffer
4234 * @cpu: The per CPU buffer to get the entries from.
4236 unsigned long ring_buffer_entries_cpu(struct trace_buffer *buffer, int cpu)
4238 struct ring_buffer_per_cpu *cpu_buffer;
4240 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4243 cpu_buffer = buffer->buffers[cpu];
4245 return rb_num_of_entries(cpu_buffer);
4247 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
4250 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
4251 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
4252 * @buffer: The ring buffer
4253 * @cpu: The per CPU buffer to get the number of overruns from
4255 unsigned long ring_buffer_overrun_cpu(struct trace_buffer *buffer, int cpu)
4257 struct ring_buffer_per_cpu *cpu_buffer;
4260 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4263 cpu_buffer = buffer->buffers[cpu];
4264 ret = local_read(&cpu_buffer->overrun);
4268 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
4271 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
4272 * commits failing due to the buffer wrapping around while there are uncommitted
4273 * events, such as during an interrupt storm.
4274 * @buffer: The ring buffer
4275 * @cpu: The per CPU buffer to get the number of overruns from
4278 ring_buffer_commit_overrun_cpu(struct trace_buffer *buffer, int cpu)
4280 struct ring_buffer_per_cpu *cpu_buffer;
4283 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4286 cpu_buffer = buffer->buffers[cpu];
4287 ret = local_read(&cpu_buffer->commit_overrun);
4291 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
4294 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
4295 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
4296 * @buffer: The ring buffer
4297 * @cpu: The per CPU buffer to get the number of overruns from
4300 ring_buffer_dropped_events_cpu(struct trace_buffer *buffer, int cpu)
4302 struct ring_buffer_per_cpu *cpu_buffer;
4305 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4308 cpu_buffer = buffer->buffers[cpu];
4309 ret = local_read(&cpu_buffer->dropped_events);
4313 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
4316 * ring_buffer_read_events_cpu - get the number of events successfully read
4317 * @buffer: The ring buffer
4318 * @cpu: The per CPU buffer to get the number of events read
4321 ring_buffer_read_events_cpu(struct trace_buffer *buffer, int cpu)
4323 struct ring_buffer_per_cpu *cpu_buffer;
4325 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4328 cpu_buffer = buffer->buffers[cpu];
4329 return cpu_buffer->read;
4331 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
4334 * ring_buffer_entries - get the number of entries in a buffer
4335 * @buffer: The ring buffer
4337 * Returns the total number of entries in the ring buffer
4340 unsigned long ring_buffer_entries(struct trace_buffer *buffer)
4342 struct ring_buffer_per_cpu *cpu_buffer;
4343 unsigned long entries = 0;
4346 /* if you care about this being correct, lock the buffer */
4347 for_each_buffer_cpu(buffer, cpu) {
4348 cpu_buffer = buffer->buffers[cpu];
4349 entries += rb_num_of_entries(cpu_buffer);
4354 EXPORT_SYMBOL_GPL(ring_buffer_entries);
4357 * ring_buffer_overruns - get the number of overruns in buffer
4358 * @buffer: The ring buffer
4360 * Returns the total number of overruns in the ring buffer
4363 unsigned long ring_buffer_overruns(struct trace_buffer *buffer)
4365 struct ring_buffer_per_cpu *cpu_buffer;
4366 unsigned long overruns = 0;
4369 /* if you care about this being correct, lock the buffer */
4370 for_each_buffer_cpu(buffer, cpu) {
4371 cpu_buffer = buffer->buffers[cpu];
4372 overruns += local_read(&cpu_buffer->overrun);
4377 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
4379 static void rb_iter_reset(struct ring_buffer_iter *iter)
4381 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4383 /* Iterator usage is expected to have record disabled */
4384 iter->head_page = cpu_buffer->reader_page;
4385 iter->head = cpu_buffer->reader_page->read;
4386 iter->next_event = iter->head;
4388 iter->cache_reader_page = iter->head_page;
4389 iter->cache_read = cpu_buffer->read;
4392 iter->read_stamp = cpu_buffer->read_stamp;
4393 iter->page_stamp = cpu_buffer->reader_page->page->time_stamp;
4395 iter->read_stamp = iter->head_page->page->time_stamp;
4396 iter->page_stamp = iter->read_stamp;
4401 * ring_buffer_iter_reset - reset an iterator
4402 * @iter: The iterator to reset
4404 * Resets the iterator, so that it will start from the beginning
4407 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
4409 struct ring_buffer_per_cpu *cpu_buffer;
4410 unsigned long flags;
4415 cpu_buffer = iter->cpu_buffer;
4417 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4418 rb_iter_reset(iter);
4419 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4421 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
4424 * ring_buffer_iter_empty - check if an iterator has no more to read
4425 * @iter: The iterator to check
4427 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
4429 struct ring_buffer_per_cpu *cpu_buffer;
4430 struct buffer_page *reader;
4431 struct buffer_page *head_page;
4432 struct buffer_page *commit_page;
4433 struct buffer_page *curr_commit_page;
4438 cpu_buffer = iter->cpu_buffer;
4439 reader = cpu_buffer->reader_page;
4440 head_page = cpu_buffer->head_page;
4441 commit_page = cpu_buffer->commit_page;
4442 commit_ts = commit_page->page->time_stamp;
4445 * When the writer goes across pages, it issues a cmpxchg which
4446 * is a mb(), which will synchronize with the rmb here.
4447 * (see rb_tail_page_update())
4450 commit = rb_page_commit(commit_page);
4451 /* We want to make sure that the commit page doesn't change */
4454 /* Make sure commit page didn't change */
4455 curr_commit_page = READ_ONCE(cpu_buffer->commit_page);
4456 curr_commit_ts = READ_ONCE(curr_commit_page->page->time_stamp);
4458 /* If the commit page changed, then there's more data */
4459 if (curr_commit_page != commit_page ||
4460 curr_commit_ts != commit_ts)
4463 /* Still racy, as it may return a false positive, but that's OK */
4464 return ((iter->head_page == commit_page && iter->head >= commit) ||
4465 (iter->head_page == reader && commit_page == head_page &&
4466 head_page->read == commit &&
4467 iter->head == rb_page_commit(cpu_buffer->reader_page)));
4469 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
4472 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
4473 struct ring_buffer_event *event)
4477 switch (event->type_len) {
4478 case RINGBUF_TYPE_PADDING:
4481 case RINGBUF_TYPE_TIME_EXTEND:
4482 delta = rb_event_time_stamp(event);
4483 cpu_buffer->read_stamp += delta;
4486 case RINGBUF_TYPE_TIME_STAMP:
4487 delta = rb_event_time_stamp(event);
4488 delta = rb_fix_abs_ts(delta, cpu_buffer->read_stamp);
4489 cpu_buffer->read_stamp = delta;
4492 case RINGBUF_TYPE_DATA:
4493 cpu_buffer->read_stamp += event->time_delta;
4497 RB_WARN_ON(cpu_buffer, 1);
4503 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
4504 struct ring_buffer_event *event)
4508 switch (event->type_len) {
4509 case RINGBUF_TYPE_PADDING:
4512 case RINGBUF_TYPE_TIME_EXTEND:
4513 delta = rb_event_time_stamp(event);
4514 iter->read_stamp += delta;
4517 case RINGBUF_TYPE_TIME_STAMP:
4518 delta = rb_event_time_stamp(event);
4519 delta = rb_fix_abs_ts(delta, iter->read_stamp);
4520 iter->read_stamp = delta;
4523 case RINGBUF_TYPE_DATA:
4524 iter->read_stamp += event->time_delta;
4528 RB_WARN_ON(iter->cpu_buffer, 1);
4533 static struct buffer_page *
4534 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
4536 struct buffer_page *reader = NULL;
4537 unsigned long overwrite;
4538 unsigned long flags;
4542 local_irq_save(flags);
4543 arch_spin_lock(&cpu_buffer->lock);
4547 * This should normally only loop twice. But because the
4548 * start of the reader inserts an empty page, it causes
4549 * a case where we will loop three times. There should be no
4550 * reason to loop four times (that I know of).
4552 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
4557 reader = cpu_buffer->reader_page;
4559 /* If there's more to read, return this page */
4560 if (cpu_buffer->reader_page->read < rb_page_size(reader))
4563 /* Never should we have an index greater than the size */
4564 if (RB_WARN_ON(cpu_buffer,
4565 cpu_buffer->reader_page->read > rb_page_size(reader)))
4568 /* check if we caught up to the tail */
4570 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
4573 /* Don't bother swapping if the ring buffer is empty */
4574 if (rb_num_of_entries(cpu_buffer) == 0)
4578 * Reset the reader page to size zero.
4580 local_set(&cpu_buffer->reader_page->write, 0);
4581 local_set(&cpu_buffer->reader_page->entries, 0);
4582 local_set(&cpu_buffer->reader_page->page->commit, 0);
4583 cpu_buffer->reader_page->real_end = 0;
4587 * Splice the empty reader page into the list around the head.
4589 reader = rb_set_head_page(cpu_buffer);
4592 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
4593 cpu_buffer->reader_page->list.prev = reader->list.prev;
4596 * cpu_buffer->pages just needs to point to the buffer, it
4597 * has no specific buffer page to point to. Lets move it out
4598 * of our way so we don't accidentally swap it.
4600 cpu_buffer->pages = reader->list.prev;
4602 /* The reader page will be pointing to the new head */
4603 rb_set_list_to_head(&cpu_buffer->reader_page->list);
4606 * We want to make sure we read the overruns after we set up our
4607 * pointers to the next object. The writer side does a
4608 * cmpxchg to cross pages which acts as the mb on the writer
4609 * side. Note, the reader will constantly fail the swap
4610 * while the writer is updating the pointers, so this
4611 * guarantees that the overwrite recorded here is the one we
4612 * want to compare with the last_overrun.
4615 overwrite = local_read(&(cpu_buffer->overrun));
4618 * Here's the tricky part.
4620 * We need to move the pointer past the header page.
4621 * But we can only do that if a writer is not currently
4622 * moving it. The page before the header page has the
4623 * flag bit '1' set if it is pointing to the page we want.
4624 * but if the writer is in the process of moving it
4625 * than it will be '2' or already moved '0'.
4628 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
4631 * If we did not convert it, then we must try again.
4637 * Yay! We succeeded in replacing the page.
4639 * Now make the new head point back to the reader page.
4641 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
4642 rb_inc_page(&cpu_buffer->head_page);
4644 local_inc(&cpu_buffer->pages_read);
4646 /* Finally update the reader page to the new head */
4647 cpu_buffer->reader_page = reader;
4648 cpu_buffer->reader_page->read = 0;
4650 if (overwrite != cpu_buffer->last_overrun) {
4651 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
4652 cpu_buffer->last_overrun = overwrite;
4658 /* Update the read_stamp on the first event */
4659 if (reader && reader->read == 0)
4660 cpu_buffer->read_stamp = reader->page->time_stamp;
4662 arch_spin_unlock(&cpu_buffer->lock);
4663 local_irq_restore(flags);
4666 * The writer has preempt disable, wait for it. But not forever
4667 * Although, 1 second is pretty much "forever"
4669 #define USECS_WAIT 1000000
4670 for (nr_loops = 0; nr_loops < USECS_WAIT; nr_loops++) {
4671 /* If the write is past the end of page, a writer is still updating it */
4672 if (likely(!reader || rb_page_write(reader) <= BUF_PAGE_SIZE))
4677 /* Get the latest version of the reader write value */
4681 /* The writer is not moving forward? Something is wrong */
4682 if (RB_WARN_ON(cpu_buffer, nr_loops == USECS_WAIT))
4686 * Make sure we see any padding after the write update
4687 * (see rb_reset_tail())
4695 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
4697 struct ring_buffer_event *event;
4698 struct buffer_page *reader;
4701 reader = rb_get_reader_page(cpu_buffer);
4703 /* This function should not be called when buffer is empty */
4704 if (RB_WARN_ON(cpu_buffer, !reader))
4707 event = rb_reader_event(cpu_buffer);
4709 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
4712 rb_update_read_stamp(cpu_buffer, event);
4714 length = rb_event_length(event);
4715 cpu_buffer->reader_page->read += length;
4718 static void rb_advance_iter(struct ring_buffer_iter *iter)
4720 struct ring_buffer_per_cpu *cpu_buffer;
4722 cpu_buffer = iter->cpu_buffer;
4724 /* If head == next_event then we need to jump to the next event */
4725 if (iter->head == iter->next_event) {
4726 /* If the event gets overwritten again, there's nothing to do */
4727 if (rb_iter_head_event(iter) == NULL)
4731 iter->head = iter->next_event;
4734 * Check if we are at the end of the buffer.
4736 if (iter->next_event >= rb_page_size(iter->head_page)) {
4737 /* discarded commits can make the page empty */
4738 if (iter->head_page == cpu_buffer->commit_page)
4744 rb_update_iter_read_stamp(iter, iter->event);
4747 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
4749 return cpu_buffer->lost_events;
4752 static struct ring_buffer_event *
4753 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
4754 unsigned long *lost_events)
4756 struct ring_buffer_event *event;
4757 struct buffer_page *reader;
4764 * We repeat when a time extend is encountered.
4765 * Since the time extend is always attached to a data event,
4766 * we should never loop more than once.
4767 * (We never hit the following condition more than twice).
4769 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
4772 reader = rb_get_reader_page(cpu_buffer);
4776 event = rb_reader_event(cpu_buffer);
4778 switch (event->type_len) {
4779 case RINGBUF_TYPE_PADDING:
4780 if (rb_null_event(event))
4781 RB_WARN_ON(cpu_buffer, 1);
4783 * Because the writer could be discarding every
4784 * event it creates (which would probably be bad)
4785 * if we were to go back to "again" then we may never
4786 * catch up, and will trigger the warn on, or lock
4787 * the box. Return the padding, and we will release
4788 * the current locks, and try again.
4792 case RINGBUF_TYPE_TIME_EXTEND:
4793 /* Internal data, OK to advance */
4794 rb_advance_reader(cpu_buffer);
4797 case RINGBUF_TYPE_TIME_STAMP:
4799 *ts = rb_event_time_stamp(event);
4800 *ts = rb_fix_abs_ts(*ts, reader->page->time_stamp);
4801 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4802 cpu_buffer->cpu, ts);
4804 /* Internal data, OK to advance */
4805 rb_advance_reader(cpu_buffer);
4808 case RINGBUF_TYPE_DATA:
4810 *ts = cpu_buffer->read_stamp + event->time_delta;
4811 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4812 cpu_buffer->cpu, ts);
4815 *lost_events = rb_lost_events(cpu_buffer);
4819 RB_WARN_ON(cpu_buffer, 1);
4824 EXPORT_SYMBOL_GPL(ring_buffer_peek);
4826 static struct ring_buffer_event *
4827 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4829 struct trace_buffer *buffer;
4830 struct ring_buffer_per_cpu *cpu_buffer;
4831 struct ring_buffer_event *event;
4837 cpu_buffer = iter->cpu_buffer;
4838 buffer = cpu_buffer->buffer;
4841 * Check if someone performed a consuming read to
4842 * the buffer. A consuming read invalidates the iterator
4843 * and we need to reset the iterator in this case.
4845 if (unlikely(iter->cache_read != cpu_buffer->read ||
4846 iter->cache_reader_page != cpu_buffer->reader_page))
4847 rb_iter_reset(iter);
4850 if (ring_buffer_iter_empty(iter))
4854 * As the writer can mess with what the iterator is trying
4855 * to read, just give up if we fail to get an event after
4856 * three tries. The iterator is not as reliable when reading
4857 * the ring buffer with an active write as the consumer is.
4858 * Do not warn if the three failures is reached.
4863 if (rb_per_cpu_empty(cpu_buffer))
4866 if (iter->head >= rb_page_size(iter->head_page)) {
4871 event = rb_iter_head_event(iter);
4875 switch (event->type_len) {
4876 case RINGBUF_TYPE_PADDING:
4877 if (rb_null_event(event)) {
4881 rb_advance_iter(iter);
4884 case RINGBUF_TYPE_TIME_EXTEND:
4885 /* Internal data, OK to advance */
4886 rb_advance_iter(iter);
4889 case RINGBUF_TYPE_TIME_STAMP:
4891 *ts = rb_event_time_stamp(event);
4892 *ts = rb_fix_abs_ts(*ts, iter->head_page->page->time_stamp);
4893 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4894 cpu_buffer->cpu, ts);
4896 /* Internal data, OK to advance */
4897 rb_advance_iter(iter);
4900 case RINGBUF_TYPE_DATA:
4902 *ts = iter->read_stamp + event->time_delta;
4903 ring_buffer_normalize_time_stamp(buffer,
4904 cpu_buffer->cpu, ts);
4909 RB_WARN_ON(cpu_buffer, 1);
4914 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4916 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4918 if (likely(!in_nmi())) {
4919 raw_spin_lock(&cpu_buffer->reader_lock);
4924 * If an NMI die dumps out the content of the ring buffer
4925 * trylock must be used to prevent a deadlock if the NMI
4926 * preempted a task that holds the ring buffer locks. If
4927 * we get the lock then all is fine, if not, then continue
4928 * to do the read, but this can corrupt the ring buffer,
4929 * so it must be permanently disabled from future writes.
4930 * Reading from NMI is a oneshot deal.
4932 if (raw_spin_trylock(&cpu_buffer->reader_lock))
4935 /* Continue without locking, but disable the ring buffer */
4936 atomic_inc(&cpu_buffer->record_disabled);
4941 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4944 raw_spin_unlock(&cpu_buffer->reader_lock);
4949 * ring_buffer_peek - peek at the next event to be read
4950 * @buffer: The ring buffer to read
4951 * @cpu: The cpu to peak at
4952 * @ts: The timestamp counter of this event.
4953 * @lost_events: a variable to store if events were lost (may be NULL)
4955 * This will return the event that will be read next, but does
4956 * not consume the data.
4958 struct ring_buffer_event *
4959 ring_buffer_peek(struct trace_buffer *buffer, int cpu, u64 *ts,
4960 unsigned long *lost_events)
4962 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4963 struct ring_buffer_event *event;
4964 unsigned long flags;
4967 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4971 local_irq_save(flags);
4972 dolock = rb_reader_lock(cpu_buffer);
4973 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4974 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4975 rb_advance_reader(cpu_buffer);
4976 rb_reader_unlock(cpu_buffer, dolock);
4977 local_irq_restore(flags);
4979 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4985 /** ring_buffer_iter_dropped - report if there are dropped events
4986 * @iter: The ring buffer iterator
4988 * Returns true if there was dropped events since the last peek.
4990 bool ring_buffer_iter_dropped(struct ring_buffer_iter *iter)
4992 bool ret = iter->missed_events != 0;
4994 iter->missed_events = 0;
4997 EXPORT_SYMBOL_GPL(ring_buffer_iter_dropped);
5000 * ring_buffer_iter_peek - peek at the next event to be read
5001 * @iter: The ring buffer iterator
5002 * @ts: The timestamp counter of this event.
5004 * This will return the event that will be read next, but does
5005 * not increment the iterator.
5007 struct ring_buffer_event *
5008 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
5010 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5011 struct ring_buffer_event *event;
5012 unsigned long flags;
5015 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5016 event = rb_iter_peek(iter, ts);
5017 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5019 if (event && event->type_len == RINGBUF_TYPE_PADDING)
5026 * ring_buffer_consume - return an event and consume it
5027 * @buffer: The ring buffer to get the next event from
5028 * @cpu: the cpu to read the buffer from
5029 * @ts: a variable to store the timestamp (may be NULL)
5030 * @lost_events: a variable to store if events were lost (may be NULL)
5032 * Returns the next event in the ring buffer, and that event is consumed.
5033 * Meaning, that sequential reads will keep returning a different event,
5034 * and eventually empty the ring buffer if the producer is slower.
5036 struct ring_buffer_event *
5037 ring_buffer_consume(struct trace_buffer *buffer, int cpu, u64 *ts,
5038 unsigned long *lost_events)
5040 struct ring_buffer_per_cpu *cpu_buffer;
5041 struct ring_buffer_event *event = NULL;
5042 unsigned long flags;
5046 /* might be called in atomic */
5049 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5052 cpu_buffer = buffer->buffers[cpu];
5053 local_irq_save(flags);
5054 dolock = rb_reader_lock(cpu_buffer);
5056 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
5058 cpu_buffer->lost_events = 0;
5059 rb_advance_reader(cpu_buffer);
5062 rb_reader_unlock(cpu_buffer, dolock);
5063 local_irq_restore(flags);
5068 if (event && event->type_len == RINGBUF_TYPE_PADDING)
5073 EXPORT_SYMBOL_GPL(ring_buffer_consume);
5076 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
5077 * @buffer: The ring buffer to read from
5078 * @cpu: The cpu buffer to iterate over
5079 * @flags: gfp flags to use for memory allocation
5081 * This performs the initial preparations necessary to iterate
5082 * through the buffer. Memory is allocated, buffer recording
5083 * is disabled, and the iterator pointer is returned to the caller.
5085 * Disabling buffer recording prevents the reading from being
5086 * corrupted. This is not a consuming read, so a producer is not
5089 * After a sequence of ring_buffer_read_prepare calls, the user is
5090 * expected to make at least one call to ring_buffer_read_prepare_sync.
5091 * Afterwards, ring_buffer_read_start is invoked to get things going
5094 * This overall must be paired with ring_buffer_read_finish.
5096 struct ring_buffer_iter *
5097 ring_buffer_read_prepare(struct trace_buffer *buffer, int cpu, gfp_t flags)
5099 struct ring_buffer_per_cpu *cpu_buffer;
5100 struct ring_buffer_iter *iter;
5102 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5105 iter = kzalloc(sizeof(*iter), flags);
5109 iter->event = kmalloc(BUF_MAX_DATA_SIZE, flags);
5115 cpu_buffer = buffer->buffers[cpu];
5117 iter->cpu_buffer = cpu_buffer;
5119 atomic_inc(&cpu_buffer->resize_disabled);
5123 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
5126 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
5128 * All previously invoked ring_buffer_read_prepare calls to prepare
5129 * iterators will be synchronized. Afterwards, read_buffer_read_start
5130 * calls on those iterators are allowed.
5133 ring_buffer_read_prepare_sync(void)
5137 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
5140 * ring_buffer_read_start - start a non consuming read of the buffer
5141 * @iter: The iterator returned by ring_buffer_read_prepare
5143 * This finalizes the startup of an iteration through the buffer.
5144 * The iterator comes from a call to ring_buffer_read_prepare and
5145 * an intervening ring_buffer_read_prepare_sync must have been
5148 * Must be paired with ring_buffer_read_finish.
5151 ring_buffer_read_start(struct ring_buffer_iter *iter)
5153 struct ring_buffer_per_cpu *cpu_buffer;
5154 unsigned long flags;
5159 cpu_buffer = iter->cpu_buffer;
5161 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5162 arch_spin_lock(&cpu_buffer->lock);
5163 rb_iter_reset(iter);
5164 arch_spin_unlock(&cpu_buffer->lock);
5165 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5167 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
5170 * ring_buffer_read_finish - finish reading the iterator of the buffer
5171 * @iter: The iterator retrieved by ring_buffer_start
5173 * This re-enables the recording to the buffer, and frees the
5177 ring_buffer_read_finish(struct ring_buffer_iter *iter)
5179 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5180 unsigned long flags;
5183 * Ring buffer is disabled from recording, here's a good place
5184 * to check the integrity of the ring buffer.
5185 * Must prevent readers from trying to read, as the check
5186 * clears the HEAD page and readers require it.
5188 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5189 rb_check_pages(cpu_buffer);
5190 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5192 atomic_dec(&cpu_buffer->resize_disabled);
5196 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
5199 * ring_buffer_iter_advance - advance the iterator to the next location
5200 * @iter: The ring buffer iterator
5202 * Move the location of the iterator such that the next read will
5203 * be the next location of the iterator.
5205 void ring_buffer_iter_advance(struct ring_buffer_iter *iter)
5207 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5208 unsigned long flags;
5210 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5212 rb_advance_iter(iter);
5214 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5216 EXPORT_SYMBOL_GPL(ring_buffer_iter_advance);
5219 * ring_buffer_size - return the size of the ring buffer (in bytes)
5220 * @buffer: The ring buffer.
5221 * @cpu: The CPU to get ring buffer size from.
5223 unsigned long ring_buffer_size(struct trace_buffer *buffer, int cpu)
5226 * Earlier, this method returned
5227 * BUF_PAGE_SIZE * buffer->nr_pages
5228 * Since the nr_pages field is now removed, we have converted this to
5229 * return the per cpu buffer value.
5231 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5234 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
5236 EXPORT_SYMBOL_GPL(ring_buffer_size);
5239 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
5241 rb_head_page_deactivate(cpu_buffer);
5243 cpu_buffer->head_page
5244 = list_entry(cpu_buffer->pages, struct buffer_page, list);
5245 local_set(&cpu_buffer->head_page->write, 0);
5246 local_set(&cpu_buffer->head_page->entries, 0);
5247 local_set(&cpu_buffer->head_page->page->commit, 0);
5249 cpu_buffer->head_page->read = 0;
5251 cpu_buffer->tail_page = cpu_buffer->head_page;
5252 cpu_buffer->commit_page = cpu_buffer->head_page;
5254 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
5255 INIT_LIST_HEAD(&cpu_buffer->new_pages);
5256 local_set(&cpu_buffer->reader_page->write, 0);
5257 local_set(&cpu_buffer->reader_page->entries, 0);
5258 local_set(&cpu_buffer->reader_page->page->commit, 0);
5259 cpu_buffer->reader_page->read = 0;
5261 local_set(&cpu_buffer->entries_bytes, 0);
5262 local_set(&cpu_buffer->overrun, 0);
5263 local_set(&cpu_buffer->commit_overrun, 0);
5264 local_set(&cpu_buffer->dropped_events, 0);
5265 local_set(&cpu_buffer->entries, 0);
5266 local_set(&cpu_buffer->committing, 0);
5267 local_set(&cpu_buffer->commits, 0);
5268 local_set(&cpu_buffer->pages_touched, 0);
5269 local_set(&cpu_buffer->pages_lost, 0);
5270 local_set(&cpu_buffer->pages_read, 0);
5271 cpu_buffer->last_pages_touch = 0;
5272 cpu_buffer->shortest_full = 0;
5273 cpu_buffer->read = 0;
5274 cpu_buffer->read_bytes = 0;
5276 rb_time_set(&cpu_buffer->write_stamp, 0);
5277 rb_time_set(&cpu_buffer->before_stamp, 0);
5279 memset(cpu_buffer->event_stamp, 0, sizeof(cpu_buffer->event_stamp));
5281 cpu_buffer->lost_events = 0;
5282 cpu_buffer->last_overrun = 0;
5284 rb_head_page_activate(cpu_buffer);
5287 /* Must have disabled the cpu buffer then done a synchronize_rcu */
5288 static void reset_disabled_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
5290 unsigned long flags;
5292 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5294 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
5297 arch_spin_lock(&cpu_buffer->lock);
5299 rb_reset_cpu(cpu_buffer);
5301 arch_spin_unlock(&cpu_buffer->lock);
5304 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5308 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
5309 * @buffer: The ring buffer to reset a per cpu buffer of
5310 * @cpu: The CPU buffer to be reset
5312 void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
5314 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5316 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5319 /* prevent another thread from changing buffer sizes */
5320 mutex_lock(&buffer->mutex);
5322 atomic_inc(&cpu_buffer->resize_disabled);
5323 atomic_inc(&cpu_buffer->record_disabled);
5325 /* Make sure all commits have finished */
5328 reset_disabled_cpu_buffer(cpu_buffer);
5330 atomic_dec(&cpu_buffer->record_disabled);
5331 atomic_dec(&cpu_buffer->resize_disabled);
5333 mutex_unlock(&buffer->mutex);
5335 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
5338 * ring_buffer_reset_online_cpus - reset a ring buffer per CPU buffer
5339 * @buffer: The ring buffer to reset a per cpu buffer of
5340 * @cpu: The CPU buffer to be reset
5342 void ring_buffer_reset_online_cpus(struct trace_buffer *buffer)
5344 struct ring_buffer_per_cpu *cpu_buffer;
5347 /* prevent another thread from changing buffer sizes */
5348 mutex_lock(&buffer->mutex);
5350 for_each_online_buffer_cpu(buffer, cpu) {
5351 cpu_buffer = buffer->buffers[cpu];
5353 atomic_inc(&cpu_buffer->resize_disabled);
5354 atomic_inc(&cpu_buffer->record_disabled);
5357 /* Make sure all commits have finished */
5360 for_each_online_buffer_cpu(buffer, cpu) {
5361 cpu_buffer = buffer->buffers[cpu];
5363 reset_disabled_cpu_buffer(cpu_buffer);
5365 atomic_dec(&cpu_buffer->record_disabled);
5366 atomic_dec(&cpu_buffer->resize_disabled);
5369 mutex_unlock(&buffer->mutex);
5373 * ring_buffer_reset - reset a ring buffer
5374 * @buffer: The ring buffer to reset all cpu buffers
5376 void ring_buffer_reset(struct trace_buffer *buffer)
5378 struct ring_buffer_per_cpu *cpu_buffer;
5381 /* prevent another thread from changing buffer sizes */
5382 mutex_lock(&buffer->mutex);
5384 for_each_buffer_cpu(buffer, cpu) {
5385 cpu_buffer = buffer->buffers[cpu];
5387 atomic_inc(&cpu_buffer->resize_disabled);
5388 atomic_inc(&cpu_buffer->record_disabled);
5391 /* Make sure all commits have finished */
5394 for_each_buffer_cpu(buffer, cpu) {
5395 cpu_buffer = buffer->buffers[cpu];
5397 reset_disabled_cpu_buffer(cpu_buffer);
5399 atomic_dec(&cpu_buffer->record_disabled);
5400 atomic_dec(&cpu_buffer->resize_disabled);
5403 mutex_unlock(&buffer->mutex);
5405 EXPORT_SYMBOL_GPL(ring_buffer_reset);
5408 * ring_buffer_empty - is the ring buffer empty?
5409 * @buffer: The ring buffer to test
5411 bool ring_buffer_empty(struct trace_buffer *buffer)
5413 struct ring_buffer_per_cpu *cpu_buffer;
5414 unsigned long flags;
5419 /* yes this is racy, but if you don't like the race, lock the buffer */
5420 for_each_buffer_cpu(buffer, cpu) {
5421 cpu_buffer = buffer->buffers[cpu];
5422 local_irq_save(flags);
5423 dolock = rb_reader_lock(cpu_buffer);
5424 ret = rb_per_cpu_empty(cpu_buffer);
5425 rb_reader_unlock(cpu_buffer, dolock);
5426 local_irq_restore(flags);
5434 EXPORT_SYMBOL_GPL(ring_buffer_empty);
5437 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
5438 * @buffer: The ring buffer
5439 * @cpu: The CPU buffer to test
5441 bool ring_buffer_empty_cpu(struct trace_buffer *buffer, int cpu)
5443 struct ring_buffer_per_cpu *cpu_buffer;
5444 unsigned long flags;
5448 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5451 cpu_buffer = buffer->buffers[cpu];
5452 local_irq_save(flags);
5453 dolock = rb_reader_lock(cpu_buffer);
5454 ret = rb_per_cpu_empty(cpu_buffer);
5455 rb_reader_unlock(cpu_buffer, dolock);
5456 local_irq_restore(flags);
5460 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
5462 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
5464 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
5465 * @buffer_a: One buffer to swap with
5466 * @buffer_b: The other buffer to swap with
5467 * @cpu: the CPU of the buffers to swap
5469 * This function is useful for tracers that want to take a "snapshot"
5470 * of a CPU buffer and has another back up buffer lying around.
5471 * it is expected that the tracer handles the cpu buffer not being
5472 * used at the moment.
5474 int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
5475 struct trace_buffer *buffer_b, int cpu)
5477 struct ring_buffer_per_cpu *cpu_buffer_a;
5478 struct ring_buffer_per_cpu *cpu_buffer_b;
5481 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
5482 !cpumask_test_cpu(cpu, buffer_b->cpumask))
5485 cpu_buffer_a = buffer_a->buffers[cpu];
5486 cpu_buffer_b = buffer_b->buffers[cpu];
5488 /* At least make sure the two buffers are somewhat the same */
5489 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
5494 if (atomic_read(&buffer_a->record_disabled))
5497 if (atomic_read(&buffer_b->record_disabled))
5500 if (atomic_read(&cpu_buffer_a->record_disabled))
5503 if (atomic_read(&cpu_buffer_b->record_disabled))
5507 * We can't do a synchronize_rcu here because this
5508 * function can be called in atomic context.
5509 * Normally this will be called from the same CPU as cpu.
5510 * If not it's up to the caller to protect this.
5512 atomic_inc(&cpu_buffer_a->record_disabled);
5513 atomic_inc(&cpu_buffer_b->record_disabled);
5516 if (local_read(&cpu_buffer_a->committing))
5518 if (local_read(&cpu_buffer_b->committing))
5521 buffer_a->buffers[cpu] = cpu_buffer_b;
5522 buffer_b->buffers[cpu] = cpu_buffer_a;
5524 cpu_buffer_b->buffer = buffer_a;
5525 cpu_buffer_a->buffer = buffer_b;
5530 atomic_dec(&cpu_buffer_a->record_disabled);
5531 atomic_dec(&cpu_buffer_b->record_disabled);
5535 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
5536 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
5539 * ring_buffer_alloc_read_page - allocate a page to read from buffer
5540 * @buffer: the buffer to allocate for.
5541 * @cpu: the cpu buffer to allocate.
5543 * This function is used in conjunction with ring_buffer_read_page.
5544 * When reading a full page from the ring buffer, these functions
5545 * can be used to speed up the process. The calling function should
5546 * allocate a few pages first with this function. Then when it
5547 * needs to get pages from the ring buffer, it passes the result
5548 * of this function into ring_buffer_read_page, which will swap
5549 * the page that was allocated, with the read page of the buffer.
5552 * The page allocated, or ERR_PTR
5554 void *ring_buffer_alloc_read_page(struct trace_buffer *buffer, int cpu)
5556 struct ring_buffer_per_cpu *cpu_buffer;
5557 struct buffer_data_page *bpage = NULL;
5558 unsigned long flags;
5561 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5562 return ERR_PTR(-ENODEV);
5564 cpu_buffer = buffer->buffers[cpu];
5565 local_irq_save(flags);
5566 arch_spin_lock(&cpu_buffer->lock);
5568 if (cpu_buffer->free_page) {
5569 bpage = cpu_buffer->free_page;
5570 cpu_buffer->free_page = NULL;
5573 arch_spin_unlock(&cpu_buffer->lock);
5574 local_irq_restore(flags);
5579 page = alloc_pages_node(cpu_to_node(cpu),
5580 GFP_KERNEL | __GFP_NORETRY, 0);
5582 return ERR_PTR(-ENOMEM);
5584 bpage = page_address(page);
5587 rb_init_page(bpage);
5591 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
5594 * ring_buffer_free_read_page - free an allocated read page
5595 * @buffer: the buffer the page was allocate for
5596 * @cpu: the cpu buffer the page came from
5597 * @data: the page to free
5599 * Free a page allocated from ring_buffer_alloc_read_page.
5601 void ring_buffer_free_read_page(struct trace_buffer *buffer, int cpu, void *data)
5603 struct ring_buffer_per_cpu *cpu_buffer;
5604 struct buffer_data_page *bpage = data;
5605 struct page *page = virt_to_page(bpage);
5606 unsigned long flags;
5608 if (!buffer || !buffer->buffers || !buffer->buffers[cpu])
5611 cpu_buffer = buffer->buffers[cpu];
5613 /* If the page is still in use someplace else, we can't reuse it */
5614 if (page_ref_count(page) > 1)
5617 local_irq_save(flags);
5618 arch_spin_lock(&cpu_buffer->lock);
5620 if (!cpu_buffer->free_page) {
5621 cpu_buffer->free_page = bpage;
5625 arch_spin_unlock(&cpu_buffer->lock);
5626 local_irq_restore(flags);
5629 free_page((unsigned long)bpage);
5631 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
5634 * ring_buffer_read_page - extract a page from the ring buffer
5635 * @buffer: buffer to extract from
5636 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
5637 * @len: amount to extract
5638 * @cpu: the cpu of the buffer to extract
5639 * @full: should the extraction only happen when the page is full.
5641 * This function will pull out a page from the ring buffer and consume it.
5642 * @data_page must be the address of the variable that was returned
5643 * from ring_buffer_alloc_read_page. This is because the page might be used
5644 * to swap with a page in the ring buffer.
5647 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
5648 * if (IS_ERR(rpage))
5649 * return PTR_ERR(rpage);
5650 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
5652 * process_page(rpage, ret);
5654 * When @full is set, the function will not return true unless
5655 * the writer is off the reader page.
5657 * Note: it is up to the calling functions to handle sleeps and wakeups.
5658 * The ring buffer can be used anywhere in the kernel and can not
5659 * blindly call wake_up. The layer that uses the ring buffer must be
5660 * responsible for that.
5663 * >=0 if data has been transferred, returns the offset of consumed data.
5664 * <0 if no data has been transferred.
5666 int ring_buffer_read_page(struct trace_buffer *buffer,
5667 void **data_page, size_t len, int cpu, int full)
5669 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5670 struct ring_buffer_event *event;
5671 struct buffer_data_page *bpage;
5672 struct buffer_page *reader;
5673 unsigned long missed_events;
5674 unsigned long flags;
5675 unsigned int commit;
5680 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5684 * If len is not big enough to hold the page header, then
5685 * we can not copy anything.
5687 if (len <= BUF_PAGE_HDR_SIZE)
5690 len -= BUF_PAGE_HDR_SIZE;
5699 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5701 reader = rb_get_reader_page(cpu_buffer);
5705 event = rb_reader_event(cpu_buffer);
5707 read = reader->read;
5708 commit = rb_page_commit(reader);
5710 /* Check if any events were dropped */
5711 missed_events = cpu_buffer->lost_events;
5714 * If this page has been partially read or
5715 * if len is not big enough to read the rest of the page or
5716 * a writer is still on the page, then
5717 * we must copy the data from the page to the buffer.
5718 * Otherwise, we can simply swap the page with the one passed in.
5720 if (read || (len < (commit - read)) ||
5721 cpu_buffer->reader_page == cpu_buffer->commit_page) {
5722 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
5723 unsigned int rpos = read;
5724 unsigned int pos = 0;
5728 * If a full page is expected, this can still be returned
5729 * if there's been a previous partial read and the
5730 * rest of the page can be read and the commit page is off
5734 (!read || (len < (commit - read)) ||
5735 cpu_buffer->reader_page == cpu_buffer->commit_page))
5738 if (len > (commit - read))
5739 len = (commit - read);
5741 /* Always keep the time extend and data together */
5742 size = rb_event_ts_length(event);
5747 /* save the current timestamp, since the user will need it */
5748 save_timestamp = cpu_buffer->read_stamp;
5750 /* Need to copy one event at a time */
5752 /* We need the size of one event, because
5753 * rb_advance_reader only advances by one event,
5754 * whereas rb_event_ts_length may include the size of
5755 * one or two events.
5756 * We have already ensured there's enough space if this
5757 * is a time extend. */
5758 size = rb_event_length(event);
5759 memcpy(bpage->data + pos, rpage->data + rpos, size);
5763 rb_advance_reader(cpu_buffer);
5764 rpos = reader->read;
5770 event = rb_reader_event(cpu_buffer);
5771 /* Always keep the time extend and data together */
5772 size = rb_event_ts_length(event);
5773 } while (len >= size);
5776 local_set(&bpage->commit, pos);
5777 bpage->time_stamp = save_timestamp;
5779 /* we copied everything to the beginning */
5782 /* update the entry counter */
5783 cpu_buffer->read += rb_page_entries(reader);
5784 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
5786 /* swap the pages */
5787 rb_init_page(bpage);
5788 bpage = reader->page;
5789 reader->page = *data_page;
5790 local_set(&reader->write, 0);
5791 local_set(&reader->entries, 0);
5796 * Use the real_end for the data size,
5797 * This gives us a chance to store the lost events
5800 if (reader->real_end)
5801 local_set(&bpage->commit, reader->real_end);
5805 cpu_buffer->lost_events = 0;
5807 commit = local_read(&bpage->commit);
5809 * Set a flag in the commit field if we lost events
5811 if (missed_events) {
5812 /* If there is room at the end of the page to save the
5813 * missed events, then record it there.
5815 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
5816 memcpy(&bpage->data[commit], &missed_events,
5817 sizeof(missed_events));
5818 local_add(RB_MISSED_STORED, &bpage->commit);
5819 commit += sizeof(missed_events);
5821 local_add(RB_MISSED_EVENTS, &bpage->commit);
5825 * This page may be off to user land. Zero it out here.
5827 if (commit < BUF_PAGE_SIZE)
5828 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
5831 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5836 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
5839 * We only allocate new buffers, never free them if the CPU goes down.
5840 * If we were to free the buffer, then the user would lose any trace that was in
5843 int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
5845 struct trace_buffer *buffer;
5848 unsigned long nr_pages;
5850 buffer = container_of(node, struct trace_buffer, node);
5851 if (cpumask_test_cpu(cpu, buffer->cpumask))
5856 /* check if all cpu sizes are same */
5857 for_each_buffer_cpu(buffer, cpu_i) {
5858 /* fill in the size from first enabled cpu */
5860 nr_pages = buffer->buffers[cpu_i]->nr_pages;
5861 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
5866 /* allocate minimum pages, user can later expand it */
5869 buffer->buffers[cpu] =
5870 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
5871 if (!buffer->buffers[cpu]) {
5872 WARN(1, "failed to allocate ring buffer on CPU %u\n",
5877 cpumask_set_cpu(cpu, buffer->cpumask);
5881 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
5883 * This is a basic integrity check of the ring buffer.
5884 * Late in the boot cycle this test will run when configured in.
5885 * It will kick off a thread per CPU that will go into a loop
5886 * writing to the per cpu ring buffer various sizes of data.
5887 * Some of the data will be large items, some small.
5889 * Another thread is created that goes into a spin, sending out
5890 * IPIs to the other CPUs to also write into the ring buffer.
5891 * this is to test the nesting ability of the buffer.
5893 * Basic stats are recorded and reported. If something in the
5894 * ring buffer should happen that's not expected, a big warning
5895 * is displayed and all ring buffers are disabled.
5897 static struct task_struct *rb_threads[NR_CPUS] __initdata;
5899 struct rb_test_data {
5900 struct trace_buffer *buffer;
5901 unsigned long events;
5902 unsigned long bytes_written;
5903 unsigned long bytes_alloc;
5904 unsigned long bytes_dropped;
5905 unsigned long events_nested;
5906 unsigned long bytes_written_nested;
5907 unsigned long bytes_alloc_nested;
5908 unsigned long bytes_dropped_nested;
5909 int min_size_nested;
5910 int max_size_nested;
5917 static struct rb_test_data rb_data[NR_CPUS] __initdata;
5920 #define RB_TEST_BUFFER_SIZE 1048576
5922 static char rb_string[] __initdata =
5923 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
5924 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
5925 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
5927 static bool rb_test_started __initdata;
5934 static __init int rb_write_something(struct rb_test_data *data, bool nested)
5936 struct ring_buffer_event *event;
5937 struct rb_item *item;
5944 /* Have nested writes different that what is written */
5945 cnt = data->cnt + (nested ? 27 : 0);
5947 /* Multiply cnt by ~e, to make some unique increment */
5948 size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
5950 len = size + sizeof(struct rb_item);
5952 started = rb_test_started;
5953 /* read rb_test_started before checking buffer enabled */
5956 event = ring_buffer_lock_reserve(data->buffer, len);
5958 /* Ignore dropped events before test starts. */
5961 data->bytes_dropped += len;
5963 data->bytes_dropped_nested += len;
5968 event_len = ring_buffer_event_length(event);
5970 if (RB_WARN_ON(data->buffer, event_len < len))
5973 item = ring_buffer_event_data(event);
5975 memcpy(item->str, rb_string, size);
5978 data->bytes_alloc_nested += event_len;
5979 data->bytes_written_nested += len;
5980 data->events_nested++;
5981 if (!data->min_size_nested || len < data->min_size_nested)
5982 data->min_size_nested = len;
5983 if (len > data->max_size_nested)
5984 data->max_size_nested = len;
5986 data->bytes_alloc += event_len;
5987 data->bytes_written += len;
5989 if (!data->min_size || len < data->min_size)
5990 data->max_size = len;
5991 if (len > data->max_size)
5992 data->max_size = len;
5996 ring_buffer_unlock_commit(data->buffer);
6001 static __init int rb_test(void *arg)
6003 struct rb_test_data *data = arg;
6005 while (!kthread_should_stop()) {
6006 rb_write_something(data, false);
6009 set_current_state(TASK_INTERRUPTIBLE);
6010 /* Now sleep between a min of 100-300us and a max of 1ms */
6011 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
6017 static __init void rb_ipi(void *ignore)
6019 struct rb_test_data *data;
6020 int cpu = smp_processor_id();
6022 data = &rb_data[cpu];
6023 rb_write_something(data, true);
6026 static __init int rb_hammer_test(void *arg)
6028 while (!kthread_should_stop()) {
6030 /* Send an IPI to all cpus to write data! */
6031 smp_call_function(rb_ipi, NULL, 1);
6032 /* No sleep, but for non preempt, let others run */
6039 static __init int test_ringbuffer(void)
6041 struct task_struct *rb_hammer;
6042 struct trace_buffer *buffer;
6046 if (security_locked_down(LOCKDOWN_TRACEFS)) {
6047 pr_warn("Lockdown is enabled, skipping ring buffer tests\n");
6051 pr_info("Running ring buffer tests...\n");
6053 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
6054 if (WARN_ON(!buffer))
6057 /* Disable buffer so that threads can't write to it yet */
6058 ring_buffer_record_off(buffer);
6060 for_each_online_cpu(cpu) {
6061 rb_data[cpu].buffer = buffer;
6062 rb_data[cpu].cpu = cpu;
6063 rb_data[cpu].cnt = cpu;
6064 rb_threads[cpu] = kthread_run_on_cpu(rb_test, &rb_data[cpu],
6065 cpu, "rbtester/%u");
6066 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
6067 pr_cont("FAILED\n");
6068 ret = PTR_ERR(rb_threads[cpu]);
6073 /* Now create the rb hammer! */
6074 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
6075 if (WARN_ON(IS_ERR(rb_hammer))) {
6076 pr_cont("FAILED\n");
6077 ret = PTR_ERR(rb_hammer);
6081 ring_buffer_record_on(buffer);
6083 * Show buffer is enabled before setting rb_test_started.
6084 * Yes there's a small race window where events could be
6085 * dropped and the thread wont catch it. But when a ring
6086 * buffer gets enabled, there will always be some kind of
6087 * delay before other CPUs see it. Thus, we don't care about
6088 * those dropped events. We care about events dropped after
6089 * the threads see that the buffer is active.
6092 rb_test_started = true;
6094 set_current_state(TASK_INTERRUPTIBLE);
6095 /* Just run for 10 seconds */;
6096 schedule_timeout(10 * HZ);
6098 kthread_stop(rb_hammer);
6101 for_each_online_cpu(cpu) {
6102 if (!rb_threads[cpu])
6104 kthread_stop(rb_threads[cpu]);
6107 ring_buffer_free(buffer);
6112 pr_info("finished\n");
6113 for_each_online_cpu(cpu) {
6114 struct ring_buffer_event *event;
6115 struct rb_test_data *data = &rb_data[cpu];
6116 struct rb_item *item;
6117 unsigned long total_events;
6118 unsigned long total_dropped;
6119 unsigned long total_written;
6120 unsigned long total_alloc;
6121 unsigned long total_read = 0;
6122 unsigned long total_size = 0;
6123 unsigned long total_len = 0;
6124 unsigned long total_lost = 0;
6127 int small_event_size;
6131 total_events = data->events + data->events_nested;
6132 total_written = data->bytes_written + data->bytes_written_nested;
6133 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
6134 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
6136 big_event_size = data->max_size + data->max_size_nested;
6137 small_event_size = data->min_size + data->min_size_nested;
6139 pr_info("CPU %d:\n", cpu);
6140 pr_info(" events: %ld\n", total_events);
6141 pr_info(" dropped bytes: %ld\n", total_dropped);
6142 pr_info(" alloced bytes: %ld\n", total_alloc);
6143 pr_info(" written bytes: %ld\n", total_written);
6144 pr_info(" biggest event: %d\n", big_event_size);
6145 pr_info(" smallest event: %d\n", small_event_size);
6147 if (RB_WARN_ON(buffer, total_dropped))
6152 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
6154 item = ring_buffer_event_data(event);
6155 total_len += ring_buffer_event_length(event);
6156 total_size += item->size + sizeof(struct rb_item);
6157 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
6158 pr_info("FAILED!\n");
6159 pr_info("buffer had: %.*s\n", item->size, item->str);
6160 pr_info("expected: %.*s\n", item->size, rb_string);
6161 RB_WARN_ON(buffer, 1);
6172 pr_info(" read events: %ld\n", total_read);
6173 pr_info(" lost events: %ld\n", total_lost);
6174 pr_info(" total events: %ld\n", total_lost + total_read);
6175 pr_info(" recorded len bytes: %ld\n", total_len);
6176 pr_info(" recorded size bytes: %ld\n", total_size);
6178 pr_info(" With dropped events, record len and size may not match\n"
6179 " alloced and written from above\n");
6181 if (RB_WARN_ON(buffer, total_len != total_alloc ||
6182 total_size != total_written))
6185 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
6191 pr_info("Ring buffer PASSED!\n");
6193 ring_buffer_free(buffer);
6197 late_initcall(test_ringbuffer);
6198 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */