1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
7 #include <linux/trace_recursion.h>
8 #include <linux/trace_events.h>
9 #include <linux/ring_buffer.h>
10 #include <linux/trace_clock.h>
11 #include <linux/sched/clock.h>
12 #include <linux/trace_seq.h>
13 #include <linux/spinlock.h>
14 #include <linux/irq_work.h>
15 #include <linux/security.h>
16 #include <linux/uaccess.h>
17 #include <linux/hardirq.h>
18 #include <linux/kthread.h> /* for self test */
19 #include <linux/module.h>
20 #include <linux/percpu.h>
21 #include <linux/mutex.h>
22 #include <linux/delay.h>
23 #include <linux/slab.h>
24 #include <linux/init.h>
25 #include <linux/hash.h>
26 #include <linux/list.h>
27 #include <linux/cpu.h>
28 #include <linux/oom.h>
30 #include <asm/local.h>
33 * The "absolute" timestamp in the buffer is only 59 bits.
34 * If a clock has the 5 MSBs set, it needs to be saved and
37 #define TS_MSB (0xf8ULL << 56)
38 #define ABS_TS_MASK (~TS_MSB)
40 static void update_pages_handler(struct work_struct *work);
43 * The ring buffer header is special. We must manually up keep it.
45 int ring_buffer_print_entry_header(struct trace_seq *s)
47 trace_seq_puts(s, "# compressed entry header\n");
48 trace_seq_puts(s, "\ttype_len : 5 bits\n");
49 trace_seq_puts(s, "\ttime_delta : 27 bits\n");
50 trace_seq_puts(s, "\tarray : 32 bits\n");
51 trace_seq_putc(s, '\n');
52 trace_seq_printf(s, "\tpadding : type == %d\n",
53 RINGBUF_TYPE_PADDING);
54 trace_seq_printf(s, "\ttime_extend : type == %d\n",
55 RINGBUF_TYPE_TIME_EXTEND);
56 trace_seq_printf(s, "\ttime_stamp : type == %d\n",
57 RINGBUF_TYPE_TIME_STAMP);
58 trace_seq_printf(s, "\tdata max type_len == %d\n",
59 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
61 return !trace_seq_has_overflowed(s);
65 * The ring buffer is made up of a list of pages. A separate list of pages is
66 * allocated for each CPU. A writer may only write to a buffer that is
67 * associated with the CPU it is currently executing on. A reader may read
68 * from any per cpu buffer.
70 * The reader is special. For each per cpu buffer, the reader has its own
71 * reader page. When a reader has read the entire reader page, this reader
72 * page is swapped with another page in the ring buffer.
74 * Now, as long as the writer is off the reader page, the reader can do what
75 * ever it wants with that page. The writer will never write to that page
76 * again (as long as it is out of the ring buffer).
78 * Here's some silly ASCII art.
81 * |reader| RING BUFFER
83 * +------+ +---+ +---+ +---+
92 * |reader| RING BUFFER
93 * |page |------------------v
94 * +------+ +---+ +---+ +---+
103 * |reader| RING BUFFER
104 * |page |------------------v
105 * +------+ +---+ +---+ +---+
107 * | +---+ +---+ +---+
110 * +------------------------------+
114 * |buffer| RING BUFFER
115 * |page |------------------v
116 * +------+ +---+ +---+ +---+
118 * | New +---+ +---+ +---+
121 * +------------------------------+
124 * After we make this swap, the reader can hand this page off to the splice
125 * code and be done with it. It can even allocate a new page if it needs to
126 * and swap that into the ring buffer.
128 * We will be using cmpxchg soon to make all this lockless.
132 /* Used for individual buffers (after the counter) */
133 #define RB_BUFFER_OFF (1 << 20)
135 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
137 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
138 #define RB_ALIGNMENT 4U
139 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
140 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
142 #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
143 # define RB_FORCE_8BYTE_ALIGNMENT 0
144 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
146 # define RB_FORCE_8BYTE_ALIGNMENT 1
147 # define RB_ARCH_ALIGNMENT 8U
150 #define RB_ALIGN_DATA __aligned(RB_ARCH_ALIGNMENT)
152 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
153 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
156 RB_LEN_TIME_EXTEND = 8,
157 RB_LEN_TIME_STAMP = 8,
160 #define skip_time_extend(event) \
161 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
163 #define extended_time(event) \
164 (event->type_len >= RINGBUF_TYPE_TIME_EXTEND)
166 static inline bool rb_null_event(struct ring_buffer_event *event)
168 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
171 static void rb_event_set_padding(struct ring_buffer_event *event)
173 /* padding has a NULL time_delta */
174 event->type_len = RINGBUF_TYPE_PADDING;
175 event->time_delta = 0;
179 rb_event_data_length(struct ring_buffer_event *event)
184 length = event->type_len * RB_ALIGNMENT;
186 length = event->array[0];
187 return length + RB_EVNT_HDR_SIZE;
191 * Return the length of the given event. Will return
192 * the length of the time extend if the event is a
195 static inline unsigned
196 rb_event_length(struct ring_buffer_event *event)
198 switch (event->type_len) {
199 case RINGBUF_TYPE_PADDING:
200 if (rb_null_event(event))
203 return event->array[0] + RB_EVNT_HDR_SIZE;
205 case RINGBUF_TYPE_TIME_EXTEND:
206 return RB_LEN_TIME_EXTEND;
208 case RINGBUF_TYPE_TIME_STAMP:
209 return RB_LEN_TIME_STAMP;
211 case RINGBUF_TYPE_DATA:
212 return rb_event_data_length(event);
221 * Return total length of time extend and data,
222 * or just the event length for all other events.
224 static inline unsigned
225 rb_event_ts_length(struct ring_buffer_event *event)
229 if (extended_time(event)) {
230 /* time extends include the data event after it */
231 len = RB_LEN_TIME_EXTEND;
232 event = skip_time_extend(event);
234 return len + rb_event_length(event);
238 * ring_buffer_event_length - return the length of the event
239 * @event: the event to get the length of
241 * Returns the size of the data load of a data event.
242 * If the event is something other than a data event, it
243 * returns the size of the event itself. With the exception
244 * of a TIME EXTEND, where it still returns the size of the
245 * data load of the data event after it.
247 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
251 if (extended_time(event))
252 event = skip_time_extend(event);
254 length = rb_event_length(event);
255 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
257 length -= RB_EVNT_HDR_SIZE;
258 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
259 length -= sizeof(event->array[0]);
262 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
264 /* inline for ring buffer fast paths */
265 static __always_inline void *
266 rb_event_data(struct ring_buffer_event *event)
268 if (extended_time(event))
269 event = skip_time_extend(event);
270 WARN_ON_ONCE(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
271 /* If length is in len field, then array[0] has the data */
273 return (void *)&event->array[0];
274 /* Otherwise length is in array[0] and array[1] has the data */
275 return (void *)&event->array[1];
279 * ring_buffer_event_data - return the data of the event
280 * @event: the event to get the data from
282 void *ring_buffer_event_data(struct ring_buffer_event *event)
284 return rb_event_data(event);
286 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
288 #define for_each_buffer_cpu(buffer, cpu) \
289 for_each_cpu(cpu, buffer->cpumask)
291 #define for_each_online_buffer_cpu(buffer, cpu) \
292 for_each_cpu_and(cpu, buffer->cpumask, cpu_online_mask)
295 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
296 #define TS_DELTA_TEST (~TS_MASK)
298 static u64 rb_event_time_stamp(struct ring_buffer_event *event)
302 ts = event->array[0];
304 ts += event->time_delta;
309 /* Flag when events were overwritten */
310 #define RB_MISSED_EVENTS (1 << 31)
311 /* Missed count stored at end */
312 #define RB_MISSED_STORED (1 << 30)
314 struct buffer_data_page {
315 u64 time_stamp; /* page time stamp */
316 local_t commit; /* write committed index */
317 unsigned char data[] RB_ALIGN_DATA; /* data of buffer page */
321 * Note, the buffer_page list must be first. The buffer pages
322 * are allocated in cache lines, which means that each buffer
323 * page will be at the beginning of a cache line, and thus
324 * the least significant bits will be zero. We use this to
325 * add flags in the list struct pointers, to make the ring buffer
329 struct list_head list; /* list of buffer pages */
330 local_t write; /* index for next write */
331 unsigned read; /* index for next read */
332 local_t entries; /* entries on this page */
333 unsigned long real_end; /* real end of data */
334 struct buffer_data_page *page; /* Actual data page */
338 * The buffer page counters, write and entries, must be reset
339 * atomically when crossing page boundaries. To synchronize this
340 * update, two counters are inserted into the number. One is
341 * the actual counter for the write position or count on the page.
343 * The other is a counter of updaters. Before an update happens
344 * the update partition of the counter is incremented. This will
345 * allow the updater to update the counter atomically.
347 * The counter is 20 bits, and the state data is 12.
349 #define RB_WRITE_MASK 0xfffff
350 #define RB_WRITE_INTCNT (1 << 20)
352 static void rb_init_page(struct buffer_data_page *bpage)
354 local_set(&bpage->commit, 0);
357 static __always_inline unsigned int rb_page_commit(struct buffer_page *bpage)
359 return local_read(&bpage->page->commit);
362 static void free_buffer_page(struct buffer_page *bpage)
364 free_page((unsigned long)bpage->page);
369 * We need to fit the time_stamp delta into 27 bits.
371 static inline bool test_time_stamp(u64 delta)
373 return !!(delta & TS_DELTA_TEST);
376 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
378 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
379 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
381 int ring_buffer_print_page_header(struct trace_seq *s)
383 struct buffer_data_page field;
385 trace_seq_printf(s, "\tfield: u64 timestamp;\t"
386 "offset:0;\tsize:%u;\tsigned:%u;\n",
387 (unsigned int)sizeof(field.time_stamp),
388 (unsigned int)is_signed_type(u64));
390 trace_seq_printf(s, "\tfield: local_t commit;\t"
391 "offset:%u;\tsize:%u;\tsigned:%u;\n",
392 (unsigned int)offsetof(typeof(field), commit),
393 (unsigned int)sizeof(field.commit),
394 (unsigned int)is_signed_type(long));
396 trace_seq_printf(s, "\tfield: int overwrite;\t"
397 "offset:%u;\tsize:%u;\tsigned:%u;\n",
398 (unsigned int)offsetof(typeof(field), commit),
400 (unsigned int)is_signed_type(long));
402 trace_seq_printf(s, "\tfield: char data;\t"
403 "offset:%u;\tsize:%u;\tsigned:%u;\n",
404 (unsigned int)offsetof(typeof(field), data),
405 (unsigned int)BUF_PAGE_SIZE,
406 (unsigned int)is_signed_type(char));
408 return !trace_seq_has_overflowed(s);
412 struct irq_work work;
413 wait_queue_head_t waiters;
414 wait_queue_head_t full_waiters;
416 bool waiters_pending;
417 bool full_waiters_pending;
422 * Structure to hold event state and handle nested events.
424 struct rb_event_info {
429 unsigned long length;
430 struct buffer_page *tail_page;
435 * Used for the add_timestamp
437 * EXTEND - wants a time extend
438 * ABSOLUTE - the buffer requests all events to have absolute time stamps
439 * FORCE - force a full time stamp.
442 RB_ADD_STAMP_NONE = 0,
443 RB_ADD_STAMP_EXTEND = BIT(1),
444 RB_ADD_STAMP_ABSOLUTE = BIT(2),
445 RB_ADD_STAMP_FORCE = BIT(3)
448 * Used for which event context the event is in.
455 * See trace_recursive_lock() comment below for more details.
466 #if BITS_PER_LONG == 32
470 /* To test on 64 bit machines */
475 struct rb_time_struct {
482 #include <asm/local64.h>
483 struct rb_time_struct {
487 typedef struct rb_time_struct rb_time_t;
492 * head_page == tail_page && head == tail then buffer is empty.
494 struct ring_buffer_per_cpu {
496 atomic_t record_disabled;
497 atomic_t resize_disabled;
498 struct trace_buffer *buffer;
499 raw_spinlock_t reader_lock; /* serialize readers */
500 arch_spinlock_t lock;
501 struct lock_class_key lock_key;
502 struct buffer_data_page *free_page;
503 unsigned long nr_pages;
504 unsigned int current_context;
505 struct list_head *pages;
506 struct buffer_page *head_page; /* read from head */
507 struct buffer_page *tail_page; /* write to tail */
508 struct buffer_page *commit_page; /* committed pages */
509 struct buffer_page *reader_page;
510 unsigned long lost_events;
511 unsigned long last_overrun;
513 local_t entries_bytes;
516 local_t commit_overrun;
517 local_t dropped_events;
520 local_t pages_touched;
523 long last_pages_touch;
524 size_t shortest_full;
526 unsigned long read_bytes;
527 rb_time_t write_stamp;
528 rb_time_t before_stamp;
529 u64 event_stamp[MAX_NEST];
531 /* pages removed since last reset */
532 unsigned long pages_removed;
533 /* ring buffer pages to update, > 0 to add, < 0 to remove */
534 long nr_pages_to_update;
535 struct list_head new_pages; /* new pages to add */
536 struct work_struct update_pages_work;
537 struct completion update_done;
539 struct rb_irq_work irq_work;
542 struct trace_buffer {
545 atomic_t record_disabled;
547 cpumask_var_t cpumask;
549 struct lock_class_key *reader_lock_key;
553 struct ring_buffer_per_cpu **buffers;
555 struct hlist_node node;
558 struct rb_irq_work irq_work;
562 struct ring_buffer_iter {
563 struct ring_buffer_per_cpu *cpu_buffer;
565 unsigned long next_event;
566 struct buffer_page *head_page;
567 struct buffer_page *cache_reader_page;
568 unsigned long cache_read;
569 unsigned long cache_pages_removed;
572 struct ring_buffer_event *event;
579 * On 32 bit machines, local64_t is very expensive. As the ring
580 * buffer doesn't need all the features of a true 64 bit atomic,
581 * on 32 bit, it uses these functions (64 still uses local64_t).
583 * For the ring buffer, 64 bit required operations for the time is
586 * - Reads may fail if it interrupted a modification of the time stamp.
587 * It will succeed if it did not interrupt another write even if
588 * the read itself is interrupted by a write.
589 * It returns whether it was successful or not.
591 * - Writes always succeed and will overwrite other writes and writes
592 * that were done by events interrupting the current write.
594 * - A write followed by a read of the same time stamp will always succeed,
595 * but may not contain the same value.
597 * - A cmpxchg will fail if it interrupted another write or cmpxchg.
598 * Other than that, it acts like a normal cmpxchg.
600 * The 60 bit time stamp is broken up by 30 bits in a top and bottom half
601 * (bottom being the least significant 30 bits of the 60 bit time stamp).
603 * The two most significant bits of each half holds a 2 bit counter (0-3).
604 * Each update will increment this counter by one.
605 * When reading the top and bottom, if the two counter bits match then the
606 * top and bottom together make a valid 60 bit number.
608 #define RB_TIME_SHIFT 30
609 #define RB_TIME_VAL_MASK ((1 << RB_TIME_SHIFT) - 1)
610 #define RB_TIME_MSB_SHIFT 60
612 static inline int rb_time_cnt(unsigned long val)
614 return (val >> RB_TIME_SHIFT) & 3;
617 static inline u64 rb_time_val(unsigned long top, unsigned long bottom)
621 val = top & RB_TIME_VAL_MASK;
622 val <<= RB_TIME_SHIFT;
623 val |= bottom & RB_TIME_VAL_MASK;
628 static inline bool __rb_time_read(rb_time_t *t, u64 *ret, unsigned long *cnt)
630 unsigned long top, bottom, msb;
634 * If the read is interrupted by a write, then the cnt will
635 * be different. Loop until both top and bottom have been read
636 * without interruption.
639 c = local_read(&t->cnt);
640 top = local_read(&t->top);
641 bottom = local_read(&t->bottom);
642 msb = local_read(&t->msb);
643 } while (c != local_read(&t->cnt));
645 *cnt = rb_time_cnt(top);
647 /* If top, msb or bottom counts don't match, this interrupted a write */
648 if (*cnt != rb_time_cnt(msb) || *cnt != rb_time_cnt(bottom))
651 /* The shift to msb will lose its cnt bits */
652 *ret = rb_time_val(top, bottom) | ((u64)msb << RB_TIME_MSB_SHIFT);
656 static bool rb_time_read(rb_time_t *t, u64 *ret)
660 return __rb_time_read(t, ret, &cnt);
663 static inline unsigned long rb_time_val_cnt(unsigned long val, unsigned long cnt)
665 return (val & RB_TIME_VAL_MASK) | ((cnt & 3) << RB_TIME_SHIFT);
668 static inline void rb_time_split(u64 val, unsigned long *top, unsigned long *bottom,
671 *top = (unsigned long)((val >> RB_TIME_SHIFT) & RB_TIME_VAL_MASK);
672 *bottom = (unsigned long)(val & RB_TIME_VAL_MASK);
673 *msb = (unsigned long)(val >> RB_TIME_MSB_SHIFT);
676 static inline void rb_time_val_set(local_t *t, unsigned long val, unsigned long cnt)
678 val = rb_time_val_cnt(val, cnt);
682 static void rb_time_set(rb_time_t *t, u64 val)
684 unsigned long cnt, top, bottom, msb;
686 rb_time_split(val, &top, &bottom, &msb);
688 /* Writes always succeed with a valid number even if it gets interrupted. */
690 cnt = local_inc_return(&t->cnt);
691 rb_time_val_set(&t->top, top, cnt);
692 rb_time_val_set(&t->bottom, bottom, cnt);
693 rb_time_val_set(&t->msb, val >> RB_TIME_MSB_SHIFT, cnt);
694 } while (cnt != local_read(&t->cnt));
698 rb_time_read_cmpxchg(local_t *l, unsigned long expect, unsigned long set)
700 return local_try_cmpxchg(l, &expect, set);
705 /* local64_t always succeeds */
707 static inline bool rb_time_read(rb_time_t *t, u64 *ret)
709 *ret = local64_read(&t->time);
712 static void rb_time_set(rb_time_t *t, u64 val)
714 local64_set(&t->time, val);
719 * Enable this to make sure that the event passed to
720 * ring_buffer_event_time_stamp() is not committed and also
721 * is on the buffer that it passed in.
723 //#define RB_VERIFY_EVENT
724 #ifdef RB_VERIFY_EVENT
725 static struct list_head *rb_list_head(struct list_head *list);
726 static void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
729 struct buffer_page *page = cpu_buffer->commit_page;
730 struct buffer_page *tail_page = READ_ONCE(cpu_buffer->tail_page);
731 struct list_head *next;
733 unsigned long addr = (unsigned long)event;
737 /* Make sure the event exists and is not committed yet */
739 if (page == tail_page || WARN_ON_ONCE(stop++ > 100))
741 commit = local_read(&page->page->commit);
742 write = local_read(&page->write);
743 if (addr >= (unsigned long)&page->page->data[commit] &&
744 addr < (unsigned long)&page->page->data[write])
747 next = rb_list_head(page->list.next);
748 page = list_entry(next, struct buffer_page, list);
753 static inline void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
760 * The absolute time stamp drops the 5 MSBs and some clocks may
761 * require them. The rb_fix_abs_ts() will take a previous full
762 * time stamp, and add the 5 MSB of that time stamp on to the
763 * saved absolute time stamp. Then they are compared in case of
764 * the unlikely event that the latest time stamp incremented
767 static inline u64 rb_fix_abs_ts(u64 abs, u64 save_ts)
769 if (save_ts & TS_MSB) {
770 abs |= save_ts & TS_MSB;
771 /* Check for overflow */
772 if (unlikely(abs < save_ts))
778 static inline u64 rb_time_stamp(struct trace_buffer *buffer);
781 * ring_buffer_event_time_stamp - return the event's current time stamp
782 * @buffer: The buffer that the event is on
783 * @event: the event to get the time stamp of
785 * Note, this must be called after @event is reserved, and before it is
786 * committed to the ring buffer. And must be called from the same
787 * context where the event was reserved (normal, softirq, irq, etc).
789 * Returns the time stamp associated with the current event.
790 * If the event has an extended time stamp, then that is used as
791 * the time stamp to return.
792 * In the highly unlikely case that the event was nested more than
793 * the max nesting, then the write_stamp of the buffer is returned,
794 * otherwise current time is returned, but that really neither of
795 * the last two cases should ever happen.
797 u64 ring_buffer_event_time_stamp(struct trace_buffer *buffer,
798 struct ring_buffer_event *event)
800 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[smp_processor_id()];
804 /* If the event includes an absolute time, then just use that */
805 if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
806 ts = rb_event_time_stamp(event);
807 return rb_fix_abs_ts(ts, cpu_buffer->tail_page->page->time_stamp);
810 nest = local_read(&cpu_buffer->committing);
811 verify_event(cpu_buffer, event);
812 if (WARN_ON_ONCE(!nest))
815 /* Read the current saved nesting level time stamp */
816 if (likely(--nest < MAX_NEST))
817 return cpu_buffer->event_stamp[nest];
819 /* Shouldn't happen, warn if it does */
820 WARN_ONCE(1, "nest (%d) greater than max", nest);
823 /* Can only fail on 32 bit */
824 if (!rb_time_read(&cpu_buffer->write_stamp, &ts))
825 /* Screw it, just read the current time */
826 ts = rb_time_stamp(cpu_buffer->buffer);
832 * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
833 * @buffer: The ring_buffer to get the number of pages from
834 * @cpu: The cpu of the ring_buffer to get the number of pages from
836 * Returns the number of pages used by a per_cpu buffer of the ring buffer.
838 size_t ring_buffer_nr_pages(struct trace_buffer *buffer, int cpu)
840 return buffer->buffers[cpu]->nr_pages;
844 * ring_buffer_nr_dirty_pages - get the number of used pages in the ring buffer
845 * @buffer: The ring_buffer to get the number of pages from
846 * @cpu: The cpu of the ring_buffer to get the number of pages from
848 * Returns the number of pages that have content in the ring buffer.
850 size_t ring_buffer_nr_dirty_pages(struct trace_buffer *buffer, int cpu)
856 read = local_read(&buffer->buffers[cpu]->pages_read);
857 lost = local_read(&buffer->buffers[cpu]->pages_lost);
858 cnt = local_read(&buffer->buffers[cpu]->pages_touched);
860 if (WARN_ON_ONCE(cnt < lost))
865 /* The reader can read an empty page, but not more than that */
867 WARN_ON_ONCE(read > cnt + 1);
874 static __always_inline bool full_hit(struct trace_buffer *buffer, int cpu, int full)
876 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
880 nr_pages = cpu_buffer->nr_pages;
881 if (!nr_pages || !full)
884 dirty = ring_buffer_nr_dirty_pages(buffer, cpu);
886 return (dirty * 100) > (full * nr_pages);
890 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
892 * Schedules a delayed work to wake up any task that is blocked on the
893 * ring buffer waiters queue.
895 static void rb_wake_up_waiters(struct irq_work *work)
897 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
899 wake_up_all(&rbwork->waiters);
900 if (rbwork->full_waiters_pending || rbwork->wakeup_full) {
901 rbwork->wakeup_full = false;
902 rbwork->full_waiters_pending = false;
903 wake_up_all(&rbwork->full_waiters);
908 * ring_buffer_wake_waiters - wake up any waiters on this ring buffer
909 * @buffer: The ring buffer to wake waiters on
910 * @cpu: The CPU buffer to wake waiters on
912 * In the case of a file that represents a ring buffer is closing,
913 * it is prudent to wake up any waiters that are on this.
915 void ring_buffer_wake_waiters(struct trace_buffer *buffer, int cpu)
917 struct ring_buffer_per_cpu *cpu_buffer;
918 struct rb_irq_work *rbwork;
923 if (cpu == RING_BUFFER_ALL_CPUS) {
925 /* Wake up individual ones too. One level recursion */
926 for_each_buffer_cpu(buffer, cpu)
927 ring_buffer_wake_waiters(buffer, cpu);
929 rbwork = &buffer->irq_work;
931 if (WARN_ON_ONCE(!buffer->buffers))
933 if (WARN_ON_ONCE(cpu >= nr_cpu_ids))
936 cpu_buffer = buffer->buffers[cpu];
937 /* The CPU buffer may not have been initialized yet */
940 rbwork = &cpu_buffer->irq_work;
943 rbwork->wait_index++;
944 /* make sure the waiters see the new index */
947 rb_wake_up_waiters(&rbwork->work);
951 * ring_buffer_wait - wait for input to the ring buffer
952 * @buffer: buffer to wait on
953 * @cpu: the cpu buffer to wait on
954 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
956 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
957 * as data is added to any of the @buffer's cpu buffers. Otherwise
958 * it will wait for data to be added to a specific cpu buffer.
960 int ring_buffer_wait(struct trace_buffer *buffer, int cpu, int full)
962 struct ring_buffer_per_cpu *cpu_buffer;
964 struct rb_irq_work *work;
969 * Depending on what the caller is waiting for, either any
970 * data in any cpu buffer, or a specific buffer, put the
971 * caller on the appropriate wait queue.
973 if (cpu == RING_BUFFER_ALL_CPUS) {
974 work = &buffer->irq_work;
975 /* Full only makes sense on per cpu reads */
978 if (!cpumask_test_cpu(cpu, buffer->cpumask))
980 cpu_buffer = buffer->buffers[cpu];
981 work = &cpu_buffer->irq_work;
984 wait_index = READ_ONCE(work->wait_index);
988 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
990 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
993 * The events can happen in critical sections where
994 * checking a work queue can cause deadlocks.
995 * After adding a task to the queue, this flag is set
996 * only to notify events to try to wake up the queue
999 * We don't clear it even if the buffer is no longer
1000 * empty. The flag only causes the next event to run
1001 * irq_work to do the work queue wake up. The worse
1002 * that can happen if we race with !trace_empty() is that
1003 * an event will cause an irq_work to try to wake up
1006 * There's no reason to protect this flag either, as
1007 * the work queue and irq_work logic will do the necessary
1008 * synchronization for the wake ups. The only thing
1009 * that is necessary is that the wake up happens after
1010 * a task has been queued. It's OK for spurious wake ups.
1013 work->full_waiters_pending = true;
1015 work->waiters_pending = true;
1017 if (signal_pending(current)) {
1022 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
1025 if (cpu != RING_BUFFER_ALL_CPUS &&
1026 !ring_buffer_empty_cpu(buffer, cpu)) {
1027 unsigned long flags;
1034 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
1035 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
1036 done = !pagebusy && full_hit(buffer, cpu, full);
1038 if (!cpu_buffer->shortest_full ||
1039 cpu_buffer->shortest_full > full)
1040 cpu_buffer->shortest_full = full;
1041 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
1048 /* Make sure to see the new wait index */
1050 if (wait_index != work->wait_index)
1055 finish_wait(&work->full_waiters, &wait);
1057 finish_wait(&work->waiters, &wait);
1063 * ring_buffer_poll_wait - poll on buffer input
1064 * @buffer: buffer to wait on
1065 * @cpu: the cpu buffer to wait on
1066 * @filp: the file descriptor
1067 * @poll_table: The poll descriptor
1068 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
1070 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
1071 * as data is added to any of the @buffer's cpu buffers. Otherwise
1072 * it will wait for data to be added to a specific cpu buffer.
1074 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
1077 __poll_t ring_buffer_poll_wait(struct trace_buffer *buffer, int cpu,
1078 struct file *filp, poll_table *poll_table, int full)
1080 struct ring_buffer_per_cpu *cpu_buffer;
1081 struct rb_irq_work *work;
1083 if (cpu == RING_BUFFER_ALL_CPUS) {
1084 work = &buffer->irq_work;
1087 if (!cpumask_test_cpu(cpu, buffer->cpumask))
1090 cpu_buffer = buffer->buffers[cpu];
1091 work = &cpu_buffer->irq_work;
1095 poll_wait(filp, &work->full_waiters, poll_table);
1096 work->full_waiters_pending = true;
1097 if (!cpu_buffer->shortest_full ||
1098 cpu_buffer->shortest_full > full)
1099 cpu_buffer->shortest_full = full;
1101 poll_wait(filp, &work->waiters, poll_table);
1102 work->waiters_pending = true;
1106 * There's a tight race between setting the waiters_pending and
1107 * checking if the ring buffer is empty. Once the waiters_pending bit
1108 * is set, the next event will wake the task up, but we can get stuck
1109 * if there's only a single event in.
1111 * FIXME: Ideally, we need a memory barrier on the writer side as well,
1112 * but adding a memory barrier to all events will cause too much of a
1113 * performance hit in the fast path. We only need a memory barrier when
1114 * the buffer goes from empty to having content. But as this race is
1115 * extremely small, and it's not a problem if another event comes in, we
1116 * will fix it later.
1121 return full_hit(buffer, cpu, full) ? EPOLLIN | EPOLLRDNORM : 0;
1123 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
1124 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
1125 return EPOLLIN | EPOLLRDNORM;
1129 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
1130 #define RB_WARN_ON(b, cond) \
1132 int _____ret = unlikely(cond); \
1134 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
1135 struct ring_buffer_per_cpu *__b = \
1137 atomic_inc(&__b->buffer->record_disabled); \
1139 atomic_inc(&b->record_disabled); \
1145 /* Up this if you want to test the TIME_EXTENTS and normalization */
1146 #define DEBUG_SHIFT 0
1148 static inline u64 rb_time_stamp(struct trace_buffer *buffer)
1152 /* Skip retpolines :-( */
1153 if (IS_ENABLED(CONFIG_RETPOLINE) && likely(buffer->clock == trace_clock_local))
1154 ts = trace_clock_local();
1156 ts = buffer->clock();
1158 /* shift to debug/test normalization and TIME_EXTENTS */
1159 return ts << DEBUG_SHIFT;
1162 u64 ring_buffer_time_stamp(struct trace_buffer *buffer)
1166 preempt_disable_notrace();
1167 time = rb_time_stamp(buffer);
1168 preempt_enable_notrace();
1172 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
1174 void ring_buffer_normalize_time_stamp(struct trace_buffer *buffer,
1177 /* Just stupid testing the normalize function and deltas */
1178 *ts >>= DEBUG_SHIFT;
1180 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
1183 * Making the ring buffer lockless makes things tricky.
1184 * Although writes only happen on the CPU that they are on,
1185 * and they only need to worry about interrupts. Reads can
1186 * happen on any CPU.
1188 * The reader page is always off the ring buffer, but when the
1189 * reader finishes with a page, it needs to swap its page with
1190 * a new one from the buffer. The reader needs to take from
1191 * the head (writes go to the tail). But if a writer is in overwrite
1192 * mode and wraps, it must push the head page forward.
1194 * Here lies the problem.
1196 * The reader must be careful to replace only the head page, and
1197 * not another one. As described at the top of the file in the
1198 * ASCII art, the reader sets its old page to point to the next
1199 * page after head. It then sets the page after head to point to
1200 * the old reader page. But if the writer moves the head page
1201 * during this operation, the reader could end up with the tail.
1203 * We use cmpxchg to help prevent this race. We also do something
1204 * special with the page before head. We set the LSB to 1.
1206 * When the writer must push the page forward, it will clear the
1207 * bit that points to the head page, move the head, and then set
1208 * the bit that points to the new head page.
1210 * We also don't want an interrupt coming in and moving the head
1211 * page on another writer. Thus we use the second LSB to catch
1214 * head->list->prev->next bit 1 bit 0
1217 * Points to head page 0 1
1220 * Note we can not trust the prev pointer of the head page, because:
1222 * +----+ +-----+ +-----+
1223 * | |------>| T |---X--->| N |
1225 * +----+ +-----+ +-----+
1228 * +----------| R |----------+ |
1232 * Key: ---X--> HEAD flag set in pointer
1237 * (see __rb_reserve_next() to see where this happens)
1239 * What the above shows is that the reader just swapped out
1240 * the reader page with a page in the buffer, but before it
1241 * could make the new header point back to the new page added
1242 * it was preempted by a writer. The writer moved forward onto
1243 * the new page added by the reader and is about to move forward
1246 * You can see, it is legitimate for the previous pointer of
1247 * the head (or any page) not to point back to itself. But only
1251 #define RB_PAGE_NORMAL 0UL
1252 #define RB_PAGE_HEAD 1UL
1253 #define RB_PAGE_UPDATE 2UL
1256 #define RB_FLAG_MASK 3UL
1258 /* PAGE_MOVED is not part of the mask */
1259 #define RB_PAGE_MOVED 4UL
1262 * rb_list_head - remove any bit
1264 static struct list_head *rb_list_head(struct list_head *list)
1266 unsigned long val = (unsigned long)list;
1268 return (struct list_head *)(val & ~RB_FLAG_MASK);
1272 * rb_is_head_page - test if the given page is the head page
1274 * Because the reader may move the head_page pointer, we can
1275 * not trust what the head page is (it may be pointing to
1276 * the reader page). But if the next page is a header page,
1277 * its flags will be non zero.
1280 rb_is_head_page(struct buffer_page *page, struct list_head *list)
1284 val = (unsigned long)list->next;
1286 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
1287 return RB_PAGE_MOVED;
1289 return val & RB_FLAG_MASK;
1295 * The unique thing about the reader page, is that, if the
1296 * writer is ever on it, the previous pointer never points
1297 * back to the reader page.
1299 static bool rb_is_reader_page(struct buffer_page *page)
1301 struct list_head *list = page->list.prev;
1303 return rb_list_head(list->next) != &page->list;
1307 * rb_set_list_to_head - set a list_head to be pointing to head.
1309 static void rb_set_list_to_head(struct list_head *list)
1313 ptr = (unsigned long *)&list->next;
1314 *ptr |= RB_PAGE_HEAD;
1315 *ptr &= ~RB_PAGE_UPDATE;
1319 * rb_head_page_activate - sets up head page
1321 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
1323 struct buffer_page *head;
1325 head = cpu_buffer->head_page;
1330 * Set the previous list pointer to have the HEAD flag.
1332 rb_set_list_to_head(head->list.prev);
1335 static void rb_list_head_clear(struct list_head *list)
1337 unsigned long *ptr = (unsigned long *)&list->next;
1339 *ptr &= ~RB_FLAG_MASK;
1343 * rb_head_page_deactivate - clears head page ptr (for free list)
1346 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
1348 struct list_head *hd;
1350 /* Go through the whole list and clear any pointers found. */
1351 rb_list_head_clear(cpu_buffer->pages);
1353 list_for_each(hd, cpu_buffer->pages)
1354 rb_list_head_clear(hd);
1357 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
1358 struct buffer_page *head,
1359 struct buffer_page *prev,
1360 int old_flag, int new_flag)
1362 struct list_head *list;
1363 unsigned long val = (unsigned long)&head->list;
1368 val &= ~RB_FLAG_MASK;
1370 ret = cmpxchg((unsigned long *)&list->next,
1371 val | old_flag, val | new_flag);
1373 /* check if the reader took the page */
1374 if ((ret & ~RB_FLAG_MASK) != val)
1375 return RB_PAGE_MOVED;
1377 return ret & RB_FLAG_MASK;
1380 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
1381 struct buffer_page *head,
1382 struct buffer_page *prev,
1385 return rb_head_page_set(cpu_buffer, head, prev,
1386 old_flag, RB_PAGE_UPDATE);
1389 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
1390 struct buffer_page *head,
1391 struct buffer_page *prev,
1394 return rb_head_page_set(cpu_buffer, head, prev,
1395 old_flag, RB_PAGE_HEAD);
1398 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
1399 struct buffer_page *head,
1400 struct buffer_page *prev,
1403 return rb_head_page_set(cpu_buffer, head, prev,
1404 old_flag, RB_PAGE_NORMAL);
1407 static inline void rb_inc_page(struct buffer_page **bpage)
1409 struct list_head *p = rb_list_head((*bpage)->list.next);
1411 *bpage = list_entry(p, struct buffer_page, list);
1414 static struct buffer_page *
1415 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1417 struct buffer_page *head;
1418 struct buffer_page *page;
1419 struct list_head *list;
1422 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1426 list = cpu_buffer->pages;
1427 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1430 page = head = cpu_buffer->head_page;
1432 * It is possible that the writer moves the header behind
1433 * where we started, and we miss in one loop.
1434 * A second loop should grab the header, but we'll do
1435 * three loops just because I'm paranoid.
1437 for (i = 0; i < 3; i++) {
1439 if (rb_is_head_page(page, page->list.prev)) {
1440 cpu_buffer->head_page = page;
1444 } while (page != head);
1447 RB_WARN_ON(cpu_buffer, 1);
1452 static bool rb_head_page_replace(struct buffer_page *old,
1453 struct buffer_page *new)
1455 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1458 val = *ptr & ~RB_FLAG_MASK;
1459 val |= RB_PAGE_HEAD;
1461 return try_cmpxchg(ptr, &val, (unsigned long)&new->list);
1465 * rb_tail_page_update - move the tail page forward
1467 static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1468 struct buffer_page *tail_page,
1469 struct buffer_page *next_page)
1471 unsigned long old_entries;
1472 unsigned long old_write;
1475 * The tail page now needs to be moved forward.
1477 * We need to reset the tail page, but without messing
1478 * with possible erasing of data brought in by interrupts
1479 * that have moved the tail page and are currently on it.
1481 * We add a counter to the write field to denote this.
1483 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1484 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1486 local_inc(&cpu_buffer->pages_touched);
1488 * Just make sure we have seen our old_write and synchronize
1489 * with any interrupts that come in.
1494 * If the tail page is still the same as what we think
1495 * it is, then it is up to us to update the tail
1498 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1499 /* Zero the write counter */
1500 unsigned long val = old_write & ~RB_WRITE_MASK;
1501 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1504 * This will only succeed if an interrupt did
1505 * not come in and change it. In which case, we
1506 * do not want to modify it.
1508 * We add (void) to let the compiler know that we do not care
1509 * about the return value of these functions. We use the
1510 * cmpxchg to only update if an interrupt did not already
1511 * do it for us. If the cmpxchg fails, we don't care.
1513 (void)local_cmpxchg(&next_page->write, old_write, val);
1514 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1517 * No need to worry about races with clearing out the commit.
1518 * it only can increment when a commit takes place. But that
1519 * only happens in the outer most nested commit.
1521 local_set(&next_page->page->commit, 0);
1523 /* Again, either we update tail_page or an interrupt does */
1524 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1528 static void rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1529 struct buffer_page *bpage)
1531 unsigned long val = (unsigned long)bpage;
1533 RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK);
1537 * rb_check_pages - integrity check of buffer pages
1538 * @cpu_buffer: CPU buffer with pages to test
1540 * As a safety measure we check to make sure the data pages have not
1543 static void rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1545 struct list_head *head = rb_list_head(cpu_buffer->pages);
1546 struct list_head *tmp;
1548 if (RB_WARN_ON(cpu_buffer,
1549 rb_list_head(rb_list_head(head->next)->prev) != head))
1552 if (RB_WARN_ON(cpu_buffer,
1553 rb_list_head(rb_list_head(head->prev)->next) != head))
1556 for (tmp = rb_list_head(head->next); tmp != head; tmp = rb_list_head(tmp->next)) {
1557 if (RB_WARN_ON(cpu_buffer,
1558 rb_list_head(rb_list_head(tmp->next)->prev) != tmp))
1561 if (RB_WARN_ON(cpu_buffer,
1562 rb_list_head(rb_list_head(tmp->prev)->next) != tmp))
1567 static int __rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1568 long nr_pages, struct list_head *pages)
1570 struct buffer_page *bpage, *tmp;
1571 bool user_thread = current->mm != NULL;
1576 * Check if the available memory is there first.
1577 * Note, si_mem_available() only gives us a rough estimate of available
1578 * memory. It may not be accurate. But we don't care, we just want
1579 * to prevent doing any allocation when it is obvious that it is
1580 * not going to succeed.
1582 i = si_mem_available();
1587 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1588 * gracefully without invoking oom-killer and the system is not
1591 mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1594 * If a user thread allocates too much, and si_mem_available()
1595 * reports there's enough memory, even though there is not.
1596 * Make sure the OOM killer kills this thread. This can happen
1597 * even with RETRY_MAYFAIL because another task may be doing
1598 * an allocation after this task has taken all memory.
1599 * This is the task the OOM killer needs to take out during this
1600 * loop, even if it was triggered by an allocation somewhere else.
1603 set_current_oom_origin();
1604 for (i = 0; i < nr_pages; i++) {
1607 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1608 mflags, cpu_to_node(cpu_buffer->cpu));
1612 rb_check_bpage(cpu_buffer, bpage);
1614 list_add(&bpage->list, pages);
1616 page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu), mflags, 0);
1619 bpage->page = page_address(page);
1620 rb_init_page(bpage->page);
1622 if (user_thread && fatal_signal_pending(current))
1626 clear_current_oom_origin();
1631 list_for_each_entry_safe(bpage, tmp, pages, list) {
1632 list_del_init(&bpage->list);
1633 free_buffer_page(bpage);
1636 clear_current_oom_origin();
1641 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1642 unsigned long nr_pages)
1648 if (__rb_allocate_pages(cpu_buffer, nr_pages, &pages))
1652 * The ring buffer page list is a circular list that does not
1653 * start and end with a list head. All page list items point to
1656 cpu_buffer->pages = pages.next;
1659 cpu_buffer->nr_pages = nr_pages;
1661 rb_check_pages(cpu_buffer);
1666 static struct ring_buffer_per_cpu *
1667 rb_allocate_cpu_buffer(struct trace_buffer *buffer, long nr_pages, int cpu)
1669 struct ring_buffer_per_cpu *cpu_buffer;
1670 struct buffer_page *bpage;
1674 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1675 GFP_KERNEL, cpu_to_node(cpu));
1679 cpu_buffer->cpu = cpu;
1680 cpu_buffer->buffer = buffer;
1681 raw_spin_lock_init(&cpu_buffer->reader_lock);
1682 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1683 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1684 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1685 init_completion(&cpu_buffer->update_done);
1686 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1687 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1688 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1690 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1691 GFP_KERNEL, cpu_to_node(cpu));
1693 goto fail_free_buffer;
1695 rb_check_bpage(cpu_buffer, bpage);
1697 cpu_buffer->reader_page = bpage;
1698 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1700 goto fail_free_reader;
1701 bpage->page = page_address(page);
1702 rb_init_page(bpage->page);
1704 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1705 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1707 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1709 goto fail_free_reader;
1711 cpu_buffer->head_page
1712 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1713 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1715 rb_head_page_activate(cpu_buffer);
1720 free_buffer_page(cpu_buffer->reader_page);
1727 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1729 struct list_head *head = cpu_buffer->pages;
1730 struct buffer_page *bpage, *tmp;
1732 irq_work_sync(&cpu_buffer->irq_work.work);
1734 free_buffer_page(cpu_buffer->reader_page);
1737 rb_head_page_deactivate(cpu_buffer);
1739 list_for_each_entry_safe(bpage, tmp, head, list) {
1740 list_del_init(&bpage->list);
1741 free_buffer_page(bpage);
1743 bpage = list_entry(head, struct buffer_page, list);
1744 free_buffer_page(bpage);
1747 free_page((unsigned long)cpu_buffer->free_page);
1753 * __ring_buffer_alloc - allocate a new ring_buffer
1754 * @size: the size in bytes per cpu that is needed.
1755 * @flags: attributes to set for the ring buffer.
1756 * @key: ring buffer reader_lock_key.
1758 * Currently the only flag that is available is the RB_FL_OVERWRITE
1759 * flag. This flag means that the buffer will overwrite old data
1760 * when the buffer wraps. If this flag is not set, the buffer will
1761 * drop data when the tail hits the head.
1763 struct trace_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1764 struct lock_class_key *key)
1766 struct trace_buffer *buffer;
1772 /* keep it in its own cache line */
1773 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1778 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1779 goto fail_free_buffer;
1781 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1782 buffer->flags = flags;
1783 buffer->clock = trace_clock_local;
1784 buffer->reader_lock_key = key;
1786 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1787 init_waitqueue_head(&buffer->irq_work.waiters);
1789 /* need at least two pages */
1793 buffer->cpus = nr_cpu_ids;
1795 bsize = sizeof(void *) * nr_cpu_ids;
1796 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1798 if (!buffer->buffers)
1799 goto fail_free_cpumask;
1801 cpu = raw_smp_processor_id();
1802 cpumask_set_cpu(cpu, buffer->cpumask);
1803 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1804 if (!buffer->buffers[cpu])
1805 goto fail_free_buffers;
1807 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1809 goto fail_free_buffers;
1811 mutex_init(&buffer->mutex);
1816 for_each_buffer_cpu(buffer, cpu) {
1817 if (buffer->buffers[cpu])
1818 rb_free_cpu_buffer(buffer->buffers[cpu]);
1820 kfree(buffer->buffers);
1823 free_cpumask_var(buffer->cpumask);
1829 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1832 * ring_buffer_free - free a ring buffer.
1833 * @buffer: the buffer to free.
1836 ring_buffer_free(struct trace_buffer *buffer)
1840 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1842 irq_work_sync(&buffer->irq_work.work);
1844 for_each_buffer_cpu(buffer, cpu)
1845 rb_free_cpu_buffer(buffer->buffers[cpu]);
1847 kfree(buffer->buffers);
1848 free_cpumask_var(buffer->cpumask);
1852 EXPORT_SYMBOL_GPL(ring_buffer_free);
1854 void ring_buffer_set_clock(struct trace_buffer *buffer,
1857 buffer->clock = clock;
1860 void ring_buffer_set_time_stamp_abs(struct trace_buffer *buffer, bool abs)
1862 buffer->time_stamp_abs = abs;
1865 bool ring_buffer_time_stamp_abs(struct trace_buffer *buffer)
1867 return buffer->time_stamp_abs;
1870 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1872 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1874 return local_read(&bpage->entries) & RB_WRITE_MASK;
1877 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1879 return local_read(&bpage->write) & RB_WRITE_MASK;
1883 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1885 struct list_head *tail_page, *to_remove, *next_page;
1886 struct buffer_page *to_remove_page, *tmp_iter_page;
1887 struct buffer_page *last_page, *first_page;
1888 unsigned long nr_removed;
1889 unsigned long head_bit;
1894 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1895 atomic_inc(&cpu_buffer->record_disabled);
1897 * We don't race with the readers since we have acquired the reader
1898 * lock. We also don't race with writers after disabling recording.
1899 * This makes it easy to figure out the first and the last page to be
1900 * removed from the list. We unlink all the pages in between including
1901 * the first and last pages. This is done in a busy loop so that we
1902 * lose the least number of traces.
1903 * The pages are freed after we restart recording and unlock readers.
1905 tail_page = &cpu_buffer->tail_page->list;
1908 * tail page might be on reader page, we remove the next page
1909 * from the ring buffer
1911 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1912 tail_page = rb_list_head(tail_page->next);
1913 to_remove = tail_page;
1915 /* start of pages to remove */
1916 first_page = list_entry(rb_list_head(to_remove->next),
1917 struct buffer_page, list);
1919 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1920 to_remove = rb_list_head(to_remove)->next;
1921 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1923 /* Read iterators need to reset themselves when some pages removed */
1924 cpu_buffer->pages_removed += nr_removed;
1926 next_page = rb_list_head(to_remove)->next;
1929 * Now we remove all pages between tail_page and next_page.
1930 * Make sure that we have head_bit value preserved for the
1933 tail_page->next = (struct list_head *)((unsigned long)next_page |
1935 next_page = rb_list_head(next_page);
1936 next_page->prev = tail_page;
1938 /* make sure pages points to a valid page in the ring buffer */
1939 cpu_buffer->pages = next_page;
1941 /* update head page */
1943 cpu_buffer->head_page = list_entry(next_page,
1944 struct buffer_page, list);
1946 /* pages are removed, resume tracing and then free the pages */
1947 atomic_dec(&cpu_buffer->record_disabled);
1948 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1950 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1952 /* last buffer page to remove */
1953 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1955 tmp_iter_page = first_page;
1960 to_remove_page = tmp_iter_page;
1961 rb_inc_page(&tmp_iter_page);
1963 /* update the counters */
1964 page_entries = rb_page_entries(to_remove_page);
1967 * If something was added to this page, it was full
1968 * since it is not the tail page. So we deduct the
1969 * bytes consumed in ring buffer from here.
1970 * Increment overrun to account for the lost events.
1972 local_add(page_entries, &cpu_buffer->overrun);
1973 local_sub(rb_page_commit(to_remove_page), &cpu_buffer->entries_bytes);
1974 local_inc(&cpu_buffer->pages_lost);
1978 * We have already removed references to this list item, just
1979 * free up the buffer_page and its page
1981 free_buffer_page(to_remove_page);
1984 } while (to_remove_page != last_page);
1986 RB_WARN_ON(cpu_buffer, nr_removed);
1988 return nr_removed == 0;
1992 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1994 struct list_head *pages = &cpu_buffer->new_pages;
1995 unsigned long flags;
1999 /* Can be called at early boot up, where interrupts must not been enabled */
2000 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2002 * We are holding the reader lock, so the reader page won't be swapped
2003 * in the ring buffer. Now we are racing with the writer trying to
2004 * move head page and the tail page.
2005 * We are going to adapt the reader page update process where:
2006 * 1. We first splice the start and end of list of new pages between
2007 * the head page and its previous page.
2008 * 2. We cmpxchg the prev_page->next to point from head page to the
2009 * start of new pages list.
2010 * 3. Finally, we update the head->prev to the end of new list.
2012 * We will try this process 10 times, to make sure that we don't keep
2018 struct list_head *head_page, *prev_page;
2019 struct list_head *last_page, *first_page;
2020 struct list_head *head_page_with_bit;
2021 struct buffer_page *hpage = rb_set_head_page(cpu_buffer);
2025 head_page = &hpage->list;
2026 prev_page = head_page->prev;
2028 first_page = pages->next;
2029 last_page = pages->prev;
2031 head_page_with_bit = (struct list_head *)
2032 ((unsigned long)head_page | RB_PAGE_HEAD);
2034 last_page->next = head_page_with_bit;
2035 first_page->prev = prev_page;
2037 /* caution: head_page_with_bit gets updated on cmpxchg failure */
2038 if (try_cmpxchg(&prev_page->next,
2039 &head_page_with_bit, first_page)) {
2041 * yay, we replaced the page pointer to our new list,
2042 * now, we just have to update to head page's prev
2043 * pointer to point to end of list
2045 head_page->prev = last_page;
2052 INIT_LIST_HEAD(pages);
2054 * If we weren't successful in adding in new pages, warn and stop
2057 RB_WARN_ON(cpu_buffer, !success);
2058 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2060 /* free pages if they weren't inserted */
2062 struct buffer_page *bpage, *tmp;
2063 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2065 list_del_init(&bpage->list);
2066 free_buffer_page(bpage);
2072 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
2076 if (cpu_buffer->nr_pages_to_update > 0)
2077 success = rb_insert_pages(cpu_buffer);
2079 success = rb_remove_pages(cpu_buffer,
2080 -cpu_buffer->nr_pages_to_update);
2083 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
2086 static void update_pages_handler(struct work_struct *work)
2088 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
2089 struct ring_buffer_per_cpu, update_pages_work);
2090 rb_update_pages(cpu_buffer);
2091 complete(&cpu_buffer->update_done);
2095 * ring_buffer_resize - resize the ring buffer
2096 * @buffer: the buffer to resize.
2097 * @size: the new size.
2098 * @cpu_id: the cpu buffer to resize
2100 * Minimum size is 2 * BUF_PAGE_SIZE.
2102 * Returns 0 on success and < 0 on failure.
2104 int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
2107 struct ring_buffer_per_cpu *cpu_buffer;
2108 unsigned long nr_pages;
2112 * Always succeed at resizing a non-existent buffer:
2117 /* Make sure the requested buffer exists */
2118 if (cpu_id != RING_BUFFER_ALL_CPUS &&
2119 !cpumask_test_cpu(cpu_id, buffer->cpumask))
2122 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
2124 /* we need a minimum of two pages */
2128 /* prevent another thread from changing buffer sizes */
2129 mutex_lock(&buffer->mutex);
2130 atomic_inc(&buffer->resizing);
2132 if (cpu_id == RING_BUFFER_ALL_CPUS) {
2134 * Don't succeed if resizing is disabled, as a reader might be
2135 * manipulating the ring buffer and is expecting a sane state while
2138 for_each_buffer_cpu(buffer, cpu) {
2139 cpu_buffer = buffer->buffers[cpu];
2140 if (atomic_read(&cpu_buffer->resize_disabled)) {
2142 goto out_err_unlock;
2146 /* calculate the pages to update */
2147 for_each_buffer_cpu(buffer, cpu) {
2148 cpu_buffer = buffer->buffers[cpu];
2150 cpu_buffer->nr_pages_to_update = nr_pages -
2151 cpu_buffer->nr_pages;
2153 * nothing more to do for removing pages or no update
2155 if (cpu_buffer->nr_pages_to_update <= 0)
2158 * to add pages, make sure all new pages can be
2159 * allocated without receiving ENOMEM
2161 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2162 if (__rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2163 &cpu_buffer->new_pages)) {
2164 /* not enough memory for new pages */
2174 * Fire off all the required work handlers
2175 * We can't schedule on offline CPUs, but it's not necessary
2176 * since we can change their buffer sizes without any race.
2178 for_each_buffer_cpu(buffer, cpu) {
2179 cpu_buffer = buffer->buffers[cpu];
2180 if (!cpu_buffer->nr_pages_to_update)
2183 /* Can't run something on an offline CPU. */
2184 if (!cpu_online(cpu)) {
2185 rb_update_pages(cpu_buffer);
2186 cpu_buffer->nr_pages_to_update = 0;
2188 /* Run directly if possible. */
2190 if (cpu != smp_processor_id()) {
2192 schedule_work_on(cpu,
2193 &cpu_buffer->update_pages_work);
2195 update_pages_handler(&cpu_buffer->update_pages_work);
2201 /* wait for all the updates to complete */
2202 for_each_buffer_cpu(buffer, cpu) {
2203 cpu_buffer = buffer->buffers[cpu];
2204 if (!cpu_buffer->nr_pages_to_update)
2207 if (cpu_online(cpu))
2208 wait_for_completion(&cpu_buffer->update_done);
2209 cpu_buffer->nr_pages_to_update = 0;
2214 cpu_buffer = buffer->buffers[cpu_id];
2216 if (nr_pages == cpu_buffer->nr_pages)
2220 * Don't succeed if resizing is disabled, as a reader might be
2221 * manipulating the ring buffer and is expecting a sane state while
2224 if (atomic_read(&cpu_buffer->resize_disabled)) {
2226 goto out_err_unlock;
2229 cpu_buffer->nr_pages_to_update = nr_pages -
2230 cpu_buffer->nr_pages;
2232 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2233 if (cpu_buffer->nr_pages_to_update > 0 &&
2234 __rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2235 &cpu_buffer->new_pages)) {
2242 /* Can't run something on an offline CPU. */
2243 if (!cpu_online(cpu_id))
2244 rb_update_pages(cpu_buffer);
2246 /* Run directly if possible. */
2248 if (cpu_id == smp_processor_id()) {
2249 rb_update_pages(cpu_buffer);
2253 schedule_work_on(cpu_id,
2254 &cpu_buffer->update_pages_work);
2255 wait_for_completion(&cpu_buffer->update_done);
2259 cpu_buffer->nr_pages_to_update = 0;
2265 * The ring buffer resize can happen with the ring buffer
2266 * enabled, so that the update disturbs the tracing as little
2267 * as possible. But if the buffer is disabled, we do not need
2268 * to worry about that, and we can take the time to verify
2269 * that the buffer is not corrupt.
2271 if (atomic_read(&buffer->record_disabled)) {
2272 atomic_inc(&buffer->record_disabled);
2274 * Even though the buffer was disabled, we must make sure
2275 * that it is truly disabled before calling rb_check_pages.
2276 * There could have been a race between checking
2277 * record_disable and incrementing it.
2280 for_each_buffer_cpu(buffer, cpu) {
2281 cpu_buffer = buffer->buffers[cpu];
2282 rb_check_pages(cpu_buffer);
2284 atomic_dec(&buffer->record_disabled);
2287 atomic_dec(&buffer->resizing);
2288 mutex_unlock(&buffer->mutex);
2292 for_each_buffer_cpu(buffer, cpu) {
2293 struct buffer_page *bpage, *tmp;
2295 cpu_buffer = buffer->buffers[cpu];
2296 cpu_buffer->nr_pages_to_update = 0;
2298 if (list_empty(&cpu_buffer->new_pages))
2301 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2303 list_del_init(&bpage->list);
2304 free_buffer_page(bpage);
2308 atomic_dec(&buffer->resizing);
2309 mutex_unlock(&buffer->mutex);
2312 EXPORT_SYMBOL_GPL(ring_buffer_resize);
2314 void ring_buffer_change_overwrite(struct trace_buffer *buffer, int val)
2316 mutex_lock(&buffer->mutex);
2318 buffer->flags |= RB_FL_OVERWRITE;
2320 buffer->flags &= ~RB_FL_OVERWRITE;
2321 mutex_unlock(&buffer->mutex);
2323 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
2325 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
2327 return bpage->page->data + index;
2330 static __always_inline struct ring_buffer_event *
2331 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
2333 return __rb_page_index(cpu_buffer->reader_page,
2334 cpu_buffer->reader_page->read);
2337 static struct ring_buffer_event *
2338 rb_iter_head_event(struct ring_buffer_iter *iter)
2340 struct ring_buffer_event *event;
2341 struct buffer_page *iter_head_page = iter->head_page;
2342 unsigned long commit;
2345 if (iter->head != iter->next_event)
2349 * When the writer goes across pages, it issues a cmpxchg which
2350 * is a mb(), which will synchronize with the rmb here.
2351 * (see rb_tail_page_update() and __rb_reserve_next())
2353 commit = rb_page_commit(iter_head_page);
2356 /* An event needs to be at least 8 bytes in size */
2357 if (iter->head > commit - 8)
2360 event = __rb_page_index(iter_head_page, iter->head);
2361 length = rb_event_length(event);
2364 * READ_ONCE() doesn't work on functions and we don't want the
2365 * compiler doing any crazy optimizations with length.
2369 if ((iter->head + length) > commit || length > BUF_PAGE_SIZE)
2370 /* Writer corrupted the read? */
2373 memcpy(iter->event, event, length);
2375 * If the page stamp is still the same after this rmb() then the
2376 * event was safely copied without the writer entering the page.
2380 /* Make sure the page didn't change since we read this */
2381 if (iter->page_stamp != iter_head_page->page->time_stamp ||
2382 commit > rb_page_commit(iter_head_page))
2385 iter->next_event = iter->head + length;
2388 /* Reset to the beginning */
2389 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2391 iter->next_event = 0;
2392 iter->missed_events = 1;
2396 /* Size is determined by what has been committed */
2397 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
2399 return rb_page_commit(bpage);
2402 static __always_inline unsigned
2403 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
2405 return rb_page_commit(cpu_buffer->commit_page);
2408 static __always_inline unsigned
2409 rb_event_index(struct ring_buffer_event *event)
2411 unsigned long addr = (unsigned long)event;
2413 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
2416 static void rb_inc_iter(struct ring_buffer_iter *iter)
2418 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2421 * The iterator could be on the reader page (it starts there).
2422 * But the head could have moved, since the reader was
2423 * found. Check for this case and assign the iterator
2424 * to the head page instead of next.
2426 if (iter->head_page == cpu_buffer->reader_page)
2427 iter->head_page = rb_set_head_page(cpu_buffer);
2429 rb_inc_page(&iter->head_page);
2431 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2433 iter->next_event = 0;
2437 * rb_handle_head_page - writer hit the head page
2439 * Returns: +1 to retry page
2444 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2445 struct buffer_page *tail_page,
2446 struct buffer_page *next_page)
2448 struct buffer_page *new_head;
2453 entries = rb_page_entries(next_page);
2456 * The hard part is here. We need to move the head
2457 * forward, and protect against both readers on
2458 * other CPUs and writers coming in via interrupts.
2460 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2464 * type can be one of four:
2465 * NORMAL - an interrupt already moved it for us
2466 * HEAD - we are the first to get here.
2467 * UPDATE - we are the interrupt interrupting
2469 * MOVED - a reader on another CPU moved the next
2470 * pointer to its reader page. Give up
2477 * We changed the head to UPDATE, thus
2478 * it is our responsibility to update
2481 local_add(entries, &cpu_buffer->overrun);
2482 local_sub(rb_page_commit(next_page), &cpu_buffer->entries_bytes);
2483 local_inc(&cpu_buffer->pages_lost);
2486 * The entries will be zeroed out when we move the
2490 /* still more to do */
2493 case RB_PAGE_UPDATE:
2495 * This is an interrupt that interrupt the
2496 * previous update. Still more to do.
2499 case RB_PAGE_NORMAL:
2501 * An interrupt came in before the update
2502 * and processed this for us.
2503 * Nothing left to do.
2508 * The reader is on another CPU and just did
2509 * a swap with our next_page.
2514 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2519 * Now that we are here, the old head pointer is
2520 * set to UPDATE. This will keep the reader from
2521 * swapping the head page with the reader page.
2522 * The reader (on another CPU) will spin till
2525 * We just need to protect against interrupts
2526 * doing the job. We will set the next pointer
2527 * to HEAD. After that, we set the old pointer
2528 * to NORMAL, but only if it was HEAD before.
2529 * otherwise we are an interrupt, and only
2530 * want the outer most commit to reset it.
2532 new_head = next_page;
2533 rb_inc_page(&new_head);
2535 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2539 * Valid returns are:
2540 * HEAD - an interrupt came in and already set it.
2541 * NORMAL - One of two things:
2542 * 1) We really set it.
2543 * 2) A bunch of interrupts came in and moved
2544 * the page forward again.
2548 case RB_PAGE_NORMAL:
2552 RB_WARN_ON(cpu_buffer, 1);
2557 * It is possible that an interrupt came in,
2558 * set the head up, then more interrupts came in
2559 * and moved it again. When we get back here,
2560 * the page would have been set to NORMAL but we
2561 * just set it back to HEAD.
2563 * How do you detect this? Well, if that happened
2564 * the tail page would have moved.
2566 if (ret == RB_PAGE_NORMAL) {
2567 struct buffer_page *buffer_tail_page;
2569 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2571 * If the tail had moved passed next, then we need
2572 * to reset the pointer.
2574 if (buffer_tail_page != tail_page &&
2575 buffer_tail_page != next_page)
2576 rb_head_page_set_normal(cpu_buffer, new_head,
2582 * If this was the outer most commit (the one that
2583 * changed the original pointer from HEAD to UPDATE),
2584 * then it is up to us to reset it to NORMAL.
2586 if (type == RB_PAGE_HEAD) {
2587 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2590 if (RB_WARN_ON(cpu_buffer,
2591 ret != RB_PAGE_UPDATE))
2599 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2600 unsigned long tail, struct rb_event_info *info)
2602 struct buffer_page *tail_page = info->tail_page;
2603 struct ring_buffer_event *event;
2604 unsigned long length = info->length;
2607 * Only the event that crossed the page boundary
2608 * must fill the old tail_page with padding.
2610 if (tail >= BUF_PAGE_SIZE) {
2612 * If the page was filled, then we still need
2613 * to update the real_end. Reset it to zero
2614 * and the reader will ignore it.
2616 if (tail == BUF_PAGE_SIZE)
2617 tail_page->real_end = 0;
2619 local_sub(length, &tail_page->write);
2623 event = __rb_page_index(tail_page, tail);
2626 * Save the original length to the meta data.
2627 * This will be used by the reader to add lost event
2630 tail_page->real_end = tail;
2633 * If this event is bigger than the minimum size, then
2634 * we need to be careful that we don't subtract the
2635 * write counter enough to allow another writer to slip
2637 * We put in a discarded commit instead, to make sure
2638 * that this space is not used again, and this space will
2639 * not be accounted into 'entries_bytes'.
2641 * If we are less than the minimum size, we don't need to
2644 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2645 /* No room for any events */
2647 /* Mark the rest of the page with padding */
2648 rb_event_set_padding(event);
2650 /* Make sure the padding is visible before the write update */
2653 /* Set the write back to the previous setting */
2654 local_sub(length, &tail_page->write);
2658 /* Put in a discarded event */
2659 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2660 event->type_len = RINGBUF_TYPE_PADDING;
2661 /* time delta must be non zero */
2662 event->time_delta = 1;
2664 /* account for padding bytes */
2665 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2667 /* Make sure the padding is visible before the tail_page->write update */
2670 /* Set write to end of buffer */
2671 length = (tail + length) - BUF_PAGE_SIZE;
2672 local_sub(length, &tail_page->write);
2675 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2678 * This is the slow path, force gcc not to inline it.
2680 static noinline struct ring_buffer_event *
2681 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2682 unsigned long tail, struct rb_event_info *info)
2684 struct buffer_page *tail_page = info->tail_page;
2685 struct buffer_page *commit_page = cpu_buffer->commit_page;
2686 struct trace_buffer *buffer = cpu_buffer->buffer;
2687 struct buffer_page *next_page;
2690 next_page = tail_page;
2692 rb_inc_page(&next_page);
2695 * If for some reason, we had an interrupt storm that made
2696 * it all the way around the buffer, bail, and warn
2699 if (unlikely(next_page == commit_page)) {
2700 local_inc(&cpu_buffer->commit_overrun);
2705 * This is where the fun begins!
2707 * We are fighting against races between a reader that
2708 * could be on another CPU trying to swap its reader
2709 * page with the buffer head.
2711 * We are also fighting against interrupts coming in and
2712 * moving the head or tail on us as well.
2714 * If the next page is the head page then we have filled
2715 * the buffer, unless the commit page is still on the
2718 if (rb_is_head_page(next_page, &tail_page->list)) {
2721 * If the commit is not on the reader page, then
2722 * move the header page.
2724 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2726 * If we are not in overwrite mode,
2727 * this is easy, just stop here.
2729 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2730 local_inc(&cpu_buffer->dropped_events);
2734 ret = rb_handle_head_page(cpu_buffer,
2743 * We need to be careful here too. The
2744 * commit page could still be on the reader
2745 * page. We could have a small buffer, and
2746 * have filled up the buffer with events
2747 * from interrupts and such, and wrapped.
2749 * Note, if the tail page is also on the
2750 * reader_page, we let it move out.
2752 if (unlikely((cpu_buffer->commit_page !=
2753 cpu_buffer->tail_page) &&
2754 (cpu_buffer->commit_page ==
2755 cpu_buffer->reader_page))) {
2756 local_inc(&cpu_buffer->commit_overrun);
2762 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2766 rb_reset_tail(cpu_buffer, tail, info);
2768 /* Commit what we have for now. */
2769 rb_end_commit(cpu_buffer);
2770 /* rb_end_commit() decs committing */
2771 local_inc(&cpu_buffer->committing);
2773 /* fail and let the caller try again */
2774 return ERR_PTR(-EAGAIN);
2778 rb_reset_tail(cpu_buffer, tail, info);
2784 static struct ring_buffer_event *
2785 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2788 event->type_len = RINGBUF_TYPE_TIME_STAMP;
2790 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2792 /* Not the first event on the page, or not delta? */
2793 if (abs || rb_event_index(event)) {
2794 event->time_delta = delta & TS_MASK;
2795 event->array[0] = delta >> TS_SHIFT;
2797 /* nope, just zero it */
2798 event->time_delta = 0;
2799 event->array[0] = 0;
2802 return skip_time_extend(event);
2805 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2806 static inline bool sched_clock_stable(void)
2813 rb_check_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2814 struct rb_event_info *info)
2818 WARN_ONCE(1, "Delta way too big! %llu ts=%llu before=%llu after=%llu write stamp=%llu\n%s",
2819 (unsigned long long)info->delta,
2820 (unsigned long long)info->ts,
2821 (unsigned long long)info->before,
2822 (unsigned long long)info->after,
2823 (unsigned long long)(rb_time_read(&cpu_buffer->write_stamp, &write_stamp) ? write_stamp : 0),
2824 sched_clock_stable() ? "" :
2825 "If you just came from a suspend/resume,\n"
2826 "please switch to the trace global clock:\n"
2827 " echo global > /sys/kernel/tracing/trace_clock\n"
2828 "or add trace_clock=global to the kernel command line\n");
2831 static void rb_add_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2832 struct ring_buffer_event **event,
2833 struct rb_event_info *info,
2835 unsigned int *length)
2837 bool abs = info->add_timestamp &
2838 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE);
2840 if (unlikely(info->delta > (1ULL << 59))) {
2842 * Some timers can use more than 59 bits, and when a timestamp
2843 * is added to the buffer, it will lose those bits.
2845 if (abs && (info->ts & TS_MSB)) {
2846 info->delta &= ABS_TS_MASK;
2848 /* did the clock go backwards */
2849 } else if (info->before == info->after && info->before > info->ts) {
2850 /* not interrupted */
2854 * This is possible with a recalibrating of the TSC.
2855 * Do not produce a call stack, but just report it.
2859 pr_warn("Ring buffer clock went backwards: %llu -> %llu\n",
2860 info->before, info->ts);
2863 rb_check_timestamp(cpu_buffer, info);
2867 *event = rb_add_time_stamp(*event, info->delta, abs);
2868 *length -= RB_LEN_TIME_EXTEND;
2873 * rb_update_event - update event type and data
2874 * @cpu_buffer: The per cpu buffer of the @event
2875 * @event: the event to update
2876 * @info: The info to update the @event with (contains length and delta)
2878 * Update the type and data fields of the @event. The length
2879 * is the actual size that is written to the ring buffer,
2880 * and with this, we can determine what to place into the
2884 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2885 struct ring_buffer_event *event,
2886 struct rb_event_info *info)
2888 unsigned length = info->length;
2889 u64 delta = info->delta;
2890 unsigned int nest = local_read(&cpu_buffer->committing) - 1;
2892 if (!WARN_ON_ONCE(nest >= MAX_NEST))
2893 cpu_buffer->event_stamp[nest] = info->ts;
2896 * If we need to add a timestamp, then we
2897 * add it to the start of the reserved space.
2899 if (unlikely(info->add_timestamp))
2900 rb_add_timestamp(cpu_buffer, &event, info, &delta, &length);
2902 event->time_delta = delta;
2903 length -= RB_EVNT_HDR_SIZE;
2904 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2905 event->type_len = 0;
2906 event->array[0] = length;
2908 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2911 static unsigned rb_calculate_event_length(unsigned length)
2913 struct ring_buffer_event event; /* Used only for sizeof array */
2915 /* zero length can cause confusions */
2919 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2920 length += sizeof(event.array[0]);
2922 length += RB_EVNT_HDR_SIZE;
2923 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2926 * In case the time delta is larger than the 27 bits for it
2927 * in the header, we need to add a timestamp. If another
2928 * event comes in when trying to discard this one to increase
2929 * the length, then the timestamp will be added in the allocated
2930 * space of this event. If length is bigger than the size needed
2931 * for the TIME_EXTEND, then padding has to be used. The events
2932 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2933 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2934 * As length is a multiple of 4, we only need to worry if it
2935 * is 12 (RB_LEN_TIME_EXTEND + 4).
2937 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2938 length += RB_ALIGNMENT;
2944 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2945 struct ring_buffer_event *event)
2947 unsigned long new_index, old_index;
2948 struct buffer_page *bpage;
2951 new_index = rb_event_index(event);
2952 old_index = new_index + rb_event_ts_length(event);
2953 addr = (unsigned long)event;
2956 bpage = READ_ONCE(cpu_buffer->tail_page);
2959 * Make sure the tail_page is still the same and
2960 * the next write location is the end of this event
2962 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2963 unsigned long write_mask =
2964 local_read(&bpage->write) & ~RB_WRITE_MASK;
2965 unsigned long event_length = rb_event_length(event);
2968 * For the before_stamp to be different than the write_stamp
2969 * to make sure that the next event adds an absolute
2970 * value and does not rely on the saved write stamp, which
2971 * is now going to be bogus.
2973 * By setting the before_stamp to zero, the next event
2974 * is not going to use the write_stamp and will instead
2975 * create an absolute timestamp. This means there's no
2976 * reason to update the wirte_stamp!
2978 rb_time_set(&cpu_buffer->before_stamp, 0);
2981 * If an event were to come in now, it would see that the
2982 * write_stamp and the before_stamp are different, and assume
2983 * that this event just added itself before updating
2984 * the write stamp. The interrupting event will fix the
2985 * write stamp for us, and use an absolute timestamp.
2989 * This is on the tail page. It is possible that
2990 * a write could come in and move the tail page
2991 * and write to the next page. That is fine
2992 * because we just shorten what is on this page.
2994 old_index += write_mask;
2995 new_index += write_mask;
2997 /* caution: old_index gets updated on cmpxchg failure */
2998 if (local_try_cmpxchg(&bpage->write, &old_index, new_index)) {
2999 /* update counters */
3000 local_sub(event_length, &cpu_buffer->entries_bytes);
3005 /* could not discard */
3009 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
3011 local_inc(&cpu_buffer->committing);
3012 local_inc(&cpu_buffer->commits);
3015 static __always_inline void
3016 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
3018 unsigned long max_count;
3021 * We only race with interrupts and NMIs on this CPU.
3022 * If we own the commit event, then we can commit
3023 * all others that interrupted us, since the interruptions
3024 * are in stack format (they finish before they come
3025 * back to us). This allows us to do a simple loop to
3026 * assign the commit to the tail.
3029 max_count = cpu_buffer->nr_pages * 100;
3031 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
3032 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
3034 if (RB_WARN_ON(cpu_buffer,
3035 rb_is_reader_page(cpu_buffer->tail_page)))
3038 * No need for a memory barrier here, as the update
3039 * of the tail_page did it for this page.
3041 local_set(&cpu_buffer->commit_page->page->commit,
3042 rb_page_write(cpu_buffer->commit_page));
3043 rb_inc_page(&cpu_buffer->commit_page);
3044 /* add barrier to keep gcc from optimizing too much */
3047 while (rb_commit_index(cpu_buffer) !=
3048 rb_page_write(cpu_buffer->commit_page)) {
3050 /* Make sure the readers see the content of what is committed. */
3052 local_set(&cpu_buffer->commit_page->page->commit,
3053 rb_page_write(cpu_buffer->commit_page));
3054 RB_WARN_ON(cpu_buffer,
3055 local_read(&cpu_buffer->commit_page->page->commit) &
3060 /* again, keep gcc from optimizing */
3064 * If an interrupt came in just after the first while loop
3065 * and pushed the tail page forward, we will be left with
3066 * a dangling commit that will never go forward.
3068 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
3072 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
3074 unsigned long commits;
3076 if (RB_WARN_ON(cpu_buffer,
3077 !local_read(&cpu_buffer->committing)))
3081 commits = local_read(&cpu_buffer->commits);
3082 /* synchronize with interrupts */
3084 if (local_read(&cpu_buffer->committing) == 1)
3085 rb_set_commit_to_write(cpu_buffer);
3087 local_dec(&cpu_buffer->committing);
3089 /* synchronize with interrupts */
3093 * Need to account for interrupts coming in between the
3094 * updating of the commit page and the clearing of the
3095 * committing counter.
3097 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
3098 !local_read(&cpu_buffer->committing)) {
3099 local_inc(&cpu_buffer->committing);
3104 static inline void rb_event_discard(struct ring_buffer_event *event)
3106 if (extended_time(event))
3107 event = skip_time_extend(event);
3109 /* array[0] holds the actual length for the discarded event */
3110 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
3111 event->type_len = RINGBUF_TYPE_PADDING;
3112 /* time delta must be non zero */
3113 if (!event->time_delta)
3114 event->time_delta = 1;
3117 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer)
3119 local_inc(&cpu_buffer->entries);
3120 rb_end_commit(cpu_buffer);
3123 static __always_inline void
3124 rb_wakeups(struct trace_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
3126 if (buffer->irq_work.waiters_pending) {
3127 buffer->irq_work.waiters_pending = false;
3128 /* irq_work_queue() supplies it's own memory barriers */
3129 irq_work_queue(&buffer->irq_work.work);
3132 if (cpu_buffer->irq_work.waiters_pending) {
3133 cpu_buffer->irq_work.waiters_pending = false;
3134 /* irq_work_queue() supplies it's own memory barriers */
3135 irq_work_queue(&cpu_buffer->irq_work.work);
3138 if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
3141 if (cpu_buffer->reader_page == cpu_buffer->commit_page)
3144 if (!cpu_buffer->irq_work.full_waiters_pending)
3147 cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
3149 if (!full_hit(buffer, cpu_buffer->cpu, cpu_buffer->shortest_full))
3152 cpu_buffer->irq_work.wakeup_full = true;
3153 cpu_buffer->irq_work.full_waiters_pending = false;
3154 /* irq_work_queue() supplies it's own memory barriers */
3155 irq_work_queue(&cpu_buffer->irq_work.work);
3158 #ifdef CONFIG_RING_BUFFER_RECORD_RECURSION
3159 # define do_ring_buffer_record_recursion() \
3160 do_ftrace_record_recursion(_THIS_IP_, _RET_IP_)
3162 # define do_ring_buffer_record_recursion() do { } while (0)
3166 * The lock and unlock are done within a preempt disable section.
3167 * The current_context per_cpu variable can only be modified
3168 * by the current task between lock and unlock. But it can
3169 * be modified more than once via an interrupt. To pass this
3170 * information from the lock to the unlock without having to
3171 * access the 'in_interrupt()' functions again (which do show
3172 * a bit of overhead in something as critical as function tracing,
3173 * we use a bitmask trick.
3175 * bit 1 = NMI context
3176 * bit 2 = IRQ context
3177 * bit 3 = SoftIRQ context
3178 * bit 4 = normal context.
3180 * This works because this is the order of contexts that can
3181 * preempt other contexts. A SoftIRQ never preempts an IRQ
3184 * When the context is determined, the corresponding bit is
3185 * checked and set (if it was set, then a recursion of that context
3188 * On unlock, we need to clear this bit. To do so, just subtract
3189 * 1 from the current_context and AND it to itself.
3193 * 101 & 100 = 100 (clearing bit zero)
3196 * 1010 & 1001 = 1000 (clearing bit 1)
3198 * The least significant bit can be cleared this way, and it
3199 * just so happens that it is the same bit corresponding to
3200 * the current context.
3202 * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
3203 * is set when a recursion is detected at the current context, and if
3204 * the TRANSITION bit is already set, it will fail the recursion.
3205 * This is needed because there's a lag between the changing of
3206 * interrupt context and updating the preempt count. In this case,
3207 * a false positive will be found. To handle this, one extra recursion
3208 * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
3209 * bit is already set, then it is considered a recursion and the function
3210 * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
3212 * On the trace_recursive_unlock(), the TRANSITION bit will be the first
3213 * to be cleared. Even if it wasn't the context that set it. That is,
3214 * if an interrupt comes in while NORMAL bit is set and the ring buffer
3215 * is called before preempt_count() is updated, since the check will
3216 * be on the NORMAL bit, the TRANSITION bit will then be set. If an
3217 * NMI then comes in, it will set the NMI bit, but when the NMI code
3218 * does the trace_recursive_unlock() it will clear the TRANSITION bit
3219 * and leave the NMI bit set. But this is fine, because the interrupt
3220 * code that set the TRANSITION bit will then clear the NMI bit when it
3221 * calls trace_recursive_unlock(). If another NMI comes in, it will
3222 * set the TRANSITION bit and continue.
3224 * Note: The TRANSITION bit only handles a single transition between context.
3227 static __always_inline bool
3228 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
3230 unsigned int val = cpu_buffer->current_context;
3231 int bit = interrupt_context_level();
3233 bit = RB_CTX_NORMAL - bit;
3235 if (unlikely(val & (1 << (bit + cpu_buffer->nest)))) {
3237 * It is possible that this was called by transitioning
3238 * between interrupt context, and preempt_count() has not
3239 * been updated yet. In this case, use the TRANSITION bit.
3241 bit = RB_CTX_TRANSITION;
3242 if (val & (1 << (bit + cpu_buffer->nest))) {
3243 do_ring_buffer_record_recursion();
3248 val |= (1 << (bit + cpu_buffer->nest));
3249 cpu_buffer->current_context = val;
3254 static __always_inline void
3255 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
3257 cpu_buffer->current_context &=
3258 cpu_buffer->current_context - (1 << cpu_buffer->nest);
3261 /* The recursive locking above uses 5 bits */
3262 #define NESTED_BITS 5
3265 * ring_buffer_nest_start - Allow to trace while nested
3266 * @buffer: The ring buffer to modify
3268 * The ring buffer has a safety mechanism to prevent recursion.
3269 * But there may be a case where a trace needs to be done while
3270 * tracing something else. In this case, calling this function
3271 * will allow this function to nest within a currently active
3272 * ring_buffer_lock_reserve().
3274 * Call this function before calling another ring_buffer_lock_reserve() and
3275 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
3277 void ring_buffer_nest_start(struct trace_buffer *buffer)
3279 struct ring_buffer_per_cpu *cpu_buffer;
3282 /* Enabled by ring_buffer_nest_end() */
3283 preempt_disable_notrace();
3284 cpu = raw_smp_processor_id();
3285 cpu_buffer = buffer->buffers[cpu];
3286 /* This is the shift value for the above recursive locking */
3287 cpu_buffer->nest += NESTED_BITS;
3291 * ring_buffer_nest_end - Allow to trace while nested
3292 * @buffer: The ring buffer to modify
3294 * Must be called after ring_buffer_nest_start() and after the
3295 * ring_buffer_unlock_commit().
3297 void ring_buffer_nest_end(struct trace_buffer *buffer)
3299 struct ring_buffer_per_cpu *cpu_buffer;
3302 /* disabled by ring_buffer_nest_start() */
3303 cpu = raw_smp_processor_id();
3304 cpu_buffer = buffer->buffers[cpu];
3305 /* This is the shift value for the above recursive locking */
3306 cpu_buffer->nest -= NESTED_BITS;
3307 preempt_enable_notrace();
3311 * ring_buffer_unlock_commit - commit a reserved
3312 * @buffer: The buffer to commit to
3314 * This commits the data to the ring buffer, and releases any locks held.
3316 * Must be paired with ring_buffer_lock_reserve.
3318 int ring_buffer_unlock_commit(struct trace_buffer *buffer)
3320 struct ring_buffer_per_cpu *cpu_buffer;
3321 int cpu = raw_smp_processor_id();
3323 cpu_buffer = buffer->buffers[cpu];
3325 rb_commit(cpu_buffer);
3327 rb_wakeups(buffer, cpu_buffer);
3329 trace_recursive_unlock(cpu_buffer);
3331 preempt_enable_notrace();
3335 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
3337 /* Special value to validate all deltas on a page. */
3338 #define CHECK_FULL_PAGE 1L
3340 #ifdef CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS
3341 static void dump_buffer_page(struct buffer_data_page *bpage,
3342 struct rb_event_info *info,
3345 struct ring_buffer_event *event;
3349 ts = bpage->time_stamp;
3350 pr_warn(" [%lld] PAGE TIME STAMP\n", ts);
3352 for (e = 0; e < tail; e += rb_event_length(event)) {
3354 event = (struct ring_buffer_event *)(bpage->data + e);
3356 switch (event->type_len) {
3358 case RINGBUF_TYPE_TIME_EXTEND:
3359 delta = rb_event_time_stamp(event);
3361 pr_warn(" [%lld] delta:%lld TIME EXTEND\n", ts, delta);
3364 case RINGBUF_TYPE_TIME_STAMP:
3365 delta = rb_event_time_stamp(event);
3366 ts = rb_fix_abs_ts(delta, ts);
3367 pr_warn(" [%lld] absolute:%lld TIME STAMP\n", ts, delta);
3370 case RINGBUF_TYPE_PADDING:
3371 ts += event->time_delta;
3372 pr_warn(" [%lld] delta:%d PADDING\n", ts, event->time_delta);
3375 case RINGBUF_TYPE_DATA:
3376 ts += event->time_delta;
3377 pr_warn(" [%lld] delta:%d\n", ts, event->time_delta);
3386 static DEFINE_PER_CPU(atomic_t, checking);
3387 static atomic_t ts_dump;
3390 * Check if the current event time stamp matches the deltas on
3393 static void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3394 struct rb_event_info *info,
3397 struct ring_buffer_event *event;
3398 struct buffer_data_page *bpage;
3403 bpage = info->tail_page->page;
3405 if (tail == CHECK_FULL_PAGE) {
3407 tail = local_read(&bpage->commit);
3408 } else if (info->add_timestamp &
3409 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE)) {
3410 /* Ignore events with absolute time stamps */
3415 * Do not check the first event (skip possible extends too).
3416 * Also do not check if previous events have not been committed.
3418 if (tail <= 8 || tail > local_read(&bpage->commit))
3422 * If this interrupted another event,
3424 if (atomic_inc_return(this_cpu_ptr(&checking)) != 1)
3427 ts = bpage->time_stamp;
3429 for (e = 0; e < tail; e += rb_event_length(event)) {
3431 event = (struct ring_buffer_event *)(bpage->data + e);
3433 switch (event->type_len) {
3435 case RINGBUF_TYPE_TIME_EXTEND:
3436 delta = rb_event_time_stamp(event);
3440 case RINGBUF_TYPE_TIME_STAMP:
3441 delta = rb_event_time_stamp(event);
3442 ts = rb_fix_abs_ts(delta, ts);
3445 case RINGBUF_TYPE_PADDING:
3446 if (event->time_delta == 1)
3449 case RINGBUF_TYPE_DATA:
3450 ts += event->time_delta;
3454 RB_WARN_ON(cpu_buffer, 1);
3457 if ((full && ts > info->ts) ||
3458 (!full && ts + info->delta != info->ts)) {
3459 /* If another report is happening, ignore this one */
3460 if (atomic_inc_return(&ts_dump) != 1) {
3461 atomic_dec(&ts_dump);
3464 atomic_inc(&cpu_buffer->record_disabled);
3465 /* There's some cases in boot up that this can happen */
3466 WARN_ON_ONCE(system_state != SYSTEM_BOOTING);
3467 pr_warn("[CPU: %d]TIME DOES NOT MATCH expected:%lld actual:%lld delta:%lld before:%lld after:%lld%s\n",
3469 ts + info->delta, info->ts, info->delta,
3470 info->before, info->after,
3471 full ? " (full)" : "");
3472 dump_buffer_page(bpage, info, tail);
3473 atomic_dec(&ts_dump);
3474 /* Do not re-enable checking */
3478 atomic_dec(this_cpu_ptr(&checking));
3481 static inline void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3482 struct rb_event_info *info,
3486 #endif /* CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS */
3488 static struct ring_buffer_event *
3489 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
3490 struct rb_event_info *info)
3492 struct ring_buffer_event *event;
3493 struct buffer_page *tail_page;
3494 unsigned long tail, write, w;
3498 /* Don't let the compiler play games with cpu_buffer->tail_page */
3499 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
3501 /*A*/ w = local_read(&tail_page->write) & RB_WRITE_MASK;
3503 b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3504 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3506 info->ts = rb_time_stamp(cpu_buffer->buffer);
3508 if ((info->add_timestamp & RB_ADD_STAMP_ABSOLUTE)) {
3509 info->delta = info->ts;
3512 * If interrupting an event time update, we may need an
3513 * absolute timestamp.
3514 * Don't bother if this is the start of a new page (w == 0).
3517 /* Use the sub-buffer timestamp */
3519 } else if (unlikely(!a_ok || !b_ok || info->before != info->after)) {
3520 info->add_timestamp |= RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND;
3521 info->length += RB_LEN_TIME_EXTEND;
3523 info->delta = info->ts - info->after;
3524 if (unlikely(test_time_stamp(info->delta))) {
3525 info->add_timestamp |= RB_ADD_STAMP_EXTEND;
3526 info->length += RB_LEN_TIME_EXTEND;
3531 /*B*/ rb_time_set(&cpu_buffer->before_stamp, info->ts);
3533 /*C*/ write = local_add_return(info->length, &tail_page->write);
3535 /* set write to only the index of the write */
3536 write &= RB_WRITE_MASK;
3538 tail = write - info->length;
3540 /* See if we shot pass the end of this buffer page */
3541 if (unlikely(write > BUF_PAGE_SIZE)) {
3542 check_buffer(cpu_buffer, info, CHECK_FULL_PAGE);
3543 return rb_move_tail(cpu_buffer, tail, info);
3546 if (likely(tail == w)) {
3547 /* Nothing interrupted us between A and C */
3548 /*D*/ rb_time_set(&cpu_buffer->write_stamp, info->ts);
3550 * If something came in between C and D, the write stamp
3551 * may now not be in sync. But that's fine as the before_stamp
3552 * will be different and then next event will just be forced
3553 * to use an absolute timestamp.
3555 if (likely(!(info->add_timestamp &
3556 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3557 /* This did not interrupt any time update */
3558 info->delta = info->ts - info->after;
3560 /* Just use full timestamp for interrupting event */
3561 info->delta = info->ts;
3562 check_buffer(cpu_buffer, info, tail);
3565 /* SLOW PATH - Interrupted between A and C */
3567 /* Save the old before_stamp */
3568 a_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3569 RB_WARN_ON(cpu_buffer, !a_ok);
3572 * Read a new timestamp and update the before_stamp to make
3573 * the next event after this one force using an absolute
3574 * timestamp. This is in case an interrupt were to come in
3577 ts = rb_time_stamp(cpu_buffer->buffer);
3578 rb_time_set(&cpu_buffer->before_stamp, ts);
3581 /*E*/ a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3582 /* Was interrupted before here, write_stamp must be valid */
3583 RB_WARN_ON(cpu_buffer, !a_ok);
3585 /*F*/ if (write == (local_read(&tail_page->write) & RB_WRITE_MASK) &&
3586 info->after == info->before && info->after < ts) {
3588 * Nothing came after this event between C and F, it is
3589 * safe to use info->after for the delta as it
3590 * matched info->before and is still valid.
3592 info->delta = ts - info->after;
3595 * Interrupted between C and F:
3596 * Lost the previous events time stamp. Just set the
3597 * delta to zero, and this will be the same time as
3598 * the event this event interrupted. And the events that
3599 * came after this will still be correct (as they would
3600 * have built their delta on the previous event.
3605 info->add_timestamp &= ~RB_ADD_STAMP_FORCE;
3609 * If this is the first commit on the page, then it has the same
3610 * timestamp as the page itself.
3612 if (unlikely(!tail && !(info->add_timestamp &
3613 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3616 /* We reserved something on the buffer */
3618 event = __rb_page_index(tail_page, tail);
3619 rb_update_event(cpu_buffer, event, info);
3621 local_inc(&tail_page->entries);
3624 * If this is the first commit on the page, then update
3627 if (unlikely(!tail))
3628 tail_page->page->time_stamp = info->ts;
3630 /* account for these added bytes */
3631 local_add(info->length, &cpu_buffer->entries_bytes);
3636 static __always_inline struct ring_buffer_event *
3637 rb_reserve_next_event(struct trace_buffer *buffer,
3638 struct ring_buffer_per_cpu *cpu_buffer,
3639 unsigned long length)
3641 struct ring_buffer_event *event;
3642 struct rb_event_info info;
3646 /* ring buffer does cmpxchg, make sure it is safe in NMI context */
3647 if (!IS_ENABLED(CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG) &&
3648 (unlikely(in_nmi()))) {
3652 rb_start_commit(cpu_buffer);
3653 /* The commit page can not change after this */
3655 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3657 * Due to the ability to swap a cpu buffer from a buffer
3658 * it is possible it was swapped before we committed.
3659 * (committing stops a swap). We check for it here and
3660 * if it happened, we have to fail the write.
3663 if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
3664 local_dec(&cpu_buffer->committing);
3665 local_dec(&cpu_buffer->commits);
3670 info.length = rb_calculate_event_length(length);
3672 if (ring_buffer_time_stamp_abs(cpu_buffer->buffer)) {
3673 add_ts_default = RB_ADD_STAMP_ABSOLUTE;
3674 info.length += RB_LEN_TIME_EXTEND;
3675 if (info.length > BUF_MAX_DATA_SIZE)
3678 add_ts_default = RB_ADD_STAMP_NONE;
3682 info.add_timestamp = add_ts_default;
3686 * We allow for interrupts to reenter here and do a trace.
3687 * If one does, it will cause this original code to loop
3688 * back here. Even with heavy interrupts happening, this
3689 * should only happen a few times in a row. If this happens
3690 * 1000 times in a row, there must be either an interrupt
3691 * storm or we have something buggy.
3694 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
3697 event = __rb_reserve_next(cpu_buffer, &info);
3699 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
3700 if (info.add_timestamp & (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND))
3701 info.length -= RB_LEN_TIME_EXTEND;
3708 rb_end_commit(cpu_buffer);
3713 * ring_buffer_lock_reserve - reserve a part of the buffer
3714 * @buffer: the ring buffer to reserve from
3715 * @length: the length of the data to reserve (excluding event header)
3717 * Returns a reserved event on the ring buffer to copy directly to.
3718 * The user of this interface will need to get the body to write into
3719 * and can use the ring_buffer_event_data() interface.
3721 * The length is the length of the data needed, not the event length
3722 * which also includes the event header.
3724 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
3725 * If NULL is returned, then nothing has been allocated or locked.
3727 struct ring_buffer_event *
3728 ring_buffer_lock_reserve(struct trace_buffer *buffer, unsigned long length)
3730 struct ring_buffer_per_cpu *cpu_buffer;
3731 struct ring_buffer_event *event;
3734 /* If we are tracing schedule, we don't want to recurse */
3735 preempt_disable_notrace();
3737 if (unlikely(atomic_read(&buffer->record_disabled)))
3740 cpu = raw_smp_processor_id();
3742 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
3745 cpu_buffer = buffer->buffers[cpu];
3747 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
3750 if (unlikely(length > BUF_MAX_DATA_SIZE))
3753 if (unlikely(trace_recursive_lock(cpu_buffer)))
3756 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3763 trace_recursive_unlock(cpu_buffer);
3765 preempt_enable_notrace();
3768 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3771 * Decrement the entries to the page that an event is on.
3772 * The event does not even need to exist, only the pointer
3773 * to the page it is on. This may only be called before the commit
3777 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3778 struct ring_buffer_event *event)
3780 unsigned long addr = (unsigned long)event;
3781 struct buffer_page *bpage = cpu_buffer->commit_page;
3782 struct buffer_page *start;
3786 /* Do the likely case first */
3787 if (likely(bpage->page == (void *)addr)) {
3788 local_dec(&bpage->entries);
3793 * Because the commit page may be on the reader page we
3794 * start with the next page and check the end loop there.
3796 rb_inc_page(&bpage);
3799 if (bpage->page == (void *)addr) {
3800 local_dec(&bpage->entries);
3803 rb_inc_page(&bpage);
3804 } while (bpage != start);
3806 /* commit not part of this buffer?? */
3807 RB_WARN_ON(cpu_buffer, 1);
3811 * ring_buffer_discard_commit - discard an event that has not been committed
3812 * @buffer: the ring buffer
3813 * @event: non committed event to discard
3815 * Sometimes an event that is in the ring buffer needs to be ignored.
3816 * This function lets the user discard an event in the ring buffer
3817 * and then that event will not be read later.
3819 * This function only works if it is called before the item has been
3820 * committed. It will try to free the event from the ring buffer
3821 * if another event has not been added behind it.
3823 * If another event has been added behind it, it will set the event
3824 * up as discarded, and perform the commit.
3826 * If this function is called, do not call ring_buffer_unlock_commit on
3829 void ring_buffer_discard_commit(struct trace_buffer *buffer,
3830 struct ring_buffer_event *event)
3832 struct ring_buffer_per_cpu *cpu_buffer;
3835 /* The event is discarded regardless */
3836 rb_event_discard(event);
3838 cpu = smp_processor_id();
3839 cpu_buffer = buffer->buffers[cpu];
3842 * This must only be called if the event has not been
3843 * committed yet. Thus we can assume that preemption
3844 * is still disabled.
3846 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3848 rb_decrement_entry(cpu_buffer, event);
3849 if (rb_try_to_discard(cpu_buffer, event))
3853 rb_end_commit(cpu_buffer);
3855 trace_recursive_unlock(cpu_buffer);
3857 preempt_enable_notrace();
3860 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3863 * ring_buffer_write - write data to the buffer without reserving
3864 * @buffer: The ring buffer to write to.
3865 * @length: The length of the data being written (excluding the event header)
3866 * @data: The data to write to the buffer.
3868 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3869 * one function. If you already have the data to write to the buffer, it
3870 * may be easier to simply call this function.
3872 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3873 * and not the length of the event which would hold the header.
3875 int ring_buffer_write(struct trace_buffer *buffer,
3876 unsigned long length,
3879 struct ring_buffer_per_cpu *cpu_buffer;
3880 struct ring_buffer_event *event;
3885 preempt_disable_notrace();
3887 if (atomic_read(&buffer->record_disabled))
3890 cpu = raw_smp_processor_id();
3892 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3895 cpu_buffer = buffer->buffers[cpu];
3897 if (atomic_read(&cpu_buffer->record_disabled))
3900 if (length > BUF_MAX_DATA_SIZE)
3903 if (unlikely(trace_recursive_lock(cpu_buffer)))
3906 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3910 body = rb_event_data(event);
3912 memcpy(body, data, length);
3914 rb_commit(cpu_buffer);
3916 rb_wakeups(buffer, cpu_buffer);
3921 trace_recursive_unlock(cpu_buffer);
3924 preempt_enable_notrace();
3928 EXPORT_SYMBOL_GPL(ring_buffer_write);
3930 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3932 struct buffer_page *reader = cpu_buffer->reader_page;
3933 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3934 struct buffer_page *commit = cpu_buffer->commit_page;
3936 /* In case of error, head will be NULL */
3937 if (unlikely(!head))
3940 /* Reader should exhaust content in reader page */
3941 if (reader->read != rb_page_commit(reader))
3945 * If writers are committing on the reader page, knowing all
3946 * committed content has been read, the ring buffer is empty.
3948 if (commit == reader)
3952 * If writers are committing on a page other than reader page
3953 * and head page, there should always be content to read.
3959 * Writers are committing on the head page, we just need
3960 * to care about there're committed data, and the reader will
3961 * swap reader page with head page when it is to read data.
3963 return rb_page_commit(commit) == 0;
3967 * ring_buffer_record_disable - stop all writes into the buffer
3968 * @buffer: The ring buffer to stop writes to.
3970 * This prevents all writes to the buffer. Any attempt to write
3971 * to the buffer after this will fail and return NULL.
3973 * The caller should call synchronize_rcu() after this.
3975 void ring_buffer_record_disable(struct trace_buffer *buffer)
3977 atomic_inc(&buffer->record_disabled);
3979 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
3982 * ring_buffer_record_enable - enable writes to the buffer
3983 * @buffer: The ring buffer to enable writes
3985 * Note, multiple disables will need the same number of enables
3986 * to truly enable the writing (much like preempt_disable).
3988 void ring_buffer_record_enable(struct trace_buffer *buffer)
3990 atomic_dec(&buffer->record_disabled);
3992 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
3995 * ring_buffer_record_off - stop all writes into the buffer
3996 * @buffer: The ring buffer to stop writes to.
3998 * This prevents all writes to the buffer. Any attempt to write
3999 * to the buffer after this will fail and return NULL.
4001 * This is different than ring_buffer_record_disable() as
4002 * it works like an on/off switch, where as the disable() version
4003 * must be paired with a enable().
4005 void ring_buffer_record_off(struct trace_buffer *buffer)
4008 unsigned int new_rd;
4010 rd = atomic_read(&buffer->record_disabled);
4012 new_rd = rd | RB_BUFFER_OFF;
4013 } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
4015 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
4018 * ring_buffer_record_on - restart writes into the buffer
4019 * @buffer: The ring buffer to start writes to.
4021 * This enables all writes to the buffer that was disabled by
4022 * ring_buffer_record_off().
4024 * This is different than ring_buffer_record_enable() as
4025 * it works like an on/off switch, where as the enable() version
4026 * must be paired with a disable().
4028 void ring_buffer_record_on(struct trace_buffer *buffer)
4031 unsigned int new_rd;
4033 rd = atomic_read(&buffer->record_disabled);
4035 new_rd = rd & ~RB_BUFFER_OFF;
4036 } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
4038 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
4041 * ring_buffer_record_is_on - return true if the ring buffer can write
4042 * @buffer: The ring buffer to see if write is enabled
4044 * Returns true if the ring buffer is in a state that it accepts writes.
4046 bool ring_buffer_record_is_on(struct trace_buffer *buffer)
4048 return !atomic_read(&buffer->record_disabled);
4052 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
4053 * @buffer: The ring buffer to see if write is set enabled
4055 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
4056 * Note that this does NOT mean it is in a writable state.
4058 * It may return true when the ring buffer has been disabled by
4059 * ring_buffer_record_disable(), as that is a temporary disabling of
4062 bool ring_buffer_record_is_set_on(struct trace_buffer *buffer)
4064 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
4068 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
4069 * @buffer: The ring buffer to stop writes to.
4070 * @cpu: The CPU buffer to stop
4072 * This prevents all writes to the buffer. Any attempt to write
4073 * to the buffer after this will fail and return NULL.
4075 * The caller should call synchronize_rcu() after this.
4077 void ring_buffer_record_disable_cpu(struct trace_buffer *buffer, int cpu)
4079 struct ring_buffer_per_cpu *cpu_buffer;
4081 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4084 cpu_buffer = buffer->buffers[cpu];
4085 atomic_inc(&cpu_buffer->record_disabled);
4087 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
4090 * ring_buffer_record_enable_cpu - enable writes to the buffer
4091 * @buffer: The ring buffer to enable writes
4092 * @cpu: The CPU to enable.
4094 * Note, multiple disables will need the same number of enables
4095 * to truly enable the writing (much like preempt_disable).
4097 void ring_buffer_record_enable_cpu(struct trace_buffer *buffer, int cpu)
4099 struct ring_buffer_per_cpu *cpu_buffer;
4101 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4104 cpu_buffer = buffer->buffers[cpu];
4105 atomic_dec(&cpu_buffer->record_disabled);
4107 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
4110 * The total entries in the ring buffer is the running counter
4111 * of entries entered into the ring buffer, minus the sum of
4112 * the entries read from the ring buffer and the number of
4113 * entries that were overwritten.
4115 static inline unsigned long
4116 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
4118 return local_read(&cpu_buffer->entries) -
4119 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
4123 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
4124 * @buffer: The ring buffer
4125 * @cpu: The per CPU buffer to read from.
4127 u64 ring_buffer_oldest_event_ts(struct trace_buffer *buffer, int cpu)
4129 unsigned long flags;
4130 struct ring_buffer_per_cpu *cpu_buffer;
4131 struct buffer_page *bpage;
4134 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4137 cpu_buffer = buffer->buffers[cpu];
4138 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4140 * if the tail is on reader_page, oldest time stamp is on the reader
4143 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
4144 bpage = cpu_buffer->reader_page;
4146 bpage = rb_set_head_page(cpu_buffer);
4148 ret = bpage->page->time_stamp;
4149 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4153 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
4156 * ring_buffer_bytes_cpu - get the number of bytes unconsumed in a cpu buffer
4157 * @buffer: The ring buffer
4158 * @cpu: The per CPU buffer to read from.
4160 unsigned long ring_buffer_bytes_cpu(struct trace_buffer *buffer, int cpu)
4162 struct ring_buffer_per_cpu *cpu_buffer;
4165 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4168 cpu_buffer = buffer->buffers[cpu];
4169 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
4173 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
4176 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
4177 * @buffer: The ring buffer
4178 * @cpu: The per CPU buffer to get the entries from.
4180 unsigned long ring_buffer_entries_cpu(struct trace_buffer *buffer, int cpu)
4182 struct ring_buffer_per_cpu *cpu_buffer;
4184 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4187 cpu_buffer = buffer->buffers[cpu];
4189 return rb_num_of_entries(cpu_buffer);
4191 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
4194 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
4195 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
4196 * @buffer: The ring buffer
4197 * @cpu: The per CPU buffer to get the number of overruns from
4199 unsigned long ring_buffer_overrun_cpu(struct trace_buffer *buffer, int cpu)
4201 struct ring_buffer_per_cpu *cpu_buffer;
4204 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4207 cpu_buffer = buffer->buffers[cpu];
4208 ret = local_read(&cpu_buffer->overrun);
4212 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
4215 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
4216 * commits failing due to the buffer wrapping around while there are uncommitted
4217 * events, such as during an interrupt storm.
4218 * @buffer: The ring buffer
4219 * @cpu: The per CPU buffer to get the number of overruns from
4222 ring_buffer_commit_overrun_cpu(struct trace_buffer *buffer, int cpu)
4224 struct ring_buffer_per_cpu *cpu_buffer;
4227 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4230 cpu_buffer = buffer->buffers[cpu];
4231 ret = local_read(&cpu_buffer->commit_overrun);
4235 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
4238 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
4239 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
4240 * @buffer: The ring buffer
4241 * @cpu: The per CPU buffer to get the number of overruns from
4244 ring_buffer_dropped_events_cpu(struct trace_buffer *buffer, int cpu)
4246 struct ring_buffer_per_cpu *cpu_buffer;
4249 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4252 cpu_buffer = buffer->buffers[cpu];
4253 ret = local_read(&cpu_buffer->dropped_events);
4257 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
4260 * ring_buffer_read_events_cpu - get the number of events successfully read
4261 * @buffer: The ring buffer
4262 * @cpu: The per CPU buffer to get the number of events read
4265 ring_buffer_read_events_cpu(struct trace_buffer *buffer, int cpu)
4267 struct ring_buffer_per_cpu *cpu_buffer;
4269 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4272 cpu_buffer = buffer->buffers[cpu];
4273 return cpu_buffer->read;
4275 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
4278 * ring_buffer_entries - get the number of entries in a buffer
4279 * @buffer: The ring buffer
4281 * Returns the total number of entries in the ring buffer
4284 unsigned long ring_buffer_entries(struct trace_buffer *buffer)
4286 struct ring_buffer_per_cpu *cpu_buffer;
4287 unsigned long entries = 0;
4290 /* if you care about this being correct, lock the buffer */
4291 for_each_buffer_cpu(buffer, cpu) {
4292 cpu_buffer = buffer->buffers[cpu];
4293 entries += rb_num_of_entries(cpu_buffer);
4298 EXPORT_SYMBOL_GPL(ring_buffer_entries);
4301 * ring_buffer_overruns - get the number of overruns in buffer
4302 * @buffer: The ring buffer
4304 * Returns the total number of overruns in the ring buffer
4307 unsigned long ring_buffer_overruns(struct trace_buffer *buffer)
4309 struct ring_buffer_per_cpu *cpu_buffer;
4310 unsigned long overruns = 0;
4313 /* if you care about this being correct, lock the buffer */
4314 for_each_buffer_cpu(buffer, cpu) {
4315 cpu_buffer = buffer->buffers[cpu];
4316 overruns += local_read(&cpu_buffer->overrun);
4321 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
4323 static void rb_iter_reset(struct ring_buffer_iter *iter)
4325 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4327 /* Iterator usage is expected to have record disabled */
4328 iter->head_page = cpu_buffer->reader_page;
4329 iter->head = cpu_buffer->reader_page->read;
4330 iter->next_event = iter->head;
4332 iter->cache_reader_page = iter->head_page;
4333 iter->cache_read = cpu_buffer->read;
4334 iter->cache_pages_removed = cpu_buffer->pages_removed;
4337 iter->read_stamp = cpu_buffer->read_stamp;
4338 iter->page_stamp = cpu_buffer->reader_page->page->time_stamp;
4340 iter->read_stamp = iter->head_page->page->time_stamp;
4341 iter->page_stamp = iter->read_stamp;
4346 * ring_buffer_iter_reset - reset an iterator
4347 * @iter: The iterator to reset
4349 * Resets the iterator, so that it will start from the beginning
4352 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
4354 struct ring_buffer_per_cpu *cpu_buffer;
4355 unsigned long flags;
4360 cpu_buffer = iter->cpu_buffer;
4362 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4363 rb_iter_reset(iter);
4364 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4366 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
4369 * ring_buffer_iter_empty - check if an iterator has no more to read
4370 * @iter: The iterator to check
4372 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
4374 struct ring_buffer_per_cpu *cpu_buffer;
4375 struct buffer_page *reader;
4376 struct buffer_page *head_page;
4377 struct buffer_page *commit_page;
4378 struct buffer_page *curr_commit_page;
4383 cpu_buffer = iter->cpu_buffer;
4384 reader = cpu_buffer->reader_page;
4385 head_page = cpu_buffer->head_page;
4386 commit_page = cpu_buffer->commit_page;
4387 commit_ts = commit_page->page->time_stamp;
4390 * When the writer goes across pages, it issues a cmpxchg which
4391 * is a mb(), which will synchronize with the rmb here.
4392 * (see rb_tail_page_update())
4395 commit = rb_page_commit(commit_page);
4396 /* We want to make sure that the commit page doesn't change */
4399 /* Make sure commit page didn't change */
4400 curr_commit_page = READ_ONCE(cpu_buffer->commit_page);
4401 curr_commit_ts = READ_ONCE(curr_commit_page->page->time_stamp);
4403 /* If the commit page changed, then there's more data */
4404 if (curr_commit_page != commit_page ||
4405 curr_commit_ts != commit_ts)
4408 /* Still racy, as it may return a false positive, but that's OK */
4409 return ((iter->head_page == commit_page && iter->head >= commit) ||
4410 (iter->head_page == reader && commit_page == head_page &&
4411 head_page->read == commit &&
4412 iter->head == rb_page_commit(cpu_buffer->reader_page)));
4414 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
4417 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
4418 struct ring_buffer_event *event)
4422 switch (event->type_len) {
4423 case RINGBUF_TYPE_PADDING:
4426 case RINGBUF_TYPE_TIME_EXTEND:
4427 delta = rb_event_time_stamp(event);
4428 cpu_buffer->read_stamp += delta;
4431 case RINGBUF_TYPE_TIME_STAMP:
4432 delta = rb_event_time_stamp(event);
4433 delta = rb_fix_abs_ts(delta, cpu_buffer->read_stamp);
4434 cpu_buffer->read_stamp = delta;
4437 case RINGBUF_TYPE_DATA:
4438 cpu_buffer->read_stamp += event->time_delta;
4442 RB_WARN_ON(cpu_buffer, 1);
4447 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
4448 struct ring_buffer_event *event)
4452 switch (event->type_len) {
4453 case RINGBUF_TYPE_PADDING:
4456 case RINGBUF_TYPE_TIME_EXTEND:
4457 delta = rb_event_time_stamp(event);
4458 iter->read_stamp += delta;
4461 case RINGBUF_TYPE_TIME_STAMP:
4462 delta = rb_event_time_stamp(event);
4463 delta = rb_fix_abs_ts(delta, iter->read_stamp);
4464 iter->read_stamp = delta;
4467 case RINGBUF_TYPE_DATA:
4468 iter->read_stamp += event->time_delta;
4472 RB_WARN_ON(iter->cpu_buffer, 1);
4476 static struct buffer_page *
4477 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
4479 struct buffer_page *reader = NULL;
4480 unsigned long overwrite;
4481 unsigned long flags;
4485 local_irq_save(flags);
4486 arch_spin_lock(&cpu_buffer->lock);
4490 * This should normally only loop twice. But because the
4491 * start of the reader inserts an empty page, it causes
4492 * a case where we will loop three times. There should be no
4493 * reason to loop four times (that I know of).
4495 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
4500 reader = cpu_buffer->reader_page;
4502 /* If there's more to read, return this page */
4503 if (cpu_buffer->reader_page->read < rb_page_size(reader))
4506 /* Never should we have an index greater than the size */
4507 if (RB_WARN_ON(cpu_buffer,
4508 cpu_buffer->reader_page->read > rb_page_size(reader)))
4511 /* check if we caught up to the tail */
4513 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
4516 /* Don't bother swapping if the ring buffer is empty */
4517 if (rb_num_of_entries(cpu_buffer) == 0)
4521 * Reset the reader page to size zero.
4523 local_set(&cpu_buffer->reader_page->write, 0);
4524 local_set(&cpu_buffer->reader_page->entries, 0);
4525 local_set(&cpu_buffer->reader_page->page->commit, 0);
4526 cpu_buffer->reader_page->real_end = 0;
4530 * Splice the empty reader page into the list around the head.
4532 reader = rb_set_head_page(cpu_buffer);
4535 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
4536 cpu_buffer->reader_page->list.prev = reader->list.prev;
4539 * cpu_buffer->pages just needs to point to the buffer, it
4540 * has no specific buffer page to point to. Lets move it out
4541 * of our way so we don't accidentally swap it.
4543 cpu_buffer->pages = reader->list.prev;
4545 /* The reader page will be pointing to the new head */
4546 rb_set_list_to_head(&cpu_buffer->reader_page->list);
4549 * We want to make sure we read the overruns after we set up our
4550 * pointers to the next object. The writer side does a
4551 * cmpxchg to cross pages which acts as the mb on the writer
4552 * side. Note, the reader will constantly fail the swap
4553 * while the writer is updating the pointers, so this
4554 * guarantees that the overwrite recorded here is the one we
4555 * want to compare with the last_overrun.
4558 overwrite = local_read(&(cpu_buffer->overrun));
4561 * Here's the tricky part.
4563 * We need to move the pointer past the header page.
4564 * But we can only do that if a writer is not currently
4565 * moving it. The page before the header page has the
4566 * flag bit '1' set if it is pointing to the page we want.
4567 * but if the writer is in the process of moving it
4568 * than it will be '2' or already moved '0'.
4571 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
4574 * If we did not convert it, then we must try again.
4580 * Yay! We succeeded in replacing the page.
4582 * Now make the new head point back to the reader page.
4584 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
4585 rb_inc_page(&cpu_buffer->head_page);
4587 local_inc(&cpu_buffer->pages_read);
4589 /* Finally update the reader page to the new head */
4590 cpu_buffer->reader_page = reader;
4591 cpu_buffer->reader_page->read = 0;
4593 if (overwrite != cpu_buffer->last_overrun) {
4594 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
4595 cpu_buffer->last_overrun = overwrite;
4601 /* Update the read_stamp on the first event */
4602 if (reader && reader->read == 0)
4603 cpu_buffer->read_stamp = reader->page->time_stamp;
4605 arch_spin_unlock(&cpu_buffer->lock);
4606 local_irq_restore(flags);
4609 * The writer has preempt disable, wait for it. But not forever
4610 * Although, 1 second is pretty much "forever"
4612 #define USECS_WAIT 1000000
4613 for (nr_loops = 0; nr_loops < USECS_WAIT; nr_loops++) {
4614 /* If the write is past the end of page, a writer is still updating it */
4615 if (likely(!reader || rb_page_write(reader) <= BUF_PAGE_SIZE))
4620 /* Get the latest version of the reader write value */
4624 /* The writer is not moving forward? Something is wrong */
4625 if (RB_WARN_ON(cpu_buffer, nr_loops == USECS_WAIT))
4629 * Make sure we see any padding after the write update
4630 * (see rb_reset_tail()).
4632 * In addition, a writer may be writing on the reader page
4633 * if the page has not been fully filled, so the read barrier
4634 * is also needed to make sure we see the content of what is
4635 * committed by the writer (see rb_set_commit_to_write()).
4643 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
4645 struct ring_buffer_event *event;
4646 struct buffer_page *reader;
4649 reader = rb_get_reader_page(cpu_buffer);
4651 /* This function should not be called when buffer is empty */
4652 if (RB_WARN_ON(cpu_buffer, !reader))
4655 event = rb_reader_event(cpu_buffer);
4657 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
4660 rb_update_read_stamp(cpu_buffer, event);
4662 length = rb_event_length(event);
4663 cpu_buffer->reader_page->read += length;
4664 cpu_buffer->read_bytes += length;
4667 static void rb_advance_iter(struct ring_buffer_iter *iter)
4669 struct ring_buffer_per_cpu *cpu_buffer;
4671 cpu_buffer = iter->cpu_buffer;
4673 /* If head == next_event then we need to jump to the next event */
4674 if (iter->head == iter->next_event) {
4675 /* If the event gets overwritten again, there's nothing to do */
4676 if (rb_iter_head_event(iter) == NULL)
4680 iter->head = iter->next_event;
4683 * Check if we are at the end of the buffer.
4685 if (iter->next_event >= rb_page_size(iter->head_page)) {
4686 /* discarded commits can make the page empty */
4687 if (iter->head_page == cpu_buffer->commit_page)
4693 rb_update_iter_read_stamp(iter, iter->event);
4696 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
4698 return cpu_buffer->lost_events;
4701 static struct ring_buffer_event *
4702 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
4703 unsigned long *lost_events)
4705 struct ring_buffer_event *event;
4706 struct buffer_page *reader;
4713 * We repeat when a time extend is encountered.
4714 * Since the time extend is always attached to a data event,
4715 * we should never loop more than once.
4716 * (We never hit the following condition more than twice).
4718 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
4721 reader = rb_get_reader_page(cpu_buffer);
4725 event = rb_reader_event(cpu_buffer);
4727 switch (event->type_len) {
4728 case RINGBUF_TYPE_PADDING:
4729 if (rb_null_event(event))
4730 RB_WARN_ON(cpu_buffer, 1);
4732 * Because the writer could be discarding every
4733 * event it creates (which would probably be bad)
4734 * if we were to go back to "again" then we may never
4735 * catch up, and will trigger the warn on, or lock
4736 * the box. Return the padding, and we will release
4737 * the current locks, and try again.
4741 case RINGBUF_TYPE_TIME_EXTEND:
4742 /* Internal data, OK to advance */
4743 rb_advance_reader(cpu_buffer);
4746 case RINGBUF_TYPE_TIME_STAMP:
4748 *ts = rb_event_time_stamp(event);
4749 *ts = rb_fix_abs_ts(*ts, reader->page->time_stamp);
4750 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4751 cpu_buffer->cpu, ts);
4753 /* Internal data, OK to advance */
4754 rb_advance_reader(cpu_buffer);
4757 case RINGBUF_TYPE_DATA:
4759 *ts = cpu_buffer->read_stamp + event->time_delta;
4760 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4761 cpu_buffer->cpu, ts);
4764 *lost_events = rb_lost_events(cpu_buffer);
4768 RB_WARN_ON(cpu_buffer, 1);
4773 EXPORT_SYMBOL_GPL(ring_buffer_peek);
4775 static struct ring_buffer_event *
4776 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4778 struct trace_buffer *buffer;
4779 struct ring_buffer_per_cpu *cpu_buffer;
4780 struct ring_buffer_event *event;
4786 cpu_buffer = iter->cpu_buffer;
4787 buffer = cpu_buffer->buffer;
4790 * Check if someone performed a consuming read to the buffer
4791 * or removed some pages from the buffer. In these cases,
4792 * iterator was invalidated and we need to reset it.
4794 if (unlikely(iter->cache_read != cpu_buffer->read ||
4795 iter->cache_reader_page != cpu_buffer->reader_page ||
4796 iter->cache_pages_removed != cpu_buffer->pages_removed))
4797 rb_iter_reset(iter);
4800 if (ring_buffer_iter_empty(iter))
4804 * As the writer can mess with what the iterator is trying
4805 * to read, just give up if we fail to get an event after
4806 * three tries. The iterator is not as reliable when reading
4807 * the ring buffer with an active write as the consumer is.
4808 * Do not warn if the three failures is reached.
4813 if (rb_per_cpu_empty(cpu_buffer))
4816 if (iter->head >= rb_page_size(iter->head_page)) {
4821 event = rb_iter_head_event(iter);
4825 switch (event->type_len) {
4826 case RINGBUF_TYPE_PADDING:
4827 if (rb_null_event(event)) {
4831 rb_advance_iter(iter);
4834 case RINGBUF_TYPE_TIME_EXTEND:
4835 /* Internal data, OK to advance */
4836 rb_advance_iter(iter);
4839 case RINGBUF_TYPE_TIME_STAMP:
4841 *ts = rb_event_time_stamp(event);
4842 *ts = rb_fix_abs_ts(*ts, iter->head_page->page->time_stamp);
4843 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4844 cpu_buffer->cpu, ts);
4846 /* Internal data, OK to advance */
4847 rb_advance_iter(iter);
4850 case RINGBUF_TYPE_DATA:
4852 *ts = iter->read_stamp + event->time_delta;
4853 ring_buffer_normalize_time_stamp(buffer,
4854 cpu_buffer->cpu, ts);
4859 RB_WARN_ON(cpu_buffer, 1);
4864 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4866 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4868 if (likely(!in_nmi())) {
4869 raw_spin_lock(&cpu_buffer->reader_lock);
4874 * If an NMI die dumps out the content of the ring buffer
4875 * trylock must be used to prevent a deadlock if the NMI
4876 * preempted a task that holds the ring buffer locks. If
4877 * we get the lock then all is fine, if not, then continue
4878 * to do the read, but this can corrupt the ring buffer,
4879 * so it must be permanently disabled from future writes.
4880 * Reading from NMI is a oneshot deal.
4882 if (raw_spin_trylock(&cpu_buffer->reader_lock))
4885 /* Continue without locking, but disable the ring buffer */
4886 atomic_inc(&cpu_buffer->record_disabled);
4891 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4894 raw_spin_unlock(&cpu_buffer->reader_lock);
4898 * ring_buffer_peek - peek at the next event to be read
4899 * @buffer: The ring buffer to read
4900 * @cpu: The cpu to peak at
4901 * @ts: The timestamp counter of this event.
4902 * @lost_events: a variable to store if events were lost (may be NULL)
4904 * This will return the event that will be read next, but does
4905 * not consume the data.
4907 struct ring_buffer_event *
4908 ring_buffer_peek(struct trace_buffer *buffer, int cpu, u64 *ts,
4909 unsigned long *lost_events)
4911 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4912 struct ring_buffer_event *event;
4913 unsigned long flags;
4916 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4920 local_irq_save(flags);
4921 dolock = rb_reader_lock(cpu_buffer);
4922 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4923 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4924 rb_advance_reader(cpu_buffer);
4925 rb_reader_unlock(cpu_buffer, dolock);
4926 local_irq_restore(flags);
4928 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4934 /** ring_buffer_iter_dropped - report if there are dropped events
4935 * @iter: The ring buffer iterator
4937 * Returns true if there was dropped events since the last peek.
4939 bool ring_buffer_iter_dropped(struct ring_buffer_iter *iter)
4941 bool ret = iter->missed_events != 0;
4943 iter->missed_events = 0;
4946 EXPORT_SYMBOL_GPL(ring_buffer_iter_dropped);
4949 * ring_buffer_iter_peek - peek at the next event to be read
4950 * @iter: The ring buffer iterator
4951 * @ts: The timestamp counter of this event.
4953 * This will return the event that will be read next, but does
4954 * not increment the iterator.
4956 struct ring_buffer_event *
4957 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4959 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4960 struct ring_buffer_event *event;
4961 unsigned long flags;
4964 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4965 event = rb_iter_peek(iter, ts);
4966 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4968 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4975 * ring_buffer_consume - return an event and consume it
4976 * @buffer: The ring buffer to get the next event from
4977 * @cpu: the cpu to read the buffer from
4978 * @ts: a variable to store the timestamp (may be NULL)
4979 * @lost_events: a variable to store if events were lost (may be NULL)
4981 * Returns the next event in the ring buffer, and that event is consumed.
4982 * Meaning, that sequential reads will keep returning a different event,
4983 * and eventually empty the ring buffer if the producer is slower.
4985 struct ring_buffer_event *
4986 ring_buffer_consume(struct trace_buffer *buffer, int cpu, u64 *ts,
4987 unsigned long *lost_events)
4989 struct ring_buffer_per_cpu *cpu_buffer;
4990 struct ring_buffer_event *event = NULL;
4991 unsigned long flags;
4995 /* might be called in atomic */
4998 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5001 cpu_buffer = buffer->buffers[cpu];
5002 local_irq_save(flags);
5003 dolock = rb_reader_lock(cpu_buffer);
5005 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
5007 cpu_buffer->lost_events = 0;
5008 rb_advance_reader(cpu_buffer);
5011 rb_reader_unlock(cpu_buffer, dolock);
5012 local_irq_restore(flags);
5017 if (event && event->type_len == RINGBUF_TYPE_PADDING)
5022 EXPORT_SYMBOL_GPL(ring_buffer_consume);
5025 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
5026 * @buffer: The ring buffer to read from
5027 * @cpu: The cpu buffer to iterate over
5028 * @flags: gfp flags to use for memory allocation
5030 * This performs the initial preparations necessary to iterate
5031 * through the buffer. Memory is allocated, buffer recording
5032 * is disabled, and the iterator pointer is returned to the caller.
5034 * Disabling buffer recording prevents the reading from being
5035 * corrupted. This is not a consuming read, so a producer is not
5038 * After a sequence of ring_buffer_read_prepare calls, the user is
5039 * expected to make at least one call to ring_buffer_read_prepare_sync.
5040 * Afterwards, ring_buffer_read_start is invoked to get things going
5043 * This overall must be paired with ring_buffer_read_finish.
5045 struct ring_buffer_iter *
5046 ring_buffer_read_prepare(struct trace_buffer *buffer, int cpu, gfp_t flags)
5048 struct ring_buffer_per_cpu *cpu_buffer;
5049 struct ring_buffer_iter *iter;
5051 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5054 iter = kzalloc(sizeof(*iter), flags);
5058 /* Holds the entire event: data and meta data */
5059 iter->event = kmalloc(BUF_PAGE_SIZE, flags);
5065 cpu_buffer = buffer->buffers[cpu];
5067 iter->cpu_buffer = cpu_buffer;
5069 atomic_inc(&cpu_buffer->resize_disabled);
5073 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
5076 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
5078 * All previously invoked ring_buffer_read_prepare calls to prepare
5079 * iterators will be synchronized. Afterwards, read_buffer_read_start
5080 * calls on those iterators are allowed.
5083 ring_buffer_read_prepare_sync(void)
5087 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
5090 * ring_buffer_read_start - start a non consuming read of the buffer
5091 * @iter: The iterator returned by ring_buffer_read_prepare
5093 * This finalizes the startup of an iteration through the buffer.
5094 * The iterator comes from a call to ring_buffer_read_prepare and
5095 * an intervening ring_buffer_read_prepare_sync must have been
5098 * Must be paired with ring_buffer_read_finish.
5101 ring_buffer_read_start(struct ring_buffer_iter *iter)
5103 struct ring_buffer_per_cpu *cpu_buffer;
5104 unsigned long flags;
5109 cpu_buffer = iter->cpu_buffer;
5111 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5112 arch_spin_lock(&cpu_buffer->lock);
5113 rb_iter_reset(iter);
5114 arch_spin_unlock(&cpu_buffer->lock);
5115 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5117 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
5120 * ring_buffer_read_finish - finish reading the iterator of the buffer
5121 * @iter: The iterator retrieved by ring_buffer_start
5123 * This re-enables the recording to the buffer, and frees the
5127 ring_buffer_read_finish(struct ring_buffer_iter *iter)
5129 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5130 unsigned long flags;
5133 * Ring buffer is disabled from recording, here's a good place
5134 * to check the integrity of the ring buffer.
5135 * Must prevent readers from trying to read, as the check
5136 * clears the HEAD page and readers require it.
5138 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5139 rb_check_pages(cpu_buffer);
5140 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5142 atomic_dec(&cpu_buffer->resize_disabled);
5146 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
5149 * ring_buffer_iter_advance - advance the iterator to the next location
5150 * @iter: The ring buffer iterator
5152 * Move the location of the iterator such that the next read will
5153 * be the next location of the iterator.
5155 void ring_buffer_iter_advance(struct ring_buffer_iter *iter)
5157 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5158 unsigned long flags;
5160 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5162 rb_advance_iter(iter);
5164 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5166 EXPORT_SYMBOL_GPL(ring_buffer_iter_advance);
5169 * ring_buffer_size - return the size of the ring buffer (in bytes)
5170 * @buffer: The ring buffer.
5171 * @cpu: The CPU to get ring buffer size from.
5173 unsigned long ring_buffer_size(struct trace_buffer *buffer, int cpu)
5176 * Earlier, this method returned
5177 * BUF_PAGE_SIZE * buffer->nr_pages
5178 * Since the nr_pages field is now removed, we have converted this to
5179 * return the per cpu buffer value.
5181 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5184 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
5186 EXPORT_SYMBOL_GPL(ring_buffer_size);
5188 static void rb_clear_buffer_page(struct buffer_page *page)
5190 local_set(&page->write, 0);
5191 local_set(&page->entries, 0);
5192 rb_init_page(page->page);
5197 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
5199 struct buffer_page *page;
5201 rb_head_page_deactivate(cpu_buffer);
5203 cpu_buffer->head_page
5204 = list_entry(cpu_buffer->pages, struct buffer_page, list);
5205 rb_clear_buffer_page(cpu_buffer->head_page);
5206 list_for_each_entry(page, cpu_buffer->pages, list) {
5207 rb_clear_buffer_page(page);
5210 cpu_buffer->tail_page = cpu_buffer->head_page;
5211 cpu_buffer->commit_page = cpu_buffer->head_page;
5213 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
5214 INIT_LIST_HEAD(&cpu_buffer->new_pages);
5215 rb_clear_buffer_page(cpu_buffer->reader_page);
5217 local_set(&cpu_buffer->entries_bytes, 0);
5218 local_set(&cpu_buffer->overrun, 0);
5219 local_set(&cpu_buffer->commit_overrun, 0);
5220 local_set(&cpu_buffer->dropped_events, 0);
5221 local_set(&cpu_buffer->entries, 0);
5222 local_set(&cpu_buffer->committing, 0);
5223 local_set(&cpu_buffer->commits, 0);
5224 local_set(&cpu_buffer->pages_touched, 0);
5225 local_set(&cpu_buffer->pages_lost, 0);
5226 local_set(&cpu_buffer->pages_read, 0);
5227 cpu_buffer->last_pages_touch = 0;
5228 cpu_buffer->shortest_full = 0;
5229 cpu_buffer->read = 0;
5230 cpu_buffer->read_bytes = 0;
5232 rb_time_set(&cpu_buffer->write_stamp, 0);
5233 rb_time_set(&cpu_buffer->before_stamp, 0);
5235 memset(cpu_buffer->event_stamp, 0, sizeof(cpu_buffer->event_stamp));
5237 cpu_buffer->lost_events = 0;
5238 cpu_buffer->last_overrun = 0;
5240 rb_head_page_activate(cpu_buffer);
5241 cpu_buffer->pages_removed = 0;
5244 /* Must have disabled the cpu buffer then done a synchronize_rcu */
5245 static void reset_disabled_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
5247 unsigned long flags;
5249 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5251 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
5254 arch_spin_lock(&cpu_buffer->lock);
5256 rb_reset_cpu(cpu_buffer);
5258 arch_spin_unlock(&cpu_buffer->lock);
5261 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5265 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
5266 * @buffer: The ring buffer to reset a per cpu buffer of
5267 * @cpu: The CPU buffer to be reset
5269 void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
5271 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5273 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5276 /* prevent another thread from changing buffer sizes */
5277 mutex_lock(&buffer->mutex);
5279 atomic_inc(&cpu_buffer->resize_disabled);
5280 atomic_inc(&cpu_buffer->record_disabled);
5282 /* Make sure all commits have finished */
5285 reset_disabled_cpu_buffer(cpu_buffer);
5287 atomic_dec(&cpu_buffer->record_disabled);
5288 atomic_dec(&cpu_buffer->resize_disabled);
5290 mutex_unlock(&buffer->mutex);
5292 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
5294 /* Flag to ensure proper resetting of atomic variables */
5295 #define RESET_BIT (1 << 30)
5298 * ring_buffer_reset_online_cpus - reset a ring buffer per CPU buffer
5299 * @buffer: The ring buffer to reset a per cpu buffer of
5301 void ring_buffer_reset_online_cpus(struct trace_buffer *buffer)
5303 struct ring_buffer_per_cpu *cpu_buffer;
5306 /* prevent another thread from changing buffer sizes */
5307 mutex_lock(&buffer->mutex);
5309 for_each_online_buffer_cpu(buffer, cpu) {
5310 cpu_buffer = buffer->buffers[cpu];
5312 atomic_add(RESET_BIT, &cpu_buffer->resize_disabled);
5313 atomic_inc(&cpu_buffer->record_disabled);
5316 /* Make sure all commits have finished */
5319 for_each_buffer_cpu(buffer, cpu) {
5320 cpu_buffer = buffer->buffers[cpu];
5323 * If a CPU came online during the synchronize_rcu(), then
5326 if (!(atomic_read(&cpu_buffer->resize_disabled) & RESET_BIT))
5329 reset_disabled_cpu_buffer(cpu_buffer);
5331 atomic_dec(&cpu_buffer->record_disabled);
5332 atomic_sub(RESET_BIT, &cpu_buffer->resize_disabled);
5335 mutex_unlock(&buffer->mutex);
5339 * ring_buffer_reset - reset a ring buffer
5340 * @buffer: The ring buffer to reset all cpu buffers
5342 void ring_buffer_reset(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_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_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);
5371 EXPORT_SYMBOL_GPL(ring_buffer_reset);
5374 * ring_buffer_empty - is the ring buffer empty?
5375 * @buffer: The ring buffer to test
5377 bool ring_buffer_empty(struct trace_buffer *buffer)
5379 struct ring_buffer_per_cpu *cpu_buffer;
5380 unsigned long flags;
5385 /* yes this is racy, but if you don't like the race, lock the buffer */
5386 for_each_buffer_cpu(buffer, cpu) {
5387 cpu_buffer = buffer->buffers[cpu];
5388 local_irq_save(flags);
5389 dolock = rb_reader_lock(cpu_buffer);
5390 ret = rb_per_cpu_empty(cpu_buffer);
5391 rb_reader_unlock(cpu_buffer, dolock);
5392 local_irq_restore(flags);
5400 EXPORT_SYMBOL_GPL(ring_buffer_empty);
5403 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
5404 * @buffer: The ring buffer
5405 * @cpu: The CPU buffer to test
5407 bool ring_buffer_empty_cpu(struct trace_buffer *buffer, int cpu)
5409 struct ring_buffer_per_cpu *cpu_buffer;
5410 unsigned long flags;
5414 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5417 cpu_buffer = buffer->buffers[cpu];
5418 local_irq_save(flags);
5419 dolock = rb_reader_lock(cpu_buffer);
5420 ret = rb_per_cpu_empty(cpu_buffer);
5421 rb_reader_unlock(cpu_buffer, dolock);
5422 local_irq_restore(flags);
5426 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
5428 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
5430 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
5431 * @buffer_a: One buffer to swap with
5432 * @buffer_b: The other buffer to swap with
5433 * @cpu: the CPU of the buffers to swap
5435 * This function is useful for tracers that want to take a "snapshot"
5436 * of a CPU buffer and has another back up buffer lying around.
5437 * it is expected that the tracer handles the cpu buffer not being
5438 * used at the moment.
5440 int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
5441 struct trace_buffer *buffer_b, int cpu)
5443 struct ring_buffer_per_cpu *cpu_buffer_a;
5444 struct ring_buffer_per_cpu *cpu_buffer_b;
5447 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
5448 !cpumask_test_cpu(cpu, buffer_b->cpumask))
5451 cpu_buffer_a = buffer_a->buffers[cpu];
5452 cpu_buffer_b = buffer_b->buffers[cpu];
5454 /* At least make sure the two buffers are somewhat the same */
5455 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
5460 if (atomic_read(&buffer_a->record_disabled))
5463 if (atomic_read(&buffer_b->record_disabled))
5466 if (atomic_read(&cpu_buffer_a->record_disabled))
5469 if (atomic_read(&cpu_buffer_b->record_disabled))
5473 * We can't do a synchronize_rcu here because this
5474 * function can be called in atomic context.
5475 * Normally this will be called from the same CPU as cpu.
5476 * If not it's up to the caller to protect this.
5478 atomic_inc(&cpu_buffer_a->record_disabled);
5479 atomic_inc(&cpu_buffer_b->record_disabled);
5482 if (local_read(&cpu_buffer_a->committing))
5484 if (local_read(&cpu_buffer_b->committing))
5488 * When resize is in progress, we cannot swap it because
5489 * it will mess the state of the cpu buffer.
5491 if (atomic_read(&buffer_a->resizing))
5493 if (atomic_read(&buffer_b->resizing))
5496 buffer_a->buffers[cpu] = cpu_buffer_b;
5497 buffer_b->buffers[cpu] = cpu_buffer_a;
5499 cpu_buffer_b->buffer = buffer_a;
5500 cpu_buffer_a->buffer = buffer_b;
5505 atomic_dec(&cpu_buffer_a->record_disabled);
5506 atomic_dec(&cpu_buffer_b->record_disabled);
5510 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
5511 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
5514 * ring_buffer_alloc_read_page - allocate a page to read from buffer
5515 * @buffer: the buffer to allocate for.
5516 * @cpu: the cpu buffer to allocate.
5518 * This function is used in conjunction with ring_buffer_read_page.
5519 * When reading a full page from the ring buffer, these functions
5520 * can be used to speed up the process. The calling function should
5521 * allocate a few pages first with this function. Then when it
5522 * needs to get pages from the ring buffer, it passes the result
5523 * of this function into ring_buffer_read_page, which will swap
5524 * the page that was allocated, with the read page of the buffer.
5527 * The page allocated, or ERR_PTR
5529 void *ring_buffer_alloc_read_page(struct trace_buffer *buffer, int cpu)
5531 struct ring_buffer_per_cpu *cpu_buffer;
5532 struct buffer_data_page *bpage = NULL;
5533 unsigned long flags;
5536 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5537 return ERR_PTR(-ENODEV);
5539 cpu_buffer = buffer->buffers[cpu];
5540 local_irq_save(flags);
5541 arch_spin_lock(&cpu_buffer->lock);
5543 if (cpu_buffer->free_page) {
5544 bpage = cpu_buffer->free_page;
5545 cpu_buffer->free_page = NULL;
5548 arch_spin_unlock(&cpu_buffer->lock);
5549 local_irq_restore(flags);
5554 page = alloc_pages_node(cpu_to_node(cpu),
5555 GFP_KERNEL | __GFP_NORETRY, 0);
5557 return ERR_PTR(-ENOMEM);
5559 bpage = page_address(page);
5562 rb_init_page(bpage);
5566 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
5569 * ring_buffer_free_read_page - free an allocated read page
5570 * @buffer: the buffer the page was allocate for
5571 * @cpu: the cpu buffer the page came from
5572 * @data: the page to free
5574 * Free a page allocated from ring_buffer_alloc_read_page.
5576 void ring_buffer_free_read_page(struct trace_buffer *buffer, int cpu, void *data)
5578 struct ring_buffer_per_cpu *cpu_buffer;
5579 struct buffer_data_page *bpage = data;
5580 struct page *page = virt_to_page(bpage);
5581 unsigned long flags;
5583 if (!buffer || !buffer->buffers || !buffer->buffers[cpu])
5586 cpu_buffer = buffer->buffers[cpu];
5588 /* If the page is still in use someplace else, we can't reuse it */
5589 if (page_ref_count(page) > 1)
5592 local_irq_save(flags);
5593 arch_spin_lock(&cpu_buffer->lock);
5595 if (!cpu_buffer->free_page) {
5596 cpu_buffer->free_page = bpage;
5600 arch_spin_unlock(&cpu_buffer->lock);
5601 local_irq_restore(flags);
5604 free_page((unsigned long)bpage);
5606 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
5609 * ring_buffer_read_page - extract a page from the ring buffer
5610 * @buffer: buffer to extract from
5611 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
5612 * @len: amount to extract
5613 * @cpu: the cpu of the buffer to extract
5614 * @full: should the extraction only happen when the page is full.
5616 * This function will pull out a page from the ring buffer and consume it.
5617 * @data_page must be the address of the variable that was returned
5618 * from ring_buffer_alloc_read_page. This is because the page might be used
5619 * to swap with a page in the ring buffer.
5622 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
5623 * if (IS_ERR(rpage))
5624 * return PTR_ERR(rpage);
5625 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
5627 * process_page(rpage, ret);
5629 * When @full is set, the function will not return true unless
5630 * the writer is off the reader page.
5632 * Note: it is up to the calling functions to handle sleeps and wakeups.
5633 * The ring buffer can be used anywhere in the kernel and can not
5634 * blindly call wake_up. The layer that uses the ring buffer must be
5635 * responsible for that.
5638 * >=0 if data has been transferred, returns the offset of consumed data.
5639 * <0 if no data has been transferred.
5641 int ring_buffer_read_page(struct trace_buffer *buffer,
5642 void **data_page, size_t len, int cpu, int full)
5644 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5645 struct ring_buffer_event *event;
5646 struct buffer_data_page *bpage;
5647 struct buffer_page *reader;
5648 unsigned long missed_events;
5649 unsigned long flags;
5650 unsigned int commit;
5655 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5659 * If len is not big enough to hold the page header, then
5660 * we can not copy anything.
5662 if (len <= BUF_PAGE_HDR_SIZE)
5665 len -= BUF_PAGE_HDR_SIZE;
5674 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5676 reader = rb_get_reader_page(cpu_buffer);
5680 event = rb_reader_event(cpu_buffer);
5682 read = reader->read;
5683 commit = rb_page_commit(reader);
5685 /* Check if any events were dropped */
5686 missed_events = cpu_buffer->lost_events;
5689 * If this page has been partially read or
5690 * if len is not big enough to read the rest of the page or
5691 * a writer is still on the page, then
5692 * we must copy the data from the page to the buffer.
5693 * Otherwise, we can simply swap the page with the one passed in.
5695 if (read || (len < (commit - read)) ||
5696 cpu_buffer->reader_page == cpu_buffer->commit_page) {
5697 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
5698 unsigned int rpos = read;
5699 unsigned int pos = 0;
5703 * If a full page is expected, this can still be returned
5704 * if there's been a previous partial read and the
5705 * rest of the page can be read and the commit page is off
5709 (!read || (len < (commit - read)) ||
5710 cpu_buffer->reader_page == cpu_buffer->commit_page))
5713 if (len > (commit - read))
5714 len = (commit - read);
5716 /* Always keep the time extend and data together */
5717 size = rb_event_ts_length(event);
5722 /* save the current timestamp, since the user will need it */
5723 save_timestamp = cpu_buffer->read_stamp;
5725 /* Need to copy one event at a time */
5727 /* We need the size of one event, because
5728 * rb_advance_reader only advances by one event,
5729 * whereas rb_event_ts_length may include the size of
5730 * one or two events.
5731 * We have already ensured there's enough space if this
5732 * is a time extend. */
5733 size = rb_event_length(event);
5734 memcpy(bpage->data + pos, rpage->data + rpos, size);
5738 rb_advance_reader(cpu_buffer);
5739 rpos = reader->read;
5745 event = rb_reader_event(cpu_buffer);
5746 /* Always keep the time extend and data together */
5747 size = rb_event_ts_length(event);
5748 } while (len >= size);
5751 local_set(&bpage->commit, pos);
5752 bpage->time_stamp = save_timestamp;
5754 /* we copied everything to the beginning */
5757 /* update the entry counter */
5758 cpu_buffer->read += rb_page_entries(reader);
5759 cpu_buffer->read_bytes += rb_page_commit(reader);
5761 /* swap the pages */
5762 rb_init_page(bpage);
5763 bpage = reader->page;
5764 reader->page = *data_page;
5765 local_set(&reader->write, 0);
5766 local_set(&reader->entries, 0);
5771 * Use the real_end for the data size,
5772 * This gives us a chance to store the lost events
5775 if (reader->real_end)
5776 local_set(&bpage->commit, reader->real_end);
5780 cpu_buffer->lost_events = 0;
5782 commit = local_read(&bpage->commit);
5784 * Set a flag in the commit field if we lost events
5786 if (missed_events) {
5787 /* If there is room at the end of the page to save the
5788 * missed events, then record it there.
5790 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
5791 memcpy(&bpage->data[commit], &missed_events,
5792 sizeof(missed_events));
5793 local_add(RB_MISSED_STORED, &bpage->commit);
5794 commit += sizeof(missed_events);
5796 local_add(RB_MISSED_EVENTS, &bpage->commit);
5800 * This page may be off to user land. Zero it out here.
5802 if (commit < BUF_PAGE_SIZE)
5803 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
5806 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5811 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
5814 * We only allocate new buffers, never free them if the CPU goes down.
5815 * If we were to free the buffer, then the user would lose any trace that was in
5818 int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
5820 struct trace_buffer *buffer;
5823 unsigned long nr_pages;
5825 buffer = container_of(node, struct trace_buffer, node);
5826 if (cpumask_test_cpu(cpu, buffer->cpumask))
5831 /* check if all cpu sizes are same */
5832 for_each_buffer_cpu(buffer, cpu_i) {
5833 /* fill in the size from first enabled cpu */
5835 nr_pages = buffer->buffers[cpu_i]->nr_pages;
5836 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
5841 /* allocate minimum pages, user can later expand it */
5844 buffer->buffers[cpu] =
5845 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
5846 if (!buffer->buffers[cpu]) {
5847 WARN(1, "failed to allocate ring buffer on CPU %u\n",
5852 cpumask_set_cpu(cpu, buffer->cpumask);
5856 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
5858 * This is a basic integrity check of the ring buffer.
5859 * Late in the boot cycle this test will run when configured in.
5860 * It will kick off a thread per CPU that will go into a loop
5861 * writing to the per cpu ring buffer various sizes of data.
5862 * Some of the data will be large items, some small.
5864 * Another thread is created that goes into a spin, sending out
5865 * IPIs to the other CPUs to also write into the ring buffer.
5866 * this is to test the nesting ability of the buffer.
5868 * Basic stats are recorded and reported. If something in the
5869 * ring buffer should happen that's not expected, a big warning
5870 * is displayed and all ring buffers are disabled.
5872 static struct task_struct *rb_threads[NR_CPUS] __initdata;
5874 struct rb_test_data {
5875 struct trace_buffer *buffer;
5876 unsigned long events;
5877 unsigned long bytes_written;
5878 unsigned long bytes_alloc;
5879 unsigned long bytes_dropped;
5880 unsigned long events_nested;
5881 unsigned long bytes_written_nested;
5882 unsigned long bytes_alloc_nested;
5883 unsigned long bytes_dropped_nested;
5884 int min_size_nested;
5885 int max_size_nested;
5892 static struct rb_test_data rb_data[NR_CPUS] __initdata;
5895 #define RB_TEST_BUFFER_SIZE 1048576
5897 static char rb_string[] __initdata =
5898 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
5899 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
5900 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
5902 static bool rb_test_started __initdata;
5909 static __init int rb_write_something(struct rb_test_data *data, bool nested)
5911 struct ring_buffer_event *event;
5912 struct rb_item *item;
5919 /* Have nested writes different that what is written */
5920 cnt = data->cnt + (nested ? 27 : 0);
5922 /* Multiply cnt by ~e, to make some unique increment */
5923 size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
5925 len = size + sizeof(struct rb_item);
5927 started = rb_test_started;
5928 /* read rb_test_started before checking buffer enabled */
5931 event = ring_buffer_lock_reserve(data->buffer, len);
5933 /* Ignore dropped events before test starts. */
5936 data->bytes_dropped += len;
5938 data->bytes_dropped_nested += len;
5943 event_len = ring_buffer_event_length(event);
5945 if (RB_WARN_ON(data->buffer, event_len < len))
5948 item = ring_buffer_event_data(event);
5950 memcpy(item->str, rb_string, size);
5953 data->bytes_alloc_nested += event_len;
5954 data->bytes_written_nested += len;
5955 data->events_nested++;
5956 if (!data->min_size_nested || len < data->min_size_nested)
5957 data->min_size_nested = len;
5958 if (len > data->max_size_nested)
5959 data->max_size_nested = len;
5961 data->bytes_alloc += event_len;
5962 data->bytes_written += len;
5964 if (!data->min_size || len < data->min_size)
5965 data->max_size = len;
5966 if (len > data->max_size)
5967 data->max_size = len;
5971 ring_buffer_unlock_commit(data->buffer);
5976 static __init int rb_test(void *arg)
5978 struct rb_test_data *data = arg;
5980 while (!kthread_should_stop()) {
5981 rb_write_something(data, false);
5984 set_current_state(TASK_INTERRUPTIBLE);
5985 /* Now sleep between a min of 100-300us and a max of 1ms */
5986 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
5992 static __init void rb_ipi(void *ignore)
5994 struct rb_test_data *data;
5995 int cpu = smp_processor_id();
5997 data = &rb_data[cpu];
5998 rb_write_something(data, true);
6001 static __init int rb_hammer_test(void *arg)
6003 while (!kthread_should_stop()) {
6005 /* Send an IPI to all cpus to write data! */
6006 smp_call_function(rb_ipi, NULL, 1);
6007 /* No sleep, but for non preempt, let others run */
6014 static __init int test_ringbuffer(void)
6016 struct task_struct *rb_hammer;
6017 struct trace_buffer *buffer;
6021 if (security_locked_down(LOCKDOWN_TRACEFS)) {
6022 pr_warn("Lockdown is enabled, skipping ring buffer tests\n");
6026 pr_info("Running ring buffer tests...\n");
6028 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
6029 if (WARN_ON(!buffer))
6032 /* Disable buffer so that threads can't write to it yet */
6033 ring_buffer_record_off(buffer);
6035 for_each_online_cpu(cpu) {
6036 rb_data[cpu].buffer = buffer;
6037 rb_data[cpu].cpu = cpu;
6038 rb_data[cpu].cnt = cpu;
6039 rb_threads[cpu] = kthread_run_on_cpu(rb_test, &rb_data[cpu],
6040 cpu, "rbtester/%u");
6041 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
6042 pr_cont("FAILED\n");
6043 ret = PTR_ERR(rb_threads[cpu]);
6048 /* Now create the rb hammer! */
6049 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
6050 if (WARN_ON(IS_ERR(rb_hammer))) {
6051 pr_cont("FAILED\n");
6052 ret = PTR_ERR(rb_hammer);
6056 ring_buffer_record_on(buffer);
6058 * Show buffer is enabled before setting rb_test_started.
6059 * Yes there's a small race window where events could be
6060 * dropped and the thread wont catch it. But when a ring
6061 * buffer gets enabled, there will always be some kind of
6062 * delay before other CPUs see it. Thus, we don't care about
6063 * those dropped events. We care about events dropped after
6064 * the threads see that the buffer is active.
6067 rb_test_started = true;
6069 set_current_state(TASK_INTERRUPTIBLE);
6070 /* Just run for 10 seconds */;
6071 schedule_timeout(10 * HZ);
6073 kthread_stop(rb_hammer);
6076 for_each_online_cpu(cpu) {
6077 if (!rb_threads[cpu])
6079 kthread_stop(rb_threads[cpu]);
6082 ring_buffer_free(buffer);
6087 pr_info("finished\n");
6088 for_each_online_cpu(cpu) {
6089 struct ring_buffer_event *event;
6090 struct rb_test_data *data = &rb_data[cpu];
6091 struct rb_item *item;
6092 unsigned long total_events;
6093 unsigned long total_dropped;
6094 unsigned long total_written;
6095 unsigned long total_alloc;
6096 unsigned long total_read = 0;
6097 unsigned long total_size = 0;
6098 unsigned long total_len = 0;
6099 unsigned long total_lost = 0;
6102 int small_event_size;
6106 total_events = data->events + data->events_nested;
6107 total_written = data->bytes_written + data->bytes_written_nested;
6108 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
6109 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
6111 big_event_size = data->max_size + data->max_size_nested;
6112 small_event_size = data->min_size + data->min_size_nested;
6114 pr_info("CPU %d:\n", cpu);
6115 pr_info(" events: %ld\n", total_events);
6116 pr_info(" dropped bytes: %ld\n", total_dropped);
6117 pr_info(" alloced bytes: %ld\n", total_alloc);
6118 pr_info(" written bytes: %ld\n", total_written);
6119 pr_info(" biggest event: %d\n", big_event_size);
6120 pr_info(" smallest event: %d\n", small_event_size);
6122 if (RB_WARN_ON(buffer, total_dropped))
6127 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
6129 item = ring_buffer_event_data(event);
6130 total_len += ring_buffer_event_length(event);
6131 total_size += item->size + sizeof(struct rb_item);
6132 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
6133 pr_info("FAILED!\n");
6134 pr_info("buffer had: %.*s\n", item->size, item->str);
6135 pr_info("expected: %.*s\n", item->size, rb_string);
6136 RB_WARN_ON(buffer, 1);
6147 pr_info(" read events: %ld\n", total_read);
6148 pr_info(" lost events: %ld\n", total_lost);
6149 pr_info(" total events: %ld\n", total_lost + total_read);
6150 pr_info(" recorded len bytes: %ld\n", total_len);
6151 pr_info(" recorded size bytes: %ld\n", total_size);
6153 pr_info(" With dropped events, record len and size may not match\n"
6154 " alloced and written from above\n");
6156 if (RB_WARN_ON(buffer, total_len != total_alloc ||
6157 total_size != total_written))
6160 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
6166 pr_info("Ring buffer PASSED!\n");
6168 ring_buffer_free(buffer);
6172 late_initcall(test_ringbuffer);
6173 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */